Equipment abnormal voiceprint detection system

Abstract

The invention provides an equipment abnormal voiceprint detection system. The equipment abnormal voiceprint detection system comprises a voiceprint collection module used for collecting original voiceprint signals in real time and preprocessing the original voiceprint signals; the edge end detection module is deployed on edge computing equipment and is used for carrying out real-time anomaly detection on the voiceprint features by utilizing a one-dimensional lightweight neural network model; the abnormity credibility evaluation module is used for converting a real-time abnormity detection result into a probabilistic abnormity credibility score; the incremental data screening module is used for screening high-value samples from the real-time voiceprint data based on a dynamic density trend sensing algorithm and caching the high-value samples in the edge; the cloud model evolution module is used for carrying out evolution training on the detection model by utilizing a continuous learning mechanism; the generation and playback module is used for jointly generating pseudo samples through a variational auto-encoder VAE and a generative adversarial network GAN; and the model updating module is used for compressing the evolved cloud model and then issuing and replacing the original model in the edge end detection module so as to form a cloud-edge collaborative sustainable evolution closed loop.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent operation and maintenance and industrial Internet of Things, in particular, especially relates to an equipment abnormal voiceprint detection system. BACKGROUND

[0002] Traditional equipment voiceprint anomaly detection technology mainly relies on methods based on spectrum analysis and rule matching for implementation. Common methods include short-time Fourier transform (STFT), Mel frequency cepstral coefficient (MFCC) extraction, and classification and recognition based on support vector machine (SVM), Gaussian mixture model (GMM), etc. These methods usually make abnormal judgments based on static features, and require expert experience for rule setting or manual parameter adjustment. In terms of model deployment, traditional methods mostly rely on centralized servers for batch processing analysis, which has problems of insufficient real-time performance, large communication bandwidth occupation, and poor adaptability to edge deployment. In addition, in the anomaly detection process, the traditional method mostly uses fixed threshold for judgment, which is difficult to accurately express the distribution characteristics of the abnormal level, and the response to early changes in the device state is relatively slow, which cannot meet the fine operation and maintenance requirements in complex industrial environments.

[0003] The application number 202211504455.8 discloses a sound monitoring device and method for real-time monitoring of emergency equipment generator operating state. The voiceprint information of the detected equipment is obtained, the voiceprint information is preprocessed, the voiceprint information is fitted and identified for similarity by using the voiceprint information experience library, and the noise is stripped. A multi-layer perceptron neural network is constructed, a GOA algorithm is used to iteratively optimize the model of the multi-layer perceptron neural network and deploy it to a cloud server to realize monitoring of the emergency equipment generator operating state. A visual platform is used to monitor the sound of the emergency equipment generator in real time. The running parameter data and fault data of the emergency equipment generator are displayed in real time, improving the accuracy of fault diagnosis of the emergency equipment generator. The platform is used for audio and video preview, the abnormal sound monitored is identified as a fault phenomenon by a model algorithm, and an alarm is pushed through the platform and short message.

[0004] The above scheme uses the grasshopper optimization algorithm (GOA) to extract multiple spectral envelope features for tone and volume states respectively, and builds a voiceprint model based on this. However, spectral envelope is inherently sensitive to noise, environmental interference, microphone position changes, etc., especially in actual emergency scenarios, the background is complex, which is easy to cause recognition errors. This scheme does not introduce an enhanced feature fusion mechanism or time series modeling capability, and lacks robustness improvement strategies for environmental adaptability.

[0005] The determination of the important indexes of the generator components only depends on the voiceprint similarity and the historical accumulated voiceprint information, and there is no real-time feedback or model adaptive updating mechanism. Once the device component state changes or a new type of fault occurs, the voiceprint features thereof cannot be well mapped in the original accumulated data, the system cannot timely adjust the identification model, and there is a "knowledge fossilization" problem. SUMMARY

[0006] In view of the technical problem of the non-stationary distribution of the voiceprint data in the operation of the industrial equipment leading to the degradation of the model performance, an abnormal voiceprint detection system for equipment is provided.

[0007] The technical means adopted by the present application are as follows: An abnormal voiceprint detection system for equipment comprises: A voiceprint collection module is configured to collect original voiceprint signals in real time during the operation of the industrial equipment and perform preprocessing. An edge detection module is disposed on an edge computing device and is configured to perform real-time abnormal detection on the preprocessed voiceprint features by using a one-dimensional lightweight neural network model. An abnormal confidence evaluation module is in communication connection with the edge detection module and is configured to convert the result of the real-time abnormal detection into a probabilistic abnormal confidence score. An incremental data screening module is connected with the abnormal confidence evaluation module and is configured to screen high-value samples from the real-time voiceprint data based on a dynamic density trend perception algorithm and cache the high-value samples at the edge. A cloud model evolution module is in communication connection with the edge side and is configured to perform evolution training on the detection model by using a continuous learning mechanism. A generative replay module is nested in the cloud model evolution module and is configured to generate pseudo samples by combining a variational autoencoder (VAE) and a generative adversarial network (GAN) to expand the training data. A model updating module is configured to compress the evolved cloud model and replace the original model in the edge detection module, thereby forming a continuous evolution closed loop of cloud-edge collaboration.

[0008] Further, the one-dimensional lightweight neural network model is based on the MobileNetV3 Small architecture, systematically reconstructs a two-dimensional depth separable convolution into a one-dimensional depth separable convolution, and embeds a channel-spatial dual attention mechanism (CBAM) and a Mish activation function in each inverted residual module, specifically as follows: The one-dimensional depth separable convolution processes the input signal The processing process of the one-dimensional depth separable convolution is represented as:

[0009] wherein, represents a depth convolution kernel size The temporal convolution operation, represents point-wise convolution, represents residual connection; In the channel dimension, the channel-space dual attention mechanism CBAM is introduced to enhance feature selectivity. The attention weight is generated by dual-path aggregation of global average pooling and maximum pooling. The formula is as follows:

[0010] in, is the feature map, is the Sigmoid function; In the spatial dimension, the spatial attention map is generated by feature compression on the channel axis. The formula is as follows:

[0011] in, is the Sigmoid function; The Mish activation function is used to optimize the gradient propagation path, while maintaining linear growth in the positive interval and introducing differentiable saturation characteristics. The expression is as follows:

[0012] Among them, the second-order derivative of the Mish activation function in the inverted residual module is always positive, which effectively alleviates the gradient vanishing problem during the back propagation process.

