Power module acoustic emission signal reconstruction method based on compressed sensing and deep learning

Abstract

The application discloses a power module acoustic emission signal reconstruction method based on compressed sensing and deep learning, comprising the following steps: performing compressed sensing sparse sampling on an original acoustic emission signal; performing processing on the compressed signal to obtain a preliminary reconstructed acoustic emission signal; and inputting the preliminary reconstructed acoustic emission signal into a trained VAE-GAN hybrid model to generate a final reconstructed signal. The power module acoustic emission signal reconstruction method based on compressed sensing and deep learning can more accurately restore key features and detailed information of the acoustic emission signal by combining the deep learning model of compressed sensing and the variational autoencoder-generative adversarial network, significantly improves the accuracy and noise resistance of signal reconstruction, reduces the data acquisition and processing cost, and provides reliable technical support for health monitoring and fault diagnosis of the power module.

Description

Technical Field

[0001] This invention belongs to the field of signal processing and fault diagnosis technology, and relates to a method for reconstructing acoustic emission signals of power modules based on compressed sensing and deep learning. Background Technology

[0002] With the rapid development of power electronics technology, power modules, as core components of power electronic systems, are widely used in power transmission, new energy, rail transportation, and other fields. The operating status of power modules directly affects the performance and reliability of the entire system. However, during long-term operation, especially under high power loads and extreme working environments, power modules are prone to internal component degradation due to factors such as thermal stress and electrical stress, leading to issues like bond wire breakage and solder layer fatigue. If these faults are not detected and addressed in a timely manner, they may cause a decline in module performance or even trigger system failures, seriously affecting industrial production and equipment safety.

[0003] Acoustic emission (AE) technology, as an effective non-destructive testing method, can monitor the stress release process inside materials or structures in real time and has been widely used in the health status monitoring of bearings and mechanical equipment. For power devices, AE technology has begun to be used as an effective means for power module condition monitoring. AE signals reflect the generation and propagation process of microscopic defects inside the power module, providing rich fault information. However, AE signals are typically non-stationary, high-noise, and sparsity-oriented, posing numerous challenges to signal processing and feature extraction. Traditional signal processing methods, such as Fourier transform and wavelet transform, struggle to accurately extract features from AE signals under high-noise and non-stationary conditions and usually require high sampling rates, increasing the cost of data acquisition and processing. In recent years, the development of compressed sensing (CS) theory and deep learning technology has brought new opportunities to the field of signal processing. Compressed sensing theory utilizes the sparsity of signals, reconstructing the original signal with a small number of random linear measurements, reducing data acquisition costs. Variational autoencoders (VAEs) and generative adversarial networks (GANs) in deep learning are important generative models with powerful data generation and feature extraction capabilities. However, combining compressed sensing with deep learning for efficient reconstruction of acoustic emission signals from power modules remains a challenging technical problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a method for reconstructing acoustic emission signals of power modules based on compressed sensing and deep learning, which solves the problem in the prior art that the combination of compressed sensing and deep learning in the reconstruction of acoustic emission signals of power modules cannot simultaneously achieve high reconstruction accuracy, strong noise resistance and low acquisition and processing cost.

[0005] The technical solution adopted in this invention is a method for reconstructing acoustic emission signals from power modules based on compressed sensing and deep learning, comprising: Step 1: Compressed sensing sparse sampling is performed on the original acoustic emission signal; Step 2: After processing the compressed signal, a preliminary reconstructed acoustic emission signal is obtained; Step 3: Input the initially reconstructed acoustic emission signal into the trained VAE-GAN (Variational Autoencoder-Generative Adversarial Network) hybrid model to generate the final reconstructed signal.

[0006] The invention is further characterized by: Step 1 includes: Step 1.1, Sparse Representation: Select a sparse basis to perform sparse representation on the original acoustic emission signal, and convert it into a sparse coefficient vector; Step 1.1.1: Convert the original acoustic emission signal into a sparse coefficient vector; Step 1.1.2: Obtain the sparse coefficients by performing signal transformation on the sparse coefficient vector; Step 1.2, Measurement Matrix Design: A matrix that satisfies the restricted equidistant property is used as the measurement matrix; Step 1.2.1: Generate a Gaussian random matrix; Step 1.2.2: Verify the generated measurement matrix through numerical verification or theoretical analysis to ensure that it meets the RIP conditions. If the measurement matrix meets the RIP conditions, proceed to step 1.3. If the measurement matrix does not meet the RIP conditions, take the following measures: regenerate the measurement matrix, optimize the measurement matrix design, increase the number of measurement values, or adjust the sparse basis. Step 1.3, Sparse Sampling: Perform linear projection on the sparse coefficient vector to obtain a low-dimensional observation vector.

[0007] Step 2 includes: Step 2.1: Initialize residuals and support sets; Step 2.2: In each iteration, select the basis vector most relevant to the current residual and add it to the support set; Step 2.3: Calculate the sparse solution using the least squares method; Step 2.4: Update the residuals; Step 2.5: Repeat the above steps until the preset stopping condition is met; Step 2.6: Obtain the preliminary reconstructed acoustic emission signal.

