Method for generating training data for fault diagnosis of analog circuit

Abstract

The invention discloses a method for generating training data for fault diagnosis of an analog circuit, and the method comprises the steps: firstly building a Condition-DDPM model employing a self-adaptive strategy, then collecting the real output data of a target analog circuit in a normal state and different fault states, carrying out the modal decomposition processing, and then training the built Condition-DDPM model, and finally, a training data set is generated through the trained Conditional-DDPM model.

Description

Technical Field

[0001] The present invention belongs to the technical field of circuit fault diagnosis. More specifically, it relates to a method for generating training data for analog circuit fault diagnosis. Background Art

[0002] In recent years, data-driven analog circuit fault diagnosis methods have been widely applied to various links such as the production, manufacturing, and maintenance of electronic equipment and systems. The accuracy of fault diagnosis greatly affects national defense security and all aspects of industrial production. Data-driven fault diagnosis is a widely used fault diagnosis method at present, and its accuracy highly depends on the quality of the neural network that plays a diagnostic role, and the quality of the neural network is closely related to the quality of the training data used for network training. However, for actual fault diagnosis problems, it is often very difficult and costly to obtain a large amount of high-quality training data. Therefore, generating a large amount of high-quality training data is of great significance for improving the accuracy of analog circuit fault diagnosis and reducing the network training cost.

[0003] Existing methods for generating analog circuit fault diagnosis training data include the following categories: 1. Generation methods based on Generative Adversarial Network (GAN). This type of method learns from the real data set by constructing a generator and a discriminator, so as to achieve the purpose of generating data with an approximate distribution to the real data. However, during the training process, it is easy to fall into mode collapse, resulting in poor quality of the generated results. 2. Generation methods based on Variational Auto-Encoder (VAE). This type of method constructs a set of trainable encoding-decoding mapping pairs through a neural network and uses the real data to train this set of mappings, so as to realize the mining and learning of the features of the real data. However, it cannot mine the deep high-dimensional features existing inside the data, resulting in poor quality of the finally generated results. 3. Generation methods based on Denoising Diffusion Probabilistic Model (DDPM). Its generation process is relatively stable, can avoid the problem of mode collapse, and can mine the high-dimensional hidden information of the original data through the noise addition / denoising process. However, the generation process of this method is uncontrollable, and the generated results have randomness and uncertainty. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for generating training data for analog circuit fault diagnosis. The method is based on a conditional denoising diffusion probability model (Conditional-DDPM) with strategy adaptation. By setting a conditional function including a scale coefficient, the model is guided to generate data with similar characteristics to real data, and the scale coefficient is adaptively adjusted according to the generation result, thereby improving the quality of the generated training data.

[0005] To achieve the above-mentioned object of the invention, the present invention provides a method for generating training data for simulating circuit fault diagnosis, characterized in that it comprises the following steps:

[0006] (1) Obtain the real output data of the target analog circuit under normal conditions and different fault conditions;

[0007] (2) Modal decomposition processing;

[0008] (3) Establish a Conditional-DDPM model using an adaptive strategy;

[0009] (4) Training the Conditional-DDPM model;

[0010] (5) Use the trained model to generate data;

[0011] (6) Output the training data set;

[0012] The object of the invention of the present invention is achieved in this way:

[0013] The present invention discloses a method for generating training data for analog circuit fault diagnosis. The method first establishes a Conditional-DDPM model using an adaptive strategy, then collects real output data of a target analog circuit in a normal state and in different fault states, and uses the data for training the established Conditional-DDPM model after modal decomposition processing. Finally, a training data set is generated through the trained Conditional-DDPM model.

[0014] At the same time, the method for generating training data for analog circuit fault diagnosis of the present invention also has the following beneficial effects:

[0015] (1) Using the diffusion denoising probability model architecture, we solve the problem that the data generation method based on the generative adversarial network (GAN) suffers from mode collapse, resulting in poor quality of generated results. We also solve the problem that the data generation method based on the variational autoencoder (VAE) cannot mine the deep high-dimensional features existing in the data.

[0016] (2) The conditional generation function is used to guide the generation process of the diffusion denoising probability model, solving the problems of uncontrollable direction during the generation of the unconditional diffusion probability model and the generated results not having the characteristics of the original data.

