IGBT life prediction method based on Gaussian process regression and LSTM neural network

Abstract

The invention relates to the technical field of life prediction of electronic components, in particular to an IGBT life prediction method based on Gaussian process regression and an LSTM neural network. Collecting and emitting peak voltage data of an IGBT device are collected, preprocessed and then decomposed into a plurality of IMFs reflecting measurement noise and a residual error reflecting the overall degradation trend; using an LSTM model to train the residual error, and fitting a residual error sequence; training each IMF by using a GPR model, and fitting an IMFs sequence; using the trained LSTM model to predict a residual value; predicting the value of each IMF by using the trained GPR type; synthesizing the predicted residual value and the value of the IMF to obtain a final predicted value of the set emission peak voltage and uncertainty quantization of the predicted value of the set emission peak voltage; and the predicted collector emission peak voltage is used as the input of the next step of prediction, and the prediction is continued until the decommissioning point of the IGBT is reached. According to the method, the advantages of the LSTM neural network and the GPR model are combined, so that excellent voltage degradation evaluation performance is obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic component life prediction, and specifically to an IGBT life prediction method based on Gaussian process regression and LSTM neural network. Background Art

[0002] As demand for new energy applications grows, so too does the capacity requirement for power converters. This increased power density subjects the devices to greater electrical and thermal stresses, posing even more stringent challenges to their operational reliability. Therefore, appropriate protection measures tailored to the device's operational characteristics are needed to improve operational reliability and ensure its adaptability to growing application demands.

[0003] Power converters primarily consist of power devices, capacitors, drivers, printed circuit boards (PCBs), inductors, resistors, connectors, and other components. According to surveys, the most common components that fail in power converters are power devices and capacitors, collectively responsible for more than half of all converter failures, as shown in Figure 1(a). The insulated gate bipolar transistor (IGBT) is one of the core components for power conversion and control in power electronics systems. Due to its low conduction loss, high voltage resistance, and fast switching speed, it has become the most widely used power device in industry, as shown in Figure 1(b). During operation, IGBTs can experience electrical and thermal fatigue failure, which gradually accumulates and causes device aging. As the most commonly used and most prone to failure component in power converters, the reliability of IGBTs is crucial for ensuring the safe operation of power systems.

[0004] To improve the operating life and reliability of IGBTs and avoid converter failures or even dangerous accidents, IGBT condition monitoring is required to obtain information about the current device health status. This information serves as a reference for converter and device maintenance, improving the reliability of the converter and the device as a whole. Currently, there are two main modeling approaches for IGBT condition monitoring: physics of failure (PoF)-based and data-driven. PoF models typically start from the device failure mechanism, taking into account the physical meaning of electrothermal parameters and the device failure mode, and constructing a mechanism model for prediction. However, this approach often requires offline sampling, making it difficult to apply in actual systems. Data-driven modeling is characterized by simple modeling and application. It does not require strong professional knowledge and only requires collecting device operating data to achieve model training.

[0005] Existing literature has used a recurrent neural network (RNN) to build an IGBT lifetime prediction model and explained the impact of network hyperparameters on the model. Other studies have also used an LSTM neural network to predict the lifetime of Si MOSFET devices, verifying the effectiveness of neural network-based modeling over Kalman filtering and particle filtering methods. However, data collected during device operation often contains a large amount of noise. Traditional single-model data-driven modeling methods struggle to discard this noise, resulting in low prediction accuracy and poor model robustness. Furthermore, methods such as Bayesian optimization are required to optimize hyperparameters to determine the optimal parameter combination for the model. Summary of the Invention

[0006] To address the above issues, the present invention aims to provide an IGBT lifetime prediction method based on Gaussian process regression and LSTM neural networks. This method enables reliable uncertainty quantification of the IGBT device's collector-emitter peak voltage data, accurately predicting its degradation trend and providing reference information for device health assessment. The technical solution is as follows:

[0007] Step 1: Data collection and processing:

[0008] Collect the collector-emitter peak voltage V of the IGBT device ce(peak) Data and preprocessing it, including filtering out outliers;

[0009] Step 2: Empirical Mode Decomposition:

