KR20190062739A - Method, algorithm and device for Data analytics for predictive maintenance using multiple sensors - Google Patents

Abstract

The present invention relates to a data analysis method for predicting an equipment failure in manufacturing processes by using a plurality of sensors and a device therefor. According to the present invention, the data analysis method comprises: a first step of generating a composite sensor pattern from patterns of a plurality of sensors; a second step of learning the composite sensor pattern in any one of normal and abnormal states; a third step of storing a corresponding composite sensor pattern in an abnormal state database together with error information on manufacturing equipment when learning is performed in the abnormal state; and a fourth step of performing predictive analysis whether equipment coupled with a sensor is abnormal based on the composite sensor pattern stored in the abnormal state database.

Description

Technical Field [0001] The present invention relates to a data analysis method, an algorithm and a device for predicting equipment failures in a manufacturing process by utilizing a plurality of sensors,

The present invention relates to a data analysis method and apparatus, and more particularly, to a data analysis method, an algorithm and an apparatus capable of predicting equipment failure in a manufacturing process by utilizing a plurality of sensors.

Conventionally, a specific sensor (for example, a vibration sensor) is attached to a manufacturing apparatus and a normal pattern of a normal state is confirmed by continuously learning a change pattern of the sensor.

If a non-normal pattern (exception condition) occurs, it is checked whether the actual equipment is abnormal. If there is an abnormality, it is stored in the abnormal pattern DB.

By continuously analyzing the patterns of the sensors installed in the equipment, the user can learn normal patterns and abnormal patterns. If an abnormality occurs, the user is guided to perform predictive maintenance.

1 is a view for explaining a conventional technique.

As shown in FIG. 1, in the prior art, when each of the sensors (sensors A, B, and C) connected to each of the mechanical equipment extracts a pattern, the pattern is compared to learn a normal pattern and an abnormal pattern. Can be done.

However, such a conventional technique has limited the range that can be actually applied by solving the problem using one specific sensor. For example, when using an electric sensor, there is a limitation that it can be applied only to a motor using AC current.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and it is an object of the present invention to provide a manufacturing method of a sensor, The present invention provides a data analysis method, an algorithm, and an apparatus for predicting equipment failures in a manufacturing process by utilizing a plurality of sensors capable of improving the accuracy of the predictive maintenance through analysis.

In addition, the present invention can prevent the damage due to the failure of the manufacturing equipment by checking the data analysis technique for the maintenance of the manufacturing equipment in the smart factory area and the area in which the failure occurs in the manufacturing equipment in advance Another object of the present invention is to provide a method and apparatus capable of reducing the maintenance cost of manufacturing equipment.

Further, the present invention can increase the accuracy of prognosis maintenance for checking whether there is an abnormality in the manufacturing equipment, and can minimize the loss that can be caused from the failure of the manufacturing equipment through the present invention and reduce the maintenance cost of the manufacturing equipment The present invention also provides a method and an apparatus capable of performing the above-described operations.

According to an aspect of the present invention, there is provided a data analysis method for predicting equipment failure in a manufacturing process using a plurality of sensors, the method comprising: a first step of generating a composite sensor pattern from a pattern from a plurality of sensors; A second step of learning the composite sensor pattern in either a steady state or an abnormal state; A third step of storing the complex sensor pattern in the abnormal state database together with the error information of the manufacturing equipment when it is learned in an abnormal state; And a fourth step of performing a predictive analysis on the presence or absence of abnormality of the sensor coupled device based on the complex sensor pattern stored in the abnormal state database. To provide a method for analyzing data.

Here, the fourth step may include setting an abnormal value discrimination model capable of discriminating whether or not a normal situation exists.

In addition, it may include learning a hybrid sensor pattern using a SOM (Self Organizing Map) refinement model.

In addition, a composite sensor pattern can be generated by correlation values of input signals from a plurality of sensors and presence or absence of correlation.

Further, it is preferable to perform the prediction analysis based on the correlation coefficient value and the presence or absence of correlation.

According to another aspect of the present invention, there is provided a data analysis apparatus for predicting equipment failure in a manufacturing process by utilizing a plurality of sensors, comprising: a sensor pattern generation unit for generating a complex sensor pattern from a pattern from a plurality of sensors; A learning unit for learning the composite sensor pattern in either a steady state or an abnormal state; A pattern management unit for storing the complex sensor pattern in the abnormal state database together with the error information of the manufacturing equipment when it is learned in an abnormal state; And a predictive analysis unit for performing a predictive analysis of an abnormality of a sensor combined device based on the complex sensor pattern stored in the abnormal state database. A data analysis device is provided.