[0013] The encoder part of the one-dimensional lightweight neural network model consists of an initial convolutional layer, 10 improved inverted residual modules and an adaptive pooling layer; the expansion ratio of each inverted residual module is set to The progressive structure retains more detailed information in the shallow layer and focuses on abstract features in the deep layer. The decoder uses four-level transposed convolution for signal reconstruction, and its mathematical description formula is as follows:

[0014] in, is a latent variable, is the hidden state at each decoding stage.

[0015] Furthermore, the abnormal credibility assessment module is used to map the reconstruction error of the input voiceprint signal into a probabilistic confidence score, specifically including the following steps: The kernel density estimation method is used to perform probability density modeling on the reconstruction error of normal samples and construct a statistical confidence function of the degree of abnormal deviation. The reconstruction error of the to-be-tested sample is mapped into the density function, the cumulative probability in the normal error tail distribution is calculated, and the relative density confidence is evaluated; A position weighting mechanism is introduced, all test samples are sorted in descending order of reconstruction error, and a quadratic position confidence score based on the serial number of each sample is calculated; An adaptive dynamic fusion function is set, the two score results are combined, and the final abnormal confidence score is output through a weighted method.

[0016] Further, the incremental data screening module is used to identify representative potential new general samples from real-time detection results based on a dynamic density trend perception algorithm, and specifically includes the following steps: A scoring set based on a sliding time window is constructed, the abnormal score results are grouped in time sequence, and the scores in each time period are taken as input samples; A time-varying kernel density estimation function is constructed for each scoring interval, an Epanechnikov kernel function is used, and the bandwidth is adaptively adjusted based on the standard deviation and quartile range of the window scores to adapt to the local changes of the score distribution; A density evolution energy function is constructed, and the gradient energy of the density function in the time dimension is calculated to capture the change trend; A hypothesis testing mechanism is introduced, and the significance of each scoring sub-interval is tested to determine whether there is a monotonic increasing trend of abnormal score density in the interval, and the original hypothesis is rejected to confirm the potential drift; A dynamic overlapping grid division strategy is used to divide the scoring space into multiple overlapping sub-intervals, and the trend significance of all sub-intervals is screened; The target interval with significant density growth and reaching a set relative growth threshold is screened out, and the samples in the interval are marked as candidate data and pushed to the artificial review module; After being confirmed as valid by artificial review, the target sample will be selected into the cloud model sub-knowledge base to participate in subsequent continuous learning.

[0017] Further, the continuous learning mechanism of the cloud model evolution module realizes the self-adaptive evolution and update of the model, and specifically includes the following steps: A parameter importance matrix is constructed, including: constructing a teacher model of an old task; using an old training data set, calculating the gradient square expectation value of each parameter position of the current model, and estimating the diagonal elements of the Fisher information matrix; the Fisher information matrix is used to measure the contribution of each parameter in the model to the old task, as the weight basis for subsequent parameter constraints; The teacher model is frozen and the student model is initialized, including: copying the old model as a teacher model and freezing its parameters; initialize the student model training process in a new round, and only allow the student model to update parameters; Constructing a joint loss function and performing training, including: constructing a joint loss function composed of a base reconstruction loss (e.g., the input reconstruction error from a VAE model), an EWC regularization term in the parameter space, and a LWF distillation loss in the feature space; during training, keeping the teacher model unchanged and only updating the student model parameters through error backpropagation, so that it retains the ability to generalize to old tasks while learning new tasks; Dynamically adjust loss weights to balance learning stability and transferability, including: real-time monitoring of the Catastrophic Forgetting Index (CFI); automatically increase the EWC regularization term weight coefficient when the CFI exceeds the set threshold to enhance the model's protection of key parameters; if the CFI is lower than the threshold, appropriately weaken parameter constraints to improve the model's plasticity and adaptability.

[0018] Furthermore, in the joint loss function: The parameter space EWC regularization term imposes a quadratic penalty based on the Fisher matrix on the offset between the student model parameters and the old model parameters; The LwF distillation loss in the feature space calculates the KL divergence between the teacher model output and the student model output, and introduces a temperature coefficient to adjust the distribution smoothness.

[0019] Furthermore, in the generation and replay module: The variational autoencoder (VAE) is used to learn the feature distribution of normal voiceprints and generate pseudo feature samples through latent variable sampling; The generative adversarial network (GAN) is used to learn the true distribution of original voiceprint samples and output pseudo-original data through the generator to enhance sample diversity.

[0020] Furthermore, in the generation and replay module, GAN is introduced as the second generator to make up for the limitations of the single VAE method. The specific implementation process is divided into the following three stages: In the initial stage, a VAE generation model is built based on the initial training set and a second generator GAN model is constructed, where VAE learns the latent space representation of normal voiceprint features and GAN learns the true distribution of historical data; In the incremental learning phase, from the latent space variables of VAE mid-sampling, through the decoder Generate fake samples and use GAN generator Generate high-quality pseudo-original data samples; The above two types of generated samples are mixed with incremental data and input into the detection model for joint training.

[0021] Furthermore, the model updating module also includes abnormal distribution monitoring and data screening functions, and the specific process includes the following steps: When an abnormal data distribution is detected or a periodic model update cycle is entered, a dynamic density trend perception algorithm is used to select representative new samples as incremental data based on the temporal evolution trend of the score distribution and upload them to the cloud sub-knowledge base; When the sample accumulation in the sub-knowledge base reaches the set threshold, the cloud model starts the evolution process. It first trains based on the total knowledge base to enhance the model's global generalization ability, and then performs lightweight adaptive training based on the sub-knowledge base to obtain a new edge detection model. The newly generated edge model is automatically distributed and replaces the original model, thereby completing the update of the edge model and forming a continuous evolution closed loop of cloud-edge collaboration.

