CN118707303A - Electronic direct current load fault diagnosis method based on multi-expert distillation network - Google Patents

Abstract

The application discloses an electronic direct current load fault diagnosis method based on a multi-expert distillation network, which comprises the steps of decomposing current data into K IMF signals by adopting an expert variation modal decomposition algorithm, and finishing to form an electronic direct current load diagnosis data set; extracting global features of the diagnostic dataset samples using a global expert attention network; extracting regional features of the diagnostic dataset samples using a regional expert attention network; the distillation function is adopted to monitor the mutual learning of the global expert attention network and the regional expert attention network, and the global enhancement feature and the regional enhancement feature of the diagnosis data set sample are obtained through the two learned networks; and inputting the global enhancement features and the regional enhancement features into a dual-feature classification network to obtain a final electronic direct current load fault diagnosis result. The electronic direct current load fault diagnosis method provided by the application has the advantages of high accuracy and strong robustness, and an effective tool is provided for an electronic direct current load intelligent fault diagnosis system.

Description

Electronic direct current load fault diagnosis method based on multi-expert distillation network

Technical Field

The invention belongs to the field of electronic direct current load artificial intelligent fault diagnosis, and particularly relates to an electronic direct current load fault diagnosis method based on a multi-expert distillation network.

Background

In modern industry and electronic systems, electronic dc loads are widely used in a number of fields such as power supply testing, device debugging, and system verification. They play a critical role in ensuring the reliability and performance of devices and systems. However, during long-term operation, the electronic dc load may fail for various reasons, such as open-circuit failure, short-circuit failure, overload failure, and the like. These faults can not only affect the accuracy of testing and verification, but can also lead to equipment damage or system downtime, thereby causing serious economic loss and safety hazards. Therefore, it is particularly important to perform effective fault diagnosis on the electronic dc load.

In recent years, with rapid development of electronic technology and computing power, fault diagnosis technology has also made remarkable progress. The traditional fault diagnosis method mainly depends on the observation and analysis of fault phenomena by experienced engineers, but the method has the defects of low efficiency, dependence on manpower, low diagnosis speed and the like. In order to improve the efficiency and accuracy of fault diagnosis, the modern fault diagnosis technology gradually introduces a plurality of advanced methods and means, the fault diagnosis accuracy of the electronic direct current load can be improved by adopting a deep learning network, and meanwhile, the efficiency and stability of diagnosis can be greatly improved by adopting a multi-expert distillation network to perform fault diagnosis in the deep learning network.

Disclosure of Invention

The invention aims to provide a fault diagnosis method of an electronic direct current load, which adopts a multi-expert distillation deep learning network, can rapidly, accurately and stably diagnose the fault type of the electronic direct current load, and provides an effective tool for an electronic direct current load fault diagnosis system.

In order to solve the technical problems, the invention provides an electronic direct current load fault diagnosis method, which comprises the following steps:

collecting current data of an electronic direct current load circuit, decomposing the current data into K IMF signals by adopting an expert variation modal decomposition algorithm, and finishing to form an electronic direct current load diagnosis data set;

extracting global features of the electronic direct current load diagnosis data set by adopting a global expert attention network;

Extracting regional characteristics of the electronic direct current load diagnosis data set by using a regional expert attention network;

The distillation function is adopted to monitor mutual learning of the global expert attention network and the regional expert attention network, and global enhancement features and regional enhancement features of the electronic direct current load diagnosis data set are obtained through the learned network;

and inputting the global enhancement features and the regional enhancement features of the electronic direct current load diagnosis data set into a dual-feature classification network to obtain a final electronic direct current load fault diagnosis result.

Optionally, collecting current data of the electronic direct current load circuit, decomposing the current data into K IMF signals by using an expert variation modal decomposition algorithm, and sorting the K IMF signals to form an electronic direct current load diagnosis data set, including:

Collecting current data of an electronic direct current load circuit in operation, adopting an expert variation modal decomposition algorithm to the current data, establishing an expert variation model, converting a solving process of the model into an unconstrained optimization problem by using a punishment factor and a multiplication operator, solving an optimal solution of the model problem by adopting an Alternate Direction Multiplication Method (ADMM), finally obtaining K IMF signals of the current data decomposition by iterating the solving process, and finishing according to three fault types of open-circuit faults, short-circuit faults and overload faults to form an electronic direct current load diagnosis data set.

