Inverter IGBT intelligent fault diagnosis method based on multi-scale attention network

Abstract

The invention discloses an inverter IGBT intelligent fault diagnosis method based on a multi-scale attention network, and relates to the technical field of power electronic equipment fault diagnosis. A three-phase current signal of a three-level NPC inverter is collected and preprocessed; multi-dimension fault features of the current signals are extracted in parallel through multi-scale depth separable convolution; a scale-aware efficient attention mechanism is introduced to carry out weighted optimization on each scale feature, and key fault information is strengthened; and outputting a fault diagnosis result through a feature fusion and classification network, and performing INT8 quantization processing on the model to adapt to an embedded deployment scene. The method has the advantages of being high in multi-scale feature capturing capacity, high in diagnosis precision and light in model weight, and can meet the dual requirements of industrial application for real-time performance and accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power electronic device fault diagnosis, and particularly relates to an intelligent fault diagnosis method for IGBT of a three-level NPC inverter based on a multi-scale attention network. BACKGROUND

[0002] In modern high-performance motor drive systems, the running reliability of insulated gate bipolar transistors (IGBT) as core power switching devices directly determines the safety and stability of the entire system. Especially in three-level NPC (neutral point clamped) inverters, IGBT modules work in harsh conditions such as high temperature and high voltage for a long time, and are prone to typical faults such as bond wire falling off, chip open circuit, and drive circuit failure due to the cumulative effects of electrical and thermal stress. Such open circuit faults have the characteristics of strong concealment and rapid development, and if they are not diagnosed in time and accurately, they will cause serious distortion of the current waveform, output torque fluctuation, and even system shutdown, which is particularly serious in application scenarios such as electric vehicles, rail transit, and industrial frequency converters that have extremely high safety requirements. Therefore, developing fast and accurate IGBT fault diagnosis technology has become a key problem to be solved in the field of power electronics.

[0003] At present, existing diagnostic methods mainly follow two technical routes: model-driven and data-driven. Model-driven methods, such as those based on state observers or sliding mode variable structure control, rely heavily on the accurate mathematical model of the controlled object. However, in actual industrial environments, there are uncertain factors such as parameter drift and external disturbances in inverter systems, which can easily lead to model mismatch, causing misdiagnosis or missed diagnosis of the diagnostic system, greatly limiting its engineering applicability. With the development of artificial intelligence technology, data-driven intelligent diagnosis methods have shown great potential. For example, using convolutional neural networks (CNN) or long short-term memory networks (LSTM) can automatically learn fault features from historical data, to some extent, overcoming the dependence on accurate models. However, such deep models usually have complex structures, high computational load, and large memory occupation, making it difficult to realize real-time deployment and operation on embedded hardware platforms that are strictly limited in terms of computing resources, power consumption, and response speed. In addition, existing intelligent diagnosis methods often have a single feature extraction level, failing to fully utilize the feature differences exhibited by fault signals at different time scales, and lacking effective filtering and strengthening mechanisms for key information when multiple features are fused, resulting in insufficient discrimination for some similar fault patterns.

[0004] In summary, under the current technical background, there is an urgent need for a new IGBT intelligent fault diagnosis solution that can balance high diagnostic accuracy, strong generalization ability, low computational complexity, and be suitable for industrial embedded scenarios. SUMMARY

[0005] This invention aims to address a series of problems in existing inverter IGBT fault diagnosis technologies, such as model-driven diagnostic methods being susceptible to field interference and misjudgment due to reliance on precise mathematical models, data-driven methods being difficult to deploy in real time on embedded platforms due to high model computational complexity, and traditional feature extraction methods failing to utilize multi-scale fault information and lacking effective key feature enhancement mechanisms, resulting in limited ability to distinguish similar faults.

