Open AccessArticle

Proposal and Implementation of an Integrated Monitoring Platform for Preventive Maintenance of Industrial Machines

Department of Computer Science, Nagoya Institute of Technology, Gokiso-cho, Showa-ku, Nagoya 466-8555, Aichi, Japan

Author to whom correspondence should be addressed.

Appl. Sci. 2024, 14(24), 11534; https://doi.org/10.3390/app142411534

Submission received: 12 November 2024 / Revised: 29 November 2024 / Accepted: 6 December 2024 / Published: 11 December 2024

(This article belongs to the Special Issue Industrial IoT: From Theory to Applications)

Download

Browse Figures

Versions Notes

Abstract

In order to realise the efficient maintenance of industrial machines, Small and Medium-sized Enterprises (SMEs) need a system that utilises digital technology to handle everything from data collection to the visualisation of the collected data in an integrated manner. In this paper, an integrated monitoring platform using external sensor devices is proposed and implemented for the purpose of preventive maintenance of industrial machines. The proposed system performs edge processing to calculate features effective for monitoring on the sensor device, collects only the obtained features, and visualises them on a web server. In order to determine the features required by edge processing, a cycle waveform cut-out algorithm was proposed. As an evaluation experiment, the proposed system was used to detect the loosening of bolts on the support side of a ball screw. The results of the analysis showed that the dispersion value immediately after the start of uniform motion from the right end to the left end was valid, so the system was implemented as edge processing in the sensor device. In wireless transmission experiments on a testbed, an average of 20 consecutive cycles were used to achieve a 99.9% correct response rate and high detection accuracy, demonstrating the usefulness of the proposed system.

Keywords:

preventive maintenance; Internet of Things (IoT); industrial machines

1. Introduction

Industry 4.0 is an industrial policy announced by the German government in 2011 [1]. It aims to reform the manufacturing industry by incorporating Information Technology (IT). The introduction of Industry 4.0 is a new way of using IT, such as Internet of Things (IoT) and big data analysis, to improve productivity efficiency and reduce the loss of raw materials, energy, and other resources. This will lead to increased efficiency in productivity and reduced loss of raw materials and energy. A similar evolution is occurring in asset management, which is called Maintenance 4.0. Maintenance 4.0 aims to digitize the maintenance process and automate the process of predicting and repairing equipment failures. However, even if Industry 4.0 and Maintenance 4.0 are introduced, their effects cannot be maximised if the Information and Communication Technology operating environment, which is the basis of technological innovation, is not in place.

The manufacturing industry accounts for 20.6% of Japan’s Gross Domestic Product (GDP) by sector. It is the second most important industry in terms of share of GDP, following the service industry [2]. In addition, the manufacturing industry supports Japanese industry with high productivity, as its nominal labour productivity per worker has been higher than the average for all industries in all years from 2011 to 2021. Thus, manufacturing enterprises, which are a major industry in Japan, are classified into three categories according to their scale: major enterprises, medium-sized enterprises and small enterprises. According to the Small and Medium-sized Enterprise Basic Act, Small and Medium-sized Enterprises (SMEs) are those with a capital of 300 million yen or less or with 300 or fewer employees, of which those with 20 or fewer employees are considered small businesses. In terms of the composition of the number of these enterprises, major enterprises account for only 0.3% of the total, while medium-sized enterprises account for 14.8% and small enterprises for 84.9%, making SMEs account for 99.7% of the total number of enterprises [3]. Furthermore, SMEs account for 68.8% of the total number of employees and 52.9% of the total value added. From these facts, it can be said that SMEs play an important role in the Japanese manufacturing industry.

The deterioration of equipment is a problem for SMEs, which support Japan’s manufacturing industry. A survey was conducted on the age of equipment by company size, with the index of equipment age as of FY1990 set at 100. According to this, as of 2017, the index for major enterprises was 148.0, while the index for SMEs was 191.8, a significant difference, indicating that the deterioration of equipment in SMEs is more serious than in major enterprises [4]. In addition, 36.5% of SMEs cited equipment failure as a risk that would make it difficult to continue operations. This indicates that the issue is how to maintain and continue to use equipment [5].

Furthermore, we compare capital investment by size in 2022 separately for tangible fixed assets and intangible fixed assets. With regard to tangible fixed assets, 94.6% of major enterprises and 76.6% of SMEs have invested in tangible fixed assets. Many SMEs invest in property, plant and equipment, although the proportion is smaller in SMEs [2]. Regarding the purpose of investment in tangible fixed assets, the highest proportion of investment in 2022 is for the renewal and reinforcement of obsolete equipment, at 64.3% of the total. Here, it is also possible to read the situation where the response to the deterioration of equipment is becoming more important [2]. In contrast, for intangible fixed assets in terms of capital investment by size in 2022, the figure for major enterprises is 83.1%, compared to 42.6% for SMEs. SMEs are not investing in intangible fixed assets, creating a large gap between major enterprises and SMEs [2]. In terms of the purpose of investment in intangible fixed assets in the manufacturing industry, the highest proportion of investment in 2022 will be to improve operational efficiency and reduce costs at 39.4%. This was followed by the renewal and maintenance of old systems at 35.4%, again showing the importance of measures against ageing systems. Digital Transformation (DX) related to the IoT in factories, for example, comes third at 24.9%, an increase of 5.8% compared to 2020 [2]. DX is defined by the Ministry of Economy, Trade and Industry, Trade and Industry in its Digital Governance Code as the use of data and digital technology to transform products, business models, operations and organisations to establish a competitive advantage [6]. Therefore, it can be said that although SMEs are required to utilise digital technology, capital investment in intangible fixed assets, including software, has not progressed.

There are also issues regarding the utilisation of data collected through IoT systems and day-to-day operations. In terms of the implementation status of data collection and utilisation in relation to business, 86.2% of major enterprises are doing so. In contrast, only 59.0% of SMEs do so, indicating that data utilisation is less advanced in SMEs than in major enterprises [2]. In addition, even among SMEs that have implemented data utilisation, only 34.7% have implemented the utilisation of collected data across departments and offices. However, the total of companies that are preparing or considering the use of data and those who have not yet done so but understand the need to do so amount to 92.6%, indicating that the use of data across divisions and offices is necessary even for companies that are already using data. Furthermore, only 13.7% of all companies have implemented data linkage and utilisation across other companies and industries. However, the total of companies that are preparing or considering it and companies that have not yet implemented it but understand the need for it is 70.5%, indicating that data linkage across other companies and industries is needed as well as between departments and business sites. Therefore, it can be said that a support system is needed for the collected data and its utilisation within and between companies.

Investment in digital technology also differs between major enterprises and SMEs; 89.5% of major enterprises in the manufacturing sector invest in IT, whereas only 47.8% of SMEs do so.

A survey was conducted to investigate the reasons for the lack of investment in IT and challenges in data utilisation in SMEs. In SMEs, the lack of human resources with the necessary skills for data integration was the most common reason, at 58.3%. This was followed by a lack of company-wide knowledge of data integration at 43.4%, which was the second highest, and the cost of renewing existing systems at 39.7%, which was the third highest. These responses indicate that SMEs lack the knowledge and human resources to build a system for data federation, and the cost of replacing existing systems is a further challenge. In order to solve these problems, it is necessary to have a system that can handle everything from data collection to analysis consistently in the same system, does not require complex knowledge for construction, and can be retrofitted to existing systems without having to replace existing systems.

In summary, it was found that the risk of breakdowns in SMEs in the manufacturing industry is increasing due to the deterioration of industrial machines, and although maintenance is becoming more important, the introduction of digital technology is not as advanced as in major enterprises. The main objective of this research is to contribute to improving the productivity of SMEs in the manufacturing industry through an integrated platform that enables preventive maintenance using data. In this paper, an integrated platform for the preventive maintenance of industrial machines is developed to achieve this major objective.