[0022] Compared with the prior art, the present invention has the following advantages: 1. This invention provides a system for detecting abnormal voiceprints in equipment. Addressing resource constraints on edge devices in industrial sites, a one-dimensional, lightweight voiceprint detection model based on an improved MobileNetV3 Small architecture is designed. This model achieves efficient temporal voiceprint feature extraction and anomaly detection while maintaining low computational overhead. This model employs a depthwise separable one-dimensional convolutional structure to address the dimensionality mismatch problem encountered when processing voiceprint signals with traditional two-dimensional convolutions. The Mish activation function is introduced to enhance gradient propagation. Furthermore, a channel-spatial dual attention mechanism (CBAM) is incorporated to enhance the model's ability to perceive key voiceprint regions. The encoder utilizes a progressive scaling and improved inverted residual module stacking, while the decoder achieves high-precision reconstruction through four-stage transposed convolutions, ensuring quantifiable detection errors. Experimental results show that this model achieves highly robust detection of both normal and abnormal voiceprints on four typical types of industrial equipment, outperforming unoptimized models in both F1-score and AUC, significantly improving edge detection accuracy and inference efficiency.

[0023] 2、The abnormal voiceprint detection system provided by the application constructs an abnormal credibility evaluation module for probabilistic risk quantification of the detection results, solves the problem that the traditional static threshold method cannot accurately represent the abnormality degree, and converts the reconstruction error of the input signal into a double-layer probability confidence score: the first layer constructs a reference probability distribution of the normal sample error through kernel density estimation, calculates the cumulative probability of the test sample in the tail distribution to reflect the deviation degree, the second layer introduces a position weighting mechanism to evaluate the confidence strength according to the error sorting position, and finally outputs the final abnormal credibility through an adaptive fusion function combining the two. The scoring mechanism not only enhances the sorting stability in the head of the distribution and improves the statistical significance in the tail, but also constructs a risk evaluation framework through the score output. Experimental verification shows that the scoring results provided by the module are more interpretable, effectively supporting the intelligent early warning and decision system.

[0024] 3、The abnormal voiceprint detection system provided by the application sets an incremental data screening module that uses a dynamic density trend perception algorithm to effectively deal with the non-stationary evolution problem of the voiceprint distribution in the operation of industrial equipment. The algorithm constructs a scoring time series based on a sliding window, uses a time-varying kernel density estimation to capture local density changes, and determines whether there is a significant change trend in the scoring interval through a density evolution energy function and a hypothesis testing method. In combination with a dynamic overlapping grid division method, high-value samples with density growth trend are screened out. These samples are sent to the cloud model sub-knowledge base for training after artificial verification. This mechanism makes up for the lack of adaptability of the static sampling strategy to distribution drift, significantly improves the representativeness and quality of the model update samples, and effectively supports the stable evolution of the subsequent model.

[0025] 4、The abnormal voiceprint detection system provided by the application designs a continuous learning mechanism that combines EWC and LwF to solve the problem of non-stationary drift of voiceprint features due to environmental changes and equipment degradation. A joint loss function is constructed through the Fisher information matrix in the parameter space and the knowledge distillation in the feature space to balance the protection of key parameters and the transfer of new tasks. In addition, to alleviate the catastrophic forgetting risk caused by the lack of historical data, the system introduces a generation replay module composed of VAE and GAN to generate feature vectors and pseudo-original data respectively, enhancing the model's expression and memory ability for historical data. Compared with the EWC or LwF strategy alone, this hybrid method performs better in detection performance and knowledge transfer ability in experiments, especially in the early deployment of data scarcity or large differences between new and old tasks, the model can still maintain good recognition accuracy and generalization ability, with high self-adaptive evolution characteristics.

[0026] Based on the above reasons, the application can be widely promoted in the fields of intelligent operation and maintenance and industrial Internet of Things. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments or prior art description. Obviously, the drawings described below are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0028] Figure 1 The abnormal voiceprint detection system architecture diagram for the present application.

[0029] Figure 2 The flow chart of the mixed constraint strategy combining elastic weight solidification and non-forgetting learning for the present application.

[0030] Figure 3 The specific implementation process diagram of the replay module for the present application. DETAILED DESCRIPTION

[0031] In order to make the person skilled in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0032] It should be noted that the terms "first", "second" and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0033] As shown in Figure 1 The present application provides an abnormal voiceprint detection system, comprising: A voiceprint collection module for collecting raw voiceprint signals in real time during the operation of industrial equipment and pre-processing; An edge detection module deployed on an edge computing device for real-time anomaly detection of pre-processed voiceprint features using a one-dimensional lightweight neural network model; An anomaly credibility assessment module, in communication with the edge detection module, for converting the results of real-time anomaly detection into a probabilistic anomaly credibility score; An incremental data screening module, connected to the abnormal credibility assessment module, is used to screen high-value samples from real-time voiceprint data based on a dynamic density trend perception algorithm and cache them at the edge; The cloud-based model evolution module communicates with the edge side and is used to evolve and train the detection model using a continuous learning mechanism; A generation and replay module, nested in the cloud model evolution module, is used to jointly generate pseudo samples through a variational autoencoder (VAE) and a generative adversarial network (GAN) to expand the training data; The model update module is used to compress the evolved cloud model and send it down to replace the original model in the edge detection module, thus forming a continuous evolution closed loop of cloud-edge collaboration.

[0034] In specific implementation, as a preferred embodiment of the present invention, the one-dimensional lightweight neural network model is based on the MobileNetV3 Small architecture, systematically reconstructs the two-dimensional depth-separable convolution into a one-dimensional depth-separable convolution, and embeds the channel-space dual attention mechanism CBAM and Mish activation function in each inverted residual module, as follows: The one-dimensional depth-wise separable convolution is used for the input signal The processing process is expressed as:

[0035] in, Indicates that the depth convolution kernel size is The temporal convolution operation, represents point-wise convolution, Represents residual connection; in this embodiment, by stacking 16 improved inverted residual modules, the model compresses the number of parameters while ensuring the feature extraction capability.

[0036] In the channel dimension, the channel-space dual attention mechanism CBAM is introduced to enhance feature selectivity. The attention weight is generated by dual-path aggregation of global average pooling and maximum pooling. The formula is as follows:

[0037] in, is the feature map, is the Sigmoid function; In the spatial dimension, the spatial attention map is generated by feature compression on the channel axis. The formula is as follows:

[0038] in, is a Sigmoid function; The Mish activation function is used to optimize the gradient propagation path, introduce differentiable saturation characteristics while maintaining linear growth in the positive interval, and the expression is as follows:

[0039] The second derivative of the Mish activation function in the inverted residual module is always positive, effectively alleviating the gradient vanishing problem in the back propagation process.

