IGBT module state prediction method based on domain adversarial long short-term memory network

Abstract

The application discloses an IGBT module state prediction method based on a domain adversarial long short-term memory network, and comprises the following steps: collecting aging data of different IGBT modules respectively, cutting different sample sequences from sample data of known IGBT modules respectively, and labeling after normalization processing; an adversarial transfer state prediction network for unknown IGBT modules is constructed; in the training stage, source domain and target domain data are mapped into a high-dimensional feature space to obtain data feature distribution; domain invariant features of the two domains are learned by using a domain adversarial module to perform feature distribution matching; a weighting discrimination mechanism is further provided to evaluate the similarity degree of target domain sample data and source domain data, discriminate the transferability of data, and thus effectively improve the classification performance of the state category; target domain test data is input into the trained prediction model for testing, the current running state of the equipment is discriminated by calculating the obtained weight value, and the final classification prediction result is output.

Description

Technical Field

[0001] This invention belongs to the field of intelligent state prediction of electronic devices, specifically relating to a method for predicting the state of IGBT modules based on domain adversarial long short-term memory networks. Background Technology

[0002] With increasing global environmental awareness, electric vehicles have become a crucial direction for future automotive development. IGBT technology plays a vital role in the development of electric vehicles. IGBTs are power semiconductor devices commonly used in high-voltage, high-current applications, possessing advantages such as high switching speed, low conduction voltage drop, and high short-circuit tolerance, and are widely used in EV charging stations (ACPs). IGBTs can be precisely controlled through pulse width modulation technology to achieve efficient power conversion and transmission, providing stable and reliable power output for electric vehicles. Therefore, accurately and promptly identifying the occurrence and evolution of faults in rotating machinery systems, and reducing or avoiding major accidents, is of great and far-reaching significance for ensuring the high-quality and efficient operation of equipment.

[0003] With the development of deep learning technology, deep learning-based device state prediction models have been proposed and widely applied. Deep learning is a branch of machine learning that has many neural nodes and many hidden layers between input and output, helping to learn the nonlinear representation between input and output. Deep neural networks are generated by the network itself, and they can achieve intelligent recognition and prediction of input data without complex model assumptions. Therefore, deep learning-based state prediction methods can better uncover the inherent relationships and nonlinear patterns between data, improving the accuracy of state prediction.

[0004] Using deep learning for state prediction requires a large number of labeled samples. In many cases, collecting sufficient labeled training data is costly, time-consuming, and even impractical. Building separate state prediction models for different tasks is not only time-consuming and labor-intensive but also requires a sufficient number of failure samples. However, in industrial applications, IGBT modules cannot operate with failures for extended periods, and the cost of collecting and labeling lifetime data for each task is extremely high.

[0005] Transfer learning is a method that applies knowledge or patterns learned in one domain or task to different but related domains or problems. It learns new knowledge by using existing knowledge, and its core is to find the similarities between existing and new knowledge, achieving the goal of transfer learning through the transfer of these similarities. Therefore, transfer learning methods can better uncover the degradation similarities between known and unknown devices, thereby improving the universality of lifetime prediction models.

[0006] An IGBT lifetime prediction method based on a GA-Elman-LSTM combined model (CN202211588473.9) effectively improves model prediction accuracy and stability while reducing prediction risk. However, the above method can only be trained on labeled IGBT aging data, resulting in weak model generalization ability and a lack of ability to predict the state of unknown IGBT modules. This invention acquires the health status information of IGBT modules by constructing a bidirectional long short-term memory network (BiLSTM), and introduces the idea of ​​domain adversarial transfer learning. It uses the target domain weight function to adaptively measure the differences between shared faults and unknown new faults in the source and target domains. Through weighted adversarial training and selection of target domain test sample weight thresholds, it achieves the prediction of the operating state of unknown IGBT modules. Summary of the Invention

[0007] To address the problems of existing technologies, this invention designs an unknown device state prediction method based on domain adversarial long short-term memory networks. It proposes an Adversarial Transfer Condition Prediction Network (ATCPN). This method uses a feature extraction module (feature extractor) to map source and target domain data into a high-dimensional feature space to obtain data feature distributions. A time-series prediction module (time-series predictor) extracts low-dimensional degradation features from the high-dimensional feature space to construct IGBT health factors. A state classification module (state classifier) ​​maps the health factors to operating state classes, outputting the final operating state label. A domain adversarial module (domain discriminator) learns domain-invariant features between the two domains, reducing the distribution difference between them and performing feature distribution matching. Since the target domain contains degradation information of unknown IGBT modules, a weighted discrimination mechanism is designed to evaluate the similarity between target domain sample data and source domain data, determine data transferability, and finally determine the data category based on weight values. This achieves adaptive matching between target domain samples and source domain samples, thereby realizing the prediction of the operating state of unknown IGBT modules.