[0006] To solve the above problems, the following technical solution is adopted: The intelligent fault diagnosis method for inverter IGBTs based on multi-scale attention networks includes the following steps: S1. Acquire the three-phase current signal of the three-level NPC inverter, preprocess the three-phase current signal to obtain standardized current samples; S2. The fault features of the current sample are initially extracted using ordinary convolution kernels. Then, multiple sets of depthwise separable convolutions of different scales are used to process the current sample in parallel to extract multi-scale fault features and output feature vectors of each branch. S3. Introduce a scale-aware, efficient attention mechanism to calculate the channel weights of the feature vectors of each branch, and achieve multi-scale feature weighting optimization by multiplying the channel weights with the multi-scale features channel by channel; S4. The weighted optimized features of each branch are fused, and after batch standardization and activation function processing, the features are compressed by global average pooling and then input into the classification network to obtain the fault diagnosis results. S5. The multi-scale attention network model is trained using the training set, the obtained model parameters are quantized using INT8, and the model performance is tested using the test set.

[0007] Preferably, in step S1, the three-phase current signal includes the current signal under normal operating conditions of the three-level NPC inverter and the current signal under open-circuit fault conditions of various power devices.

[0008] Preferably, the method for preprocessing the three-phase current signal in step S1 includes: The acquired signal is linearly normalized using the following formula: ; In the formula, For the first The first phase current Original values ​​of each sampling point , The first Minimum and maximum values ​​of phase current samples, The value is the normalized value; The normalized sequence is resampled at intervals of [number]. for: ; In the formula, The original sampling frequency of the data. The resampling frequency, For the corresponding interval number; The resampling frequency satisfies: , The number of normalized data points per unit fundamental period in the generated new sequence. This refers to the fundamental frequency of the NPC inverter's output. The current signal is segmented using a sliding window to obtain data segments after resampling.

[0009] Preferably, in step S2, the multiple sets of depthwise separable convolutions at different scales include at least three different scales of depthwise separable convolutions, and the kernel scale satisfies the following conditions: ,in ; For scale The depthwise separable convolution process is as follows: ; In the formula, In , For the input current signal Channel 1 One sampling point, For the first Scale-based depthwise convolution kernel Channel 1 Each weight, This is the output of the depthwise convolution at this scale; After pointwise convolution to fuse channel information, the output is the first... Scaled eigenvectors: ; In the formula, For the first Pointwise convolution kernel weights at different scales For the first Scale-based branch feature output, The channel index for the output feature. The channel index for the input feature. This indicates the number of channels for the input feature.

[0010] Preferably, in step S3, the computation process of the scale-aware efficient attention mechanism includes: Branch feature mapping of input Perform global average pooling on a channel-by-channel basis to obtain the global information description vector for each channel. The calculation formula is as follows: ; In the formula, It is the first In the feature mapping of the nth branch, the th Channel 1 Feature values ​​at each position, The length of the feature map. For the first The first branch The result after global average pooling of the channels; Using a convolution kernel size of One-dimensional convolution is used to capture local interaction information between channels, and then... The activation function maps the results to a range of 0 to 1, obtaining the weights of each channel. The calculation formula is as follows: ; In the formula, Indicates the kernel size as One-dimensional convolution operation, for Activation function It is the first The first branch Channel weights; The obtained weights Feature mapping with the original branch Multiplying each channel sequentially enhances the response of critical channels; the calculation formula is as follows: ; In the formula, It is the first After weight adjustment, each branch Channel 1 The feature values ​​at each position.

[0011] Preferably, the kernel size Determined through an adaptive function: ; In the formula, This indicates taking the nearest odd number; hyperparameter and Set as , This function ensures that the convolution kernel can cover all effective local interactions between channels.

[0012] Preferably, the feature fusion in step S4 employs a splicing operation, splicing the weighted feature maps from different branches along the channel dimension.

[0013] Preferably, the calculation formula for batch standardization in S4 is: ; In the formula, For the fusion feature Channel 1 The value at each position, , These are the mean and variance of the channel, respectively. , For learnable parameters, To prevent tiny values ​​from being divided by zero, This is the value after batch standardization.

[0014] Preferably, in step S4, the classification network... The function outputs the failure probability, and the calculation expression is: ; In the formula, For input data, The length of the input data.

[0015] Preferably, in step S5, the formula for INT8 quantization is: ; ; in, and The calculation formula is as follows: ; ; In the formula, , These are the maximum and minimum values ​​of the parameter, respectively. Scaling factor To quantize the zero point, For the original 32-bit floating-point number, It is an 8-bit fixed-point number after quantization.

