A high-voltage power battery intelligent fault diagnosis method based on wavelet denoising fusion time-frequency image and two-dimensional convolutional neural network

Abstract

The application discloses a high-voltage power battery intelligent fault diagnosis method based on wavelet denoising fusion time-frequency image and a two-dimensional convolutional neural network. First, the original time-domain acoustic detection signal is collected by means of a multi-channel acoustic sensor, and the original acoustic detection signal is pretreated by using a wavelet denoising technology, so as to improve the signal quality. Then, the denoised time-domain acoustic detection signal is converted into a two-dimensional time-frequency image by using a Wigner-Ville distribution time-frequency conversion method, and the fault characteristics in the time-frequency domain are clearly presented. Secondly, the two-dimensional time-frequency image is input into a two-dimensional convolutional neural network model for training, and the trained model is applied to the intelligent fault diagnosis of the high-voltage power battery. The fault recognition accuracy of 99.19% is achieved on the experimental data set, and the method exhibits more superior noise robustness and classification performance compared with other methods. The results show that the method has a wide application prospect in the high-voltage power battery fault diagnosis in a complex noise environment.

Description

Technical Field

[0001] This invention relates to an intelligent fault diagnosis method for high-voltage power batteries based on wavelet denoising fusion of time-frequency images and two-dimensional convolutional neural networks. It is used for noise reduction, feature extraction and intelligent recognition of acoustic detection electrical signals of high-voltage power batteries under complex working conditions, and belongs to the field of high-voltage power battery fault diagnosis technology. Background Technology

[0002] As a key energy storage medium, high-voltage power batteries have been widely used in new energy vehicles, rail transit, and energy storage power stations. With the continuous improvement of battery system energy density and the expansion of application scale, their operational safety issues are becoming increasingly prominent. Among these, the challenges posed by the decline in the insulation performance of high-voltage power batteries are particularly severe. This can not only cause energy leakage, reduced system efficiency, and equipment damage, but also potentially lead to electric shock, short circuits, and even thermal runaway fires, posing a serious threat to people's lives and property and the stable operation of equipment. Therefore, developing efficient and accurate high-voltage power battery insulation testing technology has significant theoretical and engineering value, and is crucial for ensuring the safe and stable operation of battery systems throughout their entire life cycle.

[0003] Traditional high-voltage power battery testing methods mainly include DC leakage current method, low-frequency AC injection method, and balanced bridge measurement. While these conventional methods demonstrate effectiveness in specific scenarios, they generally suffer from drawbacks such as long testing cycles, high costs, susceptibility to environmental interference, and reliance on manual feature extraction. In recent years, with the development of artificial intelligence technology, data-driven insulation testing systems have received widespread attention. Industry experts have developed various AI-based testing methods using signals collected from various sensors. These methods can effectively identify and classify characteristic signals of high-voltage power battery insulation under complex operating conditions. Representative methods include convolutional neural networks, Transformer, and random forests. Existing research shows that the fusion of time-frequency transformation and deep learning can effectively mine deep features of partial discharge signals, providing a new technical path for the insulation condition assessment of power equipment.

[0004] Despite significant progress, existing insulation diagnosis methods based on the fusion of symmetric point patterns and convolutional neural network models still face three major challenges in practical engineering applications such as insulation fault diagnosis of high-voltage power batteries: (1) Weak fault features are easily obscured under strong noise interference. The battery system environment is complex, and existing discrete wavelet transform uses fixed threshold denoising, which makes it difficult to achieve a balance between suppressing noise and preserving transient fault features. This results in subsequent symmetric point pattern images being dominated by noise, reducing the classification accuracy of the convolutional neural network model. (2) The construction of symmetric point pattern images depends on empirical parameters, limiting the generalization ability. The construction of symmetric point pattern images requires manually setting parameters such as time interval coefficients and angle magnification factors. Different parameters generate images with huge differences. These parameters are optimized for cables but cannot be adaptively transferred to battery packs or different working conditions, limiting the generalization ability of the model. (3) Battery insulation is a low-probability event, and there are very few early micro-insulation samples. Conventional convolutional neural network models rely on a large amount of balanced data and are prone to overfitting. Existing hybrid models do not adequately mine the temporal correlation features of insulation detection signals, and it is difficult to establish a reliable mapping under conditions of few samples, which limits the engineering applicability of the method. To address the aforementioned issues, there is an urgent need to research adaptive time-frequency image denoising and enhancement techniques, robust two-dimensional convolutional neural network feature extraction architectures, and fault diagnosis mechanisms suitable for small sample conditions, in order to overcome the performance bottlenecks of current methods under conditions of strong noise interference and data scarcity. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a smart fault diagnosis method for high-voltage power batteries based on wavelet denoising fusion of time-frequency images and two-dimensional convolutional neural networks, which can solve problems such as the inability to effectively extract insulation fault features with high characterization ability under strong noise interference and small sample conditions.