[0008] The present invention is achieved by at least one of the following technical solutions.

[0009] The IGBT module state prediction method based on domain adversarial long short-term memory networks includes the following steps:

[0010] Step 1: Data Acquisition and Sample Division. Aging data for different IGBT modules are collected. Different sample sequence labels are extracted from the known IGBT module sample data to construct the source domain dataset {(x... s ,y s )},x s For source domain data, y sLabels corresponding to the source domain data; Obtain a target domain sample dataset {x} that differs from the source domain degradation information under unknown devices. t}, x t For unlabeled target domain data x t ;

[0011] Step 2: Construct an adversarial transfer state prediction network. The adversarial transfer state prediction network includes a feature extractor G, a temporal predictor P, a state classifier C, and a neighborhood discriminator D. Considering that the input data is a one-dimensional signal, the feature extractor G includes a convolutional layer with a one-dimensional convolutional kernel, a max pooling layer with a one-dimensional pooling kernel, and a channel attention mechanism layer. The state classifier C and the neighborhood discriminator D include fully connected layers.

[0012] Step 3: Using the source domain dataset {(x) constructed in Step 1 s ,y s The parameters of the feature extractor G and the time series predictor P are optimized and trained using supervised learning algorithms and gradient descent.

[0013] Step 4: In the adversarial domain adaptation stage, the state classifier C is used to classify the target domain data, and the domain discriminator D is used to discriminate the target domain data x. t The transferability is assessed, and reverse weighted training is performed;

[0014] Step 5: During the testing phase, input the test sample, compare the obtained target domain weight with the set threshold, and predict the corresponding state classification.

[0015] Furthermore, in step 1, the source domain data consists of labeled samples from known devices, while the target domain consists of unlabeled samples from unknown devices. The source and target domain data follow probability distributions P and Q, respectively, but the data distributions in the two domains are different (P ≠ Q). However, the class labels of the source domain data and the target domain data are the same (C). s =C t ;

[0016] Furthermore, in step 2, the structure of the feature extractor G includes multiple feature extraction units, which are stacked sequentially.

[0017] Furthermore, each feature extraction unit includes a convolutional layer with a one-dimensional convolutional kernel, a batch normalization layer, a max pooling layer with a one-dimensional pooling kernel, and a channel self-attention mechanism layer.

[0018] Furthermore, in step 2, the time-series predictor P includes a forward long short-term memory network and a backward long short-term memory network, both with the same structural parameters; the forward long short-term memory network uses past time information to output the prediction result. Reverse Long Short-Term Memory (LSTM) networks use information from future moments to output prediction results. The Bidirectional Long Short-Term Memory (BiLSTM) network fuses the forward and backward prediction results to obtain the final output.

[0019] Furthermore, in step 2, the state classifier C has a fully connected layer structure, its output layer uses the ReLU output function, and its output label is {0, 1}; the neighborhood discriminator D also has a fully connected layer structure, and its output layer uses the Sigmoid output function.

[0020] Furthermore, in step 3, supervised learning algorithms and gradient descent are used to optimize and train the parameters of the feature extractor G and the state classifier C; the source and target domain data follow probability distributions P and Q, respectively; and it is necessary to classify the unlabeled samples x in the target domain.

[0021]

[0022] Where β(x) represents the classifier; This represents the prediction result. The purpose of transfer learning is to minimize the target risk ε using the source data. t (β):

[0023]

[0024] Where (x,y) represents the sample set in the target domain that follows a Q distribution; y represents the actual label value.

[0025] The parameters of the feature extractor G and the state classifier C are used to construct the classification loss function L. c (θ f ,θ c ):

[0026]

[0027]

[0028] Where, θ f θ represents the learnable parameters of the feature extractor G. c D represents the learnable parameters of the state classifier C; s Represents the source domain. Let X represent the i-th sample in the source domain. s Represents the set of all samples in the source domain. Y represents the i-th label of the i-th sample. s Let n represent the set of all labels. s E represents the total number of source samples; C' represents the expected value; and c represents the actual label value. This represents the output label of the state classifier C; This represents a probability function.

