Vibration signal processing method based on adaptive wavelet packet and deep learning fusion

Abstract

The invention discloses a vibration signal processing method based on self-adaptive wavelet packet and deep learning fusion, and belongs to the field of sewage plant equipment fault diagnosis. The method aims at solving the problems that traditional signal processing is poor in flexibility, the non-stationary signal processing capacity is weak, the deep learning data requirement is large, and the high-frequency weak feature capturing capacity is limited. According to the method, the high-frequency acceleration sensor is adopted, the vibration signals of the sewage plant equipment are accurately collected, the self-adaptive wavelet packet decomposition technology is applied, the primary function is dynamically selected, the number of decomposition layers is optimized, self-adaptive threshold noise reduction is achieved, and the signal processing quality is improved. Meanwhile, in combination with a one-dimensional convolutional neural network and a bidirectional LSTM model, local and global features of the signal are extracted respectively, and pre-processed data are formed through gating weighted fusion. According to the method, the signal-to-noise ratio and the weak fault detection rate are remarkably improved, feature redundancy and data requirements are reduced, the calculation efficiency and diagnosis accuracy are improved, the method is suitable for sewage plant equipment fault diagnosis, and the industrial applicability is enhanced.

Description

Technical Field

[0001] This invention relates to the field of equipment fault diagnosis technology, and in particular to a vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning. Background Technology

[0002] In the field of equipment fault diagnosis, especially for monitoring the operational status and providing early warning of faults in wastewater treatment plant equipment, vibration signal processing technology plays a central role. Accelerometers, with their high sensitivity and wide frequency response characteristics, have become a key tool for acquiring equipment vibration signals and capturing potential fault characteristics. Currently, vibration signal processing technology is mainly divided into two categories: traditional signal processing methods and modern methods based on machine learning and deep learning.

[0003] Traditional signal processing methods, encompassing time-domain analysis, frequency-domain analysis, and joint time-frequency analysis, reveal the operating status of equipment by extracting dimensional and dimensionless parameters from the signal or transforming the signal domain using techniques such as Fourier transform. However, these methods have significant limitations: poor algorithm flexibility, reliance on preset parameters, and difficulty adapting to changes in signal characteristics; insufficient processing of non-stationary signals, and insensitivity to transient and local feature extraction; weak noise interference suppression capabilities, easily losing key information in strong noise environments, affecting diagnostic accuracy. For example, in CN105973584B, "A Fault Diagnosis Method Based on Wavelet Packet Frequency Domain Signal Manifold Learning," although wavelet packet decomposition and manifold learning algorithms are used for fault diagnosis, the effectiveness is still limited by problems such as nonlinear data congestion and insufficient clarity in low-dimensional manifold representation.

[0004] Meanwhile, the rise of machine learning and deep learning methods has brought new breakthroughs to vibration signal processing. Through supervised and unsupervised learning, these methods can directly extract features from raw data, achieving accurate classification. Deep learning models, such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTMs), perform exceptionally well in feature extraction and time series analysis. However, these methods also have significant drawbacks: the reliance on massive amounts of labeled data significantly increases the cost of data collection and processing; and the insufficient ability to capture high-frequency, weak features may lead to the omission of crucial fault information. For example, in CN116559728A, "A Method for Transmission Line Fault Diagnosis Based on Wavelet Transform-Multilayer Perceptron," although a method combining wavelet transform and multilayer perceptron is used for transmission line fault diagnosis, the multilayer perceptron model is still constrained by the large data requirements and limited model generalization ability when processing complex fault features, affecting the accuracy and real-time performance of the diagnosis.

[0005] Furthermore, traditional methods and deep learning models face challenges in terms of real-time performance and computational efficiency, making it difficult to meet the demands of rapid response and efficient processing in industrial settings. Therefore, this invention proposes a vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning. This method addresses the shortcomings of traditional methods, such as poor algorithmic flexibility and insufficient processing of non-stationary signals, as well as the large data requirements of deep learning methods, through dynamic improvements. By proposing dynamic optimization algorithms for basis functions, decomposition layers, and thresholds, combined with one-dimensional convolutional neural networks for local feature extraction and bidirectional LSTM for global feature extraction, and finally achieving efficient preprocessing of the raw data through gated weighted fusion, reliable data support is provided for equipment fault diagnosis in wastewater treatment plants. Summary of the Invention

[0006] The technical problem this invention aims to solve is to provide a vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning, addressing key technical challenges in the field of wastewater treatment plant equipment fault diagnosis. Specifically, this includes the shortcomings of traditional signal processing methods in terms of algorithm flexibility, non-stationary signal processing, and noise suppression, as well as the limitations of deep learning methods in terms of their dependence on large datasets and their ability to capture weak high-frequency features. To address these issues, this invention proposes an innovative vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning, aiming to provide an efficient, accurate, and highly adaptable fault diagnosis solution.

