CN117556331B

CN117556331B - AI-enhancement-based air compressor maintenance decision method and system

Info

Publication number: CN117556331B
Application number: CN202410038605.3A
Authority: CN
Inventors: 周振华; 白纯波; 萧景业; 黄泽应
Original assignee: Guangdong Alink Energy Equipment Co ltd
Current assignee: Guangdong Alink Energy Equipment Co ltd
Priority date: 2024-01-11
Filing date: 2024-01-11
Publication date: 2024-04-26
Anticipated expiration: 2044-01-11
Also published as: CN117556331A

Abstract

The embodiment of the application provides an AI-enhanced air compressor maintenance decision method and system, which are characterized in that a plurality of sample air compressor operation data subsequences are analyzed in a sample air compressor operation data sequence, multi-element attribute embedding representation is carried out on template operation trend data, parameter learning is carried out on an initialized neural network after multi-element attribute embedding representation information of each template operation trend data in the sample air compressor operation data subsequence is generated, a reference neural network is generated, a target air compressor maintenance decision network is configured based on functional layer parameter information of the reference neural network, decision effect indexes are generated for decision effects of each fault maintenance category by the target air compressor maintenance decision network, and the fault maintenance category of the target air compressor is decided in the air compressor operation data of the target air compressor based on the target air compressor maintenance decision network according to the decision effect indexes, so that the accuracy and efficiency of air compressor maintenance decision can be improved.

Description

AI-enhancement-based air compressor maintenance decision method and system

Technical Field

The application relates to the technical field of artificial intelligence, in particular to an AI-enhanced air compressor maintenance decision method and system.

Background

The air compressor is an important device in industrial production, and the running state of the air compressor directly influences the production efficiency and the safety. Therefore, the method has extremely important significance in maintenance and fault prediction of the air compressor. The traditional air compressor maintenance method mainly depends on manual inspection and experience judgment, and has certain limitations, such as time consumption, high misjudgment rate, incapability of realizing real-time monitoring and the like, namely, the method is easily influenced by subjective factors and can not adapt to complex and changeable environments.

Disclosure of Invention

In order to at least overcome the defects in the prior art, the application aims to provide an AI-enhanced air compressor maintenance decision method and system.

In a first aspect, the present application provides an AI-enhanced air compressor maintenance decision method, applied to an air compressor maintenance system, the method comprising:

Obtaining a sample air compressor operation data sequence, and analyzing a plurality of sample air compressor operation data subsequences in the sample air compressor operation data sequence, wherein the sample air compressor operation data subsequences comprise template operation trend data of at least one sample air compressor;

Performing multi-attribute embedded representation on the template operation trend data to generate multi-attribute embedded representation information of each template operation trend data in the sample air compressor operation data subsequence;

according to the multi-element attribute embedded representation information, parameter learning is carried out on the initialized neural network, and a reference neural network corresponding to each sample air compressor operation data subsequence is generated;

configuring a target air compressor maintenance decision network based on the functional layer parameter information of the reference neural network, and verifying the decision effect of the target air compressor maintenance decision network for each fault maintenance category to generate a decision effect index;

And deciding the fault maintenance category of the target air compressor in the air compressor operation data of the target air compressor based on the maintenance decision network of the target air compressor according to the decision effect index.

In a possible implementation manner of the first aspect, the acquiring a sample air compressor operation data sequence includes:

Acquiring the operation data of the air compressors of each sample in the sample air compressor sequence under the condition that at least one operation process is executed, and generating template air compressor operation data of each sample air compressor;

Acquiring a priori diagnosis fault maintenance category of the sample air compressor, and determining a current fault maintenance category of the sample air compressor according to the priori diagnosis fault maintenance category;

marking the current fault maintenance category in the operation data of the template air compressor, and generating a sample air compressor operation data sequence;

the embedding of the representation information according to the multiple attributes carries out parameter learning on the initialized neural network to generate a reference neural network corresponding to each sample air compressor operation data subsequence, and the method comprises the following steps:

Deciding a target sample air compressor corresponding to each multi-element attribute embedded representation information in the sample air compressor, generating a directed knowledge graph according to the initialized neural network, generating directed knowledge graph data of the target sample air compressor, performing fault maintenance type prediction on the directed knowledge graph data, and generating decision fault maintenance types of the target sample air compressor;

Determining a target decision error parameter corresponding to each sample air compressor operation data subsequence according to the decision fault maintenance category;

and based on the target decision error parameters, performing parameter learning on the initialized neural network to generate a reference neural network corresponding to each sample air compressor operation data subsequence.

In a possible implementation manner of the first aspect, the performing parameter learning on the initialized neural network based on the target decision error parameter, to generate a reference neural network corresponding to each sample air compressor operation data subsequence, includes:

Extracting at least one undetermined function layer parameter information from a preset function layer parameter information sequence of the initialized neural network according to the target decision error parameter, and generating an undetermined function layer parameter information sequence;

Optimizing the undetermined function layer parameter information sequence based on the target decision error parameter to generate an optimized function layer parameter information sequence and an iterative air compressor maintenance decision network corresponding to the optimized function layer parameter information sequence;

And taking the optimized function layer parameter information sequence as the preset function layer parameter information sequence, taking the iterative air compressor maintenance decision network as the initialized neural network, and continuing to perform iterative operation until the target decision error parameter is smaller than a threshold error parameter, so as to generate a reference neural network corresponding to each sample air compressor operation data subsequence.

In a possible implementation manner of the first aspect, the configuring the target air compressor maintenance decision network based on the functional layer parameter information of the reference neural network includes:

acquiring performance measurement indexes of the reference neural network;

Extracting at least one target function layer parameter information from the function layer parameter information according to the performance measurement index;

and configuring the parameter information of the target functional layer in the initialized neural network to generate a maintenance decision network of the target air compressor.

acquiring performance measurement indexes of the reference neural network;

according to the performance measurement index, carrying out priority arrangement on the reference neural network;

and extracting a maintenance decision network of the target air compressor from the reference neural network based on the priority sorting result.

