Lifetime prediction method for electronic load MOS tube based on multi-modal domain invariant features

Abstract

The present application proposes a life prediction method for an electronic load MOS tube based on multi-modal domain invariant features. The method AD collects the sampling resistor voltage, the MOS tube drain voltage and the gate current, adaptively slides and slices to obtain three modal slice samples, establishes an LSTM‑WGAN‑GP network model, and performs data preprocessing. The three modal slice samples are spliced ​​to obtain spliced ​​time series data, and then the time series data is spectrally transformed. The core spectrum sequence and the three modal slice samples are input into the multi-modal domain invariant feature extraction model S‑DAN through multi-channel convolution to extract key features, and then the proposed features are input into a recurrent network prediction model MC‑RNN combined with random noise to predict the remaining service life of the MOS tube. The present application proposes a life prediction method for an electronic load MOS tube with multi-modal domain invariant features, which can improve the reliability and service life of electronic load circuits.

Description

Technical Field

[0001] The present application belongs to the field of artificial intelligence deep learning, and specifically relates to a method for predicting the life of an electronic load MOS tube based on multi-modal domain invariant features. Background Art

[0002] With the development of modern industry and electronic systems, electronic loads have been widely used in power supply testing, equipment operation and other fields. As the operating time increases, the internal components of the electronic load, such as power semiconductor devices, capacitors, resistors, etc., will gradually age and eventually lead to device damage. This damage is often a gradual degradation process with a degradation trend. By establishing an effective life prediction model, preventive maintenance measures can be taken in time before the MOS tube is damaged, avoiding economic losses and safety hazards caused by sudden failures.

[0003] Traditional methods are highly dependent on a large amount of historical data and are difficult to apply to scenarios where data is scarce or new equipment is needed. In order to improve the accuracy of life prediction, modern life prediction technology often uses deep learning networks to achieve fast and accurate life prediction of electronic loads by training neural network models. At the same time, the MC-RNN prediction model combined with random noise can greatly improve the efficiency and stability of diagnosis. Summary of the invention

[0004] The purpose of this application is to provide a method for predicting the life of an electronic load MOS tube, which adopts an improved timing deep learning network model and can quickly and accurately diagnose the remaining service life of the electronic load.

[0005] In order to solve the above technical problems, the present application proposes an electronic load MOS tube life prediction method based on multi-modal domain invariant features, including:

[0006] The sampling resistor voltage, drain voltage and gate current of the electronic load MOS tube are collected, and three modal slice samples are obtained through adaptive sliding slicing;

[0007] LSTM-WGAN-GP network model, performs data preprocessing on three modal slice samples;

[0008] Data splicing is used to enhance the correlation of the three modal slice samples, and spectrum transformation and core spectrum value screening are performed;

[0009] A new multimodal domain-invariant feature extraction model S-DAN is established to extract domain-invariant features for MOS tube life prediction.

[0010] A recurrent network prediction model MC-RNN combined with random noise is trained, and domain-invariant features are input into MC-RNN to predict the life of MOS tubes.

[0011] Optionally, the sampling resistor voltage, the drain voltage and the gate current of the MOS tube of the electronic load are collected, and three modal slice samples are obtained by adaptive sliding slicing, including:

[0012] The high-frequency signal collector collects the sampling resistor voltage, drain voltage and gate current of the electronic load MOS tube, and then dynamically adjusts the sliding window size according to the changing trend of the data to slice the sampled data to better capture the dynamic characteristics of the data and obtain three modal slice samples.

[0013] Optionally, an LSTM-WGAN-GP network model is established to perform data preprocessing on the three modal slice samples, including:

[0014] The long short-term memory neural network (LSTM) is used as the model architecture. Based on its ability to obtain temporal correlation information of the previous and next moments, the Wasserstein distance is used as the objective function to train the generative adversarial network model WGAN. A gradient penalty term WGAN-GP is added to the objective function of the Wasserstein distance to avoid excessive data gradients during model training, which may lead to the inability to find the global optimum for model parameters.