[0006] To solve the above-mentioned technical problems, one technical solution adopted by the present invention is as follows: The present invention provides a high-voltage power battery intelligent fault diagnosis method based on wavelet denoising fusion of time-frequency images and two-dimensional convolutional neural networks. By using custom wavelet denoising and Wigner distribution function time-frequency technology on the acoustic detection signal of high-voltage power battery insulation detection, deep time-frequency features that resist noise interference are extracted. Furthermore, a two-dimensional convolutional neural network is fused to mine the multi-level spatial patterns of the acoustic detection signal, thereby constructing a feature set with strong representational capabilities for fault diagnosis.

[0007] A smart fault diagnosis method for high-voltage power batteries based on wavelet denoising fusion of time-frequency images and two-dimensional convolutional neural networks includes the following steps:

[0008] 1) Acquire two channels of acoustic detection electrical signals using an acoustic sensor and discretize them into digital acoustic detection signals.

[0009] 2) The wavelet denoising algorithm is used to denoise the acoustic detection digital signal in step 1) to eliminate noise interference, and then the denoised data is batch normalized.

[0010] 3) Convert the preprocessed acoustic detection digital signal into a time-frequency graph using the Wigner distribution function to construct a training sample set and a test sample set.

[0011] 4) A two-dimensional convolutional neural network is used to train the training sample set in step 3) to construct a high-voltage power battery intelligent fault diagnosis model.

[0012] 5) Using the intelligent fault diagnosis model for high-voltage power batteries in step 4), test the test sample set in step 3).

[0013] 6) Evaluate and visually demonstrate the performance of the high-voltage power battery intelligent fault diagnosis model in step 4) by using confusion matrix and classification result comparison chart.

[0014] In step 2), the wavelet denoising algorithm decomposes the acoustic detection signal and noise components through discrete wavelet transform at multiple scales. Given a noisy original acoustic detection signal S, the decomposed acoustic detection signal obtained through discrete wavelet transform can be expressed as:

[0015]

[0016]

[0017] Here, and These are the scaling function and the wavelet function, respectively. Denotes the approximation coefficients of the Lth layer. Let be the detail coefficient of the j-th layer.

[0018] The decomposed acoustic detection signal is reconstructed by inverse discrete wavelet transform (IDT), which utilizes a threshold function. The processed low-frequency approximation coefficients c and high-frequency detail coefficients at various scales , can be represented as:

[0019]

[0020] In step 3), the Wigner distribution function is a time-frequency analysis tool used to analyze non-stationary acoustic detection signals. For the acoustic detection signal x(t), its Wigner distribution is defined as:

[0021]

[0022] Where t represents time and f represents frequency. Let τ represent the complex conjugate of the acoustic detection signal, and τ be the time delay variable.

[0023] The Wigner distribution can achieve theoretically optimal resolution in both the time and frequency domains, clearly characterizing instantaneous frequency changes. The boundary conditions satisfied by its time-frequency energy distribution can be expressed as:

[0024]

[0025] Let the acoustic detection signal be x(t) = x1(t) + x2(t), and its Wigner distribution function be expressed as:

[0026]

[0027] Among them W x1 , x2 It is a cross term between components, which manifests as oscillating pseudo-energy.

[0028] The two-dimensional convolutional neural network model in step 4) is a deep learning architecture, comprising convolutional layers, activation functions, pooling layers, fully connected layers, and an output layer. Through the convolutional and pooling layers, convolution and pooling operations are performed, and the two-dimensional convolutional neural network extracts input data features layer by layer, ultimately achieving intelligent fault diagnosis of high-voltage power batteries via the fully connected layer.