Optionally, the global feature process of extracting the electronic dc load diagnostic dataset using the global expert attention network includes:

The method comprises the steps that an electronic direct current load diagnosis data set sample is input into a global expert attention network, the global expert attention network comprises a convolution feature layer, a double-channel attention layer and a fusion layer, window features of K IMF signals of the diagnosis sample are simultaneously extracted through a sliding convolution window by the convolution feature layer, weighting and summing are carried out on the window features of the K IMF signals of the diagnosis sample through double-channel keys, values and queries to obtain two channel features, and nonlinear mapping fusion is carried out on the two channel features through full connection by the fusion layer to obtain the global features of the electronic direct current load diagnosis data set sample.

Optionally, extracting the regional features of the electronic dc load diagnostic dataset using a regional expert attention network includes:

The method comprises the steps that an electronic direct current load diagnosis data set sample is input into a regional expert attention network, the regional expert attention network comprises a regional attention layer and a conversion matrix, the regional attention layer comprises K attention blocks with the same structure, each attention block extracts component characteristics corresponding to 1 IMF signal, the regional attention layer extracts K component characteristics corresponding to the K IMF signals in the sample, and the K component characteristics are subjected to simultaneous point multiplication with the conversion matrix to obtain the regional characteristics of the electronic direct current load diagnosis data set sample.

Optionally, the distillation function is adopted to supervise mutual learning of the global expert attention network and the regional expert attention network, and global enhancement features and regional enhancement features of the electronic direct current load diagnosis data set are obtained through the learned network, including:

And calculating the feature distance between the global feature and the regional feature by adopting a distillation function, continuously updating respective network parameters by the global expert attention network and the regional expert attention network according to the counter-propagating feature distance, realizing mutual learning of the global expert attention network and the regional expert attention network, inputting the electronic direct current load diagnosis data set into the learned global expert attention network and regional expert attention network again, and obtaining the global enhancement feature and the regional enhancement feature of the electronic direct current load diagnosis data set sample.

Optionally, inputting the global enhancement feature and the regional enhancement feature of the electronic direct current load diagnosis data set into a dual feature classification network to obtain a final electronic direct current load fault diagnosis result, including:

The method comprises the steps of inputting global enhancement features and regional enhancement features of an electronic direct current load diagnosis data set sample into a dual-feature classification network, wherein the dual-feature classification network comprises convolution diagnostic layers and full-connection diagnostic layers which run in parallel, the convolution diagnostic layers classify the input global features and output convolution classification probabilities, the full-connection diagnostic layers classify the input local features and output full-connection classification probabilities, and finally the dual-feature classification network performs weighted summation on the convolution classification probabilities and the full-connection classification probabilities to obtain an electronic direct current load sample fault diagnosis result.

The application discloses an electronic direct current load fault diagnosis method based on a multi-expert distillation network, which comprises the steps of decomposing current data into K IMF signals by adopting an expert variation modal decomposition algorithm, and finishing to form an electronic direct current load diagnosis data set; inputting the diagnostic data set sample into a global expert attention network to extract global features of the sample; extracting regional features of the diagnostic dataset samples using a regional expert attention network; the distillation function is adopted to monitor the mutual learning of the global expert attention network and the regional expert attention network, and the diagnostic sample is input into the two learned networks again to obtain the global enhancement feature and the regional enhancement feature of the diagnostic data set sample; and inputting the global enhancement features and the regional enhancement features into a dual-feature classification network to obtain a final electronic direct current load fault diagnosis result. The electronic direct current load fault diagnosis method provided by the application has the advantages of high accuracy and strong robustness, and an effective tool is provided for an electronic direct current load intelligent fault diagnosis system.

Drawings

For a clearer description of embodiments of the invention or of the prior art, the drawings that are used in the description of the embodiments or of the prior art will be briefly described, it being apparent that the drawings in the description below are only some embodiments of the invention, and that other drawings can be obtained from them without inventive effort for a person skilled in the art.

