An IGBT fault prediction method and system based on inverted Transformer network
By using the Invert-Transformer-LSTM model of the inverted Transformer network in IGBT fault prediction, combining the multi-head attention mechanism and LSTM module, the problem of both prediction accuracy and rate in the existing technology is solved, and high-precision and efficient fault prediction are achieved. .
Patent Information
- Application Number
- CN202411236939.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-05
- Publication Date
- 2024-12-27
- Estimated Expiration
- 2044-09-05
AI Technical Summary
The existing IGBT fault prediction model cannot take into account both prediction accuracy and prediction rate. The traditional LSTM model has low prediction accuracy, while the Transformer model has complex structure and low prediction accuracy improvement.
The Invert-Transformer-LSTM neural network model based on the inverted Transformer network is adopted. By selecting the collector-emitter peak voltage as characteristic parameters, combining the multi-head attention mechanism and the LSTM module, hyperparameter optimization is performed to improve prediction accuracy and stability.
It realizes that while ensuring the accuracy of prediction results, it significantly improves the calculation speed, weakens the lag effect of LSTM-like models, obtains more accurate prediction results, and improves prediction accuracy and stability.
Smart Images
Figure CN118761444B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of IGBT fault prediction, and in particular relates to an IGBT fault prediction method and system based on an inverted Transformer network. Background Art
[0002] Insulated gate bipolar transistors (IGBTs) are the core components of key converter equipment in low-voltage DC distribution systems in urban rail transit, and their reliability is a key factor in determining the safe operation of station low-voltage DC distribution systems. If the IGBT status can be scientifically and accurately evaluated, the optimal replacement time of the IGBT can be determined, the full life cycle management of the equipment can be achieved, and the key converter equipment in the DC distribution system can be inspected and maintained in a timely manner, reducing potential faults and accidents in the operation of the DC distribution system.
[0003] In the prior art, the research on IGBT failure prediction is mainly based on three methods: physical model, analytical model and data-driven. The method based on analytical model is to achieve failure prediction by establishing a mathematical model to describe the operating characteristics and behavior of IGBT. The method based on physical model is to achieve module failure prediction by constructing an IGBT finite element model. The method based on data-driven is to extract the degradation characteristics of the device from the data and determine the aging trend and remaining life of the device in combination with the model. This method has the advantages of saving time, improving accuracy and accelerating the design process. Focusing on the above three methods, domestic and foreign experts have carried out the following research on IGBT failure prediction:
[0004] M. Baharani used Long Short-Term Memory (LSTM) deep learning for SiMOSFET life prediction. Its prediction method has better performance than Kalman filter and particle filter methods, but the prediction accuracy is low. In order to improve the prediction accuracy of LSTM model, Ren Hongyu used the SVMD technology to extract degradation features and optimized the hyperparameters of LSTM model through Bayesian optimization (BO) method. The prediction accuracy was significantly improved, but the complexity of data processing and model training required more computing resources and time. Ge Jianwen predicted the remaining life of IGBT based on Transformer model. Compared with LSTM model, it performed well in handling long-distance dependencies, had faster training speed and fewer parameters, but its public Transformer model had a complex structure and the prediction accuracy was lower than that of LSTM model.
[0005] Therefore, existing models cannot take into account both prediction accuracy and prediction rate at the same time. Summary of the invention
[0006] The present invention provides an IGBT fault prediction method and system based on an inverted Transformer network, which are used to solve the problems existing in the prior art.
[0007] In a first aspect, to achieve the above-mentioned purpose, the present invention provides an IGBT fault prediction method based on an inverted Transformer network, comprising:
[0008] S1: Select the collector-emitter peak voltage in the IGBT aging data as the characteristic parameter;
[0009] S2: Collect the collector-emitter voltage signal of the target IGBT, extract the transient peak voltage, and use a quadratic exponential smoothing filter function to filter the transient peak voltage to obtain sample data;
[0010] S3: Divide the sample data into training set and test set;
[0011] S4: Build an Invert-Transformer-LSTM (Invert-Transformer-Long Short-Term Memory) neural network model with a built-in multi-head attention mechanism;
[0012] S5: Optimize the hyperparameters of the Invert-Transformer-LSTM neural network model;
[0013] S6: Place the training set and test set into the Invert-Transformer-LSTM neural network model after hyperparameter optimization for training and testing, obtain the test results and perform performance evaluation on the test results;
[0014] S7: Repeat steps S5-S6 to obtain an optimal prediction model, and use the optimal prediction model to predict the actual collector-emitter peak voltage of the target IGBT in real time to obtain a real-time prediction result.
[0015] Preferably, the quadratic exponential smoothing filter function in step S2 is specifically:
[0016] (1)
[0017] (2)
[0018] in, For the time node i The steady-state quantity of the transient peak voltage; For the time node i The actual value of the upper transient peak voltage; the first smoothing coefficient A In the interval [0, 1], A When the value is close to 0, Get the previous time nodei-1 The smoothing result of A When the value is close to 1, it is retained Original feature information; is the trend quantity of transient peak voltage after double exponential smoothing; C is the second smoothing coefficient that affects the smoothing trend.