[0029] In step 5), the test sample set test visualizes the diagnostic results by comparing the confusion matrix and the classification results. On the experimental dataset, the fault identification accuracy reached 99.19%, demonstrating excellent performance.

[0030] Step 2) of batch normalization includes the following steps:

[0031] (1) Calculate the mean and variance:

[0032] For a batch of input data X={x (1) ,x (2) ,...,x (m) The batch normalized mean and variance can be expressed as follows:

[0033]

[0034] in This represents the average value of the input in the current batch; The variance of the current batch input is represented by m; m is the number of samples in the batch; x (i) This represents the input value of the i-th sample in a certain feature dimension of this layer; i is the index of the sample in the batch, from 1 to m; B refers to the label of the current batch, used to distinguish different batches in training.

[0035] (2) Normalization:

[0036]

[0037] in This represents the standardized value of the i-th sample; It is a very small constant introduced to avoid division by zero.

[0038] (3) Scaling and translation:

[0039]

[0040] γ and β are trainable parameters used to recover the expressive power of the model.

[0041] During the reconstruction process, three main types of deviations exist: threshold processing bias, boundary effect error, and wavelet basis mismatch error. The threshold processing bias is introduced by adjusting the high-frequency coefficients through a threshold function and can be quantified as the original high-frequency coefficients at each scale. With processed coefficients The L2 norm of the difference between them can be expressed as:

[0042]

[0043] Boundary effect error is caused by the boundary processing of the acoustic detection signal in wavelet transform, and its magnitude is negatively correlated with the number of transform layers L, which can be expressed as:

[0044]

[0045] in It is a positive coefficient related to the wavelet basis and acoustic detection signal characteristics.

[0046] Wavelet basis mismatch error originates from the selected wavelet basis. The mismatch between the acoustic detection signal S and its projection onto the wavelet basis is the difference between the actual acoustic detection signal S and its characteristics. The minimum deviation between them can be expressed as:

[0047]

[0048] Where B represents the set of wavelet basis functions, reflecting the optimal approximation of the acoustic detection signal by the wavelet basis.

[0049] The two-dimensional convolutional neural network model employs two sets of convolutional layers with kernel sizes of {(3,3)} and (5,5)}, and compares them with combinations of {(3,3), (3,3)}, {(5,5), (5,5)}, {(7,7), (7,7)}, {(3,3), (7,7)}, and {(5,5), (7,7)}. The model consists of multiple convolutional layers, pooling layers, and fully connected layers, and can extract multi-level features from the input data to ultimately complete the fault diagnosis task.

[0050] The beneficial effects of this invention are as follows: This invention proposes an intelligent fault diagnosis method for high-voltage power batteries based on wavelet denoising fusion of time-frequency images and two-dimensional convolutional neural networks. This method introduces cross-domain acoustic detection signal data from noisy and ideal scenarios to assist in the diagnostic task, effectively solving the problems of scarce target data and insufficient labels under small sample conditions. Deep time-frequency features are extracted using the Wigner distribution function, and wavelet denoising technology is fused with spatial pattern features of the convolutional neural network model to construct a feature set with strong insulation detection capabilities. The acoustic detection signal distribution structure is reconstructed through wavelet decomposition, achieving inter-domain distribution alignment between noisy and ideal scenarios, reducing the feature distribution differences under complex noise interference, and improving the accuracy of high-voltage insulation identification. This method enhances the environmental adaptability of high-voltage power battery insulation diagnosis under small-sample, high-noise operating conditions, and its computational efficiency is superior to traditional hybrid models, reducing industrial deployment costs. Attached Figure Description

[0051] Figure 1 The accompanying drawings, which are part of the abstract of this invention, illustrate the algorithm flow of the method.

[0052] Figure 2 The original acoustic detection signal sample of the present invention

[0053] Figure 3 The sample of this invention after being denoised using wavelet denoising technology

[0054] Figure 4 The two-dimensional time-frequency graph converted by the short-time Fourier transform comparison method of this invention.