[0010] Using the empirical mode decomposition method, the pre-processed transient spike voltage V ce(peak) The data is decomposed into several intrinsic mode functions reflecting the measurement noise and a residual Res reflecting the overall degradation trend;

[0011] Step 3: Model training:

[0012] Select the kernel function of the Gaussian process regression model, set the structure of the long short-term memory neural network model and the Gaussian process regression model, and initialize the parameters;

[0013] Use the long short-term memory neural network model to train the residual Res and fit the residual sequence;

[0014] Use the Gaussian process regression model to train each intrinsic mode function and fit the IMFs sequence;

[0015] Step 4: Model testing:

[0016] Use the trained long short-term memory neural network model to predict the residual value;

[0017] For the noise and error parts, the trained Gaussian process regression model is used to predict the values ​​of each intrinsic mode function;

[0018] Step 5: Result synthesis:

[0019] The residual value predicted by the long short-term memory neural network model and the value of the intrinsic mode function predicted by the Gaussian process regression model are synthesized to obtain the final collected spike voltage V ce(peak) Quantification of predicted values ​​and their uncertainties;

[0020] Step 6: Iterate prediction:

[0021] Using the predicted collector-emitter peak voltage V ce(peak) As input for the next prediction, the prediction continues until the retirement point of the IGBT is reached.

[0022] The beneficial effects of the present invention are:

[0023] The present invention combines the advantages of LSTM neural network and GPR model to obtain excellent voltage degradation assessment performance. First, V ce(peak) The data is decomposed to obtain the residuals reflecting the overall degradation trend and multiple intrinsic mode functions reflecting the measurement noise. Then different models are used to perform data-driven modeling and training on the residuals and intrinsic mode functions. Finally, the V is obtained by combining the evaluation results of multiple models. ce(peak) The model was validated using spike voltage data from a NASA-published IGBT accelerated aging dataset. The results show that the proposed hybrid model, after modal decomposition, achieves a root mean square error of only 0.0185V and a maximum error of 0.2316V compared to a single model. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1(a) shows the failure distribution of converter components.

[0025] Figure 1(b) shows the utilization rate of common power devices.

[0026] Figure 2(a) is a diagram illustrating NASA data - a schematic diagram of the collector-emitter spike voltage.

[0027] Figure 2(b) shows the V ce(peak) Dataset.

[0028] Figure 3 It is the LSTM neural network model structure.

[0029] Figure 4 Framework diagram for predicting the peak voltage of the emitter based on the hybrid data-driven model

[0030] Figure 5Flowchart for estimating emitter spike voltage and device retirement time based on a hybrid data-driven model.

[0031] Figure 6 V ce(peak) Data and EMD decomposition results.

[0032] Figure 7(a) shows the comparison of the prediction effects of GPR models with different kernel functions on IMF1.

[0033] Figure 7(b) shows the comparison of the prediction effects of GPR models with different kernel functions on IMF2.

[0034] Figure 7(c) shows the comparison of the prediction effects of GPR models with different kernel functions on IMF3.

[0035] Figure 7(d) shows the comparison of the prediction effects of GPR models with different kernel functions on IMF4.

[0036] Figure 7(e) shows the comparison of the prediction effects of GPR models with different kernel functions on IMF5.

[0037] Figure 8 is the prediction result of the LSTM model for the residual Res.

[0038] Figure 9(a) is a comparison of the prediction effects of the EMD+LSTM+GPR model.

[0039] Figure 9(b) is a comparison of the prediction effects of the EMD+GPR model.

[0040] Figure 9(c) is a comparison of the prediction effects of the EMD+LSTM model.

[0041] Figure 9(d) is a comparison of the prediction effects of the Solo GPR model.