[0016] Compared with existing technologies, this invention firstly uses a multi-scale depthwise separable convolutional structure to extract fault features in parallel. This design can capture fault information at all scales, from local transients to global trends, while leveraging the inherent advantages of low parameter count and low computational complexity of depthwise separable convolutions. This significantly reduces the computational burden from the source of the model, laying a solid foundation for achieving lightweight design.

[0017] Building upon this foundation, the introduced Scale-Aware Efficient Attention (ECA) mechanism plays a crucial role in accurately optimizing feature quality. It adaptively evaluates and strengthens the channel features most relevant to the fault in each scale branch, effectively suppressing irrelevant noise interference, thereby significantly improving the model's ability to represent features and its discriminative power against similar fault patterns. Furthermore, the fusion strategy of "attention-weighted first, feature concatenation later" ensures that each feature has been purified before fusing information from different scales. This makes the subsequent fused features more discriminative. Combined with batch normalization and activation function processing, the training process is further stabilized and the model's nonlinear fitting ability is enhanced.

[0018] Finally, by performing INT8 quantization on the trained model, the model parameters were converted from 32-bit floating-point numbers to 8-bit fixed-point numbers. This operation significantly reduced the model size and computational resource consumption during inference, ultimately solving the bottleneck problem of high-precision deep learning models running in real time on resource-constrained embedded devices.

[0019] Therefore, this solution ultimately achieves an excellent balance between diagnostic accuracy, model efficiency, and deployment adaptability. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0021] Figure 1 This is a general framework diagram of the diagnostic method of the present invention; Figure 2 This is a schematic diagram of the three-level NPC inverter topology for the diagnostic method of the present invention; Figure 3 This is a simulation model diagram of a three-level NPC inverter for the diagnostic method of this invention; Figure 4 The figure shows the simulation results of the three-level NPC inverter model of the diagnostic method of this invention, wherein: Figure 4 (a) shows the output of the three-phase current under normal conditions; Figure 4 (b) in the text is T a1 The output of the three-phase current under open-circuit conditions; Figure 4 (c) in the text represents T. a2 The output of the three-phase current under open-circuit conditions; Figure 5 This is a structural diagram of the diagnostic method model of the present invention; Figure 6 This is the loss function curve of the diagnostic method of the present invention; Figure 7 This is the average accuracy curve of the diagnostic method of the present invention. Detailed Implementation

[0022] This invention provides an intelligent fault diagnosis method for inverter IGBTs based on multi-scale attention networks. The specific embodiments of this invention will be described in detail below with reference to the accompanying drawings. It should be understood that these embodiments are only used to help those skilled in the art better understand the core concept of this invention and do not constitute any limitation on the scope of protection of this invention.

[0023] like Figure 1 As shown, the diagnostic method provided in this embodiment mainly includes five core steps: data acquisition and preprocessing, multi-scale feature extraction, attention-weighted optimization, feature fusion and classification, and model quantization deployment. Specifically, the diagnostic method first collects the three-phase current signals of a three-level NPC inverter through a current sensor. These signals then enter the preprocessing stage for standardization. The preprocessed data is first subjected to preliminary feature extraction through ordinary convolutional layers, and then input in parallel into multiple depthwise separable convolutional branches of different scales to extract rich multi-scale fault features. The features extracted by each branch are then weighted and optimized through a scale-aware, efficient attention mechanism to highlight key fault information. The weighted multi-scale features are fused by concatenation and then processed by batch standardization, activation function, and global average pooling before finally being input into the classification network to complete the identification of fault types. The trained model will also undergo INT8 quantization to adapt to the deployment requirements of embedded devices.

[0024] Compared with existing fault diagnosis methods based on standard CNNs (typically 1.03M parameters, computational complexity of 11.38MFLOPs, memory usage of 0.25MB, and accuracy of approximately 96%) or general lightweight networks such as MobileNetV3 (378K parameters, computational complexity of 4.05MFLOPs, and accuracy of only 95% in NPC inverter fault diagnosis tasks), this invention features a dedicated architecture designed for the fault characteristics of NPC inverters. It effectively captures fault features at different time scales through multi-scale depth separable convolutional kernels and strengthens key feature channels by combining channel attention mechanisms. It achieves 99% accuracy with a 51% reduction in parameters, demonstrating the customization advantages for specific application scenarios.