It aims to realise the maintenance of industrial machines and quality control of products by utilising digital technology in SMEs through a monitoring system that covers everything from data collection using sensor devices to data analysis and data visualisation. Furthermore, the system aims to reduce the burden on workers and support the trial-and-error process of IoT utilisation in SMEs by integrating the management of sensor devices, machines to which sensors are installed and the installation positions.

The rest of the paper is structured as follows: Section 2 provides related works; Section 3 presents the proposed system and device configuration; Section 4 presents the experimental results and evaluation using the testbed; Section 5 discusses the effectiveness of the proposed methodology and finally, Section 6 concludes the paper and suggests prospects for future research.

2. Related Works

Today, systems for monitoring industrial machines are sold and provided by companies that manufacture sensors and industrial machines. The common feature of these systems is that they collect data from industrial machines and inform the user of the occurrence or signs of failure. The differences are that some systems use dedicated external sensor devices for data acquisition, while others have data acquisition functions on the industrial machine itself.

An example of a monitoring system using a dedicated external sensor device is the ACOUS NAVI [7] from NSK Ltd. (Tokyo, Japan) [8]. The system uses a sensor device with a vibration pick-up to diagnose industrial machine components such as bearing flaws, ball screw wear and flaking in linear guides using vibration data.

The main functions of the sensor devices used in this system are listed below. Bluetooth connection: The sensor device has Bluetooth functionality and the measured data and the features obtained from the analysis can be acquired and displayed via an application on an Android smartphone.

Battery-powered: It is equipped with a rechargeable lithium-ion battery and can be recharged using the Universal Serial Bus (USB) cable and USB charging adaptor supplied, enabling a continuous standby time of 8 h from a fully charged state. This allows vibration measurements to be made even where a power supply for the measurement device is not available.

Multiple analysis modes: Multiple analysis results can be displayed in O.A. mode, which displays basic statistics such as RMS and peak values of vibration data, FFT mode, which analyses frequency-domain features of vibration acceleration, and WAVE mode, which displays vibration acceleration signals as time-domain waveforms and applications.

One advantage of this system is that data collection and power supply are wireless due to the combination of Bluetooth connectivity and battery operation so that data measurement and analysis can be easily carried out without any prior preparation. However, the battery lasts for only eight hours in standby mode, so continuous data collection over a long period of time, such as a day or a month, is not possible. In this system, the sensor device is made lightweight and compact, weighing 170 g, and can be easily attached to and detached from machines by means of magnetic fixation. By setting the patrol route in advance, i.e., which machines to visit and in what order to measure data, the smartphone application can perform a series of data collection tasks without any additional operations. In this system, the data measured by the sensor devices are collected and stored on the server via the smartphone. The data stored on the server can be visualised using management software for PCs. This allows the operator to check how the analysis results are changing. One of the challenges of this system is how to aggregate the measurement data. Although the data measurement itself is easy with wireless sensor devices, it is necessary for the operator to go to the target machine, install the sensor device and operate the smartphone application each time the data are measured, making the system a burden on the operator for data collection.

Another monitoring system that uses similar external sensor devices is THK Corporation’s (Tokyo, Japan) [9] OMNI edge [10]. In this system, specially designed sensors are attached to linear moving parts such as Linear Motion Guides and ball screws; the sensors quantify damage and lubrication conditions, and the data are stored on a server using a Long Term Evolution (LTE) connection. The collected data are analysed by the predictive detection software and visualised as an abnormality score. The system features pre- and post-introduction support. In the pre-introduction phase, the system provides the sensor modules and attachments required for proof of concept a dedicated amplifier with built-in edge Artificial Intelligence (AI), a gateway for data transmission, a Subscriber Identity Module Card (SIM) card and a web application in a single package. This eliminates the need to select sensors and edge devices, build networks and study data processing methods. In addition, support is provided for the selection of the installation model, installation location, on-site confirmation and radio wave survey. After installation, advice on analysing acquired data and reports on AI diagnostic services are available. This support is provided by dedicated technical staff. Although it is an advantage to be able to receive extensive support before and after the introduction of the system, there are only a few items that the company introducing the system actually performs, making it an outsourcing of IoT. Therefore, the system is not suitable for SMEs to collect and analyse data from industrial machines on a trial basis.

Hiruta et al. proposed a system for detecting insufficient grease in the bearings of the three-phase motors used in industrial machines [11]. In this system, the power spectrum is calculated from the current waveform of the motor, grouped like histogram bins, and then the normal state of the motor is learnt by using a Gaussian mixture model (GMM). This allows the likelihood to be calculated from the GMM and thus the deviation from the normal state can be evaluated. In conventional systems, the specifications and state of the bearing are taken into account to construct an abnormality detection model, and data in the normal state and data in the abnormal state are required as teacher data, respectively. However, to collect data on abnormal bearing conditions at actual production sites, it is necessary to keep the machine running until a failure occurs, which is impractical. In this research, an unsupervised learning method was proposed to detect bearing faults by analysing only the data in the normal state.

In this system, clamp-on current sensors were installed on each of the three wires between the inverter and the three-phase motor, and data were measured using a data logger. Data were collected for two different cases: one with a normal bearing and one in which the grease in the bearing had been removed by cleaning it with acetone. The power spectrum (0 to 500 Hz) of the current waveform obtained from the experiment is shown in Figure 1, and the log-likelihood obtained from the GMM is shown in Figure 2.

A comparison of the power spectra of the measured data shows that, in the normal case, the values are locally higher in the 50 Hz part. In contrast, in the acetone-cleaned case, the spectra were detected separately in the 50 Hz ± 10 Hz and ± 15 Hz ranges. The log-likelihood calculated from the GMM also shows that the values for the acetone cleaning case are larger than those for the normal case, indicating that a threshold can be used for detection. In this research, expensive data loggers were used to measure data from clamp-on current sensors attached to three-phase motors. Therefore, multiple data loggers are required when this system is installed in the factories of SMEs with multiple industrial machines, making the installation cost an issue.

3. Materials and Methods

3.1. System Architecture

The system overview of the integrated monitoring platform proposed in this research is shown in Figure 3.

In this system, sensor devices attached to industrial machines are used to collect data on the industrial machines being manufactured. The collected data are sent to the gateway using LoRa communication, a type of Low Power Wide Area (LPWA), after edge processing at the sensor device. The gateway sends the packets received from the sensor device to the web server by Hypertext Transfer Protocol (HTTP) POST. The web server visualises the POSTed data in the form of a graph, linking it to the sensor device information. The operator can check the status of the equipment from the graph, and if any abnormality is found, the operator can provide feedback by maintaining the machine or changing the production conditions.

Here, the data sent by the sensor device to the gateway are not the data as measured, but data that have been processed by edge processing. This edge processing calculates features for estimating the state of industrial machines from the data obtained from the sensors. To determine the content of edge processing, the measured data themselves must be analysed once. It is not suitable for the sensor device to send all the measured data to the gateway due to the communication capacity limitation of LoRa communication. Therefore, it is necessary to reduce the amount of data by adding some data processing to the measured data as edge processing in the sensor device and then transmitting the data. Based on the above, the operational phase of the integrated platform is defined in this system by dividing it into three phases: sensor data collection phase, sensor data analysis phase and industrial machine monitoring phase.

In the sensor data collection phase, external sensor devices are used to collect data. This is because SMEs in the manufacturing industry tend to have legacy equipment, and legacy industrial machines without network functionality are still in use. The amount of data acquired per unit of time depends on the sensor used and the sampling rate. As these different data cannot be handled in one format, three data formats are defined for this system: Comma Separated Values (CSV) files, sensor data files and radio packets.

A CSV file is a format in which data are recorded in comma-separated lines, where one line is the data for one measurement. Because a single line can have multiple columns separated by commas, multiple sensor data can be stored together in a single file.

