IGBT bonding wire fault detection method and system based on stress wave two-dimensional image signal

Abstract

The invention discloses an IGBT bonding wire fault detection method and system based on stress wave two-dimensional image signals. The method comprises the steps that original one-dimensional stress wave signals generated in the switching-on and switching-off processes of an IGBT device in different states and background noise in the switching-off state are collected through an acoustic emission sensor; decomposing the original one-dimensional stress wave signal through a variational mode decomposition algorithm, and removing a frequency band where background noise is located to obtain a denoised one-dimensional stress wave signal; performing two-dimensional imaging processing on the noise-reduced one-dimensional stress wave signal to obtain a two-dimensional image; and constructing a two-dimensional image data set to train a convolutional neural network so as to realize IGBT bonding wire fault detection. The invention aims to realize non-intrusive, efficient and intelligent nondestructive detection on the bonding wire fault of the key component IGBT of the converter valve, and provides effective support for reducing the operation and maintenance cost of the converter valve in a flexible direct-current power transmission system and ensuring the safe and reliable operation of the IGBT of the converter valve component.

Description

Technical Field

[0001] The present invention relates to the technical field of IGBT detection, and particularly to a method and system for detecting IGBT bonding wire faults based on stress wave two-dimensional image signals. Background Art

[0002] The electromagnetic force generated during the switching process of IGBT devices is coupled with their internal structures to release one-dimensional stress waves. Abroad, P. Davari et al. from Aalborg University in Denmark and Sebastian Müller et al. from Chemnitz University of Technology in Germany have both explored the feasibility of non-invasive methods based on stress waves to detect the aging of power semiconductor modules. Their research results all show that there is a strong correlation between stress wave signals and the aging of power modules. As the deterioration degree of the IGBT module increases, the low-frequency peak in the detected stress wave signal spectrum will increase accordingly. However, this method for detecting IGBT bonding wire detachment based on the shift of the frequency domain peak does not have a specific quantification standard, and the detachment of the IGBT bonding wire may lead to misjudgment or missed judgment due to insignificant changes in the frequency domain peak. The differential-mode current in the buck converter will flow through the IGBT module, and the change in the differential-mode current can be used for the state monitoring of the IGBT module. Domestically, Du Mingxing et al. from Tianjin University of Technology detected the IGBT module bonding wire detachment fault based on the one-dimensional electromagnetic interference signal spectrum characteristics of the buck converter differential-mode current. However, this method may not have a good detection effect on the bonding wire faults of multi-core parallel IGBT modules because when the total number of detached bonding wires inside the module is large and the number of detached bonding wires per core is small, the change in the spectrum characteristics of the electromagnetic interference signal may not be obvious. Summary of the Invention

[0003] The technical problem to be solved by the present invention: Aiming at the above problems of the prior art, a method and system for detecting IGBT bonding wire faults based on stress wave two-dimensional image signals are provided. The present invention aims to achieve non-invasive, efficient, and intelligent non-destructive detection of the bonding wire faults of IGBT, a key component of the converter valve, and provide effective support for reducing the operation and maintenance costs of the converter valve in the flexible DC transmission system and ensuring the safe and reliable operation of the IGBT components of the converter valve assembly.

[0004] To solve the above technical problems, the technical solution adopted by the present invention is as follows: A method for detecting IGBT bonding wire faults based on stress wave two-dimensional image signals includes the following steps: For IGBT devices in different types of faults and healthy states, respectively collect the original one-dimensional stress wave signals generated during the switching process of the IGBT device and the background noise in the off state through an acoustic emission sensor; The frequency band distribution of the background noise is determined through fast Fourier transform of the background noise; the original one-dimensional stress wave signal is decomposed by the variational mode decomposition algorithm to obtain the intrinsic mode functions in different frequency ranges, the intrinsic mode functions in the frequency band where the background noise is located are removed, and the remaining intrinsic mode functions are combined to obtain the one-dimensional stress wave signal after noise reduction; The one-dimensional stress wave signal after noise reduction is processed into a two-dimensional image; The two-dimensional image is appended with the fault or health state of the IGBT device to construct a two-dimensional image dataset, and the convolutional neural network is trained using the two-dimensional image dataset to establish the mapping relationship between the two-dimensional image and the IGBT bonding wire fault detection result for realizing the IGBT bonding wire fault detection.

[0005] Optionally, before decomposing the original one-dimensional stress wave signal by the variational mode decomposition algorithm to obtain the intrinsic mode functions in different frequency ranges, the particle swarm optimization algorithm is also used to optimize the penalty factor and the decomposition layer number in the variational mode decomposition algorithm to obtain the ideal intrinsic mode functions more quickly. In the ideal intrinsic mode functions, the noise signals are distributed in one or more independent mode functions and do not overlap with the effective signals. When using the particle swarm optimization algorithm for the penalty factor and the decomposition layer number in the variational mode decomposition algorithm, the fitness function used is: , , , , where, represents the fitness function, , , are weights, is the reconstruction error feature, is the modal bandwidth feature, is the penalty term feature, is the decomposition layer number, is the penalty factor, is the original signal, is the reconstructed signal after variational mode decomposition, is the modal set after decomposition, is the number of signals in the original signal, is the nth signal in the original signal, is the reconstructed signal the nth signal in, and are the frequency centers of the ith and jth modes respectively, is a small constant used to avoid division by zero errors, To obtain the maximum value, is the preset maximum number of modes.

[0006] Optionally, the process of two-dimensionally imaging the denoised one-dimensional stress wave signal to obtain a two-dimensional image includes: Scaling the denoised one-dimensional stress wave signal to the range of [-1, 1] according to the following formula: , In the above formula, is the i-th sampling point of the scaled denoised one-dimensional stress wave signal in is the i-th sampling point of the denoised one-dimensional stress wave signal before scaling in is the denoised one-dimensional stress wave signal, ; Calculating the angle value after polar coordinate transformation of the denoised one-dimensional stress wave signal according to the following formula: , In the above formula, is the angle value after polar coordinate transformation of the i-th sampling point, , is the denoised one-dimensional stress wave signal scaled to the range of [-1, 1]; Generating a Gram angle and field image GASF or a Gram angle difference field image GADF as the obtained two-dimensional image, where the functional expression for generating the Gram angle and field image GASF is: , The functional expression for generating the Gram angle difference field image GADF is: , where, is the Gram angle and field image GASF, ~ are respectively the angle values after polar coordinate transformation of the 1st to n-th sampling points, is the unit row vector, is the average value of the denoised one-dimensional stress wave signal scaled to the range of [-1, 1] in is the transposed vector of, is the Gram angle difference field image GADF.

[0007] Optionally, the convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer that are connected in sequence. The input layer is used to input the two-dimensional stress wave image, and the output layer is used to output the bonding wire fault detection result of the IGBT device.

[0008] Optionally, the different types of faults include some or all of normal bonding wire, cracked bonding wire, single bonding wire detachment, and multiple bonding wire detachment.