According to the present invention, a plurality of different sensors can be installed in the manufacturing equipment, and it is possible to check whether there is an abnormality in the manufacturing equipment through the composite pattern which recognizes the patterns of the individual sensors in a complex manner. Algorithms, and devices that utilize a plurality of sensors that can predict the failure of equipment in the manufacturing process.

In addition, the present invention can prevent the damage due to the failure of the manufacturing equipment by checking the data analysis technique for the maintenance of the manufacturing equipment in the smart factory area and the area in which the failure occurs in the manufacturing equipment in advance , And a method and an apparatus capable of reducing the maintenance cost of the manufacturing equipment can be provided.

In addition, according to the present invention, it is possible to increase the accuracy of predictive maintenance for checking whether there is an abnormality in the manufacturing equipment, minimize the loss that may arise from the error of the manufacturing equipment through the present invention, Can be saved.

1 is a view for explaining a conventional technique.
2 is a diagram for explaining the overall configuration and operation of the present invention.
3 is a diagram for explaining a model workflow.
4 is a diagram for explaining a model workflow.
5 is a diagram for explaining a function of managing data in a device according to the present invention.
6 is a screen showing a facility management function.
FIG. 7 is a diagram for explaining a process of detecting a process status abnormality by applying a big data technique.
8 is a diagram for explaining a process of performing a failure prediction and a cause search by applying a big data technique.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

The present invention is a data analysis technology for predictive maintenance of manufacturing equipment in a smart factory area. It can prevent the damage caused by the failure of the manufacturing equipment in advance by confirming the area in which the failure can occur in the manufacturing equipment, It is a cost-saving technology.

2 is a diagram for explaining the overall configuration and operation of the present invention.

As shown in FIG. 2, the present invention improves the precautionary maintenance method of the manufacturing equipment through the existing single sensor, and it is understood that the multiple sensor method using a plurality of sensors is applied.

That is, in the present invention, the patterns of the individual sensors are integrated into a composite sensor pattern, and the composite sensor pattern is learned in a normal / abnormal state.

The abnormal state is stored in the abnormal state DB together with the error of the manufacturing equipment, and the maintenance of the manufacturing equipment is possible based on this.

In such a state, the apparatus according to the present invention can perform the prediction analysis by the following method.

First, the prediction analysis model should pay attention to the following points.

That is, there is a realistic consideration that predictive analysis should accumulate data for a sufficient period of time. In addition, it is common to analyze the process efficiency and predictability of the failure, but in some cases it is not clear whether the efficiency is low or the failure state. Therefore, it is necessary to discriminate the outliers that can discriminate the normal situation regardless of the failure or not, and clustering (subdivision) is required to distinguish the states from various kinds of data.

Next, the outlier detection model will be described.

 o Basic Theory

        <When there is little data>

- 'Third Quartile + 1.5 × (Third Quartile - First Quartile) / 2' Upper Limit

- 'First quartile -1.5 × (third quartile - first quartile) / 2' is set as the lower limit.

- Data beyond the upper limit value and data below the lower limit value become abnormal values.

        <Many data and normal distribution>

      - Set the 'average + 2 × standard error' as the upper limit and 'the average -2 × standard error' as the lower limit.

<For data with multiple variables (dimensions)>

- Use PCA (Principal Component Analysis) to reduce the dimension

- The characteristics of the data should be differentiated among the other data. This means that there is a large variance between the data in terms of the attributes. In terms of dimensions (variables, attributes), not characteristics, the data are not well distinguished, so the distance, or variance, is small. Then, the dimension reduction of the PCA is a process of discarding a meaningless dimension, and the result is the same as that of creating new data by selecting only the dimensions that serve as the characteristic.

- Draw the distribution with the reduced parameters, and see the part that is distant from the main distribution as the outliers.

o Model Workflow

3 is a diagram for explaining a model workflow.

Referring to FIG. 3, the following will be described. The model workflow encompasses both a single variable and multiple variables. After identifying the input data type, the single work variable and multiple variable processing are divided.