FIG. 1 is a schematic flow chart of an electronic DC load fault diagnosis method provided by the invention;

FIG. 2 is a diagram of a global expert attention network framework of the electronic DC load fault diagnosis method provided by the invention;

FIG. 3 is a diagram of a regional expert attention network framework of the electronic DC load fault diagnosis method provided by the invention;

fig. 4 is a network flow chart of the electronic direct current load fault diagnosis method provided by the invention.

Detailed Description

In order to better understand the aspects of the present invention, the present invention will be described in further detail with reference to the accompanying drawings and detailed description. It will be apparent that the described embodiments are only some, but not all, embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

As shown in fig. 1, fig. 1 is a flow chart of a method for extracting motor breath non-uniformity fault characteristics, which includes:

and S11, collecting current data of the electronic direct current load circuit, decomposing the current data into K IMF signals by adopting an expert variation modal decomposition algorithm, and finishing to form an electronic direct current load diagnosis data set.

It should be noted that, the present application introduces the adoption of the expert variation modal decomposition algorithm to decompose the current data into K intrinsic modal components (IMF) signals, because the decomposed K IMF signals contain more stable, more easily identifiable and more comprehensive electronic direct current load fault characteristics, and the robustness and accuracy of the subsequent characteristic extraction network are ensured.

The current sensor is arranged in an electronic direct current load circuit which operates to collect a circuit current signal, and the sampling frequency is N kHz.

The method comprises the steps of decomposing current data into K IMF signals by adopting an expert variation modal decomposition algorithm, and finishing to form an electronic direct current load diagnosis data set, wherein the method comprises 5 steps:

Step 1: in order to decompose the inherent modal component representing the fault characteristics of the electronic direct current load from the original current data, the application adopts an expert variation model shown in formulas (1 and 2):

In the method, in the process of the invention, In the case of a partial derivative operation,For the dirichlet distribution function ∗ denotes the convolution,Represents K IMFs obtained by expert variation modal decomposition algorithm,The center frequency of each IMF is represented, f represents the original current signal, and K is the number of IMFs.

Step 2: converting the solving process of the model of the formula (1) into an unconstrained optimization problem by using a punishment factor and a multiplication operator, and obtaining an augmented Lagrangian function as shown in the formula (3) as follows:

In the method, in the process of the invention, Is the lagrangian constant and,Is a penalty factor.

Step 3: in order to solve the optimal solution of the expert variation model problem, the application adopts an Alternate Direction Multiplier Method (ADMM) to update variables, and establishes a solving process shown as a formula (4):

In the middle of For the purpose of noise tolerance,AndRespectively correspond toIs a fourier transform of (a).

Step 4: initialization ofIteration is performed by means of n=n+1, where n is the number of iterations, updating the parametersRepeating the steps until the iteration stop condition is met, and obtaining the optimalIn the followingIs constant and is；

Step 5: finally obtaining K IMF signals of current data decompositionAnd sorting and forming an electronic direct current load diagnosis data set according to the three fault types of the open circuit fault, the short circuit fault and the overload fault.

Based on the discussion above, in an alternative embodiment of the present application, decomposing the current data into K IMF settings using the expert variation modal decomposition algorithm described above may include:

In an alternative embodiment of S11, the current sampling frequency is 15kHz, the current data is decomposed into 9 IMF signals, the number of iterations n is 120, and the iteration stop condition epsilon=0.5.

And S12, extracting global characteristics of the electronic direct current load diagnosis data set by adopting a global expert attention network.

It should be noted that the present application introduces a global expert attention network to extract global features of the sample because the global expert attention network has a global field of view and the attention mechanisms in the network can allocate more computing resources to important feature information.

The global expert attention network structure is shown in fig. 2, the network comprises a convolution characteristic layer, a double-channel attention layer and a fusion layer, and the network extracts global characteristics of an electronic direct current load diagnosis data set sample, and specifically comprises 3 steps:

Step 1: the method comprises the steps that an electronic direct current load diagnosis data set sample is firstly input into a convolution feature layer of a global expert attention network, and the convolution feature layer carries out sliding convolution on K IMF signals of the sample simultaneously to obtain window features.