[0029] Furthermore, in step 4, the Conditional Domain Adversarial Network (CDAN) is used to achieve knowledge transfer. CDAN first needs to define multiple linear mappings. That is, the outer product of multiple random vectors; if two random vectors x and y are given, the mean mapping... Fully capture the complex multimodal structure and cross-covariance within the data. The loss function for modeling the joint distribution P(x,y) is:

[0030]

[0031] Where φ(·) represents the mapping of the sample in the Hilbert space, L CDAN (θ f ,θ d ) represents the conditional adversarial loss function; θ f θ represents the learnable parameters of the feature extractor G. d G represents the learnable parameters of the domain discriminator D; d Represents the neighborhood discriminator D; The source domain samples extracted by feature extractor G. Features; G c D represents a state classifier; s and D t Represent the source domain and the target domain, respectively; This represents the i-th sample in the target domain;

[0032] A gradient inversion layer is added before the domain adversarial discriminator D to optimize the network parameters of all modules in an end-to-end manner.

[0033] Furthermore, in step 4, the category label y obtained from the input sample is... s Using the principle of minimum entropy, the information entropy H(p) is constructed, which is defined as:

[0034]

[0035] Where C' represents all possible labels; p c It is the probability of the prediction result corresponding to the label, and then the entropy value of the target domain sample is calculated;

[0036] Based on the analysis of prediction uncertainty and domain similarity, the weighted calculation of the target domain is expressed as follows:

[0037] ω(H(p))=1+e -H(p) (5)

[0038] According to the definition of the entropy-aware weighting function shown in equation (5), in the modified conditional adversarial loss equation (6), samples that are difficult to transfer are reweighted with lower weights:

[0039]

[0040] The final optimization process for the two loss functions is as follows:

[0041] L(θ f ,θ c ,θ d ) = L c (θ f ,θ c )-λ CDAN L CDAN (θ f ,θ d (7) Where λ CDAN These are weight parameters.

[0042] Further, step 5 involves inputting the target domain sample data into the feature extractor G to extract the feature distribution z of the target domain samples, and then further calculating the sample weights using the state classifier C:

[0043]

[0044] Where, ω t ω represents the target domain weights output by the state classifier C, and ω0 represents the empirical weight values ​​obtained from multiple experiments. During the testing process, if ω t If ω < 0, then the input sample is classified as label 0; if ω t If ≥ω0, then the input sample belongs to label 1.

[0045] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0046] 1. This invention introduces a channel self-attention mechanism layer to optimize the construction of the degenerate feature set, and uses one-dimensional convolution to efficiently realize local cross-channel interaction, thereby improving the correlation between degenerate features and original data.

[0047] 2. This invention constructs an IGBT state predictor based on BiLSTM, which can effectively utilize the forward and backward feature information of the input, reduce the dependence of the prediction model on manual features, and improve the robustness of the model.

[0048] 3. This invention constructs a conditional domain adversarial transfer model, which promotes feature distribution matching between the two domains by quantifying the similarity between the target domain samples and the source domain samples, so that the label classifier trained using the source domain sample data can be well applied to target domain discrimination.

[0049] 4. The adversarial transfer state prediction network proposed in this invention utilizes the adversarial domain adaptation algorithm in deep transfer learning to mine deep information from existing data, providing a feasible solution for end-to-end state monitoring and prediction of unknown IGBT modules.

[0050] 5. The state prediction network model constructed in this invention fully considers the actual situation of industrial applications. It can be trained using existing IGBT module aging data and transferred to state prediction instances of different IGBT modules in industrial sites, greatly improving the diversity of state prediction and providing important information for evaluating the comprehensive performance of equipment. Attached Figure Description

[0051] Figure 1 This is a flowchart of the state prediction process in an embodiment;

[0052] Figure 2 This is a schematic diagram of the feature extractor G in the embodiment;

[0053] Figure 3 This is a schematic diagram of the time-series predictor P and the state classifier C in the embodiment;

[0054] Figure 4 This is a schematic diagram illustrating the principle of conditional domain adversarial migration state prediction in an embodiment. Detailed Implementation

[0055] To make the technical solutions and objectives of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific implementation steps described herein are only used to better illustrate the application of the present invention, but the technical features involved in the embodiments of the present invention are not limited thereto.