[0025] like Figure 2 The diagram shows the topology of an NPC three-level inverter circuit. This circuit mainly includes a DC power supply and 12 IGBT switching transistors T. a1 To T c4It consists of 12 anti-parallel diodes and two DC-side supporting capacitors C1 and C2. During implementation, we built a system in the MATLAB / Simulink environment, such as... Figure 3 The simulation model shown is used to obtain training and testing data. The model includes a drive pulse generation module, an NPC main circuit module, and a load module.

[0026] Table 1: Parameter information used for fault diagnosis model simulation As shown in Table 1 above, during the data acquisition process, the DC side voltage U dc The voltage was set to 100V, and the load consisted of a 15Ω resistor and a 5mH inductor connected in series. Space Vector Pulse Width Modulation (SVPWM) was used for modulation. To simulate actual fault conditions, we simulated an open-circuit fault by removing the drive signal of a specific IGBT. Data was collected for 14 operating conditions, including the normal state. The specific fault type correspondences are shown in Table 2 below. For example, label 1 corresponds to T... a1 Open circuit fault, tag 2 corresponds to T a2 Open circuit faults, and so on.

[0027] Table 2: Fault Types In the preprocessing stage, the acquired raw current signal is first subjected to linear normalization. The normalization formula is as follows: In the formula, For the first The first phase current Original values ​​of each sampling point , The first Minimum and maximum values ​​of phase current samples, This is the normalized value.

[0028] Unlike traditional multi-scale attention networks that typically only process data through normalization, this embodiment further introduces a resampling design, significantly reducing the size of the input data and effectively lowering the subsequent computational load. After normalization, the sequence is resampled at intervals of [insert interval here]. The resampling frequency satisfies: , The number of normalized data points per unit fundamental period in the generated new sequence. This is the fundamental frequency of the NPC inverter's output. For example, when the original sampling frequency... The output fundamental frequency is 10kHz. At 50Hz, if we expect each fundamental frequency cycle to contain 250 data points, then the resampling frequency is... It should be set to 12.5kHz, corresponding to The value is 0.8. After resampling, the current signal is segmented using a sliding window with a window length of 250 points (one fundamental cycle) and a sliding step size of 10 points to generate a large number of training and test samples.

[0029] Therefore, the data preprocessing method in this embodiment eliminates the influence of signal amplitude differences through linear normalization, ensures the uniformity of data length through resampling, and increases the number of samples through sliding window segmentation. These processes not only lay a good foundation for subsequent feature extraction, but also effectively improve the generalization ability of the model.

[0030] like Figure 5 As shown, the preprocessed current samples are first subjected to preliminary feature extraction through a conventional convolutional layer. This design not only performs preliminary feature extraction, but more importantly, it can effectively suppress or filter out residual environmental noise and irrelevant interference in the signal, while enhancing the significance of primary features related to faults, providing cleaner and more prominent input for subsequent complex multi-scale analysis.

[0031] Optionally, the convolutional layer uses a 1×3 kernel with a stride of 2, followed by a batch normalization layer and a ReLU activation function, and then downsampled through a 1×2 max pooling layer. Subsequently, the data is input in parallel to three depthwise separable convolutional branches of different scales, which can use kernel sizes of 1×3, 1×5, and 1×9 respectively, corresponding to the following formula: K1 , K2 and K3 .

[0032] Each depthwise separable convolutional branch contains two steps: First, depthwise convolution is performed, and the calculation process is as follows: In the formula, In , For the input current signal Channel 1 One sampling point, For the first Scale-based depthwise convolution kernel Channel 1 Each weight, This is the output of the depthwise convolution at this scale; Then perform pointwise convolution, calculated using the following formula: In the formula, For the first Pointwise convolution kernel weights at different scales For the first Scale-based branch feature output, The channel index for the output feature. The channel index for the input feature. This indicates the number of channels in the input features. This embodiment uses depthwise separable convolution instead of ordinary convolution kernels in parallel multi-scale convolution, efficiently completing the aforementioned multi-scale feature extraction task with extremely low parameter count and computational cost.

