IGBT life prediction method fusing time-frequency domain branch and autoregressive neural network

Abstract

The invention discloses an IGBT life prediction method fusing time-frequency domain branches and an autoregressive neural network, and the method comprises the following steps: S1, obtaining Vce (on) historical data of an IGBT in a continuous thermal cycle test, the historical data comprising a Vce (on) value of an IGBT initial operation state; s2, carrying out preprocessing on the historical data; s3, respectively extracting time domain features and frequency domain features from the preprocessed data, and constructing an autoregressive neural network model fusing the time domain branches and the frequency domain branches; s4, training the autoregressive neural network model by adopting a dynamic teacher mandatory strategy; and S5, reasoning by using the trained autoregressive neural network model to obtain the life prediction of the IGBT. According to the method, the time-frequency domain features are fused, the lightweight GRU network is adopted, a dynamic teacher forced training strategy is innovated, the residual life of the IGBT can be accurately and efficiently predicted at low cost without complex modeling, and the convenience and generalization of life management of a power device are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remaining useful life prediction of Insulated Gate Bipolar Transistor in power electronic system, and particularly relates to an IGBT life prediction method fusing time-frequency domain branch and autoregressive neural network. BACKGROUND

[0002] The reliability of Insulated Gate Bipolar Transistor (IGBT) is closely related to its working conditions. Temperature fluctuations caused by different reasons will cause package aging, and ultimately affect the service life of the device. Due to the characteristics of high power density and high frequency, IGBT is widely used in high-voltage, large-current control and high switching frequency fields. This inevitably makes IGBT need to withstand tens of thousands or even millions of temperature shocks during its entire working stage. The most important failure mechanism of IGBT is the peeling of the suture line, which accounts for about 70% of the failure reasons of IGBT module. During thermal cycling, shear stress is generated at the junction of the suture line and the IGBT chip, causing thermal mechanical fatigue. With the repeated occurrence of thermal cycling, the contact area between the suture line and the chip decreases, the current density increases, and the stress gradually accumulates, eventually causing the suture line to peel off the IGBT chip. At the same time, the on-state resistance of IGBT gradually increases, causing the on-state voltage drop Vce(on) to continue to rise until the IGBT completely fails.

[0003] Before the IGBT completely fails, the device can still continue to work normally for a period of time, but during this period of normal operation, the IGBT often shows some signs of aging failure. Therefore, if the remaining useful life of the device can be accurately predicted in advance before the device completely fails, timely maintenance and replacement can be carried out, thereby effectively avoiding failures. The collector-emitter on-state voltage Vce(on) can accurately reflect the remaining useful life of the IGBT, and due to its easy-to-measure characteristics, Vce(on) can be used as the main electrical parameter to predict the remaining useful life of the IGBT. During the aging process of IGBT, Vce(on) has multiple periodic fluctuation components, such as: stable periodic fluctuations caused by thermal cycles produced by the continuous switching of IGBT; low-frequency oscillations caused by the change in thermal resistance due to the continuous breaking of the bonding wire; after the bonding wire falls off, the mismatch of the thermal expansion coefficient between the chip and the substrate intensifies, causing periodic changes in mechanical stress in the temperature cycle, thereby generating a sub-harmonic component in Vce(on). These frequency domain features are difficult to model and capture based on the time domain IGBT aging model. In the model architecture, designing a module that can directly model the frequency domain features will help improve the prediction accuracy of the module and optimize the generalization performance of the model.

[0004] In the prior art, the invention with the name "IGBT life prediction method and device based on CNN-LSTM model", patent number CN202410972518.5, its characteristics are to select the off peak voltage as the aging characteristic parameter of IGBT, and select the CNN-LSTM hybrid network as the IGBT life prediction network, which can automatically extract features from the collected off peak voltage data sequence, deeply mine parameter features, and thus obtain high confidence IGBT aging state. However, it uses LSTM as a time series prediction neural network, which has a large computational overhead on real-time devices, making it difficult to implement and deploy, and it uses pure time domain processing without modeling the signal in the frequency domain, which may overlook some important periodic features. Although the method based on sliding window and CNN network can extract periodic features within a period of time, it lacks explicit modeling ability for frequency domain information, resulting in incomplete feature space, weak working condition adaptability, and unbalanced calculation and interpretation. The invention with the name "IGBT residual life prediction method", patent number CN114707423A, uses an improved runoff flow direction algorithm optimized extreme learning machine (IFDA-ELM) model to predict IGBT life, but it uses a machine learning approach and does not involve a time series neural network approach, which has low prediction accuracy. The invention with the name "IGBT residual service life prediction method based on GRU network", patent number CN116338405A, its characteristics are to select the on-state saturation voltage drop as the module aging failure characteristic quantity, and to construct an optimal parameter Gated Recurrent Unit (GRU) neural network residual life prediction model through experiments to complete the prediction of the aging failure characteristic quantity. It uses a GRU network, which is a pure time series network prediction method, and does not construct a frequency domain branch, lacking effective fusion of time-frequency domain features.