[0042] Figure 9(e) is a comparison of the prediction effects of the Solo LSTM model. DETAILED DESCRIPTION

[0043] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] 1. IGBT aging characteristics analysis

[0045] For IGBT devices, different failure modes can be reflected by characteristic parameters such as voltage and current. thjc , on-state voltage drop V ce(on) And the transient peak voltage V ce(peak) It is often used in IGBT life prediction research. thjc Mainly reflects the solder aging at the device packaging level, V ce(on) Mainly reflects the breakage and falling of the bonding wire, Vce(peak) This embodiment uses the IGBT accelerated aging data set published by the National Aeronautics and Space Administration (NASA) Center of Excellence for Prediction, with V ce(peak) This data was collected during the power cycling of an IRG4BC30K discrete IGBT device, with a switching frequency and duty cycle of 10 kHz and 40%, respectively. Due to the presence of device parasitics, the IGBT generates a transient spike voltage in addition to the turn-off voltage when it turns off, as shown in Figure 2(a). This voltage gradually decreases as the IGBT performance degrades until latch-up occurs, reflecting the degradation of the IGBT's parasitic parameters and overall performance. Figure 2(b) shows the spike voltage extracted from the NASA dataset.

[0046] 2. Hybrid Data-Driven Model Construction

[0047] To achieve reliable V ce(peak) Degradation prediction requires attention to three points. First, the original V ce(peak) The data exhibits highly nonlinear trends and regenerative phenomena, making it unsuitable for accurate health prediction. Second, learning the correlations of time series is crucial for updating long-term dependencies. Third, the level of uncertainty is a key component that cannot be ignored.

[0048] To address these challenges, the data-driven method proposed in this paper mainly adopts three technologies: EMD (Empirical Mode Decomposition) method to decompose the original V ce(peak) The dataset is decomposed, the LSTM sub-model captures long-term dependencies, and the GPR sub-model generates the uncertainty of each prediction result.

[0049] 2.1 Empirical Mode Decomposition

[0050] EMD is an effective signal processing technique that has been applied to many practical fields, such as ocean waves, rotating machinery, etc., due to its powerful ability to extract low-frequency and high-frequency components from high-dynamic signals. ce(peak) The non-stationary data set can be decomposed into a residual sequence and a series of intrinsic mode functions (IMFs). Both the residual sequence and the intrinsic mode functions are orthogonal basis components. The intrinsic mode functions need to meet the following two criteria: first, for the entire data set, the number of zero crossings differs from the number of extreme values ​​by at most one; second, at any time, the envelope defined by the local extreme values ​​must produce a zero mean. Considering that sampling errors and noise can be regarded as high-frequency signals, and V ce(peak)The global attenuation trend of is a low-frequency signal. The EMD method is used to decompose the original capacity attenuation data into several IMFs and a residual Res. The detailed screening process of the decomposed capacity dataset is described as follows:

[0051] For i=1 to i max Do the following:

[0052] 1) Set up dataset C Vp For the first screening process, the original V ce(peak) The data was selected as C Vp .

[0053] 2) By comparing all adjacent values ​​of fluctuations, in C Vp Search for all local maxima and minima in the data. Here, a local minimum or maximum represents the smallest or largest data point in a local fluctuation scale. Then, use splines to connect these local extremes and construct an upper envelope e respectively. up and a lower envelope e low .

[0054] 3) Calculate the local mean m by the following formula e :

[0055]

[0056] 4) Check and calculate C using the following formula Vp and m e The difference between c :

[0057] d c =C Vp -m e (2)

[0058] 5) According to the difference d c Whether the IMF standards are met to set up the IMF pool. (Judge the difference d c The standard way to determine whether the eigenmode function is satisfied is: for the entire data series, first, the number of zero crossings must differ from the number of extreme values ​​by at most one; second, at any time, the envelope defined by the local extreme values ​​must produce zero mean.) If d c It is proved to be an IMF signal, from C Vp Remove this d c , the corresponding difference r is obtained from formula (3) c The obtained r c It will also be represented as the new C in the next screening process bat :

[0059] r c =C Vp -d c(3)

[0060] 6) Repeat 2)-5) until the difference r is obtained c Becomes a monotonic function. If a predefined number i is reached max , the filtering loop will be terminated.

[0061] After the above process, IMFs carry information related to measurement noise. After obtaining m IMFs and a monotonic residual Res, C Vp This formula can be obtained:

[0062]

[0063] 2.2 Gaussian Process Regression

[0064] The GPR model can be seen as an effective method for regression with Gaussian processes. The probability distribution of GPR can be expressed as follows:

[0065] f(k)~GPR(m(k),K(k,k'))(5)

[0066] Where m(k) and K(k,k') represent the mean function and covariance function, respectively. k is the known training data, and k' is the future data to be predicted.

