A method for evaluating inverter switching performance degradation

Abstract

The present invention proposes a method for evaluating the performance degradation of inverter switches, which belongs to the field of inverter switches. The raw data of six switching tubes are collected; the raw data are preprocessed, and a degradation index sequence that expresses the degree of degradation of a single switching tube is constructed; a CNN-LSTM model is constructed and trained to obtain a prediction model corresponding to each switching tube; the data of the switching tube to be tested is input into the prediction model to obtain a degradation index of the switching tube group; the degradation indexes are fused to obtain a unified degradation evaluation index. The present invention utilizes the high reliability of the CNN-LSTM model in the field of time series prediction, combined with the preprocessing of the pre-parameters, to greatly improve the accuracy of the switch tube aging prediction, which can improve the operational safety of the electric drive system and the cost-effectiveness and feasibility of the predictive maintenance guarantee plan, bringing better market competitiveness.

Description

Technical Field

[0001] The present invention relates to the field of inverter switches, and in particular to a method for evaluating performance degradation of inverter switches. Background Art

[0002] As a key component of the electric drive system, the inverter performs energy conversion and output, converting DC power into constant-frequency, constant-voltage, or frequency- and voltage-regulated AC power, used to drive motors for motion control or energy conversion. Under the influence of multi-physical field stresses in the working profile of the inverter power switch, aging of the chip's internal components, package bonds, and pins can cause gate failure, bond cracking, and chip thermal breakdown. Ultimately, this can lead to an open or short circuit in the switch tube, drawing high current and burning the inverter. Inverter failures can cause equipment downtime or loss of functionality, impacting the normal operation of production and equipment tasks. In more serious cases, they can lead to serious consequences such as loss of control, fire, and burning of the equipment or equipment.

[0003] IGBT (Insulated Gate Bipolar Transistor) chip fatigue failure mechanism:

[0004] IGBT chips are one of the core components of inverters, responsible for converting direct current into alternating current. Within the inverter, IGBT chips control the direction and magnitude of current through rapid switching, thereby achieving efficient power conversion. Because inverters frequently switch high voltages and currents during operation, IGBT chips are subject to significant electrothermal stress, increasing the risk of chip fatigue failure. Therefore, the reliability and lifespan of IGBT chips directly impact the performance and stability of the inverter.

[0005] IGBT chip fatigue failure modes primarily involve interface fatigue and silicon material fatigue. When exposed to external stresses such as temperature (T), alternating temperature (ΔT), voltage (V), current (I), and humidity (H), IGBT devices can experience solid-solid diffusion, ion charge migration, hot electron injection, and even cracks at the interface between the chip's gate and emitter. This causes slow variations in the electrical, thermal, and mechanical properties of the interface, leading to device parameter degradation and eventual failure.

[0006] At the Si-SIO2 interface, dangling bonds of silicon atoms can exchange electrons and holes with the silicon surface layer, modulating the silicon surface potential and causing device parameter instability. They can also simultaneously capture an electron and hole, acting as recombination centers, causing the gate channel threshold voltage of the IGBT device to drift and leakage current to increase. At the Al-Si interface, after being subjected to a strong current shock, physical processes such as solid-state dissolution of Si into Al, electromigration of Si at the Al grain boundary, and thermal electromigration of Al in Si occur. These accelerate the interpenetration of aluminum and silicon atoms at the Al-Si interface of the IGBT device emitter, resulting in increased emitter interface leakage current and even device failure.

[0007] Under electrothermal stress, the IGBT gate interface fatigues, increasing the gate charge density. This change accumulates with the duration of stress application, leading to a gradual increase in the IGBT device threshold voltage and leakage current. Fatigue also occurs at the IGBT emitter interface, increasing the emitter interface leakage current. The PN junction breakdown voltage decreases, increasing the leakage current. This fatigue effect accumulates with the duration of stress application.

[0008] Principle of fault precursor parameters:

[0009] Because the gate leakage current I G and V th Obviously related, the tail current I CE and T on / T off Obviously related, so I G and I CE Characterize the degradation of IGBT. Each time IGBT goes through a turn-on and turn-off cycle, it means a cycle. In this paper, the gate leakage current I G As a failure precursor parameter, when the IGBT is turned off, the small current flowing through the gate is I G .