A sensor data file is a file in which the data from a single measurement, such as sound data, image data or video data, are written in one file according to the respective file protocol. It is not possible to include data from several sensors in one file, such as a CSV file.

Wireless packets are packets sent by the sensor device to the gateway via LoRa communication. In the LoRa communication used in this system, the payload of a packet transmitted at a time should be limited to about 50 [B] to ensure stable communication. For the transmission frequency, the transmission interval should be at least one second, because high-frequency transmission from sensor devices will affect the communication quality in an environment with multiple sensor devices. Due to the above requirements, it is difficult for wireless packets to continue transmitting the data measured by the sensors as they are. Therefore, wireless packets are used to transmit the results of processing measured data. However, if the data acquisition intervals using sensors are spaced, e.g., once every minute or once every ten minutes, it is possible to transmit the measurement data as they are.

Of these data formats, CSV files and sensor data files are stored on the Secure Digital (SD) card installed in the sensor device. Therefore, to retrieve the data, the operator needs to take out the SD card and upload it to the system. This is a procedure that places a burden on the operator, but since LoRa communication is used in the industrial machine monitoring phase, this is not an operational problem.

The sensor data analysis phase uses the data obtained by the sensor data collection phase to analyse whether any abnormalities or signs of failure appear in industrial machines. Here, not only the simple size of the sensor data is taken into account, but also the cycles in the manufacturing process. A cycle is the unit of measurement from the beginning to the end of a process when the same product is manufactured or processed repeatedly. By analysing the data in consideration of cycles, it is possible to analyse differences in the data from the point of view of the manufacturing process.

The industrial machine monitoring phase collects the valid feature values obtained from the sensor data analysis phase from the sensor devices via LoRa communication and displays the transition in the web application. The feature values of the sensors attached to industrial machines are automatically collected by LPWA communication, and the values can be immediately checked.

3.2. Integrated Monitoring Platform Web Application

In this research, Django 4.2 [12] is used as the framework for the web application. It is a framework that uses Python 3.12 as its programming language and is characterised by its database-based development. It allows for the easy connection of databases and web pages by defining a data model. It comes standard with a function to manage user accounts, and it is possible to control who is accessing the site by using password authentication and limiting the range of pages that can be accessed by that user for security purposes. The reason for using a Django for the integrated monitoring platform proposed in this research is that the operation of the application does not depend on the terminal environment or operating system. Since this system is expected to be used by workers in SMEs in the manufacturing industry in factories and offices, a Web application that can be accessed and used from a variety of terminals is suitable. Another advantage is that there is no need to install the application on the terminal, and the system can be used simply by accessing it from a Web browser, eliminating the need to build an environment on the terminal.

Besides, a web server is built to run the web application. The web server consists of nginx 1.26.1 [13], a web server application, Django, a web application framework, and MySQL 8.4 [14], a relational database management system. As Django is a web application and does not have a function to process HTTP requests at high speed on a web server, nginx, which is open source and is characterised by fast request processing, is used as the web server application. In addition, MySQL is used to build a relational database based on Django’s data model. MySQL supports robust transaction features, offering a strong mechanism for maintaining data consistency. It is also highly compatible with various operating systems, making it excellent in terms of interoperability and compatibility. Additionally, since MySQL is open-source, it can be operated at a lower cost compared to other database management systems. This makes it accessible for SMEs, which can implement it without the need for significant investment.

In the integrated monitoring platform proposed in this research, external sensor devices are attached to industrial machines to collect data. In this case, it is assumed that several industrial machines that manufacture the same products are installed in a factory, and sensor devices with the same configuration are installed on each of them. Therefore, it is necessary to control which sensor device is attached to which industrial machine and which part of the industrial machine the sensor is attached to. The Entity Relationship (ER) diagram of the database designed to realise these requirements is shown in Figure 4.

The sensor_base_info manages the sensors used in a sensor device. By managing these on the platform, data can be compared between different sensors. The operator registers the sensor model number and physical quantity as information on the sensors used in the sensor device. This makes it possible to obtain data from a device equipped with a certain sensor or to obtain information on what kind of analysis has been carried out with a similar sensor in the past.

The device_base_info manages the configuration of measurement devices with sensors installed on industrial machines and consists of the device model number and sensor installation information. This sensor installation information consists of a combination of the basic sensor information and the sensor number for the device. The reason for the sensor number is to distinguish between two or more sensors that are mounted on a device. The operator registers sensor installation information as the number of sensors installed in the device. By selecting the basic device information, a list of devices with that device configuration can be displayed. This enables data analysis focused on the data measured in devices with similar configurations.

The machine manages the industrial machines that are actually installed in the factory. The machine information consists of machine identification (ID), machine name, installation location and category. The machine ID is an identifier to uniquely identify a machine in the system. The machine name is set to the model number of the machine or an identification name in the company. The installation location is information that manages the installation location of industrial machines and sets the information on which floor and location in which factory the machine is installed. By selecting a machine, the sensors installed on the machine can be listed. This makes it easy to compare and analyse multiple sensors installed on a machine.

The device manages the measurement devices actually installed based on the registered basic device information. The device information consists of a device ID, device name, basic device information, additional information and gateway. The device ID is an identifier to uniquely identify a device in the system. The device name is a name for workers to identify devices associated with the same basic device information. In the device basic information, the device configuration of the device to be registered is selected. In the additional information, points that are specific to the device and that should be shared can be entered.

The sensor manages the sensors possessed by the actually installed devices. The sensor information consists of the sensor mounting information and the device combination. The sensor information consists of the sensor mounting information and the combination of devices. This sensor information is used to manage the locations where the sensors are installed.

The mount manages where on the industrial machine the sensors of the measurement device that are actually installed are located. The installation information consists of the target industrial machine, the sensors installed, the installation point indicating where the sensor was installed on the machine, the installation start date and time, and the installation end date and time. This makes it easy to manage where the sensors of each device are installed in an environment where several devices of the same configuration are installed. The installation information also includes the installation end date and time, so that the history of past installation information can be traced back.

The device_data_csv manages the CSV files of the measurement data stored on the SD card of the measurement device. The purpose of this function is to enable easy management of the conditions under which the CSV files generated when using multiple devices of the same configuration were measured. Even when only one device is used, this function facilitates data analysis during the trial-and-error phase of data acquisition, when considering which part of an industrial machine to install a sensor.

The sensor_installation_info manages the individual data measured by sensors. The target data here are data such as sound data, image data and video data, which do not contain multiple sensor data in one file. These data cannot be registered in the database as they are because the file size is expected to be large. The system uploads sensor data files to a folder on the server and registers the path to the files in the database. This makes it possible to download sensor data later and analyse them under specific measurement conditions.

The gateways manage data aggregation gateways used during data collection. The gateways described here are devices that receive data sent from multiple measurement devices using wired or wireless communication, aggregate them and send them to the server. When sending data from the gateway to the server, the HTTP request is first sent using a TCP connection. The gateway first establishes the TCP connection by using the server’s IP address and port number. The server is waiting on the specified port, and once the TCP connection is established, it receives the HTTP request. The HTTP POST request is sent from the gateway to the server via the TCP/IP network. This request contains header information and the data to be sent. Therefore, it is necessary to interpret the payload of the received HTTP POST packet and convert it into the device’s data In this system, the format of the packet body part of the HTTP POST is not handled by the operator, but by the system implementation. In other words, the system does not support a gateway with a different HTTP POST packet body part format from the assumed gateway. This is because there is no common data format for gateways. However, it is possible to support this by additionally defining which part of the HTTP POST packet body part is the transmission data on the server side. As the format of the transmitted data depends on the implementation of the measurement device, the definition of the format of the transmitted data can be edited by the operator of this system.

