A voiceprint anomaly detection method and system based on wavelet convolution time domain feature extraction

Abstract

The present application belongs to the field of acoustic signal processing, deep learning and intelligent operation and maintenance of industrial equipment, and discloses a voiceprint anomaly detection method and system based on wavelet convolution time domain feature extraction. In view of the technical bottlenecks such as poor robustness of voiceprint feature extraction and weak model generalization ability in the industrial scene, the multi-resolution analysis ability of wavelet packet transform and the end-to-end learning ability of convolutional neural network are deeply integrated to construct a wavelet convolution time domain feature extraction module with strong robustness, and an unsupervised anomaly classification module is built in combination with a variational autoencoder to form an end-to-end integrated model of feature extraction and anomaly classification. The present application only needs voiceprint samples of normal operation of equipment to complete model training, and in the industrial scene with strong background noise, it can not only accurately capture the transient time domain features of early weak faults of equipment, but also realize unsupervised and high-accuracy anomaly state classification, effectively solving the industry pain points of low detection accuracy under strong noise interference and lack of abnormal samples in the industrial scene.

Description

Technical Field

[0001] This invention relates to the fields of acoustic signal processing, deep learning and intelligent operation and maintenance of industrial equipment, and in particular to a method and system for detecting voiceprint anomalies based on wavelet convolution temporal feature extraction. Background Technology

[0002] With the development of Industry 4.0 and intelligent manufacturing, industrial equipment is becoming increasingly large-scale and sophisticated. Its operational status directly impacts production safety and efficiency, making the construction of an efficient equipment fault early warning and diagnosis system a core requirement for intelligent operation and maintenance. Voiceprint detection, due to its advantages of being non-invasive, having good real-time performance, and being easy to deploy, has been widely used in anomaly monitoring of various industrial equipment and is a crucial support for intelligent operation and maintenance. Equipment faults cause changes in the temporal characteristics of acoustic signals; the accuracy of voiceprint feature extraction and the accuracy of anomaly classification directly determine the reliability of early fault identification and early warning.

[0003] Current industrial voiceprint anomaly detection mainly consists of two stages: feature extraction and anomaly classification, but both have significant limitations. In feature extraction, while traditional signal processing methods can achieve multi-scale decomposition, they lack adaptive learning capabilities, exhibit poor noise resistance, and are prone to losing fault information. CNN-based deep learning methods, although capable of end-to-end learning, suffer from fixed convolution kernels, making them difficult to adapt to non-stationary voiceprint signals and multi-scale faults, and insufficiently capturing transient features. In anomaly classification, mainstream supervised models rely on a large number of labeled samples, while industrial scenarios suffer from scarce anomaly samples and unknown fault types, leading to overfitting and weak generalization. Furthermore, existing technologies struggle to achieve end-to-end collaboration between robust multi-scale feature extraction and unsupervised anomaly classification with limited samples; the modules are fragmented, failing to meet practical deployment requirements.

[0004] In summary, existing technologies cannot simultaneously address the dual challenges of insufficient robustness in feature extraction and poor model generalization ability under limited sample sizes, becoming a key bottleneck for the engineering implementation of industrial voiceprint anomaly detection. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a method and system for voiceprint anomaly detection based on wavelet convolutional temporal feature extraction. This invention deeply integrates the multi-resolution analysis capabilities of wavelet packet transform with the end-to-end learning capabilities of convolutional neural networks to construct a robust wavelet convolutional temporal feature extraction module. Simultaneously, it combines a variational autoencoder (VAE) to build an unsupervised anomaly classification module, forming an end-to-end integrated model of feature extraction and anomaly classification. This invention requires only voiceprint samples from normally operating equipment to complete model training. In industrial scenarios with strong background noise, it can accurately capture the transient temporal features of early, subtle equipment faults and achieve unsupervised, high-accuracy anomaly state classification, effectively solving the industry pain points of scarce anomaly samples and low detection accuracy under strong noise interference in industrial scenarios.

[0006] To achieve the above technical objectives, the technical solution adopted by the present invention is as follows: a method for detecting voiceprint anomalies based on wavelet convolution temporal feature extraction, comprising the following steps; Step 1: Acquire and preprocess the industrial equipment acoustic print dataset to generate a standardized single-channel time-domain acoustic print sequence; Step 2: Construct a one-dimensional wavelet convolution module WTConv1d to perform multi-scale feature extraction and adaptive enhancement on the standardized single-channel temporal voiceprint sequence, and output wavelet convolution features that fuse multi-scale information. Step 3: Construct a Time-Domain Wavelet Convolution Network (TDWCN) based on a one-dimensional wavelet convolution module, perform high-level semantic abstraction and feature encoding on the wavelet convolution features, and generate a voiceprint time-domain feature vector. Step 4: Construct a variational autoencoder anomaly classification module (VAE), perform end-to-end unsupervised training based on the temporal feature vectors of normal samples' voiceprints, learn the latent space probability distribution of normal voiceprint features, and determine the anomaly judgment threshold. Step 5: Input the voiceprint signal to be detected into the trained TDWCN and variational autoencoder anomaly classification module. The TDWCN generates a time-domain feature vector, and the variational autoencoder anomaly classification module calculates the anomaly score. Based on the preset threshold, a binary classification output of the device's normal and abnormal states is obtained.