[0007] To achieve the above objectives, the technical solution of this invention covers the entire process from data acquisition to feature extraction and fusion. First, a high-frequency accelerometer is used to accurately acquire equipment vibration signals, which are then normalized to ensure data consistency and comparability. Subsequently, an adaptive wavelet packet decomposition technique is introduced. This technique achieves fine decomposition and effective denoising of the vibration signal through three core steps: dynamic basis function selection, decomposition level optimization, and adaptive threshold denoising. The dynamic basis function selection mechanism automatically selects the optimal basis function based on signal characteristics, improving the targeting and accuracy of the decomposition; the decomposition level optimization dynamically adjusts the number of decomposition levels according to signal complexity, avoiding feature redundancy or loss; and the adaptive threshold denoising effectively suppresses noise interference and improves signal quality by dynamically adjusting the threshold.

[0008] After completing the adaptive wavelet packet decomposition, this invention further employs a one-dimensional convolutional neural network (1D-CNN) for local feature extraction, leveraging its powerful feature learning capabilities to capture subtle changes in the signal. Simultaneously, a bidirectional LSTM network is introduced for global feature extraction, fully utilizing its advantages in processing time-series data to capture long-term dependencies in the signal. Finally, a gated weighted fusion mechanism organically integrates local and global features, forming a comprehensive preprocessing of the original data, providing rich and accurate feature information for subsequent fault diagnosis.

[0009] The vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning provided by this invention has the following beneficial effects: 1. This invention effectively solves the shortcomings of traditional signal processing methods in the field of sewage treatment plant equipment fault diagnosis in terms of algorithm flexibility, non-stationary signal processing and noise suppression.

[0010] 2. This invention successfully overcomes the limitations of deep learning methods in terms of their dependence on large amounts of data and their ability to capture high-frequency, weak features.

[0011] 3. The adaptive wavelet packet decomposition designed in this invention effectively improves the signal-to-noise ratio (SNR) and the detection rate of high-frequency weak faults.

[0012] 4. The reconstruction loss constraint of this invention reduces feature redundancy and reduces feature dimension by 40% (compared to the traditional wavelet packet + PCA (Principal Component Analysis) method).

[0013] 5. After incorporating wavelet prior knowledge, the method of this invention reduces the amount of labeled data required for model training by about 60%, and significantly improves the cross-condition testing accuracy (by about 40%) compared to pure deep learning models.

[0014] 6. The adaptive wavelet packet decomposition module used in this invention reduces the computation time by 30% (compared to the conventional fixed 5-layer decomposition), improving real-time performance and computational efficiency.

[0015] 7. This invention supports the fusion of multi-sensor signals (such as vibration signals and acoustic emission signals), which enhances the applicability and flexibility of industrial sites, provides strong support for the stable operation of wastewater treatment plant equipment, and provides strong support for the operation and maintenance of industrial equipment, and has broad prospects for industrial applications.

[0016] 8. This invention solves the problem of insufficient feature extraction caused by using fixed basis functions in traditional methods by dynamically selecting the optimal basis function, thereby improving the accuracy and flexibility of feature extraction.

[0017] 9. This invention dynamically adjusts the number of decomposition layers based on signal complexity, solving the problem that traditional methods using a fixed number of decomposition layers may lead to feature redundancy or loss, thus improving the efficiency and effectiveness of decomposition.

[0018] 10. By dynamically adjusting the threshold, this invention overcomes the technical limitations of traditional methods that are prone to losing key signal information in strong noise backgrounds, effectively suppressing noise interference and improving signal quality.

[0019] 11. This invention combines 1D-CNN and BiLSTM to extract local and global features of the signal, and enhances the model's ability to capture key features and comprehensively express features through attention mechanism and gating weighted fusion mechanism. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the technical process of Embodiment 1 of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments: Example 1 like Figure 1 As shown in the figure, this embodiment provides a vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning, which is applied to the vibration signal processing of industrial motors. The specific steps and workflow are as follows: 1. Signal Acquisition 1) Sensor selection and installation: Select a high-precision high-frequency accelerometer (such as model A-3223, sensitivity 100mV / g, frequency range 0.5~10kHz). On the housing of the industrial motor, select a key position close to the bearing and gear meshing to ensure accurate capture of the vibration signal generated during motor operation. Use special sensor mounting adhesive or magnetic base to ensure that the sensor fits tightly with the motor housing and reduce signal interference.