[0008] The VAE-GAN hybrid model consists of an encoder, a reparameterization node, a decoder, and a discriminator connected in sequence. The encoder maps the initially reconstructed signal to the latent space and outputs the mean and variance of the latent variables. The reparameterization node uses reparameterization techniques to convert the mean and variance of the latent variables into latent vectors. The decoder generates a reconstructed signal based on the latent vectors. The discriminator determines the authenticity of the generated reconstructed signal and the real signal.

[0009] The encoder includes an encoder input layer, a first convolutional layer, a first batch normalization layer, a second convolutional layer, a second batch normalization layer, and a pooling layer, which are connected in sequence. The pooling layer is connected to the first fully connected layer and the second fully connected layer, and the first fully connected layer and the second fully connected layer are connected to the reparameterization node. The decoder consists of a decoder input layer, a fully connected layer 3, a deconvolution layer 1, a batch normalization layer 3, a deconvolution layer 2, and a decoder output layer, which are connected in sequence. The fully connected layer 3 is connected to the reparameterization node, and the decoder output layer is connected to the discriminator. The discriminator consists of a discriminator input layer, a third convolutional layer, a fourth convolutional layer, a fourth batch normalization layer, a fourth fully connected layer, and a discriminator output layer, connected in sequence.

[0010] The trained VAE-GAN hybrid model is obtained through the following method: Step A1: Initialize model parameters: Randomly initialize all trainable parameters of the encoder, decoder, and discriminator; Step A2: Set the optimizer and learning rate; Step A3, Iterative Training Process: Within each training cycle, the training dataset is iterated multiple times. The following operations are performed in each iteration: Step A3.1: Train the discriminator; Step A3.2: Train the encoder and decoder; Step A4, Model Saving: After training is complete, save the trained model parameters; Step A5: Performance evaluation and analysis.

[0011] Step A3.1 includes: Step A3.1.1: Sample a batch of acoustic emission signals from a real dataset; Step A3.1.2: Sample a batch of noise vectors from the latent space; Step A3.1.3: Input the noise vector into the decoder to generate a fake signal; Step A3.1.4: Calculate the loss of the discriminator; Step A3.1.5: Fix the parameters of the encoder and decoder, repeat steps A3.1.1 to A3.1.5, and use the optimizer to update the parameters of the discriminator to minimize the loss of the discriminator.

[0012] Step A3.2 includes: Step A3.2.1: Sample a batch of acoustic emission signals from a real dataset; Step A3.2.2: Input the acoustic emission signal into the encoder to obtain the mean and logarithmic variance of the latent variables; Step A3.2.3: Obtain the latent vector by sampling from the mean and log-variance of the latent variables using the reparameterization technique; Step A3.2.4: Input the latent vector into the decoder to generate the reconstructed acoustic emission signal; Step A3.2.5: Calculate the joint loss function; Step A3.2.6: Fix the parameters of the discriminator, repeat steps A3.2.1 to A3.2.6, and use the optimizer to update the parameters of the encoder and decoder to minimize the joint loss function.

[0013] The joint loss function includes the reconstruction loss function, the KL divergence loss function (KL divergence loss function), and the adversarial loss function. The reconstruction loss function measures the difference between the reconstructed signal and the original acoustic emission signal, the KL divergence loss function constrains the distribution of latent variables, and the adversarial loss function improves the realism of the generated signal through adversarial training of the discriminator and the decoder.

[0014] Step A5 includes: Step A5.1, Performance Evaluation: Using the set indicators, evaluate the quality of the reconstructed signal and the noise resistance performance of the model; Step A5.2, Robustness Analysis: Under different noise levels and fault types, verify the noise resistance and generalization ability of the model, and compare the model's set indicators under different noise conditions; Step A5.2.1, Noise Level Setting: Add noise with different signal-to-noise ratios to the original acoustic emission signal; Step A5.2.2, Model Evaluation: Record the reconstruction effect of the model under different noise conditions using the set indicators; Step A5.2.3, Result Analysis: Compare the changes in the model's set indicators under different noise levels. If the set indicators meet the preset conditions, the training is complete; if the set indicators do not meet the preset conditions, repeat steps A1 to A5.