In the case of multiple variables, PCA is used to distinguish the normal / abnormal state, and normal / abnormal data learning is determined by logistic regression.

o Role of model stage

step role Remarks Input Bring input data Data Type Identify whether the data configuration is a single variable or multiple variables Univariate Outliers In the case of a single variable, normal / abnormal is classified according to the setting rule EDA for Outlier by PCA In the case of data composed of several variables, the data distribution before PCA application is visually examined Outlier by PCA PCA-applied new variables distinguish normal / abnormal Logistic Regression Learn from labeled data and test normal / abnormal Save R Object Store the learned discriminant model as an object Output DB Normal / abnormal data is stored

<State subdivision model>

o Basic Theory

In the present invention, subdivision reflects learning characteristics as continuous data accumulates instead of fixed data. The SOM (Self Organizing Map) is selected as a subdivision model.

SOM was first introduced by Teuvo Kohonen of the University of Helsinki and is also known as the Kohonen map as a method of autonomous learning based on weighted artificial neural cell grids that matched the input vector to the training set.

SOM is most easily described as a collection of artificial neurons that have their physical location in the output map and participate in the winner-take process. The winner-take-all process, a kind of competitive neural network, declares a node with the weight vector closest to the input vector as the winner, and the weight is adjusted so that the value is closer to the input vector.

Each node has a set of neighbors. When these nodes win in competition, the weights of neighbors also change. They do not change equally. As the neighbors move away from the winner, their weights change smaller. This process is repeated for each input vector to perform many cycles. If the input is different, the winner is also different.

The SOM network is terminated by associating output nodes with groups or patterns in the input data set. If such a pattern can be given a name, its name can be attached to the relevant node in the training network.

o Model Workflow

4 is a diagram for explaining a model workflow.

Referring to FIG. 4, the following will be described. That is, a step of processing raw data and processing an ideal value is added so as to be suitable for SOM learning. Since SOM is dynamic and changing characteristics, we add tasks to control the number of clusters, map size, and initial value setting. Since the number of subdivisions coming out of SOM is usually large, we need to add a second subdivision to reduce the number of subdivisions.

o Role of model stage

step role Remarks Missing value Remove Missing Value Derived variable Generate calculated data based on source data Class Indicator Change categorization data to Matrix type suitable for learning Remove variable Identify and remove variables that are not suitable for learning by applying rules Missing value Remove Missing Value A null value generated from a derived variable. Correlation coefficient Identify highly correlated variables and remove one of them Ideal value Correct anomalies that can have a negative effect on learning Cluster variable Selection of optimal parameters for clustering algorithm scale Scale learning data. Cluster number Select the optimal number of clusters Map Size Choose the best map size Initial value Optimum seed value selected SOM Data clustering Secondary community Subdivided into general k-means Selection task

&Lt; Basic analysis function according to the present invention &

o Basic analysis functions

The basic analysis function according to the present invention is applicable not only to a manufacturing process but also to general sensor analysis in general.

Such basic analysis functions may be, for example, as follows.

(1) Future N period trend line

(2) Data Distribution and Outliers

(3) Correlation of two or more data

<Future N period trend line>

o Application Scenario

1) Select the object (sensor).

2) Select the period.

3) Pass the input information to the prediction module.

4) Calculate the predicted value and deliver it to the device according to the invention.

5) Show the forecast trend line.

In this case, since it is a single variable, it is necessary to apply a simple regression model or a time series, and it is necessary to test it by type because the appropriate model and parameters are different for each sensor type. Also, it is necessary to periodically update the model (parameter).

<Data distribution and outliers>

o Single Sensor Comparison

- Time zone, day of the week, seasonal comparison of one sensor

o Comparison between homologous sensors

- Comparison between the same sensors (temperature sensors 1, 2, 3, 4)

- Comparison by location (machine)

o Application Scenario

1) Select the object (sensor).

2) Select whether to select single sensor or homogeneous sensor.

3) Select the condition (place, time, etc.).

4) Select the fault range parameter.

5) Pass the input information to the analysis module.

6) Calculate and transmit anomalies to the device according to the invention.

7) The apparatus according to the present invention receives the abnormal value information calculated in the distribution display and analysis module in its own chart,

<Correlation for two or more variables>

o correlation between two variables

When two variables are selected, a comparison between two data (eg temperature and humidity) is performed, showing the correlation coefficient value (eg 67%) and the presence of correlation (eg with / without correlation).

o Comparison between three or more variables

Using a matrix or a network chart, the correlation coefficient values for each data pair and the presence or absence of correlation are displayed.

o Application Scenario

1) Select the target.

2) Pass the input information to the analysis module.

3) The correlation is calculated and transmitted to the device according to the present invention.

