Industrial equipment and progressive distillation unsupervised anomaly detection method, device and medium thereof

Abstract

Industrial equipment and its progressive distillation unsupervised anomaly detection method, device and medium, relate to the technical field of equipment anomaly detection. The method comprises: acquiring monitoring time series data and preprocessing. The time neighbor set of each sample is calculated in the time dimension. The spatial similarity is measured in the spatial dimension, and the BallTree index structure is constructed through the hierarchical spatial division strategy, and the spatial neighbor set is searched. The current sample is normalized by the spatiotemporal weighted fusion strategy. After normalization, the adaptive threshold unsupervised anomaly detection teacher model is input for training. The student network is trained in stages by using the progressive distillation strategy, the teacher model is used to guide the student model learning, and the distillation student model is obtained. The data to be detected is input into the anomaly detection model, the reconstruction error between the input sample and the reconstructed sample is calculated, and the adaptive threshold judgment mechanism is combined to obtain the anomaly detection result.

Description

Technical Field

[0001] This invention relates to the field of equipment anomaly detection technology, and more specifically, to an industrial equipment and a method, apparatus, and medium for unsupervised anomaly detection of progressive distillation. Background Technology

[0002] Industrial equipment plays a crucial role in industrial production processes. However, due to long-term operation and the influence of complex environments, equipment performance gradually declines, leading to malfunctions and downtime, thereby threatening enterprise production efficiency and product quality. Therefore, there is an urgent need in the industrial sector for efficient anomaly detection technologies to achieve real-time equipment status monitoring, fault prediction, and preventative maintenance, ensuring the safe and stable operation of industrial systems and creating significant economic benefits for enterprises.

[0003] Currently, industrial equipment management primarily relies on periodic maintenance and manual diagnostics. However, these methods have limitations, resulting in insufficient fault prediction capabilities and an inability to monitor equipment operating status in real time. Furthermore, existing anomaly detection technologies are widely used in equipment management, mainly including purely supervised anomaly detection models, purely unsupervised anomaly detection models, and hybrid anomaly detection models. Supervised models use machine learning or deep learning models for anomaly detection, achieving high accuracy. Unsupervised models do not rely on labeled data and can identify unknown anomalies. Hybrid methods attempt to combine the advantages of both supervised and unsupervised methods to improve detection capabilities.

[0004] However, these existing anomaly detection methods have significant drawbacks. Supervised models require large amounts of labeled data and expert experience, and can only detect known anomalies, making them ill-suited for handling new or unknown anomalies. Unsupervised models, while not relying on labeled data, have lower detection accuracy and are less capable of identifying known anomalies than supervised methods. Hybrid methods often fail to effectively detect both known and unknown anomalies simultaneously in multi-condition environments, and their high computational cost makes them unsuitable for the real-time requirements of industrial systems, resulting in insufficient adaptability in practical applications. Summary of the Invention

[0005] The present invention provides an industrial equipment and a method, apparatus and medium for unsupervised anomaly detection in progressive distillation, to improve at least one of the above-mentioned technical problems.

[0006] In a first aspect, the present invention provides an industrial equipment and a method for detecting unsupervised anomalies in progressive distillation, comprising steps S1 to S7.

[0007] S1. Acquire monitoring time series data of multi-dimensional sensors of industrial equipment under various operating conditions and perform preprocessing.

[0008] S2. Based on the continuity characteristics of time series, calculate the time nearest neighbor set of each sample in the time dimension.

[0009] S3. Perform spatial similarity measurement in the spatial dimension, and construct a BallTree index structure through a hierarchical spatial partitioning strategy, and then search to obtain the set of spatial nearest neighbors.

[0010] S4. Based on the time nearest neighbor set and the spatial nearest neighbor set, perform spatiotemporal joint normalization processing on the current sample using a spatiotemporal weighted fusion strategy.

[0011] S5. Input the normalized feature samples into the adaptive threshold unsupervised anomaly detection teacher model based on GAT-DiffusionVAE for training.

[0012] S6. A progressive distillation strategy is adopted to train the student network in stages. The intermediate denoised distribution of the teacher model output is used to guide the student model learning. Knowledge transfer and feature alignment are achieved by calculating the Wasserstein distance between the output distribution of the student network and the teacher network, and a lightweight distilled student model as an anomaly detection model is obtained.