[0010] Figure 1 Shows the IGBT gate current (I G) control circuit waveforms in the on and off states. This waveform shows a significant difference in gate current between the on and off states of the IGBT: the gate current is high during on-state, while it drops rapidly and stabilizes during off-state. This characteristic provides a basis for identifying the on and off states of the IGBT. Based on this, code can be written to analyze the gate current waveform, using the current difference between the on and off states to accurately identify the IGBT's operating state. Furthermore, gate leakage current (the tiny current in the off state) can be used as a precursor parameter to IGBT aging failure. As the IGBT ages, the gate leakage current gradually increases. Therefore, by monitoring and extracting the changing trend of the gate leakage current, early warning of IGBT failure risks can be provided, providing an important reference for device health assessment.

[0011] Principles of Mahalanobis distance and Kalman filter algorithm:

[0012] Mahalanobis distance is a measure of the distance between two vectors (or data points), particularly useful for high-dimensional data and when there is correlation. Unlike Euclidean distance, Mahalanobis distance takes into account the covariance structure of the dataset, making it more suitable for analyzing variables with different scales and correlations.

[0013] The Mahalanobis distance measures the correlation between different dimensions in the data through the covariance matrix. If there is a strong correlation between the variables, the distance in that direction will be shortened, while the distance in the uncorrelated direction will be expanded. The Mahalanobis distance formula is as follows:

[0014]

[0015] x and y are two sample vectors; S is the covariance matrix of the data; S -1 is the inverse of the covariance matrix; is the difference between two sample vectors.

[0016] The Kalman filter, based on a linear Gaussian system model, recursively estimates the system state through two steps: prediction and update. The core idea of the Kalman filter is to use prior knowledge (the system model and noise characteristics) and current observations to continuously revise the estimate of the system state, thereby obtaining the optimal estimate.

[0017] The basic Kalman filter process consists of two phases: prediction and update. The prediction phase uses the system's state transition model to predict the system state and state covariance at the next moment based on the current state estimate and known inputs. The update phase uses new observations combined with the predicted values to correct the predicted state estimate and update the system's covariance matrix.

[0018] The mathematical model of the Kalman filter is based on two equations:

[0019] State equation: describes the dynamic behavior of the system. The state equation is as follows.

[0020]

[0021] x k is the state vector at time k. A is the state transition matrix. B is the control matrix, describing the input u k-1 Impact on the system. k-1 is the process noise, which is assumed to be Gaussian noise with mean 0 and covariance Q.

[0022] Observation equation: describes the relationship between the observed value and the state of the system. The observation equation is as follows.

[0023]

[0024] Z k is the observation vector at time k. H is the observation matrix, which describes the relationship between the state and the observation value. k is the observation noise, which is assumed to be Gaussian noise with mean 0 and covariance R.

[0025] CNN-LSTM model

[0026] The CNN-LSTM model combines a convolutional neural network (CNN) with a long short-term memory (LSTM) network to process complex data with spatiotemporal dependencies. This model can simultaneously extract both spatial and temporal features of the data, making it ideal for sequence prediction of inverter switching transistors.

[0027] Through operations such as convolutional layers and pooling layers, CNN can effectively extract local spatial features. Its principle is to scan the input data through convolution kernels (filters) to extract local patterns and features.

[0028] LSTM is a special type of recurrent neural network (RNN) that excels at processing time series data and can capture long-term and short-term dependencies. Through its memory cells and gating mechanism, LSTM effectively memorizes long-term information and ignores less important information. The LSTM is primarily used to process the temporal dependencies of features extracted by CNN, understand how these features change over time, and predict the future lifespan of switching transistors.

[0029] The failure of the inverter's power switching tubes is caused by device aging and shows a gradual degradation trend. Evaluation and early warning can be performed by monitoring the sensitive characterization parameters corresponding to the fault. How to accurately predict the performance of a single switching tube and comprehensively evaluate the overall performance of the entire inverter are urgent industry needs to ensure the controllable health status of the system and intelligent operation and maintenance.

[0030] Existing methods for predicting failure of power switching tubes in inverters primarily collect sensitive parameters such as collector leakage current or gate threshold voltage as precursor parameters for the switching tubes. These parameters are then processed and analyzed using machine learning algorithms such as SVM and LSTM, ultimately effectively predicting the lifespan of the switching tubes. For inverter module lifespan assessment, existing technologies primarily focus on predicting the lifespan of individual switching tubes and require the collection of precursor parameters. In real-world engineering applications, a comprehensive prediction and assessment of the lifespan of the entire inverter is often required. Furthermore, the collector leakage current detection device is relatively large, placing high demands on the acquisition equipment. Summary of the Invention

[0031] To solve the above technical problems. The present invention proposes a method for evaluating the performance degradation of inverter switches, which belongs to the field of inverter switches. Collect the original data of six switching tubes; pre-process the original data, and construct a degradation index sequence that expresses the degree of degradation of a single switching tube; construct and train a CNN-LSTM model to obtain a degradation prediction model corresponding to each switching tube; input the data of the switching tube to be tested into the degradation prediction model to obtain the degradation index of each switching tube; fuse the degradation indexes to obtain the degradation evaluation index of the switching tube group. The present invention utilizes the high reliability characteristics of the CNN-LSTM model in the field of time series prediction, combined with the pre-processing of the pre-parameters, greatly improving the accuracy of the switch tube aging prediction, which can improve the operating safety of the electric drive system and the cost-effectiveness and feasibility of the predictive maintenance guarantee plan, and bring better market competitiveness.