[0009] In addition, the present invention also provides an IGBT bonding wire fault detection system based on two-dimensional stress wave image signals, including: A data acquisition program unit, which is used to collect the original one-dimensional stress wave signals generated during the switching process of the IGBT device and the background noise in the off state through an acoustic emission sensor for different types of faults and healthy states of the IGBT device; A data noise reduction program unit, which is used to determine the frequency band distribution of the background noise through fast Fourier transform of the background noise; decompose the original one-dimensional stress wave signal into intrinsic mode functions in different frequency ranges by using the variational mode decomposition algorithm, remove the intrinsic mode functions in the frequency band where the background noise is located, and combine the remaining intrinsic mode functions to obtain the denoised one-dimensional stress wave signal; An image generation program unit, which is used to perform two-dimensional image processing on the denoised one-dimensional stress wave signal to obtain a two-dimensional image; A model training program unit, which constructs a two-dimensional image data set by attaching the fault or health state of the IGBT device to the two-dimensional image, and uses the two-dimensional image data set to train the convolutional neural network to establish a mapping relationship between the two-dimensional image and the IGBT bonding wire fault detection result for realizing the IGBT bonding wire fault detection.

[0010] Optionally, before the data noise reduction program unit decomposes the original one-dimensional stress wave signal into intrinsic mode functions in different frequency ranges by using the variational mode decomposition algorithm, it further includes optimizing the penalty factor and the decomposition layer number in the variational mode decomposition algorithm by using the particle swarm optimization algorithm to obtain the ideal intrinsic mode functions more quickly. In the ideal intrinsic mode functions, the noise signals are distributed in one or more independent mode functions and do not overlap with the effective signals. The fitness function used when optimizing the penalty factor and the decomposition layer number in the variational mode decomposition algorithm by using the particle swarm optimization algorithm is: , , , , Wherein, represents the fitness function, , , is the weight, is the reconstruction error feature, is the modal bandwidth feature, is the penalty term feature, is the decomposition level, is the penalty factor, is the original signal, is the reconstructed signal after variational mode decomposition, is the set of modes after decomposition, is the number of signals in the original signal, is the nth signal in the original signal, is the reconstructed signal the nth signal in, and are the frequency centers of the ith and jth modes respectively, is a small constant used to avoid division by zero errors, is to take the maximum value, is the preset maximum number of modes.

[0011] Optionally, the image generation program unit includes: A normalization program unit for scaling the denoised one-dimensional stress wave signal to the range [-1, 1] according to the following formula: , In the above formula, is the ith sampling point of the scaled denoised one-dimensional stress wave signal in, is the ith sampling point of the denoised one-dimensional stress wave signal before scaling in, is the denoised one-dimensional stress wave signal, ; An angle conversion program unit for calculating the angle value after polar coordinate conversion of the denoised one-dimensional stress wave signal according to the following formula: , In the above formula, is the angle value after polar coordinate conversion of the ith sampling point, , is the denoised one-dimensional stress wave signal scaled to the range [-1, 1]; An image calculation program unit for generating a Gram angle and field image GASF or a Gram angle difference field image GADF as the obtained two-dimensional image, where the functional expression for generating the Gram angle and field image GASF is: , The functional expression for generating the Gram Angular Difference Field Image (GADF) is as follows: , where is the Gram Angular and Field Image (GASF), ~ are the angular values after polar coordinate transformation for the 1st to nth sampling points respectively, is the unit row vector, is the one-dimensional stress wave signal after noise reduction scaled to the range [-1, 1], is the average value of is the transposed vector of and is the Gram Angular Difference Field Image (GADF).

[0012] Optionally, the convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer connected in sequence. The input layer is used to input the two-dimensional stress wave image, and the output layer is used to output the bonding wire fault detection result of the IGBT device.

[0013] Optionally, the different types of faults collected by the data acquisition program unit include some or all of normal bonding wire, cracked bonding wire, single bonding wire detachment, and multiple bonding wire detachments.

[0014] In addition, the present invention also provides an IGBT bonding wire fault detection system based on two-dimensional stress wave image signals, including a microprocessor and a memory connected to each other. The microprocessor is programmed or configured to execute the IGBT bonding wire fault detection method based on two-dimensional stress wave image signals.

[0015] In addition, the present invention also provides a computer-readable storage medium storing a computer program or instruction, which is programmed or configured to execute the IGBT bonding wire fault detection method based on two-dimensional stress wave image signals through a processor.

[0016] In addition, the present invention also provides a computer program product including a computer program or instruction, which is programmed or configured to execute the IGBT bonding wire fault detection method based on two-dimensional stress wave image signals through a processor.

[0017] Compared with the prior art, the present invention mainly has the following advantages: 1. The method of the present invention includes collecting the original one-dimensional stress wave signals generated during the switching process of IGBT devices in different states and the background noise in the off state through acoustic emission sensors respectively; decomposing the original one-dimensional stress wave signals by using the variational mode decomposition algorithm and removing the frequency band where the background noise is located to obtain the denoised one-dimensional stress wave signals; performing two-dimensional imaging processing on the denoised one-dimensional stress wave signals to obtain two-dimensional images; constructing a two-dimensional image dataset to train a convolutional neural network for realizing IGBT bonding wire fault detection. The present invention can realize non-invasive, efficient and intelligent non-destructive detection of the bonding wire faults of the key component IGBT of the converter valve, and provide effective support for reducing the operation and maintenance costs of the converter valve in the flexible DC transmission system and ensuring the safe and reliable operation of the IGBT of the converter valve assembly.

[0018] 2. Signals often contain multiple superimposed components, such as noise, periodicity, transients, and non-linear characteristics, etc. Traditional frequency domain analysis methods may be difficult to effectively separate these components. The stress waves generated during the IGBT switching process contain background noise generated by environmental factors (such as the working noise generated by the DC power supply, the electromagnetic interference signals generated by the changes in the surrounding electromagnetic field, etc.), and these background noises can be collected by acoustic emission sensors through air coupling and other means. Different from randomly distributed white noise, this background noise is usually distributed in a relatively concentrated frequency range and has a relatively high time domain amplitude, which may affect the accuracy of subsequent two-dimensional image signal analysis and processing. Therefore, it is necessary to perform background noise removal processing on the original signal. The present invention decomposes the signal into several intrinsic mode functions (IMFs) through the variational mode decomposition algorithm, so that each mode can independently reflect the specific frequency characteristics or local dynamic behavior of the signal. And the variational mode decomposition algorithm can achieve precise positioning of the signal frequency, so that each mode component has a relatively independent frequency band. Therefore, the variational mode decomposition algorithm can decompose the background noise generated during the IGBT switching process into one or more independent frequency bands, and then remove the signals in these frequency bands from all independent mode signals, thereby achieving the purpose of noise reduction. And converting the one-dimensional signals of different modes into two-dimensional images for feature extraction can also increase the number of the two-dimensional image signal dataset and improve the accuracy of subsequent neural network model training.