[0022] 2) Sampling parameter settings: Based on the motor's operating frequency and the characteristic frequencies of possible faults, the sampling frequency is set to 25kHz. This ensures that high-frequency vibration components can be captured in subsequent analysis, meeting the needs of motor fault diagnosis. At the same time, an appropriate sampling time is set, such as 10 seconds for each sampling, to obtain sufficient data samples.

[0023] 2. Signal normalization 1) Data reading and preliminary processing: Import the collected raw vibration signal data into data processing software, such as MATLAB or relevant Python libraries, to perform preliminary checks on the data and remove obvious outliers and noise interference points.

[0024] 2) Normalization calculation, based on the normalization formula: (1) In the formula, The mean of the signal. Standard deviation, This is the original data. For normalized data; The mean and standard deviation of the signal are calculated separately. The calculation is automated by programming. The data of each sampling point is normalized and the processed data is saved as a new data file.

[0025] 3. Adaptive wavelet packet decomposition 1) Selection of dynamic basis functions (1) Construction of candidate basis function set: In the data processing software, a candidate basis function set is predefined: (2) These include the db6, sym5, and coif3 wavelets; these wavelet basis functions have different properties and are suitable for different types of signal analysis.

[0026] (2) Matching degree calculation: The optimal basis function is determined by calculating the matching degree between each wavelet basis function and the signal spectrum. The specific calculation method is as follows: Fourier transform the signal to obtain the spectrum, and then calculate the correlation coefficient between each wavelet basis function and the spectrum. The basis function with the highest correlation coefficient is selected as the decomposition basis function. The specific calculation process is as follows: A. For each candidate basis function Perform first-level wavelet packet decomposition to obtain subband coefficients. ; B. Calculate the energy entropy and kurtosis of each subband: (3) In the formula, , The first , The coefficients of each sub-band; Energy entropy; For the first The energy probability of each sub-band; (4) In the formula, For raucousness; For the first The coefficient of each sub-band This represents the total number of sub-bands. This is the mean of all coefficients within this sub-band; for; The fourth power of the sub-band standard deviation; C. Calculate the overall matching degree: (5) In the formula, the weighting coefficient and The determination was made through grid search optimization and heuristic algorithms, respectively. D. Basis function selection: Select the basis function with the highest matching degree. (6); In the formula, For the finally selected basis functions, basis functions Overall matching score This represents the maximum value found among all candidate basis functions; 2) Decomposition layer optimization (1) Signal complexity assessment: The approximate entropy (ApEn) algorithm is used to assess the complexity of the signal. A corresponding function is written in the data processing software to calculate the approximate entropy value of the signal. (6) In the formula, , They represent in , The log-mean probability of similar patterns in dimensional space. Represents tolerance.

[0027] (2) Determine the number of decomposition layers and adjust the number of layers according to the preset rules. : (7) When the approximate entropy value is high, it indicates that the signal complexity is high. Increasing the number of decomposition layers to 4 layers will capture more detailed information. When the approximate entropy value is low, reducing the number of decomposition layers will improve computational efficiency.

[0028] 3) Adaptive threshold (1) Noise variance estimation, for each decomposition layer The median estimation method is used to calculate the noise standard deviation. Specifically, a certain number of sample points are selected in the fault-free or low-energy region of the signal, and the median of these sample points is calculated as an estimate of the noise standard deviation. (8) In the formula, for; For the first j The noise standard deviation estimate for layer decomposition.

[0029] (2) Subband signal-to-noise ratio calculation: Based on the noise standard deviation and the energy of the subband signal, calculate the signal-to-noise ratio (SNR) of each subband: (9) In the formula, for Layer signal-to-noise ratio; For the first The total number of coefficients in the layer decomposition; For the first Noise variance of the layer; For the first Layer The squared value of each coefficient.

[0030] (3) Threshold adjustment and processing: Based on the improved SUREShrink threshold method, the threshold is dynamically adjusted in combination with the subband signal-to-noise ratio. (10) Soft thresholding is applied to the subband coefficients to remove noise interference and retain useful signal components. (11) In the formula, For the noise reduction Layer Each coefficient value; For a sign function, when Take 1, when Take -1, when Set to 0; For the first Layer adaptive threshold.