[0013] S7. Input the preprocessed data to be detected into the anomaly detection model, calculate the reconstruction error between the input sample and the reconstructed sample, and obtain the anomaly detection result by combining the adaptive threshold determination mechanism.

[0014] Secondly, the present invention provides an unsupervised anomaly detection device for progressive distillation in industrial equipment, which includes a training data acquisition module, a time dimension module, a spatial dimension module, a normalization module, a teacher model module, a student model module, and a detection module.

[0015] The training data acquisition module is used to acquire and preprocess the monitoring time series data of multi-dimensional sensors of industrial equipment under various operating conditions.

[0016] The time dimension module is used to calculate the time nearest neighbor set of each sample in the time dimension based on the continuous characteristics of the time series.

[0017] The spatial dimension module is used to measure spatial similarity in the spatial dimension and to build a BallTree index structure through a hierarchical spatial partitioning strategy, and then search to obtain the set of spatial nearest neighbors.

[0018] The normalization module is used to perform spatiotemporal joint normalization processing on the current sample based on the temporal nearest neighbor set and the spatial nearest neighbor set through a spatiotemporal weighted fusion strategy.

[0019] The teacher model module is used to input normalized feature samples into the adaptive threshold unsupervised anomaly detection teacher model based on GAT-DiffusionVAE for training.

[0020] The student model module is used to train the student network in stages using a progressive distillation strategy. It uses the intermediate denoised distribution of the teacher model output to guide the student model's learning. Knowledge transfer and feature alignment are achieved by calculating the Wasserstein distance between the student network and the teacher network output distribution, resulting in a lightweight distilled student model as an anomaly detection model.

[0021] The detection module is used to input the preprocessed data to be detected into the anomaly detection model, calculate the reconstruction error between the input sample and the reconstructed sample, and obtain the anomaly detection result by combining the adaptive threshold judgment mechanism.

[0022] Thirdly, the present invention provides an industrial device comprising a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement an unsupervised anomaly detection method for progressive distillation in an industrial device as described in any paragraph of the first aspect.

[0023] Fourthly, the present invention provides a computer-readable storage medium. The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform a progressive distillation unsupervised anomaly detection method for industrial equipment as described in any paragraph of the first aspect.

[0024] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0025] This invention combines multiple advanced time-series data analysis techniques, such as dynamic spatiotemporal feature modeling, graph attention network GAT algorithm, diffusion variational autoencoder algorithm, and progressive knowledge distillation, to mine the spatiotemporal dependencies between sensors in industrial equipment. At the same time, it adaptively constructs a dynamic nearest neighbor structure to achieve a unified representation of multi-condition data, thus making up for the shortcomings of existing methods in adaptability to multi-condition scenarios. Attached Figure Description

[0026] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the specific embodiments of the present invention will be briefly introduced below. It should be understood that the following drawings only show some specific embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is a flowchart illustrating the unsupervised anomaly detection method for progressive distillation.

[0028] Figure 2 This is a flowchart of the teacher model training and testing process.

[0029] Figure 3 This is a flowchart of the unsupervised anomaly detection method for progressive distillation. Detailed Implementation

[0030] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention.

[0031] Example 1, please refer to Figures 1 to 3 The first embodiment of the present invention provides a method for unsupervised anomaly detection of progressive distillation in industrial equipment, which can be implemented by an unsupervised anomaly detection device for progressive distillation in multi-condition industrial equipment (hereinafter referred to as an anomaly detection device), and in particular, executed by one or more processors within the anomaly detection device.

[0032] In this embodiment, the anomaly detection device may be an electronic device equipped with a processor. The processor carries a computer program for the progressive distillation unsupervised anomaly detection method for multi-condition industrial equipment, and the computer program can be executed. For example, a computer, smartphone, smart tablet, workstation, etc. are not limited here.

[0033] like Figures 1 to 3 As shown, the progressive distillation unsupervised anomaly detection method includes steps S1 to S7.

[0034] S1. Acquire monitoring time series data of multi-dimensional sensors of industrial equipment under various working conditions, and preprocess them to construct a sample dataset for training anomaly detection models.

[0035] Preprocessing includes data cleaning and outlier removal, global preliminary normalization, multi-condition identification and partitioning, time window slicing, and feature smoothing and noise suppression.