In this research, a LoRa gateway, ES920GWX2 manufactured by EASEL Inc. (Kanagawa, Japan) [15], is used as the gateway. This LoRa gateway receives data sent by the company’s LoRa module ES920LR2 and sends the data to the server via HTTP POST using an Ethernet connection or LTE communication using a SIM card. The body of the packet sent by the LoRa gateway to the server is a character string and has the following format.

gwid=XX&ch=XX&sf=XX&rssi=XX&panid=XX&id=XX&Sensor unit transmission data

where gwid is the ID of the ES920GWX2, ch is the radio channel, sf is the spreading factor, rssi is the received signal strength, panid is the network address and id is the ID of the source module. This format combines information on the source and communication with the transmitted data using the ampersand as a delimiter. Therefore, when data in this format are received by a server, the strings below the sixth ampersand can be treated as transmission data. However, as the extracted transmission data may contain multiple data, it is necessary to cast these further. Here, the first to fourth characters of the payload are interpreted as the device_id, the fifth to seventeenth characters as the average of the norm during uniform motion from the left end to the right end (left_to_right_norm_mean) and the eighteenth to thirtieth characters as the average of the norm during uniform motion from the right end to the left end (right_to_left_norm_mean).

3.3. External Attachable Sensor Device

The integrated monitoring platform proposed in this research uses sensor devices that can be attached externally to industrial machines during the sensor data collection phase and the industrial machine monitoring phase. This section describes the configuration and operation of these devices.

First, the configuration and operation of the SD card-compatible accelerometer device used in the sensor data collection phase of the proposed system are described. The appearance of the SD card-compatible accelerometer device is shown in Figure 5.

The main element of this device is a microcontroller board called Grand Central M4 Express featuring the SAMD51 (hereafter referred to as ‘Grand Central’) [16], manufactured by Adafruit Industries (New York, NY, USA). The reason for selecting this microcontroller board was the richness of its interfaces. Grand Central has 62 General Purpose Input Output (GPIO) ports. Among them, 16 ports can be used for analogue input and support many communication standards such as hardware Serial Peripheral Interface (SPI), Inter-Integrated Circuit (I2C) and Universal Asynchronous Receiver Transmitter (UART). This makes it possible to control a large number of sensors and acquire data from a single microcontroller board.

As an acceleration sensor, an ICM-20948 [17] from TDK Corporation (Tokyo, Japan) is used. This sensor is a nine-axis inertial measurement unit with not only a three-axis accelerometer but also a three-axis gyro sensor and a three-axis magnetometer. As this device only measures acceleration, only the three-axis accelerometer is enabled. The measurable range of acceleration can be selected from four levels, ±2 g, ±4 g, ±8 g and ±16 g, and can be set according to the machine to be measured. The device is also equipped with a First-in First-out (FIFO) for temporarily storing sensor data, making it suitable for applications where sensor data are read out at regular intervals. In the implementation of this device, only the accelerometer is installed in the industrial machine to be measured, while the microcontroller board is installed in a remote location. Therefore, the acceleration sensor is connected to the microcontroller board by extending it with shielded conductors. I2C is used for communication with the microcontroller.

A module with a PCF8523, manufactured by Adafruit Industries (New York, NY, USA) is used as the real-time clock (RTC). The purpose of installing a RTC is to record the measurement time of the data. When a CSV file containing measurement data is saved on an SD card, the measurement time can be recorded in the file so that the data can be checked against the data of which product was being manufactured using the production record data. This makes it possible to compare data from the same product production. As with the acceleration sensor, I2C is used for communication with the microcontroller.

The power supply to the device is via an Alternating Current (AC) adapter from a 100 V AC power supply. Grand Central is equipped with a 5.5 mm/2.1 mm centre plus Direct Current (DC) connector and can accept input voltages from 6 V to 12 V. Simply connecting the power supply to the device is not enough to check whether the sensor device itself has started working properly. A reset button with a Light Emitting Diode (LED) is, therefore, used. The LED inside the button is illuminated when the operation starts using the GPIO pins of Grand Central to inform the operator that the device is working properly. Conversely, if the connection to the sensor fails or the SD card recognition fails, the LED blinks to inform the operator that there is a fault with the device. This button is connected to the reset pin of Grand Central and can be reset by pressing the button.

Here, the firmware of the accelerometer device that supports SD cards is described. The device starts up and begins operating when connected to a power source via an AC adapter. After start-up, the RTC, SD card and accelerometer are initialised in the setup function in Arduino.

The initialisation of the RTC requires the correct time to be set. The time when the firmware is written to the computer is used for this time. However, if this time is set every time the sensor device is reset, the time of the RTC reverts to the time at which the firmware was written when the device is reset. Therefore, Grand Central uses Electrically Erasable Programmable Read-Only Memory (EEPROM), which is a non-volatile memory. A flag is stored on the EEPROM to determine whether or not the device has been started up for the first time, and this flag is set after the device has been started up for the first time. This ensures that the RTC time is set only at the first start-up when the firmware is written.

In the initialisation of SD, the file number in the SD card is acquired. The file number is the serial number of the CSV file created in the sensor device, and the file name of each CSV is set to this file number to prevent duplicate file names. This value is reset to 0 when the firmware is written and is incremented by 1 when the CSV file is saved. However, if the file number is managed as a variable in the firmware, it returns to 0 when the power is turned off or reset, so it is also stored in the non-volatile memory EEPROM to make it persistent.

In the initialisation of the accelerometer, various parameters are set. The parameters to be set are the range of acceleration measured by the accelerometer, the digital low-pass filter, the sampling rate and the FIFO mode. In this research, the acceleration measurement range is set to ±2 g, the cut-off frequency of the low-pass filter to 246 [Hz], the sampling rate to 562.5 [Hz] and the FIFO mode to stream mode, which overwrites the FIFO when it overflows. If the initialisation process for these modules fails, the LED on the reset button blinks.

The flow of operations of this accelerometer device after the initialisation process is completed is shown below.

Creating CSV files
Extracting data from the accelerometer FIFO
Writing measurement data to CSV files
Saving CSV files

The acceleration device repeats these operations for each measurement time per file. When creating a CSV file, the CSV file number obtained in the initialisation process is used as the file name to create the CSV file. It also acquires the time from the RTC and sets it as the file creation time. This allows the creation time of the CSV file to be checked. Write one line each for device ID, measurement start time and sampling interval as header information in the created CSV file. This allows the web application to retrieve information on which device the CSV file is from, when the measurement start time of the CSV file is, and how long the sampling interval is. Finally, ‘x, y, z’ is written as the CSV column labels. Here, these labels correspond to the x- y- and z-axes of the triaxial acceleration.

Next, data extraction from the accelerometer’s FIFO and measurement data writing to the CSV file are repeated alternately. In data extraction from the accelerometer’s FIFO, the number of data in the FIFO is first obtained. When data are extracted from the FIFO here, it is necessary to extract the data one by one, but the extracted data are not written to the CSV file one by one and are instead written to a variable array in the microcontroller memory once. The data are inserted once into a variable array in the microcontroller memory. This is because data are added to the FIFO at any time in the accelerometer while writing to the CSV file, and the FIFO may overflow due to variations in the writing time to the SD card. By writing the measurement data to the CSV file after all the data in the FIFO have been extracted once, the data can be stored in the FIFO while writing to the CSV file. The acceleration data extracted to the variable array are written to the CSV file in comma-delimited order in the x- y- and z-axis. After the measurement time per file has elapsed since the CSV file was created, the CSV file is saved. At this time, the next file number is written to the EEPROM.

Next, the configuration and operation of the LPWA communication-compatible accelerometer device used in the industrial machine monitoring phase are described. The appearance of the accelerometer device supporting LPWA communication is shown in Figure 6.

As the main element of this device, a microcontroller board equipped with an ATSAME53J20A, manufactured by Microchip Technology Inc. (Chandler, AZ, USA) [18] is used. In addition, an ES920LR2 [19] from EASEL Corporation (Kanagawa, Japan) [20] is mounted as the LoRa module that transmits the features calculated by edge processing.