[0007] Step 1 specifically involves: Step 1.1: Acquire raw acoustic signature signals from the industrial equipment using acoustic sensors. Save it as a WAV file; Step 1.2: Convert the original voiceprint signal Resampling to a preset fixed sampling frequency All voiceprint signals have the same bandwidth; multi-channel audio signals are converted into mono signals, reducing data dimensionality while preserving core voiceprint time-domain information, resulting in a voiceprint time-domain sequence. ; Step 1.3: Convert the voiceprint time-domain sequence The length is fixed to a preset value. Generate standardized single-channel time-domain acoustic signature sequences ,in It is the set of real numbers.

[0008] Step 2 specifically involves: Step 2.1: Perform wavelet packet decomposition based on one-dimensional convolution operation, and process the standardized single-channel temporal acoustic sequence obtained in Step 1.3. Perform hierarchical multi-scale decomposition and extract the first... Low-frequency components corresponding to layer decomposition With high frequency components The number of decomposition layers, , Define the maximum number of decomposition layers; define the decomposition filter tensor. , The number of feature channels, Given the filter kernel length, decompose the filter coefficients corresponding to each channel in the filter tensor, and then compare them with the low-pass coefficients of the wavelet packet decomposition. High-pass decomposition coefficient One-to-one correspondence; for the first Low-frequency components obtained by scale wavelet decomposition Perform one-dimensional convolution operation to complete the first... The wavelet decomposition of the scale is calculated using the following formula: Where the fill length Ensure that the length of the decomposed sequence is half the length of the original sequence; number of groups A channel-group convolution strategy is adopted to ensure that each channel independently completes the wavelet decomposition operation; convolution stride... This achieves binary downsampling, halving the length of the decomposed sequence, which aligns with the core characteristic of wavelet packet decomposition: step-by-step binary scaling. The decomposition output tensor... , Indicates batch size, The standardized single-channel time-domain acoustic signature sequence from step 1.3 The original length, The decomposition level is the number of layers, including the first layer. Low-frequency components of scale wavelet decomposition With high frequency components The above decomposition process is executed recursively to complete the preset task. Multiscale wavelet decomposition of layers yields Low-frequency components corresponding to each scale With high frequency components ; Step 2.2: For the low-frequency components at each decomposition scale With high frequency components Feature recombination, grouped convolution, and scale modulation are performed to enhance fault-related features and suppress noise interference; firstly, the first... Tensors after scale decomposition Remodeling , making Low-frequency components at scale With high frequency components Unify and integrate to the channel dimension; for the reorganized tensor The formula for performing a one-dimensional grouped convolution operation is as follows: in, For the first The convolution kernel corresponding to the scale, To ensure that the sequence length remains unchanged after convolution, the number of groups... This allows low-frequency and high-frequency components to complete feature transformations in their respective independent spaces, avoiding the aliasing of features in different sub-bands and preserving the frequency domain resolution characteristics brought about by wavelet decomposition. The features are output through hierarchical convolution; finally, a scale modulation module is introduced. The output features of the grouped convolution are adaptively weighted channel by channel, and the calculation formula is as follows: ,in, To be consistent with scale Bound learnable scaling parameters, This represents element-wise multiplication; through the scale modulation module, it autonomously learns the importance weights of features at different scales and in different sub-bands, and performs adaptive enhancement and dynamic calibration of multi-scale wavelet features; Step 2.3: Deconvolution implementation of wavelet reconstruction. Wavelet packet reconstruction is performed based on transposed one-dimensional convolution, applying the scale-modulated features from Step 2.2. A step-by-step inverse transform fusion is performed to restore the time resolution to the original input, and the reconstructed multi-scale fused features are output. First, the reconstruction filter tensor is defined. The reconstructed filter tensor and the decomposed filter tensor in step 2.1 Satisfying the bioorthogonality condition ensures signal fidelity during the reconstruction process; Decomposed into enhanced low-frequency components according to channel dimension With enhanced high-frequency components The input tensor is then reassembled along the frequency domain dimension to obtain the reconstructed input tensor. Based on reconstruction filter tensor For reconstructing the input tensor Performing a transposed one-dimensional convolution operation completes the single-scale inverse decomposition process. The calculation formula is as follows: Among them, step size This achieves doubling of the sequence length, forming the inverse operation of doubling downsampling in the decomposition process; number of groups. Continuing the channel grouping strategy from the decomposition phase, ensure that each channel completes reconstruction independently; fill length To maintain consistency with the decomposition stage, scale matching is ensured during the reconstruction process. Finally, the reconstructed tensor is pruned to the original scale sequence length to ensure dimensionality matching. The above reconstruction process is then recursively executed, reconstructing in reverse order from the highest scale to the lowest scale, outputting a feature sequence of the same length as the original input. .

[0009] Step 3 includes: Step 3.1: Construct a two-layer cascaded one-dimensional wavelet convolution module. The first-layer one-dimensional wavelet convolution module extracts the overall operating status and long-term trend features of the equipment. The second-layer one-dimensional wavelet convolution module analyzes micro-scale distortion and local anomalous disturbances, and outputs multi-scale enhanced temporal features. ; Step 3.2: Construct a three-layer cascaded convolutional feature transformation module to enhance the temporal features after multi-scale enhancement. High-level semantic abstraction and feature encoding are performed. Each convolutional feature transformation module consists of a normalization layer, an activation function layer, and a one-dimensional convolutional layer cascaded in sequence, outputting high-dimensional discriminative features. ; Step 3.3: Construct an adaptive pooling layer for high-dimensional discriminative features. Adaptive compression of the temporal dimension is performed, mapping the feature temporal domain length to a preset fixed dimension, and outputting a standardized voiceprint temporal domain feature vector. ,in Number of output channels The preset fixed sequence length is used.