[0017] (3) The joint loss function that takes into account both the MMD of the generated data and the variance is used to quantify the quality of the generated data, and the conditional scale coefficient matrix is adaptively modified using the Adam strategy according to the value of the joint loss function, improving the quality of the finally generated data. Brief Description of the Drawings

[0018] Figure 1 is a flowchart of a method for generating training data for analog circuit fault diagnosis according to the present invention;

[0019] Figure 2 is a schematic diagram of fault data and modal decomposition data;

[0020] Figure 3 is a schematic diagram of the Conditional-DDPM model structure;

[0021] Figure 4 is Figure 3 a schematic diagram of the U-Net3+ model structure shown; Detailed Description of the Invention

[0022] The following describes the specific embodiments of the present invention with reference to the drawings, so that those skilled in the art can better understand the present invention. It should be particularly noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0023] Embodiment

[0024] Figure 1 is a flowchart of a method for generating training data for analog circuit fault diagnosis according to the present invention.

[0025] In this embodiment, as Figure 1 shown, a method for generating training data for analog circuit fault diagnosis according to the present invention includes the following steps:

[0026] (1) Obtain the true output data of the target analog circuit in the normal state and different fault states;

[0027] Assume that the target analog circuit has a total of n fault modes, and each fault mode F i corresponds to b output data, and the length of each output data is N. Then the true fault data set in different fault states is represented as: F = {F1, F2,..., F i ,..., F n}, Fi = {f ij}, i = 1, 2, …, n, j = 1, 2, …, b, f ij represents the j-th output data under the i-th fault state;

[0028] In this embodiment, it is assumed that there are 5 types of faults, 20 real fault data are collected for each type of fault, and the length of each output data is 2048;

[0029] (2), Modal decomposition;

[0030] Perform modal decomposition on each output data under each fault state, and then take the first m main modes to form a fault data training set represents the k-th modal data after decomposition of the j-th output data under the i-th fault state;

[0031] Perform modal decomposition on each output data in the normal state, and then take the first m main modes to form a normal data training set represents the k-th modal data after decomposition of the j-th output data;

[0032] In this embodiment, the modal data of the first 4 main modes are taken to form a fault data training set; as Figure 2 shown, from top to bottom are the original data of the 2nd data of fault type 1, and the data of the first 4 modes after modal decomposition, and the length of each modal data is 2048.

[0033] (3), Establish a Conditional-DDPM model using an adaptive strategy;

[0034] As Figure 3 shown, the Conditional-DDPM model includes a learnable scale coefficient matrix W, a conditional function H for guiding the generation direction, and a U-Net3+ model for noise estimation;

[0035] Among them, as Figure 4As shown in the figure, the U-Net3+ model successively includes, in the direction of signal flow: an encoding module 1, a downsampling module 1, an encoding module 2, a downsampling module 2, an encoding module 3, a downsampling module 3, an intermediate module, a channel connection module 1, a decoding module 1, a channel connection module 2, a decoding module 2, a channel connection module 3, a decoding module 3, and finally a variable-channel convolution module; among them, all encoding modules are implemented with three consecutive convolutional structures. The output of each module is the same as the shape of the input data, and the number of output channels successively becomes 10 times, 2 times, and 2 times that of the input data; the downsampling module is implemented with a convolutional module with a stride value of 2, which does not change the number of channels of the input data, but the output size is half of the input data size; the three convolutional structures of the intermediate module first change the number of channels of the input data to half, and then double it, while the data size remains unchanged; the channel connection module 1 performs size matching on the outputs of the downsampling module 1, the downsampling module 2, the downsampling module 3, and the intermediate module through convolutional operations, then connects them according to the channels, and inputs them into the decoding module 1; the channel connection module 2 performs size matching on the outputs of the downsampling module 1, the downsampling module 2, the intermediate module, and the decoding module 1 through convolutional operations, then connects them according to the channels, and inputs them into the decoding module 2; the channel connection module 3 performs size matching on the outputs of the downsampling module 1, the intermediate module, the decoding module 1, and the decoding module 2 through convolutional operations, then connects them according to the channels, and inputs them into the decoding module 3; the decoding modules 1, 2, and 3 are each composed of 3 convolutional structures, which respectively change the channels and sizes of the output data of the channel connection modules 1, 2, and 3 to be the same as those of the encoding modules 3, 2, and 1. However, among them, the number of channels of the output data of the decoding module 3 is 10, and then a final convolutional module is used to restore the number of channels to 1; for all encoding, decoding, and intermediate modules, their inputs and outputs are connected in a residual manner, and the input time t is also added as a common input;