[0019] It should be explained that during the IGBT shutdown period, due to the presence of parasitic elements such as stray inductance, a transient voltage will be generated, which together with the device's shutdown voltage forms a transient peak voltage. As the IGBT performance degrades, this transient peak voltage (characteristic parameter) will gradually decrease until the latch effect occurs. This characteristic parameter can well reflect the degradation law of IGBT performance, so it can be used as a general case to study its fault prediction.
[0020] Preferably, the Invert-Transformer-LSTM neural network model includes:
[0021] The Invert-Transformer module is used to capture the global features and long-distance dependencies of sample data in the time series through a multi-head attention mechanism, establish connections between time nodes, focus on important features, and obtain the first high-dimensional feature vector;
[0022] The LSTM (Long Short-Term Memory) module is used to perform secondary learning on the global features extracted by the Invert-Transformer module, enhance the temporal correlation of the global features, and obtain the second high-dimensional feature vector; the LSTM module has a built-in fully connected layer, which is used to convert the first high-dimensional feature vector and the second high-dimensional feature vector extracted by the Invert-Transformer module and the LSTM module into prediction results.
[0023] Preferably, the Invert-Transformer module includes:
[0024] The embedding layer is used to map high-dimensional input data to a low-dimensional space to obtain an input sequence;
[0025] The multi-head attention mechanism layer is used to perform weighted aggregation on each part of the input sequence and output the compressed transformation result;
[0026] The feedforward network layer consists of two fully connected layers and an activation function, which is used to perform nonlinear transformation on the input sequence;
[0027] Normalization layer, with built-in residual connection function and layer standardization function.
[0028] Preferably, the expression of the multi-head attention mechanism layer is:
[0029] (3)
[0030] (4)
[0031] (5)
[0032] in, Q To query data, K is the key value, V is the value corresponding to the query data, is a normalizing function, is the scaling factor, , and They all represent parameter matrices, For a certain time point No. j The output of the head, is the number of attention heads, is a function for concatenating strings. is a linear matrix, is the attention mechanism function.
[0033] Preferably, the expression of the feedforward network layer is:
[0034] (6)
[0035] in, and is the weight parameter of the fully connected layer, and is the bias term, is the input value.
[0036] Preferably, the expression of the normalization layer is:
[0037] (7)
[0038] in, and are network parameters, is the vector dot product, and Input The mean and variance of is a minimum constant.
[0039] Preferably, the LSTM module is composed of multiple levels of LSTM units and fully connected layers, and the LSTM unit includes a forget gate, an input gate, an output gate and a unit state;
[0040] The forget gate is used to and the current momentt The cell state at the previous time step determines the information that needs to be discarded. The expression of the forget gate is:
[0041] (8)
[0042] in, is the weight matrix of the forget gate; Forget the current moment of the gate t The intermediate state of the previous time step; is the input of the current cell; is the bias term of the forget gate; for Activation function;
[0043] The input gate is used to select t To write new information to the cell state and update the cell state Integrating the selected information, the expression of the input gate is:
[0044] (9)
[0045] (10)
[0046] (11)
[0047] in, is the weight matrix of the input gate, is the input gate bias term; Formula (10) is used to describe the current input unit state, is the bias term of the input gate; Formula (11) is used to update the old cell state to the new cell state, is the input gate at the current moment t The intermediate state calculation results are: is the input gate at the current moment t The new cell state calculation result is is the weight matrix of the input gate; for Activation function, is the hyperbolic tangent function;
[0048] The output gate is used to determine the current moment t The intermediate state The cell state information that should be included, the expression of the output gate is:
[0049] (12)
[0050] (13)
[0051] in, is the weight matrix of the output gate, is the bias term of the output gate, for Activation function;
[0052] The expression of the fully connected layer is:
[0053] (14)
[0054] in, is the weight matrix of the fully connected layer; is the bias vector of the fully connected layer; For the current moment t The intermediate state is passed from the LSTM module; is the activation function of the fully connected layer; is the final output prediction value.
[0055] Preferably, step S6 comprises:
[0056] S61: placing the training set into the Invert-Transformer-LSTM neural network model after hyperparameter optimization for training, and then using the trained Invert-Transformer-LSTM neural network model to perform IGBT fault prediction on the test set to obtain test results;
[0057] S62: Perform performance evaluation on the test results according to the evaluation indicators;
[0058] The loss function formula used in the training process in step S61 is:
[0059] (15)
[0060] in, is the number of samples, is the true value of the sample data, is the predicted value of the sample data;
[0061] The evaluation indicators include root mean square error, mean absolute error and coefficient of determination.