[0056] like Figure 1 As shown, the IGBT module state prediction method based on domain adversarial long short-term memory networks provided in this embodiment includes the following steps:

[0057] Step 1: In some embodiments of the present invention, aging data (collector-emitter voltage, collector current, junction-to-case thermal resistance, and junction temperature) of different IGBT modules are collected. During the aging process, the performance characteristic parameters of the IGBT are constantly changing. When the performance characteristic parameter values ​​reach the failure threshold, the IGBT is considered to have failed.

[0058] In some embodiments of the present invention, a source domain data set is constructed. A source domain sample dataset of IGBT modules is known. and, where D s Represents the source domain. X is the i-th sample. s It is the set of all samples. Y is the i-th label of the i-th sample. s It is the set of all distinct labels, and n s This represents the total number of source samples.

[0059] In some embodiments of the present invention, a target domain dataset is constructed. This includes a target domain sample dataset of unknown IGBT modules. Where D t Indicates the target domain. Let X represent the i-th sample. t Let n represent the set of all samples. t This represents the total number of target samples.

[0060] In some embodiments of the present invention, in step 1, the source domain data consists of labeled samples from known devices, and the target domain data consists of unlabeled samples from unknown devices. The source and target domain data follow probability distributions P and Q, respectively, but the data distributions of the two domains are different (P ≠ Q); however, the category labels of the two domains are the same (C). s =C t Because the data distributions in the two domains are different, traditional domain adaptation methods are prone to the problem of domain gap, that is, the difference in degenerate feature information can easily cause negative transfer problems.

[0061] Step 2: The Adversarial Transfer Condition Prediction Network (ATCPN) is used to simultaneously identify the degradation characteristics of unknown IGBTs. It includes a feature extractor G, a temporal predictor P, a state classifier C, and a neighborhood discriminator D. The feature extractor G uses a one-dimensional convolutional neural network structure, with added batch normalization and max-pooling layers. The temporal predictor P consists of multiple long short-term memory network units, capable of acquiring full-cycle aging data. Both the state classifier C and the neighborhood discriminator D use fully connected layer structures, with the output layer of the state classifier C using the ReLU activation function and the output layer of the neighborhood discriminator D using the Sigmoid activation function.

[0062] In some embodiments of the present invention, taking into account the characteristic that the input data is a one-dimensional signal, the structure of the constructed feature extractor G includes multiple feature extraction units, which are stacked sequentially. Each feature extraction unit includes a convolutional layer with a one-dimensional convolutional kernel, a batch normalization layer, a max pooling layer with a one-dimensional pooling kernel, and a channel self-attention mechanism layer.

[0063] like Figure 2As shown, in this embodiment, the feature extractor G has two feature extraction units. The input signal dimension is 64×1×1024. The first feature extraction unit has a convolution kernel size of 3×1, 16 filters, and a max pooling layer size of 2×2. The second feature extraction unit has a convolution kernel size of 3×16, 32 filters, and a max pooling layer size of 3×3. The channel self-attention mechanism layer uses the ECA (Efficient Channel Attention) model, which is added after each feature extraction unit to enhance the correlation between features and data. The ECA model avoids the impact of dimensionality reduction on the data by introducing one-dimensional convolution. The convolution kernel size is set to 3, and the number of filters is set to 1. The output signal dimension of the feature extractor G is 64×25×32.

[0064] The time-series predictor P consists of a forward long short-term memory network (F_LSTM) and a backward long short-term memory network (B_LSTM), which have the same structural parameters; the forward long short-term memory network uses information from past time points to output the prediction result. Reverse Long Short-Term Memory Network

[0065] In some embodiments of the present invention, the timing predictor P consists of a forward long short-term memory network (F_LSTM) and a backward long short-term memory network (B_LSTM). See [link to relevant documentation]. Figure 3 F_LSTM uses information from past time points to output prediction results. B_LSTM uses information from future moments to output prediction results. Bidirectional Long Short-Term Memory (BiLSTM) networks fuse forward and backward predicted features to obtain the final output. Its calculation expression is:

[0066]

[0067] Where w and v represent the F_LSTM output prediction results in the BiLSTM model, respectively. And B_LSTM output prediction results The weights are determined by the model structure. To reduce model complexity, F_LSTM and B_LSTM are given the same structural parameters, meaning the output feature weights of both F_LSTM and B_LSTM are 0.5. Additionally, a Dropout layer is added to the unidirectional LSTM network to eliminate overfitting. The output signal dimension of the time series predictor P is 64×25×128.