[0015] The beneficial effects of this invention are: Significantly improves reconstruction accuracy: By combining compressed sensing with a variational autoencoder-generative adversarial network deep learning model, the key features and details of acoustic emission signals can be recovered more accurately, significantly improving the accuracy and noise resistance of signal reconstruction, reducing data acquisition and processing costs, and providing reliable technical support for the health monitoring and fault diagnosis of power modules. Enhanced noise resistance: The model can maintain high signal reconstruction quality even in high-noise environments, has good robustness, can effectively suppress noise interference, and adapt to complex signal conditions in real industrial environments; Reduce data acquisition and processing costs: Compressed sensing technology is used to achieve sparse sampling at low sampling rates, which reduces the cost of data acquisition and storage. At the same time, efficient deep learning models reduce the computational resource requirements for signal processing. Improve real-time monitoring capabilities: By optimizing the model structure and training strategy, the computational efficiency of the model has been improved, enabling it to meet the needs of real-time health monitoring and fault diagnosis of the power module, thereby improving the system's response speed and reliability. Wide range of applications: It is not only suitable for acoustic emission signal reconstruction of power modules, but can also be extended to signal processing tasks in other industrial fields, such as mechanical fault diagnosis and biomedical signal analysis, with broad application prospects and market value; Flexible system integration: The modular design and low computational complexity of the model make it easy to integrate into existing industrial monitoring systems, facilitating deployment and expansion, and improving the overall performance and adaptability of the system; Scalable technical framework: By introducing technologies such as multimodal data fusion, attention mechanisms, and multi-scale feature extraction, the performance and application scope of the model can be further improved. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the power module acoustic emission signal reconstruction method based on compressed sensing and deep learning according to the present invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 This embodiment proposes a power module acoustic emission signal reconstruction method based on compressed sensing and deep learning, such as... Figure 1 As shown, it includes the following steps: Step 1: Compressed sensing sparse sampling is performed on the original acoustic emission signal; Step 2: After processing the compressed signal, a preliminary reconstructed acoustic emission signal is obtained; The present invention adopts an Orthogonal Matching Pursuit (OMP) algorithm to process the compressed signal and obtain a preliminarily reconstructed acoustic emission signal; Step 3: Input the preliminarily reconstructed acoustic emission signal into the trained VAE-GAN hybrid model to generate the final reconstructed signal.

[0019] Example 2 Based on Example 1, this embodiment proposes Step 1.1, Sparse representation: select a sparse basis (such as wavelet basis, discrete cosine basis, etc.) to perform sparse representation on the original acoustic emission signal and convert it into a sparse coefficient vector; Step 1.1.1: Convert the original acoustic emission signal into a sparse coefficient vector , where , x is the original acoustic emission signal; is a set of real numbers; Q is the dimension of the original acoustic emission signal; s is the sparse coefficient vector; K is the dimension of the sparse coefficient vector, and ; is a sparse basis matrix; Step 1.1.2: Adopt an appropriate sparse transformation tool (such as a wavelet transform tool kit) to perform signal transformation on the sparse coefficient vector to obtain sparse coefficients; Step 1.2, Measurement matrix design: adopt a Gaussian random matrix that satisfies the Restricted Isometry Property (RIP) or another matrix with low coherence as the measurement matrix and M < Q, wherein, represents the measurement matrix, M is the number of measurement values, and Q is the dimension of the original acoustic emission signal; indicates that the number of measurement values is less than the dimension of the original acoustic emission signal, which embodies the idea of compressed sampling and ensures accurate signal reconstruction; Step 1.2.1: Generate a Gaussian random matrix, whose matrix elements , wherein, represents the element at the -th row and -th column in the measurement matrix ; represents a Gaussian distribution with a mean of 0 and a variance of , and M is the number of measurement values; Step 1.2.2: Ensure the generated measurement matrix through numerical verification or theoretical analysis satisfies the RIP condition to guarantee accurate signal reconstruction; if the measurement matrix satisfies the RIP condition, perform step 1.3; if the measurement matrix Failure to meet the RIP conditions may lead to decreased signal reconstruction accuracy or even failure. The following measures should be taken: Regenerate the measurement matrix: Try to regenerate the measurement matrix using different parameters or different types of random matrices (such as Bernoulli random matrices, partial Fourier matrices, etc.) until the RIP condition is met or the set coherence condition is reached. Optimize measurement matrix design: Use optimization algorithms to design the measurement matrix to achieve better RIP characteristics or lower coherence; Increase the number of measurements If conditions permit, appropriately increase the number of measured values. This can increase the probability that the measurement matrix satisfies the RIP condition; Adjusting the sparse basis: Reselecting a sparse basis that is more suitable for the signal characteristics to reduce the measurement matrix. and sparse basis matrix Coherence; RIP condition: For a sparse coefficient vector There exists a constant , making , in, It is a measurement matrix; It is a sparse basis matrix; To constrain the equidistant constant; when The closer to 0, the better the measurement matrix. The longer the sparse signal length is preserved, the better the reconstruction performance; in practical applications, this is usually achieved by calculating the measurement matrix. and sparse basis matrix The coherence of the coefficients is used to indirectly evaluate the degree to which the RIP conditions are met. Low coherence is a sufficient condition for satisfying the RIP conditions; s is the sparse coefficient vector; K is the dimension of the sparse coefficient vector. Step 1.3, Sparse Sampling: Using the designed measurement matrix, linear projection is performed on the sparse coefficient vector to obtain the compressed signal, i.e., the low-dimensional observation vector, thereby reducing data acquisition and storage costs. The formula for calculating the low-dimensional observation vector is: , in, This represents a low-dimensional observation vector, which is the signal after compressed sampling; Represents the measurement matrix; Represents the original acoustic emission signal; Represents a sparse basis matrix; Represents a sparse coefficient vector; Representing the perceptual matrix, the sparse coefficient vector Direct mapping to low-dimensional observation vectors .