[0005] In general, the prior art has at least the following defects: 1) limited prediction accuracy: whether it is a physical model-based method or a traditional data-driven method, it is difficult to comprehensively and accurately capture the complex characteristics and laws in the IGBT aging process. The physical model-based method has a large deviation from the actual situation due to insufficient consideration of various complex factors in actual operation; while the traditional machine learning and deep learning methods are not fine enough in processing time-frequency feature fusion, and cannot fully utilize the potential information in the data, resulting in that the prediction accuracy is difficult to meet the actual engineering requirements. For example, in the actual industrial environment, the aging process of IGBT is affected by the coupling of multiple factors, and the existing methods are difficult to accurately describe these complex relationships, resulting in a large error between the predicted remaining useful life and the actual value; 2) high computational complexity: some deep learning models, such as complex structure LSTM network, contain a large number of parameters and complex calculation processes, resulting in a large amount of calculation resources and time required for training and prediction. In actual application, especially in scenarios with high real-time requirements, such as online monitoring and fault warning of power systems, this high computational complexity model cannot quickly give the prediction result, limiting its application range. For example, in a large-scale power system, a large number of IGBT modules need to be monitored and life predicted in real time, and the high computational complexity model is difficult to meet the real-time requirements of the system, which may delay the fault warning and affect the safe and stable operation of the power system; 3) poor adaptability: the existing methods have poor adaptability when facing different operating conditions and environmental conditions. IGBTs in different application scenarios, such as electric vehicles, wind power generation and industrial variable frequency devices, have great differences in operating conditions, including temperature, humidity, load characteristics, etc. The existing model is difficult to adaptively adjust according to different conditions and environments, resulting in weak generalization ability in actual application. For example, IGBTs running in harsh environments of high temperature and high humidity, the existing prediction model cannot accurately adapt to this special environment, and the reliability of the prediction result is greatly reduced. The existing pure time domain-based method is difficult to capture the signal characteristics that change under different conditions; 4) lack of effective feature fusion: IGBT aging data contains rich time and frequency domain information, but the existing methods often fail to fully exploit and fuse these information. Some methods rely only on time domain features for prediction, ignoring the important information about the internal failure mechanism of IGBT contained in the frequency domain features; while other methods that attempt to fuse time and frequency, the fusion method is simple and cannot fully utilize the complementary advantages of time and frequency information, thereby affecting the performance of the prediction model; 5) high measurement cost: the acquisition equipment of the dependent signal features is expensive and not convenient for mass deployment.

[0006] Therefore, those skilled in the art are committed to developing an IGBT life prediction method that fuses time-frequency domain branches and autoregressive neural networks. SUMMARY

[0007] In view of the above-mentioned defects of the prior art, the present invention at least aims to solve the following technical problems: solving the following technical problems in the existing IGBT life prediction methods: limited prediction accuracy (difficult to fully capture the complex characteristics of aging), high computational complexity (such as poor real-time performance of models such as LSTM), insufficient adaptability (weak generalization ability under different working conditions), insufficient fusion of time-frequency features (relying only on the time domain or simple fusion), and expensive and inconvenient deployment of signal feature acquisition equipment.

[0008] To achieve the above object, the present invention provides an IGBT life prediction method that integrates time-frequency domain branching and autoregressive neural network, the method comprising the following steps:

[0009] S1: Acquire historical Vce(on) data of the IGBT in a continuous thermal cycle test, wherein the historical data includes the Vce(on) value of the IGBT in the initial operating state;

[0010] S2: Preprocessing the historical data;

[0011] S3: extracting time domain features and frequency domain features from the preprocessed data, respectively, and constructing an autoregressive neural network model that integrates the time domain branch and the frequency domain branch;

[0012] S4: using a dynamic teacher forcing strategy to train the autoregressive neural network model;

[0013] S5: Obtaining a life prediction of the IGBT by using the trained autoregressive neural network model for inference;

[0014] Furthermore, the preprocessing in step S2 includes performing sliding exponential average filtering and Min-Max normalization processing on the historical data to generate a standardized sequence;

[0015] The sliding exponential average filtering is specifically:

[0016] EMA t =α×x t +(1-α)×EMA t-1 .