[0015] Optionally, data splicing is used to enhance the correlation of the three modal slice samples, and spectrum transformation and core spectrum value screening are performed, including:

[0016] The three pre-processed interference-free modal slice samples are spliced ​​to obtain spliced ​​time series data, and then the time series data is spectrally transformed. Based on big data statistics, the core spectrum values ​​that are strongly correlated with the life of the MOS tube are screened out from the spectrum sequence to obtain the core spectrum sequence.

[0017] Optionally, a new multimodal domain invariant feature extraction model S-DAN is established to extract domain invariant features for MOS tube life prediction, including:

[0018] A multimodal domain-invariant feature extraction model S-DAN is established. The core spectrum sequence and three modal slice samples are input into the model. The high-dimensional features are mapped to the low-dimensional space through the sparse self-attention mechanism algorithm to reduce the computational complexity. Then, the multimodal domain-invariant MOS tube life prediction features are extracted through the domain adversarial algorithm.

[0019] Optionally, a recurrent network prediction model MC-RNN combined with random noise is trained, and the domain invariant features are input into the MC-RNN to predict the life of the MOS tube, including:

[0020] A recurrent network prediction model MC-RNN combined with random noise is trained, and the RNN model parameters are randomly updated by adding random noise to ensure the global search capability of the model. The extracted multi-modal domain invariant MOS tube life prediction features are input into MC-RNN to predict the life of the electronic load MOS tube.

[0021] The present application proposes a life prediction method for an electronic load MOS tube based on multi-modal domain invariant features. The method AD collects the sampling resistor voltage, the MOS tube drain voltage and the gate current, adaptively slides and slices to obtain three modal slice samples, establishes an LSTM-WGAN-GP network model, and performs data preprocessing. The three modal slice samples are spliced ​​to obtain spliced ​​time series data, and then the time series data is spectrally transformed. The core spectrum sequence and the three modal slice samples are input into the multi-modal domain invariant feature extraction model S-DAN through multi-channel convolution to extract key features, and then the proposed features are input into a recurrent network prediction model MC-RNN combined with random noise to predict the remaining service life of the MOS tube. The present application proposes a life prediction method for an electronic load MOS tube with multi-modal domain invariant features, which can improve the reliability and service life of the electronic load circuit. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0023] Figure 1 A schematic diagram of the process flow of the method for predicting the service life of an electronic load MOS tube provided by the present invention;

[0024] Figure 2 The framework diagram of the multimodal domain invariant feature extraction model S-DAN provided by the present invention;

[0025] Figure 3 This is a framework diagram of the recurrent network prediction model combined with random noise provided by the present invention. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0027] like Figure 1 As shown, Figure 1 The present invention provides a flow chart of a method for predicting the service life of an electronic load MOS tube, which includes five parts.

[0028] S11: Collect the sampling resistor voltage, MOS tube drain voltage and gate current of the electronic load MOS tube, and obtain three modal slice samples through adaptive sliding slicing.

[0029] It should be noted that this application adopts an improved sliding slicing method to dynamically adjust the window size according to the data characteristics, rather than using a fixed-size window. This method can better capture the dynamic change characteristics of the voltage data and current data of the electronic load MOS tube by flexibly adjusting the window length in different time periods or data areas, and specifically includes 5 steps.

[0030] Step 11: Change detection. The collected data is the time series x(t). The size a of the adaptive sliding window is adjusted according to the local changes of the sequence. The local mean u calculated by formula (1) is used. a To detect whether there is a local change in the data. When a local mean deviates from the normal range, there is a local change in the data.

[0031]

[0032] Step 12: Variance threshold detection. When local changes are detected, the local variance of the window data is calculated using formula (2). if If the value exceeds a certain threshold θ, the window size is adjusted.