[0019] 3. The present invention includes constructing a two-dimensional image dataset by attaching the fault or health state of the IGBT device to a two-dimensional image, training a convolutional neural network using the two-dimensional image dataset to establish a mapping relationship between the two-dimensional image and the IGBT bond wire fault detection result for IGBT bond wire fault detection. By converting the one-dimensional signal into a two-dimensional image, the time-frequency characteristics of the signal can be more intuitively displayed, thereby enhancing the visual understanding of signal changes. This conversion method usually combines techniques such as the short-time Fourier transform, continuous wavelet transform, and Gramian angular field transform to present the time-frequency domain information of the one-dimensional signal in the form of an image. The two-dimensional image can not only effectively capture the non-stationary characteristics in the signal but also, with the help of image processing techniques and deep learning algorithms, achieve automatic feature extraction and pattern recognition, further improving the accuracy and efficiency of fault detection.

[0020] 4. The existing IGBT stress wave detection method detects by attaching a piezoelectric ceramic sensor to the surface of the device package, without the need for an additional hardware detection circuit, and has the advantages of non-invasive and non-destructive detection. Therefore, the one-dimensional stress wave signal generated during the IGBT switching process can be collected through the stress wave detection technology. The present invention intends to perform noise reduction processing on the stress wave signal generated during the IGBT switching process through the variational mode decomposition algorithm, and then realize the intelligent detection of the IGBT bond wire fault based on the two-dimensional image feature changes of different modal stress wave signals generated during the IGBT switching process. Description of the Drawings

[0021] Figure 1 It is a schematic diagram of the basic process of the method in the embodiment of the present invention.

[0022] Figure 2 It is the network structure diagram of the convolutional neural network in the embodiment of the present invention.

[0023] Figure 3 It is the structure diagram of the original one-dimensional stress wave signal acquisition system in the embodiment of the present invention.

[0024] Figure 4 It is a schematic diagram of the detailed process of the method in the embodiment of the present invention.

[0025] Figure 5 It is the typical background noise signal generated by the operation of the DC power supply in the embodiment of the present invention.

[0026] Figure 6 It is the typical one-dimensional stress wave signal in the embodiment of the present invention.

[0027] Figure 7 It is a schematic diagram of stress wave signals of different modes in the embodiment of the present invention.

[0028] Figure 8Gram angles and field images corresponding to sine waves of different frequencies in the embodiments of the present invention, where the frequency of (a) is 20 kHz, the frequency of (b) is 50 kHz, and the frequency of (c) is 80 kHz.

[0029] Figure 9 Gram angle difference field images corresponding to sine waves of different frequencies in the embodiments of the present invention, where the frequency of (a) is 20 kHz, the frequency of (b) is 50 kHz, and the frequency of (c) is 80 kHz. Detailed implementation manners

[0030] To enable those skilled in the art of the present technology to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention. As Figure 1 shown, the IGBT bonding wire fault detection method based on the two-dimensional image signal of stress waves in this embodiment includes the following steps: For IGBT devices in different types of faults and healthy states, respectively collect the original one-dimensional stress wave signals generated during the switching process of the IGBT device and the background noise in the off state through an acoustic emission sensor, and determine the frequency band distribution of the background noise through fast Fourier transform; Determine the frequency band distribution of the background noise through fast Fourier transform of the background noise. Decompose the original one-dimensional stress wave signal into intrinsic mode functions in different frequency ranges by using the variational mode decomposition algorithm, remove the intrinsic mode functions in the frequency band where the background noise is located, and combine the remaining intrinsic mode functions to obtain the one-dimensional stress wave signal after noise reduction; Perform two-dimensional imaging processing on the one-dimensional stress wave signal after noise reduction to obtain a two-dimensional image; Construct a two-dimensional image dataset by attaching the fault or healthy state of the IGBT device to the two-dimensional image, and use the two-dimensional image dataset to train a convolutional neural network to establish a mapping relationship between the two-dimensional image and the IGBT bonding wire fault detection result for realizing IGBT bonding wire fault detection.

[0031] The stress wave detection technology is a non-invasive non-destructive testing technology, which is used for IGBT bonding wire fault detection in this embodiment. However, the existing IGBT bonding wire fault detection methods mainly rely on one-dimensional stress wave signals for frequency domain feature analysis. The change intuitiveness of the characteristic parameters selected by this traditional analysis method is poor, making it possible for technicians to face difficulties when interpreting the results and unable to quickly identify potential fault modes. And this method may also ignore the non-linear relationship in the original signal, which is crucial for capturing complex fault characteristics, and may lead to misjudgment or missed judgment. In addition, most of the existing frequency domain feature analysis methods are based on the frequency domain feature analysis of the complete signal, and the accuracy of reflecting the dynamic changes of the local features of the signal is poor. Moreover, the existing noise processing methods such as wavelet denoising algorithms have poor noise reduction effects on the background noise of the stress wave signal, and may not be able to completely filter out the noise signal or lose the effective signal during the noise reduction process. Aiming at these shortcomings, the purpose of the present invention is to analyze the potential bonding wire faults based on the two-dimensional image features of the stress wave signal to improve the accuracy and intuitiveness of fault detection. The two-dimensional image signal can more intuitively display the frequency characteristics of the original signal and enhance the ability to identify complex features and changes in the signal. This method can not only effectively capture the non-linear features in the original signal, but also combine neural network algorithms for image processing and feature learning, automatically extract the high-dimensional features in the original signal, and achieve intelligent detection. This will provide important support for the reliability monitoring and maintenance of IGBT bonding wires and promote the development and application of related condition detection technologies. The variational mode decomposition algorithm is mainly used in the present invention to achieve the precise positioning and removal of the background noise frequency band, avoiding the influence of background noise on subsequent signal analysis. Analyzing the stress wave signal from multiple modes improves the ability of the subsequent stress wave two-dimensional image signal to reflect the dynamic changes of local features and can also reduce the complexity of signal processing and analysis.

[0032] Before the original one-dimensional stress wave signal is decomposed into intrinsic mode functions in different frequency ranges by the variational mode decomposition algorithm in this embodiment, it also includes optimizing the penalty factor and the decomposition layer number in the variational mode decomposition algorithm by using the particle swarm optimization algorithm to obtain the ideal intrinsic mode function more quickly. The noise signal in the ideal intrinsic mode function is distributed in one or more independent mode functions and does not overlap with the effective signal. And the fitness function used when using the particle swarm optimization algorithm for the penalty factor and the decomposition layer number in the variational mode decomposition algorithm is: , , , , Wherein, represents the fitness function, 、 , is the weight, is the reconstruction error feature, is the modal bandwidth feature, is the penalty term feature, is the decomposition level, is the penalty factor, is the original signal, is the reconstructed signal after variational mode decomposition, is the set of modes after decomposition, is the number of signals in the original signal, is the nth signal in the original signal, is the reconstructed signal the nth signal in, and are the frequency centers of the ith and jth modes respectively, is a small constant used to avoid division-by-zero errors, is to take the maximum value, is the preset maximum number of modes. Variational mode decomposition is an adaptive signal processing method used to decompose complex signals into a series of simple intrinsic mode functions (IMFs) to extract important components and features in the signals. The particle swarm optimization algorithm (PSO) is an optimization algorithm based on swarm intelligence that finds the optimal solution to a problem by simulating the foraging behavior of bird flocks and gradually approaches the optimal solution using information sharing and update strategies among particles. In this embodiment, based on the parameter optimization algorithm, the penalty factor and decomposition level of variational mode decomposition are selected, and the stress wave signal is decomposed into independent stress wave signals of different modes through the variational mode decomposition algorithm for noise reduction processing and highlighting the local features of the signal. Among them, the background noise signal is decomposed within one or more independent frequency bands, and the original stress wave signal generated at the IGBT switching moment can be denoised by deleting the corresponding noisy signal frequency bands. Then, the Gramian angular field transform is used to convert the stress wave signals of different modes into two-dimensional image signals, and based on the feature changes in the two-dimensional images, the intelligent detection of IGBT bonding wire faults is realized by combining the neural network algorithm.