[0055] Figure 5 The two-dimensional time-frequency graph transformed by the Hilbert-Huang transform comparison method of the present invention

[0056] Figure 6 Two-dimensional time-frequency graph converted by the Mel spectrum transform comparison method of the present invention

[0057] Figure 7 Two-dimensional time-frequency diagram of the conversion of the method proposed in this invention.

[0058] Figure 8 This is a diagram of the two-dimensional convolutional neural network model of the present invention.

[0059] Figure 9 The fault diagnosis results of the short-time Fourier transform comparison method of the present invention

[0060] Figure 10 The fault diagnosis results of the Hilbert-Huang transform comparison method of the present invention

[0061] Figure 11 The fault diagnosis results of the Mel spectrum transformation comparison method of the present invention

[0062] Figure 12 The fault diagnosis results of the method proposed in this invention

[0063] Figure 13 The confusion matrix for fault diagnosis of the convolutional neural network model-support vector machine model of this invention.

[0064] Figure 14 The confusion matrix for fault diagnosis of the convolutional neural network model-long short-term memory recurrent neural network model of this invention.

[0065] Figure 15 The confusion matrix for fault diagnosis of the convolutional neural network model-bidirectional gated recurrent unit model of this invention.

[0066] Figure 16 The fault diagnosis confusion matrix of the proposed model in this invention Detailed Implementation

[0067] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0068] like Figure 1 As shown in the figure, the intelligent fault diagnosis method for high-voltage power batteries based on wavelet denoising fusion of time-frequency images and two-dimensional convolutional neural networks in this embodiment includes the following steps:

[0069] 1) Acquire two-channel acoustic detection electrical signals using an acoustic sensor and discretize them into digital acoustic detection signals. The original acoustic detection signal samples are shown below. Figure 2 As shown.

[0070] 2) The wavelet denoising algorithm is used to denoise the acoustic detection digital signal in step 1) to eliminate noise interference. Then, the denoised data is batch normalized. The samples after wavelet denoising are as follows: Figure 3 As shown.

[0071] Step 2) The wavelet denoising algorithm decomposes the acoustic detection signal and noise components through discrete wavelet transform at multiple scales. Given a noisy original acoustic detection signal S, the decomposed acoustic detection signal obtained through discrete wavelet transform can be expressed as:

[0072]

[0073]

[0074] Here, and These are the scaling function and the wavelet function, respectively. Denotes the approximation coefficients of the Lth layer. Let be the detail coefficients of the j-th layer. Table 1 shows the frequency domain distribution characteristics of the wavelet coefficients:

[0075] Table 1. Frequency distribution characteristics of wavelet coefficients

[0076] coefficient type Frequency range Physical meaning noise sensitivity <![CDATA[cD1]]> <![CDATA[[f s / 4, f s / 2]]]> The finest level of detail Extremely high <![CDATA[cA1]]> <![CDATA[[f s / 8, f s / 4]]]> Medium-sized details high <![CDATA[cD j ]]> <![CDATA[[f s / 4 L+1 , f s / 2 L ]]]> Roughest scale details middle <![CDATA[cA L ]]> <![CDATA[[0, f s / 2 L+1 ]]]> Acoustic detection signal characteristics Low

[0077] The decomposed acoustic detection signal is reconstructed by inverse discrete wavelet transform (IDT), which utilizes a threshold function. The processed low-frequency approximation coefficients c and high-frequency detail coefficients at various scales , can be represented as:

[0078]

[0079] During the reconstruction process, three main types of deviations exist: threshold processing bias, boundary effect error, and wavelet basis mismatch error. The threshold processing bias is introduced by adjusting the high-frequency coefficients through a threshold function and can be quantified as the original high-frequency coefficients at each scale. With processed coefficients The L2 norm of the difference between them can be expressed as:

[0080]

[0081] Boundary effect error is caused by the boundary processing of the acoustic detection signal in wavelet transform, and its magnitude is negatively correlated with the number of transform layers L, which can be expressed as:

[0082]

[0083] in It is a positive coefficient related to the wavelet basis and acoustic detection signal characteristics.

[0084] Wavelet basis mismatch error originates from the selected wavelet basis. The mismatch between the acoustic detection signal S and its projection onto the wavelet basis is the difference between the actual acoustic detection signal S and its characteristics. The minimum deviation between them can be expressed as:

[0085]

[0086] Where B represents the set of wavelet basis functions, reflecting the optimal approximation of the acoustic detection signal by the wavelet basis.