In a possible implementation manner of the first aspect, after the configuring the target air compressor maintenance decision network based on the functional layer parameter information of the reference neural network, the method further includes:

Acquiring at least one extended template operation trend data, wherein the extended template operation trend data comprises a current tag fault maintenance category;

Optimizing the functional layer parameter information of the target air compressor maintenance decision network according to the current tag fault maintenance category to generate a current air compressor maintenance decision network;

and based on the operation trend data of the expansion template, parameter learning is carried out on the current air compressor maintenance decision network, and the air compressor maintenance decision network after parameter learning is used as the target air compressor maintenance decision network.

In a possible implementation manner of the first aspect, the verifying the decision effect of the target air compressor maintenance decision network for each fault maintenance category generates a decision effect indicator, including:

The sample air compressor operation data sequence is separated into N sample air compressor operation data subsequences, N test sample data sequences are generated, the test sample data sequences comprise template operation trend data of the current sample air compressor corresponding to each fault maintenance type, and N is a preset positive integer greater than 2;

Determining a check sample data sequence from the test sample data sequence, and taking the test sample data sequences except the check sample data sequence in the test sample data sequence as a learning sample data sequence;

Parameter learning is carried out on the target air compressor maintenance decision network according to the learning sample data sequence, and the fault maintenance category of the current sample air compressor corresponding to the checking sample data sequence is decided by the target air compressor maintenance decision network after the parameter learning, so that the current decision fault maintenance category of the current sample air compressor is generated;

Acquiring a label fault maintenance category of the current sample air compressor, determining at least one basic decision effect index corresponding to the check sample data sequence according to the label fault maintenance category and the current decision fault maintenance category, acquiring an influence coefficient corresponding to each basic decision effect index, fusing the basic decision effect indexes according to the influence coefficient, calculating the average value of the fused basic decision effect indexes, and generating an initial decision effect index corresponding to the check sample data sequence;

returning to the step of determining the check sample data sequence in the test sample data sequence until each test sample data sequence is the check sample data sequence, and generating an initial decision effect index corresponding to each test sample data sequence;

And carrying out weighted calculation on the initial decision effect index to generate the decision effect index of the maintenance decision network of the target air compressor.

In a possible implementation manner of the first aspect, the determining, according to the tag fault maintenance category and the current decision fault maintenance category, at least one basic decision effect indicator corresponding to the check sample data sequence includes:

comparing the label fault maintenance category with the current decision fault maintenance category to generate fault maintenance category decision detail information of each current sample air compressor;

According to the detailed information of the fault maintenance class decision, statistics of effective instance statistics and misjudgment instance statistics are carried out in the check sample data sequence, wherein the effective instance statistics comprise instance statistics with reliable fault maintenance class decision, and the misjudgment instance statistics comprise instance statistics with unreliable fault maintenance class decision;

Acquiring instance statistics of the check sample data sequence, and determining at least one basic decision effect index corresponding to the check sample data sequence according to the instance statistics, the effective instance statistics and the misjudgment instance statistics;

The instance statistics comprise positive instance statistics and negative instance statistics, the effective instance statistics comprise effective positive instance statistics and effective negative instance statistics, the misjudgment instance statistics comprise misjudgment positive instance statistics and misjudgment negative instance statistics, and the determining at least one basic decision effect index corresponding to the check sample data sequence according to the instance statistics, the effective instance statistics and the misjudgment instance statistics comprises:

Determining at least one basic verification performance index corresponding to the verification sample data sequence according to the positive instance statistic, the negative instance statistic, the effective positive instance statistic, the effective negative instance statistic, the misjudgment positive instance statistic and the misjudgment negative instance statistic;

Extracting at least one target verification performance index from the basic verification performance indexes, and carrying out weighted calculation on the target verification performance indexes to generate fusion verification performance indexes;

And determining a current performance measurement index corresponding to the check sample data sequence according to the effective example statistic, the misjudgment negative example statistic and the example statistic, and taking the basic check performance index, the fusion check performance index and the current performance measurement index as the basic decision effect index.

In a possible implementation manner of the first aspect, the determining, according to the decision effect index, the fault maintenance category of the target air compressor in the air compressor operation data of the target air compressor based on the target air compressor maintenance decision network includes:

When the decision effect index characterizes that the maintenance decision network of the target air compressor achieves a preset decision effect, acquiring air compressor operation data of the target air compressor when at least one operation process is executed;

performing multi-attribute embedded representation on the operation data of the air compressor to generate target multi-attribute embedded representation information of the target air compressor;

and deciding the fault maintenance category of the target air compressor pair in the target multi-element attribute embedded representation information according to the target air compressor maintenance decision network.

In a second aspect, an embodiment of the present application further provides an air compressor maintenance system, where the air compressor maintenance system includes a processor and a machine-readable storage medium, where the machine-readable storage medium stores a computer program, and the computer program is loaded and executed in conjunction with the processor to implement the AI-enhanced air compressor maintenance decision method of the first aspect above.

According to the technical scheme, the sample air compressor operation data sequence is obtained, the sample air compressor operation data subsequences are analyzed in the sample air compressor operation data sequence, the multi-element attribute embedding representation is carried out on the template operation trend data, multi-element attribute embedding representation information of each template operation trend data in the sample air compressor operation data subsequences is generated, parameter learning is carried out on the initialized neural network according to the multi-element attribute embedding representation information, the reference neural network corresponding to each sample air compressor operation data subsequence is generated, the target air compressor maintenance decision network is configured based on the functional layer parameter information of the reference neural network, the decision effect index of the target air compressor maintenance decision network for each fault maintenance class is generated, and the fault maintenance class of the target air compressor is decided in the air compressor operation data of the target air compressor according to the decision effect index, so that the accuracy and the efficiency of the air compressor maintenance decision can be improved.

That is, according to the embodiment of the application, the operation data sequence of the sampling example air compressor is obtained, the plurality of subsequences are analyzed, and the operation trend data of the templates are expressed in a multi-element attribute embedded manner, so that each operation trend data of the templates can be accurately and comprehensively described, and the accuracy of the fault prediction and maintenance decision of the air compressor is improved. And the initialized neural network is subjected to parameter learning by utilizing the multi-attribute embedded representation information to generate a reference neural network, so that the learning capacity and the prediction accuracy of the neural network are improved, and the complexity of model training is reduced. The target air compressor maintenance decision network is configured based on the functional layer parameter information of the reference neural network, and the decision effect is verified, so that the performance of the decision network can be evaluated, and the network parameters can be adjusted according to actual needs to further optimize the decision effect. Finally, according to the decision effect index, the fault maintenance type of the target air compressor is decided by using a maintenance decision network of the target air compressor, so that automatic and intelligent fault diagnosis and maintenance of the air compressor are realized, and the running efficiency and stability of the air compressor are improved. Therefore, by adopting the air compressor maintenance decision method based on the neural network, the accurate prediction and the timely maintenance of the faults of the air compressor are realized, the operation efficiency and the stability of the air compressor are obviously improved, and the maintenance cost is reduced.