[0068] In some embodiments of the present invention, the output of the time series predictor P is flattened to obtain a signal dimension of 64×3200, which is then input to a state classifier C with a fully connected layer structure, such as... Figure 3As shown, its output layer uses a ReLU output function, and its output labels are {0, 1}.

[0069] In some embodiments of the present invention, the domain discriminator D also has a fully connected layer structure, and the output layer adopts the Sigmoid output function.

[0070] Step 3: Using the source domain dataset D constructed in Step 1 s The parameters of the feature extractor G and the state classifier C are optimized and trained using traditional supervised learning algorithms and gradient descent. The source and target domain data follow probability distributions P and Q, respectively. Now, it is necessary to classify the unlabeled samples x in the target domain:

[0071]

[0072] Where β(x) represents the classifier; This represents the prediction result. Therefore, the goal of transfer learning is to minimize the target risk ε using the source data. t (β):

[0073]

[0074] Where (x, y) represents the sample set in the target domain that follows a Q-distribution; y represents the actual label value. In some embodiments of the present invention, the parameters of the feature extractor G and the state classifier C are used to construct the classification loss function:

[0075]

[0076] Where, θ f θ represents the learnable parameters of the feature extractor G. c D represents the learnable parameters of the state classifier C; s Represents the source domain. Let X represent the i-th sample in the source domain. s Represents the set of all samples in the source domain. Y represents the i-th label of the i-th sample. s Let n represent the set of all labels. s E represents the total number of source samples; C' represents the expected value; and c represents the actual label value. This represents the output label of the state classifier C; This represents a probability function.

[0077] It should be noted that this step differs from other existing methods in that the network is first trained with a large learning rate (set to 0.001) for supervised training. After a certain classification accuracy is achieved, the network is fine-tuned with a small learning rate (set to 0.0001), which can accelerate the convergence of the network.

[0078] In some embodiments of the present invention, Conditional Domain Adversarial Networks (CDANs) are used to achieve knowledge transfer. CDANs first require defining multiple linear mappings. That is, the outer product of multiple random vectors. Given two random vectors x and y, the mean map... It can fully capture the complex multimodal structure within the data. Furthermore, cross-covariance... (φ(·) represents the mapping of the sample in the Hilbert space) can successfully model the joint distribution P(x,y). Its loss function is:

[0079]

[0080] Among them, L CDAN (θ f ,θ d ) represents the conditional adversarial loss function; θ f θ represents the learnable parameters of the feature extractor G. d G represents the learnable parameters of the domain discriminator D; d Represents the domain discriminator D; The source domain samples extracted by feature extractor G. Features; G c Represents the state classifier C; D s and D t Represent the source domain and the target domain, respectively; This represents the i-th sample in the target domain.

[0081] This method adds a gradient inversion layer (GRL layer) before the domain adversarial discriminator D to optimize the network parameters of all modules in an "end-to-end" manner.

[0082] Step 4: In the domain adaptation phase, the trained state classifier C is used to classify the target domain data, and the adversarial domain discriminator D is used to classify the source domain data x. s and target domain data x t To make a judgment, the target domain weight function w is obtained during the training phase. t Weighted training is performed using the obtained weights.

[0083] In some embodiments of the present invention, step 4 provides a computational expression for the reverse weighted training of the network. Entropy is a measure of uncertainty; the smaller the entropy value, the higher the reliability of its prediction. Since the source domain samples are labeled, while the target domain data is unlabeled, for the state classifier C, the source domain labels are input during training. Therefore, the state classifier C predicts the source domain samples x... s The prediction has higher reliability compared to the target domain samples, and therefore lower entropy; based on the category label y obtained from the input samples... s Using the principle of minimum entropy, information entropy is constructed, which is defined as: Where p c This represents the probability of the predicted result corresponding to the label. Then, the entropy value of the target domain sample is calculated:

[0084] Therefore, based on the analysis of prediction uncertainty and domain similarity, the weighted calculation of the target domain can be expressed as:

[0085] ω(H(p))=1+e -H(p) (6)

[0086] According to the definition of the entropy-aware weighting function shown in equation (6), in the modified conditional adversarial loss equation (7), those samples that are difficult to transfer are reweighted with lower weights.

[0087]

[0088] Therefore, the optimization process of the two loss functions mentioned above can be summarized as follows:

[0089] L(θ f ,θ c ,θ d ) = L c (θ f ,θ c )-λ CDAN L CDAN (θ f ,θ d (8)

[0090] Where λ CDAN These are weight parameters.