[0040] The encoder part of the one-dimensional lightweight neural network model is composed of an initial convolution layer, 10 improved inverted residual modules, and an adaptive pooling layer; the expansion ratio of each inverted residual module is set to The progressive structure, which retains more detailed information in the shallow layer and focuses on abstract features in the deep layer, is used; the decoder uses four-level transpose convolution for signal reconstruction, and the mathematical description formula is as follows:

[0041] wherein, is a latent variable, is the hidden state of each decoding stage. In this embodiment, through the design of the symmetric encoding and decoding structure, the model realizes accurate reconstruction of the input signal while maintaining lightweight.

[0042] In summary, the edge detection module realizes efficient feature extraction and anomaly detection of the time sequence voiceprint signal at the edge, builds a one-dimensional lightweight neural network model based on the MobileNetV3 Small architecture, combines the channel-spatial dual attention mechanism (CBAM) and the Mish activation function, solves the dimension mismatch problem when the traditional two-dimensional convolution processes the voiceprint signal, and realizes signal reconstruction and abnormal error calculation through the progressive encoding and decoding structure; the purpose is to complete the accurate identification of voiceprint features with low computational overhead under the resource constraints of edge devices, and through the dynamic data selection mechanism and the cloud-edge collaborative update strategy, the model realizes the adaptive optimization of the data distribution change, reduces the cloud computing load and data transmission cost; the final effect is to greatly compress the parameter amount under the premise of ensuring the detection accuracy of the model, improve the edge inference efficiency through one-dimensional depth separable convolution and lightweight design, enhance the key voiceprint feature expression by using CBAM, and realize the continuous evolution of the model through cloud-edge collaborative closed loop, forming a voiceprint analysis solution that takes into account real-time, accuracy and adaptability. In this embodiment, the edge detection module uses a one-dimensional lightweight neural network model improved based on the MobileNetV3 Small architecture, and the overall structure includes multiple functional levels, including an initial convolution layer, an inverted residual (bneck) module stack, a convolution layer, a pooling layer, and a fully connected layer, and the specific configuration is as follows:

[0043] In a specific implementation, as a preferred embodiment of the present application, the abnormal confidence evaluation module is used to map the reconstruction error of the input voiceprint signal to a probabilistic confidence score, specifically including the following steps: The kernel density estimation method (Kernel Density Estimation) is used to model the probability density of the normal sample reconstruction error, and a statistical confidence function of the abnormal deviation degree is constructed; The reconstruction error of the to-be-tested sample is mapped into the density function, and the cumulative probability of the to-be-tested sample in the tail distribution of the normal error is calculated, so as to evaluate the relative density confidence thereof; Meanwhile, a position weighting mechanism is introduced, all test samples are sorted in descending order of reconstruction error, and a quadratic position confidence score based on the serial number of each sample is calculated; An adaptive dynamic fusion function is set, the two score results are combined, and the final abnormal confidence score is output through a weighted manner.

[0044] The fusion mechanism enhances the position differentiation ability in the front of the error sorting and enhances the statistical significance in the tail of the error, so as to improve the interpretability and stability of the overall confidence score.

[0045] In the embodiment, given an input signal The reconstruction output obtained through the variational autoencoder The original reconstruction error thereof is as follows:

[0046] Wherein, is the feature dimension. The traditional anomaly detection method directly uses the reconstruction error or simple normalization processing, ignores the nonlinear relationship between the relative position and the sorting position of the error and the abnormal degree, and cannot effectively reflect the rich information contained in the error distribution. To solve this problem, the present application constructs a double-layer probability conversion model as follows: The first layer conversion establishes the baseline probability density distribution of the normal sample error through the kernel density estimation (Kernel Density Estimation, KDE), and the formula is as follows:

[0047] Wherein, is the number of normal samples in the training set, is the bandwidth parameter, is the Gaussian kernel function. Thus, the above formula is the relative density confidence of the test sample , and the calculation formula is as follows:

[0048] In the formula, Minimum value, is the mode position of the normal sample error distribution, is the cumulative distribution function. The formula effectively describes the degree of deviation from the normal mode by calculating the cumulative probability of the test sample in the tail of the normal distribution. When deviates from the main body of the normal distribution, tends to 1, showing more significant abnormality.

[0049] The second layer conversion introduces a position weighting mechanism, further considering the nonlinear correlation between error ranking and abnormal detection confidence. After arranging the test samples in descending order of error, the position confidence is calculated as follows:

[0050] Wherein, is the sample ranking number, is the total number of test samples. The position weight of the quadratic function reflects the actual experience rule that abnormal events are usually concentrated in the top of the ranking, and the credibility of the samples at the back quickly decays.

[0051] In order to dynamically balance the contribution of relative density estimation and position information to the credibility, the application designs an adaptive weight function, as shown in the following formula:

[0052] Wherein, is a coefficient for controlling the decay rate. The final abnormal confidence calculation formula is as follows:

[0053] In this embodiment, the above fusion strategy mainly relies on position information to ensure the stability of the ranking at the head of the error distribution, and enhances the statistical significance of density estimation at the tail, so as to improve the interpretability of the credibility score as a whole. The application combines non-parametric estimation with ranking theory, overcomes the threshold sensitivity problem caused by error scale difference in traditional methods through relative density estimation, and realizes effective complementation of statistical distribution and position information through adaptive dynamic weight mechanism. In addition, the probabilistic output form provides a clear risk assessment framework for anomaly detection tasks, effectively quantifying the uncertainty in the decision-making process.

[0054] In specific implementation, as a preferred embodiment of the application, the incremental data screening module is used to identify representative potential new general samples from real-time detection results based on a dynamic density trend perception algorithm, and specifically includes the following steps: A scoring set based on a sliding time window is constructed, the abnormal score results are grouped in time sequence, and the score of each time period is taken as an input sample; For each score interval, a time-varying kernel density estimation function is constructed, with an Epanechnikov kernel function and adaptive bandwidth based on the standard deviation and interquartile range of the window score to adapt to local changes in the score distribution; A density evolution energy function is constructed to capture the trend of change by calculating the gradient energy of the density function in the time dimension; A hypothesis testing mechanism is introduced to perform significance testing on each score sub-interval to determine whether there is a monotonic increasing trend of abnormal score density in the interval, and to reject the null hypothesis to confirm potential drift; A dynamic overlapping grid partitioning strategy is used to divide the score space into multiple overlapping sub-intervals, and trend significance screening is performed on all sub-intervals; The target interval with significant density growth and reaching the set relative growth threshold is screened out, and the samples in the interval are marked as candidate data and pushed to the artificial review module; After artificial confirmation of effectiveness, the target sample will be selected into the cloud model sub-knowledge base to participate in subsequent continuous learning.