[0010] Step 4 includes: Step 4.1: Feature flattening: Flatten the temporal feature vectors of the voiceprints of normal samples in the training set. Flatten the vector to convert it into a one-dimensional flattened feature vector. ,in To flatten the feature dimensions; Step 4.2: Construct an encoder consisting of a three-layer fully connected network to perform nonlinear mapping on the one-dimensional flattened feature vector, learn the Gaussian probability distribution of normal voiceprint features in the latent space, and output the mean vector corresponding to the latent variables. With log-variance vector The core calculation formula is: in, For the nonlinear mapping function of the encoder, For the hidden layer output of the encoder, , The weight matrix and bias terms for the mean branch are shown. , Here are the weight matrix and bias terms for the variance branch; Step 4.3: Use reparameterization techniques to sample random noise from the standard normal distribution. Combined with the mean vector output by the encoder With variance vector The latent variables were calculated. The calculation formula is: ; Step 4.4: Construct a decoder with a symmetric structure to the encoder. Based on the latent variable z obtained by reparameterized sampling, perform nonlinear mapping through a three-layer fully connected network to reconstruct the reconstructed feature vector with the same dimension as the input. ; Step 4.5: Construct a joint loss function consisting of reconstruction loss and KL divergence loss, and achieve end-to-end optimization of the encoder and decoder through backpropagation. The formula for calculating the joint loss function is as follows: in, For the reconstruction loss, the mean squared error (MSE) is used to calculate the difference between the input features and the reconstructed features; KL divergence loss is used to constrain the distribution of latent variables to approximate the standard normal distribution; Step 4.6: Anomaly Threshold Determination: Input all normal samples from the training set into the model, calculate the anomaly score for each sample, and determine the anomaly threshold using the quantile method. The anomaly score is the mean squared error between the input features and the reconstructed features, calculated using the following formula: .

[0011] Step 5 includes: Step 5.1: Perform standardized preprocessing operations on the voiceprint signals to be detected collected in real time at the industrial site, consistent with the training set, to generate a standardized time-domain voiceprint sequence; Step 5.2: Temporal feature extraction: Input the standardized temporal voiceprint sequence into the trained TDWCN network to generate the corresponding temporal feature vector of the voiceprint. Step 5.3: Flatten the temporal feature vector of the voiceprint and input it into the variational autoencoder anomaly classification module. After the encoder and decoder, the reconstructed features are obtained, and the anomaly score of the sample to be detected is calculated. ; Step 5.4: Anomaly Classification and Result Output: A threshold judgment is made on the anomaly score of the sample to be tested. If... If the sample is in an abnormal equipment state, an abnormal warning message will be output; if If the sample is found to be in normal condition, the normal condition detection result will be output.

[0012] A voiceprint anomaly detection system based on wavelet convolution temporal feature extraction, which implements the voiceprint anomaly detection method based on wavelet convolution temporal feature extraction, includes a voiceprint preprocessing module 21, a wavelet convolution temporal feature extraction module 22, and a variational autoencoder anomaly classification module 23; The voiceprint preprocessing module 21 is used to acquire the original voiceprint signal of the industrial equipment operation, perform sampling rate standardization, channel unification and sequence length normalization processing to generate a standardized single-channel time-domain voiceprint sequence. The wavelet convolution temporal feature extraction module 22 inputs the standardized temporal voiceprint sequence into the built-in two-layer one-dimensional wavelet convolution unit and three-layer convolution feature transformation unit to complete multi-scale wavelet decomposition, intra-scale feature transformation, wavelet reconstruction and high-level feature abstraction encoding, and outputs a fixed-dimensional voiceprint temporal feature vector. The variational autoencoder anomaly classification module 23 takes the voiceprint temporal feature vector as input and sequentially performs feature flattening, latent space distribution learning, feature reconstruction, and anomaly score calculation; it determines whether the device is in normal or abnormal state based on a preset threshold and outputs the detection result.

[0013] The voiceprint preprocessing module 21 includes: The sampling rate standardization unit 211 is used to perform unified sampling rate processing on the acoustic fingerprint signal of industrial equipment to eliminate the influence of differences in sampling parameters of different acquisition devices. The channel unification unit 212 is used to unify multi-channel voiceprint signals into mono signals to ensure that the input format is consistent in subsequent processing. The sequence length normalization unit 213 is used to normalize the length of the voiceprint signal sequence and output a standardized time-domain voiceprint sequence with matching dimensions and uniform format.

[0014] The wavelet convolution temporal feature extraction module 22 includes: The dual-layer one-dimensional wavelet convolutional unit 221 is used to convolutionalize and learn the multi-resolution time-frequency decomposition process of wavelet packet transform, realize the multi-scale decomposition, feature enhancement and reconstruction fusion of voiceprint signal, and effectively separate fault features from background noise. The three-layer convolutional feature transformation unit 222 is used to perform hierarchical modeling of wavelet convolution output features from coarse to fine, fully characterizing the long-term trend features and local transient fault features of equipment operation, and outputting a highly discriminative voiceprint time-domain feature vector.

[0015] The variational autoencoder anomaly classification module 23 includes: The feature flattening unit 231 is used to convert the two-dimensional voiceprint temporal feature vector into a one-dimensional feature vector to adapt to the input requirements of the fully connected network. The encoder unit 232 is used to perform nonlinear mapping on the flattened features, learn the probability distribution of normal voiceprints in the latent space, and output the mean vector and variance vector. Decoder unit 233 is used to reconstruct a feature vector of the same dimension as the input based on the reparameterized latent variables. Loss calculation unit 234 is used to calculate the joint loss function of the model to achieve end-to-end optimization of the encoder and decoder; The anomaly determination unit 235 is used to calculate the reconstruction error between the input and the reconstruction features and generate an anomaly score, and complete the binary classification of the device as normal or abnormal based on a preset threshold.