The same ICM-20948 as the accelerometer device in the sensor data collection phase is used as the accelerometer. Power is supplied to the device via the micro USB port.

The firmware of the accelerometer device supporting LPWA communication is described here. The device starts up and starts operating when it is connected to a power source via a microUSB port. After start-up, the LoRa module and accelerometer are first initialised in Ardunio’s setup function. The initialisation process of the LoRa module sets the parameters required for LoRa communication. The parameters to be set include those related to the transmission method and those related to the network. The parameters related to the transmission method include the communication bandwidth, spread ratio, radio channel, existence of acks and the number of retransmissions. The communication bandwidth and radio channel number must be standardised between the LoRa modules that want to communicate. In this system, the bandwidth is set to 500 kHz, which is the maximum value. This is to shorten the communication time and reduce communication failures due to noise during transmission. Increasing the bandwidth increases the communication speed, but there is a trade-off in that the communication range is reduced. Therefore, a smaller bandwidth is necessary in an environment with many obstacles. The spread factor is another parameter that is a trade-off between communication distance and communication speed and is set here to 7, which is the default value of the LoRa module. For communication over a wider area, a higher spreading factor should be set. The radio channel is set to 4. The ack function confirms that the data sent in the communication between LoRa modules have been received by means of an ack packet. In this system, this function is enabled. This ensures that the packets sent by the sensor device reach the gateway. The retransmission frequency parameter can be set to 5 so that up to five retransmissions are performed if no acks are received. These settings reduce the number of data collection failures due to temporary poor signal conditions. The ones related to the network are PANID, OWNID and DSTID. PANID is the address of the network itself, expressed as a four-digit hexadecimal number, and must be standardised between LoRa modules that want to communicate. The OWNID is the address of the LoRa module itself, represented by a four-digit hexadecimal number, and must be unique within the same network. The DSTID is the OWNID of the destination LoRa module, expressed as a four-digit hexadecimal number. In this system, the OWNID of the LoRa module installed in the gateway is set. This allows a single gateway to aggregate data from multiple sensor devices. The initialisation process for the accelerometer is the same as for the accelerometer device in the data collection phase using CSV.

After the initialisation process in this acceleration sensor device is completed, the operating cycle of the industrial machine is detected, the feature values are calculated from the measured data and the feature values are transmitted.

4. Results

4.1. Evaluation Experiments Using Testbed

The system proposed in this research can be applied to a wide range of industrial machines used in manufacturing sites of SMEs by flexibly changing the configuration of the sensor devices. Here, evaluation experiments are carried out on the looseness detection of bolts fixing the support side of ball screws, which are used in industrial machines to transport products and for precise positioning.

The testbed used in the experiment is shown in Figure 7.

The testbed consists of four elements: ball screw, servo motor, servo amplifier and stage. A 1 [kg] weight is placed on the stage on top of a plastic case, thus reproducing the product to be transported on the production floor. The testbed was set to move in a reciprocating motion between the left and right ends. This reciprocating motion consists of four movements: starting from the leftmost position of the stage, uniform motion from left to right, stopping at the rightmost position, uniform motion from right to left and stopping at the leftmost position. Here, the stopping time at the left end is longer than the stopping time at the right end, so the stopping time between reciprocating motions is longer. The testbed repeats this reciprocating motion at intervals of approximately 5.5 [s] per round trip. In the following, this one round trip is treated as one cycle in the manufacturing process.

The flow of this evaluation experiment is shown next.

Data collection using SD card
Cutting out to cyclic waveforms
Determination of edge processing
Wireless transmission of features using edge processing

First, data are collected using SD cards to measure the data required to determine the edge processing to be performed on the sensor devices of the proposed system. At this stage, before the edge processing is determined, the appropriate sampling rate for measuring the data for edge processing is unknown. Therefore, the data are measured at the maximum sampling rate that can be measured by the sensor device. Next, the data recorded on the SD card are cut into cycle waveforms. The cycle waveforms are then analysed between different bolt tightening torque conditions to derive features that show differences. Finally, the cycle extraction and feature calculation processes obtained in the experiments up to this point are implemented in the sensor device as edge processing, and the feature values collected using LPWA communication are visualised in a web application.

Here, the experimental conditions of this evaluation experiment are described. In this evaluation experiment, the bolts on the support side of the ball screw are tightened appropriately to 11 [N] and the bolts are loosened to 1 [N]. Data measurements were carried out at two different tightening torques. A torque spanner is used to adjust to the respective tightening torques.

4.2. Data Collection Experiments Using SD Cards

In this experiment, an acceleration sensor device that saves the measurement data to an SD card as described earlier is used. The acceleration sensor of the sensor device was fixed to the side of the support using magnets and tape in order to install it as close as possible to the support. Data measurements were carried out for one hour at each tightening torque. The triaxial acceleration data measured at a tightening torque of 11 N and the triaxial acceleration data measured at a tightening torque of 1 N are shown in Figure 8, Figure 9, Figure 10, Figure 11, Figure 12 and Figure 13 for each axis.

The experimental results showed that vibrations associated with the reciprocating motion of the testbed were measured in each axis at tightening torques of 11 N and 1 N, respectively. The acceleration sensor value is 16,384, which is the same acceleration as the gravitational acceleration. In actual production, it is assumed that the mounting posture of the accelerometers may change slightly when the accelerometers are removed. In addition, if the three-axis accelerations are handled independently of each other, the cut-out and analysis of the cycle waveforms becomes more complicated. Therefore, in this evaluation experiment, the norm

| a |

of the three-axis acceleration is analysed according to the Formula (1).

| a | = \sqrt{{a_{x}}^{2} + {a_{y}}^{2} + {a_{z}}^{2}}

(1)

By treating the three-axis acceleration measured by the accelerometer device as a one-dimensional vibration waveform, the influence of differences in the installation posture of the accelerometer on the measured values can be reduced. Part of the norm of the three-axis acceleration data measured at a tightening torque of 11 N is shown in Figure 14, and part of the norm of the three-axis acceleration data measured at a tightening torque of 1 N is shown in Figure 15. In the norm of the triaxial acceleration, it was confirmed that cyclic vibrations with reciprocating motion of the testbed were measured in each axis at tightening torques of 11 N and 1 N, respectively. Next, the norm waveforms obtained here were cut out into waveforms in cycles by applying a cycle cut-out algorithm.

4.3. Cutting to Cycle Waveforms

Data collection experiments using SD cards have enabled vibration waveforms on the testbed to be obtained. However, as these data are one continuous waveform for each tightening torque, it is not possible to analyse them from the point of view of which part of a cycle the change appears. In this evaluation experiment, a cut-out algorithm for cycle waveforms is proposed to enable comparison between testbed reciprocating cycles from a single continuous vibration waveform, and its usefulness is demonstrated by applying it to the measured norm waveforms. The cycle waveform cut-out algorithm used in this research consists of the following steps:

Detection of outages between cycles
Detection of cycle start point
Data cut-out at cycle length

In the detection of the stop state between cycles, the range of values of the vibration waveform at the time of the ball screw stop is first defined from the graph of the norm waveform. Next, the part where the value within the defined norm range at the time of ball screw stop continues for a certain period of time or longer is judged to be the stop state of the ball screw. The stop state determined here includes both the stop when the ball screw stage is at the right end and the stop when the stage is at the left end. The fact that the time for the stage to stop at the left end is longer than the time for the stage to stop at the right end in each cycle is used. In other words, a stop at the left end is judged to be a stop when the duration of the stop is longer than the stop at the right end of the ball screw. Stopping at the left end can be treated as stopping between cycles, and thus the stopping state between cycles can be detected.

After detecting the stop between cycles, the cycle start point is detected. Here, the fact that the stop always takes place at the left-hand end before the start of the cycle is used. In other words, a new cycle of the ball screw is judged to have started when a norm exceeding the norm range at the time of the ball screw stop is detected immediately after the detected stop state between cycles. With this method, even if a norm exceeding the range of the norm at ball screw stop is observed as noise during the stop at the right end or during the stop at the left end, it does not satisfy the condition that it is immediately after the stop state between cycles, so it does not incorrectly cut out the norm.