[0091] Step 5: During the testing phase, input the test sample, obtain the corresponding target domain weight and compare it with the set threshold to determine whether the sample belongs to a known fault or a new fault.

[0092] In some embodiments of the present invention, step 5 involves that the model structures input from the source domain and the target domain are identical, i.e., the model structures are shared. (See also...) Figure 4The target domain sample data is input into the feature extractor G to extract the feature distribution z of the target domain samples, and then the sample weights are calculated by the state classifier C.

[0093]

[0094] Where, ω t Let ω represent the target domain weights output by the state classifier C, and ω0 represent the empirical weight values ​​obtained from multiple experiments. During the testing process, if ω t If ω < 0, then the input sample is classified as label 0; if ω t If ≥ω0, then the input sample belongs to label 1.

[0095] The invention will be further described with reference to the accompanying drawings and experimental examples. To evaluate the performance of the proposed method, a publicly available IGBT aging dataset was used. An IGBT aging dataset from four devices at the University of Padua, Italy, was selected to test the proposed prediction network. The experiment controlled the packaging temperature outside the device's rated temperature to accelerate device aging. Several parameters were monitored, such as collector current, collector voltage, gate voltage, and packaging temperature. Performance characteristic parameters reflecting the IGBT's aging state are called aging parameters, and the failure thresholds for these aging parameters are shown in Table 1.

[0096] Table 1 Failure Thresholds for Aging Parameters

[0097]

[0098] Among them, V ce This represents the collector-emitter voltage; a 5% increase in its value from the initial value can be used as a criterion for IGBT failure. c R and T j and represent collector current, junction-to-case thermal resistance, and junction temperature, respectively. A 20% increase in their values ​​compared to their initial values ​​can also be used as a criterion for IGBT failure.

[0099] During IGBT operation, the junction-to-case thermal resistance, junction temperature, saturation voltage drop, and collector current are not only affected by the aging of the IGBT itself, but also by the interactions between these parameters, accelerating the aging process of the IGBT. ce Not only is it easy to measure, but its overall trend is also constantly increasing. Therefore, V is selected. ce As an aging parameter for predicting the remaining lifespan of an IGBT, an IGBT is considered to have failed when the predicted saturation voltage drop exceeds 1.05 times the original value. Therefore, from V... ce A data segment containing at least one cycle length is extracted from the signal to form a sample, which is then labeled.

[0100] The dataset includes three different IGBT modules (Device2, Device3, and Device5), which were subjected to the same stimulus but exhibited different aging processes. In most cases, a large amount of monitoring data can be collected, but not all IGBTs have corresponding label information. Therefore, three transfer learning tasks can be set up, and the corresponding source and target domains are constructed as shown in Table 2.

[0101] Table 2 Dataset Partitioning

[0102]

[0103] Taking migration task 0 as an example, the V of Device2 ce For the source domain dataset, the V of the Device3 device ce The target domain dataset is used. Known device state prediction involves three steps: data input, feature extraction, and state prediction. In this invention, the original vibration signal is used as input, requiring no human intervention. The operational state assessment of new devices aims to match the degradation information of the new devices. The goal is to accurately predict the operational state of the new devices. The main difficulty of this task is that, since the state of the new devices is not present in the source domain, the voltage data of the new devices has not been trained. Therefore, a state classifier C is trained to identify some target samples as new samples.

[0104] The feature extractor G consists of an input layer, a convolutional layer, a pooling layer, and a self-attention mechanism layer. The input layer uses a time window function to compress the input voltage signal. or To improve model training speed, convolutional and pooling layers extract degradation features from voltage data; a self-attention mechanism layer addresses long dependencies in sequences, improving prediction accuracy. The time-series predictor P consists of a two-layer BiLSTM with independent unit states, enabling more sophisticated memory management learning. The state classifier C consists of a two-layer linear layer, ultimately outputting device operating status indicators. Parameters for each module are shown in Table 3.

[0105] Table 3 ATCPN Model Structure

[0106]

[0107]

[0108] To verify the superiority of the proposed method, several classic transfer algorithms were compared on three transfer tasks, including AdaBN, MMD, JMMD, CORAL, and DDAN.

[0109] To ensure fairness in the experiment, all methods used the same convolutional neural network architecture. To prevent randomness in the experiment, each transfer task was run 10 times, and the average accuracy was taken.

[0110] Table 4 compares the CDAN of the present invention with other methods.