[0033] In this embodiment, the two-dimensional image processing of the denoised one-dimensional stress wave signal to obtain a two-dimensional image includes: S3.1, scale the denoised one-dimensional stress wave signal to the range [-1, 1] according to the following formula: , In the above formula, is the scaled denoised one-dimensional stress wave signal the i-th sampling point in is the one-dimensional stress wave signal after noise reduction before scaling the i-th sampling point in is the one-dimensional stress wave signal after noise reduction ; S3.2. Calculate the angular value after polar coordinate transformation of the one-dimensional stress wave signal after noise reduction according to the following formula: , In the above formula, is the angular value after polar coordinate transformation of the i-th sampling point, , is the one-dimensional stress wave signal after noise reduction scaled to the range of [-1, 1]; S3.3. Generate the Gram angle and field image GASF or the Gram angle difference field image GADF as the obtained two-dimensional image. The functional expression for generating the Gram angle and field image GASF is: , The functional expression for generating the Gram angle difference field image GADF is: , where is the Gram angle and field image GASF, ~ are respectively the angular values after polar coordinate transformation of the 1st to n-th sampling points, is the unit row vector, is the one-dimensional stress wave signal after noise reduction scaled to the range of [-1, 1] is the average value of is the transposed vector of is the Gram angle difference field image GADF. The Gram angle field transform is a technique for converting time series data into images, commonly used in time series analysis and deep learning. By converting the polar coordinate representation of a time series into a two-dimensional image, the Gram angle field transform can preserve the characteristics and change patterns of the time series. This method can utilize convolutional neural networks (CNNs) for classification and regression tasks when dealing with sequential data.

[0034] Convolutional Neural Network (CNN) is a deep learning model specifically designed to process data with grid structures, such as two-dimensional images. CNN automatically extracts features through convolutional layers, reduces the feature dimension using pooling layers, and performs classification or regression through fully connected layers. Its advantage lies in being able to effectively process high-dimensional data, reduce the number of parameters, and improve the generalization ability of the model. Convolutional Neural Network (CNN) is suitable for processing grid-like input data such as two-dimensional images. Due to the continuity between pixels in an image, adjacent pixels often have similar color or brightness values, and this spatial correlation is an important characteristic of image data. The network structure diagram of the convolutional neural network in this embodiment is as Figure 2 shown. The convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer connected in sequence. The input layer is used to input the two-dimensional stress wave image, and the output layer is used to output the bonding wire fault detection result of the IGBT device. The convolutional neural network in this embodiment captures this local spatial dependence by introducing the structures of convolutional layers and pooling layers. The convolutional layer extracts the feature information of the image by performing convolutional calculations on the image pixels. The pooling layer is generally used to reduce the dimension of the feature map and improve the calculation speed. In a CNN, each convolutional layer generates a set of feature maps, and each feature map is obtained by performing convolutional calculations on the input image with multiple sets of convolutional kernels. Therefore, each set of convolutional kernels corresponds to a specific feature, thus extracting different information in the image. At the same time, in the fully connected layer of the CNN, all the features of the previous layer are input in the form of vectors to obtain the corresponding output vector to indicate the final classification result.

[0035] The different types of faults in step S1 of this embodiment include some or all of normal bonding wire, cracked bonding wire, single bonding wire detachment, and multiple bonding wire detachment.

[0036] As Figure 4 shown, the method of this embodiment specifically includes the following steps: Data acquisition: Stress wave: During the normal switching process of the IGBT device, there are carrier changes, and the electromagnetic force caused thereby is coupled with the internal structure to generate an electromagnetic acoustic emission phenomenon, forming a stress wave signal that propagates outward from the inside of the device. This stress wave signal is a one-dimensional time series signal. As Figure 3 shown, where u ce and i ceThey are the voltage and current signals during the switching process of the IGBT device respectively. The electromagnetic force caused by the voltage and current signals during the switching process of the IGBT device is coupled with the internal structure to generate an electromagnetic acoustic emission phenomenon, forming a stress wave signal propagating from the inside of the device to the outside. The original one-dimensional stress wave signal acquisition system consists of a piezoelectric ceramic acoustic emission sensor, a preamplifier, an oscilloscope, and a laptop computer. Attaching the piezoelectric ceramic sensor to the packaging surface on the heat dissipation side of the IGBT device can detect the original one-dimensional stress wave signal generated during the switching process of the IGBT device. Among them, the IGBT device to be measured is used to produce fault samples by artificially creating bonding wire defects such as bonding wire cracks, single bonding wire detachment, and multiple bonding wire detachment. The original one-dimensional stress wave signal collected by the piezoelectric ceramic acoustic emission sensor is amplified by the preamplifier and displayed on the oscilloscope, and finally the original one-dimensional stress wave signal is recorded on the laptop computer. Keep the IGBT device to be measured in the off state, and collect background noise through the piezoelectric ceramic acoustic emission sensor attached to the IGBT packaging surface, and confirm the main frequency band distribution of the background noise signal through fast Fourier transform. In this embodiment, the stress wave signal generated during the operation of the IGBT is detected by the piezoelectric ceramic sensor. Based on the modal decomposition algorithm, the original stress wave signal is decomposed to obtain stress wave signals in different modes, and the IGBT bonding wire fault detection is realized by extracting the two-dimensional image features of the stress wave signal in different modes. The piezoelectric ceramic sensor in this solution can be replaced by a piezoelectric film sensor as an alternative sensing solution for stress wave data acquisition. The acoustic emission instrument can also display the stress wave signal collected by the piezoelectric ceramic sensor in real time, so the acoustic emission instrument can replace the oscilloscope in the original technical solution.