[0032] A method for evaluating inverter switch performance degradation, wherein the specific steps of the evaluation method are as follows:

[0033] Step 1: Use USB-DSO2 digital storage oscilloscope to collect the original data of the inverter switch tube group: the original data includes the pre-parameter I of the switch tube group G and I CE ;I G is the gate leakage current, I CE is the collector current; the switch tube group includes six switch tubes; the USB-DSO2 digital storage oscilloscope simultaneously collects the six switch tubes; the pre-parameter I G Contains the time domain feature sequence X i , (0<i≤13, i is an integer), pre-parameter I CE Contains the time domain feature sequence Y j , (0<j≤13, j is an integer);

[0034] Step 2: The time domain characteristic sequences X of the six switching tubes are respectively i and Y j Perform data preprocessing to obtain the degradation index sequences of the six switching transistors; divide the degradation index sequence data set of each switching transistor into a training set and a test set in a 7:3 ratio;

[0035] Step 3: Build a CNN-LSTM model;

[0036] Step 4: Use the training set to train the CNN-LSTM model: Use the degradation indicator sequences of the six switching transistors to independently train the CNN-LSTM model to obtain the CNN-LSTM models of the six switching transistors.

[0037] Step 5: Use the test set to test the trained CNN-LSTM model to obtain the degradation prediction models of the six switching transistors;

[0038] Step 6: Input the test data of the six switching tubes into six degradation prediction models respectively to obtain degradation index prediction values of the six switching tubes; the degradation index prediction values are life assessment values;

[0039] Step 7: Perform Mahalanobis distance fusion on the degradation index prediction values of the six switch tubes to obtain the degradation index prediction value of the switch group;

[0040] Step 8: Compare the predicted value of the switch tube group degradation index with the failure threshold to determine whether the inverter has failed. If the predicted value of the switch tube group degradation index is greater than or equal to the failure threshold, the inverter has failed and the switch tube of the inverter needs to be replaced.

[0041] Furthermore, in step 1, the time domain feature sequence X i and Y jThey all include 13 time domain features, including mean, variance (var), maximum (max), minimum (min), standard deviation (std), root mean square (fgf), root mean square (rms), peak-to-peak (pp), skewness (skew), kurtosis (kur), crest factor (crest), margin factor (clear) and impulse index (imp).

[0042] Furthermore, in step 2, the specific steps of the data preprocessing are as follows:

[0043] Step 2.1, Grey relational analysis

[0044] Through grey relational analysis, the time domain feature sequence X i Select any three sequences to form a triplet, and calculate the sum of the grey correlation degrees corresponding to the triplet, and take the three sequences corresponding to the triplet with the smallest sum of correlation degrees as the time domain feature sequence group G1;

[0045] Among them, the grey correlation The calculation formula is as follows:

[0046]

[0047] Where, is the time domain feature mother sequence, is the time domain feature subsequence; express The value at point k, express The value at point k; express and Grey correlation coefficient at point k; n represents the number of switching cycles of the switch tube, which is the sampling point;

[0048] Grey correlation coefficient The formula is as follows:

[0049] (2)

[0050] Where, represents the resolution coefficient, The value range is (0,1); s means ,(0 <s≤12); express and The minimum value of the distance; express and The maximum value of the distance; express and The distance at point k;

[0051] Similarly, through grey correlation analysis, from the time domain feature sequence Y j The time domain feature sequence group G2 is obtained;

[0052] Step 2.2, Mahalanobis distance fusion

[0053] The six time domain feature sequences in the time domain feature sequence groups G1 and G2 are fused using Mahalanobis distance to obtain the Mahalanobis distance sequence, which is the health deviation sequence.