Finally, the data are cut out at a fixed length of cycle from the detected cycle start point. Here, the cycle length is defined manually from the norm waveform as well as the norm range when the ball screw stops. The fixed-length cut-out has the disadvantage that it cannot cope with dynamically changing cycle lengths. However, fixed-length cut-outs are an advantage when similar products are repeatedly manufactured at the same cycle length on the production floor, as changes in the approximate shape of the cut-out waveform can detect anomalies in manufacturing, even though the cut-outs are made at the same cycle length.

This cut-out algorithm was applied to the norm waveforms obtained in the data collection experiment using the SD card. The cycle waveforms at a tightening torque of 11 N are shown in Figure 16 and at a tightening torque of 1 N in Figure 17. The experimental results confirmed that each of the extracted waveforms was a cycle-by-cycle waveform, with the vibration matching the cyclic motion of the testbed. The use of the norm of triaxial acceleration, rather than triaxial acceleration itself, as the data to be extracted enables extraction without having to set the range of the stop state for each axis.

4.4. Data Analysis for Edge Processing Content Determination

From the cycle waveform obtained by the cycle waveform cutting algorithm, features effective for detecting bolt loosening on the ball screw support side support are extracted, and the edge processing content of the sensor device is determined. In this evaluation experiment, the stage moves in two parts during one cycle of the ball screw of the testbed: uniform motion from the left end to the right end and uniform motion from the right end to the left end. Therefore, in order to focus on the vibration during the stage movement, the data for the relevant part are extracted from the respective cycle waveforms. Figure 18 shows the parts to be extracted as the parts of uniform motion from the left end to the right end and from the right end to the left end. Here, in extracting the data, the 700th sample from the beginning of the cycle waveform is considered as the constant velocity motion from the left end to the right end, and the 1300th sample to the 2000th sample as the constant velocity motion from the right end to the left end, because each cycle waveform is cut out with a similar length.

The evolution of the variance values for the uniform motion part of the extracted cycles was then plotted on a graph. The evolution of the dispersion values for the uniform motion from the leftmost to the rightmost is shown in Figure 19 and that for the uniform motion from the rightmost to the leftmost is shown in Figure 20. In each graph, a moving average of 50 points is applied to reduce the influence of cycle-to-cycle variation. In the case of uniform motion from the left end to the right end, the dispersion value tends to be higher for a tightening torque of 1 N up to around 300 cycles. However, after 300 cycles, the dispersion value for a tightening torque of 11 N repeatedly increased and decreased, and the dispersion value for a particularly large portion was similar to that for a tightening torque of 1 N. It was confirmed that the threshold value could not be used to determine the dispersion value. In contrast, in the case of uniform motion from the right end to the left end, it was confirmed that the dispersion values for a tightening torque of 11 N tended to be larger than those for a tightening torque of 1 N overall. However, in the vicinity of 200, 480 and 600 cycles, where the dispersion value for a tightening torque of 1 N is larger, the dispersion value for a tightening torque of 11 N is larger than that for a smaller torque, leading to misjudgment by the threshold value in these areas. Based on the above results, attention was paid to the dispersion during uniform motion from the right end to the left end, where the greatest difference was observed between a tightening torque of 11 N and a tightening torque of 1 N.

Here, a difference in the variance of the waveform immediately after the start of uniform motion from the right end to the left end was observed in the cycle waveform of Figure 16, which is a cycle waveform with a tightening torque of 11 N, and in the cycle waveform of Figure 17, which is a cycle waveform with a tightening torque of 1 N. Therefore, 100 samples from the 1300th sample to the 1400th sample in each cycle waveform were extracted as the data immediately after the start of uniform motion from the right end to the left end, and the variance was determined. However, a moving average of 50 points is applied to reduce the effect of cycle-to-cycle variations.

In the dispersion value focused immediately after the start of uniform motion from the right end to the left end, it was confirmed that the dispersion value for a tightening torque of 11 N was greater than that for a tightening torque of 1 N in all cycles. Therefore, the dispersion value focused immediately after the start of uniform motion from the right end to the left end was found to be suitable for detecting bolt loosening on the ball screw support. By obtaining this feature value as edge processing in the sensor device of this system, only the data necessary for detecting bolt loosening of the ball screw support can be transmitted, thereby reducing the amount of data.

4.5. Experimental Wireless Transmission of Features Using Edge Processing

Finally, the edge processing obtained in the experiments up to this point is implemented in a sensor device that supports LPWA communication, and only the feature values are aggregated via LPWA communication and visualised in a web application.

First, we describe the edge processing procedure for the sensor device. Edge processing repeats the following steps in accordance with the testbed cycle.

Detection of outages between cycles
Measurement of uniform motion from left end to right end
Measuring stops at the right-hand end
Measurement of uniform motion from the right end to the left end
Calculation of features
Transmission of features

In order to be able to respond either when the testbed stage is stopped at the left end or when it has already been cycled, the sensor device first detects the stop condition between cycles. This is determined by whether the norm of the three-axis acceleration measured at the sensor device falls within the norm of the ballscrew stop. If the norm at ballscrew stop is measured continuously for longer than the stop time at the right end, the ballscrew is judged to be in an inter-cycle stop state. If the norm is measured for a period longer than the stopping time at the ballscrew stop, it is judged to be the start of a cycle and the measurement of uniform motion from the left end to the right end is started. As the time taken for uniform motion from the left end to the right end is constant, the fixed length data are measured and stored in memory. Next, after stopping the measurement during the stop at the right end, the measurement of the uniform motion from the right end to the left end is carried out with a fixed length.

As the norm of the three-axis acceleration has been measured in the previous steps, these values are used to calculate the feature values. Here, the variance is calculated for the 1300th sample to the 1400th sample in the cycle waveform, i.e., the first 100 samples of the uniform motion from the right end to the left end. The calculation results are then sent to the gateway via LPWA communication as the feature values for this cycle. The packets sent by the sensor device include the device ID and the feature values. This allows the web server that receives the packet from the gateway to identify which sensor device the data comes from.

Using the sensor device with the above edge processing implemented, data were collected on the testbed for one hour at a tightening torque of 11 N, followed by one hour at a tightening torque of 1 N. The data were visualised in a web application. The actual visualisation page of the web application is shown in Figure 21.

In this graph, the dispersion values at a tightening torque of 11 N are displayed from 16:13 to 17:13 on 30 October 2023, and the dispersion values at a tightening torque of 1 N are displayed from 17:14 to 18:14. The graph shows that the dispersion values in the section with a tightening torque of 11 N tend to be larger than those in the section with a tightening torque of 1 N. However, the values vary from cycle to cycle, with some cycles having a dispersion value in the range of the dispersion value of the fastening torque 1 N when the dispersion value of the fastening torque 11 N is small, and some cycles having a dispersion value in the range of the dispersion value of the fastening torque 11 N when the dispersion value of the fastening torque 1 N is large. The value of the dispersion value of the tightening torque 1 N is the range of the dispersion value of the tightening torque 11 N. Here, it is considered that the cycles of tightening torque 11 N and tightening torque 1 N are discriminated by a threshold value. As a threshold value, the average of the dispersion values of each tightening torque interval is calculated and a value halfway between these averages is used. From the experimental data, the mean of the 11 N tightening torque interval was 307,968.7 and the mean of the 1 N tightening torque interval was 204,376.7, so a threshold value of 256,172.7 was set. A cycle with a tightening torque of 11 N is judged to be a cycle with a tightening torque of 11 N if it is greater than or equal to this threshold value, and a cycle with a tightening torque of 1 N is judged to be a section with a tightening torque of 1 N if it is less than this threshold value. The confusion matrices resulting from the threshold judgement for the 641 cycles in each section are shown in Table 1.