[0017] Among them, x t Represents the data value at time t; EMA t represents the exponential average at time t; α is the smoothing coefficient, preferably 0.3, that is, the new data accounts for 30% of the weight and the filtered data at the previous moment accounts for 70% of the weight;

[0018] The Min-Max normalization process maps the data to between [0, 1], expressed as:

[0019]

[0020] Among them, x min and x max are the x t The minimum and maximum values ​​in the data;

[0021] Furthermore, obtaining the time domain features includes: extracting the time series features of the data using a GRU neural network, wherein the GRU neural network is a variant of an LSTM, and reducing training parameters by merging an input gate and a forget gate in the LSTM into an update gate;

[0022] Furthermore, the GRU neural network is used to extract the temporal features of the data through the calculation logic of the reset gate, the update gate and the candidate hidden state, which is expressed as follows:

[0023] r t =σ(W γ ·[h t-1 ,x t ]).

[0024] z t =σ(W z ·[h t-1 ,x t ]).

[0025]

[0026] h t =(1-z t )⊙h t-1 +z t ⊙h′ t .

[0027] Where: x t and h t are the input information of the GRU unit in the GRU neural network at the current moment and the new hidden state to be output; h t-1 is the output information of the previous moment; [h t-1 ,x t ] means concatenating the hidden state of the previous moment with the input of the current moment in the feature dimension; W ht ' and W z is the weight matrix corresponding to the reset gate, candidate hidden state and update gate; σ() is the sigmoid activation function; h t ' is a candidate hidden state; · represents matrix multiplication; ⊙ represents element-by-element multiplication;

[0028] Acquiring the frequency domain features includes:

[0029] The data is subjected to FFT transformation to obtain frequency domain features, and the frequency domain features are normalized by truncation or zero padding and then input into an MLP to output a frequency domain embedding vector; the MLP shares parameters between each time step, thereby omitting reinitialization;

[0030] Further, the fusion of the time domain branch and the frequency domain branch employs a feature fusion function to fuse the end hidden state of the GRU neural network with the MLP output frequency domain embedding vector to output a prediction value, and the feature fusion function is a fully connected layer;

[0031] The fully connected layer fusion representation formula is:

[0032]

[0033] wherein, represents the frequency domain information of the sequence related to the round, and F is a fully connected layer function;

[0034] Further, the step S3 further comprises adding the prediction value to the end of the preprocessed historical sequence as new input data by recursive iteration;

[0035] Further, the input data for training by the dynamic teacher forcing strategy is randomly selected real value at the previous time and the prediction value, and the teacher forcing ratio is a tunable hyperparameter of 0-1;

[0036] Further, the training in step S4 further comprises applying an upward constraint to ensure that the prediction value is monotonically increasing, and the representation formula is:

[0037]

[0038] wherein, is the original prediction value of the model at time t; is the prediction value adjusted after the upward constraint; is the prediction value at the previous time; and the preferred value of ε is 1x10 -3 ;

[0039] Further, the method further comprises testing and evaluating the autoregressive neural network model, and the model performance is evaluated by mean square error, mean absolute error, and determination coefficient index, the mean square error is the mean value of the square of the difference between the prediction value and the real value, the mean absolute error is the mean value of the absolute difference between the prediction value and the real value, and the determination coefficient is used to measure the goodness of fit of the autoregressive neural network model to the data;

[0040] Further, the autoregressive neural network model supports end-side deployment or cloud-side deployment.