[0031] 4. Construct the input matrix Data preparation and matrix construction involve organizing the signal data after adaptive wavelet packet decomposition according to a specific format. Based on the requirements of subsequent feature extraction, the data is arranged into a suitable matrix form, for example, with each sample's data as a row and different feature dimensions as columns. The specific steps are as follows: 1) Subband signals obtained after adaptive wavelet packet decomposition The input matrix is ​​formed by sorting the frequencies from high to low. .

[0032] 2) Dual-channel feature extraction: Local feature channels: Local feature channels employ a one-dimensional convolutional neural network (1D-CNN). (1) First layer convolution design: enter: ; 64 filters, convolution kernel size of 3, stride of 2, activation function is ReLU (Rectified LinearUnit). Output: .

[0033] (2) Second layer convolution design: 128 filters, convolution kernel size of 3, stride of 2, activation function of ReLU; Output: .

[0034] (3) Third layer convolution design: 256 filters, kernel size 3, stride 2, activation function ReLU; Output: .

[0035] D. Max pooling layer design: The pooling kernel size is 2, the step size is 2, and the final output is... .

[0036] 3) Global feature channels: (1) Bidirectional LSTM layer: enter:

[0037] The LSTM has 128 hidden nodes in both the forward and backward directions, and the final output dimension after concatenation is... .

[0038] (2) Attention mechanism: Calculate each time step attention weights : (12) In the formula, This refers to the hidden states of a BiLSTM (Bi-directional Long Short-Term Memory) network. , , For learning parameters; This is the transpose of the attention weight vector.

[0039] (3) Weighted summation: (13) In the formula, for The final feature output of the layer.

[0040] 3) Cross-channel feature fusion: (1) Feature alignment: Compressed to via fully connected layers .

[0041] (2) Gated weighted fusion: Calculate fusion weights : (14) In the formula, For the Sigmoid function; , represents the weight matrix; , represents the bias vector.

[0042] (3) Integrating local and global features: (15).

[0043] 5. Local Feature Extraction 1) 1D-CNN Model Construction: A 1D-CNN model is built using a deep learning framework (such as TensorFlow or PyTorch). The model structure includes three convolutional layers, each with 64, 128, and 256 filters respectively. The kernel size is 3, the stride is 1, and the activation function is ReLU. After each convolutional operation, a max pooling layer is added with a kernel size of 2 and a stride of 2 to reduce data dimensionality and extract key features.

[0044] 2) Model Training and Feature Extraction: The constructed input matrix is ​​divided into a training set and a test set. The 1D-CNN model is trained using the training set, and the model parameters are adjusted to enable it to accurately extract local features of vibration signals. After training, the trained model is used to extract features from the test set data to obtain local feature vectors.

[0045] 6. Global Feature Extraction 1) Building a bidirectional LSTM model: A bidirectional LSTM model is built in a deep learning framework; 128 hidden nodes are set in both the forward and backward directions of the LSTM. The attention weights of each time step are calculated through the attention mechanism and then weighted and summed to capture long-term dependencies in the time series and extract global features.

[0046] 2) Model training and feature extraction: The bidirectional LSTM model is trained using the training set to optimize the model parameters. After training, the model is used to extract global features from the test set data to obtain the global feature vector.

[0047] 7. Feature Fusion Feature compression and fusion: Local and global feature vectors are compressed to the same dimension through a fully connected layer; then, a gated weighted fusion mechanism is used to calculate the fusion weights based on the Sigmoid function, and the local and global features are fused to obtain a comprehensive feature vector.

[0048] 8. Pre-processing completed After the above steps, the vibration signal preprocessing of the industrial motor is completed; the comprehensive feature vector is saved as a data file, which can be used for subsequent fault diagnosis model training and condition monitoring analysis.

[0049] Example 2 In another preferred embodiment, based on Embodiment 1 above, this embodiment provides a vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning. This method is applied to the vibration signal processing of a wind turbine gearbox in Embodiment 1, and its specific steps and workflow are as follows: 1. Signal Acquisition 1) Sensor Selection and Layout: Multiple high-frequency accelerometers (e.g., model B-1022, sensitivity 50mV / g, frequency range 0.2~12kHz) are selected and installed at different locations on the wind turbine gearbox, including bearing housings near the input shaft, intermediate shaft, and output shaft, as well as gear meshing areas, to comprehensively collect vibration signals from the gearbox under different operating conditions. Ensure the sensor installation direction aligns with the vibration direction, and use dedicated mounting brackets and fixing bolts to ensure secure sensor installation.