[0033] With this design, smaller-scale convolutional kernels can be achieved. K1 Primarily captures localized transient fault characteristics of current signals, at a medium scale. K2 Capturing transitional phase characteristics, on a larger scale K3 It focuses on global trend characteristics, thereby achieving full-scale coverage of inverter fault characteristics.

[0034] like Figure 5 As shown, the feature map output from each depthwise separable convolution branch is input into the corresponding scale-aware efficient attention module for feature optimization. This scale-aware efficient attention mechanism weights the massive features extracted by multi-scale depthwise separable convolutions, automatically identifying and strengthening the most critical scale and spatial location features for current fault diagnosis, while weakening unimportant or redundant features.

[0035] This module first maps the branch features of the input. Perform global average pooling, the calculation formula is as follows: ,here The length of the feature map.

[0036] Then, a convolution kernel of size... One-dimensional convolution is used to capture local interaction information between channels, and then... The activation function yields the weight values ​​for each channel, calculated using the following formula: , Indicates the kernel size as One-dimensional convolution operation, for Activation function It is the first The first branch Channel weights.

[0037] Finally, the obtained weights Feature mapping with the original branch Channel-by-channel multiplication is used to achieve feature recalibration; the calculation formula is as follows: .

[0038] Kernel size in the attention module It is determined by an adaptive function, and the specific formula is as follows: In the formula, This indicates taking the nearest odd number; hyperparameter and Set as , This function ensures that the convolution kernel can cover all effective local interactions between channels. This adaptive mechanism ensures that the kernel size can be automatically adjusted according to changes in the number of channels, thereby effectively capturing local correlations between channels.

[0039] Therefore, by independently calculating channel weights for different scale branches, specific optimization of multi-scale features is achieved. Compared with the global attention mechanism, this scale-aware design can better preserve the unique features of each scale branch while suppressing interference from irrelevant noise.

[0040] like Figure 5 As shown, the feature maps at three scales are attention-weighted. , , The components will be merged through a splicing operation to form a fused feature. Specifically, it is expressed as ,in, This indicates concatenation along the channel dimension (the last dimension). If the number of channels in each branch is... c The number of channels after fusion is 3. c For example, if each branch has 64 feature map channels, then the fused feature map will have 192 channels.

[0041] Subsequently, the fusion features underwent batch standardization, and their calculation formula is as follows: In the formula, For the fusion feature Channel 1 The value at each position, , These are the mean and variance of the channel, respectively. , For learnable parameters, To prevent tiny values ​​from being divided by zero, it can usually be set to 0.001. This is the value after batch standardization.

[0042] After batch standardization, a nonlinear transformation is introduced using the ReLU activation function. Then, a global average pooling layer is used to compress the spatial information of each channel into a scalar value, ultimately generating a compact global feature vector.

[0043] This global feature vector is input into a fully connected classification layer, through... The function outputs the probability of failure. The calculation formula is In the formula, For input data, The length of the input data.

[0044] The number of output neurons in a fully connected layer is the same as the number of fault categories, and the output value of each neuron represents the predicted probability of the corresponding fault category.

[0045] Therefore, by employing a fusion strategy of first weighting features by attention and then concatenating them, the uniqueness of features at each scale is preserved, while cross-scale feature interaction enhancement is achieved. Combined with batch normalization and global average pooling operations, the discriminative expressive power of features and the generalization performance of the model are effectively improved.

[0046] Optionally, model training can use a training set and a test set randomly divided in a 2:1 ratio, with a training period of 300 epochs and a batch size of 32. During training, changes in the loss function and accuracy are recorded, such as... Figure 6 and Figure 7 As shown, the loss function decreases steadily with the increase of training cycles, while the accuracy increases synchronously. Furthermore, the model performs stably on the validation set, indicating that there is no overfitting phenomenon.

[0047] After training, the model parameters are quantized using INT8 to adapt for embedded deployment. Finally, the trained floating-point model (usually FP32) is quantized using INT8, significantly reducing the space occupied by the model during storage and the computational resources during inference, thereby greatly improving the inference speed on embedded hardware. Simultaneously, because the preceding steps (including resampling, noise suppression, depthwise separable convolution, and attention weighting) have already greatly optimized the model structure and computational cost, the model itself is insensitive to quantization operations, ensuring that there is almost no significant loss of accuracy after quantization. The quantization formula is: ; ; in, and The calculation formula is as follows: ; ; here and These represent the maximum and minimum values ​​of the parameter, respectively. This indicates the rounding operation. For the original 32-bit floating-point number, It is an 8-bit fixed-point number after quantization.