[0055] In this embodiment, the current time is , the historical time window is , where is the observation period length. Define as the abnormal score of sample , the lower the score, the closer to the normal state. A sliding window is used to represent a score set containing consecutive time slots, and each time slot contains sample scores. To quantify the density evolution in a specific score interval, a time-varying kernel density estimation function is constructed:

[0056] where is the Epanechnikov kernel function, and the adaptive bandwidth is calculated as follows:

[0057] In the formula, is the standard deviation of the scores in the window , and is the interquartile range of the scores. This bandwidth selection strategy effectively adapts to the dispersion of the score distribution while ensuring estimation accuracy.

[0058] To further describe the evolution intensity of the target score interval , a density evolution energy function is defined:

[0059] The function captures the local change intensity of the score distribution by calculating the gradient energy of the density function in the time domain.

[0060] To identify the significant evolution region, a hypothesis testing framework is introduced, and the null hypothesis is set as The density trend in the interval is not significant, and the alternative hypothesis is that there is a monotonic increasing trend. The corresponding test statistic is defined as follows:

[0061] where the cumulative density The corresponding variance estimate is When , the null hypothesis is rejected , confirming that the interval has a significant density growth.

[0062] In specific implementation, based on the abnormal confidence algorithm mentioned above, the score space [0, 100] is dynamically divided into overlapping intervals , where each interval has a width of , and the adjacent intervals overlap by 50%. For each , the density trend significance level is calculated, and intervals that meet the conditions and are selected as candidate regions. Finally, the target region is determined as:

[0063] where is the relative growth threshold, and samples within will enter the artificial verification queue and be selected into the training set after artificial review.

[0064] In summary, in this embodiment, by using the time-varying kernel density estimation method, the overfitting problem of the traditional fixed bandwidth kernel smoothing method is solved. A trend identification mechanism based on hypothesis testing is constructed, and the statistical significant selection of incremental data is realized. The dynamic overlapping grid scanning method is adopted, which enhances the sensitivity of the algorithm to changes in any distribution form.

[0065] In specific implementation, as a preferred embodiment of the present application, the continuous learning mechanism of the cloud model evolution module realizes the self-adaptive evolution and update of the model, which specifically includes the following steps: The parameter importance matrix is constructed, including: constructing a teacher model of an old task; calculating the gradient square expectation value of the current model at each parameter position by using an old training data set, and estimating the diagonal elements of the Fisher information matrix; the Fisher information matrix is used to measure the contribution of each parameter in the model to the old task, as the weight basis for subsequent parameter constraints; The teacher model is frozen and the student model is initialized, including: copying the old model as the teacher model and freezing its parameters; initializing a new round of student model training process, and only allowing the student model to update parameters; The joint loss function is constructed and training is performed, including: constructing a joint loss function composed of a basic reconstruction loss (such as an input reconstruction error from a VAE model), an EWC regularization term in the parameter space, and an LwF distillation loss in the feature space; during the training process, the teacher model is kept unchanged, and only the student model parameters are updated through error back propagation, so that the student model can learn new tasks while retaining the generalization ability for old tasks; The loss weight is dynamically adjusted to balance learning stability and migration ability, including: real-time monitoring of the catastrophic forgetting index (CFI); when the catastrophic forgetting index (CFI) exceeds the set threshold, the EWC regularization term weight coefficient is automatically increased to enhance the protection of key parameters of the model; if the catastrophic forgetting index (CFI) is lower than the threshold, the parameter constraint is appropriately weakened to improve the plasticity and adaptability of the model.

[0066] In the embodiment, specifically, in the cloud model evolution module self-adaptive evolution method based on teacher-student collaborative training, the design of the continuous learning mechanism needs to consider parameter stability and knowledge transfer efficiency. The present application proposes a hybrid constraint strategy combining elastic weight consolidation (EWC) and learning without forgetting (LwF), as shown in Figure 2 As shown in the figure, through the dual mechanisms of parameter space regularization and feature space knowledge distillation, the progressive optimization of the voiceprint feature extraction model is realized; The core idea of the cloud model evolution module self-adaptive evolution method based on teacher-student collaborative training is to establish a joint optimization goal of parameter importance measurement and feature distribution preservation, which can be mathematically expressed as:

[0067] Among them, is the reconstruction loss of the variational autoencoder, and These are the elastic weight curing loss and the forgetting-free learning loss. The innovation of this hybrid constraint strategy lies in the fact that EWC selectively curates network parameters using the Fisher information matrix to address catastrophic forgetting, while LwF maintains consistency in the feature space through knowledge distillation between the teacher and student models. The synergy between these two allows the model to preserve key parameters of the old task while maintaining the continuity of feature discriminative power when adapting to new voiceprint features.

[0068] In the mathematical implementation of EWC, the parameter importance is quantified by the diagonal elements of the Fisher information matrix. parameters , the Fisher information calculation formula is:

[0069] This expected value is obtained from the old training data set. In the continuous learning phase, the EWC regularization term imposes a quadratic penalty on the parameter offset, as shown in the following formula:

[0070] The LwF module realizes the progressive migration of feature space through knowledge distillation technology.

[0071] Set up teacher model The voiceprint reconstruction output is: , student model The corresponding output is: ,in is the temperature coefficient. The knowledge distillation loss is measured using the KL divergence as shown in the following formula:

[0072] Among them, the temperature coefficient The regulatory effect is reflected in When , the probability distribution tends to be smooth, which helps to capture the characteristic distribution pattern of the teacher model; when , it degenerates into the original classification task.

[0073] The training mechanism and execution process of the cloud model evolution module are as follows: The algorithm uses the EWC-LwF continuous learning training process, including: Input is old model and the new training set ; Initialize the manager: ; Calculate the Fisher information matrix: ; Clone the teacher model and initialize the new model; The training cycle includes: calculating the reconstruction loss: ; Calculate the EWC regularization term: ; Calculate LwF distillation loss: ; Total loss back propagation: ; Update parameters: ; Output updated model .