[0062] In a second aspect, in order to achieve the above-mentioned purpose of the invention, the present invention further discloses an IGBT fault prediction system based on an inverted Transformer network, using the above-mentioned IGBT fault prediction method based on an inverted Transformer network, the system of the present invention comprises:
[0063] A characteristic parameter selection module is used to select the collector-emitter peak voltage in the IGBT aging data as a characteristic parameter;
[0064] The sample data acquisition module is used to collect the collector-emitter voltage signal of the target IGBT, extract the transient peak voltage, and filter the transient peak voltage using a quadratic exponential smoothing filter function to obtain sample data;
[0065] A sample data partitioning module is used to partition the sample data into a training set and a test set;
[0066] Model building module, used to build the Invert-Transformer-LSTM neural network model with built-in multi-head attention mechanism;
[0067] Hyperparameter optimization module, used to optimize the hyperparameters of the Invert-Transformer-LSTM neural network model;
[0068] The performance evaluation module is used to place the training set and the test set into the Invert-Transformer-LSTM neural network model after hyperparameter optimization for training and testing, obtain the test results and perform performance evaluation on the test results;
[0069] The fault prediction module is used to repeatedly execute the built-in steps of the hyperparameter optimization module and the performance evaluation module to obtain the optimal prediction model, and use the optimal prediction model to predict the actual collector-emitter peak voltage of the target IGBT in real time to obtain real-time prediction results.
[0070] Compared with the prior art, the present invention has at least the following advantages:
[0071] 1. The Invert-Transformer-LSTM neural network model has obvious advantages in prediction accuracy and stability. Unlike the traditional LSTM model, the predicted value and actual value of the Invert-Transformer-LSTM neural network model do not have a time lag offset, which effectively weakens the lag effect of the LSTM model and makes the prediction results more accurate.
[0072] 2. The Invert-Transformer-LSTM neural network model is equivalent to the Transformer model, but its structure is simpler. It takes into account higher computing speed while ensuring the accuracy of prediction results. BRIEF DESCRIPTION OF THE DRAWINGS
[0073] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0074] Figure 1 This is a flow chart of an IGBT fault prediction method based on an inverted Transformer network in Embodiment 1 of the present invention.
[0075] Figure 2 This is a schematic diagram of the Invert-Transformer-LSTM neural network model structure in Example 1 of the present invention.
[0076] Figure 3 Graphs showing the initial peak voltage curve and the peak voltage curve after filtering in step S2 in the first embodiment of the present invention.
[0077] Figure 4 Schematic diagram of the multi-head attention mechanism in Example 1 of the present invention.
[0078] Figure 5 Schematic diagram of the LSTM memory unit structure in Example 1 of the present invention.
[0079] Figure 6 This is a graph of prediction results after being processed by the LSTM model in the prior art.
[0080] Figure 7 This is a graph of prediction results after being processed by the Transformer model in the prior art.
[0081] Figure 8 This is a prediction structure diagram after being processed by the iTransformer model in the prior art.
[0082] Fig. 9 This is a diagram of the prediction results of the Invert-Transformer-LSTM neural network model in Example 1 of the present invention.
[0083] Fig.10 Schematic diagram of the structure of an IGBT fault prediction system based on an inverted Transformer network in Embodiment 2 of the present invention. DETAILED DESCRIPTION
[0084] In order to make the purpose, technical scheme and advantages of the present invention clearer, the specific embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings. It is understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for the convenience of description, only the parts related to the present invention are shown in the accompanying drawings, rather than all the contents. Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow chart describes each operation (or step) as a sequential process, many of the operations therein can be implemented in parallel, concurrently or simultaneously. In addition, the order of each operation can be rearranged. The above process can be terminated when its operation is completed, but it can also have additional steps not included in the accompanying drawings. The above process can correspond to methods, functions, procedures, subroutines and subprograms, etc.
[0085] The invention discloses an IGBT fault prediction method and system based on an inverted Transformer network, which are used to solve the problems existing in the prior art.
[0086] Embodiment 1
[0087] like Figure 1 As shown, this embodiment discloses an IGBT fault prediction method based on an inverted Transformer network, comprising:
[0088] S1: Select the collector-emitter peak voltage in the IGBT aging data as the characteristic parameter.
[0089] In this embodiment, the collector-emitter peak voltage is used as a characteristic parameter to assist in predicting the fault of the IGBT.
[0090] It should be noted that during the IGBT shutdown period, parasitic components such as stray inductance will instantly generate a transient voltage, which together with the device's shutdown voltage will generate a transient peak voltage. This voltage will decrease to the point where a latching effect occurs as the IGBT's performance degrades. This feature can better reflect the law of IGBT performance degradation and can be regarded as a case standard for fault prediction.
[0091] S2: Collect the collector-emitter voltage signal of the target IGBT, extract the transient peak voltage, and use a quadratic exponential smoothing filter function to filter the transient peak voltage to obtain sample data.