[0087] Step 2) of batch normalization includes the following steps:

[0088] (1) Calculate the mean and variance:

[0089] For a batch of input data X={x (1) ,x (2) ,...,x (m) The batch normalized mean and variance can be expressed as follows:

[0090]

[0091] in This represents the average value of the input in the current batch; The variance of the current batch input is represented by m; m is the number of samples in the batch; x (i) This represents the input value of the i-th sample in a certain feature dimension of this layer; i is the index of the sample in the batch, from 1 to m; B refers to the label of the current batch, used to distinguish different batches in training.

[0092] (2) Normalization:

[0093]

[0094] in This represents the standardized value of the i-th sample; It is a very small constant introduced to avoid division by zero.

[0095] (3) Scaling and translation:

[0096]

[0097] γ and β are trainable parameters used to recover the expressive power of the model.

[0098] 3) The preprocessed acoustic detection digital signal is converted into a time-frequency graph using the Wigner distribution function to construct training and testing sample sets. The time-frequency graph is then compared with those generated by other methods, including short-time Fourier transform, Hilbert-Huang transform, and Mel-spectral transform. The results are as follows: Figure 4 , Figure 5 , Figure 6 and Figure 7 As shown.

[0099] In step 3), the Wigner distribution function is a time-frequency analysis tool used to analyze non-stationary acoustic detection signals. For the acoustic detection signal x(t), its Wigner distribution is defined as:

[0100]

[0101] Where t represents time and f represents frequency. Let τ represent the complex conjugate of the acoustic detection signal, and τ be the time delay variable.

[0102] The Wigner distribution achieves theoretically optimal resolution in both the time and frequency domains, clearly characterizing instantaneous frequency changes. As a quadratic representation of the energy of a sound detection signal, the Wigner distribution inherently contains cross terms: its advantage lies in revealing the interactions between the components of the sound detection signal, while its disadvantage is that it can lead to spurious energy distributions in multi-component sound detection signals. The boundary conditions satisfied by this time-frequency energy distribution can be expressed as:

[0103]

[0104] Let the acoustic detection signal be x(t) = x1(t) + x2(t), and its Wigner distribution function be expressed as:

[0105]

[0106] Among them W x1 , x2 It is a cross term between components, which manifests as oscillating pseudo-energy.

[0107] 4) A two-dimensional convolutional neural network is used to train the training sample set in step 3), thereby constructing an intelligent fault diagnosis model for high-voltage power batteries, such as... Figure 8 As shown.

[0108] The two-dimensional convolutional neural network model in step 4) is a deep learning architecture, comprising convolutional layers, activation functions, pooling layers, fully connected layers, and an output layer. Through the convolutional and pooling layers, convolution and pooling operations are performed. The two-dimensional convolutional neural network extracts input data features layer by layer, and finally, through the fully connected layer, intelligent fault diagnosis of the high-voltage power battery is achieved. The structural parameters are shown in Table 2.

[0109] Table 2. Structural parameters of the 2D Convolutional Neural Network (2D-CNN) model

[0110] Layer type kernel size stride number of cores Two-dimensional convolutional layer (3,3) 1 32 Max pooling layer (2,2) 2 32 Two-dimensional convolutional layer (5,5) 1 64 Max pooling layer (2,1) 2 64 Fully connected layer - - 64 Fully connected layer - - 32 Fully connected layer - - 12 softmax layer

[0111] The two-dimensional convolutional neural network model employs two sets of convolutional layers with kernel sizes of {(3,3)} and (5,5)}, and compares them with combinations of {(3,3), (3,3)}, {(5,5), (5,5)}, {(7,7), (7,7)}, {(3,3), (7,7)}, and {(5,5), (7,7)}. The model consists of multiple convolutional layers, pooling layers, and fully connected layers, and can extract multi-level features from the input data to ultimately complete the fault diagnosis task. The proposed method employs two sets of convolutional layers with kernel sizes of {(3,3)} and (5,5)}, and compares them with combinations of {(3,3), (3,3)}, {(5,5), (5,5)}, {(7,7), (7,7)}, {(3,3), (7,7)}, and {(5,5), (7,7)}. The fault diagnosis accuracy rates are shown in Table 3 as 98.44%, 91.48%, 87.33%, 94.74%, 97.06%, and 99.19%, respectively.