[0074] Compared to methods using either EWC or LwF alone, this hybrid strategy demonstrates significant advantages in voiceprint feature evolution scenarios. When using EWC alone, the model easily falls into local optimality and struggles to adapt to non-stationary drifts in the distribution of voiceprint features. When there are significant differences between new and old voiceprint samples, applying LwF alone can lead to parameter drift. Furthermore, EWC constraints protect the key network parameters corresponding to common voiceprint features, while LwF ensures that individual-specific features can be gradually transferred through knowledge distillation.

[0075] To further improve the stability and adaptability of the model during continuous learning, this paper designs a phased optimization parameter update strategy based on the EWC-LwF joint training mechanism to dynamically adjust the parameter constraint strength. The strategy is as follows: During the initial training phase, the model primarily relies on the LwF loss to guide knowledge transfer. This involves maintaining consistency in the feature space distribution through feature distillation between the teacher and student models, enabling rapid adaptation to new voiceprint tasks. During this phase, the EWC regularization term is weighted relatively low to mitigate constraints on parameter updates and improve the efficiency of new knowledge absorption.

[0076] As training progresses, the system gradually increases the strength of the EWC regularization term to enhance the stability of key parameters of the old tasks. The focus of this stage shifts to preventing the model from forgetting knowledge of old tasks.

[0077] To achieve the above dynamic switching, the system introduces the Catastrophic Forgetting Index (CFI) as an adjustment basis:

[0078] The CFI metric is used to assess the degree of performance degradation of the model on the old task validation set, and its expression is shown in Equation X. During training, the system periodically evaluates the CFI value. When the value exceeds the set threshold, the EWC weight is automatically increased to strengthen the parameter fixed constraint. When the CFI remains low, the constraint is appropriately weakened to ensure the model has good transferability.

[0079] In this embodiment, through the above steps, the cloud model evolution module realizes the double protection mechanism in the knowledge transfer process of new and old tasks, so that the model has the ability of continuous learning, and can maintain the detection accuracy and stability in the scene where the voiceprint feature distribution drifts.

[0080] In specific implementation, as a preferred embodiment of the present application, the parameter space EWC regularization term imposes a quadratic penalty based on the Fisher matrix on the parameter offset of the student model and the old model parameter; The LwF distillation loss of the feature space calculates the KL divergence of the teacher model output and the student model output, and introduces a temperature coefficient to adjust the distribution smoothness.

[0081] In specific implementation, as a preferred embodiment of the present application, in the generation replay module: The variational autoencoder VAE is used to learn the feature distribution of normal voiceprints, and pseudo feature samples are generated through hidden variable sampling; The generative adversarial network GAN is used to learn the real distribution of original voiceprint samples, and pseudo original data is output by the generator to enhance sample diversity.

[0082] In this embodiment, compared with the traditional parameter constraint (EWC) and knowledge distillation (LwF) strategy, the generation replay mechanism is more suitable for the following scenarios: 1) the original training samples are scarce when the system is initially deployed; 2) the feature space difference between new and old tasks is significant, resulting in distillation failure; 2) the feature space difference between new and old tasks is significant, resulting in distillation failure; In specific implementation, as a preferred embodiment of the present application, in the generation replay module, the introduction of GAN as a second generator compensates for the limitations of the single VAE method, and the specific implementation process is as shown in Figure 3 The technical route can be divided into the following three stages: In the initial stage, the VAE generation model is constructed based on the initial training set, and the second generator GAN model is constructed, wherein the VAE learns the latent space representation of normal voiceprint features, and the GAN learns the real distribution of historical data; In the incremental learning stage, the hidden space variables are sampled from the VAE, pseudo samples are generated through the decoder , and high-quality pseudo original data samples are generated by the GAN generator . The above two types of generated samples are mixed with the incremental data and input into the detection model for joint training.

[0083] In this embodiment, the mathematical essence of this process is to maintain the approximate expression of the historical data distribution under the Bayesian framework:

[0084] where, VAE decoder parameters are frozen after initial training, ensuring that generated samples always reflect the statistical properties of historical data. GAN generator parameters are also kept unchanged, ensuring high authenticity of pseudo data. The model dynamically generates pseudo samples aligned with the distribution of historical data through a dual generation mechanism, where VAE achieves feature reconstruction through normal distribution sampling in latent space, and GAN generates high-quality pseudo raw data. The final output dimensions are feature reconstruction data set and raw data replay data set .

[0085] In summary, compared with a single generation model, the dual generation replay method can more comprehensively reproduce the diversity and authenticity of historical data, effectively covering the low-density area of the feature space, thereby enhancing the robustness of the model to edge cases. By dynamically adjusting the mixing ratio of generated data and incremental data, the absorption intensity of new and old knowledge can be flexibly controlled. This mechanism is particularly suitable for industrial device voiceprint detection scenarios, effectively avoiding model overfitting caused by direct use of EWC or LwF.

[0086] In specific implementation, as a preferred embodiment of the present application, the model updating module further comprises an abnormal distribution monitoring and data screening function, and the specific process comprises the following steps: When data distribution anomalies are detected or a periodic model updating cycle is entered, a dynamic density trend perception algorithm is used to select representative new samples as incremental data based on the time evolution trend of score distribution, and upload them to the cloud sub-knowledge base; When the sample accumulation in the sub-knowledge base reaches a set threshold, the cloud model starts the evolution process, first trains the global generalization ability of the enhanced model based on the total knowledge base, and then performs lightweight adaptation training based on the sub-knowledge base to obtain a new edge detection model; The newly generated edge model is automatically distributed and replaces the original model, thereby completing the update of the edge model and forming a continuous evolution closed loop of cloud-edge collaboration.

[0087] Example 1 To verify the effectiveness and practicability of the proposed equipment abnormal voiceprint detection method and system, a series of embodiments are designed, and the detection accuracy, continuous learning ability, model updating performance and other key dimensions of the system are quantitatively analyzed and visually compared by combining public data sets and self-constructed evaluation indicators. The embodiments use the MIMII (Malfunctioning Industrial Machine Investigation and Inspection) data set for training and testing. This data set is widely used in industrial voiceprint anomaly detection tasks and covers multiple types of equipment and multiple abnormal situations, which can truly reflect the changes in voiceprint features during equipment operation.