[0037] In this embodiment, keep the IGBT device to be measured in the off state (the IGBT has no switching action and other conditions remain unchanged), and collect background noise through the piezoelectric ceramic acoustic emission sensor attached to the IGBT packaging surface (typical background noise signals generated during the operation of the DC power supply such as Figure 5As shown, the main frequency band distribution of the background noise signal is confirmed through fast Fourier transform. IGBT devices with different bond wire fault types and health states are selected as the devices under test. Among them, the bond wire fault types of IGBT devices include bond wire cracks, single bond wire detachment, multiple bond wire detachment, etc. These faults can be artificially created by cutting the bond wires. The stress wave data acquisition system consists of a piezoelectric ceramic sensor, a preamplifier, an oscilloscope, and a laptop computer. The piezoelectric ceramic sensor is attached to the packaging surface on the heat dissipation side of the IGBT device to detect the original one-dimensional stress wave signal generated during the switching process of the IGBT device. Among them, the piezoelectric ceramic sensor directly acquires the stress wave signal based on the piezoelectric effect. The piezoelectric ceramic sensor can perform stress wave detection without electrical contact with the working circuit of the IGBT, and has the advantage of non-invasive non-destructive detection. Since the amplitude order of magnitude of the stress wave signal is only a few tens of millivolts, a preamplifier is used to amplify the weak stress wave signal. The display effect of the stress wave signal on the oscilloscope can be adjusted according to different amplification multiples of the preamplifier. The original one-dimensional stress wave signal is recorded through the oscilloscope data saving software on the laptop computer. A typical one-dimensional stress wave signal is as Figure 6 shown.

[0038] b. One-dimensional signal modal decomposition and denoising: Since the change of the stress wave signal generated by the IGBT device with bond wire faults during the switching process may not be obvious in the complete frequency spectrum range, it is considered to use the variational mode decomposition algorithm to decompose the stress wave signal in different frequency ranges to highlight the local characteristic information of the original one-dimensional stress wave signal and remove the background noise in a specific frequency band at the same time. In the data analysis software MATLAB, the variational mode decomposition algorithm is used to identify different frequency components of the original one-dimensional signal, and the original one-dimensional stress wave signal is decomposed into several intrinsic mode functions (IMF) according to different frequency components. Then, different intrinsic mode functions are respectively extracted as independent one-dimensional modal signals. Among them, the variational mode decomposition algorithm can use global optimization algorithms such as the particle swarm optimization algorithm to optimize the parameter selection of the penalty factor and the decomposition layer number to obtain the ideal intrinsic mode function more quickly. In the ideal intrinsic mode function, the noise signal is distributed in one or more independent mode functions and does not overlap with the effective signal.

[0039] Specifically, in this embodiment, the variational mode decomposition algorithm is used to implement the modal decomposition of the original one-dimensional stress wave signal, and the particle swarm optimization algorithm is used to optimize the selection of the penalty factor and the decomposition layer number parameters in the variational mode decomposition algorithm. The setting ranges of the penalty factor and the decomposition layer number will affect the modal decomposition result, so they can be set according to the actual situation. In this embodiment, the range of the penalty factor is taken from 500 to 5000, and the range of the decomposition layer number is taken from 3 to 10 layers. When using the particle swarm optimization algorithm to optimize the penalty factor and the decomposition layer number , first define the fitness function, which comprehensively evaluates the independence of the modes and the reconstruction error. Then, initialize the particle swarm, where each particle represents a set of parameters ( , ), and then randomly generate the initial position and velocity. During the iteration process, update the velocity and position of the particles, and at the same time calculate the fitness and update the individual optimal position and the global optimal position of each particle. In this way, the particles continuously adjust their positions to approach the optimal parameters, and finally obtain the best penalty factor and decomposition layer number. The advantage of variational mode decomposition is that it does not need to pre-construct fixed basis functions or frequencies, and it can dynamically determine the decomposition level of the signal and the number of intrinsic mode functions according to the characteristics of the signal. Based on the variational mode decomposition algorithm, the original one-dimensional stress wave signal can be decomposed into multiple groups of different modal signals that can reflect information in different frequency ranges. Among them, the intrinsic mode function of each original signal corresponds to a group of modal signals as Figure 7 shown. Specifically, the present invention designs a fitness function applicable to the variational mode decomposition of stress wave signals based on minimizing the reconstruction error of the decomposition result and considering the spectral characteristics to ensure that each mode can effectively capture the target of each frequency peak ; , , , , where, is the decomposition layer number, is the penalty factor, is the original signal, is the reconstructed signal after variational mode decomposition, is the set of modes after decomposition, and are the frequency centers of the i-th and j-th modes respectively, is a small constant used to avoid division by zero errors, is the n-th signal in the original signal, is the reconstructed signal the n-th signal in is the number of signals in the original signal, is the preset maximum number of modes, , , is the weight. In the above function expression, To reconstruct the error characteristic, the difference between the original signal and the reconstructed signal is measured, and the mean square error is usually used. For the modal bandwidth feature, the distance between them is evaluated by calculating the spectral center of each mode, which is mainly used to calculate the frequency bandwidth of each mode to minimize the overlap between modes. It is a penalty feature to prevent overfitting caused by excessive modal decomposition. , , The sum of the weights is 1 to ensure that the relative importance of each feature can be directly reflected in the fitness function. , , All must be positive numbers. , , The weights are optimized based on the value of , and a suitable weight combination is quickly found based on a simple linear search. A weight range is set (from 0 to 1), and then each step is 0.1 in this range. The fitness function is calculated for each weight combination, and the weight combination with the best effect is recorded.

[0040] c. One-dimensional signal two-dimensional imaging processing: Apply the Gram angular field transform to map the one-dimensional signal into two-dimensional space by calculating the phase information and frequency information of the signal, thereby generating the corresponding Gram angular field image. First, convert the one-dimensional modal signal in the rectangular coordinate system into X Scaling to the range [-1,1] by formula (1) : , In the above formula, is the scaled and denoised one-dimensional stress wave signal The i-th sampling point in is the one-dimensional stress wave signal after noise reduction before scaling The i-th sampling point in is the one-dimensional stress wave signal after noise reduction, ; Then, the stress wave signal based on the time series is converted into a vector according to the polar coordinate transformation formula for subsequent inner product operation.

[0041] , In the above formula, is the angle value of the i-th sampling point after polar coordinate conversion, , The one-dimensional stress wave signal after noise reduction scaled to the range of [-1, 1]; Then, according to the formula, the Gram angle and field image GASF and the Gram angle difference field image GADF are generated. Among them, the GASF is obtained by using the cosine function of the sum of two angles, and the Gram angle and field image GADF is obtained by using the sine function of the difference of two angles. Add (subtract) each pair of values, then take the cosine value and sum them up to generate two two-dimensional feature images, namely the Gram angle and field image GASF and the Gram angle difference field image GADF.

[0042] , , wherein, is the Gram angle and field image GASF, ~ are the angle values after polar coordinate transformation of the 1st to nth sampling points respectively, is the unit row vector, is the one-dimensional stress wave signal after noise reduction scaled to the range of [-1, 1] the average value of is the transposed vector of is the Gram angle difference field image GADF.