[0092] Wherein, before step S2, the method of the present invention further comprises:
[0093] S21: Build a DC power cycle accelerated aging test platform and apply a square wave voltage signal to the gate of the target IGBT.
[0094] In this embodiment, an accelerated aging strategy of applying a square wave voltage signal to the gate is used to obtain electrical parameter data of the target IGBT during the accelerated aging process. The model of the target IGBT is IRG4BC30K, the application frequency of the square wave voltage signal is 1kHz, the duty cycle is 40%, and the amplitude is 0-8V. By applying a square wave voltage signal to the gate of the target IGBT, the device is in a continuous overcurrent and high temperature state.
[0095] Specifically, Table 1 shows the test environment configuration of step S21:
[0096] Table 1
[0097]
[0098] After the above step S21, the gate voltage, collector voltage, collector current, emitter voltage, emitter current, package temperature and acquisition time of the target IGBT are recorded by an instrument when the square wave application device is turned on and off. The occurrence of latch-up effect of the IGBT is used as a sign to determine whether the device is failed.
[0099] Furthermore, the quadratic exponential smoothing filter function in the above step S2 is specifically:
[0100] (1)
[0101] (2)
[0102] in, For the time node i The steady-state quantity of the transient peak voltage; For the time node i The actual value of the upper transient peak voltage; the first smoothing coefficient In the interval [0, 1], A When the value is close to 0, Get the previous time node i-1 The smoothing result of A When the value is close to 1, it is retained Original feature information; is the trend quantity of transient peak voltage after double exponential smoothing; C is the second smoothing coefficient that affects the smoothing trend.
[0103] By using the above formulas (1) and (2), the steady-state quantity and trend quantity of the transient peak voltage can be combined and a sequence filtering operation can be performed on the transient peak voltage with a weak trend period.
[0104] After the above step S2, we can get Figure 3 The sample data curve is shown in red.
[0105] S3: Divide the sample data into training set and test set.
[0106] S4: Build an Invert-Transformer-LSTM neural network model with a built-in multi-head attention mechanism.
[0107] like Figure 2 As shown, in this embodiment, the Invert-Transformer-LSTM neural network model includes an Invert-Transformer module and an LSTM module. The Invert-Transformer module captures the global features and long-distance dependencies of sample data in the time series through a multi-head attention mechanism, establishes connections between time nodes and pays attention to important features. However, relying solely on global features is not enough to fully grasp the dynamic change process of sample data, so it is necessary to further process the sample data. The LSTM module processes the short-term and long-term dependencies of sample data in the time series by memorizing and utilizing the information of the previous time step, performs secondary learning on the global features extracted by the Invert-Transformer module, and integrates more time series information, thereby enhancing the time series correlation of the features. Finally, the model of the present invention converts the high-dimensional feature vectors extracted by the Invert-Transformer module and the LSTM module into predicted values through a fully connected layer.
[0108] More specifically, the above Invert-Transformer module includes an embedding layer, a multi-head attention mechanism layer, a feedforward network layer, and a normalization layer. The following is a detailed explanation of each layer of the Invert-Transformer module:
[0109] 1) Multi-head attention mechanism layer
[0110] The self-attention mechanism is a key feature of the Invert-Transformer module, which enables the model to perform weighted aggregation on parts of the input sequence. ),key( ) and value ( ), the model of the present invention can determine the importance of each position to other positions. After the attention scores are scaled, the value matrix is fused by weighted summation to generate the final output.
[0111] The weighted aggregation calculation formula of the multi-head attention mechanism is:
[0112] (3)
[0113] Formula (3) is obtained by The function is normalized to obtain the weight distribution. To prevent Q and The product of is too large, resulting in The value of tends to zero, causing the gradient to disappear, and a scaling factor needs to be introduced .
[0114] It should be explained that the multi-head attention mechanism is composed of multiple independent and parallel attention mechanisms, which can focus on the information of multiple subspaces, so that the attention mechanism can capture the multi-dimensional characteristics of the input data. Based on formula (3), the high-dimensional operation process of the multi-head self-attention mechanism includes:
[0115] (4)
[0116] (5)
[0117] in, Q To query data, K is the key value, V is the value corresponding to the query data, is a normalizing function, is the scaling factor, , and They all represent parameter matrices, For the i The time node j Head output, is the number of attention heads, is a function for concatenating strings. is a linear matrix, is the attention mechanism function; in formulas (4)-(5) , and The query ( ),key( ) and value ( ) is mapped to j Perform attention mechanism operation on each of them; For a certain time point No. j Head output; h is the number of attention heads; The function concatenates multiple attention heads together; is a linear matrix, and the combined matrix is compressed and transformed and output, as shown in detail Figure 4 Schematic diagram of the multi-head attention mechanism shown.