[0088] The experimental platform selects Microsoft Visual Studio Code as the main development tool, the backend uses the PyTorch 1.8.2 deep learning framework, the main programming language is Python, and the running environment is Windows system. The specific experimental hardware and software configurations are shown in the following table.

[0089]

[0090] Through the above environment and data configuration, this paper carries out system experiments from the aspects of model detection performance, continuous learning robustness, credibility score interpretability, etc. The experimental results and analysis of the voiceprint detection model self-adaptive evolution method are as follows:

[0091] According to the above table, the detection performance of the teacher model and the EWC+LwF continuous learning model on the "normal" and "abnormal" voiceprint samples of four types of equipment (pump, fan, slider, valve) in the MIMII data set is compared. This embodiment uses F1-score and AUC value as evaluation indicators to compare the accuracy and robustness of the two models under each detection mode. The teacher model represents the basic training model without continuous learning adjustment; the EWC+LwF model uses the elastic weight solidification and no forgetting distillation joint strategy proposed in this invention to realize model self-adaptive evolution. This embodiment aims to verify whether the continuous learning mechanism can effectively improve the recognition ability of the model for abnormal voiceprint features and enhance its generalization performance in complex voiceprint environments.

[0092] In this scheme, the functional modules form a closed-loop deployment architecture in the order of (edge collection-edge detection-confidence evaluation-data screening-cloud evolution-model update): the voiceprint collection module is deployed on the industrial equipment body or its near-end data collection unit, used to acquire voiceprint signals in real time and perform preprocessing; the edge detection module and the abnormal confidence evaluation module are deployed on the edge computing device (such as an industrial gateway or an edge AI terminal), wherein the lightweight detection model is constructed based on the MobileNetV3 Small structure, adapted to a low-power environment, realizing local rapid judgment and probabilistic risk scoring; the incremental data screening module is located at the edge, intelligently selects representative new samples by monitoring the evolution trend of scoring density; these high-value samples are uploaded to the cloud sub-knowledge base through a secure interface, received by the cloud model evolution module deployed on the cloud server or data center and triggering a continuous learning process, using the EWC and LwF strategy combination to generate a replay mechanism to complete model update training; the newly generated edge model is distributed to the edge device through the model update module, replacing the old version, completing the adaptive iterative deployment of cloud-edge collaboration. The overall system realizes efficient, stable, and evolvable industrial voiceprint anomaly detection function without relying on original data storage, and adapts to data distribution changes in complex industrial sites.

[0093] Embodiment 2 In actual industrial applications, a certain manufacturing workshop deploys the system of the present scheme on a key device. When the device is running, the voiceprint collection module collects voiceprint signals in real time during the operation of the device, and completes denoising and standardization processing; these signals are transmitted to the edge detection module, which is quickly identified by the lightweight neural network model, and if a suspected anomaly or a score deviating from the historical distribution is detected, the abnormal confidence evaluation module will output a risk value with probability for on-site personnel or system to judge; at the same time, the incremental data screening module dynamically captures the change of scoring density through trend perception algorithm, and selects representative general data samples to upload to the cloud; when the cloud sub-knowledge base sample accumulation reaches the threshold, the cloud model evolution module automatically starts continuous learning, optimizes the model combined with historical knowledge and new data, generates a replay module to provide historical pseudo data to prevent forgetting during the period, and the updated model is distributed to the edge through the model update module, replacing the original model, realizing accurate monitoring and adaptive optimization of device voiceprint anomalies.

[0094] In the above embodiments of the present application, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0095] In several embodiments provided in the present application, it should be understood that the disclosed technical contents can be implemented by other manners. Among them, the above-described device embodiments are only illustrative, for example, the division of the units can be a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units or modules shown or discussed can be indirect coupling or communication connection through some interfaces, units or modules, which can be electrical or other forms.

[0096] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place or distributed on multiple units. Part or all of the units can be selected to achieve the purpose of the embodiment scheme according to actual needs.

[0097] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0098] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0099] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An abnormal voiceprint detection system for equipment, characterized in that: include: Voiceprint collection module, used to collect original voiceprint signals in real time and perform preprocessing during the operation of industrial equipment; The edge detection module is deployed on edge computing devices and is used to perform real-time anomaly detection on pre-processed voiceprint features using a one-dimensional lightweight neural network model; An anomaly credibility assessment module, in communication with the edge detection module, for converting the results of real-time anomaly detection into a probabilistic anomaly credibility score; An incremental data screening module, connected to the abnormal credibility assessment module, is used to screen high-value samples from real-time voiceprint data based on a dynamic density trend perception algorithm and cache them at the edge; The cloud-based model evolution module communicates with the edge side and is used to evolve and train the detection model using a continuous learning mechanism; A generation and replay module, nested in the cloud model evolution module, is used to jointly generate pseudo samples through a variational autoencoder (VAE) and a generative adversarial network (GAN) to expand training data; The model update module is used to compress the evolved cloud model and send it down to replace the original model in the edge detection module, thus forming a continuous evolution closed loop of cloud-edge collaboration.

2. The abnormal voiceprint detection system for equipment according to claim 1, characterized in that: The one-dimensional lightweight neural network model is based on the MobileNetV3 Small architecture. It systematically reconstructs the two-dimensional depth-wise separable convolution into a one-dimensional depth-wise separable convolution and embeds the channel-spatial dual attention mechanism (CBAM) and the Mish activation function in each inverted residual module, as follows: The one-dimensional depth-wise separable convolution is used for the input signal The processing process is expressed as: in, Indicates that the depth convolution kernel size is The temporal convolution operation, represents point-wise convolution, represents residual connection; In the channel dimension, the channel-space dual attention mechanism CBAM is introduced to enhance feature selectivity. The attention weight is generated by dual-path aggregation of global average pooling and maximum pooling. The formula is as follows: in, is the feature map, is the Sigmoid function; In the spatial dimension, the spatial attention map is generated by feature compression on the channel axis. The formula is as follows: in, is the Sigmoid function; The Mish activation function is used to optimize the gradient propagation path, while maintaining linear growth in the positive interval and introducing differentiable saturation characteristics. The expression is as follows: Among them, the second-order derivative of the Mish activation function in the inverted residual module is always positive, which alleviates the gradient disappearance problem during the back propagation process. The encoder part of the one-dimensional lightweight neural network model consists of an initial convolutional layer, 10 improved inverted residual modules and an adaptive pooling layer; the expansion ratio of each inverted residual module is set to The progressive structure retains more detailed information in the shallow layer and focuses on abstract features in the deep layer. The decoder uses four-level transposed convolution for signal reconstruction, and its mathematical description formula is as follows: in, is a latent variable, is the hidden state at each decoding stage.