[0067] In practical applications, there are many kernel functions that can be selected for K(k,k'). ce(peak) In degradation assessment, a suitable kernel function has a significant impact on the prediction performance and therefore needs to be carefully selected.

[0068] A kernel function square exponential (SE) function commonly used by GPR is as follows:

[0069]

[0070] Where, σ SE and l SE is a scale factor that controls the magnitude and dispersion of the covariance.

[0071] Another common kernel function is the Matern function, as shown below:

[0072]

[0073] Where γ is a hyperparameter reflecting smoothness. A widely used Matern covariance is Matern5 / 2 (M5 / 2), which is obtained by fixing the value of γ to 5 / 2. MA is the scale factor used to control the amplitude of the covariance, Γ(γ) is the gamma function, ρ is the length scale parameter, is the Bessel function.

[0074] By summing up some kernels with different length scales and the squared exponent SE of the kernel function, we get a new popular kernel function called rational quadratic (RQ):

[0075]

[0076] Among them, α reflects the relative weight of the size change, σ RQ and l RQ is a hyperparameter that affects the axis scale.

[0077] For a regression process, its output is a function plus an additive noise To model, σ n is the variance of the noise signal.

[0078] Then the prior distribution of the known output y can be expressed as follows:

[0079]

[0080] Among them, I n is the n×n unit matrix;

[0081] Assuming that the future data k' follows a Gaussian distribution similar to the training data k, the total joint prior distribution of the known output y and the predicted output y' will be expressed as follows:

[0082]

[0083] Then, by calculating the conditional distribution P(y'|k,y,k') we can predict the output as follows:

[0084] P(y'|k,y,k')=N(y'|m',cov(y')) (11)

[0085] Where m' is the predicted value of the predicted output y', and cov(y') is the covariance matrix reflecting uncertainty, expressed as follows:

[0086]

[0087] Among them, P(y'|k,y,k') obeys Gaussian distribution.

[0088] In the prediction of radiated spike voltage, the noise signal in the voltage degradation dataset is a high-frequency signal with large uncertainty. The GPR model is used to fit the IMFs, and the uncertainty of the predicted high-frequency values ​​can also be considered through the covariance matrix.

[0089] 2.3 LSTM Neural Network

[0090] To alleviate the exploding and vanishing gradient problems of data-driven models, LSTM blocks are often used to embed three gates into the hidden neurons of RNNs. In this sense, one benefit of the LSTM framework is that key information can be stored or updated by manipulating the introduced gates. Furthermore, LSTM models are able to retain information over long periods of time without experiencing vanishing gradients.

[0091] A typical RNN-based LSTM model can be divided into three gate parts, such as Figure 3 The states of all these gates are represented by x k (the input of k at the current moment) and h k-1 (The output of the previous moment k-1) is determined by an S-shaped unit. The input gate determines whether the LSTM model can receive new state information. The forget gate is responsible for forgetting the previous state s in the hidden layer k-1 The output gate determines which information calculated by the RNN model can be output as h k The detailed process of each gate part can be summarized as follows:

[0092] Step a): For the input gate, update and input gate i k The status is as follows:

[0093]

[0094] Step b): Forget gate part, update the forget gate f k And calculate the state s k as follows:

[0095]

[0096] Step c): For the output gate part, update the output gate o k And calculate the output state h k as follows:

[0097]

[0098] Among them, x k is the input, and the output state is h k is the hidden state, representing the short-term memory capacity, and the subscript s is the cell state, representing the long-term memory capacity; represents element-by-element multiplication, σ and tanh are sigmoid and hyperbolic tangent activation functions respectively; W s 、U s and b s are the input x of the cell state information update process in the input gate respectively k The weight matrix, the output state h at the previous moment k-1The weight matrix and corresponding bias vector of W i 、W f and W o are the input x of the input gate, forget gate and output gate at the current moment respectively k The weight matrix, U i 、U f and U o are the output states h of the input gate, forget gate, and output gate at the previous moment respectively k-1 The weight matrix, b i 、b f and b o are the bias vectors of the input gate, forget gate, and output gate respectively.