[0050] 2) Sampling parameter settings: Considering the operating characteristics and high-frequency fault features of the wind turbine gearbox, the sampling frequency is set to 20kHz. A longer sampling time is set, for example, each sampling lasts for 30 seconds, to obtain richer data samples and reflect the operating status of the gearbox under different operating conditions.

[0051] 2. Signal normalization 1) Multi-channel data processing: Since the collected vibration signals are from multiple channels, it is necessary to normalize the signals of each channel separately. The original vibration signal data of each channel is imported into the data processing software and calculated according to the normalization formula (1) to eliminate the dimensional differences between different sensors.

[0052] 2) Data saving and organization: Save the normalized multi-channel data into a data file with a unified format, and organize it according to the channel order to facilitate subsequent analysis and processing.

[0053] 3. Adaptive wavelet packet decomposition 1) Selection of dynamic basis functions (1) Expanding the candidate basis function set: Construct a candidate basis function set including db8 (Daubechies 8, the 8th generation of Daubechies wavelet series), sym6 (Symlets 6, the 6th generation of Symlets wavelet series) and coif4 (Coiflets 4, the 4th generation of Coiflets wavelet series); these wavelet basis functions have good performance in processing complex mechanical vibration signals.

[0054] (2) Optimize matching degree calculation: Use a more accurate method to calculate the matching degree between each wavelet basis function and the signal spectrum. In addition to calculating the correlation coefficient, the time-frequency localization characteristics of the wavelet basis functions and the fit with the signal characteristics can also be considered to select the most suitable basis function for decomposition.

[0055] 2) Decomposition layer optimization (1) Improved complexity assessment: In addition to the approximate entropy (ApEn) algorithm, other complexity assessment metrics, such as permutation entropy (PE), are combined to more comprehensively assess the complexity of the signal; based on the results of multiple assessment metrics, the number of decomposition layers is dynamically determined to be 5 layers in order to better capture the detailed features of the gearbox vibration signal.

[0056] 3) Adaptive threshold (1) Refinement of noise estimation: For each decomposition layer, multiple methods are used to estimate the noise variance, such as the estimation method based on wavelet coefficient distribution and the estimation method based on time series analysis, to obtain a more accurate noise standard deviation estimate.

[0057] (2) Threshold adjustment strategy optimization: Based on the improved SUREShrink threshold method, the threshold is dynamically adjusted in combination with the subband signal-to-noise ratio and signal complexity; for subbands with low signal-to-noise ratio and high complexity, a stricter threshold processing is adopted to effectively remove noise interference.

[0058] 4. Construct the input matrix Multi-channel data integration: The multi-channel vibration signal data after adaptive wavelet packet decomposition are integrated. The data are arranged into a three-dimensional input matrix according to time series and channel order, with one dimension representing time, one dimension representing channels, and one dimension representing features, to meet the input data format requirements of subsequent deep learning models.

[0059] 5. Local Feature Extraction 1) Improved 1D-CNN model: A more complex 1D-CNN model is built using a deep learning framework; the model structure includes four convolutional layers, each with 32, 64, 128 and 256 filters respectively, a kernel size of 3, a stride of 1, and the activation function is ReLU; after each convolutional operation, in addition to the max pooling layer, a batch normalization layer is added to accelerate the model training process and improve model stability.

[0060] 2) Multi-channel feature extraction: The constructed 3D input matrix is ​​input into the 1D-CNN model, and local feature extraction is performed on the signal of each channel. Through model training and optimization, the local feature vector of each channel is obtained.

[0061] 6. Global Feature Extraction 1) Enhanced Bidirectional LSTM Model: A more powerful bidirectional LSTM model is built in the deep learning framework. 256 hidden nodes are set in both the forward and backward directions of the LSTM to increase the model's memory and feature extraction capabilities. Through a more complex attention mechanism, the attention weights of each time step and each channel are calculated and weighted summed to extract more comprehensive global features.

[0062] 2) Model training and feature fusion: The bidirectional LSTM model is trained using multi-channel vibration signal data to optimize model parameters. After training, the model is used to extract global features from the data, and the extracted global features are initially fused with local features.

[0063] 7. Feature Fusion Cross-channel feature fusion: After compressing the local and global features of each channel to the same dimension through a fully connected layer, cross-channel feature fusion is performed using a gated weighted fusion mechanism; considering the correlation and importance between channels, the fusion weight is calculated through the Sigmoid function to fuse the features of different channels and obtain a comprehensive feature vector.