[0111] AdaBN MMD JMMD CORAL DDAN CDAN 0→1 63.27 71.43 73.47 79.59 91.84 91.84 0→2 61.22 63.27 71.55 67.35 73.47 77.55 1→2 12.38 20.49 16.94 24.49 42.02 51.02

[0112] The transfer learning performance on the IGBT dataset did not reach the optimal level. In particular, the prediction accuracy of the AdaBN method was only 45.62%. This is because the changes in parameters such as current and voltage are not strongly time-varying but transient, which leads to a decrease in the prediction accuracy of traditional transfer learning models.

[0113] For the CDAN method, the prediction accuracy of the transfer task "[0]→[1]" is 91.84%, while the prediction accuracy of the transfer tasks "[0]→[2]" and "[1]→[2]" is the highest compared to other transfer learning methods, but it is still only 77.55% and 51.02%, respectively. This is because the working conditions of the task have changed completely, and different devices have different degradation characteristics, which is more demanding than other tasks. The prediction effect of the DDAN transfer task "[0]→[2]" is also not good. The prediction results are not concentrated in a label space, but different prediction results appear in the same set of sample features. However, CDAN effectively utilizes the distribution of the learned invariant features and perfectly identifies the degradation characteristics of the new device in the target domain. At the same time, there is a clear boundary between the samples from the source domain and the samples from the target domain.

[0114] This invention proposes a novel intelligent transfer lifetime prediction network to address the state prediction problem of different IGBT modules. The proposed raw data input method requires less prior knowledge and can effectively complete the transfer task of unlabeled target samples. Although the source domain lacks relevant feature information about the new device, it can still predict the current device's operating state with relatively high accuracy. Results show that the adversarial transfer lifetime prediction network using the CDAN method achieves better performance than other transfer learning methods on different transfer tasks, and CDAN exhibits stronger robustness. Therefore, the proposed ATCPN model holds promise for future applications in predicting the operating state of IGBT modules.

[0115] It should be noted that although the implementation of the present invention has been described in detail with reference to examples, those skilled in the art will readily understand that any modifications, substitutions, and improvements made without departing from the spirit and principles of the invention as set forth in the appended claims should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the state of IGBT modules based on domain-adversarial long short-term memory networks, characterized in that, Includes the following steps: Step 1: Data Acquisition and Sample Division. Aging data of different IGBT modules are collected separately. Different sample sequence labels are extracted from the sample data of known IGBT modules to construct the source domain dataset. , For source domain data, Labels corresponding to the source domain data; obtaining target domain sample datasets with degradation information different from the source domain under unknown devices. , For unlabeled target domain data ; Step 2: Construct an adversarial transfer state prediction network, which includes a feature extractor. G Timing predictor P State classifier C Neighborhood discriminator D ; Considering the characteristic that the input data is a one-dimensional signal, the feature extractor G It includes convolutional layers with one-dimensional kernels, max pooling layers with one-dimensional pooling kernels, and channel attention mechanism layers, as well as a state classifier. C Neighborhood discriminator D Includes fully connected layers; Step 3: Use the source domain dataset constructed in Step 1 Supervised learning algorithms and gradient descent are used to improve the feature extractor. G The parameters, and the timing predictor P The parameters are optimized during training; Step 4: In the adversarial domain adaptation stage, utilize the state classifier. C Classify the target domain data using a neighborhood discriminator. D Discriminating target domain data The transferability is assessed, and reverse weighted training is performed; Step 5: During the testing phase, input the test sample, compare the obtained target domain weight with the set threshold, and predict the corresponding state classification.

2. The IGBT module state prediction method based on domain adversarial long short-term memory network according to claim 1, characterized in that, In step 1, the source domain data consists of labeled samples from known devices, while the target domain data consists of unlabeled samples from unknown devices; the source and target domain data each follow a probability distribution. P and Q However, the data distributions in the two domains are different, that is... However, the category labels of the source domain data and the target domain data are the same, that is... .

3. The IGBT module state prediction method based on domain adversarial long short-term memory network according to claim 1, characterized in that, In step 2, the feature extractor G The structure includes multiple feature extraction units, which are stacked sequentially.

4. The IGBT module state prediction method based on domain adversarial long short-term memory network according to claim 3, characterized in that, Each feature extraction unit includes a convolutional layer with a one-dimensional convolutional kernel, a batch normalization layer, a max pooling layer with a one-dimensional pooling kernel, and a channel self-attention mechanism layer.