[0033]

[0034] Step 13: Window resizing, window size W a Dynamically adjust according to the changes. In the data area with large variance, reduce the window size; in the area with relatively stable data, increase the window size. Adjust the rule (3), where α is the step size of window reduction, β is the step size of window expansion, and W min and W max are the minimum and maximum window lengths, respectively.

[0035]

[0036] Step 14: Sliding step size adjustment: In the adaptive sliding window, the sliding step size S a The adjustment of is related to the window size through equation (4), where γ is a constant proportional to the step size and the window size.

[0037] Sa =γ·W a (4)

[0038] Step 15: Repeat steps 12-14 until all data are processed to form a new data set.

[0039] Based on the above discussion, in an optional embodiment of the present application, for the above-mentioned collection of the electronic load MOS tube sampling resistor voltage, the MOS tube drain voltage and the gate current, obtaining a multimodal slice sample through adaptive sliding slicing may include: in an optional embodiment of S11, the current sampling frequency is 20kHz, the default window size is set to 12000, the default sliding step size is set to 1000, the step size α of the window reduction is set to 1000, the step size β of the window expansion is set to 1000, and the proportional constant γ of the step size to the window size is set to 0.083.

[0040] S12: Establish an LSTM-WGAN-GP network model and perform data preprocessing on the three modal slice samples.

[0041] It should be noted that this application establishes the LSTM-WGAN-GP model for data preprocessing because the model can effectively deal with the problem of missing data to ensure data quality, which specifically includes three steps.

[0042] Step 21: Use the long short-term memory neural network LSTM as the model architecture, based on which the temporal correlation information of the previous and next moments can be obtained, and use the Wasserstein gradient penalty adversarial model WGAN-GP to extract the data distribution characteristics of the normal modal slice samples.

[0043] Step 22: In order to train the LSTM-WGAN-GP model using semi-supervised learning, the normal historical modal slice sample points are randomly removed, and the removed sample points are used as labels to establish a missing value sample training data set.

[0044] Step 23: (1) The GAN model consists of a generator and a discriminator. In the adversarial training between the two, the samples generated by the GAN model are infinitely close to the real samples; (2) When the data distribution of the samples generated by the GAN is completely disjoint with that of the real samples, the JS divergence of the sample data distribution is a constant. At this time, the gradient is 0 and the model cannot be optimized further. To make up for the limitations of the GAN model, WGAN-GP optimizes the data distribution distance between the real samples and the generated samples through the Wasserstein distance of formula (5), and adds a penalty term to avoid the data gradient being too large during model training, which causes the model parameters to be unable to find the global optimum. is the data distribution characteristic of the true sample, is the data distribution feature of the generated sample, D is the discriminator, x is the true sample, θ is the discriminator network parameter, G is the discriminator, c is the generated sample, ρ is the generator network parameter, is the discriminator gradient, τ is the learning rate; (3) the missing value samples are input into the generator, and the generator supplements the missing values ​​and then enters the discriminator. At this time, the distribution distance between the normal historical modal slice samples and the missing value samples is calculated by the Wasserstein distance, and a penalty term is added in the calculation process. The discriminator outputs the difference loss between the generated samples and the original samples, and the model is trained with minimizing this difference loss as the optimization goal.

[0045]

[0046] Step 24: After the model is trained, the sample set with missing data is input into the model. Based on the normal data distribution characteristics that have been learned, the missing data can be supplemented and a multimodal slice sample set without missing values ​​can be output.

[0047] Based on the above discussion, in an optional embodiment of the present application, the training of the above LSTM-WGAN-GP network model may include: in an optional embodiment of S13, the generator uses a 5-layer LSTM network, the discriminator uses a 3-layer LSTM network, the hidden layer activation function uses a Relu function, and the SGD optimization algorithm is used when training the model, the learning rate τ is 0.001, and the number of iterations is 500.

[0048] S13: Data splicing is used to enhance the correlation of the three modal slice samples, and spectrum transformation and core spectrum value screening are performed.