[0054] The six time domain feature sequences are taken as sample set A 6×n , take the first 100 data points of the sample set as the healthy sample set B 6×100 , sample set A 6×n For the healthy sample set B 6×100 Mahalanobis distance The calculation formula is as follows:

[0055] (3)

[0056] Where, It's B 6×100 The average value of Representative sample set A 6×n ; B represents the healthy sample set B 6×100 ; for The transpose of ; Q represents the covariance matrix, is the inverse matrix of Q; the calculation formula of the covariance matrix Q is as follows:

[0057] (4)

[0058] in, The mean vector representing the time domain features;

[0059] Each MD value in the Mahalanobis distance sequence corresponds to a sampling point relative to the healthy sample set B 6×100 The Mahalanobis distance represents the degree of deviation of each sampling period relative to the healthy sample set;

[0060] Step 2.3, Dimensionality reduction of Kalman filter

[0061] The observation noise covariance is set to 0.1, the first sampling point of the Mahalanobis distance sequence is set to the initial state, the state transition matrix is set to 1, and the prediction noise covariance (transition covariance) is set to 0.01; the Mahalanobis distance sequence is first filtered using a one-dimensional Kalman filter, and then Kalman filter smoothing is performed to obtain the degradation indicator sequence of the switching tube.

[0062] Furthermore, in step 3, the CNN-LSTM model is constructed as follows:

[0063] The CNN-LSTM model sequentially comprises a feature encoding module, a feature enhancement module, a time series prediction module and a prediction output module;

[0064] The feature encoding module uses a one-dimensional convolution module to perform preliminary feature learning on the input degradation indicator sequence. First, 64 convolution kernels are set to perform sliding window scanning in the time dimension to capture local dynamic features in the input signal. Then, the results of the one-dimensional convolution module are processed by a nonlinear activation function to enhance the expressive power of the feature encoding model.

[0065] The feature enhancement module uses a two-stage fully connected network to reconstruct the output of the feature encoding module. First, the first-stage fully connected network projects the high-dimensional features into a medium-dimensional space, and then uses an activation function to implement a nonlinear transformation to obtain medium-dimensional features. Then, the second-stage fully connected network compresses the medium-dimensional features into a low-dimensional representation space to obtain a low-dimensional feature sequence with temporal correlation.

[0066] The time series prediction module uses a two-layer LSTM network to model time dependencies. The low-dimensional feature sequence is input into the memory unit of the first-layer LSTM to obtain the long-term dependency relationship of the low-dimensional feature sequence in the time dimension. The second-layer LSTM performs deep feature extraction on the output of the first-layer LSTM and finally outputs a feature representation sequence containing time series context information.

[0067] The prediction output module uses a fully connected network to map the feature representation sequence output by the time series prediction module to the target dimensional space, and generates the predicted value of the degradation indicator through dimensional transformation.

[0068] Furthermore, in step 4, the training process of the CNN-LSTM model is as follows:

[0069] Step 4.1: Use the sliding window method to group the degradation index sequence: group the data into groups of five consecutive time steps, with each time step containing a one-dimensional feature vector, and each group containing one-dimensional feature vectors for five consecutive time steps. The degradation index values of the first four time steps of each group serve as the input of the CNN-LSTM model, and the degradation index value of the fifth time step serves as the comparison value. The output of the CNN-LSTM model is compared with the comparison value.

[0070] Step 4.2: Adjust the degradation indicator sequence of the training set: adjust the degradation indicator sequence to a three-dimensional tensor structure to meet the input requirements of the one-dimensional convolution layer;

[0071] In step 4.3, backpropagation is performed using the Adam optimizer: the learning rate of the Adam optimizer is initially set to 0.001, and early stopping is used to prevent overfitting. Training is terminated if there is no improvement in the validation set loss for 10 consecutive epochs.

[0072] Furthermore, in step 7, the Mahalanobis distance fusion process of the degradation index prediction value is as follows:

[0073] The four most recent historical degradation index values of each of the six switching tubes are input into the six degradation prediction models respectively to obtain the degradation index prediction values Z of the six switching tubes. 6×1 ;

[0074] The first 100 degradation index sequences of each of the six switching tubes are taken as the healthy sample set H of the six switching tubes. 6×100 ; Calculate the degradation index prediction value Z by Mahalanobis distance 6×1 To the healthy sample set H 6×100 The distance between the switch tube group and the load is calculated to obtain the predicted degradation index value T of the switch tube group.

[0075] The beneficial effects of the present invention are as follows:

[0076] This invention cleverly utilizes the characteristics of the switching tube's pre-measurement parameters IG and ICE. Gate current IG offers low detection cost and good characterization results; collector current ICE can sensitively reflect IGBT package structure aging and electrothermal stress accumulation. This invention solves the industry's difficult problem of early warning for inverter switching tube groups and provides technical support for the reliability of electrically driven new energy products.