[0036] Data cleaning and outlier removal involve denoising, missing value imputation, and invalid sample removal of the acquired multi-dimensional sensor monitoring time series data of industrial equipment. Three methods are employed. The principle is to eliminate outliers.

[0037] The initial global normalization process involves Z-score standardization of sensor signals in different dimensions.

[0038] Multi-condition identification partitioning is based on equipment operation logs, control parameters, or environmental labels to divide historical data into stable operating condition intervals and transitional operating condition intervals.

[0039] Time window slicing is a process of dividing time series samples into segments using a sliding time window of adaptive length. Each window sample contains multidimensional features and their temporal order.

[0040] Feature smoothing and noise suppression are achieved by using Savitzky-Golay smoothing to suppress noise within the window.

[0041] Specifically, historical time-series monitoring data from multi-dimensional sensors of industrial equipment is acquired. Data cleaning and outlier removal are performed, employing the 3σ principle to eliminate obvious anomalies and mitigate the impact of sensor jitter or communication errors. Subsequently, global preliminary normalization is conducted, using Z-score standardization to map feature values ​​across dimensions to a unified numerical range, eliminating feature bias caused by inconsistencies in units. Based on this, multi-condition partitioning is performed according to equipment operation logs, control parameters, or environmental labels, forming stable and transitional operating condition intervals to ensure relative consistency in data distribution within each condition. Next, the sequence is sliced ​​using an adaptive-length sliding time window to construct sample segments containing multi-dimensional features and their temporal order, effectively characterizing the temporal evolution of equipment status. Finally, within the window, Savitzky-Golay smoothing is used for feature smoothing and high-frequency noise suppression, preserving the main trend information of the equipment's operating signals.

[0042] S2. Based on the continuous characteristics of time series, calculate the nearest neighbor set of each sample in the time dimension to obtain the temporal nearest neighbor set, so as to characterize the dynamic change relationship of samples on the time axis and capture time-dependent features. Preferably, S2 includes specific steps S21 to S26.

[0043] S21. Model the preprocessed multidimensional time series data of industrial equipment to construct time features.

[0044] .

[0045] In the formula It is a feature set. For time step The sample. This represents the length of the time series (i.e., the total number of samples). Represents a real number. Indicates the time step as Samples at that time. This represents the number of sensors.

[0046] The feature set represents the operating status of the equipment over a continuous time period, forming a time-series sample set under multiple operating conditions, reflecting the continuous operating status of the equipment under multiple operating conditions.

[0047] S22. Based on the continuity characteristic of time series, on the time axis for each... Establish a time neighborhood window And select candidate samples from within the window. This represents the half-width or time offset of the time neighborhood window.

[0048] S23. By calculating the similarity between samples over time, the dynamic relationship between adjacent time points is characterized. Weighted Euclidean distance is used to calculate the temporal similarity between the current sample and candidate samples. And select the nearest one according to the distance from smallest to largest. Each sample constitutes a time nearest neighbor set. .

[0049] .

[0050] .

[0051] In the formula For the sample Its time neighbor Temporal similarity between them (i.e., weighted Euclidean distance). This represents the time offset step. This is the sensor's serial number. This represents the number of sensors. These are the weighting coefficients for different channels. For time step The The monitoring values ​​of each sensor. For time step The The monitoring values ​​of each sensor. For the most recent One sample.

[0052] S24. Calculate the weighted mean of the samples in the nearest neighbor set at each time step. With variance This is used to characterize the dynamic changes and fluctuation range of a sample over time. The standard deviation is denoted as . For the first The standard deviation of each sample.

[0053] .

[0054] .

[0055] In the formula This represents the number of samples in the time nearest neighbor set. The first in the nearest neighbor set A close neighbor.

[0056] S25. Integrate the statistical characteristics of the current sample with those of the time neighborhood (i.e., weighted mean). With variance ), to obtain time-dependent feature representation .

[0057] .

[0058] In the formula This is a time feature mapping function used to fuse the original samples with their temporal neighborhood features, thereby enhancing the model's ability to represent temporal continuity and dynamic pattern changes.

[0059] S26. Set the time nearest neighbors of each sample and... and This serves as the input basis for subsequent spatial nearest neighbor search and dynamic spatiotemporal joint standardization.