[0016] The beneficial effects of adopting the above technical solution are as follows: 1. This invention proposes a one-dimensional wavelet convolution WTConv1d unit, which convolutions and makes the multi-resolution time-frequency decomposition process of wavelet packet transform learnable. It combines the multi-scale analysis capability of wavelet transform for non-stationary signals with the end-to-end adaptive learning capability of convolutional neural networks. It solves the technical pain points of traditional wavelet transform lacking task adaptability and traditional CNNs having difficulty capturing multi-scale fault features. It can accurately separate fault features and background noise in industrial acoustic signals, significantly improve the anti-interference capability and robustness of time-domain feature extraction, and has a strong ability to capture transient features of early weak faults in equipment.

[0017] 2. This invention deeply integrates wavelet convolution temporal feature extraction with VAE unsupervised anomaly classification to construct an end-to-end voiceprint anomaly detection model. The model can be trained using only voiceprint samples from equipment in normal operation, which completely solves the industry pain point of insufficient generalization ability of supervised models caused by the scarcity of anomaly samples and unpredictable fault types in industrial scenarios. It also significantly reduces the deployment threshold and data annotation cost of the model in industrial sites.

[0018] 3. This invention achieves end-to-end joint training of feature extraction and anomaly classification through a collaborative optimization architecture of wavelet convolution feature extraction and VAE classification. The wavelet convolution module provides highly discriminative temporal features for the VAE classification module, significantly reducing the difficulty of VAE distribution learning and improving the accuracy of anomaly detection. The gradient backpropagation of the VAE module can simultaneously optimize the feature extraction capability of the wavelet convolution module, making feature extraction more aligned with the task objective of anomaly detection. The two form a positive synergistic effect, significantly reducing the false alarm rate and false negative rate in industrial scenarios with strong background noise, and have extremely strong engineering practical value. Attached Figure Description

[0019] Figure 1 is a schematic diagram of the voiceprint anomaly detection method based on wavelet convolution temporal feature extraction provided in an embodiment of the present invention.

[0020] Figure 2 is a block diagram of the voiceprint anomaly detection system based on wavelet convolution temporal feature extraction used in this invention.

[0021] In the figure: 20 - Voiceprint anomaly detection system based on wavelet convolution temporal feature extraction; 21 - Voiceprint preprocessing module; 22 - Wavelet convolution temporal feature extraction module; 23 - Variational autoencoder anomaly classification module; 211 - Sampling rate normalization unit; 212 - Channel unification unit; 213 - Sequence length normalization unit; 221 - Two-layer one-dimensional wavelet convolution unit; 222 - Three-layer convolution feature transformation unit; 231 - Feature flattening unit; 232 - Encoder unit; 233 - Decoder unit; 234 - Loss calculation unit; 235 - Anomaly determination unit. Detailed Implementation

[0022] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0023] As shown in Figure 1, the voiceprint anomaly detection system based on wavelet convolution temporal feature extraction is as follows: the system includes a voiceprint preprocessing module, a wavelet convolution temporal feature extraction module, and a variational autoencoder (VAE) anomaly classification module; The methods include: The original acoustic signature signal of the industrial equipment is obtained through the acoustic signature preprocessing module. Sampling rate standardization, channel unification and sequence length normalization are performed to generate a standardized single-channel time-domain acoustic signature sequence. The standardized temporal voiceprint sequence is input into the wavelet convolution temporal feature extraction module. Through the built-in WTConv1d unit and TDWCN network, multi-scale feature extraction and high-level feature encoding are completed, and a fixed-dimensional voiceprint temporal feature vector is output. The voiceprint temporal feature vector is input into the variational autoencoder (VAE) anomaly classification module to complete feature reconstruction and anomaly score calculation. Based on a preset threshold, the classification result of the device's normal or abnormal state is output, and the processed structured data is stored in the database.

[0024] The voiceprint preprocessing module is responsible for acquiring WAV format voiceprint data from industrial equipment, extracting audio time-domain waveform information, and performing sampling rate standardization, channel unification, and sequence length normalization on the voiceprint signal to provide a standardized data foundation for subsequent processing. The wavelet convolutional temporal feature extraction module completes multi-scale wavelet decomposition, adaptive feature enhancement and reconstruction fusion of the voiceprint signal through the WTConv1d unit, and completes high-level semantic abstraction and encoding of the features through the TDWCN network structure, outputting a highly discriminative voiceprint temporal feature vector. The variational autoencoder anomaly classification module (VAE) includes a feature flattening unit, an encoder, a decoder, a loss calculation unit, and an anomaly determination unit. It can complete unsupervised training with only normal samples and achieve accurate identification of abnormal device states by reconstructing errors.