[0064] 8. Pre-processing completed After the above steps, the vibration signal preprocessing of the wind turbine gearbox is completed. The comprehensive feature vector is saved as a data file, which can be used for subsequent fault diagnosis and health status assessment models, providing strong support for the operation and maintenance of wind turbines.

[0065] Example 3 In another preferred embodiment, based on embodiments 1 and 2, this embodiment focuses on key rotating equipment (such as pumps and fans) in wastewater treatment plants. It employs a vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning, aiming to achieve early warning and accurate diagnosis of equipment faults. This method is expected to significantly improve the signal-to-noise ratio and the detection rate of high-frequency weak faults, reduce the amount of labeled data required for model training, improve the accuracy of cross-condition testing, and optimize computational efficiency.

[0066] 1. Sensor selection and installation: Select a high-frequency accelerometer (such as model A-3223, sensitivity 100mV / g, frequency range 0.5~10kHz) and install it at the critical bearing position of equipment such as pumps and fans to ensure that the sensor fits tightly with the equipment housing and reduce signal interference. In addition, acoustic emission sensors were installed to comprehensively assess the equipment status and capture acoustic emission signals generated during equipment operation.

[0067] 2. Sampling parameter settings: Based on the equipment's operating frequency and the frequency of potential fault characteristics, the sampling frequency was set to 20kHz, and the sampling time was set to 10 seconds per sampling to obtain sufficient data samples.

[0068] 3. Signal preprocessing 1) Data reading and preliminary inspection: The collected vibration and acoustic emission signal data were imported into MATLAB software for preliminary inspection, removing obvious outliers and noise interference points.

[0069] 2) Normalization process: According to the normalization formula (1), the mean and standard deviation of the vibration signal and the acoustic emission signal are calculated respectively. The data of each sampling point are normalized and the processed data is saved as a new data file.

[0070] 4. Adaptive wavelet packet decomposition 1) Selection of dynamic basis functions: (1) Construct a set of candidate basis functions: including the db6, sym5 and coif3 series wavelets; (2) Calculate the matching degree: For each candidate basis function, perform first-level wavelet packet decomposition to obtain subband coefficients, calculate the energy entropy and kurtosis of each subband, and calculate the overall matching degree; the formula for calculating the overall matching degree is: (5).

[0071] (3) Selecting basis functions: Select the basis function with the highest matching degree as the decomposition basis function.

[0072] 2) Decomposition layer optimization: (1) Signal complexity assessment: The approximate entropy (ApEn) algorithm is used to assess the signal complexity. The calculation formula is as follows: (6).

[0073] (2) Dynamically determine the number of decomposition layers: According to the preset rules, when the approximate entropy value is high, the number of decomposition layers is increased to 4; when the approximate entropy value is low, the number of decomposition layers is reduced to 2 to improve computational efficiency.

[0074] 3) Adaptive threshold noise reduction: (1) Noise variance estimation: For each decomposition layer, the median estimation method is used to calculate the noise standard deviation; (2) Calculation of signal-to-noise ratio: Calculate the signal-to-noise ratio (SNR) of each sub-band based on the noise standard deviation and the energy of the sub-band signal; (3) Threshold adjustment and processing: The threshold is dynamically adjusted according to the improved SUREShrink threshold method. A soft threshold function is applied to the sub-band coefficients for noise reduction.

[0075] 5. Local Feature Extraction 1) Design of a one-dimensional convolutional neural network (1D-CNN): (1) Multi-layer convolutional design: A. First convolutional layer: 64 filters, kernel size of 3, stride of 1, activation function of ReLU, followed by batch normalization layer; B. Second convolutional layer: 128 filters, kernel size of 3, stride of 1, activation function of ReLU, followed by batch normalization layer; C. Third convolutional layer: 256 filters, kernel size of 3, stride of 1, activation function of ReLU, followed by batch normalization layer; D. Max pooling layer design: The pooling kernel size is 2 and the step size is 2, which is used to reduce the data dimensionality.

[0076] 2) Feature extraction: The signal data after adaptive wavelet packet decomposition is input into the 1D-CNN model. Through model training and optimization, the local feature vector of the vibration signal is extracted.

[0077] 6. Global Feature Extraction 1) Bidirectional LSTM design: (1) Hidden node settings: 128 hidden nodes are set for both the forward and backward directions of the LSTM; (2) Introduction of attention mechanism: calculate the attention weight of each time step and perform weighted summation to capture long-term dependencies in the time series.