[0060] Specifically, the temporal nearest neighbor set of each sample at each time step... Its corresponding weighted statistical features Serving as the input foundation for spatial nearest neighbor search and dynamic spatiotemporal joint standardization in subsequent steps, it is used to achieve comprehensive modeling of temporal and spatial dependencies in multi-condition scenarios. It enables comprehensive modeling of temporal dependency information and spatial distribution characteristics, providing temporal dependency support for unsupervised anomaly detection in multi-condition scenarios.

[0061] S3. Perform spatial similarity measurement in the spatial dimension, and construct a BallTree index structure through a hierarchical spatial partitioning strategy, then search for the spatial nearest neighbor set. Specifically, in the high-dimensional feature space, search for the spatial nearest neighbor set of the sample using Euclidean distance, and then use a hierarchical spatial partitioning strategy to replace the traditional nearest neighbor algorithm to construct the BallTree index structure. Preferably, S3 specifically includes S31 to S35.

[0062] S31. Construct a high-dimensional feature space. Assume the industrial equipment is in time step... The running status is determined by High-dimensional sensor feature vector representation. Within a time window, all samples constitute a multi-condition high-dimensional feature set. This is used to comprehensively reflect the spatiotemporal operating characteristics of the equipment under different working conditions.

[0063] .

[0064] In the formula, For time step The sample. for The The monitoring values ​​of each sensor. This represents the number of sensors. express The Middle The monitoring values ​​of each sensor. Represents a real number.

[0065] .

[0066] In the formula, It is a feature set. For the first Samples at each time step.

[0067] S32. Use Euclidean distance to measure the spatial similarity between samples. The smaller the Euclidean distance, the higher the spatial similarity between samples.

[0068] .

[0069] In the formula Indicates sample and samples The spatial distance between them. This is the sensor's serial number. This represents the number of sensors. For the first The first sample The monitoring values ​​of each sensor. For the first The first sample The monitoring values ​​of each sensor. For the first One sample. For the first One sample.

[0070] Traditional brute-force nearest neighbor search requires calculating the distance between all samples, with a time complexity of O(n). In high-dimensional and large-sample scenarios, the computational cost is extremely high. Therefore, this embodiment introduces a Ball Tree index structure to accelerate nearest neighbor retrieval in high-dimensional space.

[0071] S33. The feature space is organized using a hierarchical partitioning method based on spheres to construct a Ball Tree index structure. Each node in the index structure... Represents a spherical region, which is defined by a central point. and radius definition.

[0072] .

[0073] The steps for constructing a Ball Tree index structure are as follows.

[0074] Calculate the mean of the current node's sample set as the center of the sphere. : .

[0075] Calculate radius In the formula Indicates sample and center point The distance.

[0076] Choose the two points that are farthest apart as the partition center to divide the left and right subsets.

[0077] Continue recursively constructing the left and right child nodes until the number of node samples is less than the threshold. Get the Ball Tree index structure.

[0078] S34, Spatial Nearest Neighbor Search. In the query sample... When considering spatial nearest neighbors, the Ball Tree employs a heuristic pruning strategy: starting from the root node, it calculates the distance between the query sample and the center of the ball. ,like And it is known that the nearest neighbor distance is less than If the node is found to be a valid node, then prune that node; otherwise, recursively search its child nodes. This is the distance threshold. Ultimately, we obtain... A set of spatial nearest neighbors. Its search complexity is... Reduced to It is far superior to the traditional KNN method. .

[0079] .

[0080] In the formula It is a set of spatial nearest neighbors. For query samples The A close neighbor.

[0081] S35. Convert the spatial nearest neighbor set obtained by the search into a similarity matrix.

[0082] .

[0083] In the formula, Indicates sample With sample Spatial correlation. It is a natural exponential function. τ is a scale parameter used to control the distance decay rate. For the sample The spatial nearest neighbor set.

[0084] The obtained spatial similarity matrix It will serve as the input to the dynamic spatiotemporal nearest neighbor standardization module, providing spatial dependency information for subsequent spatiotemporal joint normalization and anomaly detection, and laying the groundwork for the subsequent multi-condition feature normalization.

[0085] S4. Based on the time nearest neighbor set and the spatial nearest neighbor set, establish a dynamic spatiotemporal nearest neighbor standardization model, and then perform spatiotemporal joint normalization processing on the current sample through a spatiotemporal weighted fusion strategy.