[0077] The present invention utilizes the high reliability of the CNN-LSTM model in the field of time series prediction, combined with the preprocessing of the pre-parameters, to greatly improve the accuracy of switch tube aging prediction, which can improve the operational safety of the electric drive system and the cost-effectiveness and feasibility of the predictive maintenance guarantee plan, bringing better market competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 It is the dynamic curve of the gate current part;

[0079] Figure 2 Provide a degradation assessment scheme for inverter systems;

[0080] Figure 3 Flowchart for training a single-tube degradation assessment model;

[0081] Figure 4 Construct a flow chart for single-tube performance degradation indicators;

[0082] Figure 5 This is the flowchart of the CNN-LSTM model;

[0083] Figure 6 Flowchart for CNN-LSTM model training;

[0084] Figure 7 This is the inverter comprehensive evaluation flow chart. DETAILED DESCRIPTION

[0085] The present invention will be further described below with reference to the accompanying drawings and examples.

[0086] The present invention proposes a method for evaluating the performance degradation of inverter switches, which belongs to the field of inverter switches. The original data of six switching tubes are collected; the original data are preprocessed, and a degradation index sequence that expresses the degree of degradation of a single switching tube is constructed; a CNN-LSTM model is constructed and trained to obtain a degradation prediction model corresponding to each switching tube; the data of the switching tube to be tested is input into the degradation prediction model to obtain the degradation index of each switching tube; the degradation indexes are fused to obtain the degradation evaluation index of the switching tube group. The present invention utilizes the high reliability characteristics of the CNN-LSTM model in the field of time series prediction, combined with the preprocessing of the pre-parameters, to greatly improve the accuracy of the switch tube aging prediction, which can improve the operational safety of the electric drive system and the cost-effectiveness and feasibility of the predictive maintenance guarantee plan, bringing better market competitiveness.

[0087] A method for evaluating inverter switch performance degradation, wherein the specific steps of the evaluation method are as follows:

[0088] like Figure 2 and Figure 3 As shown:

[0089] Step 1: Use USB-DSO2 digital storage oscilloscope to collect the original data of the inverter switch tube group: the original data includes the pre-parameter I of the switch tube group G and I CE ;I G is the gate leakage current, I CE is the collector current; the switch tube group includes six switch tubes; the USB-DSO2 digital storage oscilloscope simultaneously collects the six switch tubes; the pre-parameter I G Contains the time domain feature sequence X i , (0<i≤13, i is an integer), pre-parameter I CE Contains the time domain feature sequence Y j , (0<j≤13, j is an integer); time domain feature sequence X i and Y jThey all include 13 time domain features: mean (mean), variance (var), maximum (max), minimum (min), standard deviation (std), root mean square (fgf), root mean square (rms), peak-to-peak (pp), skewness (skew), kurtosis (kur), crest factor (crest), margin factor (clear) and impulse index (imp);

[0090] Step 2: The time domain characteristic sequences X of the six switching tubes are respectively i and Y j Perform data preprocessing to obtain the degradation index sequences of the six switching transistors; divide the degradation index sequence data set of each switching transistor into a training set and a test set in a 7:3 ratio;

[0091] Step 3: Build a CNN-LSTM model;

[0092] Step 4: Use the training set to train the CNN-LSTM model: Use the degradation indicator sequences of the six switching transistors to independently train the CNN-LSTM model to obtain the CNN-LSTM models of the six switching transistors.

[0093] Step 5: Use the test set to test the trained CNN-LSTM model to obtain the degradation prediction models of the six switching transistors;

[0094] Step 6: Input the test data of the six switching tubes into six degradation prediction models respectively to obtain degradation index prediction values of the six switching tubes; the degradation index prediction values are life assessment values;

[0095] Step 7: Perform Mahalanobis distance fusion on the degradation index prediction values of the six switching tubes to obtain the degradation index prediction value of the switching tube group;

[0096] Step 8: Compare the predicted value of the switch tube group degradation index with the failure threshold to determine whether the inverter has failed. If the predicted value of the switch tube group degradation index is greater than or equal to the failure threshold, the inverter has failed and the switch tube of the inverter needs to be replaced.