[0043] Specifically, the signal obtained by decomposing the above step b is normalized to keep its amplitude in the range of -1 to 1 to eliminate the possible dimensional influence between different signals. Then, the Gram angle field transform is used to convert the normalized signal into a two-dimensional image signal. Each group of modal signals corresponds to two Gram images, namely the Gram angle and field image and the Gram angle difference field image. Based on these Gram image data, a two-dimensional image signal dataset corresponding to the IGBT original one-dimensional stress wave signal is constructed. The Gram angle field images of typical sine signals are as shown in Figure 8 and Figure 9 . Among them Figure 8 is the Gram angle and field image (GASF), Figure 9 is the Gram angle difference field image (GADF). As the frequency increases, the number of points in both images increases correspondingly, indicating that the Gram images can characterize the frequency differences of the signals.

[0044] d. Data annotation: Annotate the samples of the two-dimensional image signal dataset according to the actual state of the bonding wires of the IGBT device. Annotate the image samples of different bonding wire fault types or health states, and these samples will be used as the supervision signals for the convolutional neural network. The supervision signals usually refer to the data labels used to guide the model learning during the training process. These labels are the correct outputs corresponding to the input data, which tell the model what it should predict given the input. Supervised learning is a learning algorithm in which the model learns from the labeled training data, which includes the input and the corresponding output labels.

[0045] e. Model setting: Divide the training set, test set, and validation set according to the ratio of 7:1.5:1.5. Use the training set to train the convolutional neural network model, and at the same time adjust the hyperparameters such as the learning rate and batch size of the model to optimize the model performance. The learning rate controls the speed of the model weight update. Too high may lead to unstable training, while too low may result in slow training speed. Usually, the learning rate is set to start from 0.01. The batch size is the number of samples used in each training. A smaller batch size can improve the generalization ability of the model, but the computational efficiency is low; a larger batch size can improve the computational efficiency, but may affect the generalization of the model. The batch size can be appropriately adjusted according to the convergence situation of the model and the computing resources. Sometimes increasing the batch size can speed up the training, but the learning rate may need to be readjusted.

[0046] f. Model evaluation: Loss is an indicator to measure the difference between the predicted value and the actual value of the model. Common loss functions include mean squared error, etc. The fitting situation of the model can be estimated according to the mean squared error (MSE) formula:

[0047] where, is the actual value, is the predicted value, is the number of samples. During the training process, after each iteration, calculate the mean squared error of the training set and the validation set for each iteration. Ideally, as the training progresses, both the training loss and the validation loss should gradually decrease. If the training loss continues to decrease while the validation loss starts to increase after a certain point, this may be a sign of overfitting. If the training loss or the validation loss becomes flat or increases during the training process, it may be a sign of underfitting. Both overfitting and underfitting can be solved by adding more training data.

[0048] g. Model application: Use the trained convolutional neural network model to perform fault detection on new two-dimensional Gramian angular field image data. The model will predict the health status of the IGBT device based on the learned features, thereby realizing the identification of defect types. Among them, the new Gramian angular field image is obtained by performing Gramian angular field transformation on the stress wave signal detection data of the new device under test after being processed by the empirical mode decomposition algorithm.

[0049] The IGBT bonding wire fault detection method based on stress wave two-dimensional image signals in this embodiment mainly has the following advantages: (1) A fault diagnosis method for IGBT bonding wires based on two-dimensional images is proposed. Traditional frequency analysis methods usually detect the state of IGBT bonding wires based on changes in statistical data. By using two-dimensional image processing technology, the state of the IGBT bonding wire can be presented in a visual way, enabling fault diagnosis to shift from traditional electrical tests to image analysis, which provides a new idea for signal detection. Through image processing algorithms, features related to faults can be extracted to improve the accuracy of fault recognition. At the same time, two-dimensional image data can also be combined with technologies such as deep learning algorithms to achieve automatic analysis and classification of two-dimensional images. Such an automated process can improve detection efficiency, reduce human errors, and enhance the reliability of fault diagnosis. Through rapid image acquisition and analysis, real-time detection and response to faults can be achieved, timely maintenance measures can be taken, downtime can be reduced, and production efficiency can be improved. (2) The variational mode decomposition algorithm is used to denoise and extract local features in the original stress wave signal of the IGBT. The acoustic emission sensor can collect background noise generated by other equipment and instruments in the IGBT working environment. This noise usually has a relatively concentrated frequency range and a high time-domain amplitude, and may be misidentified as a valid signal, thus affecting the subsequent analysis process. The variational mode algorithm can accurately locate the frequency of the noise signal, making the noise modal components distributed in relatively independent frequency bands, and then achieving noise reduction processing for the original signal. Optimization algorithms are used to select the penalty factor and decomposition layer parameters of the variational mode decomposition algorithm to accelerate the decomposition process of the original signal. In addition, the IGBT is affected by various stresses during use, and its original stress wave signal usually contains complex frequency components, which makes it difficult to directly analyze these signals and effectively extract useful features. Through variational mode decomposition, it is possible to focus on the local features contained in each mode, thereby extracting specific information related to faults. This local analysis makes it more accurate to capture the dynamic changes and subtle features of the signal. In addition, the stress wave signal of the IGBT usually exhibits non-linear and non-stationary characteristics. The variational mode decomposition algorithm can also effectively handle these characteristics, making the extracted features more in line with the dynamic characteristics of the actual signal. (3) The stress wave detection technology can achieve non-invasive and non-destructive detection of IGBT bonding wire faults. The stress wave detection technology can use the change of characteristic parameters of the stress wave signal generated during the normal switching operation of the IGBT to detect the state and potential faults of the bonding wire. Since it is not necessary to damage the packaging structure of the IGBT, the stress wave detection technology can ensure the integrity and continued use of the IGBT device. In summary, in the IGBT bonding wire fault detection method based on stress wave two-dimensional image signals in this embodiment, the variational mode decomposition algorithm can be combined with the optimization algorithm to achieve noise reduction processing for the background noise in specific frequency bands of the IGBT working environment, which is difficult to handle by traditional denoising algorithms such as wavelet denoising algorithms.The present invention is an IGBT bonding wire fault diagnosis method based on stress wave two-dimensional image signals. Stress wave is a one-dimensional signal, and the two-dimensionalization of stress wave signals can be achieved through a one-dimensional signal two-dimensionalization algorithm. Two-dimensional image features can intuitively display whether there is a fault in the bonding wire state of the IGBT. Combining neural network algorithms such as deep learning can realize the automatic and rapid detection of IGBT faults. The IGBT bonding wire fault detection method based on stress wave two-dimensional image signals in this embodiment can expand the existing IGBT bonding wire fault detection system based on frequency domain features and increase the characteristic parameters that can characterize the IGBT fault type. The present invention also has the advantages of non-invasive, non-destructive detection, rapid detection, intelligent detection, etc., and can provide a reliable and rapid detection means for the operation and maintenance of IGBT, a key component of the converter valve in the flexible DC transmission system, which helps to improve the stability of the flexible DC transmission system.