[0086] Specifically, combining the time nearest neighbor set Spatial Proximity Set A dynamic spatiotemporal nearest neighbor standardization model is established to achieve unified feature distribution under multiple operating conditions.

[0087] First, calculate the weighted statistics for the temporal nearest neighbor set and the spatial nearest neighbor set, respectively. The weighted statistics include the mean and variance.

[0088] .

[0089] .

[0090] .

[0091] .

[0092] In the formula The mean of the time nearest neighbor set, The standard deviation of the time nearest neighbor set, The number of samples in the time nearest neighbor set, For time nearest neighbor set, The mean of the spatial nearest neighbor set, The standard deviation of the spatial nearest neighbor set, This represents the number of spatial nearest neighbor samples. It is a set of spatial nearest neighbors.

[0093] Then, the current sample is processed using a spatiotemporal weighted fusion strategy. Normalization is performed:

[0094] .

[0095] In the formula Normalized time step The sample. This is the spatiotemporal tradeoff coefficient. It is the mean of the time nearest neighbor set. Let be the standard deviation of the time nearest neighbor set. To prevent smoothing terms from being divided by zero. It is the mean of the set of spatial nearest neighbors. Let be the standard deviation of the set of nearest neighbors in space.

[0096] The above steps achieve joint standardization of temporal dynamic features and spatial distribution features, eliminate the dimensional differences between different sensor channels and samples under multiple operating conditions, and ensure that the feature space remains consistent in both time and space dimensions, providing stable and comparable input features for subsequent unsupervised anomaly detection models based on GAT-DiffusionVAE.

[0097] S5. Input the normalized feature samples into the adaptive threshold unsupervised anomaly detection teacher model based on GAT-DiffusionVAE for training.

[0098] Specifically, in this embodiment, the feature samples after dynamic spatiotemporal nearest neighbor standardization are... The input is used to train an adaptive threshold unsupervised anomaly detection teacher model based on GAT-DiffusionVAE. The teacher model for anomaly detection is constructed based on the GAT-DiffusionVAE model, which combines graph attention network and diffusion variational autoencoder. The specific construction steps are as follows.

[0099] First, the multidimensional time series samples after dynamic spatiotemporal nearest neighbor standardization are regarded as the node features of a graph structure. The nodes represent samples at different time steps, and the edges are determined by the temporal and spatial nearest neighbor relationships, forming an adjacency matrix.

[0100] .

[0101] In the formula For the sample With sample The adjacency matrix. This is the spatiotemporal tradeoff coefficient. For time similarity. For spatial similarity.

[0102] A graph attention network (GAT) is applied to the graph structure to adaptively weight and aggregate node features. By fusing feature correlations at different neighborhood scales through a multi-head attention mechanism, the graph structure embedding representation of the nodes is obtained. The attention coefficient of a multi-head attention mechanism can help learn the dynamic correlations between nodes.

[0103] .

[0104] .

[0105] In the formula For the updated node For nodes The nearest neighbor set. For nodes With nodes The spatiotemporal tradeoff coefficient is calculated by the attention function. Let be a linear transformation matrix. It is a natural exponential function. This is the ReLU function. This is the attention vector. for The transpose of . For nodes . For nodes . For nodes .

[0106] Subsequently, a diffusion process is introduced into the embedding space to model noise perturbation and inverse denoising for normally distributed samples. The inverse mapping is then learned through a neural network. This enables the generation and recovery of the distribution from noisy samples to normal samples.

[0107] .

[0108] .

[0109] In the formula This represents the forward diffusion transition probability. At time step The hidden variable state. At time step The hidden variable state. This represents the time step of the diffusion process. These are noise scheduling parameters. It is an identity matrix. The parameterized inverse denoising transition probability. This is the mean value predicted by the neural network. This represents the variance of the neural network predictions.

[0110] Simultaneously, a variational encoder and decoder (VAE) structure is introduced onto the latent space representation output by the diffusion module to learn the distribution of latent variables. With the generated distribution Minimizing reconstruction error and KL divergence loss: This process enables low-dimensional latent distribution modeling and anomaly representation learning of device features. Represents conditional distribution; Indicates a given Under the conditions, The distribution situation.

[0111] .