[0020] Example 3 Based on Example 1, this example proposes step 2.1, initializing the residual. low-dimensional observation vector Supports collection empty set ; Step 2.2: In each iteration, select the basis vector most relevant to the current residual and add it to the support set. ; Step 2.3: Calculate the sparse solution using the least squares method. That is, in the current support set Optimal sparse coefficient estimation on; Step 2.4: Update residuals ,in, Represents a low-dimensional observation vector; For measurement matrix Corresponding support set The submatrix; For sparse solutions, i.e., in the current support set Optimal sparse coefficient estimation on; Step 2.5: Repeat the above steps until the preset stopping condition is met (such as the residual being less than the preset threshold or the maximum number of iterations being reached). Step 2.6: Obtain the preliminary reconstructed acoustic emission signal .

[0021] Example 4 Based on Example 1, this example proposes a VAE-GAN hybrid model including an encoder, a reparameterized node, a decoder, and a discriminator connected in sequence; The encoder includes an encoder input layer, a first convolutional layer, a first batch normalization layer, a second convolutional layer, a second batch normalization layer, and a pooling layer, which are connected in sequence. The pooling layer is connected to the first fully connected layer and the second fully connected layer, and the first fully connected layer and the second fully connected layer are connected to the reparameterization node. Reparameterized nodes utilize reparameterization techniques to transform the mean and variance of latent variables into latent vectors. The calculation formula is as follows: , in, This represents the latent vector obtained by sampling from the latent space; The vector representing the mean of the latent variables is output by the encoder; The standard deviation vector (usually the square root of the variance) of the latent variables is output by the encoder; This represents element-wise multiplication; Indicates the distribution from the standard normal distribution The noise vector sampled in the middle; The decoder consists of a decoder input layer, a fully connected layer 3, a deconvolution layer 1, a batch normalization layer 3, a deconvolution layer 2, and a decoder output layer, which are connected in sequence. The fully connected layer 3 is connected to the reparameterization node, and the decoder output layer is connected to the discriminator. The discriminator consists of a discriminator input layer, a third convolutional layer, a fourth convolutional layer, a fourth batch normalization layer, a fourth fully connected layer, and a discriminator output layer, connected in sequence. In this invention, the encoder consists of multiple one-dimensional convolutional layers and fully connected layers, which map the initially reconstructed signal to the latent space and output the mean and variance of the latent variables. The reparameterization node uses reparameterization techniques to convert the mean and variance of the latent variables into latent vectors, which facilitates the subsequent generation process. The decoder consists of fully connected layers and multiple one-dimensional deconvolutional layers, which generate a high-quality reconstructed signal based on the latent vectors. The discriminator consists of multiple one-dimensional convolutional layers and fully connected layers, which distinguishes the authenticity of the generated reconstructed signal from the real signal.

[0022] The specific structure of the encoder in one embodiment of the present invention is as follows: Encoder input layer: Receives the initially reconstructed acoustic emission signal ; Convolutional layer 1: One-dimensional convolution, with 1 input channel and 16 output channels, a kernel size of 3, a stride of 1, padding of 1, and the activation function is Leaky ReLU (Leaky Rectified Linear Unit). First batch normalization layer: Performs batch normalization on the output of the first convolutional layer to improve training stability; Convolutional layer 2: One-dimensional convolution, with 16 input channels and 32 output channels, a kernel size of 3, a stride of 1, padding of 1, and the activation function Leaky ReLU; Second batch normalization layer: Performs batch normalization on the output of the second convolutional layer; Pooling layer: max pooling, pooling window size of 2, stride of 2, to reduce feature dimensionality; Fully connected layer 1: Maps the convolutional features to the latent space and outputs the mean vector of the latent variables; The second fully connected layer maps the convolutional features to the latent space and outputs the log-variance vector of the latent variables. ; The specific structure of the decoder in one embodiment of the present invention is as follows: Third fully connected layer: This layer stores the latent vectors... Mapped to a high-dimensional feature space, the output size is 32 × (signal length / 4). Deconvolution layer 1: One-dimensional deconvolution, with 32 input channels, 16 output channels, a kernel size of 3, a stride of 2, padding of 1, output padding of 1, and the activation function is ReLU (Rectified Linear Unit). Third Batch Normalization Layer: Performs batch normalization on the output of the first deconvolution layer to improve the stability of the generated signal; Deconvolutional layer 2: One-dimensional deconvolution with 16 input channels, 1 output channel, 3 kernels, 1 stride, 1 padding, and Tanh (Hyperbolic Tangent) activation function, used to generate the final reconstructed signal; The specific structure of the discriminator in one embodiment of the present invention is as follows: Discriminator input layer: Receives raw acoustic emission signals Or preliminary reconstructed acoustic emission signal ; Convolutional layer 3: One-dimensional convolution, with 1 input channel and 16 output channels, a kernel size of 3, a stride of 2, padding of 1, and the activation function is Leaky ReLU; Convolutional layer 4: One-dimensional convolution, with 16 input channels, 32 output channels, a kernel size of 3, a stride of 2, padding of 1, and the activation function Leaky ReLU; Fourth Batch Normalization Layer: Performs batch normalization on the output of the fourth convolutional layer to improve the stability and convergence speed of the discriminator; The fourth fully connected layer maps the convolutional features to a single discriminant probability. The activation function is the sigmoid function, used to output the discriminant result. .