[0036] (4), Train the Conditional-DDPM model;

[0037] (4.1), Set the training parameters;

[0038] Set the noise addition time t, t ∈ [1, 1000], and initialize t = 1000;

[0039] Set a Gaussian white noise ε with a length of N = 2048 and obeying the standard Gaussian distribution N(0, 1) t ;

[0040] Set the noise addition coefficient β t , β t = 0.001t;

[0041] Set the noise addition coefficient α t , αt =1-β t ;

[0042] Set the cumulative noise factor

[0043] (4.2) From the fault data training set Randomly select a type of fault data from take out All modal data of the kth mode in As training data;

[0044] (4.3), the modal data Treated as initial data Then the initial data Add noise to get the noise data at time t

[0045]

[0046] (4.4) The corresponding time t is input into the U-Net3+ neural network model to output the estimated noise

[0047] (4.5) Calculate the loss value;

[0048]

[0049] (4.6) Let j = j + 1, then return to step (4.3) and repeat the training until the loss value converges, and obtain the U-Net3+ network model corresponding to the k-th modality training;

[0050] (4.7) Traverse all m modes and train a total of m U-Net3+ network models corresponding to the fault type data Each mode of

[0051] (4.8), let i = i + 1, and then return to step (4.2) until all fault type data training is completed;

[0052] (5) Use the trained model to generate data;

[0053] (5.1), set the reference data set;

[0054] Set b reference data Y for each fault i ij , j = 1, 2, ..., b, and then for each reference data Y ij Perform modal decomposition and take each reference data Y ij The first m modal data of constitute the reference data set;

[0055] (5.2) Take the reference data Y ij The first m modal data of [data] are formed into a reference matrix Y0 row by row, and its size is m×N;

[0056] (5.3) Initialize a scale coefficient matrix W, which is an m×m diagonal matrix, and initialize the initial values of all diagonal elements to 0.001;

[0057] (5.4) Initialize an unconditional denoising matrix X t , whose size is m×N, and each row of it is a noise sequence obeying the standard Gaussian distribution;

[0058] (5.5) At time t, add noise to each row of the reference matrix Y0 to obtain the noisy reference matrix Y t ;

[0059] (5.6) Use the conditional function H to splice the unconditional denoising matrix X t and the noisy reference matrix Y t in the following splicing method to obtain the conditional denoising matrix X′ t ;

[0060] X′ t = H(X t , Y t ) = W·Y t +(I - W)·X t

[0061] where I is an m×m identity matrix;

[0062] (5.7) Input each row of the conditional denoising matrix X′ t at time t into the corresponding U-Net3+ network model to obtain m groups of estimated noises

[0063] (5.8) Substitute the m groups of estimated noises and the conditional denoising matrix X′ t into the following formula to obtain the unconditional denoising matrix X at time t - 1 t-1 ;

[0064]

[0065] where Z is a matrix of size m×N, and each row of it is a noise sequence obeying the standard Gaussian distribution;

[0066] (5.9) Judge whether the current time t satisfies: t > 1; if t > 1 is satisfied, then let t = t - 1, and return to step (5.2); otherwise, the obtained denoising matrix Xt And put X t into the generated dataset G corresponding to the fault mode i i , then determine whether the amount of data in G i at this time reaches b. If the amount of data is less than b, return to step (5.1) and continue to generate data for the fault mode i. Otherwise, the generation ends and enter step (5.10);

[0067] (5.10), Adjust the scale coefficient matrix W according to the adaptive function;

[0068] (5.10.1), Calculate the maximum mean difference loss L i between the fault dataset F i and the generated dataset G MMD and the variance loss L i of the generated dataset G itself var ;

[0069]

[0070] Among them, indicates that the calculation is carried out in the Hilbert space, and g ik , g ij respectively represent the k-th row and the j-th row elements of G i ;

[0071] (5.10.2), Determine whether the generated dataset G i simultaneously satisfies the following conditions:

[0072]