[0112] In the formula Let be the loss function of the variational autoencoder (VAE). This is the expected value. These are the parameters for the encoder. These are latent variables. This is the output of the decoder. The reconstruction error... This represents the L2 norm. These are the balancing parameters. Let KL divergence be denoted as KL divergence. It is a standard normal prior. The graph structure embedding representation for nodes. KL divergence. is the standard notation for relative entropy, used to measure the degree of difference between two probability distributions.

[0113] Adaptive threshold anomaly detection based on K-means.

[0114] After the model training is complete, the input samples are... Perform encoding-decoding reconstruction and calculate the reconstruction error as an anomaly score. In the formula Indicates time step Reconstruction error at any given moment. To reconstruct the sample.

[0115] Reconstruction error of all time steps Composition of anomaly score dataset The K-means clustering algorithm was used to... Perform unsupervised clustering with K=2 to obtain normal clusters. with abnormal clusters Cluster center and The anomaly detection threshold is dynamically determined based on the mean of the two cluster centers. .

[0116] When the anomaly score of the sample satisfies If the condition is met, the sample is determined to be in an abnormal state; otherwise, it is determined to be in a normal state.

[0117] S6. A progressive distillation strategy is adopted to train the student network in stages. The intermediate denoised distribution of the teacher model's output is used to guide the student model's learning. Knowledge transfer and feature alignment are achieved by calculating the Wasserstein distance between the student network and the teacher network's output distribution, resulting in a lightweight distilled student model for anomaly detection. Specific steps include S61 to S63.

[0118] S61. To measure the difference in the generation distribution between the teacher model and the student model on the same noise layer, Wasserstein distance is introduced for distribution alignment.

[0119] .

[0120] In the formula, For probability distribution and The distance. This is the infimum. This represents the set of all joint distributions with P and Q as marginal distributions. This is the expected value. For a pair of samples drawn from a joint distribution. Indicates sample pairs From the set Obtained from sampling. This indicates sampling from a specified set. The distance is Euclidean.

[0121] In actual training, the sample mean is used as an approximation:

[0122] .

[0123] In the formula This is an estimate of the Wasserstein distance. and The teacher model and student model respectively represent the first two models in the second stage. Output distribution on each noise layer. This represents the number of samples in the batch. This is the sample number from the distillation. and These are the output results for the teacher and the student, respectively, under the same noise input.

[0124] S62, progressive staged distillation design.

[0125] The backdiffusion process of the teacher model (total) Steps) are divided into Distillation stages Each stage corresponds to a set of noise layers, progressing from high noise to low noise. The student model learns the feature mapping relationships of different noise layers at each stage. Specifically, Stage 1 (high noise stage): learns the coarse-grained distribution of the signal recovered by the device under strong interference conditions. Stage 2 (medium noise stage): captures stable temporal patterns under normal operating conditions. Stage 3 (low noise stage): focuses on fine-grained features and the minute shifts of outliers.

[0126] Student model in The stage uses the output of the noise layer corresponding to the teacher model as the "guidance label," and its learning objective is: By reducing the number of diffusion steps in stages. ( This reduces reasoning complexity, thereby accelerating distillation. In the formula, For the student model in the first The generation mapping of the noise layer corresponding to the stage learning teacher model. For the first The hidden variable states of the noise layer. This is the generative mapping for the teacher model. This represents the total number of steps in the reverse diffusion process. For the first Each distillation stage. Specifically, the reasoning complexity ranges from... Down to .

[0127] S63. Single-stage distillation objective. Specifically, the comprehensive distillation optimization objective is set at each distillation stage. In this model, the student model is guided by the teacher model and minimizes the following weighted distillation loss.

[0128] .

[0129] In the formula, This represents the weighted distillation loss. These are the noise layer weighting coefficients.

[0130] This is used to reflect the importance of different noise levels. Different weights reflect the importance of the noise layer in industrial scenarios. The high noise layer corresponds to operating conditions with large abnormal fluctuations in equipment and has a higher weight. The low noise layer corresponds to stable operating conditions and has a lower weight.

[0131] To balance distribution alignment and feature reconstruction capabilities, the student network simultaneously optimizes the following comprehensive objective function at each stage:

[0132] S64. Comprehensive Training Objective. Specifically, to balance distribution alignment and feature reconstruction capabilities, and considering the constraints of distribution alignment, signal reconstruction, and potential regularization, a total loss function is defined. The student network optimizes the total loss function at each stage.

[0133] .

[0134] .

[0135] .