[0097] like Figure 4 As shown:

[0098] In step 2, the data preprocessing includes the following steps:

[0099] Step 2.1, Grey relational analysis

[0100] Through grey relational analysis, the time domain feature sequence X i Select any three sequences to form a triplet, and calculate the sum of the grey correlation degrees corresponding to the triplet, and take the three sequences corresponding to the triplet with the smallest sum of correlation degrees as the time domain feature sequence group G1;

[0101] Among them, the grey correlation The calculation formula is:

[0102]

[0103] Where, is the time domain feature mother sequence, is the time domain feature subsequence; express The value at point k, express The value at point k; express and Grey correlation coefficient at point k; n represents the number of switching cycles of the switch tube, which is the sampling point;

[0104] Grey correlation coefficient The formula is as follows:

[0105] (2)

[0106] Where, represents the resolution coefficient, The value range is (0,1); s means ,(0 <s≤12); express and The minimum value of the distance; express and The maximum value of the distance; express and The distance at point k;

[0107] Similarly, through grey correlation analysis, from the time domain feature sequence Y j The time domain feature sequence group G2 is obtained;

[0108] Step 2.2, Mahalanobis distance fusion

[0109] The six time domain feature sequences in the time domain feature sequence groups G1 and G2 are fused using Mahalanobis distance to obtain the Mahalanobis distance sequence, which is the health deviation sequence.

[0110] The six time domain feature sequences are taken as sample set A 6×n , take the first 100 data points of the sample set as the healthy sample set B 6×100 , sample set A 6×n For the healthy sample set B 6×100 Mahalanobis distance The calculation formula is as follows:

[0111] (3)

[0112] Where, It's B 6×100 The average value of Representative sample set A 6×n ; B represents the healthy sample set B 6×100 ; for The transpose of ; Q represents the covariance matrix, Q -1 is the inverse matrix of Q; the calculation formula of the covariance matrix Q is as follows:

[0113] (4)

[0114] in, The mean vector representing the time domain features;

[0115] Each MD value in the Mahalanobis distance sequence corresponds to a sampling point relative to the healthy sample set B 6×100 The Mahalanobis distance represents the degree of deviation of each sampling cycle from the healthy sample set; it represents the degree of deviation of a single IGBT switch from the healthy state; the MD value is used to describe the degradation degree of the IGBT, that is, the degradation index;

[0116] Step 2.3, Dimensionality reduction of Kalman filter

[0117] The observation noise covariance is set to 0.1, the first sampling point of the Mahalanobis distance sequence is set to the initial state, the state transition matrix is set to 1, and the prediction noise covariance (transition covariance) is set to 0.01; the Mahalanobis distance sequence is first filtered using a one-dimensional Kalman filter, and then Kalman filter smoothing is performed to obtain the degradation indicator sequence of the switching tube.

[0118] like Figure 5 As shown:

[0119] In step 3, the CNN-LSTM model is constructed as follows:

[0120] The CNN-LSTM model sequentially comprises a feature encoding module, a feature enhancement module, a time series prediction module and a prediction output module;

[0121] The feature encoding module uses a one-dimensional convolution module to perform preliminary feature learning on the input degradation indicator sequence. First, 64 convolution kernels are set to perform sliding window scanning in the time dimension to capture local dynamic features in the input signal. Then, the results of the one-dimensional convolution module are processed by a nonlinear activation function to enhance the expressive power of the feature encoding model.

[0122] The feature enhancement module uses a two-stage fully connected network to reconstruct the output of the feature encoding module. First, the first-stage fully connected network projects the high-dimensional features into a medium-dimensional space, and then uses an activation function to implement a nonlinear transformation to obtain medium-dimensional features. Then, the second-stage fully connected network compresses the medium-dimensional features into a low-dimensional representation space to obtain a low-dimensional feature sequence with temporal correlation.

[0123] The time series prediction module uses a two-layer LSTM network to model time dependencies. The low-dimensional feature sequence is input into the memory unit of the first-layer LSTM to obtain the long-term dependency relationship of the low-dimensional feature sequence in the time dimension. The second-layer LSTM performs deep feature extraction on the output of the first-layer LSTM and finally outputs a feature representation sequence containing time series context information.

[0124] The prediction output module uses a fully connected network to map the feature representation sequence output by the time series prediction module to the target dimensional space, and generates the predicted value of the degradation indicator through dimensional transformation, thereby completing the end-to-end time series data processing flow.

[0125] The CNN-LSTM model achieves end-to-end mapping from multi-step input to single-step prediction through the collaboration of convolution-fully connected feature abstraction and LSTM temporal dynamic modeling.

[0126] like Figure 6 As shown:

[0127] In step 4, the training process of the CNN-LSTM model is as follows:

[0128] For a degradation index sequence of length n, the first four time steps in each set of data are used as the input of the CNN-LSTM model. Each time step contains a 1-dimensional feature, and the degradation index value at the fifth time step is used as the prediction target.