5. The IGBT module state prediction method based on domain adversarial long short-term memory network according to claim 1, characterized in that, In step 2, the timing predictor P This includes forward long short-term memory networks and backward long short-term memory networks, both of which have the same structural parameters; forward long short-term memory networks use information from past moments to output prediction results. Inverse Long Short-Term Memory (LSTM) networks use information from future moments to output prediction results. The final output is obtained by fusing the forward and backward prediction results of the Bidirectional Long Short-Term Memory (BiLSTM) network. .

6. The IGBT module state prediction method based on domain adversarial long short-term memory network according to claim 1, characterized in that, In step 2, the state classifier C It has a fully connected layer structure, and its output layer uses the ReLU output function, with output labels of {0, 1}; neighborhood discriminator D It also has a fully connected layer structure, and the output layer uses the Sigmoid output function.

7. The IGBT module state prediction method based on domain adversarial long short-term memory network according to claim 2, characterized in that, In step 3, a supervised learning algorithm and gradient descent method are used to refine the feature extractor. G Parameters and state classifier C The parameters are optimized during training; the source and target domain data follow probability distributions respectively. P and Q It is necessary to process unlabeled samples in the target domain. Categorize: (1) in, Represents the state classifier C; This indicates the prediction result; the purpose of transfer learning is to minimize the target risk using the source data. : (2) in, Indicates that the target domain conforms to Q A sample set with a distribution; Indicates the actual label value; express The probability of occurrence; Feature extractor G Parameters and state classifier C Constructing a classification loss function : (3) in, Feature extractor G Learnable parameters; State classifier C Learnable parameters; Represents the source domain. Represents the source domain. i One sample, Represents the set of all samples in the source domain. Indicates the first i The label of each sample, Represents the set of all labels. Represents the total number of source samples; E represents the expected value. C ’ Indicates all possible labels; Indicates the actual label value; This represents the output label of the state classifier C; This represents a probability function.

8. The IGBT module state prediction method based on domain adversarial long short-term memory network according to claim 1, characterized in that, In step 4, the Conditional Domain Adversarial Network (CDAN) is used to achieve knowledge transfer. CDAN first needs to define multiple linear mappings. , that is, the outer product of multiple random vectors; Given two random vectors and Mean mapping Fully capture the complex multimodal structure and cross-covariance within the data. For joint distribution The model is performed, and its loss function is: (4) in, This represents the mapping of samples in Hilbert space. This represents the conditional adversarial loss function; Feature extractor G Learnable parameters; Neighborhood discriminator D Learnable parameters; Neighborhood discriminator D ; Feature extractor G Extracted source domain samples Features; Represents a state classifier; and Represent the source domain and the target domain, respectively; Indicates the target domain. i One sample; In the neighborhood discriminator D Previously, a gradient inversion layer was added to optimize the network parameters of all modules in an end-to-end manner.

9. The IGBT module state prediction method based on domain adversarial long short-term memory network according to claim 7, characterized in that, In step 4, the category labels obtained from the input samples are... Using the principle of minimum entropy, information entropy is constructed. Its definition is: in, C ’ Indicates all possible labels; It is the probability of the prediction result corresponding to the label, and then the entropy value of the target domain sample is calculated; Based on the analysis of prediction uncertainty and domain similarity, the weighted calculation of the target domain is expressed as follows: (5) According to the definition of the entropy-aware weighting function shown in equation (5), in the modified conditional adversarial loss equation (6), samples that are difficult to transfer are reweighted with lower weights: (6) The final optimization process for formulas (3) and (6) is as follows: (7) in Represents the classification loss function. These are weight parameters. This represents the conditional adversarial loss function; Feature extractor G Learnable parameters; State classifier C Learnable parameters; Neighborhood discriminator D Learnable parameters; Neighborhood discriminator D ; Feature extractor G Extracted source domain samples Features; Represents a state classifier; and Represent the source domain and the target domain, respectively; Indicates the target domain. i One sample; Represents the source domain. i One sample.

10. The IGBT module state prediction method based on domain adversarial long short-term memory networks according to any one of claims 1-9, characterized in that, Step 5 involves inputting the target domain sample data into the feature extractor. G In this process, the feature distribution z of the target domain samples is extracted and further processed by a state classifier. C Perform sample weight calculation: (8) in, State classifier C Output target domain weights, This represents the initial threshold of the weights obtained from multiple experiments. During the testing process, if... If the input sample is identified as label 0; if If the input sample belongs to label 1, then the input sample belongs to label 1.

Images