[0041] The application provides an IGBT life prediction method fusing a time-frequency domain branch and an autoregressive neural network, the method trains an autoregressive time series prediction model, recursively decodes an electrical parameter, collector-emitter on-state voltage Vce(on) representing the remaining life of the IGBT from high-dimensional time-frequency features, and uses the Vce(on) as a criterion for judging the failure of the IGBT life, when the predicted Vce(on) is higher than an empirical threshold (initial value 20%), it is determined that the IGBT device fails, the application fuses time-frequency domain features, adopts a lightweight GRU network, and innovatively uses a dynamic teacher forced training strategy, so that the remaining life of the IGBT can be decoded from historical data without complex electrical or mechanical modeling of the device, and a more convenient deployment and higher generalization technical means are provided for the life management of the power device. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 It is a schematic diagram of the method architecture of the application;

[0043] Figure 2 It is a schematic diagram of the overall implementation scheme of a preferred embodiment of the application;

[0044] Figure 3 It is a schematic diagram of the autoregressive neural network model structure of a preferred embodiment of the application;

[0045] Figure 4 It is a comparison diagram of the prediction results of the model of the application and other different models;

[0046] Figure 5 It is a visualization schematic diagram of different teacher forced proportions under the dynamic teacher forced strategy of the application. DETAILED DESCRIPTION

[0047] The following reference description of the drawings introduces a plurality of preferred embodiments of the application, so that the technical content thereof is clearer and easier to understand. The application can be embodied in many different forms of embodiments, and the protection scope of the application is not limited to the embodiments mentioned herein.

[0048] In the drawings, the same numbers are used to represent the same components throughout the drawings, and components with similar structures or functions are represented by similar numbers. The size and thickness of each component shown in the drawings are arbitrarily shown, and the size and thickness of each component are not limited in the application. In order to make the drawing clearer, the thickness of some components is appropriately exaggerated in some places in the drawing.

[0049] The application provides an IGBT life prediction method fusing a time-frequency domain branch and an autoregressive neural network, as shown in Figure 1 The method comprises the following steps:

[0050] The method comprises the following steps:

[0051] S1: obtaining Vce(on) historical data of IGBT in continuous thermal cycle test, the historical data including Vce(on) value of initial running state of IGBT;

[0052] S2: preprocessing the historical data;

[0053] S3: extracting time domain features and frequency domain features from the preprocessed data respectively, and constructing an autoregressive neural network model fusing the time domain branch and the frequency domain branch;

[0054] S4: training the autoregressive neural network model by using a dynamic teacher forced strategy;

[0055] S5: inferring and obtaining the life prediction of the IGBT by using the trained autoregressive neural network model.

[0056] One preferred embodiment of the present application combines Figure 2 The overall implementation scheme of the present application includes:

[0057] First step, data preparation phase: Vce(on) historical data of IGBT in continuous thermal cycle test, in this embodiment, representative data set collected by NASA researchers is temporarily used.

[0058] Second step, data preprocessing. First, in order to effectively suppress the influence of random noise, exponential moving average (EMA) method is used for smoothing processing of data. Unlike the sliding average with fixed window, EMA gives different weights to recent data and historical data, so as to effectively maintain the signal trend while realizing dynamic smoothing. The calculation formula of EMA is:

[0059] EMA t = alpha * x t + (1-alpha) * EMA t-1

[0060] Wherein, x t represents the data value at time t; EMA t represents the exponential average value at time t; alpha is the smoothing coefficient, and in this embodiment, alpha is preferably 0.3. In each step of calculation, new data is given a weight of 30%, while the previous filtering result accounts for 70% weight, so as to effectively suppress random noise in data and retain the signal trend.

[0061] Then, the filtered data is normalized, and in this embodiment, the Min-Max normalization method is used to process the smoothed data according to the formula:

[0062]

[0063] Among them, x min and x max are the x t The minimum and maximum values ​​in the data values. Mapping the data to the interval [0,1] makes data of different scales in the same comparable range, which facilitates subsequent model learning and processing.

[0064] The third step is model construction. The structure of the autoregressive neural network model is as follows Figure 3 As shown, specifically including:

[0065] First, define and initialize the autoregressive neural network model. The task of predicting the remaining life of an IGBT is a time series prediction problem. In order to accurately evaluate the health status and remaining life of the device, based on the degradation characterization of the on-state voltage drop Vce(on), this embodiment defines the remaining life prediction as a single variable time series prediction problem: Given a historical sequence {x1, x2, ..., x n}, estimate the on-state voltage drop change trend at the next moment and in a longer time range to assist in device life management and fault prevention. Assume that the degradation data of D IGBT devices in the continuous thermal cycle test are collected, and all degradation curves can be expressed as

[0066]

[0067] in Indicates that the ith IGBT device is in the n∈{1, 2, ..., N i} rounds of measurement, where D is the number of devices.