3. The abnormal voiceprint detection system for equipment according to claim 1, characterized in that: The abnormal credibility assessment module is used to map the reconstruction error of the input voiceprint signal into a probabilistic confidence score, specifically including the following steps: The kernel density estimation method is used to model the probability density of the reconstruction error of normal samples and construct a statistical confidence function of the degree of abnormal deviation; Map the reconstruction error of the sample to be tested into the density function, calculate its cumulative probability in the normal error tail distribution, and thus evaluate its relative density confidence; At the same time, a position weighting mechanism is introduced to sort all test samples in descending order of reconstruction error, and calculate the secondary position confidence score of each sample based on the sorting sequence number; An adaptive dynamic fusion function is set to combine the two scoring results and output the final anomaly credibility score in a weighted manner.

4. The abnormal voiceprint detection system for equipment according to claim 1, characterized in that: The incremental data screening module is used to identify representative potential new general samples from real-time detection results based on a dynamic density trend perception algorithm, specifically including the following steps: Construct a scoring set based on a sliding time window, group the abnormal scoring results in chronological order, and use the scores of each time period as input samples; A time-varying kernel density estimation function is constructed for each score interval. The Epanechnikov kernel function is used and the bandwidth is adaptively adjusted based on the standard deviation and interquartile range of the window score to adapt to local changes in the score distribution. Construct a density evolution energy function and capture the change trend by calculating the gradient energy of the density function in the time dimension; A hypothesis testing mechanism is introduced to perform a significance test on each score sub-interval to determine whether there is a monotonic growth trend of abnormal score density in the interval, and the null hypothesis is rejected accordingly to confirm potential drift; The scoring space is divided into multiple overlapping subintervals using a dynamic overlapping grid partitioning strategy, and all subintervals are screened for trend significance. Filter out target intervals with significant density growth and reaching the set relative growth threshold, mark the samples in the interval as candidate data, and push them to the manual review module; After manual confirmation of validity, the target sample will be selected into the cloud-based model sub-knowledge base to participate in subsequent continuous learning.

5. The abnormal voiceprint detection system for equipment according to claim 1, characterized in that: The continuous learning mechanism of the cloud-based model evolution module implements the adaptive evolution and update of the model, which specifically includes the following steps: Constructing a parameter importance matrix involves: building a teacher model for the old task; using the old training dataset, calculating the expected squared gradient of the current model at each parameter position and estimating the diagonal elements of the Fisher information matrix; using the Fisher information matrix to measure the contribution of each parameter in the model to the old task, which serves as the weight basis for subsequent parameter constraints; Freeze the teacher model and initialize the student model, including: copying the old model as the teacher model and freezing its parameters; initializing a new round of student model training process, allowing only the student model to update parameters; Constructing a joint loss function and performing training, including: constructing a joint loss function consisting of a base reconstruction loss, an EWC regularization term in the parameter space, and a LWF distillation loss in the feature space; during training, keeping the teacher model unchanged and only updating the student model parameters through error backpropagation, so that it retains the ability to generalize to old tasks while learning new tasks; Dynamically adjust the loss weight to balance learning stability and transfer ability, including: real-time monitoring of the catastrophic forgetting index; when the catastrophic forgetting index exceeds the set threshold, automatically increase the EWC regularization term weight coefficient to enhance the model's protection of key parameters; if the catastrophic forgetting index is lower than the threshold, weaken the parameter constraints to improve the model's plasticity and adaptability.

6. The abnormal voiceprint detection system for equipment according to claim 5, characterized in that: In the joint loss function: The parameter space EWC regularization term imposes a quadratic penalty based on the Fisher matrix on the offset between the student model parameters and the old model parameters; The LwF distillation loss in the feature space calculates the KL divergence between the teacher model output and the student model output, and introduces a temperature coefficient to adjust the distribution smoothness.

7. The abnormal voiceprint detection system for equipment according to claim 1, characterized in that: In the generation and replay module: The variational autoencoder (VAE) is used to learn the feature distribution of normal voiceprints and generate pseudo feature samples through latent variable sampling; The generative adversarial network (GAN) is used to learn the true distribution of original voiceprint samples and output pseudo-original data through the generator to enhance sample diversity.

8. The abnormal voiceprint detection system for equipment according to claim 7, characterized in that: In the generative replay module, the generative adversarial network (GAN) is introduced as the second generator to overcome the limitations of the single variational autoencoder (VAE) method. The specific implementation process is divided into the following three stages: In the initial stage, a VAE generation model is built based on the initial training set and a second generator GAN model is constructed, where VAE learns the latent space representation of normal voiceprint features and GAN learns the true distribution of historical data; In the incremental learning phase, from the latent space variables of VAE mid-sampling, through the decoder Generate fake samples and use GAN generator Generate high-quality pseudo-original data samples; The above two types of generated samples are mixed with incremental data and input into the detection model for joint training.

9. The abnormal voiceprint detection system for equipment according to claim 1, characterized in that: The model update module also includes abnormal distribution monitoring and data screening functions, and the specific process includes the following steps: When an abnormal data distribution is detected or a periodic model update cycle is entered, a dynamic density trend perception algorithm is used to select representative new samples as incremental data based on the temporal evolution trend of the score distribution and upload them to the cloud sub-knowledge base; When the sample accumulation in the sub-knowledge base reaches the set threshold, the cloud model starts the evolution process. It first trains based on the total knowledge base to enhance the model's global generalization ability, and then performs lightweight adaptive training based on the sub-knowledge base to obtain a new edge detection model. The newly generated edge model is automatically distributed and replaces the original model, thereby completing the update of the edge model and forming a continuous evolution closed loop of cloud-edge collaboration.

Images