Step 2: the window features are input into a dual-channel attention layer of the global expert attention network, parameters of dual channels in the dual-channel attention layer are mutually independent, and window features are weighted and summed by keys, values and queries of the dual channels to obtain two channel features.

Step 3: and inputting the two channel characteristics into a fusion layer of the network, and carrying out nonlinear mapping fusion on the two channel characteristics by the fusion layer through full connection to obtain the global characteristics of the electronic direct current load diagnosis data set sample.

Based on the discussion above, in an alternative embodiment of the present application, the global feature setting process for extracting the electronic dc load diagnostic dataset sample described above may include:

in an alternative embodiment of S12, adam optimization algorithm is used during training, learning rate is 0.001, input epoch is 64, and iteration number is 1000.

And S13, extracting the regional characteristics of the electronic direct current load diagnosis data set by using a regional expert attention network.

It should be noted that, the present application introduces the regional expert attention network to extract the regional features of the sample, because the regional expert attention network has a more precise and finer regional field of view than the global expert attention network, and can extract more fine regional features in the sample.

The structure of the regional expert attention network is shown in fig. 3, the regional expert attention network comprises a regional attention layer and a conversion matrix, and the regional expert attention network extracts regional characteristics of the electronic direct current load diagnosis data set sample and specifically comprises 2 steps.

Step 1: the electronic direct current load diagnosis data set sample firstly enters a regional attention layer of a regional expert attention network, the regional attention layer comprises K attention blocks with the same structure, each attention block extracts component features corresponding to 1 IMF signal, and the regional attention layer extracts K component features corresponding to K IMF signals in the sample.

Step 2: the K component features of the samples of the electronic DC load diagnostic dataset are linearly transformed together into the region features of the samples.

Based on the foregoing discussion, in an alternative embodiment of the present application, the region feature setting process for extracting the electronic dc load diagnostic dataset sample described above may include:

In an alternative embodiment of S13, adam optimization algorithm is adopted during training, learning rate is 0.003, input epoch is 64, and iteration number is 500.

And S14, supervising the mutual learning of the global expert attention network and the regional expert attention network by adopting a distillation function, and obtaining global enhancement features and regional enhancement features of the electronic direct current load diagnosis data set through the learned network.

It should be noted that, the distance between the global feature and the regional feature is calculated by introducing the distillation function, because the global expert attention network and the regional expert attention network make up the respective defects by learning each other through distillation loss, and the global expert attention network and the regional expert attention network with high-accuracy feature extraction capability are obtained.

The distance between the global feature and the regional feature of the sample is calculated by adopting a distillation function, the mutual learning of a global expert attention network and a regional expert attention network is supervised, and finally the global enhancement feature and the regional enhancement feature of the diagnosis sample are obtained, which comprises the following steps:

Step 1: in order to calculate the distance between the global features and the regional features of the diagnostic sample, the present application builds a distillation function shown in (6):

In the middle of 、Global features and regional features of the diagnostic sample respectively,The JS divergence is indicated,Representing the mean square error.

Step 2: in order to translate the distance of the diagnostic sample global features from the regional features into optimized parameters for the global expert attention network and the regional expert attention network, calculations are requiredThe gradient of (2) is calculated as shown in formula (7):

In the middle of Representing the derivative, W representing the network parameter.

Step 3: the global expert attention network and the regional expert attention network continuously update the network parameters according to the gradient calculated in the formula (7) until iteration is stopped, wherein the network parameter update formula is shown in the formula (8):

In the middle of Representing the learning rate.

Step 4: and inputting the electronic direct current load diagnosis data set into a global expert attention network and a regional expert attention network after updating network parameters to obtain global enhancement features and regional enhancement features of the diagnosis sample.

Based on the foregoing discussion, in an alternative embodiment of the present application, for the above calculation of the distance between the global feature and the regional feature of the sample using the distillation function, the process of setting the global enhancement feature and the regional enhancement feature of the finally obtained diagnostic sample by supervising the mutual learning of the global expert attention network and the regional expert attention network may include:

in an alternative embodiment of S14, adam optimization algorithm is used during training, learning rate is 0.001, input epoch is 64, and iteration number is 1000.