Drawings

For a clearer description of the technical solutions of the embodiments of the present application, reference will be made to the accompanying drawings, which are needed to be activated in the embodiments, and it should be understood that the following drawings only illustrate some embodiments of the present application and should not be considered as limiting the scope, and that other related drawings can be obtained by those skilled in the art without the inventive effort.

Fig. 1 is a schematic flow chart of an AI-enhanced air compressor maintenance decision method according to an embodiment of the present application;

fig. 2 is a schematic block diagram of network composition data of an air compressor maintenance system for implementing the AI-enhanced air compressor maintenance decision method according to an embodiment of the present application.

Detailed Description

The following description is presented to enable one of ordinary skill in the art to make and use the application and is provided in the context of a particular application and its requirements. It will be apparent to those having ordinary skill in the art that various changes can be made to the disclosed embodiments and that the general principles defined herein may be applied to other embodiments and applications without departing from the principles and scope of the application. Therefore, the present application is not limited to the described embodiments, but is to be accorded the widest scope consistent with the claims.

Referring to fig. 1, the application provides an AI-enhancement-based air compressor maintenance decision method, which comprises the following steps.

Step S110, a sample air compressor operation data sequence is obtained, and a plurality of sample air compressor operation data subsequences are analyzed in the sample air compressor operation data sequence, wherein the sample air compressor operation data subsequences comprise template operation trend data of at least one sample air compressor.

For example, an air compressor operation database of a certain factory in the past year can be stored in the cloud server, so that a sample air compressor operation data sequence can be constructed based on the air compressor operation databases, and the sample air compressor operation data sequence can comprise parameters such as operation time, operation temperature, exhaust pressure and the like of each sample air compressor, and maintenance records of each fault occurrence. The aim of the embodiment is to construct a model capable of automatically identifying the fault maintenance type of the air compressor by utilizing the subsequences of the operation data of the air compressor of the samples.

For example, first, the air compressor operation data may be arranged in a time sequence to form a sample air compressor operation data sequence, where the sample air compressor operation data sequence may include operation data of each sample air compressor in the past year. Next, a plurality of subsequences containing at least one sample air compressor operation data are parsed from this sample air compressor operation data sequence. For example, each month of data may be used as a sub-sequence of sample air compressor operation data, thus obtaining 12 sub-sequences of sample air compressor operation data.

That is, the sample air compressor operation data sequence may refer to a set of sequences including operation data of a plurality of air compressors. For example, assuming that a set of operation data including 10 air compressors in the past year is provided, the set of data is a sample air compressor operation data sequence.

The sample air compressor operation data subsequence refers to a subset extracted from the sample air compressor operation data sequence, and includes template operation trend data of at least one sample air compressor. For example, assuming that data of each month is extracted from the above-described sample air compressor operation data sequence as one subsequence, 12 subsequences are obtained.

The template operation trend data is data which is extracted from the operation data of the sample air compressor and can reflect the operation state of the air compressor. For example, dynamic parameters such as operation time, operation temperature, exhaust pressure and the like can be extracted from the operation data of the air compressor as the template operation trend data. That is, the operation trend data can reflect a change trend of the operation state of the air compressor. For example, a trend of change in the operation temperature per hour may be extracted from the operation data of the air compressor as the operation trend data.

And step S120, performing multi-attribute embedded representation on the template operation trend data, and generating multi-attribute embedded representation information of each template operation trend data in the sample air compressor operation data subsequence.

For example, for each sample air compressor operation data subsequence, the template operation trend data therein needs to be represented by multi-attribute embedding. For example, each operational trend data may be mapped into a high-dimensional space using word embedding techniques. Specifically, each operational trend data may be represented as a high-dimensional vector using the Word2Vec algorithm. Thus, the high-dimensional space coordinates of each running trend data can be embedded as its multivariate attribute into the representation information. Specifically, the high-dimensional vector of each operational trend data may be embedded as its multivariate attribute with the representation information.

For example, assuming that there is a sub-sequence of operating data for one example air compressor, which contains operating data for two air compressors, each air compressor has 10 operating time data, it is now necessary to embed multiple attributes into the data.

First, it is necessary to determine which properties can be used as the multivariate properties. In this example, operating time, operating temperature, and exhaust pressure may be selected as the multivariate attributes.

Next, a multivariate attribute embedded representation needs to be made for each runtime data. Specifically, the data at each runtime may be represented as a high-dimensional vector using Word embedding techniques, such as Word2 Vec. For example, the run time, the run temperature, and the exhaust pressure for each run time may be represented as a high-dimensional vector, respectively, and then the three vectors may be stitched into a final multi-element attribute embedded representation vector. Finally, the multivariate attribute embedded representation information is required to be generated for each template operational trend data. Specifically, the multi-attribute embedded representation vector for each runtime may be used as multi-attribute embedded representation information for the template runtime trend data for that runtime. For example, for the first air compressor, the operation time, the operation temperature and the exhaust pressure of each operation time can be respectively represented as a high-dimensional vector, and then the three vectors are spliced into a final multi-attribute embedded representation vector, and the final multi-attribute embedded representation vector is used as multi-attribute embedded representation information of the template operation trend data of the operation time. Similarly, for the second air compressor, the same steps may be taken to generate the multi-attribute embedded representation vector for each run-time. In this way, a multi-attribute embedded representation of the template operational trend data is achieved, and multi-attribute embedded representation information for each of the template operational trend data is generated.

And step S130, carrying out parameter learning on the initialized neural network according to the multi-element attribute embedded representation information, and generating a reference neural network corresponding to each sample air compressor operation data subsequence.

For example, next, representation information needs to be embedded according to multiple attributes of each template operation trend data, parameter learning is performed on the initialized neural network, and a reference neural network corresponding to each sample air compressor operation data subsequence is generated. For example, a back propagation technique may be used to minimize the prediction error by training the neural network. Specifically, a three-layer feedforward neural network may be used, and a high-dimensional vector of each operation trend data is taken as an input, and a fault maintenance category corresponding to the operation trend data is output.