[0099] In spike voltage degradation, the long-term dependence is expressed in V ce(peak) The correlation between the current spike voltage value and the historical spike voltage during long-term degradation. The spike voltage degradation dataset generally covers hundreds of voltage sampling cycles, and the degradation information between these cycles is highly correlated. In order to accurately capture the V ce(peak) The decline trend of V requires effective learning of long-term dependencies. ce(peak) To this end, consider using LSTM neural network to fit the residual sequence and extract V ce(peak) Long-term characteristics of degradation.

[0100] 2.4 IGBT Collector-Emitter Peak Voltage Degradation Assessment Model Framework

[0101] The future capacity prediction framework and process based on the LSTM+GPR combined model are respectively Figure 4 and Figure 5 Given in.

[0102] The model framework can be divided into two parts. ce(peak) Prediction, based on current and historical peak voltage vector [C Vp (ti),...,C Vp (t)] is the model input, and GPR and LSTM are used to study the potential mapping of the corresponding IMFs and residuals to predict the output C Vp (t+k). For V ce(peak) Prediction, iteratively perform a recursive prediction process, using the previously predicted voltage value as the next input of the model to further predict the new V ce(peak) value until the end of life is reached.

[0103] For the data-driven model, predictions are made based on historical spike voltage information. The detailed steps of the entire prediction process are as follows:

[0104] Step 1: Data collection and processing: Data preprocessing before any training process. ce(peak) The original data needs to be filtered out of the outliers, and the processed model is processed by EMD to obtain several IMFs and a residual.

[0105] Step 2: Model Training: First, select a kernel function suitable for GPR. Set the structure of the LSTM and GPR models and initialize the parameters. For the decomposed residual sequence Res, train the LSTM model to fit the residual sequence. For the obtained IMFs, train the GPR model to fit each IMF sequence.

[0106] Step 3: Model testing: For Res, use the trained LSTM model to predict the Res value. For the noise and error parts, use the trained GPR model to predict the values ​​of each IMF. Combine the outputs of the two models to get the predicted V ce(peak) and the corresponding uncertainty quantification until the device reaches retirement.

[0107] 3 IGBT life prediction model verification

[0108] The modeling results were verified. First, empirical mode decomposition (EMD) was performed on the spike voltage data. The decomposition quantities were then trained and predicted using different data-driven models. Finally, the different data-driven models were compared to test their predictive performance. All tests were performed in MATLAB 2022b using an AMD Ryzen 7 5800H CPU and an RTX-3060 GPU. For all tests, GPRs were trained using gradient methods to maximize the log-marginal likelihood based on hyperparameter optimization.

[0109] 3.1 Empirical Mode Decomposition Results

[0110] V ce(peak) The intrinsic mode function and residual data after empirical mode decomposition are as follows Figure 6 As shown. Among them, the residual Res trend is relatively gentle, reflecting that V ce(peak) The overall degradation trend of IMF5 is large, reflecting the noise information in the data.

[0111] 3.2 Analysis of IMFs evaluation results under different GPR kernel functions

[0112] For the GPR model, selecting an appropriate kernel function can effectively improve prediction accuracy. For the intrinsic mode function (IMFs), the curve fluctuations are much greater than those of the residual curve (Res). To ensure that the model has a good ability to capture the intrinsic mode function, it is necessary to select a kernel function that can better describe its characteristics. This embodiment selects three kernel functions: RQ, SE, and M5 / 2, which can well describe local characteristics, to achieve the GPR fitting effect for IMFs. The experiment uses the first 70% of the IMFs data as the training set and the last 30% as the test set. Figure 7(a)-Figure 7(e) The prediction effects of Gaussian process regression models with three kernel functions, RQ, SE, and M5 / 2, on IMF1-IMF5 are demonstrated.