And S15, inputting the global enhancement features and the regional enhancement features of the electronic direct current load diagnosis data set into a dual-feature classification network to obtain a final electronic direct current load fault diagnosis result.

It should be noted that, the dual-feature classification network introduced by the application can perform high-accuracy and high-robustness fault diagnosis on the electronic direct current load circuit, because the dual-feature classification network has a convolution diagnosis layer and a full-connection diagnosis layer, the network can fully utilize global enhancement feature and local enhancement feature information in the diagnosis sample.

The method comprises the following steps of performing fault diagnosis on the electronic direct current load circuit based on global enhancement features and regional enhancement features of a diagnosis sample by adopting a dual-feature classification network, and specifically comprising the following 4 steps:

Step 1: the method comprises the steps of establishing a dual-feature classification network, wherein the dual-feature network comprises a convolution diagnostic layer and a full-connection diagnostic layer which run in parallel, inputting global enhanced features and regional enhanced features of diagnostic samples into the dual-feature classification network, performing multi-layer convolution on the global features by the convolution diagnostic layer in the dual-feature classification network to extract convolution features, and inputting the convolution features into a softmax classifier to obtain convolution softmax classification probability.

Step 2: and carrying out nonlinear mapping on the region enhanced features of the diagnosis sample by using the full-connection diagnosis layer in the dual-feature classification network to extract full-connection features, and inputting the full-connection features into a softmax classifier to obtain full-connection softmax classification probability.

Step 3: and the double-feature classification network performs weighted summation on the convolution classification probability and the full-connection classification probability obtained in the steps to obtain an electronic direct current load sample fault diagnosis result.

Based on the discussion above, in an alternative embodiment of the present application, for the above-described fault diagnosis of an electronic dc load circuit based on the global enhancement feature and the regional enhancement feature of the diagnostic sample using the dual feature classification network, the setup process may include:

in an alternative embodiment of S15, adam optimization algorithm is adopted during training, learning rate is 0.001, input epoch is 32, iteration number is 100, and the ratio weights of convolution classification probability and full connection classification probability are respectively 0.3 and 0.7.

FIG. 4 is a network flow chart of the method for diagnosing the fault of the electronic direct current load, which is provided by the application, wherein the method firstly adopts an expert variation modal decomposition algorithm to decompose current data into K intrinsic modal components (IMFs) and collates the K intrinsic modal components into an electronic direct current load diagnosis data set; then, inputting the electronic direct current load diagnosis data set sample into a global expert attention network to extract global characteristics of the sample; extracting regional characteristics of the sample by using a regional expert attention network; calculating the distance between the global feature and the regional feature of the diagnostic sample through a distillation function, wherein the distance is used for supervising mutual learning of a global expert attention network and a regional expert attention network, and inputting the diagnostic sample into the learned global expert attention network and regional expert attention network again to obtain the global enhancement feature and the regional enhancement feature of the diagnostic sample; and finally, inputting the global enhanced features and the regional enhanced features of the diagnosis sample into a dual-feature classification network to obtain a final electronic direct current load fault diagnosis result. The fault diagnosis method of the electronic direct current load provided by the application can rapidly, accurately and stably diagnose the fault type of the electronic direct current load, and provides an effective tool for an electronic direct current load fault diagnosis system.

It is noted that relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. The principles and embodiments of the present invention have been described herein with reference to specific examples, the description of which is intended only to facilitate an understanding of the method of the present invention and its core ideas. It should be noted that it will be apparent to those skilled in the art that various modifications and adaptations of the invention can be made without departing from the principles of the invention and these modifications and adaptations are intended to be within the scope of the invention as defined in the following claims.

Claims (6)

1. The electronic direct current load fault diagnosis method based on the multi-expert distillation network is characterized by comprising the following steps of:

collecting current data of an electronic direct current load circuit, decomposing the current data into K IMF signals by adopting an expert variation modal decomposition algorithm, and finishing to form an electronic direct current load diagnosis data set;

extracting global features of the electronic direct current load diagnosis data set by adopting a global expert attention network;

Extracting regional characteristics of the electronic direct current load diagnosis data set by using a regional expert attention network;

The distillation function is adopted to monitor mutual learning of the global expert attention network and the regional expert attention network, and global enhancement features and regional enhancement features of the electronic direct current load diagnosis data set are obtained through the learned network;

and inputting the global enhancement features and the regional enhancement features of the electronic direct current load diagnosis data set into a dual-feature classification network to obtain a final electronic direct current load fault diagnosis result.