[0025] On the other hand, the present invention also provides a voiceprint anomaly detection method based on wavelet convolution temporal feature extraction, comprising the following steps: Step 1: Acquire and preprocess the industrial equipment acoustic signature dataset, specifically including: Step 1.1: Data Acquisition and Format Standardization: The MIMII Industrial Voiceprint Public Dataset is used, which includes normal and abnormal operation voiceprint signals of four typical industrial equipment: fans, pumps, valves, and slide rails. The signal-to-noise ratio is 0dB, simulating a moderate background noise environment in an industrial setting. The audio time-domain waveform data is extracted and saved as a WAV file. Step 1.2: Sampling Rate and Channel Standardization: Resample all voiceprint audio signals to a fixed sampling frequency of 16kHz to ensure that the bandwidth of all signals is consistent; convert the two-channel audio signals into mono signals by calculating the average of the left and right channels to reduce the data dimensionality. Step 1.3: Sequence Length Normalization: The length of the single-channel time-domain acoustic sequence is fixed at 160,000 sampling points (corresponding to 10 seconds of audio). Sequences longer than 160,000 sampling points are truncated at the end, and sequences shorter than 160,000 sampling points are zero-padding at the end to generate a standardized single-channel time-domain acoustic sequence. The dataset is divided into a training set and a test set. The training set contains only voiceprint samples of devices in normal operation, accounting for 70%, while the test set contains normal samples and abnormal samples, accounting for 30%.

[0026] Step 2: Construct a one-dimensional wavelet convolution WTConv1d module to perform multi-scale feature extraction and adaptive enhancement on the standardized temporal voiceprint sequence, specifically including: Step 2.1: Convolutional Implementation of Multi-Scale Wavelet Decomposition: This embodiment employs a two-layer wavelet decomposition. A decomposition filter tensor is constructed based on the db4 wavelet basis function, with a filter kernel length K=8. The low-pass and high-pass filter coefficients corresponding to each channel in the decomposition filter tensor correspond to the decomposition coefficients of the db4 wavelet. The first-layer wavelet decomposition is performed on the input standardized time-domain acoustic signature sequence. The stride of the one-dimensional convolution operation is stride=2, the number of groups is the same as the number of input channels, and the padding length P=3. After decomposition, the low-frequency components of the first layer are output. With high frequency components The sequence length becomes 80000; for the low-frequency components of the first layer... Perform a second-level wavelet decomposition and output the low-frequency components of the second level. With high frequency components The sequence length becomes 40000; the low-frequency components before each decomposition are length checked, and if the sequence length is odd, zeros are padded at the end of the sequence to ensure that the sequence length matches during the decomposition process.

[0027] Step 2.2: Intra-scale feature transformation: Reshape the tensors after decomposition at each scale along the channel dimension, integrating low-frequency and high-frequency components into the channel dimension to double the number of channels in the feature tensor; Perform one-dimensional grouped convolution on the recombined features, with a kernel size of 3×1 and a padding length of 1, and the number of groups is consistent with the number of channels after recombination; Bind learnable scaling parameters to each scale, and perform adaptive weighting on the convolution output features channel by channel to enhance fault-related features and suppress noise interference.

[0028] Step 2.3: Deconvolution Implementation of Wavelet Reconstruction: Construct a reconstruction filter tensor based on the reconstruction filter coefficients of the db4 wavelet basis, which satisfies the biorthogonality condition with the decomposition filter tensor; concatenate the high and low frequency sub-band features of the second layer after feature transformation along the frequency domain dimension, and perform a transposed one-dimensional convolution operation with stride=2 to reconstruct the feature sequence of the first layer scale, restoring the length to 80000; then concatenate the reconstructed first layer features with the high and low frequency features of the first layer after feature transformation, and perform a second transposed one-dimensional convolution to reconstruct the feature sequence of the same length as the original input, restoring the length to 160000; prune the reconstructed sequence to ensure that it completely matches the length of the original input sequence.

[0029] Step 3: Construct a temporal wavelet convolutional network (TDWCN) based on the WTConv1d module to perform high-level semantic abstraction and feature encoding on the wavelet convolutional features, specifically including: Step 3.1: Cascaded feature extraction of two-layer WTConv1d modules: The first-layer WTConv1d module has 1 input channel, 64 output channels, an input sequence length of 160,000, and an output feature size of 64×160,000; the second-layer WTConv1d module has 64 input channels, 128 output channels, and an output feature size of 128×160,000.

[0030] Step 3.2: Multi-layer convolutional feature transformation encoding: Construct a three-layer cascaded convolutional feature transformation module. Each layer consists of a LayerNorm layer, a LeakyReLU activation function layer (slope of 0.2 in the negative interval), and a Conv1D convolutional layer cascaded in sequence. The first Conv1D convolutional kernel size is 5×1, stride is 4, number of output channels is 128, and output feature size is 128×40000. The second Conv1D convolutional kernel size is 3×1, stride is 4, number of output channels is 128, and output feature size is 128×10000. The third Conv1D convolutional kernel size is 3×1, stride is 8, number of output channels is 128, and output feature size is 128×1250.

[0031] Step 3.3: Adaptive Pooling and Feature Output: Through an adaptive average pooling layer, the 128×1250 feature tensor is adaptively mapped to a 128×313 voiceprint temporal feature vector, completing the temporal feature extraction process.

[0032] Step 4: Construct the Variational Autoencoder (VAE) anomaly classification module to complete model training and anomaly threshold determination, specifically including: Step 4.1: Feature flattening process: Flatten the 128×313 voiceprint temporal feature vector of normal samples in the training set into a one-dimensional flattened feature vector with a dimension of 40064.

[0033] Step 4.2: VAE Encoder Distribution Learning: Construct a three-layer fully connected encoder. The first fully connected layer maps the 40064-dimensional features to 2048 dimensions, the second layer maps to 256 dimensions, and the third layer maps to a 64-dimensional hidden layer. The hidden layer output is divided into two parallel branches, each outputting a mean vector of dimension 16. With log-variance vector Each fully connected layer is followed by a LayerNorm normalization layer and a LeakyReLU activation function layer.