[0023] Example 5 Based on Example 1, this example proposes that the trained VAE-GAN hybrid model is obtained in the following way: This invention employs a joint optimization strategy, minimizing the joint loss function by alternately training the encoder, decoder, and discriminator. The specific training steps are as follows: Step A1: Initialize model parameters: Randomly initialize all trainable parameters of the encoder, decoder, and discriminator; Step A2, Set the optimizer and learning rate: Use the Adam optimizer (Adaptive Moment Estimation) and set the initial learning rate (e.g., ), momentum parameters =0.5、 =0.999; Step A3, Iterative Training Process: Within each training cycle (iteration), the training dataset is iterated multiple times. The following operations are performed in each iteration: Step A3.1: Training the discriminator: Step A3.1.1: Sample a batch of acoustic emission signals from a real dataset; Step A3.1.2: Sample a batch of noise vectors from the latent space; Step A3.1.3: Input the noise vector into the decoder to generate a fake signal; Step A3.1.4: Calculate the loss of the discriminator; , in, This represents the loss function of the discriminator; Represents the original acoustic emission signal; Represents the actual data distribution; Indicates to Follows the true data distribution Expectations at that time; Discriminator For the original acoustic emission signal The output; This represents the logarithm of the discriminator's judgment result on the real data; Represents the latent vector; Represents the latent vector The prior distribution; Indicates to Follows prior distribution Expectations at that time; Represents generator According to the latent vector Generated fake data; Discriminator For generator Generated fake data The output; : Represents the logarithm of the discriminator's judgment result for false data, when When approaching 0, Approaching 1, Approaching 0; Step A3.1.5: With the encoder and decoder parameters fixed, repeat steps A3.1.1 to A3.1.5 to update the discriminator using the Adam optimizer. The parameters are set to minimize the discriminator's loss. ; Step A3.2, Training the encoder and decoder: Step A3.2.1: Sample a batch of acoustic emission signals from a real dataset; Step A3.2.2: Input the acoustic emission signal into the encoder. Obtain the mean and log-variance of the latent variables; Step A3.2.3: Obtain the latent vector by sampling from the mean and log-variance of the latent variables using the reparameterization technique; Step A3.2.4: Input the latent vector into the decoder to generate the reconstructed acoustic emission signal; Step A3.2.5: Calculate the joint loss function; The reconstruction loss measures the mean square error between the original acoustic emission signal and the initially reconstructed acoustic emission signal; the KL divergence loss constrains the distribution of latent variables to approximate a standard normal distribution; the adversarial loss, through feedback from the discriminator, prompts the decoder to generate a more realistic signal; Step A3.2.6: Fix the discriminator Given the parameters, repeat steps A3.2.1 to A3.2.6 to update the encoder using the Adam optimizer. and decoder The parameters are set to minimize the joint loss function. ; Step A4, Model Saving: After training is complete, save the trained model parameters for subsequent model performance evaluation. Regularization techniques such as batch normalization and random deactivation are applied to prevent overfitting and improve the model's generalization ability and training stability: Batch normalization layers are added after convolutional and deconvolutional layers to stabilize the training process and accelerate convergence; random deactivation layers are added to fully connected layers to prevent overfitting and improve the model's generalization ability; gradient clipping is introduced to clip gradients during backpropagation to prevent gradient explosion and ensure the stability of the training process; an early stopping strategy is used in model training and optimization: after each iteration, the model's performance metrics are evaluated on the validation set, and the loss on the validation set is monitored. When the loss on the validation set no longer decreases within a few consecutive iterations (e.g., 10-20 iterations), training is stopped early to prevent overfitting and to save the current optimal model parameters. The joint loss function includes the reconstruction loss function, the KL divergence loss function, and the adversarial loss function; The reconstruction loss function uses mean squared error (MSE) loss to measure the reconstructed signal. With the original acoustic emission signal The difference between them is expressed by the following formula: , in, Indicates the reconstruction loss; This indicates that, given the input, i.e., the original acoustic emission signal... Under the condition of latent variables posterior distribution Seeking expectations; Represents the original acoustic emission signal; This indicates the initial reconstruction of the acoustic emission signal; This represents the square of the L2 norm, i.e., the mean square error. Indicates encoder; Indicates the parameters of the encoder; The KL divergence loss function constrains the distribution of latent variables to approximate a standard normal distribution. The formula is as follows:

[0024] in, Indicates the KL divergence loss; Represents the posterior distribution of the latent variables With prior distribution KL divergence between them; It is the posterior distribution of the latent variables output by the encoder, usually assumed to be a Gaussian distribution, and its mean and variance are calculated by the encoder; It is the prior distribution of the latent variables, usually assumed to be a standard normal distribution. Representing latent variables The The mean of each dimension; Representing latent variables The Variance in each dimension; Is when the posterior distribution and prior distribution All are analytical forms of the KL divergence for a diagonal covariance Gaussian distribution, where... The dimension of the potential vector; Indicates encoder; Indicates the parameters of the encoder, Represents the latent vector; i represents the dimension index of the latent vector; The adversarial loss function improves the realism of the generated signal through adversarial training between the discriminator and the decoder. The specific formula is as follows: , in, Indicates resistance to loss; Represents the latent vector; Represents the latent vector The prior distribution; Indicates a generator; Discriminator For generator Generated fake data The output; The joint loss function combines the three types of losses mentioned above, and its formula is as follows: , in, For joint losses, To reconstruct the loss, For KL divergence loss, To counteract the loss, β and γ are weighting coefficients used to balance the contributions of various losses.

[0025] Example 6 Based on Example 5, this example proposes step A5.1, performance evaluation: using indicators such as root mean square error, signal-to-noise ratio and structural similarity index to evaluate the quality of the reconstructed signal and the noise resistance performance of the model; Step A5.2, Robustness Analysis: Under different noise levels and fault types, the noise resistance and generalization ability of the model are verified. By comparing the root mean square error, signal-to-noise ratio and structural similarity index of the model under different noise conditions, it is proved that the model can still maintain high reconstruction quality in high noise environment.

[0026] Step A5.2.1, Noise Level Setting: Add noise with different signal-to-noise ratios, such as 10dB, 20dB, 30dB, etc., to the original acoustic emission signal; Step A5.2.2, Model Evaluation: Record the reconstruction effect of the model under different noise conditions using the above performance indicators; Step A5.2.3, Result Analysis: Compare the changes in the root mean square error, signal-to-noise ratio, and structural similarity index of the model under different noise levels to verify the robustness of the model in a high-noise environment; if the set indicators meet the preset conditions, the training is complete; if the set indicators do not meet the preset conditions, repeat steps A1 to A5.

[0027] The specific analysis includes: calculating the root mean square error, signal-to-noise ratio, and structural similarity index under different noise levels; plotting the performance indicators as a function of noise level; comparing the performance differences of the present invention with traditional compressed sensing methods and other deep learning methods under various noise levels; and verifying through statistical analysis that the present invention can maintain high reconstruction quality under different noise conditions and has wide applicability.

[0028] The present invention has achieved significant technical advantages and economic benefits in actual implementation. The following will elaborate on the implementation effect of the present invention in detail with specific numerical data.

[0029] Original acoustic emission signal parameters in one embodiment of the present invention: Sampling frequency: The sampling frequency for acquiring the original acoustic emission signal is 1MHz.

[0030] Signal length: The signal length for a single acquisition is 2048 sampling points.

[0031] Dynamic range: The signal dynamic range is ±5V.

[0032] In one embodiment of the present invention, the compressed sensing sparse sampling parameters are as follows: Sparse basis: The Daubechies4 (db4) wavelet basis is used for sparse representation.

[0033] Sparsity: The sparsity of the signal in the db4 wavelet domain is approximately 128 (i.e., the first 128 largest coefficients can represent more than 95% of the energy of the original acoustic emission signal).

[0034] Measurement matrix: An M×Q Gaussian random matrix is ​​used as the measurement matrix.

[0035] Number of measurements: Set the number of measurements to 512, which corresponds to a compression ratio of [value missing]. =512 / 2048=0.25.

[0036] Constraint equidistant constant: After numerical verification, the generated measurement matrix has... The value is approximately 0.3, which meets the requirements for accurate signal reconstruction.

[0037] Preliminary signal reconstruction parameters in one embodiment of the present invention: Number of iterations: The maximum number of iterations is set to 128.

[0038] Stop threshold: residual norm less than .

[0039] VAE-GAN hybrid model parameters: Potential vector dimension: 64.

[0040] In one embodiment of the present invention, the weights of the loss function are: (KL divergence loss weights) (Adversarial loss weights).

[0041] In one embodiment of the present invention, the optimizer is the Adam optimizer, and the learning rate is... , .

[0042] Training batch size: 64.

[0043] Training cycles (number of iterations): 200.

[0044] Early stopping strategy: Monitor the loss on the validation set, and stop training if the loss does not decrease after 20 consecutive iterations.