[0049] It should be noted that the accuracy of MOS tube life prediction based on the features extracted from single-modal samples is not high. This application integrates the feature information in multiple modal samples to improve the model prediction accuracy and proposes a data splicing and fusion method to improve the data correlation of multi-modal samples. The purpose of performing spectrum transformation on the spliced ​​data is to decompose the collected electronic load data into different frequency components to obtain more obvious periodic changes in the signal. Because the spectrum characteristics of normal data are weak, when the MOS tube is in a poor health state, the spectrum value will surge. By screening out these surged spectrum values, the remaining useless information can be eliminated, the data redundancy can be reduced, and the model training speed and accuracy can be improved. Specifically, there are 4 steps.

[0050] Step 31: First, the three modal slice samples are spliced ​​end to end to improve the correlation between the data.

[0051] Step 32: To reduce spectrum leakage of the spliced ​​samples during spectrum transformation, the Hanning window function shown in equation (6) is used to intercept the data to reduce the discontinuity at both ends of the signal, thereby reducing spectrum leakage, where T is the window function width and z is the spliced ​​data.

[0052]

[0053] Step 33: Perform spectrum transformation on the sample data intercepted by the Hanning window function through the fast Fourier transform algorithm of formula (7) to obtain the spectrum sequence of the sliced ​​samples, where z(o) represents the oth frequency component in the frequency domain signal.

[0054]

[0055] Step 34: Using the MOS tube life prediction accuracy as the evaluation index, the spectrum peak value threshold ε is learned from the historical data through formula (8), where u f represents the mean of the historical spectrum values, σ f Represents the standard deviation of the historical spectrum value, and n is the standard deviation multiple of statistical learning. Take the spectrum peak value greater than the threshold ε, and form a two-dimensional array of the selected spectrum peak value and its corresponding frequency point in the form of [spectrum peak value, frequency point]. Finally, perform a one-dimensional conversion on the two-dimensional array to obtain the core spectrum value sequence.

[0056] ε=nσ f +u f (8)

[0057] Based on the above discussion, in an optional embodiment of the present application, the above-mentioned data splicing is used to enhance the correlation of the three modal slice samples, and the spectrum transformation and core spectrum value screening can include: in an optional embodiment of S13, the window function width T is set to 8192, the frequency interval of the spectrum sequence is 2.4Hz, the optimized standard deviation multiple n is selected to be 6, and the least squares method is selected as the objective function of the MOS tube life prediction accuracy.

[0058] S14: Establish a new multimodal domain-invariant feature extraction model S-DAN to extract domain-invariant features for MOS tube life prediction.

[0059] It should be noted that the multimodal domain-invariant feature extraction model S-DAN established in this application calculates the similarity between the elements of each input sequence and other elements through the feature sparse self-attention mechanism algorithm, obtains the importance of each element to other elements, reduces the computational complexity, and finally extracts the multimodal domain-invariant MOS tube life prediction features through the domain adversarial algorithm, which specifically includes two steps.

[0060] Step 41: The core spectrum sequence and the three modal slice samples are sent to S-DAN through multi-channel convolution. The weights of non-important elements are reduced through the sparse self-attention mechanism algorithm to reduce the amount of calculation. The weight calculation formula and sparse feature generation formula of the sparse self-attention mechanism are as shown in formulas (9)-(13), where W Q is the projection matrix that generates the query vector, W K is the projection matrix that generates the key vector, W V is the projection matrix that generates the value vector, q i is the core spectrum sequence element p i The query element is to find the part related to itself from the input, k j Indicates the jth key element that matches the query. A (i,j) represents the i-th query q i For the jth key k j The attention weight, Represents the query q i Only the set of keys that are calculated is obtained by filtering the similarity with all keys, d k is the dimension of the key vector and exp(·) is the exponential operation of the attention score, which is used to measure the similarity between the query and the key.