[0112] Table 3. Convolution Kernel Size Comparison Table

[0113] kernel size Accuracy (%) (3,3) (3,3) 98.44 (5,5) (5,5) 91.48 (7,7) (7,7) 87.33 (3,3) (7,7) 94.74 (5,5) (7,7) 97.06 (3,3) (5,5) 99.19

[0114] While maintaining low computational complexity, the network structure possesses powerful feature extraction and representation capabilities, making it suitable for processing complex image data tasks. The training parameter settings for the model are detailed in Table 4.

[0115] Table 4 Training Parameters

[0116] optimization parameter Maximum number of iterations 200 Gradient threshold 1 Initial learning rate 0.005 Learning rate decline factor 0.01

[0117] 5) Using the intelligent fault diagnosis model for high-voltage power batteries in step 4), test the test sample set in step 3).

[0118] In step 5), the test sample set test visualizes the diagnostic results by comparing the confusion matrix and the classification results. On the experimental dataset, the fault identification accuracy reached 99.19%, demonstrating excellent performance.

[0119] 6) The performance of the intelligent fault diagnosis model for high-voltage power batteries in step 4) is evaluated and visually displayed through confusion matrix and classification result comparison chart. A comparison chart of the classification results for the four time-frequency conversion technologies is shown below. Figure 9 , Figure 10 , Figure 11 and Figure 12 As shown,

[0120] The target feature sample set is input into the training model, and the performance of the classification model is evaluated and visually displayed through confusion matrix and classification result comparison chart. The classification result comparison chart of four time-frequency transformation techniques is shown below. Figure 6As shown in the figure, the confusion matrix comparison diagram of the four network models is as follows: Figure 13 , Figure 14 , Figure 15 and Figure 16 As shown.

[0121] Multiple methods were employed to compare and analyze the denoising performance of the proposed method, including no denoising, SGolayFilt denoising, one-dimensional correlation denoising, Fourier denoising, and threshold denoising. Subsequently, the proposed Wigner distribution function was used to transform the acoustic detection signals denoised by different methods into two-dimensional feature maps, which were then trained and tested using the proposed two-dimensional convolutional neural network model. The results are shown in Table 5. The accuracy of the proposed method is significantly higher than all the compared methods.

[0122] Table 5 Performance results of different denoising methods

[0123] Noise reduction methods Accuracy (%) No noise reduction 84.35 SGolayFilt noise reduction method 97.06 One-dimensional correlation denoising method 97.63 Fourier noise reduction method 95.83 Threshold denoising method 94.29 The proposed noise reduction method 99.19

[0124] Three methods—short-time Fourier transform, Hilbert-Huang transform, and Mel-spectrum transform—were used to convert the acoustic detection signal of high-voltage power battery insulation into a two-dimensional time-frequency image, and the results were compared with the proposed method. Furthermore, the two-dimensional time-frequency images generated by different methods were input into the proposed two-dimensional convolutional neural network model for training and testing. The experimental results are shown in Table 6. The results show that the proposed method achieves an accuracy of 99.19%, significantly higher than the other three comparative methods, which confirms the robustness of the proposed method in high-voltage power battery fault detection.

[0125] Table 6 Performance Evaluation of Different Two-Dimensional Time-Frequency Image Transformation Methods

[0126] Time-frequency transformation method Accuracy (%) Short-time Fourier Transform Method 96.16 Hilbert-Huang Transform Method 84.35 Mel spectrum transform method 97.73 The proposed method 99.19

[0127] Furthermore, the proposed model was compared and analyzed using three different models: a convolutional neural network model-support vector machine model, a convolutional neural network model-long short-term memory recurrent neural network model, and a convolutional neural network model-bidirectional gated recurrent unit model. The proposed two-dimensional feature maps were used for training and testing, and the results are shown in Table 7. The results demonstrate that the proposed model outperforms the other comparative models with an accuracy of 99.72%.