[0045] In a certain embodiment of the present invention, when actually deployed in a power module health monitoring system, the following quantitative results were achieved by simulating different operating conditions and fault types in a laboratory environment and comparing them with existing mainstream methods: Improved signal reconstruction performance: Reconstruction accuracy: In a noisy environment with a signal-to-noise ratio (SNR) of 20 dB, the root mean square error of the reconstructed signal of this invention is reduced by 62.5% compared to the traditional compressed sensing method (from 0.080 to 0.030), and the structural similarity index is improved by 14.3% (from 0.84 to 0.96). In a strong noise environment with a SNR of 10 dB, the reconstruction SNR of this invention reaches 26.5 dB, which is much higher than the 19.8 dB of the traditional method, indicating that it can still effectively recover signals under harsh conditions.

[0046] Noise robustness: During the process of signal-to-noise ratio decreasing from 30dB to 10dB, the reconstruction performance index of the present invention decreased by less than 15%, while the traditional method decreased by more than 30%, which proves the excellent robustness of the present invention over a wide noise range.

[0047] Cost-effectiveness of data acquisition and processing: Data acquisition cost: By using a compression ratio of 0.25, the amount of data collected for each signal is reduced from 2048 points to 512 points. This means that under the same monitoring duration, the required data storage space and transmission bandwidth are reduced by 75%. For example, a monitoring point that generates 1GB of raw acoustic emission data per day only needs to store 250MB of data per day after adopting this invention, saving approximately 270GB of storage space per year.

[0048] Computational resource consumption: When deployed on an embedded platform (such as NVIDIA Jetson Nano), the inference time of the model of this invention is 25 milliseconds for a single signal reconstruction, the CPU (Central Processing Unit) utilization is less than 15%, and the memory usage is approximately 80MB. This makes it possible to achieve real-time monitoring on microcontrollers or edge computing devices, avoiding reliance on expensive cloud computing resources.

[0049] Real-time monitoring and fault diagnosis efficiency: Real-time performance: The average end-to-end delay from signal acquisition to final reconstructed signal output in this invention is 30 milliseconds. This fully meets the millisecond-level fault response requirements of power modules, enabling near real-time fault early warning.

[0050] Diagnostic accuracy: When the high-quality acoustic emission signal reconstructed by the present invention is input into the subsequent fault diagnosis classifier, the accuracy of fault diagnosis is improved by 5% to 8% (e.g., from 90% to 95%) compared with the original noise signal, and the false alarm rate is reduced by 10%.

[0051] Generalization ability and application prospects: Multiple fault type adaptation: The present invention exhibits similar excellent performance in the reconstruction of acoustic emission signals for three typical power module faults: bonding wire breakage, solder layer fatigue, and chip cracking. The similarity index of the reconstructed structure is above 0.90, which verifies its good generalization ability for different fault modes.

[0052] Cross-domain potential: The versatility of this invention allows it to be extended to signal processing tasks in other industrial fields, such as vibration signal analysis of mechanical equipment and denoising and reconstruction of biomedical signals (such as electrocardiogram and electroencephalogram), which has broad market application prospects.

[0053] In summary, this invention, through the optimization of key parameters and the innovative combination of deep learning models, not only theoretically solves the problem of reconstructing acoustic emission signals from power modules, but also demonstrates outstanding performance advantages and significant economic benefits in practical implementation, providing strong technical support for intelligent monitoring and fault diagnosis in the industrial field.

Claims

1. A method for reconstructing acoustic emission signals from power modules based on compressed sensing and deep learning, characterized in that, include: Step 1: Compressed sensing sparse sampling is performed on the original acoustic emission signal; Step 2: After processing the compressed signal, a preliminary reconstructed acoustic emission signal is obtained; Step 3: Input the initially reconstructed acoustic emission signal into the trained VAE-GAN hybrid model to generate the final reconstructed signal; The VAE-GAN hybrid model includes an encoder, a reparameterization node, a decoder, and a discriminator connected in sequence; the encoder maps the initially reconstructed signal to the latent space and outputs the mean and variance of the latent variables; The reparameterization node uses reparameterization techniques to convert the mean and variance of latent variables into latent vectors; the decoder generates a reconstructed signal based on the latent vectors; and the discriminator determines the authenticity of the generated reconstructed signal compared to the real signal. The encoder includes an encoder input layer, a first convolutional layer, a first batch normalization layer, a second convolutional layer, a second batch normalization layer, and a pooling layer connected in sequence. The pooling layer is connected to the first fully connected layer and the second fully connected layer, and the first fully connected layer and the second fully connected layer are connected to the reparameterization node. The decoder includes a decoder input layer, a fully connected layer 3, a deconvolution layer 1, a batch normalization layer 3, a deconvolution layer 2, and a decoder output layer connected in sequence. The fully connected layer 3 is connected to the reparameterization node, and the decoder output layer is connected to the discriminator. The discriminator comprises a discriminator input layer, a third convolutional layer, a fourth convolutional layer, a fourth batch normalization layer, a fourth fully connected layer, and a discriminator output layer, connected in sequence.