Step S140, configuring a target air compressor maintenance decision network based on the functional layer parameter information of the reference neural network, and verifying the decision effect of the target air compressor maintenance decision network for each fault maintenance category to generate a decision effect index.

For example, some functional layer parameter information in the reference neural network may be directly copied into the target air compressor maintenance decision network, or parameters in the reference neural network may be used as initial parameters of the target air compressor maintenance decision network. Specifically, the first layer and the second layer of each reference neural network may be used as the first layer and the second layer of the target air compressor maintenance decision network, and the parameters in the reference neural network may be used as initial parameters of the corresponding layer in the target air compressor maintenance decision network.

After the maintenance decision network of the target air compressor is configured, the decision effect of the network aiming at each fault maintenance type needs to be verified, and a decision effect index is generated. For example, some test data may be used to evaluate the classification accuracy of the network for different fault maintenance categories. Specifically, the last operation trend data in each sub-sequence can be used as test data, the output result of the maintenance decision network of the target air compressor is compared with the actual fault maintenance category, and the classification accuracy is calculated.

And step S150, deciding the fault maintenance type of the target air compressor in the air compressor operation data of the target air compressor based on the target air compressor maintenance decision network according to the decision effect index.

For example, the output result of the target air compressor maintenance decision network may be used as a final fault maintenance category decision. Specifically, if the classification accuracy of the target air compressor maintenance decision network is higher than a certain threshold, the target air compressor maintenance decision network can be considered to be well learned, and the output result of the target air compressor maintenance decision network can be used as a final fault maintenance class decision. If the classification accuracy is below the threshold, it may be necessary to retrain the target air compressor maintenance decision network or adjust the network structure of the target air compressor maintenance decision network.

Based on the steps, a sample air compressor operation data sequence is obtained, a plurality of sample air compressor operation data subsequences are analyzed in the sample air compressor operation data sequence, multi-element attribute embedding representation is carried out on the template operation trend data, multi-element attribute embedding representation information of each template operation trend data in the sample air compressor operation data subsequences is generated, parameter learning is carried out on an initialized neural network according to the multi-element attribute embedding representation information, a reference neural network corresponding to each sample air compressor operation data subsequence is generated, a target air compressor maintenance decision network is configured based on functional layer parameter information of the reference neural network, decision effect indexes are generated according to the decision effect indexes according to the decision effect of the target air compressor maintenance decision network for each fault maintenance category, and the fault maintenance category of the target air compressor is decided in the air compressor operation data of the target air compressor based on the target air compressor maintenance decision network, so that the accuracy and the efficiency of the air compressor maintenance decision can be improved.

In one possible implementation, step S110 may include:

Step S111, acquiring air compressor operation data of each sample air compressor in the sample air compressor sequence under the condition that at least one operation process is executed, and generating template air compressor operation data of each sample air compressor.

Step S112, obtaining a priori diagnosis fault maintenance category of the sample air compressor, and determining a current fault maintenance category of the sample air compressor according to the priori diagnosis fault maintenance category.

And step S113, marking the current fault maintenance category in the operation data of the template air compressor, and generating a sample air compressor operation data sequence.

For example, assume that a set of operation data including 10 air compressors has 10 operation time data for each air compressor. These data may be used as a sample air compressor operation data sequence. Then, parameters such as operation time, operation temperature, exhaust pressure and the like can be extracted from the data of each operation time as the operation data of the template air compressor.

On this basis, it is assumed that 5 of the 10 air compressors are known to have failed in different types, and the other 5 have not failed. Such fault information may be used as a priori diagnostic fault maintenance category. Then, according to the fault information, the current fault maintenance type of each air compressor, namely, which type of fault occurs, can be determined.

And finally, marking the corresponding current fault maintenance category in the operation data of each template air compressor to generate a sample air compressor operation data sequence. For example, for an air compressor that has failed, the corresponding failure type may be noted in its template air compressor operation data, while for an air compressor that has not failed, the "normal" may be noted in its template air compressor operation data. Thus, a sample air compressor operation data sequence containing 10 air compressors is obtained, and the data of each air compressor is marked with the current fault maintenance category.

In one possible implementation, step S130 may include:

step S131, deciding a target sample air compressor corresponding to each multi-attribute embedded representation information in the sample air compressor, generating a directed knowledge graph according to the initialized neural network, generating directed knowledge graph data of the target sample air compressor, performing fault maintenance type prediction on the directed knowledge graph data, and generating a decision fault maintenance type of the target sample air compressor.

For example, assume that the sample air compressor operation data subsequence contains 10 operation time data. The 10 pieces of data at the running time can be subjected to multi-attribute embedded representation by using an initialized neural network, so that multi-attribute embedded representation information at each running time is obtained. Then, a decision algorithm can be used to decide the target sample air compressor corresponding to each multi-attribute embedded representation information from the 10 pieces of data at the running time. For example, the 10 runtime data may be classified into 5 classes, each class representing a target sample air compressor, using a K-means clustering algorithm. And then, using an initialized neural network to generate a directed knowledge graph of the multi-attribute embedded representation information of each target sample air compressor, and generating directed knowledge graph data of each target sample air compressor.

Illustratively, the multi-attribute embedded representation information of the target sample air compressor is a two-dimensional vector. The two-dimensional vector may be mapped into a high-dimensional space using an initializing neural network to obtain a high-dimensional vector. The high-dimensional vector can then be converted into a directed knowledge graph data using a knowledge graph generation algorithm.

In particular, a knowledge graph generation algorithm based on a self-attention mechanism may be used to convert the high-dimensional vector into a directed knowledge graph data. This algorithm can be divided into the following steps:

1. using a self-attention mechanism, the correlation between each element and other elements in the high-dimensional vector is calculated.

2. Each element in the high-dimensional vector is divided into different categories according to the correlation.

3. For each category, a self-attention mechanism is used to calculate the correlation between each element and other elements in the category.

4. Each element in each category is connected to other related elements according to the correlation to form a directed knowledge graph data.