[0034] Step 4.3: Reparameterized sampling: Sampling 16-dimensional random noise from a standard normal distribution. Calculate latent variables by combining the mean and variance of the encoder output. The dimension is 16.

[0035] Step 4.4: VAE Decoder Feature Reconstruction: Construct a three-layer fully connected decoder symmetrical to the encoder. The first layer maps the 16-dimensional latent variables to 256 dimensions, the second layer maps to 204 dimensions, and the third layer maps back to a 40064-dimensional reconstructed feature vector. Each fully connected layer is followed by a LayerNorm normalization layer and a LeakyReLU activation function layer. The final output layer uses a linear activation function.

[0036] Step 4.5: Joint Loss Function Calculation and Model Optimization: In this embodiment, the weight balancing coefficient of the joint loss function is... The value is 0.5, the Adam optimizer is used, the initial learning rate is set to 1e-4, the batch size is set to 32, the number of training epochs is set to 50, and all weight parameters of the model are updated through backpropagation to achieve end-to-end unsupervised training. After each training epoch, the model reconstruction accuracy is verified on the validation set, and the model weights with the lowest reconstruction loss are saved as the optimal model.

[0037] Step 4.6: Anomaly Threshold Determination: Input all normal samples from the training set into the trained optimal model, calculate the anomaly score for each sample, and take the 95th percentile of all anomaly scores as the final anomaly threshold. .

[0038] Step 5: Anomaly detection and classification output of the voiceprint signal to be detected, specifically including: Step 5.1: Preprocessing of the voiceprint signal to be detected: Perform the same standardized preprocessing operation as the training set on the voiceprint signals to be detected in the test set to generate a standardized time-domain voiceprint sequence; Step 5.2: Temporal feature extraction: Input the standardized temporal voiceprint sequence into the trained TDWCN network to generate a 128×313 voiceprint temporal feature vector; Step 5.3: Feature Reconstruction and Anomaly Score Calculation: The speaker temporal feature vector is flattened and input into the VAE module. After passing through the encoder and decoder, the reconstructed features are obtained, and the anomaly score of the sample to be detected is calculated. .

[0039] Step 5.4: Anomaly Classification and Result Output: A threshold judgment is made on the anomaly score of the sample to be tested. If... If the sample is in an abnormal equipment state, an abnormal warning message will be output; if If the sample is found to be in normal condition, the normal condition detection result will be output.

[0040] In this embodiment, the model performance was verified using the MIMII dataset. Experimental results show that the model of the present invention achieves an average AUC of 96.27% on four types of equipment: fans, pumps, valves, and slide rails. This is an improvement of 22.76 percentage points compared to the traditional AE model and 13.42 percentage points compared to the model using only VAE. It also maintains extremely high detection accuracy in strong background noise scenarios, fully verifying the effectiveness and superiority of the method of the present invention.

[0041] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for voiceprint anomaly detection based on wavelet convolution temporal feature extraction, characterized in that, Includes the following steps; Step 1: Acquire and preprocess the industrial equipment acoustic print dataset to generate a standardized single-channel time-domain acoustic print sequence; Step 2: Construct a one-dimensional wavelet convolution module to perform multi-scale feature extraction and adaptive enhancement on the standardized single-channel temporal voiceprint sequence, and output wavelet convolution features that fuse multi-scale information; Step 3: Construct a temporal wavelet convolutional network TDWCN based on a one-dimensional wavelet convolution module, perform high-level semantic abstraction and feature encoding on the wavelet convolution features, and generate a voiceprint temporal feature vector. Step 4: Construct a variational autoencoder anomaly classification module, perform end-to-end unsupervised training based on the voiceprint temporal feature vector of normal samples, learn the latent space probability distribution of normal voiceprint features, and determine the anomaly judgment threshold. Step 5: Input the voiceprint signal to be detected into the trained TDWCN and variational autoencoder anomaly classification module. The TDWCN generates a time-domain feature vector, and the variational autoencoder anomaly classification module calculates the anomaly score. Based on the preset threshold, a binary classification output of the device's normal and abnormal states is obtained.

2. The voiceprint anomaly detection method based on wavelet convolution temporal feature extraction according to claim 1, characterized in that, Step 1 specifically involves: Step 1.1: Acquire raw acoustic signature signals from the industrial equipment using acoustic sensors. Save it as a WAV file; Step 1.2: Convert the original voiceprint signal Resampling to a preset fixed sampling frequency All voiceprint signals have the same bandwidth; multi-channel audio signals are converted into mono signals, reducing data dimensionality while preserving core voiceprint time-domain information, resulting in a voiceprint time-domain sequence. ; Step 1.3: Convert the voiceprint time-domain sequence The length is fixed to a preset value. Generate standardized single-channel time-domain acoustic signature sequences ,in It is the set of real numbers.

3. The voiceprint anomaly detection method based on wavelet convolution temporal feature extraction according to claim 1, characterized in that, Step 2 specifically involves: Step 2.1: Perform wavelet packet decomposition based on one-dimensional convolution operation, and process the standardized single-channel temporal acoustic sequence obtained in Step 1.