2. The method for reconstructing acoustic emission signals of a power module based on compressed sensing and deep learning according to claim 1, characterized in that, Step 1 includes: Step 1.1, Sparse Representation: Select a sparse basis to perform sparse representation on the original acoustic emission signal, and convert it into a sparse coefficient vector; Step 1.1.1: Convert the original acoustic emission signal into a sparse coefficient vector; Step 1.1.2: Obtain the sparse coefficients by performing signal transformation on the sparse coefficient vector; Step 1.2, Measurement Matrix Design: A matrix that satisfies the restricted equidistant property is used as the measurement matrix; Step 1.2.1: Generate a Gaussian random matrix; Step 1.2.2: Verify the generated measurement matrix through numerical verification or theoretical analysis to ensure that it meets the RIP conditions. If the measurement matrix meets the RIP conditions, proceed to step 1.

3. If the measurement matrix does not meet the RIP conditions, take the following measures: regenerate the measurement matrix, optimize the measurement matrix design, increase the number of measurement values, or adjust the sparse basis. Step 1.3, Sparse Sampling: Perform linear projection on the sparse coefficient vector to obtain a low-dimensional observation vector.

3. The method for reconstructing acoustic emission signals of a power module based on compressed sensing and deep learning according to claim 1, characterized in that, Step 2 includes: Step 2.1: Initialize residuals and support sets; Step 2.2: In each iteration, select the basis vector most relevant to the current residual and add it to the support set; Step 2.3: Calculate the sparse solution using the least squares method; Step 2.4: Update the residuals; Step 2.5: Repeat the above steps until the preset stopping condition is met; Step 2.6: Obtain the preliminary reconstructed acoustic emission signal.

4. The method for reconstructing acoustic emission signals of a power module based on compressed sensing and deep learning according to claim 1, characterized in that, The trained VAE-GAN hybrid model is obtained through the following method: Step A1: Initialize model parameters: Randomly initialize all trainable parameters of the encoder, decoder, and discriminator; Step A2: Set the optimizer and learning rate; Step A3, Iterative Training Process: Within each training cycle, the training dataset is iterated multiple times. The following operations are performed in each iteration: Step A3.1: Train the discriminator; Step A3.2: Train the encoder and decoder; Step A4, Model Saving: After training is complete, save the trained model parameters; Step A5: Performance evaluation and analysis.

5. The method for reconstructing acoustic emission signals of a power module based on compressed sensing and deep learning according to claim 4, characterized in that, Step A3.1 includes: Step A3.1.1: Sample a batch of acoustic emission signals from a real dataset; Step A3.1.2: Sample a batch of noise vectors from the latent space; Step A3.1.3: Input the noise vector into the decoder to generate a fake signal; Step A3.1.4: Calculate the loss of the discriminator; Step A3.1.5: Fix the parameters of the encoder and decoder, repeat steps A3.1.1 to A3.1.5, and use the optimizer to update the parameters of the discriminator to minimize the loss of the discriminator.

6. The method for reconstructing acoustic emission signals of a power module based on compressed sensing and deep learning according to claim 4, characterized in that, Step A3.2 includes: Step A3.2.1: Sample a batch of acoustic emission signals from a real dataset; Step A3.2.2: Input the acoustic emission signal into the encoder to obtain the mean and logarithmic variance of the latent variables; Step A3.2.3: Obtain the latent vector by sampling from the mean and log-variance of the latent variables using the reparameterization technique; Step A3.2.4: Input the latent vector into the decoder to generate the reconstructed acoustic emission signal; Step A3.2.5: Calculate the joint loss function; Step A3.2.6: Fix the parameters of the discriminator, repeat steps A3.2.1 to A3.2.6, and use the optimizer to update the parameters of the encoder and decoder to minimize the joint loss function.

7. The method for reconstructing acoustic emission signals of a power module based on compressed sensing and deep learning according to claim 6, characterized in that, The joint loss function includes a reconstruction loss function, a KL divergence loss function, and an adversarial loss function. The reconstruction loss function measures the difference between the reconstructed signal and the original acoustic emission signal, the KL divergence loss function constrains the distribution of latent variables, and the adversarial loss function improves the realism of the generated signal through adversarial training of the discriminator and the decoder.

8. The method for reconstructing acoustic emission signals of a power module based on compressed sensing and deep learning according to claim 4, characterized in that, Step A5 includes: Step A5.1, Performance Evaluation: Using the set indicators, evaluate the quality of the reconstructed signal and the noise resistance performance of the model; Step A5.2, Robustness Analysis: Under different noise levels and fault types, verify the noise resistance and generalization ability of the model, and compare the model's set indicators under different noise conditions; Step A5.2.1, Noise Level Setting: Add noise with different signal-to-noise ratios to the original acoustic emission signal; Step A5.2.2, Model Evaluation: Record the reconstruction effect of the model under different noise conditions using the set indicators; Step A5.2.3, Result Analysis: Compare the changes in the model's set indicators under different noise levels. If the set indicators meet the preset conditions, the training is complete; if the set indicators do not meet the preset conditions, repeat steps A1 to A5.

Images