[0136] In the formula, This is the total loss function. The reconstruction loss is used to constrain the consistency between the student's output signal and the original input signal. This is the balancing coefficient for reconstruction loss. It is a potential regularization term that ensures the latent space maintains a stable distribution. These are the balance coefficients for the potential regularization term. The balance coefficients are used to balance the three losses. This is the expected value. These are samples taken from the joint distribution. The reconstructed samples output for the student model. This represents the L2 norm. Let KL divergence be denoted as KL divergence. The distribution of latent variables learned by the encoder. It is a standard normal prior.

[0137] S7. Input the preprocessed data to be detected into the anomaly detection model, calculate the reconstruction error between the input sample and the reconstructed sample, and combine it with the adaptive threshold judgment mechanism to obtain the anomaly detection result, so as to realize fast unsupervised anomaly detection for multiple working conditions.

[0138] In this embodiment, the preprocessed multidimensional time series data to be detected is input into the anomaly detection model trained by progressive distillation. The model first analyzes the input samples... Encode and decode to reconstruct the corresponding reconstructed sample. The reconstruction error between the two is calculated as the anomaly score.

[0139] .

[0140] In the formula, Indicates time step The reconstruction error at any given moment reflects the model's ability to reproduce the current state of the device.

[0141] When the equipment is in normal working condition, the model can accurately reconstruct its feature distribution with a small reconstruction error. However, when the equipment malfunctions, the sensor drifts, or the system fails, the model cannot restore the original features, resulting in a significant increase in reconstruction error.

[0142] To achieve adaptive anomaly detection under multiple operating conditions, a dynamic threshold mechanism based on K-means clustering is introduced. The reconstruction errors from all time steps are combined into an anomaly score set. :

[0143] .

[0144] right Perform unsupervised clustering with cluster size of 2 to obtain normal clusters. with abnormal clusters Their cluster centers are denoted as follows: and The anomaly threshold is dynamically determined based on the central mean of the two clusters.

[0145] .

[0146] When the reconstruction error of the sample satisfies If the condition is abnormal, the device status corresponding to the sample is determined to be abnormal; otherwise, it is determined to be normal.

[0147] After the model is trained, it is tested and evaluated. If the test evaluation results meet the set requirements, the anomaly detection model is deemed qualified and can be used for subsequent anomaly detection. Otherwise, the model is retrained. The test evaluation results of the anomaly detection model are calculated using the following indicators.

[0148] The formula for calculating accuracy is: In the formula, Accuracy represents the accuracy rate.

[0149] The formula for calculating accuracy is: In the formula, Precision represents the accuracy rate.

[0150] The formula for calculating recall rate is: In the formula, Recall represents the recall rate.

[0151] The formula for calculating the F1 score is: In the formula, F1-score represents the F1 score.

[0152] In the formula This represents the true number of cases, i.e., the number of samples where both the actual situation and the test results are normal. This represents the number of false positives, which is the number of samples that are actually abnormal but are detected as normal. This represents the number of true negative examples, which is the number of samples that are both abnormal in the actual situation and in the test results. This represents the number of false negatives, which is the number of samples that are actually normal but are detected as abnormal.

[0153] In another preferred embodiment, to verify the effectiveness of the proposed progressive distillation unsupervised anomaly detection method and model solution for multi-condition industrial equipment, the publicly available CWRU (Case Western Reserve University) bearing fault dataset was selected as the experimental object. This dataset, collected by the Vibration and Diagnostics Laboratory at Case Western Reserve University, is a classic standard dataset in the field of industrial equipment fault diagnosis and is widely used in health monitoring and anomaly detection research of rotating machinery.

[0154] The dataset was generated by collecting vibration signals from the motor drive and fan bearings under different operating conditions using accelerometers mounted on a test bench. The experimental platform consisted of a motor, coupling, torque sensor, and loading device, simulating different working conditions through external load and speed variations. Data sampling frequencies were set at 12 kHz and 48 kHz. Bearing failures were simulated by introducing artificial damage with diameters of 0.007 inch, 0.014 inch, and 0.021 inch on the inner ring, outer ring, and rolling elements via electrical spark discharge, respectively, to construct abnormal samples of varying severity.