[0118] It should be explained that the goal of the multi-head attention mechanism is to process the input data in parallel through multiple attention heads, capturing the relationship between variables in the time series from different perspectives. In the Invert-Transformer module, time dynamics are considered by incorporating time information when embedding variable tokens and explicitly modeling the dependencies between time nodes in the multi-head attention mechanism. In this way, the model can not only pay attention to the multivariate information within each time node, but also capture the dynamic changes between time nodes, thereby improving the accuracy and robustness of the prediction.
[0119] 2) Feedforward network layer
[0120] The feedforward network layer usually consists of two fully connected layers and an activation function. Its purpose is to perform nonlinear transformation on sequence data and enhance the representation ability of the model. The specific expression formula of this layer is:
[0121] (6)
[0122] in, and is the weight parameter of the fully connected layer, and is the bias term.
[0123] 3) Normalization layer
[0124] The residual normalization layer includes residual connection and layer normalization. The residual connection essentially adds input features to the output features to prevent the model from experiencing gradient vanishing. The layer normalization function converts the input of each layer of neurons into a vector with the same mean and variance to reduce data deviation and prevent overfitting. The specific formula is:
[0125] (7)
[0126] in, and are network parameters, is the vector dot product, and Input The mean and variance of is a minimum constant.
[0127] More specifically, the LSTM module consists of multiple levels of LSTM units and fully connected layers. The following is a detailed explanation of the LSTM unit and the fully connected layer:
[0128] 1) LSTM unit
[0129] The LSTM unit consists of a forget gate, an input gate, and an output gate, which are used to update the cell state and retain historical information. The specific connection relationship between its components is as follows: Figure 5 shown.
[0130] 11) Forget Gate Usage and Description
[0131] The forget gate is used to and the current moment t The cell state at the previous time step determines the information that needs to be discarded. The expression of the forget gate is:
[0132] (8)
[0133] in, is the weight matrix of the forget gate; Forget the current moment of the gate t The intermediate state of the previous time step; is the input of the current cell; is the bias term of the forget gate; for Activation function.
[0134] 12) Input gate purpose and description
[0135] The input gate is used to select t To write new information to the cell state and update the cell state Integrating the selected information, the expression of the input gate is:
[0136] (9)
[0137] (10)
[0138] (11)
[0139] in, is the weight matrix of the input gate, is the input gate bias term; Formula (10) is used to describe the current input unit state, is the bias term of the input gate; Formula (11) is used to update the old cell state to the new cell state, is the input gate at the current moment t The intermediate state calculation results are: is the input gate at the current moment t The new cell state calculation result is is the weight matrix of the input gate; for Activation function, is the hyperbolic tangent function.
[0140] 13) Output gate purpose and description
[0141] The output gate is used to determine the current moment t The intermediate state The cell state information that should be included, the expression of the output gate is:
[0142] (12)
[0143] (13)
[0144] in, is the weight matrix of the output gate, is the bias term of the output gate, for Activation function.
[0145] 2) Fully connected layer
[0146] The fully connected layer receives the output vectors of the LSTM unit as input. These vectors contain the time series features processed by the LSTM layer. The fully connected layer maps these feature vectors to the output space through the weight matrix and bias term to obtain the final prediction result. The fully connected layer is calculated as:
[0147] (14)
[0148] in, is the weight matrix of the fully connected layer; is the bias vector of the fully connected layer; For the current moment t The intermediate state is passed from the LSTM module; is the activation function of the fully connected layer; is the final output prediction value.
[0149] S5: Optimize the hyperparameters of the Invert-Transformer-LSTM neural network model;
[0150] Specifically, the hyperparameters of the built Invert-Transformer-LSTM neural network model are optimized, mainly for the number of training rounds, learning rate, sliding window size, number of attention heads, and batch size, so as to make the model have a faster prediction speed while ensuring the accuracy of the model. The above hyperparameter optimization adopts the optimal learning strategy and stores the current optimal model.
[0151] S6: The training set and the test set are placed in the Invert-Transformer-LSTM neural network model after hyperparameter optimization for training and testing, and the test results are obtained and the performance of the test results is evaluated.
[0152] Specifically, the above step S6 includes:
[0153] S61: The training set is placed in the Invert-Transformer-LSTM neural network model after hyperparameter optimization for training, and then the trained Invert-Transformer-LSTM neural network model is used to predict IGBT faults on the test set to obtain test results.
[0154] The loss function formula used in the training process in step S61 is:
[0155] (15)
[0156] in, is the number of samples, is the true value of the sample data, is the predicted value of the sample data.
[0157] S62: Perform performance evaluation on the test results according to the evaluation indicators.
[0158] In this embodiment, the above evaluation indicators include root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R2). The following is a detailed description of the three evaluation indicators:
[0159] 1) Root Mean Square Error (RMSE)
[0160] RMSE calculates the squared mean of the error, which reflects the stability of the model. The smaller the RMSE, the lower the deviation of the model prediction, indicating that the model has good prediction accuracy. The calculation formula of RMSE is as follows:
[0161] (16)
[0162] in, is the actual value, is the predicted value, The predicted number.