[0129] After traversing all degradation indicator sequence groups, the degradation indicator sequence of the training set is adjusted to a three-dimensional tensor structure to meet the input requirements of the one-dimensional convolutional layer. The three-dimensional tensor structure is (batch_size, 4, 1), where 4 is the input sequence length, 1 is the single-channel degradation indicator, and the batch size batch_size represents the number of samples for forward propagation and backpropagation of a single input model. The batch_size value is 1.

[0130] The root mean square error (RMSE) was used as the loss function, and backpropagation was performed through the Adam optimizer. The learning rate was initially set to 0.001, and early stopping was used to prevent overfitting. Training was terminated if there was no improvement in the validation set loss after 10 consecutive epochs. Independent CNN-LSTM models were established and trained for the six IGBT switches of the inverter, ultimately obtaining degradation prediction models for the six switches.

[0131] like Figure 7 As shown:

[0132] In step 7, the Mahalanobis distance fusion process of the degradation index prediction value is as follows:

[0133] The four most recent historical operating data of each of the six switching tubes are preprocessed to obtain the four most recent degradation indicators of each switching tube. The four degradation indicators of each switching tube are input into their respective CNN-LSTM models to obtain the degradation indicator prediction values Z of the six switching tubes. 6×1 .

[0134] The first 100 degradation index sequences of each of the six switching tubes are taken as the healthy sample set H of the six switching tubes. 6×100 ; Calculate the degradation index prediction value Z by Mahalanobis distance 6×1 To the healthy sample set H 6×100 The distance between the switch tube group and the degradation index prediction value T is obtained;

[0135] The predicted value T of the switch tube group degradation index is compared with the failure threshold Tthreshold. If the predicted value T of the switch tube group degradation index is greater than or equal to the failure threshold Tthreshold, it means that the switch tube of the inverter is about to fail. The switch tube of the inverter needs to be replaced to ensure safe operation of the equipment.