[0113] As can be seen, the GPR model, using all three kernel functions, exhibits excellent forecasting performance for IMFs. The GPR model exhibits excellent forecasting performance for the low-volatility IMF3–IMF5 time series, with model output values ​​generally consistent with the true reference values. For the more volatile IMF1 and IMF2 curves, the RQ, SE, and M5 / 2 kernel functions all effectively capture the overall fluctuation trend, with only limited error. Given the minimal difference in forecasting performance between the three kernel functions, the RQ kernel was selected for subsequent comparisons of the output performance of the different models.

[0114] 3.3 Analysis of LSTM prediction residual Res prediction results

[0115] For residual Res data prediction, the output effect of the LSTM model is as follows Figure 8 As shown in the figure, since the Res curve is relatively flat, the output value of the LSTM model is basically consistent with the true reference value, and the mean square error and maximum absolute value error are 0.0034 and 0.005 respectively.

[0116] 3.4 Comparison of the effects of hybrid model and single model

[0117] To highlight the effectiveness of the proposed LSTM+GPR hybrid data-driven model, we compared it with a single GPR model, a single LSTM model, a single GPR+EMD model, and a single LSTM+EMD model. The results are shown in Figures 9(a)-(e), and the evaluation errors for different models are shown in Table 1. Specifically, the first two models used a single GPR model or a single LSTM model to predict the raw spike voltage data. The latter two models applied multiple GPR models and multiple LSTM models, respectively, to process all components obtained after EMD decomposition.

[0118] Table 1 Evaluation errors of different models for the collected spike voltage data

[0119]

[0120] As can be seen, the GPR or LSTM model alone is unable to capture the degradation trend of spike voltage. This is because the raw spike voltage data fluctuates significantly, and the relatively simple GPR or LSTM model cannot capture effective information about voltage degradation during training. Similarly, for the EMD+LSTM method, due to the large fluctuations in the IMF1 component, the LSTM model performs poorly in prediction, resulting in poor overall prediction of the voltage data. However, for the EMD+GPR method, the GPR model's strong fitting capabilities enable it to perform well on the IMFs data, ensuring overall prediction accuracy for the voltage data. For the proposed EMD+GPR+LSTM hybrid model, the LSTM model further improves the prediction accuracy of the residual Res data, resulting in overall better prediction results than other models.

Claims

1. A method for predicting IGBT lifespan based on Gaussian process regression and LSTM neural network, characterized in that: The following steps are involved: Step 1: Data collection and processing: Collect the collector-emitter peak voltage V of the IGBT device ce(peak) Data and preprocessing it, including filtering out outliers; Step 2: Empirical Mode Decomposition: Using the empirical mode decomposition method, the pre-processed transient spike voltage V ce(peak) The data is decomposed into several intrinsic mode functions reflecting the measurement noise and a residual Res reflecting the overall degradation trend; Step 3: Model training: Select the kernel function of the Gaussian process regression model, set the structure of the LSTM neural network model and the Gaussian process regression model, and initialize the parameters; Use the LSTM neural network model to train the residual Res and fit the residual sequence; Use the Gaussian process regression model to train each intrinsic mode function and fit the IMFs sequence; Step 4: Model testing: Use the trained LSTM neural network model to predict the residual value; For the noise and error parts, the trained Gaussian process regression model is used to predict the values ​​of each intrinsic mode function; Step 5: Result synthesis: The residual value predicted by the LSTM neural network model and the value of the intrinsic mode function predicted by the Gaussian process regression model are synthesized to obtain the final collected spike voltage V ce(peak) Quantification of predicted values ​​and their uncertainties; Step 6: Iterate prediction: Using the predicted collector-emitter peak voltage V ce(peak) As input for the next prediction, the prediction continues until the retirement point of the IGBT is reached.

2. The IGBT life prediction method based on Gaussian process regression and LSTM neural network according to claim 1 is characterized in that: The pre-processing in step 1 also includes the collection and emission peak voltage V ce(peak) The data are normalized so that the range is between 0 and 1.