[0068] 31) Construct a time-domain branch. This branch uses a standard GRU neural network. The GRU neural network is a further optimization of the Long Short-Term Memory Network (LSTM). As a variant of the LSTM, the GRU does not introduce additional parameters. Instead, it combines the input gate and forget gate in the LSTM into a single update gate and also merges the cell state and hidden state. While maintaining accuracy, the GRU model has fewer training parameters and faster convergence than the LSTM.

[0069] The reset gate is achieved by combining the hidden state h of the previous moment t-1 and the current input x t , after the tanh activation function outputs the candidate hidden state h t ', thereby controlling whether the information of the previous step needs to be forgotten.

[0070] The update gate is used to control the degree of information retention, and the z generated by the sigmoid activation functiont The value range is between 0 and 1. When z t is closer to 1, it indicates that the information is retained more; on the contrary, when z t is closer to 0, it indicates that the information is retained less. The GRU network effectively extracts the time domain features according to the calculation logic of the reset gate, the update gate and the candidate hidden state, and the calculation formula is as follows:

[0071] r t =σ(W γ ·[h t-1 ,x t ]).

[0072] z t =σ(W z ·[h t-1 ,x t ]).

[0073]

[0074] h t =(1-zt)⊙h t-1 +z t ⊙h′ t .

[0075] Wherein: x t and h t are the input information and the new hidden state to be output of the GRU unit in the GRU neural network at the current moment; h t-1 is the output information of the last moment; [h t-1 , x t ] represents that the last moment hidden state and the input at the current moment are spliced in the feature dimension; W ht ' and W z are the weight matrices corresponding to the reset gate, the candidate hidden state and the update gate; σ() is a sigmoid activation function; h t ' is the candidate hidden state; · represents matrix multiplication; and represents element-by-element multiplication.

[0076] The GRU neural network comprises a plurality of recurrent GRU units, the input of each GRU unit is the IGBT collector-emitter on-state voltage Vce(on) value predicted by the last GRU unit at the current time step and the hidden state of the last GRU neural network, and the output of each GRU unit is the predicted Vce(on) value. The hidden state is a high-dimensional matrix describing the current signal characteristics in the field of autoregressive time series neural networks.

[0077] 32) Construct a frequency domain branch. The frequency domain branch uses a multi-layer perceptron (MLP) to obtain the embedding representation of the frequency domain features. The frequency domain features are obtained by performing a fast Fourier transform (FFT) on the input sequence, extracting the amplitude and phase information of each frequency component of the signal. Then, the resulting spectral features are standardized to a fixed length by truncation or zero padding for subsequent processing; the standardized frequency domain features are input into the MLP, which automatically learns and extracts key frequency domain features through nonlinear transformation of multiple layers of neurons, i.e., an MLP is introduced to comprehensively analyze the frequency domain features of the entire historical sequence.

[0078] Specifically, the multi-layer perceptron has only one unit, but the multi-layer perceptron is repeatedly used between each time step. The input of the multi-layer perceptron (MLP) is the output of the fast Fourier transform (FFT) of a Vce(on) sequence composed of all Vce(on) sampling points from the initial time step to the current time step, which is a frequency domain vector. After the frequency domain vector is input into the MLP, it will be converted into an extracted frequency domain feature vector (feature vector, also called embedding vector). During the training phase, instead of reinitializing a new MLP at each time step, the MLP that has been parameter-adjusted through backpropagation at previous time steps is directly used for continuous training; during the inference phase, the Vce(on) sequence at each time step is input into the same MLP with adjusted parameters after FFT for frequency domain feature extraction.

[0079] 33) Time domain and frequency domain feature fusioner

[0080] The last hidden state of the GRU unit of the time domain branch is concatenated with the frequency domain feature matrix of the frequency domain branch, fully utilizing the complementary advantages of time domain and frequency domain information. With the fused features, the Vce(on) value at the next time step is predicted, and a recursive iteration method is used to add each predicted value to the end of the historical sequence as new input, realizing long-term prediction of the remaining service life of the IGBT.

[0081] The time domain and frequency domain feature fusioner is a fully connected layer. After the end hidden state of the GRU and the MLP output are fused in the fully connected layer fusion, the conduction voltage drop under the next thermal cycle can be predicted by the feature fusion function The process can be represented as a function F, which is:

[0082]

[0083] Where: represents the frequency domain information related to the sequence at this round. F is a feature fusion function, and in this embodiment, a fully connected layer is used as the feature fusion function.