Claims

1. A method for evaluating inverter switch performance degradation, characterized in that: The specific steps of the evaluation method are as follows: Step 1: Use USB-DSO2 digital storage oscilloscope to collect the original data of the inverter switch tube group: the original data includes the pre-parameter I of the switch tube group G and I CE ;I G is the gate leakage current, I CE is the collector current; the switch tube group includes six switch tubes; The USB-DSO2 digital storage oscilloscope collects data from six switching tubes at the same time; the pre-parameter I G Contains the time domain feature sequence X i , pre-parameter I CE Contains the time domain feature sequence Y j ; The time domain feature sequences Xi and Yj each contain 13 time domain features, namely, mean value, variance, maximum value, minimum value, standard deviation, root square amplitude, root mean square value, peak-to-peak value, skewness, kurtosis, crest factor, margin factor and pulse index; Step 2: The time domain characteristic sequences X of the six switching tubes are respectively i and Y j Perform data preprocessing to obtain the degradation index sequences of the six switching transistors; divide the degradation index sequence data set of each switching transistor into a training set and a test set in a 7:3 ratio; The specific steps of data preprocessing are as follows: Step 2.1, Grey relational analysis Through grey relational analysis, the time domain feature sequence X i Select any three sequences to form a triplet, and calculate the sum of the grey correlation degrees corresponding to the triplet, and take the three sequences corresponding to the triplet with the smallest sum of correlation degrees as the time domain feature sequence group G1; Among them, the grey relational degree γ(X M ,X m ) is calculated as: Where, X M is the time domain feature mother sequence, X m is the time domain feature subsequence; x M (k) represents X M The value at point k, x m (k) represents X m The value at point k; ξ(x M (k),x m (k)) represents X M and X m Grey correlation coefficient at point k; n represents the number of switching cycles of the switch tube, which is the sampling point; Grey relational coefficient ξ(x M (k),x m The formula for (k)) is as follows: Where, ρ represents the resolution coefficient, and the value range of ρ is (0,1); s represents X m the number of x m (k)| represents X M With X m The minimum value of the distance; Represents X M With X m The maximum value among the distances of |x M (k)-x m (k)| represents X M and X m The distance at point k; Similarly, through grey correlation analysis, from the time domain feature sequence Y j The time domain feature sequence group G2 is obtained; Step 2.2, Mahalanobis distance fusion The six time domain feature sequences in the time domain feature sequence groups G1 and G2 are fused using Mahalanobis distance to obtain the Mahalanobis distance sequence, which is the health deviation sequence. The six time domain feature sequences are taken as sample set A 6×n , take the first 100 data points of the sample set as the healthy sample set B 6×100 , sample set A 6×n For the healthy sample set B 6×100 The Mahalanobis distance MD calculation formula is as follows: Where, It's B 6×100 The average value of; A represents the sample set A 6×n ; B represents the healthy sample set B 6×100 ; for The transpose of ; Q represents the covariance matrix, Q -1 is the inverse matrix of Q; the calculation formula of the covariance matrix Q is as follows: Among them, μ represents the mean vector of time domain features; Step 2.3, dimensionality reduction processing of Kalman filter: Set the observation noise covariance to 0.1, set the first sampling point of the Mahalanobis distance sequence to the initial state, set the state transfer matrix to 1, and set the prediction noise covariance to 0.01; use a one-dimensional Kalman filter to first filter the Mahalanobis distance sequence, and then perform Kalman filter smoothing to obtain the degradation indicator sequence of the switching tube; Step 3: Build a CNN-LSTM model; The CNN-LSTM model is constructed as follows: The CNN-LSTM model sequentially comprises a feature encoding module, a feature enhancement module, a time series prediction module and a prediction output module; The feature encoding module uses a one-dimensional convolution module to perform preliminary feature learning on the input degradation indicator sequence. First, 64 convolution kernels are set to perform sliding window scanning in the time dimension to capture local dynamic features in the input signal. Then, the results of the one-dimensional convolution module are processed by a nonlinear activation function to enhance the expressive power of the feature encoding model. The feature enhancement module uses a two-stage fully connected network to reconstruct the output of the feature encoding module. First, the first-stage fully connected network projects the high-dimensional features into a medium-dimensional space, and then uses an activation function to implement a nonlinear transformation to obtain medium-dimensional features. Then, the second-stage fully connected network compresses the medium-dimensional features into a low-dimensional representation space to obtain a low-dimensional feature sequence with temporal correlation. The time series prediction module uses a two-layer LSTM network to model time dependencies. The low-dimensional feature sequence is input into the memory unit of the first-layer LSTM to obtain the long-term dependency relationship of the low-dimensional feature sequence in the time dimension. The second-layer LSTM performs deep feature extraction on the output of the first-layer LSTM and finally outputs a feature representation sequence containing time series context information. The prediction output module uses a fully connected network to map the feature representation sequence output by the time series prediction module to the target dimensional space, and generates the predicted value of the degradation indicator through dimensional transformation; Step 4: Use the training set to train the CNN-LSTM model: Use the degradation indicator sequences of the six switching transistors to independently train the CNN-LSTM model to obtain the CNN-LSTM models of the six switching transistors. The training process of the CNN-LSTM model is as follows: Step 4.1: Use the sliding window method to group the degradation index sequence: group the data into groups of five consecutive time steps, with each time step containing a one-dimensional feature vector, and each group containing one-dimensional feature vectors for five consecutive time steps. The degradation index values of the first four time steps of each group serve as the input of the CNN-LSTM model, and the degradation index value of the fifth time step serves as the comparison value. The output of the CNN-LSTM model is compared with the comparison value. Step 4.2: Adjust the degradation indicator sequence of the training set: adjust the degradation indicator sequence to a three-dimensional tensor structure to meet the input requirements of the one-dimensional convolution layer; Step 4.3: Perform backpropagation using the Adam optimizer. The learning rate of the Adam optimizer is initially set to 0.001, and early stopping is used to prevent overfitting. Training is terminated if there is no improvement in the validation set loss for 10 consecutive epochs. Step 5: Use the test set to test the trained CNN-LSTM model to obtain the degradation prediction models of the six switching transistors; Step 6: Input the test data of the six switching tubes into six degradation prediction models respectively to obtain degradation index prediction values of the six switching tubes; the degradation index prediction values are life assessment values; Step 7: Perform Mahalanobis distance fusion on the degradation index prediction values of the six switch tubes to obtain the degradation index prediction value of the switch group; The Mahalanobis distance fusion process of the degradation index prediction value is as follows: The four most recent historical degradation index values of each of the six switching tubes are input into the six degradation prediction models respectively to obtain the degradation index prediction values Z of the six switching tubes. 6×1 ; The first 100 degradation index sequences of each of the six switching tubes are taken as the healthy sample set H of the six switching tubes. 6×100 ; Calculate the degradation index prediction value Z by Mahalanobis distance 6×1 To the healthy sample set H 6×100 The distance between the switch tube group and the degradation index prediction value T is obtained; Step 8: Compare the predicted value of the switch tube group degradation index with the failure threshold to determine whether the inverter has failed. If the predicted value of the switch tube group degradation index is greater than or equal to the failure threshold, the inverter has failed and the switch tube of the inverter needs to be replaced.

2. An electronic device, characterized in that: include: one or more processors; Memory; One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs being configured to perform the method of claim 1.

Images