3. The IGBT life prediction method based on Gaussian process regression and LSTM neural network according to claim 1 is characterized in that: The empirical mode decomposition method in step 2 comprises the following steps: Step 2.1: Set up the dataset: For the first screening process, the pre-processed transient spike voltage V ce(peak) Data is selected as data to be decomposed C Vp ; Step 2.2: Search for local extrema: By comparing all the adjacent values ​​of the fluctuation, the data to be decomposed C Vp Search for all local maxima and minima in the vector, connect these local extreme values ​​using splines, and construct an upper envelope e up and a lower envelope e low ; Step 2.3: Calculate the local mean: The local mean m is calculated by the following formula e : Step 2.4: Calculate the difference: Check and calculate the data to be decomposed C by the following formula Vp and the local mean m e The difference between c : d c =C Vp -m e (2) Step 2.5: Check the eigenmode function criteria If the difference d c If the standard of the eigenmode function is met, then the data to be decomposed C Vp Subtract the difference d c , and get the corresponding difference r c , the difference r c It will be used as the new data to be decomposed in the next screening process C bat ; Difference r c The calculation is as follows: r c =C Vp -d c (3) Step 2.6: Repeat the screening: Repeat steps 2.2-2.5 until the difference r is obtained. c becomes a monotonic function, then the difference r c As the final residual Res, m intrinsic mode functions carrying information related to the measurement noise are obtained at the same time; The data to be decomposed C Vp As shown in the following formula: Where, IMF j is the jth eigenmode function.

4. The IGBT life prediction method based on Gaussian process regression and LSTM neural network according to claim 1 is characterized in that: The kernel function of the Gaussian process regression model used in step 3 is a square exponential function as shown below: In the formula, k is the known training data, k' is the future data to be predicted, σ SE and l SE are scale factors, which are used to control the magnitude and dispersion of the covariance respectively.

5. The IGBT life prediction method based on Gaussian process regression and LSTM neural network according to claim 1 is characterized in that: The kernel function of the Gaussian process regression model used in step 3 is the Matern function, as shown below: Among them, γ is a hyperparameter reflecting smoothness; σ MA is the scale factor used to control the amplitude of the covariance, Γ(γ) is the gamma function, ρ is the length scale parameter, is the Bessel function.

6. The IGBT life prediction method based on Gaussian process regression and LSTM neural network according to claim 1 is characterized in that: The kernel function of the Gaussian process regression model used in step 3 is a rational quadratic kernel function, and its expression is: Among them, k is the known training data, k' is the future data to be predicted, α reflects the relative weight of the size change, σ RQ and l RQ is a hyperparameter that affects the axis scale.

7. The IGBT life prediction method based on Gaussian process regression and LSTM neural network according to claim 1 is characterized in that: The Gaussian process regression model used in step 3 is a regression process, whose output is a function plus an additive noise To model, σ n is the variance of the noise signal; Then the prior distribution of the known output y is expressed as follows: Among them, I n is the n×n unit matrix; Assuming that the future data k' follows a Gaussian distribution similar to that of the training data k, the total joint prior distribution of the known output y and the predicted output y' will be expressed as follows: Then, by calculating the conditional distribution P(y'|k,y,k') the predicted output is as follows: P(y'|k,y,k')=N(y'|m',cov(y')) (10) Where m' is the predicted value of the predicted output y', and cov(y') is the covariance matrix reflecting uncertainty, expressed as follows: Among them, P(y'|k,y,k') obeys Gaussian distribution.

8. The IGBT life prediction method based on Gaussian process regression and LSTM neural network according to claim 1 is characterized in that: The LSTM neural network model used in step 3 includes an input gate, a forget gate, and an output gate, and the calculation formulas are: Step a): For the input gate part, update the state information and input gate i k The status is as follows: Step b): Forget gate part, update the forget gate f k And calculate the forget gate state s k as follows: Step c): For the output gate part, update the output gate o k And calculate the output state h k as follows: in, represents element-by-element multiplication, σ and tanh are sigmoid and hyperbolic tangent activation functions respectively; W s 、U s and b s are the input x of the cell state information update process in the input gate respectively k The weight matrix, the output state h at the previous moment k-1 The weight matrix and corresponding bias vector of W i 、W f and W o are the input x of the input gate, forget gate and output gate at the current moment respectively k The weight matrix, U i 、U f and U o are the output states h of the input gate, forget gate, and output gate at the previous moment respectively k-1 The weight matrix, b i 、b f and b o are the bias vectors of the input gate, forget gate, and output gate respectively.

Images