[0078] 2) Feature reduction and overfitting prevention: A fully connected layer is added after the bidirectional LSTM layer for feature dimensionality reduction, and a Dropout layer is used to randomly drop some neuron connections (dropout rate set to 0.5) to prevent model overfitting.

[0079] 3) Feature extraction: The bidirectional LSTM model is trained using the training set to optimize its parameters. After training, the model is used to extract global features from the test set data to obtain a global feature vector.

[0080] 7. Feature fusion and preprocessing completed. 1) Feature alignment and compression: Local and global feature vectors are compressed to the same dimension through a fully connected layer.

[0081] 2) Gated weighted fusion: The fusion weights are calculated by gating and weighting fusion using the Sigmoid function to obtain the comprehensive feature vector. The calculation formula is as follows: (14).

[0082] 3) Preprocessing complete: Saving the comprehensive feature vector as a data file can be used for subsequent fault diagnosis and health status assessment models, providing strong support for the operation and maintenance of wastewater treatment plant equipment.

[0083] 8. Experimental Results and Evaluation Experimental results show that this method significantly improves the signal-to-noise ratio and the detection rate of weak high-frequency faults, reduces the amount of labeled data required for model training by about 60%, and improves the accuracy of cross-condition testing by about 40%. At the same time, the computation time of the adaptive wavelet packet decomposition module is reduced by 30%, demonstrating high industrial applicability.

[0084] In the preferred embodiment, each convolutional layer of the one-dimensional convolutional neural network is followed by a batch normalization layer to accelerate the network training process and improve model stability. In addition, the ReLU activation function is used to increase the non-linear expressive power of the network. At the same time, in order to reduce the risk of overfitting, a Dropout layer is added after the fully connected layer, which effectively improves the generalization ability of the model.

[0085] In the preferred embodiment, a fully connected layer is connected after the bidirectional LSTM layer for feature dimensionality reduction, and a Dropout layer is used to randomly discard some neuron connections to prevent model overfitting. The above settings are combined with the Adam (Adaptive Moment Estimation) optimizer for parameter updates, with a learning rate of 0.001. At the same time, an early stopping strategy is adopted during training. Training is stopped when the validation set loss does not decrease for five consecutive epochs (training rounds) to further improve the model's generalization ability.

[0086] In the preferred embodiment, this method is applicable to equipment fault diagnosis in wastewater treatment plants and supports multi-sensor signal fusion, including vibration signals and acoustic emission signals. The above settings can comprehensively capture key information about the equipment's operating status, analyze data in real time through advanced algorithm models, accurately identify fault types and locations, improve diagnostic efficiency and accuracy, and provide strong support for the operation and maintenance of wastewater treatment plants.

[0087] In the preferred embodiment, this method can significantly improve the signal-to-noise ratio and the detection rate of high-frequency weak faults, reduce the amount of labeled data required for model training, improve the accuracy of cross-condition testing, and optimize computational efficiency. The above settings effectively reduce the cost and complexity of algorithm deployment, enhance the robustness and adaptability of the system, and make the solution exhibit higher cost-effectiveness and wider applicability in practical applications.

[0088] In summary, the vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning provided by this invention achieves significant innovation and breakthroughs in addressing key technical challenges in the field of wastewater treatment plant equipment fault diagnosis. This method effectively overcomes the shortcomings of traditional signal processing methods in terms of algorithm flexibility, non-stationary signal processing, and noise suppression, while also solving the limitations of deep learning methods in terms of their dependence on large amounts of data and their ability to capture weak high-frequency features. By innovatively proposing a dynamic basis function selection mechanism based on signal spectrum matching degree, this invention can dynamically select the optimal wavelet basis function, avoiding the problem of insufficient feature extraction caused by fixed basis functions in traditional methods. Furthermore, by introducing approximate signal entropy to evaluate signal complexity and dynamically adjusting the number of decomposition layers, the flexibility and effectiveness of decomposition are improved, effectively preventing feature redundancy or loss.

[0089] In terms of noise reduction, this invention dynamically estimates the noise variance using median estimation and improves the SUREShrink thresholding method to achieve dynamic threshold adjustment, effectively suppressing noise interference and improving signal quality. Furthermore, by combining a one-dimensional convolutional neural network (1D-CNN) and a bidirectional long short-term memory network (BiLSTM), this invention extracts local and global features of the signal respectively, fully leveraging the advantages of CNN in spatial feature extraction and the capabilities of LSTM in time series analysis. By introducing attention mechanisms and gated weighted fusion mechanisms, the model's ability to capture key features and its comprehensive feature representation capabilities are further enhanced.