For example, assume the following high-dimensional vectors:

[1, 2, 3, 4, 5]

The correlation between each element and the other elements in the high-dimensional vector can be calculated using a self-attention mechanism, resulting in the following correlation matrix:

[1, 0.5, 0.3, 0.2, 0.1]

[0.5, 1, 0.6, 0.4, 0.3]

[0.3, 0.6, 1, 0.8, 0.7]

[0.2, 0.4, 0.8, 1, 0.9]

[0.1, 0.3, 0.7, 0.9, 1]

Each element in the high-dimensional vector may be divided into different categories according to the correlation matrix. For example, a first element may be divided into a first class, a second element into a second class, and so on.

For each category, a self-attention mechanism may be used to calculate the correlation between each element and other elements in the category. For example, for the first class, the following correlation matrix may be calculated:

[1, 0.5, 0.3, 0.2, 0.1]

[0.5, 1, 0.6, 0.4, 0.3]

[0.3, 0.6, 1, 0.8, 0.7]

[0.2, 0.4, 0.8, 1, 0.9]

[0.1, 0.3, 0.7, 0.9, 1]

Each element in each category may be connected to other related elements according to the correlation matrix to form a directed knowledge graph data. For example, a first element may be connected to a second element and a third element, a second element may be connected to the first element, the third element, and the fourth element, and so on.

In this way, directed knowledge graph data of a target sample air compressor is generated.

Finally, a classification algorithm can be used for predicting the fault maintenance category of the directed knowledge graph data of each target sample air compressor, and the decision fault maintenance category of each target sample air compressor is generated.

And step S132, determining a target decision error parameter corresponding to each sample air compressor operation data subsequence according to the decision fault maintenance category.

For example, assuming that a decision fault maintenance class has been obtained for each target sample air compressor, a loss function may be used to calculate a target decision error parameter for each target sample air compressor operation data subsequence. For example, a cross entropy loss function may be used to calculate a target decision error parameter corresponding to each target sample air compressor operation data subsequence.

The method of calculation of the target decision error parameter is illustratively dependent on the selected loss function. Common loss functions include Mean Square Error (MSE), cross entropy loss (Cross Entropy Loss), and the like.

Taking the mean square error as an example, it is assumed that there is a target decision error parameter θ, which represents the difference between the predicted value and the actual value. The target decision error parameter θ may be calculated using the following formula:

θ = ∑(y_true - y_pred)^2 / N

Where y_true represents the actual value, y_pred represents the predicted value, and N represents the number of samples.

Taking the cross entropy loss as an example, it is assumed that there is a target decision error parameter θ, which represents the difference between the predicted value and the actual value. The target decision error parameter θ may be calculated using the following formula:

θ = - ∑(y_true * log(y_pred) + (1 - y_true) * log(1 - y_pred)) / N

And step S133, based on the target decision error parameters, performing parameter learning on the initialized neural network to generate a reference neural network corresponding to each sample air compressor operation data subsequence.

Finally, a gradient descent algorithm can be used for carrying out parameter learning on the initialized neural network based on each target decision error parameter, and a reference neural network corresponding to each sample air compressor operation data subsequence is generated. For example, an Adam optimization algorithm may be used to perform parameter learning on the initialized neural network based on each target decision error parameter, so as to generate a reference neural network corresponding to each sample air compressor operation data subsequence.

In one possible embodiment, step S133 may include:

Step S1331, extracting at least one undetermined function layer parameter information from the preset function layer parameter information sequence of the initialized neural network according to the target decision error parameter, and generating an undetermined function layer parameter information sequence.

For example, a gradient descent algorithm may be used to update the parameter information of each functional layer in the initialized neural network on the basis of the target decision error parameter corresponding to the operation data subsequence of each sample air compressor, and generate the pending functional layer parameter information sequence.

Assuming that the initialized neural network has a preset functional layer parameter information sequence of W and a target decision error parameter of E, the undetermined functional layer parameter information sequence may be expressed as:

W' = W - α * ∇E(W)

Where α is the learning rate and ∇ E (W) is the gradient of the target decision error parameter with respect to the pre-set functional layer parameter information sequence.

Step S1332, optimizing the undetermined functional layer parameter information sequence based on the target decision error parameter, and generating an optimized functional layer parameter information sequence and an iterative air compressor maintenance decision network corresponding to the optimized functional layer parameter information sequence.

For example, a back propagation algorithm may be used to update the parameter information of each functional layer in the pending functional layer parameter information sequence based on the target decision error parameter corresponding to the operation data subsequence of each sample air compressor, and generate the optimized functional layer parameter information sequence. Meanwhile, a new iterative air compressor maintenance decision network can be constructed by using the optimized functional layer parameter information sequence.

Assuming that the pending function layer parameter information sequence is W', the optimized function layer parameter information sequence may be expressed as:

W'' = W' - β * ∇E(W')

where β is the learning rate and ∇ E (W') is the gradient of the target decision error parameter with respect to the pending functional layer parameter information sequence.

Meanwhile, a new iterative air compressor maintenance decision network can be constructed by using the optimized functional layer parameter information sequence.

Step S1333, taking the optimized functional layer parameter information sequence as the preset functional layer parameter information sequence, taking the iterative air compressor maintenance decision network as the initialized neural network, and continuing to perform iterative operation until the target decision error parameter is smaller than a threshold error parameter, so as to generate a reference neural network corresponding to each sample air compressor operation data subsequence.

For example, a gradient descent algorithm may be used to update parameter information of each functional layer in the iterative air compressor maintenance decision network on the basis of the target decision error parameter corresponding to the operation data subsequence of each sample air compressor, and generate a new optimized functional layer parameter information sequence. Meanwhile, a new iterative air compressor maintenance decision network can be constructed by using a new optimized functional layer parameter information sequence. The process is repeated until the target decision error parameter is smaller than the threshold error parameter, and a reference neural network corresponding to each sample air compressor operation data subsequence is generated.

For example, assuming that the optimized function layer parameter information sequence is W ", the new preset function layer parameter information sequence may be expressed as:

W''' = W''

Meanwhile, a new iterative air compressor maintenance decision network can be constructed by using a new preset functional layer parameter information sequence. The process is repeated until the target decision error parameter is smaller than the threshold error parameter, and a reference neural network corresponding to each sample air compressor operation data subsequence is generated.