3. Perform hierarchical multi-scale decomposition and extract the first... Low-frequency components corresponding to layer decomposition With high frequency components The number of decomposition layers, , Define the maximum number of decomposition layers; define the decomposition filter tensor. , The number of feature channels, Given the filter kernel length, decompose the filter coefficients corresponding to each channel in the filter tensor, and then compare them with the low-pass coefficients of the wavelet packet decomposition. High-pass decomposition coefficient One-to-one correspondence; for the first Low-frequency components obtained by scale wavelet decomposition Perform one-dimensional convolution operation to complete the first... The wavelet decomposition of the scale is calculated using the following formula: Where the fill length Ensure that the length of the decomposed sequence is half the length of the original sequence; number of groups A channel-group convolution strategy is adopted to ensure that each channel independently completes the wavelet decomposition operation; convolution stride... This achieves binary downsampling, halving the length of the decomposed sequence, which aligns with the core characteristic of wavelet packet decomposition: step-by-step binary scaling. The decomposition output tensor... , Indicates batch size, The standardized single-channel time-domain acoustic signature sequence from step 1.3 The original length, The decomposition level is the number of layers, including the first layer. Low-frequency components of scale wavelet decomposition With high frequency components ; Recursively execute the above decomposition process to complete the preset... Multiscale wavelet decomposition of layers yields Low-frequency components corresponding to each scale With high frequency components ; Step 2.2: For the low-frequency components at each decomposition scale With high frequency components Feature recombination, grouped convolution, and scale modulation are performed to enhance fault-related features and suppress noise interference; firstly, the first... Tensors after scale decomposition Remodeling , making Low-frequency components at scale With high frequency components Unify and integrate to the channel dimension; for the reorganized tensor The formula for performing a one-dimensional grouped convolution operation is as follows: in, For the first The convolution kernel corresponding to the scale, To ensure that the sequence length remains unchanged after convolution, the number of groups... This allows low-frequency and high-frequency components to complete feature transformations in their respective independent spaces, avoiding the aliasing of features in different sub-bands and preserving the frequency domain resolution characteristics brought about by wavelet decomposition. The features are output through hierarchical convolution; finally, a scale modulation module is introduced. The output features of the grouped convolution are adaptively weighted channel by channel, and the calculation formula is as follows: ,in, To be consistent with scale Bound learnable scaling parameters, This represents element-wise multiplication; through the scale modulation module, it autonomously learns the importance weights of features at different scales and in different sub-bands, and performs adaptive enhancement and dynamic calibration of multi-scale wavelet features; Step 2.3: Deconvolution implementation of wavelet reconstruction. Wavelet packet reconstruction is performed based on transposed one-dimensional convolution, applying the scale-modulated features from Step 2.

2. A step-by-step inverse transform fusion is performed to restore the time resolution to the original input, and the reconstructed multi-scale fused features are output. First, the reconstruction filter tensor is defined. The reconstructed filter tensor and the decomposed filter tensor in step 2.1 Satisfying the bioorthogonality condition ensures signal fidelity during the reconstruction process; Decomposed into enhanced low-frequency components according to channel dimension With enhanced high-frequency components The input tensor is then reassembled along the frequency domain dimension to obtain the reconstructed input tensor. Based on reconstruction filter tensor For reconstructing the input tensor Performing a transposed one-dimensional convolution operation completes the single-scale inverse decomposition process. The calculation formula is as follows: Among them, step size This achieves doubling of the sequence length, forming the inverse operation of doubling downsampling in the decomposition process; number of groups. Continuing the channel grouping strategy from the decomposition phase, ensure that each channel completes reconstruction independently; fill length To maintain consistency with the decomposition stage, scale matching is ensured during the reconstruction process. Finally, the reconstructed tensor is pruned to the original scale sequence length to ensure dimensionality matching. The above reconstruction process is then recursively executed, reconstructing in reverse order from the highest scale to the lowest scale, outputting a feature sequence of the same length as the original input. .

4. The voiceprint anomaly detection method based on wavelet convolution temporal feature extraction according to claim 1, characterized in that, Step 3 includes: Step 3.1: Construct a two-layer cascaded one-dimensional wavelet convolution module. The first-layer one-dimensional wavelet convolution module extracts the overall operating status and long-term trend features of the equipment. The second-layer one-dimensional wavelet convolution module analyzes micro-scale distortion and local anomalous disturbances, and outputs multi-scale enhanced temporal features. ; Step 3.2: Construct a three-layer cascaded convolutional feature transformation module to enhance the temporal features after multi-scale enhancement. High-level semantic abstraction and feature encoding are performed. Each convolutional feature transformation module consists of a normalization layer, an activation function layer, and a one-dimensional convolutional layer cascaded in sequence, outputting high-dimensional discriminative features. ; Step 3.3: Construct an adaptive pooling layer for high-dimensional discriminative features. Adaptive compression of the temporal dimension is performed, mapping the feature temporal domain length to a preset fixed dimension, and outputting a standardized voiceprint temporal domain feature vector. ,in Number of output channels The preset fixed sequence length is used.