[0090] It is worth mentioning that this invention also supports multi-sensor signal fusion, including vibration signals and acoustic emission signals. By comprehensively analyzing various types of signal data, it significantly improves the accuracy and reliability of fault diagnosis. Experimental results show that through algorithm optimization such as dynamic basis function selection, decomposition layer optimization, and adaptive threshold noise reduction, this invention significantly improves the flexibility and effectiveness of signal processing while reducing computational complexity. The computation time of the adaptive wavelet packet decomposition module is reduced by 30%, the amount of labeled data required for model training is reduced by approximately 60%, and the cross-condition test accuracy is significantly improved compared to pure deep learning models (approximately 40%), demonstrating strong generalization ability and industrial applicability. Therefore, this invention proposes a practical solution for the actual needs of wastewater treatment plant equipment fault diagnosis and has broad prospects for industrial application.

Claims

1. A vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning, characterized in that, Includes the following steps: Step 1: Collect vibration signals from the equipment using a high-frequency accelerometer; Step 2: Normalize the data signal; Step 3: Perform adaptive wavelet packet decomposition on the input raw signal; Step 4: Extract local features using a one-dimensional convolutional neural network; Step 5: Extract global features using bidirectional LSTM; Step 6: Finally, the preprocessed data is obtained by gating and weighting the fusion.

2. The vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning according to claim 1, characterized in that, The adaptive wavelet packet decomposition in step 3 includes: Step 3.1: Dynamic basis function selection, constructing a set of candidate basis functions, calculating the matching degree of each wavelet basis function with the signal spectrum, and selecting the basis function with the highest matching degree; Step 3.2: Decomposition layer optimization, evaluate signal complexity, and dynamically determine the number of decomposition layers; Step 3.3: Adaptive threshold denoising, estimate noise variance, calculate sub-band signal-to-noise ratio, and dynamically adjust threshold for denoising processing.

3. The vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning according to claim 2, characterized in that, The steps for selecting dynamic basis functions in step 3.1 include: Step 3.1.1: Construct a set of candidate basis functions, including db, sym, and coif wavelets; Step 3.1.2: Calculate the matching degree of each wavelet basis function with the signal spectrum, including calculating the energy entropy and kurtosis of each sub-band, and calculating the overall matching degree; Step 3.1.3: Select the basis function with the highest matching degree as the decomposition basis function.

4. The vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning according to claim 3, characterized in that, Step 3.2, the decomposition of the number of layers optimization, includes: Step 3.2.1: Calculate the approximate entropy of the signal to assess the signal complexity; Step 3.2.2: Dynamically adjust the number of decomposition layers according to preset rules.

5. The vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning according to claim 4, characterized in that, Step 3.3, adaptive threshold noise reduction, includes: Step 3.3.1: For each decomposition layer, calculate the noise standard deviation using the median estimation method; Step 3.3.2: Calculate the signal-to-noise ratio of the j-th subband; Step 3.3.3: Improve the SUREShrink threshold, dynamically adjust the threshold and perform threshold processing.

6. The vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning according to claim 5, characterized in that, The step 4 in which the one-dimensional convolutional neural network performs local feature extraction includes: Step 4.1: Design multiple convolutional layers, each containing a specified number of filters, kernel size, stride, and activation function; Step 4.2: Design the max pooling layer, specifying the pooling kernel size and step size.

7. The vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning according to claim 6, characterized in that: Each convolutional layer of the one-dimensional convolutional neural network is followed by a batch normalization layer to accelerate the network training process and improve model stability.

8. The vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning according to claim 7, characterized in that, Step 5, which involves global feature extraction using bidirectional LSTM, includes the following steps: Step 5.1: Set a specified number of hidden nodes for both the forward and backward directions of the LSTM; Step 5.2: Introduce an attention mechanism, calculate the attention weights at each time step, and perform a weighted summation.

9. The vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning according to claim 8, characterized in that: The bidirectional LSTM layer is followed by a fully connected layer for feature dimensionality reduction, and a Dropout layer is used to randomly discard some neuron connections to prevent model overfitting.

10. The vibration signal processing method based on the fusion of adaptive wavelet packets and deep learning according to claim 9, characterized in that, The gated weighted fusion step in step 6 includes: Step 6.1: Compress local and global features to the same dimension using a fully connected layer; Step 6.2: Calculate the fusion weights and perform gated weighted fusion using the Sigmoid function.

Images