[0061]

[0062] q i =p i W Q (10)

[0063] k j =p i W K (11)

[0064] v i =W V p i (12)

[0065]

[0066] Step 42: Input the sparse feature sequence generated in step 41 into the feature adversarial layer to extract the key features for life prediction, as shown in equations (14)-(16), where G y represents the classifier, G d represents the domain classifier, y c represents the true label of the sample, C represents the number of categories of the classification task, and L y represents the classification task loss of the source domain, L d represents the loss of the domain classifier, and μ represents the hyperparameter of the balanced classification loss and the domain classification loss.

[0067]

[0068] L d =-[d i log(G d (z) +(1-d i )log(1-G d (z))] (15)

[0069]

[0070] Based on the above discussion, in an optional embodiment of the present application, the multimodal domain invariant feature extraction model S-DAN may include: in an optional embodiment of S14, the SGD optimization algorithm is used when training the model, the learning rate is 0.001, the hyperparameter μ is set to 0.1, and the number of iterations is 500.

[0071] S15: Train a recurrent network prediction model MC-RNN combined with random noise, input domain-invariant features into MC-RNN, and predict the life of MOS tubes.

[0072] It should be noted that the features extracted by S14 are time series features. In order to extract the health information of the MOS tube from the time series features, this application uses a time series RNN model to predict the life of the MOS tube. Because the life information of the MOS tube is not easy to extract, it is difficult for traditional optimization methods to find the global optimal parameters. This application relies on adding random noise to randomly update the parameters, ensuring the global search capability of the model and avoiding the model parameters from falling into the local optimum during training. It specifically includes 3 steps.

[0073] Step 51: The RNN model can further extract the domain-invariant key features. The output of each hidden layer is not only related to the input at the current time point, but also to the hidden state at the previous time point. Therefore, the RNN can integrate the feature information of the previous time point to further improve the model accuracy. The output of a hidden layer at time t can be obtained by formula (17), where F is the activation function and S is t-1 is the current hidden layer input, C t is the hidden state at the previous time point.

[0074] S t =F(W C C t +W S S t-1 ) (17)

[0075] Step 52: In order to further capture the dependency between domain-invariant key feature sequences and enhance the relevance of data, the domain-invariant key feature sequences are divided into n feature vectors. Each feature vector is sequentially sent to the RNN model in units of feature vectors, so that when the RNN model extracts features from each feature vector, the output of the extracted feature information at a certain moment is not only related to the current input, but also to its state at the previous time.

[0076] Step 53: Apply a random noise to the trained model parameters through equation (18), and then perform parameter back propagation to avoid the model parameters falling into the local optimum. The model training parameters are After adding noise, the model parameters are τ is the learning rate, is the added random noise.

[0077]

[0078] Based on the above discussion, in an optional embodiment of the present application, the above-mentioned recurrent network prediction model MC-RNN combined with random noise may include: in an optional embodiment of S15, the activation function F selects the Relu function, the learning rate τ is set to 0.01, and the random noise Select Gaussian Noise.

[0079] The present application proposes a life prediction method for electronic load MOS tubes based on multi-modal domain invariant features. This method AD collects sampling resistor voltage, MOS tube drain voltage and gate current, obtains three modal slice samples through adaptive sliding slicing, establishes LSTM-WGAN-GP network model, and performs data preprocessing on the three modal slice samples. The three modal samples are spliced ​​to obtain spliced ​​time series data, and then the time series data is spectrally transformed. The core spectrum sequence and the three modal slice samples are input into the multi-modal domain invariant feature extraction model S-DAN through multi-channel convolution to extract key features, and then the proposed features are input into a recurrent network prediction model MC-RNN combined with random noise to predict the remaining service life of the MOS tube.