The number of correctly determined cycles was 1027, giving a correctness rate of 80.0%. This result shows that the correct judgement of the tightening torque can be made with an accuracy of 80% by using a threshold to judge the dispersion value of only one cycle. Focusing on the incorrect judgement results, the proportion of incorrect judgement of the tightening torque 11 N was 24.5 % and the proportion of incorrect judgement of the tightening torque 1 N was 15.3%, with the proportion of incorrect judgement of the tightening torque 11 N being higher. Therefore, there is a high possibility that a false alarm is generated to indicate that a bolt is loose even though it is not loose. To improve this result, a moving average of 20 points was applied in this evaluation experiment to reduce the influence of cycle-by-cycle variations. The visualisation page of the web application when a moving average of 20 points is applied is shown in Figure 22.

The part where the dispersion value decreased significantly around 17:15 is the part where the ball screw was stopped to change the tightening torque of the support using a torque spanner, which is the effect of the 20-point moving average. The graph shows that the dispersion values in the section with a tightening torque of 11 N exceed the dispersion values in the section with a tightening torque of 1 N. Here, a value between the minimum value in the 11 N tightening torque section and the minimum value in the 1 N tightening torque section is used as the threshold value. From the experimental data, the minimum value in the 11 N tightening torque section was 269,984.3 and the maximum value in the 1 N tightening torque section was 259,078.9, so 258,741.0 is set as the threshold value. The threshold value was used to determine the tightening torque in the same way as in the case where no moving average was applied. The confusion matrices resulting from the threshold judgement for the 621 cycles of each interval are shown in Table 2.

The number of correctly judged cycles was 1241, giving a correct answer rate of 99.999%. In particular, there was one sample in which a tightening torque of 1 N was incorrectly judged as a tightening torque of 11 N, but there was no sample in which a tightening torque of 11 N was incorrectly judged as a tightening torque of 1 N. The result was that no false alarms were generated when the bolt was not loose. These results indicate that, when detecting bolt looseness on the support side of a ball screw, the average of the dispersion values of 20 consecutive cycles can be used instead of using only the dispersion values of one cycle. The results show that detection with high accuracy is possible by using the average of the dispersion values of 20 consecutive cycles.

5. Discussion

In the cut-out to the cycle waveform performed in the evaluation experiment, the range of the norm at the ball screw stop was defined manually from the measurement data. It is assumed that the range of the norm varies depending on the industrial machine and product to be measured. Therefore, when the number of industrial machines handled by the proposed system increases, the definition and management of these ranges becomes a burden for the operator. Therefore, it is necessary to automatically detect the periodicity of the measured waveforms, taking into account the autocorrelation. Norm ranges are also used in sensor devices that perform edge processing. To change these values, the firmware needs to be written again. Therefore, the downlink function of LPWA communication is used to transmit the norm range set in the server to the sensor device, so that the range can be updated remotely.

In the proposed system, the sensor devices used in the sensor data collection phase are different from those used in the industrial machine monitoring phase. Therefore, two types of sensor devices need to be prepared for each industrial machine in order to install the proposed system. In factories where there are several industrial machines to be installed, the number of sensor devices required increases, which increases production costs and the burden on workers when changing the firmware. Therefore, we design a sensor device that integrates the functions of these two devices. This makes it easy to switch from the sensor data collection phase to the industrial machine monitoring phase.

Hsiao et al. [21] proposed the development of an industrial Internet of Things (IIoT) platform for information integration and data analysis using open-source technologies. They compared the proposed method with several industrial commercial platforms and conducted tests related to data communication and analysis for information support. They also demonstrated a web interface for system implementation. Folgado et al. [22] proposed a data acquisition and monitoring system based on IIoT for a Polymer Electrolyte Membrane hydrogen generators. Parameters such as current, voltage, pressure, and hydrogen flow are sensed and acquired by industrial controllers and stored in a database. This enables users to access the information online in real time through a web-based interface. These works provide a good idea for further improvement of our research in the future.

6. Conclusions

In this research, an integrated monitoring platform was built to handle everything from data collection to visualisation, with the aim of data-based preventive maintenance of industrial machines in SMEs. In order to efficiently manage devices with the same configuration, the web application implements a function for basic sensor information and basic device information that combines them so that devices equipped with specific sensors can be easily checked. The registration of the machine to be installed and the installation position also enables the tracing back of which machine and position the sensor was installed on at a specific date and time. Furthermore, by linking the CSV file stored on the SD card to the device, the data collected in the past and the installation conditions can be retrieved, thereby reducing the burden of analysis. In this research, two sensor devices were created: one for use in the sensor data collection phase and one for use in the industrial machine monitoring phase. The device for the sensor data collection phase stores the measured data on an SD card and enables data analysis to determine the edge processing content. The device for the industrial machine monitoring phase reduces the amount of measurement data by implementing edge processing and enables the easy collection of data from industrial machines by using LPWA communication, a type of wireless communication.

In the evaluation experiment, data were actually collected on an SD card for the detection of loosening of bolts on the support side of a ball screw, and the data were cut out into a cycle waveform using a cycle waveform cut-out algorithm. The feature values during a specific movement were calculated from the cycle waveform, and the data from different tightening torques were compared to analyse the feature values effective for detecting bolt loosening. As a result of the experiment, a difference in the dispersion value immediately after the start of uniform motion from the right end to the left end was observed, and this was implemented as edge processing in a sensor device supporting LPWA communication. In the wireless transmission experiments on the testbed, the threshold judgement for each cycle resulted in a correct rate of 80.099%, but by using the average of 20 consecutive cycles, the detection accuracy was improved to 99.999%, demonstrating the usefulness of the proposed system.

The future outlook is as follows: It is expected that the range of norms will differ depending on the industrial machinery or products being measured. By automatically detecting the periodicity of the measured waveform while considering autocorrelation, and using the downlink functionality of LPWA communication to send the norm range set on the server to the sensor device, it will be possible to remotely update the range.

In the proposed system, two types of sensor devices are used: one for the sensor data collection phase and another for the industrial machinery monitoring phase. A sensor device that integrates the functions of these two devices will be designed. This will make it easier to transition from the sensor data collection phase to the industrial machinery monitoring phase.

Improvements to the user interface of the web application are also being considered. To enable intuitive operation and result display for users in the field, the user experience will be enhanced through surveys and other methods.

Based on the above, the goal for the future is to implement the proposed system in the manufacturing sites of small and medium-sized enterprises, carrying out the entire process from data collection to visualisation, and enabling long-term operation.

Author Contributions

Conceptualisation, N.K., S.H. and K.Y.; methodology, N.K., S.H. and K.Y.; software, N.K., S.H. and K.Y.; validation, N.K., S.H. and K.Y.; formal analysis, N.K., S.H. and K.Y.; investigation, N.K., S.H. and K.Y.; resources, N.K., S.H. and K.Y.; data curation, N.K., S.H. and K.Y.; writing—original draft preparation, N.K., S.H. and K.Y.; writing—review and editing, N.K., S.H. and K.Y.; visualisation, N.K., S.H. and K.Y.; supervision, T.O.; project administration, T.O.; funding acquisition, T.O. All authors have read and agreed to the published version of the manuscript.

Funding

This article is based on results obtained from a project, JPNP20004, subsidized by the New Energy and Industrial Technology Development Organization (NEDO).

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The data presented in this research are available on individual request from the corresponding author.

Acknowledgments

The authors extend their appreciation to a project, JPNP20004, subsidized by the New Energy and Industrial Technology Development Organization (NEDO).

Conflicts of Interest

The authors declare no conflicts of interest.