[0073] If G i satisfies the conditions, the generation of the fault mode i ends. Let i = i + 1, and then return to step (5) to generate data for the (i + 1)-th fault mode; If G i does not satisfy the conditions, enter step (5.10.3);

[0074] (5.10.3), Calculate the combined loss L U :

[0075] L U = L MMD + L var

[0076] (5.10.4), According to the combined loss L U , use the Adam optimization strategy to optimize the scale coefficient matrix W, and then substitute the optimized scale coefficient matrix W back into step (5.1) to regenerate data for the fault mode i until the generated data meets the conditions;

[0077] (6), Output the training data set;

[0078] Add the modalities of all the generated fault data to the real fault data set F in the order of fault and modality type to form an augmented fault data set The data set for training the fault diagnosis classifier.

[0079] Although the above-described illustrative specific embodiments of the present invention have been described to facilitate the understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

Claims

1. A method for generating training data for analog circuit fault diagnosis, characterized in that, Including the following steps: (1) Obtain the true output data of the target analog circuit in the normal state and different fault states; Suppose the target analog circuit has a total of n fault modes, and each fault mode F i corresponds to b output data, and the length of each output data is N. Then the true fault data set under different fault states is expressed as: F = {F1, F2,, F i , …, F n}, F i = {f ij}, i = 1, 2,, n, j = 1, 2, …, b, f ij represents the j-th output data in the i-th fault state; Suppose that the target analog circuit also corresponds to b pieces of output data in the normal state, and the length of each piece of output data is N. Then the normal data set in the normal state is expressed as: P = {p1, p2,, p j ,…, p b}; (2) For each piece of output data in each fault state, perform modal decomposition, and then take the first m main modes to form a fault data training set represents the k-th modal data after decomposition of the j-th piece of output data in the i-th fault state; Perform modal decomposition on each piece of output data in the normal state, and then take the first m main modes to form a normal data training set Denote the k-th modal data after decomposition of the j-th piece of output data; (3) Establish a Conditional-DDPM model using an adaptive strategy; The Conditional-DDPM model includes a learnable scale coefficient matrix W, a conditional function H for guiding the generation direction, and a U-Net3+ model for noise estimation; (4) Train the Conditional-DDPM model; (5) Use the trained model for data generation; (5.1) Set a reference data set; Set b reference data Y for each fault i ij , j = 1, 2, …, b, and then for each reference data Y ij perform modal decomposition, and take the first m modal data of each reference data Y ij to form a reference data set; (5.2) Take the reference data Y ij The first m modal data of are formed into a reference matrix Y0 by rows, and its size is m×N; (5.3) Initialize a scale coefficient matrix W, which is an m×m diagonal matrix, and initialize the initial values of all diagonal elements to 0.001; (5.4) Initialize an unconditional denoising matrix X t , with a size of m×N, and each row of it is a noise sequence following a standard Gaussian distribution; (5.5) At time t, add noise to each row of the reference matrix Y0 to obtain the reference matrix Y after adding noise. t ; (5.6) Use the conditional function H on the unconditional denoising matrix X t and the reference matrix Y after adding noise t to perform splicing in the following way to obtain the conditional denoising matrix X' t ; X′ t = H(X t , Y t ) = W·Y t + (I - W)·X t where I is an m×m identity matrix; (5.7), input each row of the conditional denoising matrix X′ t at time t into the corresponding U-Net3+ network model to obtain m groups of estimated noise (5.8), Substitute m groups of estimated noises and the conditional denoising matrix X′ t into the following formula to obtain the unconditional denoising matrix X at time t - 1 t-1 ; where Z is a matrix of size m×N, and each row of it is a noise sequence subject to a standard Gaussian distribution; (5.9) Determine whether the current time t satisfies: t > 1; if t > 1 is satisfied, then set t = t - 1 and return to step (5.2); otherwise, obtain the denoising matrix X t and put X t into the generated data set G corresponding to the fault mode i i and then determine whether the amount of data in G i at this time reaches b. If the amount of data is less than b, return to step (5.1) and continue to generate data for the fault mode i; otherwise, the generation ends and step (5.10) is entered; (5.10) Adjust the scale coefficient matrix W according to the adaptive function; (5.10.1), calculate the fault data set F i and the generated data set G i the maximum mean difference loss L MMD between them, and the variance loss L i of the generated data set G var ; Among them, indicates that the calculation is performed in a Hilbert space, and g ik , g ij respectively represent the k-th row and j-th row elements of G i ; (5.10.2), Determine whether the generated dataset G i simultaneously satisfies the following conditions: If G i satisfies the condition, the generation of failure mode i ends. Let i = i + 1, and then return to step (5) to generate data for the (i + 1)-th failure mode. If G i does not satisfy the condition, then proceed to step (5.10.3); (5.10.3), Calculate the combined loss L U : L U = L MMD + L var (5.10.4) According to the combined loss L U , use the Adam optimization strategy to optimize the scale coefficient matrix W, and then substitute the optimized scale coefficient matrix W back into step (5.1) to regenerate data for the fault mode i until the generated data meets the conditions; (6) Output the training data set; Add the modalities of all the generated fault data to the real fault data set F in the order of fault and modality type to form the augmented fault data set. A data set for training a fault diagnosis classifier.