In one possible implementation, step S140 may include: and acquiring performance measurement indexes of the reference neural network, extracting at least one target functional layer parameter information from the functional layer parameter information according to the performance measurement indexes, configuring the target functional layer parameter information in the initializing neural network, and generating a target air compressor maintenance decision network.

For example, some performance metrics may be used to evaluate the performance of this reference neural network. For example, the performance of the reference neural network may be evaluated using indices of accuracy, recall, F1 score, etc. That is, the performance metric indicators may be used to extract at least one target functional layer parameter information from the functional layer parameter information of the reference neural network. For example, a gradient descent algorithm may be used to extract at least one target functional layer parameter information from the functional layer parameter information of the reference neural network, such that the target functional layer parameter information can improve the performance metric of the reference neural network. Finally, the parameter information of the target functional layer can be configured into an initialized neural network to generate a maintenance decision network of the target air compressor. For example, a back propagation algorithm may be used to configure the target functional layer parameter information into the initializing neural network to generate the target air compressor maintenance decision network.

In one possible implementation, step S140 may further include: and acquiring performance measurement indexes of the reference neural network, performing priority arrangement on the reference neural network according to the performance measurement indexes, and extracting a target air compressor maintenance decision network from the reference neural network based on a priority arrangement result.

In this embodiment, some performance metrics may be used to evaluate the performance of this reference neural network. For example, the performance of the reference neural network may be evaluated using indices of accuracy, recall, F1 score, etc. Then, a ranking algorithm may be used to rank each functional layer in the reference neural network such that the functional layer with better performance is ranked in front and the functional layer with worse performance is ranked in back. Therefore, the maintenance decision network of the target air compressor can be extracted from the reference neural network based on the priority sorting result of the reference neural network. For example, a greedy algorithm may be used to extract the target air compressor maintenance decision network from the reference neural network, so that the performance of the target air compressor maintenance decision network is optimal.

In a possible implementation manner, after step S140, the method further includes:

Step S141, at least one extended template operation trend data is obtained, wherein the extended template operation trend data comprises the current label fault maintenance category.

And step S142, optimizing the functional layer parameter information of the target air compressor maintenance decision network according to the current tag fault maintenance category to generate a current air compressor maintenance decision network.

And step S143, based on the operation trend data of the expansion template, parameter learning is carried out on the current air compressor maintenance decision network, and the air compressor maintenance decision network after parameter learning is used as the target air compressor maintenance decision network.

For example, some extended template operational trend data may be used to optimize a target air compressor maintenance decision network. That is, some historical data may be used, including current tag fault maintenance categories, such as fault type, fault cause, etc. When the current tag fault maintenance categories are obtained, the categories can be used for optimizing the function layer parameter information of the maintenance decision network of the target air compressor. For example, a gradient descent algorithm may be used to optimize functional layer parameter information of the target air compressor maintenance decision network, so that the target air compressor maintenance decision network may better process the current tag fault maintenance category. Therefore, the parameter learning can be performed on the current air compressor maintenance decision network by using the operation trend data of the expansion template. For example, a back propagation algorithm can be used to perform parameter learning on the current air compressor maintenance decision network, so that the current air compressor maintenance decision network can better process the operation trend data of the expansion template.

In a possible implementation manner, the verifying the decision effect of the target air compressor maintenance decision network for each fault maintenance category generates a decision effect index, including:

Step A110, the sample air compressor operation data sequence is separated into N sample air compressor operation data subsequences, N test sample data sequences are generated, the test sample data sequences comprise template operation trend data of the current sample air compressor corresponding to each fault maintenance type, and N is a preset positive integer greater than 2.

For example, the sample air compressor operation data sequence may be divided into 3 test sample data sequences, where each test sample data sequence includes template operation trend data of the current sample air compressor corresponding to three fault maintenance categories, namely normal, fault a and fault B.

And step A120, determining a check sample data sequence from the test sample data sequences, and taking the test sample data sequences except the check sample data sequence in the test sample data sequences as learning sample data sequences.

For example, one of the 3 test sample data sequences may be randomly selected as the check sample data sequence, and the remaining two are selected as the learning sample data sequences.

And step A130, performing parameter learning on the maintenance decision network of the target air compressor according to the learning sample data sequence, and deciding the fault maintenance category of the current sample air compressor corresponding to the checking sample data sequence according to the maintenance decision network of the target air compressor after parameter learning, so as to generate the current decision fault maintenance category of the current sample air compressor.

For example, the first two test sample data sequences can be used for carrying out parameter learning on the maintenance decision network of the target air compressor, and the fault maintenance category of the current sample air compressor corresponding to the test sample data sequences of the maintenance decision network decision check of the target air compressor after the parameter learning is used for generating the current decision fault maintenance category.

Step A140, obtaining a label fault maintenance category of the air compressor of the current sample, determining at least one basic decision effect index corresponding to the data sequence of the check sample according to the label fault maintenance category and the current decision fault maintenance category, obtaining an influence coefficient corresponding to each basic decision effect index, fusing the basic decision effect indexes according to the influence coefficients, calculating the average value of the fused basic decision effect indexes, and generating an initial decision effect index corresponding to the data sequence of the check sample.

For example, the accuracy, recall and F1 score may be used as the base decision effect indexes, and the influence coefficient corresponding to each index is obtained, and then the base decision effect indexes are fused by using the influence coefficients, and the average value of the fused base decision effect indexes is calculated, so as to generate the initial decision effect index corresponding to the data sequence of the check sample.

And step A150, returning to execute the step of determining the check sample data sequence in the test sample data sequences until each test sample data sequence is the check sample data sequence, and generating an initial decision effect index corresponding to each test sample data sequence.

And step A160, carrying out weighted calculation on the initial decision effect index to generate a decision effect index of the maintenance decision network of the target air compressor.

For example, the initial decision effect index corresponding to each test sample data sequence can be used for weighting calculation to generate the decision effect index of the maintenance decision network of the target air compressor.

In a possible implementation manner, in step a140, determining at least one basic decision effect index corresponding to the data sequence of the verification sample according to the tag fault maintenance class and the current decision fault maintenance class includes:

and step A141, comparing the label fault maintenance category with the current decision fault maintenance category to generate fault maintenance category decision detail information of each current sample air compressor.

And step A142, counting effective instance statistics and misjudgment instance statistics in the check sample data sequence according to the fault maintenance category decision detail information, wherein the effective instance statistics comprise instance statistics with reliable fault maintenance category decisions, and the misjudgment instance statistics comprise instance statistics with unreliable fault maintenance category decisions.