[0050] This embodiment also provides an IGBT bonding wire fault detection system based on stress wave two-dimensional image signals, including: A data acquisition program unit, which is used to collect the original one-dimensional stress wave signals generated during the switching process of the IGBT device and the background noise in the off state through an acoustic emission sensor for different types of faults and healthy states of the IGBT device respectively; A data noise reduction program unit, which is used to determine the frequency band distribution of the background noise through fast Fourier transform of the background noise, decompose the original one-dimensional stress wave signal into intrinsic mode functions in different frequency ranges by using the variational mode decomposition algorithm, remove the intrinsic mode functions in the frequency band where the background noise is located, and combine the remaining intrinsic mode functions to obtain a noise-reduced one-dimensional stress wave signal; An image generation program unit, which is used to perform two-dimensional image processing on the noise-reduced one-dimensional stress wave signal to obtain a two-dimensional image; A model training program unit, which constructs a two-dimensional image data set by attaching the fault or healthy state of the IGBT device to the two-dimensional image, and uses the two-dimensional image data set to train a convolutional neural network to establish a mapping relationship between the two-dimensional image and the IGBT bonding wire fault detection result for realizing IGBT bonding wire fault detection.

[0051] In this embodiment, before the data noise reduction program unit decomposes the original one-dimensional stress wave signal into intrinsic mode functions in different frequency ranges by using the variational mode decomposition algorithm, it also includes optimizing the penalty factor and decomposition layer number in the variational mode decomposition algorithm by using the particle swarm optimization algorithm to obtain ideal intrinsic mode functions more quickly. In the ideal intrinsic mode functions, the noise signals are distributed in one or more independent mode functions and do not overlap with the effective signals. The fitness function used when optimizing the penalty factor and decomposition layer number in the variational mode decomposition algorithm by using the particle swarm optimization algorithm is: , , , , wherein, represents the fitness function, , , are weights, is the reconstruction error feature, is the modal bandwidth feature, is the penalty term feature, is the decomposition level, is the penalty factor, is the original signal, is the reconstructed signal after variational mode decomposition, is the set of modes after decomposition, is the number of signals in the original signal, is the nth signal in the original signal, is the reconstructed signal the nth signal in, and are the frequency centers of the ith and jth modes respectively, is a small constant used to avoid division by zero errors, is to take the maximum value, is the preset maximum number of modes.

[0052] In this embodiment, the image generation program unit includes: A normalization program unit for scaling the denoised one-dimensional stress wave signal to the range of [-1, 1] according to the following formula: , In the above formula, is the ith sampling point of the scaled denoised one-dimensional stress wave signal in, is the ith sampling point of the denoised one-dimensional stress wave signal before scaling in, is the denoised one-dimensional stress wave signal, ; An angle conversion program unit for calculating the angle value after polar coordinate conversion of the denoised one-dimensional stress wave signal according to the following formula: , In the above formula, is the angle value after polar coordinate conversion of the ith sampling point, , The one-dimensional stress wave signal after noise reduction scaled to the range of [-1, 1]; An image calculation program unit for generating a Gram angle and field image GASF or a Gram angle difference field image GADF as the obtained two-dimensional image, where the functional expression for generating the Gram angle and field image GASF is: , The functional expression for generating the Gram angle difference field image GADF is: , where, is the Gram angle and field image GASF, ~ are respectively the angle values after polar coordinate transformation of the 1st to nth sampling points, is a unit row vector, is the one-dimensional stress wave signal after noise reduction scaled to the range of [-1, 1] the average value of, is the transposed vector of, is the Gram angle difference field image GADF.

[0053] In this embodiment, the convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer connected in sequence, where the input layer is used to input the stress wave two-dimensional image, and the output layer is used to output the bonding wire fault detection result of the IGBT device.

[0054] In this embodiment, the different types of faults collected by the data collection program unit include some or all of normal bonding wire, cracked bonding wire, single bonding wire detachment, and multiple bonding wire detachment.

[0055] In addition, this embodiment also provides an IGBT bonding wire fault detection system based on stress wave two-dimensional image signals, including a microprocessor and a memory connected to each other, and the microprocessor is programmed or configured to execute the IGBT bonding wire fault detection method based on stress wave two-dimensional image signals.

[0056] In addition, this embodiment also provides a computer-readable storage medium, in which a computer program or instruction is stored, and the computer program or instruction is programmed or configured to execute the IGBT bonding wire fault detection method based on stress wave two-dimensional image signals through a processor.

[0057] In addition, this embodiment also provides a computer program product, including a computer program or instruction, and the computer program or instruction is programmed or configured to execute the IGBT bonding wire fault detection method based on stress wave two-dimensional image signals through a processor.

[0058] Those skilled in the art should understand that the technical solutions provided by the embodiments of the present application can be in the form of a method, a system, or a computer program product. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for realizing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks. These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means realizes the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks. These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for realizing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0059] The above is only the preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for detecting IGBT bonding wire faults based on stress wave two-dimensional image signals, characterized in that: The steps include: For IGBT devices in different fault and healthy states, the original one-dimensional stress wave signal generated by the switching process of the IGBT device and the background noise in the off state are collected by acoustic emission sensors. The frequency distribution of the background noise is determined by fast Fourier transform of the background noise; the original one-dimensional stress wave signal is decomposed by a variational mode decomposition algorithm to obtain the intrinsic mode functions in different frequency ranges, the intrinsic mode functions in the frequency band where the background noise is located are removed, and the remaining intrinsic mode functions are combined to obtain the one-dimensional stress wave signal after noise reduction; Processing the noise-reduced one-dimensional stress wave signal into a two-dimensional image to obtain a two-dimensional image; A two-dimensional image dataset is constructed by adding the fault or healthy status of the IGBT device to the two-dimensional image. The convolutional neural network is trained using the two-dimensional image dataset to establish a mapping relationship between the two-dimensional image and the IGBT bonding wire fault detection results for realizing IGBT bonding wire fault detection.

2. The IGBT bonding wire fault detection method based on stress wave two-dimensional image signal according to claim 1, characterized in that: Before the original one-dimensional stress wave signal is decomposed by the variational modal decomposition algorithm to obtain the intrinsic mode functions in different frequency ranges, the method further includes optimizing the penalty factor and the number of decomposition layers in the variational modal decomposition algorithm by using a particle swarm optimization algorithm to obtain an ideal intrinsic mode function more quickly, in which the noise signal is distributed in one or more independent mode functions and is not aliased with the effective signal, and the fitness function used when the particle swarm optimization algorithm is used for the penalty factor and the number of decomposition layers in the variational modal decomposition algorithm is: ， ， ， ， in, represents the fitness function, , , is the weight, To reconstruct the error characteristics, is the modal bandwidth characteristic, is the penalty feature, is the number of decomposition layers, is the penalty factor, is the original signal, is the reconstructed signal after variational mode decomposition, is the decomposed mode set, is the number of signals in the original signal, is the nth signal in the original signal, To reconstruct the signal The nth signal in and are the frequency centers of the i-th and j-th modes respectively, is a small constant used to avoid division by zero errors, To obtain the maximum value, The maximum number of preset modes.