[0048] During the quantization process, we employ a symmetric quantization strategy, which determines the appropriate scaling factor by statistically analyzing the distribution range of weights and activation values ​​at each layer. and zero point .

[0049] Table 3 below shows a partial comparison of the kernel parameters of a certain layer in the FP32 format. It can be seen that the numerical changes are minimal, proving that the accuracy loss after quantization is negligible. Table 3: FP32 Parameter Examples The quantized model can be saved in TFLite format, facilitating loading and inference on embedded devices. Compared to traditional diagnostic schemes based on spectral analysis or standard deep convolutional networks, this invention achieves an optimal balance in terms of parameter count, diagnostic accuracy, and model size, thereby meeting the industrial application requirements for real-time performance, accuracy, and low resource consumption.

[0050] To verify the impact of quantization on model performance, we compared the performance of the model before and after quantization on the test set. The results show that the INT8 quantized model maintains high diagnostic accuracy while significantly reducing model size and computational load, making real-time fault diagnosis possible on embedded devices.

[0051] The model training and quantization deployment scheme in this embodiment ensures the convergence and generalization ability of the model through a carefully designed training strategy, and significantly reduces the model size and computational load through INT8 quantization technology. This combined approach successfully solves the challenge of deploying deep learning models on resource-constrained embedded devices, providing a feasible path for practical applications in industrial settings.

[0052] Figure 4 Figure (a) shows the output waveform of the three-phase current under normal conditions. It can be seen that i a i b i c The three-phase current exhibits standard sinusoidal waveform characteristics with equal amplitude and a phase difference of 120°, and the waveform is symmetrical and smooth. Figure 4 (b) shows T a1 In the open-circuit fault state, the current waveform clearly shows severe distortion in the positive half-cycle of the A-phase current, with obvious waveform loss. Figure 4 (c) shows T a2 The current waveform under open-circuit fault conditions also shows distortion during the positive half-cycle of the A-phase current, but it is different from that under T. a1 The waveform characteristics of faults exhibit subtle differences. These highly discriminative current waveform characteristics provide a reliable basis for fault diagnosis algorithms and are also the physical foundation for the high-precision diagnosis achieved by the method of this invention.

[0053] To systematically evaluate the performance of the method of this invention, we conducted tests under various operating conditions, including different modulation intensities from 0.3 to 0.9 and different output frequencies from 20Hz to 50Hz. The test results show that the method of this invention maintains a high diagnostic accuracy rate under various operating conditions, averaging over 99%, and the time for a single diagnosis is less than 10ms, fully meeting the real-time requirements of industrial environments.

Claims

1. A smart fault diagnosis method for inverter IGBTs based on multi-scale attention networks, characterized in that, Includes the following steps: S1. Acquire the three-phase current signal of the three-level NPC inverter, preprocess the three-phase current signal to obtain standardized current samples; S2. The fault features of the current samples are initially extracted using ordinary convolution kernels. Then, multiple sets of depthwise separable convolutions of different scales are used to process the current samples in parallel to extract multi-scale fault features and output feature vectors of each branch. S3. Introduce a scale-aware, efficient attention mechanism to calculate the channel weights of the feature vectors of each branch, and achieve multi-scale feature weighting optimization by multiplying the channel weights with the multi-scale features channel by channel; S4. The weighted and optimized features of each branch are fused, and after batch standardization and activation function processing, the features are compressed by global average pooling and then input into the classification network to obtain the fault diagnosis results. S5. The multi-scale attention network model is trained using the training set, the obtained model parameters are quantized using INT8, and the model performance is tested using the test set.

2. The intelligent fault diagnosis method for inverter IGBTs based on multi-scale attention networks according to claim 1, characterized in that: In step S1, the three-phase current signal includes the current signal under normal operating conditions of the three-level NPC inverter and the current signal under open-circuit fault conditions of various power devices.