[0080] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Specific examples are used herein to illustrate the principles and implementation methods of the present invention, and the description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. The electronic load MOS tube life prediction method based on multi-modal domain invariant characteristics is characterized by: include: The sampling resistor voltage, drain voltage and gate current of the electronic load MOS tube are collected, and three modal slice samples are obtained through adaptive sliding slicing; Establish LSTM-WGAN-GP network model and perform data preprocessing on three modal slice samples; Data splicing is used to enhance the correlation of the three modal slice samples, and spectrum transformation and core spectrum value screening are performed; A new multimodal domain invariant feature extraction model S-DAN is established to extract the domain invariant features of MOS tube life prediction. The core spectrum sequence and three modal slice samples are sent to S-DAN through multi-channel convolution. The sparse self-attention mechanism algorithm is used to reduce the weight of non-important elements and the amount of calculation. The weight calculation formula and sparse feature generation formula of the sparse self-attention mechanism are shown in formulas (1)-(5), where W Q is the projection matrix that generates the query vector, W K is the projection matrix that generates the key vector, W V is the projection matrix that generates the value vector, q i is the core spectrum sequence element p i The query element is to find the part related to itself from the input, k j Indicates the jth key element, matching the query, S A (i,j) represents the i-th query q i For the jth key k j The attention weight, Represents the query q i Only the set of keys that are calculated is obtained by filtering the similarity with all keys, d k is the dimension of the key vector, exp(·) is the exponential operation of the attention score, which is used to measure the similarity between the query and the key. The generated sparse feature sequence is input into the feature adversarial layer to extract the key features for life prediction, as shown in Equations (6)-(8), where G y represents the classifier, G d represents the domain classifier, y c represents the true label of the sample, C represents the number of categories of the classification task, and L y represents the classification task loss of the source domain, L d represents the loss of the domain classifier, μ represents the hyperparameter of the balanced classification loss and the domain classification loss; q i =p i W Q (2) k j =p i W K (3) v i =W V p i (4) L d =-[d i log(G d (z) +(1-d i )log(1-G d (z))] (7) A recurrent network prediction model MC-RNN combined with random noise is trained, and domain-invariant features are input into MC-RNN to predict the life of MOS tubes.

2. The method for predicting the life of an electronic load MOS tube as claimed in claim 1, characterized in that: The sampling resistor voltage, drain voltage and gate current of the electronic load MOS tube are collected, and three modal slice samples are obtained through adaptive sliding slicing, including: The high-frequency signal collector collects the sampling resistor voltage, drain voltage and gate current of the electronic load MOS tube, and then dynamically adjusts the sliding window size according to the changing trend of the data to slice the sampled data to better capture the dynamic characteristics of the data and obtain three modal slice samples.

3. The method for predicting the life of an electronic load MOS tube as claimed in claim 1, characterized in that: An LSTM-WGAN-GP network model was established to preprocess the data of three modal slice samples, including: The long short-term memory neural network LSTM is used as the model architecture. The temporal correlation information of the previous and next moments is obtained based on it. The Wasserstein distance is used as the objective function to train the adversarial generative network model WGAN, and the gradient penalty term WGAN-GP is added to the objective function of the Wasserstein distance to avoid the model parameters being unable to find the global optimum due to excessive data gradient during model training.

4. The method for predicting the life of an electronic load MOS tube as claimed in claim 1, characterized in that: Data splicing is used to enhance the correlation of the three modal slice samples, and spectrum transformation and core spectrum value screening are performed, including: The three pre-processed interference-free modal slice samples are spliced ​​to obtain spliced ​​time series data, and then the time series data is spectrally transformed. Based on big data statistics, the core spectrum values ​​that are strongly correlated with the life of the MOS tube are screened out from the spectrum sequence to obtain the core spectrum sequence.

5. The method for predicting the life of an electronic load MOS tube as claimed in claim 1, characterized in that: A recurrent network prediction model MC-RNN combined with random noise is trained, and domain-invariant features are input into MC-RNN to predict the life of MOS tubes, including: A recurrent network prediction model MC-RNN combined with random noise is trained. The RNN model parameters are randomly updated by adding random noise to ensure the global search ability of the model. The extracted multi-modal domain invariant MOS tube life prediction features are input into MC-RNN to predict the life of the electronic load MOS tube.

Images