2. A method for generating training data for analog circuit fault diagnosis according to claim 1, characterized in that The U-Net3+ model sequentially includes, according to the signal flow direction: an encoding module 1, a downsampling module 1, an encoding module 2, a downsampling module 2, an encoding module 3, a downsampling module 3, an intermediate module, a channel connection module 1, a decoding module 1, a channel connection module 2, a decoding module 2, a channel connection module 3, a decoding module 3, and finally a variable-channel convolution module; among them, all encoding modules are implemented with three consecutive convolution structures, the output of each module is consistent with the shape of the input data, and the output channel numbers sequentially become 10 times, 2 times, and 2 times that of the input data; the downsampling module is implemented with a convolution module with a stride value of 2, which does not change the channel number of the input data, but the output size is half of the input data size; the three convolution structures of the intermediate module first change the channel number of the input data to one-half, and then change it to twice, with the data size remaining unchanged; the channel connection module 1 performs size matching on the outputs of the downsampling module 1, the downsampling module 2, the downsampling module 3, and the intermediate module through convolution operations, then connects them according to channels, and inputs them into the decoding module 1; the channel connection module 2 performs size matching on the outputs of the downsampling module 1, the downsampling module 2, the intermediate module, and the decoding module 1 through convolution operations, then connects them according to channels, and inputs them into the decoding module 2; the channel connection module 3 performs size matching on the outputs of the downsampling module 1, the intermediate module, the decoding module 1, and the decoding module 2 through convolution operations, then connects them according to channels, and inputs them into the decoding module 3; the decoding modules 1, 2, and 3 are each composed of 3 convolution structures, which respectively change the channels and sizes of the output data of the channel connection modules 1, 2, and 3 to be the same as those of the encoding modules 3, 2, and 1, but among them, the channel number of the output data of the decoding module 3 is 10, and then its channel number is restored to 1 through a final convolution module; for all encoding, decoding, and intermediate modules, their inputs and outputs are connected in a residual manner, and the input time t is also added as a common input.

3. A method for generating training data for analog circuit fault diagnosis according to claim 1, characterized in that, The training process of the Conditional-DDPM model is as follows: (3.1) Set training parameters; Set the noise addition time t, t ∈ [1, 1000], and initialize t = 1000; Set Gaussian white noise ε with length N and following the standard Gaussian distribution N(0, 1) t ; Set the noise addition coefficient β t , β t = 0.001t; Set the noise addition coefficient α t , α t = 1 - β t ; Set the cumulative noise addition coefficient (3.2) Randomly extract a type of fault data from the fault data training set Extract all the modal data of the k-th mode in as training data; Extract (3.3), regard the modal data as the initial data Then, perform noise addition on the initial data to obtain the noisy data at time t ​ (3.4), Input the noisy data and the corresponding number of time instants t into the U-Net3+ neural network model, so as to output the estimated noise (3.5) Calculate the loss value; (3.6) Let j = j + 1, then return to step (3.3) to repeat the training until the loss value loss converges, and obtain the U-Net3+ network model corresponding to the k-th mode training; (3.7) Traverse all m modalities, and train a total of m U-Net3+ network models corresponding to the data of each modality of the fault type ; (3.8) Let i = i + 1, then return to step (3.2) until all fault type data training is completed.

Images