And step A143, obtaining instance statistics of the check sample data sequence, and determining at least one basic decision effect index corresponding to the check sample data sequence according to the instance statistics, the effective instance statistics and the misjudgment instance statistics.

For example, the fault maintenance category decision detail information of each current sample air compressor can be generated by using the current decision fault maintenance category and the label fault maintenance category, including information such as whether the current decision fault maintenance category is correct, the confidence of the current decision fault maintenance category and the like.

On the basis, effective instance statistics and misjudgment instance statistics can be counted in the check sample data sequence according to the fault maintenance category decision detail information. For example, the fault maintenance category decision detail information may be used to calculate valid instance statistics and misjudged instance statistics in the verification sample data sequence, including instance statistics for which the fault maintenance category decision is reliable and instance statistics for which the fault maintenance category decision is unreliable.

Then, at least one basic decision effect index corresponding to the data sequence of the check sample can be determined according to the instance statistic, the effective instance statistic and the misjudgment instance statistic. For example, the instance statistics, the effective instance statistics, and the misjudgment instance statistics may be used to determine the base decision effect indicators such as the accuracy, recall, and F1 score corresponding to the check sample data sequence.

The instance statistics include positive instance statistics and negative instance statistics, the effective instance statistics include effective positive instance statistics and effective negative instance statistics, the misjudgment instance statistics include misjudgment positive instance statistics and misjudgment negative instance statistics, and step a143 includes:

And step A1431, determining at least one basic verification performance index corresponding to the verification sample data sequence according to the positive instance statistic, the negative instance statistic, the effective positive instance statistic, the effective negative instance statistic, the misjudgment positive instance statistic and the misjudgment negative instance statistic.

And step A1432, extracting at least one target verification performance index from the basic verification performance indexes, and performing weighted calculation on the target verification performance indexes to generate a fusion verification performance index.

And step A1433, determining a current performance measurement index corresponding to the check sample data sequence according to the effective example statistic, the misjudgment negative example statistic and the example statistic, and taking the basic check performance index, the fusion check performance index and the current performance measurement index as the basic decision effect index.

After the positive instance statistic, the negative instance statistic, the effective positive instance statistic, the effective negative instance statistic, the misjudgment positive instance statistic and the misjudgment negative instance statistic have been obtained, at least one basic verification performance index corresponding to the verification sample data sequence can be determined according to the positive instance statistic, the negative instance statistic, the effective positive instance statistic, the effective negative instance statistic, the misjudgment positive instance statistic and the misjudgment negative instance statistic. For example, the positive instance statistic, the negative instance statistic, the effective positive instance statistic, the effective negative instance statistic, the misjudgment positive instance statistic and the misjudgment negative instance statistic may be used to determine the basic verification performance indexes such as the accuracy, the recall, the F1 score and the like corresponding to the verification sample data sequence.

After the basic verification performance index has been obtained, at least one target verification performance index may be extracted from the basic verification performance index, and the target verification performance index may be weighted to generate a fusion verification performance index. For example, the accuracy and recall may be extracted from the base verification performance index as the target verification performance index, and the accuracy and recall may be weighted to generate the fusion verification performance index.

After the effective example statistic, the misjudgment negative example statistic and the example statistic are obtained, the current performance measurement index corresponding to the check sample data sequence can be determined according to the effective example statistic, the misjudgment negative example statistic and the example statistic, and the basic check performance index, the fusion check performance index and the current performance measurement index are taken as basic decision effect indexes. For example, the effective instance statistics, the misjudgment negative instance statistics and the instance statistics may be used to determine the current performance metrics such as the accuracy, recall, and F1 score corresponding to the check sample data sequence, and take the basic check performance metrics, the fusion check performance metrics, and the current performance metrics as the basic decision effect metrics.

Illustratively, accuracy = (effective positive instance statistic + effective negative instance statistic)/instance statistic

Recall = valid positive instance statistic/(valid positive instance statistic + false positive instance statistic)

F1 fraction = 2 x (accuracy x recall)/(accuracy + recall)

Fusion verification performance index=α x accuracy + β x recall + γ x F1 score

Wherein, alpha, beta and gamma are weight coefficients, which can be adjusted according to practical conditions.

The current performance metric index= (effective positive example statistic + effective negative example statistic)/example statistic, and the basic verification performance index, the fusion verification performance index and the current performance metric index are used as basic decision effect indexes, so that the performance of the model can be evaluated.

In one possible implementation, step S150 may include:

And step S151, when the decision effect index characterizes that the maintenance decision network of the target air compressor reaches a preset decision effect, acquiring air compressor operation data of the target air compressor when at least one operation process is executed.

And step S152, performing multi-element attribute embedded representation on the operation data of the air compressor, and generating target multi-element attribute embedded representation information of the target air compressor.

And step 153, deciding the fault maintenance category of the target air compressor pair in the target multi-element attribute embedded representation information according to the target air compressor maintenance decision network.

For example, if the decision effect index characterizes that the maintenance decision network of the target air compressor reaches a preset decision effect, the operation data of the air compressor when the target air compressor executes at least one operation process can be obtained. For example, air compressor operation data of a target air compressor when executing an operation course of compressed air, cooling air, discharge air, and the like may be acquired.

Therefore, the multi-attribute embedded representation can be carried out on the air compressor operation data of the target air compressor, and target multi-attribute embedded representation information of the target air compressor is generated. For example, the air compressor operation data may be represented by embedding multiple attributes by using a word embedding technique, so as to generate target multiple attribute embedded representation information of the target air compressor, which may be specifically described in the foregoing embodiments.

Finally, the fault maintenance category of the target air compressor can be decided in the target multi-element attribute embedded representation information according to the maintenance decision network of the target air compressor. For example, a target air compressor maintenance decision network may be used to decide a fault maintenance category of the target air compressor in the target multi-element attribute embedded representation information, including normal, fault a, fault B, fault C, and the like.

Fig. 2 schematically illustrates an air compressor maintenance system 100 that may be used to implement various embodiments described in the present disclosure.