[0084] Fourth step, neural network model training:

[0085] GUR neural network, if the teacher forced technology is used completely, the input value of each time step GRU unit is the real value of the last step in the training stage, and the output value is compared with the real value of the current time step to calculate the error, and then the error is propagated backward. In the inference stage, the features are as described above in 31).

[0086] The difference between this embodiment and the completely teacher forced training is that the neural network model using dynamic teacher forced technology randomly selects the real value of the last time step as the input of the current time step, or uses the predicted output of the last GRU unit as the input of the current time step GRU unit during training. Then, the output value of the current GRU unit is compared with the real value of the current time step to calculate the error, and then the error is propagated backward. This strategy provides accurate context information for the model, speeds up the convergence of the model, effectively avoids the error accumulation problem caused by relying on the last step prediction result in the traditional autoregressive model, and significantly improves the training efficiency and prediction accuracy of the model.

[0087] This embodiment quantifies an indicator, defined as the teacher forced ratio, which is a hyperparameter that can be adjusted during neural network training. The teacher forced ratio measures the tendency to use the real value of the last time step as the input of the current time step, or to use the predicted output of the last GRU unit as the input of the current time step GRU unit during training. The teacher forced ratio is a floating-point number between 0 and 1. When it is 0, the neural network is in the inference state, so the hyperparameter will not be set to 0 during training. When it is 1, the neural network is in the completely teacher forced state. During inference, the teacher forced logic can be turned off or the teacher forced ratio can be set to 1, because during inference, the real value is unknown, and the neural network can only use its own output value of the last time step for recursive prediction. As shown in FIG. 8, the visualization results of different teacher forced ratios of the input are shown. Figure 5

[0088] For the frequency domain branch MLP, in the training stage, similar to the GRU unit of the time domain branch, in the completely teacher forced state, for each time step, the input of the MLP is the FFT result of the historical real Vce(on) sequence. In the dynamic teacher forced state, for each time step, the input of the MLP is converted to the FFT result of the sequence composed of the historical all time steps, and the Vce(on) predicted by the neural network.

[0089] For the frequency domain branch MLP, in the inference stage, since there is no real value available, the input of the MLP is the FFT result of the sequence composed of the historical all time steps, and the known Vce(on) and the Vce(on) predicted by the neural network.

[0090] ​Specifically, the hyperparameters of the neural network model were set. The initial learning rate was set to 0.001, which ensures rapid model convergence while avoiding instability caused by excessively large learning rates. The learning rate decay factor was set to 0.5. When the model's training performance did not improve within a certain number of epochs (the learning rate patience value was set to 5), the learning rate was halved, which facilitated more precise parameter adjustments in the later stages of training. The number of training epochs was set to 140, allowing the model to fully learn the data features through multiple iterations. The batch size was set to 64, achieving a good balance between computing resources and training efficiency. The number of GRU hidden units was set to 128 to ensure that the GRU network had sufficient capacity to capture time-domain features. The frequency-domain feature dimension was set to 16, rationally selecting the number of frequency-domain features. The number of hidden nodes in the fully connected layer was set to 32 to optimize feature fusion and prediction. During training, a dynamic teacher forcing strategy was adopted, dividing 50% of the known historical data into a training set for training the model, enabling the model to quickly learn the patterns and features in the data.

[0091] In a preferred embodiment, since the life of the IGBT should change monotonically over time, this embodiment adopts an upward constraint method to always maintain a strict upward trend for Vce(on), which is expressed as follows:

[0092]

[0093] in, is the original prediction value of the model at time t; is the forecast value adjusted after the upward constraint; is the predicted value at the previous moment; the preferred value of ε is 1×10 -3 .

[0094] Furthermore, the ADAM optimizer, dynamic teacher forcing strategy and back-propagation algorithm are used for neural network training.

[0095] Step 5: Model testing and evaluation: Use the trained model to predict the remaining test data and calculate the mean square error (MSE), mean absolute error (MAE) and coefficient of determination (R 2 Score) and other indicators to evaluate model performance, such as Figure 4 As shown in FIG, by comparing the curves (such as the prediction trends of different network models such as GRU and LSTM), the advantages of the model of the present invention in fitting the Vce(on) degradation curve can be more intuitively demonstrated.

[0096] MSE amplifies the impact of large errors by calculating the mean of the squares of the differences between the predicted values ​​and the true values, and more sensitively reflects the model prediction deviation; MAE directly calculates the mean of the absolute differences between the predicted values ​​and the true values, which can intuitively reflect the average prediction error of the model and is relatively insensitive to outliers; R2 Score is used to measure the goodness of fit of the model to the data. The closer the value is to 1, the better the model fit is.