3. The intelligent fault diagnosis method for inverter IGBTs based on multi-scale attention networks according to claim 1, characterized in that, The method for preprocessing the three-phase current signal in step S1 includes: The acquired signal is linearly normalized using the following formula: ； In the formula, For the first The first phase current Original values ​​of each sampling point , The first Minimum and maximum values ​​of phase current samples, The value is the normalized value; The normalized sequence is resampled at intervals of [number]. for: ； In the formula, The original sampling frequency of the data. The resampling frequency, For the corresponding interval number; The resampling frequency satisfies: , The number of normalized data points per unit fundamental period in the generated new sequence. This refers to the fundamental frequency of the NPC inverter's output. The current signal is segmented using a sliding window to obtain data segments after resampling.

4. The intelligent fault diagnosis method for inverter IGBTs based on multi-scale attention networks according to claim 1, characterized in that: In step S2, the multiple sets of depthwise separable convolutions at different scales include at least three different scales of depthwise separable convolutions, and the kernel scale satisfies ,in ; For scale The depthwise separable convolution process is as follows: ； In the formula, In , For the input current signal Channel 1 One sampling point, For the first Scale-based depthwise convolution kernel Channel 1 Each weight, This is the output of the depthwise convolution at this scale; After pointwise convolution to fuse channel information, the output is the first... Scaled eigenvectors: ； In the formula, For the first Pointwise convolution kernel weights at different scales For the first Scale-based branch feature output, The channel index for the output feature. The channel index for the input feature. This indicates the number of channels for the input feature.

5. The intelligent fault diagnosis method for inverter IGBTs based on multi-scale attention networks according to claim 1, characterized in that, In step S3, the computational process of the scale-aware, efficient attention mechanism includes: Branch feature mapping of input Perform global average pooling on a channel-by-channel basis to obtain the global information description vector for each channel. The calculation formula is as follows: ； In the formula, It is the first In the feature mapping of the nth branch, the th Channel 1 Feature values ​​at each position, The length of the feature map. For the first The first branch The result after global average pooling of the channels; Using a convolution kernel size of One-dimensional convolution is used to capture local interaction information between channels, and then... The activation function maps the results to a range of 0 to 1, obtaining the weights of each channel. The calculation formula is as follows: ； In the formula, Indicates the kernel size as One-dimensional convolution operation, for Activation function It is the first The first branch Channel weights; The obtained weights Feature mapping with the original branch Multiplying each channel sequentially enhances the response of critical channels; the calculation formula is as follows: ； In the formula, It is the first After weight adjustment, each branch Channel 1 The feature values ​​at each position.

6. The intelligent fault diagnosis method for inverter IGBTs based on multi-scale attention networks according to claim 5, characterized in that, The kernel size Determined through an adaptive function: ； In the formula, This indicates taking the nearest odd number; hyperparameter and Set as , This function ensures that the convolution kernel can cover all effective local interactions between channels.

7. The intelligent fault diagnosis method for inverter IGBTs based on multi-scale attention networks according to claim 1, characterized in that: In step S4, feature fusion employs a concatenation operation, which concatenates the weighted feature maps from different branches along the channel dimension.

8. The intelligent fault diagnosis method for inverter IGBTs based on multi-scale attention networks according to claim 1, characterized in that, The calculation formula for batch standardization in S4 is as follows: ； In the formula, For the fusion feature Channel 1 The value at each position, , These are the mean and variance of the channel, respectively. , For learnable parameters, To prevent tiny values ​​from being divided by zero, This is the value after batch standardization.

9. The intelligent fault diagnosis method for inverter IGBTs based on multi-scale attention networks according to claim 1, characterized in that, In step S4, the classification network... The function outputs the failure probability, and the calculation expression is: ； In the formula, For input data, The length of the input data.

10. The intelligent fault diagnosis method for inverter IGBTs based on multi-scale attention networks according to claim 1, characterized in that, In step S5, the formula for INT8 quantization is: ； ； in, and The calculation formula is as follows: ； ； In the formula, , These are the maximum and minimum values ​​of the parameter, respectively. Scaling factor To quantize the zero point, For the original 32-bit floating-point number, It is an 8-bit fixed-point number after quantization.

Images