3. The IGBT bonding wire fault detection method based on stress wave two-dimensional image signal according to claim 1, characterized in that: The processing of converting the noise-reduced one-dimensional stress wave signal into a two-dimensional image to obtain a two-dimensional image comprises: The denoised one-dimensional stress wave signal is scaled to the range of [-1,1] according to the following formula: ， In the above formula, is the scaled and denoised one-dimensional stress wave signal The i-th sampling point in is the one-dimensional stress wave signal after noise reduction before scaling The i-th sampling point in is the one-dimensional stress wave signal after noise reduction, ; The angle value of the one-dimensional stress wave signal after noise reduction after polar coordinate conversion is calculated according to the following formula: ， In the above formula, is the angle value of the i-th sampling point after polar coordinate conversion, , is the one-dimensional stress wave signal after noise reduction scaled to the range of [-1,1]; A Gram angle sum field image GASF or a Gram angle difference field image GADF is generated as the obtained two-dimensional image, wherein the function expression for generating the Gram angle sum field image GASF is: ， The function expression for generating the Gram angle difference field image GADF is: ， in, is the Gram angle and field image GASF, ~ They are the angle values ​​of the 1st to nth sampling points after polar coordinate conversion. is the unit row vector, is the one-dimensional stress wave signal after noise reduction scaled to the range of [-1,1] The average value of for The transposed vector of is the Gram angle difference field image GADF.

4. The IGBT bonding wire fault detection method based on stress wave two-dimensional image signal according to claim 1, characterized in that: The convolutional neural network includes an input layer, a convolution layer, a pooling layer, a convolution layer, a pooling layer, a fully connected layer and an output layer connected in sequence, wherein the input layer is used to input a stress wave two-dimensional image, and the output layer is used to output a bonding wire fault detection result of an IGBT device.

5. The IGBT bonding wire fault detection method based on stress wave two-dimensional image signal according to claim 1, characterized in that: The different types of faults include a normal bonding wire, a cracked bonding wire, a single bonding wire falling off, or part or all of the falling off of multiple bonding wires.

6. An IGBT bonding wire fault detection system based on stress wave two-dimensional image signal, characterized in that: include: The data acquisition program unit is used to collect the original one-dimensional stress wave signal generated by the switching process of the IGBT device and the background noise in the off state through the acoustic emission sensor for the IGBT devices in different types of faults and healthy states; The data denoising program unit is used to determine the frequency band distribution of the background noise by fast Fourier transforming the background noise; decompose the original one-dimensional stress wave signal by using the variational mode decomposition algorithm to obtain the intrinsic mode functions in different frequency ranges, remove the intrinsic mode functions in the frequency band where the background noise is located, and combine the remaining intrinsic mode functions to obtain the one-dimensional stress wave signal after denoising; An image generation program unit is used for processing the noise-reduced one-dimensional stress wave signal into a two-dimensional image to obtain a two-dimensional image; The model training program unit constructs a two-dimensional image data set by attaching the fault or healthy state of the IGBT device to the two-dimensional image, and uses the two-dimensional image data set to train a convolutional neural network so as to establish a mapping relationship between the two-dimensional image and the IGBT bonding wire fault detection result for realizing IGBT bonding wire fault detection.

7. The IGBT bonding wire fault detection system based on stress wave two-dimensional image signal according to claim 6, characterized in that: The data denoising program unit further includes, before decomposing the original one-dimensional stress wave signal by using the variational modal decomposition algorithm to obtain the intrinsic modal functions in different frequency ranges, using the particle swarm optimization algorithm to optimize the penalty factor and the number of decomposition layers in the variational modal decomposition algorithm to obtain the ideal intrinsic modal function more quickly, in which the noise signal is distributed in one or more independent modal functions and is not aliased with the effective signal, and the fitness function used when using the particle swarm optimization algorithm for the penalty factor and the number of decomposition layers in the variational modal decomposition algorithm is: ， ， ， ， in, represents the fitness function, , , is the weight, To reconstruct the error characteristics, is the modal bandwidth characteristic, is the penalty feature, is the number of decomposition layers, is the penalty factor, is the original signal, is the reconstructed signal after variational mode decomposition, is the decomposed mode set, is the number of signals in the original signal, is the nth signal in the original signal, To reconstruct the signal The nth signal in and are the frequency centers of the i-th and j-th modes respectively, is a small constant used to avoid division by zero errors, To obtain the maximum value, The maximum number of preset modes.

8. The IGBT bonding wire fault detection system based on stress wave two-dimensional image signal according to claim 6, characterized in that: The image generation program unit comprises: The normalization program unit is used to scale the denoised one-dimensional stress wave signal to the range of [-1,1] according to the following formula: ， In the above formula, is the scaled and denoised one-dimensional stress wave signal The i-th sampling point in is the one-dimensional stress wave signal after noise reduction before scaling The i-th sampling point in is the one-dimensional stress wave signal after noise reduction, ; The angle conversion program unit is used to calculate the angle value of the one-dimensional stress wave signal after noise reduction after polar coordinate conversion according to the following formula: ， In the above formula, is the angle value of the i-th sampling point after polar coordinate conversion, , is the one-dimensional stress wave signal after noise reduction scaled to the range of [-1,1]; The image calculation program unit is used to generate a Gram angle and field image GASF or a Gram angle difference field image GADF as the obtained two-dimensional image, wherein the function expression for generating the Gram angle and field image GASF is: ， The function expression for generating the Gram angle difference field image GADF is: ， in, is the Gram angle and field image GASF, ~ They are the angle values ​​of the 1st to nth sampling points after polar coordinate conversion. is the unit row vector, is the one-dimensional stress wave signal after noise reduction scaled to the range of [-1,1] The average value of for The transposed vector of is the Gram angle difference field image GADF.

9. The IGBT bonding wire fault detection system based on stress wave two-dimensional image signal according to claim 6, characterized in that: The convolutional neural network includes an input layer, a convolution layer, a pooling layer, a convolution layer, a pooling layer, a fully connected layer and an output layer connected in sequence, wherein the input layer is used to input a stress wave two-dimensional image, and the output layer is used to output a bonding wire fault detection result of an IGBT device.

10. The IGBT bonding wire fault detection system based on stress wave two-dimensional image signal according to claim 6, characterized in that: The different types of faults collected by the data collection program unit include some or all of normal bonding wires, cracked bonding wires, single bonding wire detachment, and multiple bonding wire detachment.

11. An IGBT bonding wire fault detection system based on stress wave two-dimensional image signals, comprising a microprocessor and a memory connected to each other, characterized in that: The microprocessor is programmed or configured to execute the IGBT bonding wire fault detection method based on stress wave two-dimensional image signals as recited in any one of claims 1 to 5.

12. A computer-readable storage medium having a computer program or instruction stored therein, characterized in that: The computer program or instruction is programmed or configured to execute the IGBT bonding wire fault detection method based on stress wave two-dimensional image signals as recited in any one of claims 1 to 5 through a processor.

13. A computer program product comprising a computer program or instructions, characterized in that The computer program or instruction is programmed or configured to execute the IGBT bonding wire fault detection method based on stress wave two-dimensional image signals as recited in any one of claims 1 to 5 through a processor.

Images