For one embodiment, fig. 2 illustrates an air compressor maintenance system 100, the air compressor maintenance system 100 having a plurality of processors 102, a control module (chipset) 104 coupled to one or more of the processor(s) 102, a memory 106 coupled to the control module 104, a non-volatile memory (NVM)/storage 108 coupled to the control module 104, a plurality of input/output devices 110 coupled to the control module 104, and a network interface 112 coupled to the control module 104.

Processor 102 may include a plurality of single-core or multi-core processors, and processor 102 may include any combination of general-purpose or special-purpose processors (e.g., graphics processors, application processors, baseband processors, etc.). In some alternative embodiments, the air compressor maintenance system 100 can be used as a server device such as a gateway in the embodiments of the present application.

In some alternative embodiments, the air compressor maintenance system 100 can include a plurality of computer-readable media (e.g., memory 106 or NVM/storage 108) having instructions 114 and a plurality of processors 102 combined with the plurality of computer-readable media configured to execute the instructions 114 to implement the modules to perform the actions described in this disclosure.

For one embodiment, the control module 104 may include any suitable interface controller to provide any suitable interface to one or more of the processor(s) 102 and/or any suitable device or component in communication with the control module 104.

The control module 104 may include a memory controller module to provide an interface to the memory 106. The memory controller modules may be hardware modules, software modules, and/or firmware modules.

Memory 106 may be used, for example, to load and store data and/or instructions 114 for air compressor maintenance system 100. For one embodiment, memory 106 may comprise any suitable volatile memory, such as, for example, a suitable DRAM. In some alternative embodiments, memory 106 may comprise a double data rate type four synchronous dynamic random access memory.

For one embodiment, the control module 104 may include a plurality of input/output controllers to provide interfaces to the NVM/storage 108 and the input/output device(s) 110.

For example, NVM/storage 108 may be used to store data and/or instructions 114. NVM/storage 108 may include any suitable non-volatile memory (e.g., flash memory) and/or may include any suitable non-volatile storage(s).

NVM/storage 108 may include storage resources that are physically part of the device on which air compressor maintenance system 100 is installed, or which may be accessible by the device, but may not be necessary as part of the device. For example, NVM/storage 108 may be accessed via input/output device(s) 110 in connection with a network.

The input/output device(s) 110 may provide an interface for the air compressor maintenance system 100 to communicate with any other suitable device, and the input/output device 110 may include a communication component, a pinyin component, a sensor component, and the like. The network interface 112 may provide an interface for the air compressor maintenance system 100 to communicate in accordance with a plurality of networks, and the air compressor maintenance system 100 may communicate wirelessly with a plurality of components of a wireless network based on any of a plurality of wireless network standards and/or protocols, such as accessing a wireless network in accordance with a communication standard, such as WiFi, 2G, 3G, 4G, 5G, etc., or a combination thereof.

For one embodiment, one or more of the processor(s) 102 may be packaged together with logic of a plurality of controllers (e.g., memory controller modules) of the control module 104. For one embodiment, one or more of the processor(s) 102 may be packaged together with logic of multiple controllers of the control module 104 to form a system in package. For one embodiment, one or more of the processor(s) 102 may be integrated on the same die with logic of multiple controllers of the control module 104. For one embodiment, one or more of the processor(s) 102 may be integrated on the same die with logic of multiple controllers of the control module 104 to form a system-on-chip.

In various embodiments, the air compressor maintenance system 100 may be, but is not limited to: a desktop computing device or a mobile computing device (e.g., a laptop computing device, a handheld computing device, a tablet, a netbook, etc.), and the like. In various embodiments, the air compressor maintenance system 100 may have more or fewer components and/or different architectures. For example, in some alternative embodiments, the air compressor maintenance system 100 includes a plurality of cameras, a keyboard, a liquid crystal display screen (including a touch screen display), a non-volatile memory port, a plurality of antennas, a graphics chip, an application specific integrated circuit, and a speaker.

The foregoing has outlined rather broadly the more detailed description of the application in order that the detailed description of the principles and embodiments of the application may be implemented in conjunction with the detailed description of the application that follows, the examples being merely intended to facilitate an understanding of the method of the application and its core concepts; meanwhile, as those skilled in the art will have variations in the specific embodiments and application scope in accordance with the ideas of the present application, the present description should not be construed as limiting the present application in view of the above.

Claims

1. The maintenance decision method of the air compressor based on AI enhancement is characterized by being applied to an air compressor maintenance system, and comprises the following steps:

2. The AI-enhanced-based air compressor maintenance decision method of claim 1, wherein the obtaining the sample air compressor operation data sequence includes:

3. The AI-enhanced air compressor maintenance decision method of claim 2, wherein the performing parameter learning on the initialized neural network based on the target decision error parameter to generate a reference neural network corresponding to each sample air compressor operation data subsequence includes:

4. The AI-enhanced air compressor maintenance decision method of claim 3, wherein configuring the target air compressor maintenance decision network based on the functional layer parameter information of the reference neural network includes:

acquiring performance measurement indexes of the reference neural network;

5. The AI-enhanced-based air compressor maintenance decision method of claim 1, wherein configuring the target air compressor maintenance decision network based on the functional layer parameter information of the reference neural network comprises:

acquiring performance measurement indexes of the reference neural network;

6. The AI-enhanced-based air compressor maintenance decision method of claim 1, wherein after configuring the target air compressor maintenance decision network based on the functional layer parameter information of the reference neural network, further comprises:

7. The AI-enhanced air compressor maintenance decision method of any of claims 1-5, wherein the verifying the decision effect of the target air compressor maintenance decision network for each failure maintenance category, generating a decision effect indicator, includes:

8. The AI-enhanced air compressor maintenance decision method of claim 7, wherein determining at least one base decision effect indicator corresponding to the check sample data sequence according to the tag fault maintenance category and a current decision fault maintenance category includes:

9. The AI-enhanced air compressor maintenance decision method of any of claims 1-5, wherein the deciding, based on the decision effect index, a failure maintenance category of the target air compressor in air compressor operation data of the target air compressor based on the target air compressor maintenance decision network includes:

And according to the maintenance decision network of the target air compressor, deciding the fault maintenance category corresponding to the target air compressor in the target multi-element attribute embedded representation information.

10. An air compressor maintenance system comprising a processor and a machine-readable storage medium having stored therein machine-executable instructions loaded and executed by the processor to implement the AI-enhanced air compressor maintenance decision method of any of claims 1-9.