5. The voiceprint anomaly detection method based on wavelet convolution temporal feature extraction according to claim 1, characterized in that, Step 4 includes: Step 4.1: Feature flattening: Flatten the temporal feature vectors of the voiceprints of normal samples in the training set. Flatten the vector to convert it into a one-dimensional flattened feature vector. ,in To flatten the feature dimensions; Step 4.2: Construct an encoder consisting of a three-layer fully connected network to perform nonlinear mapping on the one-dimensional flattened feature vector, learn the Gaussian probability distribution of normal voiceprint features in the latent space, and output the mean vector corresponding to the latent variables. With log-variance vector The core calculation formula is: in, For the nonlinear mapping function of the encoder, For the hidden layer output of the encoder, , The weight matrix and bias terms for the mean branch are shown. , Here are the weight matrix and bias terms for the variance branch; Step 4.3: Use reparameterization techniques to sample random noise from the standard normal distribution. Combined with the mean vector output by the encoder With variance vector The latent variables were calculated. The calculation formula is: ; Step 4.4: Construct a decoder with a symmetric structure to the encoder. Based on the latent variable z obtained by reparameterized sampling, perform nonlinear mapping through a three-layer fully connected network to reconstruct the reconstructed feature vector with the same dimension as the input. ; Step 4.5: Construct a joint loss function consisting of reconstruction loss and KL divergence loss, and achieve end-to-end optimization of the encoder and decoder through backpropagation. The formula for calculating the joint loss function is as follows: in, For the reconstruction loss, the mean squared error (MSE) is used to calculate the difference between the input features and the reconstructed features; KL divergence loss is used to constrain the distribution of latent variables to approximate the standard normal distribution; Step 4.6: Anomaly Threshold Determination: Input all normal samples from the training set into the model, calculate the anomaly score for each sample, and determine the anomaly threshold using the quantile method. The anomaly score is the mean squared error between the input features and the reconstructed features, calculated using the following formula: .

6. The voiceprint anomaly detection method based on wavelet convolution temporal feature extraction according to claim 1, characterized in that, Step 5 includes: Step 5.1: Perform standardized preprocessing operations on the voiceprint signals to be detected collected in real time at the industrial site, consistent with the training set, to generate a standardized time-domain voiceprint sequence; Step 5.2: Temporal feature extraction: Input the standardized temporal voiceprint sequence into the trained TDWCN network to generate the corresponding temporal feature vector of the voiceprint. Step 5.3: Flatten the temporal feature vector of the voiceprint and input it into the variational autoencoder anomaly classification module. After the encoder and decoder, the reconstructed features are obtained, and the anomaly score of the sample to be detected is calculated. ; Step 5.4: Anomaly Classification and Result Output: A threshold judgment is made on the anomaly score of the sample to be tested. If... If the sample is in an abnormal equipment state, an abnormal warning message will be output; if If the sample is found to be in normal condition, the normal condition detection result will be output.

7. A voiceprint anomaly detection system based on wavelet convolution temporal feature extraction, characterized in that, The voiceprint anomaly detection method based on wavelet convolution temporal feature extraction as described in any one of claims 1-6 includes a voiceprint preprocessing module (21), a wavelet convolution temporal feature extraction module (22), and a variational autoencoder anomaly classification module (23). The voiceprint preprocessing module (21) is used to acquire the original voiceprint signal of the industrial equipment operation, perform sampling rate standardization, channel unification and sequence length normalization processing to generate a standardized single-channel time-domain voiceprint sequence. The wavelet convolution temporal feature extraction module (22) inputs the standardized temporal voiceprint sequence into the built-in two-layer one-dimensional wavelet convolution unit and three-layer convolution feature transformation unit to complete multi-scale wavelet decomposition, intra-scale feature transformation, wavelet reconstruction and high-level feature abstraction encoding, and outputs a fixed-dimensional voiceprint temporal feature vector. The variational autoencoder anomaly classification module (23) takes the voiceprint temporal feature vector as input and sequentially performs feature flattening, latent space distribution learning, feature reconstruction and anomaly score calculation; it determines the normal or abnormal state of the device based on a preset threshold and outputs the detection result.

8. The voiceprint anomaly detection system based on wavelet convolution temporal feature extraction according to claim 7, characterized in that, The voiceprint preprocessing module (21) includes: The sampling rate normalization unit (211) is used to perform unified sampling rate processing on the acoustic fingerprint signal of industrial equipment to eliminate the influence of differences in sampling parameters of different acquisition devices. The channel unification unit (212) is used to unify multi-channel voiceprint signals into mono signals to ensure that the input format is consistent in subsequent processing. The sequence length normalization unit (213) is used to normalize the length of the voiceprint signal sequence and output a standardized time-domain voiceprint sequence with matching dimensions and uniform format.

9. The voiceprint anomaly detection system based on wavelet convolution temporal feature extraction according to claim 7, characterized in that, The wavelet convolution temporal feature extraction module (22) includes: A two-layer one-dimensional wavelet convolutional unit (221) is used to convolve and learn the multi-resolution time-frequency decomposition process of wavelet packet transform, realize the multi-scale decomposition, feature enhancement and reconstruction fusion of acoustic signal, and effectively separate fault features from background noise. The three-layer convolutional feature transformation unit (222) is used to perform hierarchical modeling of wavelet convolution output features from coarse to fine, fully characterizing the long-term trend features and local transient fault features of equipment operation, and outputting a highly discriminative voiceprint time-domain feature vector.

10. The voiceprint anomaly detection system based on wavelet convolution temporal feature extraction according to claim 7, characterized in that, The variational autoencoder anomaly classification module (23) includes: The feature flattening unit (231) is used to convert the two-dimensional voiceprint temporal feature vector into a one-dimensional feature vector to adapt to the input requirements of the fully connected network. The encoder unit (232) is used to perform nonlinear mapping on the flattened features, learn the probability distribution of normal voiceprints in the latent space, and output the mean vector and variance vector. The decoder unit (233) is used to reconstruct a feature vector of the same dimension as the input based on the reparameterized latent variables; The loss calculation unit (234) is used to calculate the joint loss function of the model and realize end-to-end optimization of the encoder and decoder; Anomaly determination unit (235) is used to calculate the reconstruction error between the input and the reconstruction features and generate anomaly score, and to complete the binary classification of the device as normal or abnormal based on a preset threshold.

Images