[0128] Table 7 Performance of different network models

[0129] Network Model Accuracy (%) Convolutional Neural Network Model - Support Vector Machine Model 93.29 Convolutional Neural Network Model - Long Short-Term Memory Recurrent Neural Network Model 85.26 Convolutional Neural Network Model - Bidirectional Gated Recurrent Unit Model 96.43 The proposed model 99.19

[0130] The results show that the proposed method performs excellently in high-voltage power battery insulation detection. A high-performance high-voltage power battery insulation detection and recognition framework is constructed by combining wavelet denoising, Wigner function distribution mapping, and a two-dimensional convolutional neural network with optimized kernel size, achieving excellent results on balanced datasets.

[0131] This invention innovatively employs Wigner function distribution and wavelet denoising techniques. First, in the acoustic detection electrical signal feature extraction stage, the acoustic detection electrical signal acquired by the sampling circuit is converted into an acoustic detection digital signal. Second, in the data preprocessing stage, wavelet denoising is applied to the acoustic detection digital signal to eliminate noise interference, followed by batch normalization of the processed data. Next, the Wigner function distribution is used to convert the data into a two-dimensional time-frequency image. Finally, the model is trained and tested based on the time-frequency image, and the classification results are presented through a confusion matrix and a comparison diagram.

[0132] In summary, this invention proposes an intelligent fault diagnosis method for high-voltage power batteries based on wavelet denoising fusion of time-frequency images and two-dimensional convolutional neural networks. By integrating Wigner function distribution and wavelet denoising techniques, it exhibits significant advantages in data preprocessing, feature extraction, and model construction. Compared with short-time Fourier transform, Hilbert-Huang transform, and Mel-spectrum transform methods, this method has superior classification performance. Simultaneously, wavelet denoising effectively eliminates noise, enhancing the model's fault tolerance under complex background acoustic detection signals. At the model level, the proposed model outperforms integrated support vector machine models, long short-term memory network models, and bidirectional gated recurrent unit models in terms of temporal feature learning and generalization capabilities. Overall, this method constructs a complete diagnostic process from acoustic detection signal denoising and customized transform function design to neural network optimization, demonstrating good practicality and engineering deployability. It is particularly suitable for high-voltage power battery monitoring scenarios with limited sample data and severe noise interference, possessing significant practical application value. Furthermore, through systematic comparative experiments, this paper provides a technical path and data support for the fusion of acoustic detection signal processing and deep learning models, demonstrating broad potential for widespread application.

[0133] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A method for intelligent fault diagnosis of high-voltage power batteries based on wavelet denoising fusion of time-frequency images and two-dimensional convolutional neural networks, characterized in that, Includes the following steps: 1) Acquire two channels of acoustic detection electrical signals using an acoustic sensor and discretize them into digital acoustic detection signals; 2) The wavelet denoising algorithm is used to denoise the acoustic detection digital signal in step 1) to eliminate noise interference, and then the denoised data is batch normalized. 3) Convert the preprocessed acoustic detection digital signal into a time-frequency graph using the Wigner distribution function to construct a training sample set and a test sample set; 4) A two-dimensional convolutional neural network is used to train the training sample set in step 3) to construct a high-voltage power battery intelligent fault diagnosis model; 5) Using the intelligent fault diagnosis model for high-voltage power batteries in step 4), test the test sample set in step 3); 6) Evaluate and visually demonstrate the performance of the high-voltage power battery intelligent fault diagnosis model in step 4) using confusion matrix and classification result comparison charts.

2. The intelligent fault diagnosis method for high-voltage power batteries based on wavelet denoising fusion of time-frequency images and two-dimensional convolutional neural networks according to claim 1, characterized in that: In step 2), the wavelet denoising algorithm decomposes the acoustic detection signal and noise components through discrete wavelet transform at multiple scales. Given a noisy original acoustic detection signal S, the decomposed acoustic detection signal obtained through discrete wavelet transform can be expressed as: Here, and These are the scaling function and the wavelet function, respectively. Denotes the approximation coefficients of the Lth layer. Let J be the detail coefficients for the j-th layer; The decomposed acoustic detection signal is reconstructed by inverse discrete wavelet transform (IDT), which utilizes a threshold function. The processed low-frequency approximation coefficients c and high-frequency detail coefficients at various scales , can be represented as: 。 3. The intelligent fault diagnosis method for high-voltage power batteries based on wavelet denoising fusion of time-frequency images and two-dimensional convolutional neural networks according to claim 1, characterized in that: In step 3), the Wigner distribution function is a time-frequency analysis tool used to analyze non-stationary acoustic detection signals; for the acoustic detection signal x(t), its Wigner distribution is defined as: Where t represents time and f represents frequency. The complex conjugate of the acoustic detection signal is represented by τ, where τ is the time delay variable. The Wigner distribution can achieve theoretically optimal resolution in both the time and frequency domains, clearly characterizing instantaneous frequency changes. The boundary conditions satisfied by its time-frequency energy distribution can be expressed as: Let the acoustic detection signal be x(t) = x1(t) + x2(t), and its Wigner distribution function can be expressed as: Among them W x1 , x2 It is a cross term between components, which manifests as oscillating pseudo-energy.