4) The device according to the present invention displays the correlation on its own chart, receives and displays the correlation coefficient calculated by the analysis module and the presence or absence of correlation.

5 is a diagram for explaining a function of managing data in a device according to the present invention.

Referring to FIG. 5, the following will be described. First, Query-set creates a query automatically when you specify the analysis items and conditions from the data model (indexing model) and stores it as a query set. Because the analysis uses the query set rather than the data, the analysis result and the report are always updated with the latest data.

In addition, the data set defines the execution result data of the analysis model, and a plurality of model results can be defined as one data set. It can be used for dashboard visualization and reporting in the same way as a query set.

Mart change management is to detect changes through comparison of related tables in the mart, and errors in reports often result from data abnormalities (errors due to changes in source data types are typical).

The impact analysis shows that the change or deletion of specific data tracks the impact on the analysis model using the data and ultimately on the analysis report.

The apparatus and method according to the invention are particularly useful for smart factories. The Smart Factory, which is being recently proposed, has been increasingly required for the diagnosis and the real-time monitoring / analysis of the manufacturing environment as the manufacturing process and equipment in order to improve the productivity and secure the work safety of the small and medium sized manufacturing companies. Application is very effective for production / process / environmental monitoring and analysis.

From the manager's point of view, as production management, it is possible to monitor real-time production performance by factory and line, monthly production output, and production volume to target. In terms of quality management, SPC (Statistical Process Control) The defective number and defect rate management and the process capability index (Cpk) can be effectively performed.

From the operator's point of view, in terms of process management, it can be used to identify and analyze equipment utilization rate by manufacturing line / process, analyze the cause of failure / failure of each process, Preventive maintenance and response (predictive maintenance) can be performed.

Also, in terms of environmental management, it is possible to perform environmental monitoring such as temperature, humidity, and dust in a plant that affects operators, products, and facilities.

6 is a screen showing a facility management function.

As shown in FIG. 6, preventive maintenance and predictive maintenance can be performed through real-time facility status monitoring and error alarm response.

FIG. 7 is a diagram for explaining a process of detecting a process status abnormality by applying a big data technique.

As shown in FIG. 7, abnormal process can be detected by learning the process equipment period signal and learning SOM (Self Organizing Map) clustering.

8 is a diagram for explaining a process of performing a failure prediction and a cause search by applying a big data technique.

As shown in FIG. 8, a big data technique is used to derive a failure type and an influence factor, and an abnormality is detected. Based on the current state of the influence factors and the recent change degree of influence factors, Can be performed.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments.

Claims (6)

A data analysis method for predicting equipment failure in a manufacturing process utilizing a plurality of sensors,
A first step of generating a composite sensor pattern from a pattern from a plurality of sensors;
A second step of learning the composite sensor pattern in either a steady state or an abnormal state;
A third step of storing the complex sensor pattern in the abnormal state database together with the error information of the manufacturing equipment when it is learned in an abnormal state; And
A fourth step of performing a predictive analysis on the abnormality of the equipment coupled with the sensor based on the complex sensor pattern stored in the abnormal state database
Wherein the plurality of sensors are used to predict an equipment failure in a manufacturing process.
The method according to claim 1,
Wherein the fourth step includes setting an abnormal value discrimination model capable of discriminating whether or not a normal state is present.
3. The method of claim 2,
And learning a complex sensor pattern using a SOM (Self Organizing Map) refinement model. 2. The method of claim 1,
The method according to claim 1,
Wherein a composite sensor pattern is generated by correlation values of input signals from a plurality of sensors and presence or absence of a correlation, wherein the plurality of sensors are used to predict an equipment failure in a manufacturing process.
5. The method of claim 4,
And performing a predictive analysis based on the correlation coefficient and the presence or absence of a correlation between the plurality of sensors.

1. A data analysis apparatus utilizing a plurality of sensors for predicting equipment failure in a manufacturing process,
A sensor pattern generation unit for generating a composite sensor pattern from a pattern from a plurality of sensors;
A learning unit for learning the composite sensor pattern in either a steady state or an abnormal state;
A pattern management unit for storing the complex sensor pattern in the abnormal state database together with the error information of the manufacturing equipment when it is learned in an abnormal state; And
A predictive analysis unit for performing a predictive analysis on the presence or absence of abnormality of a sensor-coupled equipment based on the complex sensor pattern stored in the abnormal state database,
Wherein the plurality of sensors are used to predict an equipment failure in a manufacturing process.

Images