[0163] 2) Mean absolute error
[0164] MAE calculates the average absolute value of the error, which can avoid the problem of errors canceling each other out and provide a direct measure of the actual prediction error of the model. The smaller the MAE, the smaller the prediction error of the model, indicating that the model is more accurate. The calculation formula of MAE is as follows:
[0165] (17)
[0166] 3) Coefficient of determination
[0167] Coefficient of determination Characterizes the degree of correlation between the predicted value and the actual value. The value is between 0 and 1. The closer the value is to 1, the better the model fit is. The calculation formula is as follows:
[0168] (18)
[0169] in, Represents the mean of the actual values.
[0170] S7: Repeat steps S5-S6 to obtain an optimal prediction model, and use the optimal prediction model to predict the actual collector-emitter peak voltage of the target IGBT in real time to obtain a real-time prediction result.
[0171] Specifically, step S7 includes:
[0172] S71: Repeat steps S5-S6 to obtain hyperparameters of the number of training rounds, learning rate, number of attention heads, and batch size, and use them as prediction configuration parameters of the Invert-Transformer-LSTM neural network model;
[0173] S72: Select the optimal prediction parameter among the prediction configuration parameters as the actual configuration hyperparameter of the Invert-Transformer-LSTM neural network model to obtain the optimal prediction model;
[0174] S73: Use the optimal prediction model to predict the actual collector-emitter peak voltage of the target IGBT in real time, and after obtaining the real-time prediction result, determine whether the target IGBT is faulty based on the characteristic parameters and the real-time prediction result.
[0175] Based on the above steps S1-S7, the Invert-Transformer-LSTM neural network model is compared with other existing models.
[0176] The LSTM, Transformer, and Invert-Transformer (iTransformer for short) models were selected as comparison models, and the IGBT degradation state prediction results of the Invert-Transformer-LSTM (iTransformer-LSTM for short) model of the present invention were compared. In order to avoid overfitting or underfitting of the model during training, the model parameters and test environment were set according to factors such as the data set size and data characteristics. The specific test environment parameters are detailed in Table 1, and the model parameter settings are referred to Table 2.
[0177] Table 2
[0178]
[0179] Based on Table 1 and Table 2, combined with the specific properties of each model, for the same target IGBT, after training each model with the training set, the prediction performance of the model was evaluated using the test set data. Figure 6-Figure 9 The specific evaluation indicators are shown in Table 3.
[0180] Table 3
[0181]
[0182] Compared with other deep learning models, the iTransformer-LSTM model has obvious advantages in prediction accuracy and stability. Its prediction residual is smaller, and it is more consistent with the changes in measured data at the deformation extreme points and local fluctuation points. In addition, unlike the LSTM model, there is no time-lag offset between the predicted value and the actual value of the iTransformer-LSTM, which effectively weakens the lag effect of the LSTM model. This shows that after integrating the attention mechanism and extracting time series information, the iTransformer-LSTM model can better mine the time series information in the data. At the same time, the iTransformer-LSTM model coupled with LSTM has a significant decrease in residuals compared to the iTransformer model, indicating that after deep processing of the LSTM layer, the iTransformer-LSTM model has a stronger ability to capture time series features and better prediction results.
[0183] Table 3 shows the prediction effect evaluation indicators of the four models at different prediction lengths. The MAE of the iTransformer-LSTM model is reduced by 53.13%, 20.12%, and 8.78% compared with the LSTM, Transformer, and iTransformer models, and the RMSE is reduced by 50.31%, 22.97%, and 14.36%, respectively. An increase of 13.17%, 2.64% and 1.41% respectively.
[0184] In summary, the iTransformer-LSTM model has higher accuracy, smaller error, and improved stability than traditional deep learning models.
[0185] Embodiment 2
[0186] like Fig.10 As shown, based on the first embodiment, this embodiment discloses an IGBT fault prediction system based on an inverted Transformer network, the system comprising:
[0187] The characteristic parameter selection module is used to select the collector-emitter peak voltage in the IGBT aging data as the characteristic parameter.
[0188] The sample data acquisition module is used to collect the collector-emitter voltage signal of the target IGBT, extract the transient peak voltage, and filter the transient peak voltage using a quadratic exponential smoothing filter function to obtain sample data.
[0189] The sample data partitioning module is used to divide the sample data into a training set and a test set.
[0190] Model building module, used to build the Invert-Transformer-LSTM neural network model with built-in multi-head attention mechanism.
[0191] Hyperparameter optimization module, used to optimize the hyperparameters of the Invert-Transformer-LSTM neural network model.
[0192] The performance evaluation module is used to place the training set and the test set into the Invert-Transformer-LSTM neural network model after hyperparameter optimization for training and testing, obtain the test results and perform performance evaluation on the test results.
[0193] The fault prediction module is used to repeatedly execute steps S5-S6 to obtain an optimal prediction model, and use the optimal prediction model to predict the actual collector-emitter peak voltage of the target IGBT in real time to obtain a real-time prediction result.