4. The intelligent fault diagnosis method for high-voltage power batteries based on wavelet denoising fusion of time-frequency images and two-dimensional convolutional neural networks according to claim 1, characterized in that: In step 4), the two-dimensional convolutional neural network model is a deep learning architecture that includes convolutional layers, activation functions, pooling layers, fully connected layers, and an output layer. Convolution and pooling operations are performed through the convolutional and pooling layers. The two-dimensional convolutional neural network extracts input data features layer by layer, and finally realizes intelligent fault diagnosis of high-voltage power batteries through the fully connected layer.

5. The intelligent fault diagnosis method for high-voltage power batteries based on wavelet denoising fusion of time-frequency images and two-dimensional convolutional neural networks according to claim 1, characterized in that: In step 5), the test sample set test visualizes the diagnostic results by comparing the confusion matrix and the classification results. On the experimental dataset, the fault identification accuracy reached 99.19%, demonstrating excellent performance.

6. The intelligent fault diagnosis method for high-voltage power batteries based on wavelet denoising fusion of time-frequency images and two-dimensional convolutional neural networks according to claim 1, characterized in that: Step 2) of batch normalization includes the following steps: (1) Calculate the mean and variance: For a batch of input data X={x (1) ,x (2) ,...,x (m) The batch normalized mean and variance can be expressed as follows: in This represents the average value of the input in the current batch; Represents the variance of the current batch of inputs; m is the number of samples in the batch; x (i) This represents the input value of the i-th sample in a certain feature dimension of this layer; i is the index of the sample in the batch, from 1 to m; B refers to the label of the current batch, used to distinguish different batches in training; (2) Normalization: in This represents the standardized value of the i-th sample; It is a very small constant introduced to avoid division by zero; (3) Scaling and translation: γ and β are trainable parameters used to recover the expressive power of the model.

7. The wavelet denoising algorithm according to claim 2, characterized in that: During the reconstruction process, three main types of deviations exist: threshold processing bias, boundary effect error, and wavelet basis mismatch error. The threshold processing bias is introduced by adjusting the high-frequency coefficients through a threshold function and can be quantified as the original high-frequency coefficients at each scale. With processed coefficients The L2 norm of the difference between them can be expressed as: Boundary effect error is caused by the boundary processing of the acoustic detection signal in wavelet transform, and its magnitude is negatively correlated with the number of transform layers L, which can be expressed as: in These are positive coefficients related to the characteristics of wavelet basis and acoustic detection signals; Wavelet basis mismatch error originates from the selected wavelet basis. The mismatch between the acoustic detection signal S and its projection onto the wavelet basis is the difference between the actual acoustic detection signal S and its characteristics. The minimum deviation between them can be expressed as: Where B represents the set of wavelet basis functions, reflecting the optimal approximation of the acoustic detection signal by the wavelet basis.

8. The two-dimensional convolutional neural network model according to claim 4, characterized in that: The two-dimensional convolutional neural network model employs two sets of convolutional layers with kernel sizes of {(3,3)} and (5,5)}, and compares them with combinations of {(3,3), (3,3)}, {(5,5), (5,5)}, {(7,7), (7,7)}, {(3,3), (7,7)}, and {(5,5), (7,7)}. The model consists of multiple convolutional layers, pooling layers, and fully connected layers, and can extract multi-level features from the input data to ultimately complete the fault diagnosis task.

Images