[0097] Step 6: Model Deployment: The model supports both on-device and cloud deployment. On-device deployment involves directly deploying the model on the control device (such as an MCU or FPGA) of the IGBT acquisition device. Cloud deployment involves deploying the model on a server and remotely acquiring the IGBT voltage values ​​collected by the on-device sensors.

[0098] Specifically, after the network is deployed and before the inference phase, there is a segment of Vce(on) signals that have been collected. The network's purpose is to infer subsequent Vce(on) signals from this segment of Vce(on) signals in order to infer when the Vce(on) signal will reach the IGBT's aging peak.

[0099] 61) After the network is deployed, the collected Vce(on) signal will not be used for parameter tuning. However, for the time domain branch, the GRU unit will use the collected Vce(on) signal sequence to obtain the hidden state of the last time step of this signal in order to start reasoning later.

[0100] 62) After the network is deployed, for the frequency domain branch, since there is no hidden state, there is no need to perform the operation 61), but the collected Vce(on) signal will be input into the FFT together with the Vce(on) signal recursively predicted by the subsequent neural network to obtain the frequency domain vector, which is further input into the MLP for processing.

[0101] Step 7: Model Reasoning:

[0102] The trained neural network is used to predict the remaining life of an IGBT device that has been in operation for a period of time, that is, whose Vce(on) has been sampled for a period of time. By predicting the number of power cycles that can be continued to be tolerated after its Vce(on) reaches the threshold (generally exceeding the initial value by 20%), the remaining life of the IGBT device is extracted.

[0103] Specifically, the Vce(on) threshold value is 20% higher than the initial value, the model selects the forward propagation mode, and according to the current collected Vce(on) sequence, the subsequent Vce(on) sequence is predicted, when any point of the subsequent Vce(on) sequence exceeds the threshold value, the model is warned, and the IGBT has the risk of aging damage. In the actual application scene, for example, the device operation and maintenance management of the power system, the maintenance and replacement plan of the IGBT can be planned in advance according to the prediction result of the model. When the prediction shows that the remaining service life of a certain IGBT module is close to the warning threshold value, the maintenance or replacement can be arranged in time by the operation and maintenance personnel, so as to avoid the power system failure caused by the unexpected failure of the IGBT, and ensure the stable and reliable operation of the power system.

[0104] The method provided by the application comprehensively uses a gated recurrent unit (GRU), time-frequency fusion and a deep learning algorithm, and has the following advantages:

[0105] 1) The on-state saturation voltage drop can be obtained by steady-state measurement, and the cost and real-time performance of the steady-state measurement are better than those of the transient-state measurement;

[0106] 2) The GRU network used has the advantage of faster inference speed in real-time deployment;

[0107] 3) The frequency domain features are directly modeled and combined with time sequence features, and have better adaptability and prediction accuracy;

[0108] 4) The teacher forced neural network training method is introduced during training, so as to realize relieving the cumulative training error during inference of the autoregressive type time sequence neural network;

[0109] 5) The deep learning is used to improve the prediction accuracy, and the lightweight time sequence neural network and the fast Fourier transform (FFT) algorithm do not significantly reduce the inference speed, so that real-time deployment can be realized.

[0110] Therefore, the application can realize accurate prediction of the remaining service life of the IGBT, provide core technical support for reliability guarantee and operation and maintenance management of the power system, and has wide application prospects in many fields such as industrial automation, smart grid and new energy power generation.

[0111] The above detailed the preferred embodiments of the application. It should be understood that those skilled in the art can make many modifications and changes without creative labor according to the concept of the application. Therefore, any technical solution obtained by logical analysis, reasoning or limited experiment on the basis of the prior art according to the concept of the application shall be within the protection scope determined by the claims.

Claims

1. A method for predicting IGBT lifespan by integrating time-frequency domain branching with autoregressive neural network, characterized in that: The method comprises the following steps: S1: Acquire historical Vce(on) data of the IGBT in a continuous thermal cycle test, wherein the historical data includes the Vce(on) value of the IGBT in the initial operating state; S2: Preprocessing the historical data; S3: extracting time domain features and frequency domain features from the preprocessed data, respectively, and constructing an autoregressive neural network model that integrates the time domain branch and the frequency domain branch; S4: using a dynamic teacher forcing strategy to train the autoregressive neural network model; S5: Utilize the trained autoregressive neural network model to infer and obtain the life prediction of the IGBT.