[0194] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above-mentioned method embodiment; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk, etc. Various media that can store program codes.
[0195] For those skilled in the art, various other corresponding changes and deformations can be made according to the technical solutions and concepts described above, and all of these changes and deformations should fall within the protection scope of the claims of the present invention.
Claims
1. An IGBT fault prediction method based on an inverted Transformer network, characterized in that: include: S1: Select the collector-emitter peak voltage in the IGBT aging data as the characteristic parameter; S2: collecting the collector-emitter voltage signal of the target IGBT, extracting the transient peak voltage, and filtering the transient peak voltage using a quadratic exponential smoothing filter function to obtain sample data; S3: Divide the sample data into a training set and a test set; S4: Build an Invert-Transformer-LSTM neural network model with a built-in multi-head attention mechanism; S5: Optimize the hyperparameters of the Invert-Transformer-LSTM neural network model; S6: placing the training set and the test set into the Invert-Transformer-LSTM neural network model after hyperparameter optimization for training and testing, obtaining test results and performing performance evaluation on the test results; S7: Repeat steps S5-S6 to obtain an optimal prediction model, and use the optimal prediction model to predict the actual collector-emitter peak voltage of the target IGBT in real time to obtain a real-time prediction result; Wherein, the Invert-Transformer-LSTM neural network model includes: Invert-Transformer module, which is used to capture the global features and long-distance dependencies of sample data in time series through a multi-head attention mechanism, establish connections between time nodes and focus on important features to obtain the first high-dimensional feature vector; The LSTM module is used to perform secondary learning on the global features extracted by the Invert-Transformer module, enhance the temporal correlation of the global features, and obtain a second high-dimensional feature vector; the LSTM module is built with a fully connected layer, and the fully connected layer is used to convert the first high-dimensional feature vector and the second high-dimensional feature vector extracted by the Invert-Transformer module and the LSTM module into a prediction result; The Invert-Transformer module includes: The embedding layer is used to map high-dimensional input data to a low-dimensional space to obtain an input sequence; A multi-head attention mechanism layer, used to perform weighted aggregation on each part of the input sequence and output a compressed transformation result; A feed-forward network layer, consisting of two fully connected layers and an activation function, for performing a nonlinear transformation on the input sequence; Normalization layer, with built-in residual connection function and layer standardization function; The expression of the multi-head attention mechanism layer is: （3） （4） （5） in, Q To query data, K is the key value, V is the value corresponding to the query data, is a normalizing function, is the scaling factor, They all represent parameter matrices, For a certain time point i No. j Head output, h is the number of attention heads, is a function for concatenating strings. is a linear matrix, is the attention mechanism function; Specifically, step S7 includes: S71: Repeat steps S5-S6 to obtain hyperparameters of the number of training rounds, learning rate, number of attention heads, and batch size, and use them as prediction configuration parameters of the Invert-Transformer-LSTM neural network model; S72: Select the optimal prediction parameter among the prediction configuration parameters as the actual configuration hyperparameter of the Invert-Transformer-LSTM neural network model to obtain the optimal prediction model; S73: Use the optimal prediction model to predict the actual collector-emitter peak voltage of the target IGBT in real time, and after obtaining the real-time prediction result, determine whether the target IGBT is faulty based on the characteristic parameters and the real-time prediction result.
2. The IGBT fault prediction method based on the inverted Transformer network according to claim 1 is characterized in that: The quadratic exponential smoothing filter function in step S2 is specifically: （1） （2） in, For the time node i The steady-state quantity of the transient peak voltage; For the time node The actual value of the upper transient peak voltage; the first smoothing coefficient A In the interval [0, 1], A When the value is close to 0, Get the previous time node i-1 The smoothing result of A When the value is close to 1, it is retained Original feature information; is the trend value of transient peak voltage after double exponential smoothing; the second smoothing coefficient C The value directly affects the smoothed trend.
3. The IGBT fault prediction method based on inverted Transformer network according to claim 1, characterized in that: The expression of the feedforward network layer is: (6) in, and is the weight parameter of the fully connected layer, and is the bias term, X is the input value.
4. The IGBT fault prediction method based on the inverted Transformer network according to claim 3 is characterized in that: The expression of the normalization layer is: (7) in, and are network parameters, is the vector dot product, and Input X The mean and variance of is a minimum constant.