Abbreviations

The following abbreviations are used in this manuscript:

SMEs	Small and Medium-sized Enterprises
IoT	Internet of Things
IT	Information Technology
GDP	Gross Domestic Product
DX	Digital Transformation
USB	Universal Serial Bus
LTE	Long Term Evolution
AI	Artificial Intelligence
SIM	Subscriber Identity Module Card
GMM	Gaussian Mixture Model
LPWA	Low Power Wide Area
HTTP	Hypertext Transfer Protocol
CSV	Comma Separated Values
SD	Secure Digital
ER	Entity Relationship
ID	Identification
GPIO	General Purpose Input Output
SPI	Serial Peripheral Interface
I2C	Inter-Integrated Circuit
UART	Universal Asynchronous Receiver Transmitter
FIFO	First-in First-out
RTC	Real Time Clock
AC	Alternating Current
DC	Direct Current
LED	Light Emitting Diode
EEPROM	Electrically Erasable Programmable Read-Only Memory
IIoT	Industrial Internet of Things

References

Folgado, F.J.; Calderón, D.; González, I.; Calderón, A.J. Review of Industry 4.0 from the Perspective of Automation and Supervision Systems: Definitions, Architectures and Recent Trends. Electronics 2024, 13, 782. [Google Scholar] [CrossRef]
Ministry of Economy, Trade and Industry. White Paper on Manufacturing Industries (FY2023); SHOBI Printing Co., Ltd.: Tokyo, Japan, 2023.
The Small and Medium Enterprise Agency. White Paper on Small and Medium Enterprises in Japan (FY2023); Nikkei Printing Inc.: Tokyo, Japan, 2023.
The Small and Medium Enterprise Agency. White Paper on Small and Medium Enterprises in Japan (FY2019); Nikkei Printing Inc.: Tokyo, Japan, 2019.
The Small and Medium Enterprise Agency. White Paper on Small and Medium Enterprises in Japan (FY2022); Nikkei Printing Inc.: Tokyo, Japan, 2022.
Ministry of Economy, Trade and Industry. Digital Governance Code 2.0. Available online: https://www.meti.go.jp/policy/it_policy/investment/dgc/dgc2.pdf (accessed on 10 December 2024).
Condition Monitoring Solution ACOUS NAVI | NSK Corporation. Available online: https://www.acousnavi.nsk.com/jp/ (accessed on 1 September 2024).
Ball Bearing and Roller Bearing Manufacturers | NSK. Available online: https://www.nsk.com/jp/ (accessed on 1 September 2024).
THK Official Web Site. Available online: https://www.thk.com/jp/ja/ (accessed on 1 September 2024).
OMNIedge OEE Optimization System | THK Official Web Site. Available online: https://www.thk.com/brand/omniedge/jp/ (accessed on 11 November 2023).
Hiruta, T.; Maki, K.; Kato, T.; Umeda, Y. Unsupervised learning based diagnosis model for anomaly detection of motor bearing with current data. In Proceedings of the Procedia CIRP, 28th CIRP Conference on Life Cycle Engineering, Jaipur, India, 10–12 March 2021; Volume 98, pp. 336–341. [Google Scholar]
The Web Framework for Perfectionists with Deadlines—Django. Available online: https://www.djangoproject.com/ (accessed on 16 September 2024).
nginx. Available online: https://nginx.org/en/ (accessed on 16 September 2024).
MySQL. Available online: https://www.mysql.com/jp/ (accessed on 16 November 2023).
Private LoRa Gateway ES920GWX2 | EASEL. Available online: https://easel5.com/service/products-information/products/lora-gateway/es920gwx2/ (accessed on 15 September 2024).
Overview—Introducing the Adafruit Grand Central M4 Express—Adafruit Learning System. Available online: https://learn.adafruit.com/adafruit-grand-central (accessed on 14 September 2024).
ICM-20948: Detailed Information | Sensors and Sensor Systems—Motion/Inertial Sensors—IMU (Inertial Measurement Unit) | TDK Product Center. Available online: https://product.tdk.com/ja/search/sensor/mortion-inertial/imu/info?part_no=ICM-20948 (accessed on 16 September 2024).
ATSAME53J20A—Microchip Technology. Available online: https://www.microchip.com/en-us/product/ATSAME53J20A (accessed on 16 September 2024).
920 MHz Band LoRa/FSK Module (Ultra-Low Power Consumption) ES920LR2 | EASEL. Available online: https://easel5.com/service/products-information/products/wireless-module/es920lr2/ (accessed on 15 September 2024).
IoT Consulting and Radio Development ODM Supply | EASEL. Available online: https://easel5.com/ (accessed on 14 September 2024).
Hsiao, C.H.; Lee, W.P. OPIIoT: Design and Implementation of an Open Communication Protocol Platform for Industrial Internet of Things. Internet Things 2021, 16, 100441. [Google Scholar] [CrossRef]
Folgado, F.J.; González, I.; Calderón, A.J. Data Acquisition and Monitoring System Framed in Industrial Internet of Things for PEM Hydrogen Generators. Internet Things 2023, 22, 100795. [Google Scholar] [CrossRef]

Figure 1. Power spectrum of current waveforms [11].

Figure 2. Log likelihood by experimental condition [11].

Figure 3. Schematic diagram of proposed system.

Figure 4. Entity Relationship (ER) diagram of the database.

Figure 5. Accelerometer devices compatible with Secure Digital (SD) cards.

Figure 6. Accelerometer devices for Low Power Wide Area (LPWA) communication.

Figure 7. Testbed.

Figure 8. x-axis acceleration at 11 N of torque.

Figure 9. x-axis acceleration at 1 N of torque.

Figure 10. y-axis acceleration at 11 N of torque.

Figure 11. y-axis acceleration at 1 N of torque.

Figure 12. z-axis acceleration at 11 N of torque.

Figure 13. z-axis acceleration at 1 N of torque.

Figure 14. Norm at 11 N of torque.

Figure 15. Norm at 1 N of torque.

Figure 16. Cycle waveform at 11 N of torque.

Figure 17. Cycle waveform at 1 N of torque.

Figure 18. Operating point of the cycle waveform at 11 N of torque.

Figure 19. Dispersion values in uniform motion from the left end to the right end.

Figure 20. Dispersion values in uniform motion from the right end to the left end.

Figure 21. Dispersion values immediately after the start of uniform motion from the right end to the left end, as displayed in the web application.

Figure 22. Variance values immediately after the start of uniform motion from the right end to the left end, as displayed in the web application (20-point average).

Table 1. Confusion matrix of decisions using thresholds.

		Results
		11 N	1 N
Actual torque	11 N	484	157
Actual torque	1N	98	543

Table 2. Confusion matrix of decisions with thresholds when 20-point moving average is applied.

		Results
		11 N	1 N
Actual torque	11 N	621	0
Actual torque	1 N	1	620

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

© 2024 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).

Share and Cite

MDPI and ACS Style

Kamiya, N.; Hibino, S.; Yoshizato, K.; Otsuka, T. Proposal and Implementation of an Integrated Monitoring Platform for Preventive Maintenance of Industrial Machines. Appl. Sci. 2024, 14, 11534. https://doi.org/10.3390/app142411534

AMA Style

Kamiya N, Hibino S, Yoshizato K, Otsuka T. Proposal and Implementation of an Integrated Monitoring Platform for Preventive Maintenance of Industrial Machines. Applied Sciences. 2024; 14(24):11534. https://doi.org/10.3390/app142411534

Chicago/Turabian Style

Kamiya, Nene, Shunya Hibino, Konosuke Yoshizato, and Takanobu Otsuka. 2024. "Proposal and Implementation of an Integrated Monitoring Platform for Preventive Maintenance of Industrial Machines" Applied Sciences 14, no. 24: 11534. https://doi.org/10.3390/app142411534

APA Style

Kamiya, N., Hibino, S., Yoshizato, K., & Otsuka, T. (2024). Proposal and Implementation of an Integrated Monitoring Platform for Preventive Maintenance of Industrial Machines. Applied Sciences, 14(24), 11534. https://doi.org/10.3390/app142411534

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Menu