2. The IGBT life prediction method integrating time-frequency domain branching and autoregressive neural network according to claim 1 is characterized in that: The preprocessing in step S2 includes performing sliding exponential average filtering and Min-Max normalization processing on the historical data to generate a standardized sequence; The sliding exponential average filtering is specifically: EMA t =α×x t +(1-a)×EMA t-1 . Among them, x t Represents the data value at time t; EMA t represents the exponential average at time t; α is the smoothing coefficient, preferably 0.3, which means that the new data accounts for 30% of the weight and the filtered data at the previous moment accounts for 70% of the weight; The Min-Max normalization process maps the data to between [0, 1], expressed as: Among them, x min and x max are the x t The minimum and maximum values ​​in the data.

3. The IGBT life prediction method integrating time-frequency domain branching and autoregressive neural network according to claim 1 is characterized in that: Acquiring the time domain features includes: extracting the time series features of the data using a GRU neural network, where the GRU neural network is a LSTM variant, and reducing training parameters by merging the input gate and the forget gate in the LSTM into an update gate.

4. The IGBT life prediction method integrating time-frequency domain branching and autoregressive neural network according to claim 3 is characterized in that: The GRU neural network is used to extract the temporal features of the data through the calculation logic of the reset gate, the update gate and the candidate hidden state, which is expressed as follows: r t =σ(W γ ·[h t-1 ,x t ]). z t =σ(W z ·[h t-1 ,x t ]). h t =(1-z t )☉h t-1 +z t ☉h t . Where: x t and h t are the input information of the GRU unit in the GRU neural network at the current moment and the new hidden state to be output; h t-1 is the output information of the previous moment; [h t-1 ,x t ] means concatenating the hidden state of the previous moment with the input of the current moment in the feature dimension; W γ 、W ht ' and W z is the weight matrix corresponding to the reset gate, candidate hidden state and update gate; σ() is the sigmoid activation function; h t ' is a candidate hidden state; · represents matrix multiplication; ⊙ represents element-by-element multiplication; Acquiring the frequency domain features includes: The data is subjected to an FFT transformation to obtain frequency domain features, which are then normalized by truncation or zero padding and then input into an MLP to output a frequency domain embedding vector; the MLP shares parameters between each time step, thereby omitting reinitialization.

5. The IGBT life prediction method integrating time-frequency domain branching and autoregressive neural network according to claim 3 is characterized in that: The fusion of the time domain branch and the frequency domain branch uses a feature fusion function to fuse the terminal hidden state of the GRU neural network with the MLP output frequency domain embedding vector to output a predicted value, and the feature fusion function is a fully connected layer; The fully connected layer fusion expression formula is: in, Represents the frequency domain information related to the sequence in this round, and F is the fully connected layer function.

6. The IGBT life prediction method integrating time-frequency domain branching and autoregressive neural network according to claim 4 is characterized in that: The step S3 further includes adding the predicted value to the end of the preprocessed historical sequence as new input data in a recursive iterative manner.

7. The IGBT life prediction method integrating time-frequency domain branching and autoregressive neural network according to claim 4 is characterized in that: The input data for training using the dynamic teacher forcing strategy is a randomly selected true value at the previous moment and the predicted value, and the teacher forcing ratio is an adjustable hyperparameter of 0-1.

8. The IGBT life prediction method integrating time-frequency domain branching and autoregressive neural network according to claim 1 is characterized in that: The training in step S4 also includes applying an ascending constraint to ensure that the predicted value increases monotonically, which is expressed as: in, is the original prediction value of the model at time t; is the forecast value adjusted after the upward constraint; is the predicted value at the previous moment; the preferred value of ε is 1×10 -3 .

9. The IGBT life prediction method integrating time-frequency domain branching and autoregressive neural network according to claim 1 is characterized in that: The method also includes testing and evaluating the autoregressive neural network model, and evaluating the model performance through mean square error, mean absolute error and determination coefficient indicators, wherein the mean square error is the mean of the squares of the differences between the predicted value and the true value, the mean absolute error is the mean of the absolute differences between the predicted value and the true value, and the determination coefficient is used to measure the goodness of fit of the autoregressive neural network model to the data.

10. The IGBT life prediction method integrating time-frequency domain branching and autoregressive neural network according to claim 1 is characterized in that: The autoregressive neural network model supports device-side deployment or cloud deployment.

Images