5. The IGBT fault prediction method based on inverted Transformer network according to claim 1, characterized in that: The LSTM module is composed of multiple levels of LSTM units and a fully connected layer, and the LSTM unit includes a forget gate, an input gate and an output gate; The forget gate is used to and the current moment t The cell state at the previous time step determines the information that needs to be discarded; the expression of the forget gate is: (8) in, is the weight matrix of the forget gate; Forget the current moment of the gate t The intermediate state of the previous time step; is the input of the current cell; is the bias term of the forget gate; for Activation function; The input gate is used to select the t To write new information to the cell state and update the cell state Integrating the selected information, the expression of the input gate is: (9) (10) (11) in, is the weight matrix of the input gate, is the input gate bias term; Formula (10) is used to describe the current input unit state, is the bias term of the input gate; Formula (11) is used to update the old cell state to the new cell state, is the input gate at the current moment t The intermediate state calculation results are: is the input gate at the current moment t The new cell state calculation result is is the weight matrix of the input gate; for Activation function, tanh is the hyperbolic tangent function; The output gate is used to determine the current moment t The intermediate state The cell state information should be included, and the expression of the output gate is: （12） （13） in, is the weight matrix of the output gate, is the bias term of the output gate, for Activation function; The expression of the fully connected layer is: （14） in, is the weight matrix of the fully connected layer; is the bias vector of the fully connected layer; For the current moment t The intermediate state; is the activation function of the fully connected layer; is the final output prediction value.
6. The IGBT fault prediction method based on inverted Transformer network according to claim 1, characterized in that: The step S6 comprises: S61: placing the training set into the Invert-Transformer-LSTM neural network model after hyperparameter optimization for training, and then using the trained Invert-Transformer-LSTM neural network model to perform IGBT fault prediction on the test set to obtain a test result; S62: Perform performance evaluation on the test results according to the evaluation indicators; The loss function formula used in the training process in step S61 is: （15） in, n is the number of samples, is the true value of the sample data, is the predicted value of the sample data; the evaluation indicators include root mean square error, mean absolute error and determination coefficient.
7. An IGBT fault prediction system based on an inverted Transformer network, characterized in that: The IGBT fault prediction method based on the inverted Transformer network according to any one of claims 1 to 6 is adopted, and the system comprises: A characteristic parameter selection module is used to select the collector-emitter peak voltage in the IGBT aging data as a characteristic parameter; A sample data acquisition module is used to collect the collector-emitter voltage signal of the target IGBT, extract the transient peak voltage, and filter the transient peak voltage using a quadratic exponential smoothing filter function to obtain sample data; A sample data division module, used for dividing the sample data into a training set and a test set; Model building module, used to build the Invert-Transformer-LSTM neural network model with built-in multi-head attention mechanism; A hyperparameter optimization module, used for optimizing the hyperparameters of the Invert-Transformer-LSTM neural network model; A performance evaluation module, used for placing the training set and the test set into the Invert-Transformer-LSTM neural network model after hyperparameter optimization for training and testing, obtaining test results and performing performance evaluation on the test results; A fault prediction module, used to repeatedly execute the built-in steps of the hyperparameter optimization module and the performance evaluation module to obtain an optimal prediction model, and use the optimal prediction model to predict the actual collector-emitter peak voltage of the target IGBT in real time to obtain a real-time prediction result; Wherein, the Invert-Transformer-LSTM neural network model includes: Invert-Transformer module, which is used to capture the global features and long-distance dependencies of sample data in time series through a multi-head attention mechanism, establish connections between time nodes and focus on important features to obtain the first high-dimensional feature vector; The LSTM module is used to perform secondary learning on the global features extracted by the Invert-Transformer module, enhance the temporal correlation of the global features, and obtain a second high-dimensional feature vector; the LSTM module is built with a fully connected layer, and the fully connected layer is used to convert the first high-dimensional feature vector and the second high-dimensional feature vector extracted by the Invert-Transformer module and the LSTM module into a prediction result; The Invert-Transformer module includes: The embedding layer is used to map high-dimensional input data to a low-dimensional space to obtain an input sequence; A multi-head attention mechanism layer, used to perform weighted aggregation on each part of the input sequence and output a compressed transformation result; A feed-forward network layer, consisting of two fully connected layers and an activation function, for performing a nonlinear transformation on the input sequence; Normalization layer, with built-in residual connection function and layer standardization function; The expression of the multi-head attention mechanism layer is: （3） （4） （5） in, Q To query data, K is the key value, V is the value corresponding to the query data, is a normalizing function, is the scaling factor, They all represent parameter matrices, For a certain time point i No. j Head output, h is the number of attention heads, is a function for concatenating strings. is a linear matrix, is the attention mechanism function; Specifically, the built-in steps of the fault prediction module include: S71: Repeat the built-in steps of the hyperparameter optimization module and the performance evaluation module to obtain hyperparameters of the number of training rounds, learning rate, number of attention heads, and batch size, and use them as prediction configuration parameters of the Invert-Transformer-LSTM neural network model; S72: Select the optimal prediction parameter among the prediction configuration parameters as the actual configuration hyperparameter of the Invert-Transformer-LSTM neural network model to obtain the optimal prediction model; S73: Use the optimal prediction model to predict the actual collector-emitter peak voltage of the target IGBT in real time, and after obtaining the real-time prediction result, determine whether the target IGBT is faulty based on the characteristic parameters and the real-time prediction result.