2. The method for diagnosing an electronic dc load fault as recited in claim 1, wherein collecting current data of the electronic dc load circuit, decomposing the current data into K IMF signals using an expert variation modal decomposition algorithm, and sorting the K IMF signals to form an electronic dc load diagnostic data set, comprising:

Collecting current data of an electronic direct current load circuit in operation, adopting an expert variation modal decomposition algorithm to the current data, establishing an expert variation model, converting a solving process of the model into an unconstrained optimizing problem by using a punishment factor and a multiplication operator, solving an optimal solution of the model problem by adopting an alternate direction multiplier method, finally obtaining K IMF signals of the current data decomposition by iterating the solving process, and finishing according to three fault types of open-circuit faults, short-circuit faults and overload faults to form an electronic direct current load diagnosis data set.

3. The method of electronic dc load fault diagnosis as claimed in claim 1, wherein extracting global features of the electronic dc load diagnostic dataset using a global expert attention network comprises:

The method comprises the steps that an electronic direct current load diagnosis data set sample is input into a global expert attention network, the global expert attention network comprises a convolution feature layer, a double-channel attention layer and a fusion layer, window features of K IMF signals of the diagnosis sample are simultaneously extracted through a sliding convolution window by the convolution feature layer, weighting and summing are carried out on the window features of the K IMF signals of the diagnosis sample through double-channel keys, values and queries to obtain two channel features, and nonlinear mapping fusion is carried out on the two channel features through full connection by the fusion layer to obtain the global features of the electronic direct current load diagnosis data set sample.

4. The method of electronic dc load fault diagnosis as claimed in claim 1, wherein extracting the regional characteristics of the electronic dc load diagnostic dataset using a regional expert attention network comprises:

The method comprises the steps that an electronic direct current load diagnosis data set sample is input into a regional expert attention network, the regional expert attention network comprises a regional attention layer and a conversion matrix, the regional attention layer comprises K attention blocks with the same structure, each attention block extracts component characteristics corresponding to 1 IMF signal, the regional attention layer extracts K component characteristics corresponding to the K IMF signals in the sample, and the K component characteristics are subjected to simultaneous point multiplication with the conversion matrix to obtain the regional characteristics of the electronic direct current load diagnosis data set sample.

5. The method for diagnosing an electronic dc load fault as recited in claim 1, wherein supervising the mutual learning of the global expert's attention network and the regional expert's attention network by using the distillation function and obtaining global enhancement features and regional enhancement features of the electronic dc load diagnosis data set through the learned network comprises:

And calculating the feature distance between the global feature and the regional feature by adopting a distillation function, continuously updating respective network parameters by the global expert attention network and the regional expert attention network according to the counter-propagating feature distance, realizing mutual learning of the global expert attention network and the regional expert attention network, inputting the electronic direct current load diagnosis data set into the learned global expert attention network and regional expert attention network again, and obtaining the global enhancement feature and the regional enhancement feature of the electronic direct current load diagnosis data set sample.

6. The method for diagnosing an electronic dc load fault as recited in claim 1, wherein inputting the global enhancement feature and the regional enhancement feature of the electronic dc load diagnostic dataset into the dual feature classification network to obtain a final electronic dc load fault diagnosis result comprises:

The method comprises the steps of inputting global enhancement features and regional enhancement features of an electronic direct current load diagnosis data set sample into a dual-feature classification network, wherein the dual-feature classification network comprises convolution diagnostic layers and full-connection diagnostic layers which run in parallel, the convolution diagnostic layers classify the input global features and output convolution classification probabilities, the full-connection diagnostic layers classify the input local features and output full-connection classification probabilities, and finally the dual-feature classification network performs weighted summation on the convolution classification probabilities and the full-connection classification probabilities to obtain an electronic direct current load sample fault diagnosis result.