[0155] In the experimental design, the acceleration signal at the drive end was selected as the main input feature to construct a multi-dimensional time series sample, with each sample segment containing 1024 continuous sampling points. Vibration signal data under normal operating conditions was used as the training set for unsupervised learning and feature distribution modeling of the GAT-DiffusionVAE teacher model. Data containing bearing fault signals (inner race, outer race, rolling elements) was used as the test set for evaluating the model's anomaly detection performance.

[0156] To ensure multi-condition characteristic verification, this experiment collected data under four load conditions (0 hp, 1 hp, 2 hp, and 3 hp), and then performed uniform normalization and segmentation to form a multi-condition sample set. The input dimension of the model corresponds to the number of sensors (single-channel vibration signal), and the time step is... The corresponding sample sequence length.

[0157] Table 1. Experimental Parameter Settings

[0158]

[0159] Table 2. Comparison of Anomaly Detection Experiments for Various Multi-condition Processing Methods

[0160]

[0161] Table 3. Experimental Results of the Progressive Distillation Procedure

[0162]

[0163] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0164] This invention combines multiple advanced time-series data analysis techniques, such as dynamic spatiotemporal feature modeling, graph attention network GAT algorithm, diffusion variational autoencoder algorithm, and progressive knowledge distillation, to mine the spatiotemporal dependencies between sensors in industrial equipment. At the same time, it adaptively constructs a dynamic nearest neighbor structure to achieve a unified representation of multi-condition data, thus making up for the shortcomings of existing methods in adaptability to multi-condition scenarios.

[0165] Compared to traditional anomaly detection methods, this invention focuses on the spatial correlation of multidimensional time-series data of industrial equipment while also considering the data's distribution characteristics under different operating conditions and its evolution over time. Through progressive distillation technology, detection efficiency is significantly improved while maintaining accuracy. Combined with an adaptive threshold calculation mechanism, reliance on expert experience is reduced. This invention enhances the accuracy and efficiency of anomaly detection in industrial equipment under various operating conditions, providing a more flexible and accurate unsupervised anomaly detection tool for practical applications. Even without labeled data, it can adapt to the dynamic changes in complex operating conditions of industrial equipment.

[0166] Example 2: The present invention provides an unsupervised anomaly detection device for progressive distillation in industrial equipment, which includes a training data acquisition module, a time dimension module, a spatial dimension module, a normalization module, a teacher model module, a student model module, and a detection module.

[0167] The training data acquisition module is used to acquire and preprocess the monitoring time series data of multi-dimensional sensors of industrial equipment under various operating conditions.

[0168] The time dimension module is used to calculate the time nearest neighbor set of each sample in the time dimension based on the continuous characteristics of the time series.

[0169] The spatial dimension module is used to measure spatial similarity in the spatial dimension and to build a BallTree index structure through a hierarchical spatial partitioning strategy, and then search to obtain the set of spatial nearest neighbors.

[0170] The normalization module is used to perform spatiotemporal joint normalization processing on the current sample based on the temporal nearest neighbor set and the spatial nearest neighbor set through a spatiotemporal weighted fusion strategy.

[0171] The teacher model module is used to input normalized feature samples into the adaptive threshold unsupervised anomaly detection teacher model based on GAT-DiffusionVAE for training.

[0172] The student model module is used to train the student network in stages using a progressive distillation strategy. It uses the intermediate denoised distribution of the teacher model output to guide the student model's learning. Knowledge transfer and feature alignment are achieved by calculating the Wasserstein distance between the student network and the teacher network output distribution, resulting in a lightweight distilled student model as an anomaly detection model.

[0173] The detection module is used to input the preprocessed data to be detected into the anomaly detection model, calculate the reconstruction error between the input sample and the reconstructed sample, and obtain the anomaly detection result by combining the adaptive threshold judgment mechanism.

[0174] Example 3: This invention provides an industrial device including a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement an unsupervised anomaly detection method for progressive distillation in an industrial device as described in any paragraph of Example 1.

[0175] Example 4: This invention provides a computer-readable storage medium. The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform a progressive distillation unsupervised anomaly detection method for industrial equipment as described in any paragraph of Example 1.

[0176] Obviously, the embodiments described above are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0177] In the several embodiments provided in this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus and method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0178] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0179] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks. It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0180] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention are also intended to include the plural forms unless the context clearly indicates otherwise.

[0181] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0182] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0183] The terms "first" and "second" used in the embodiments are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first" and "second" can be interchanged in a specific order or sequence where permitted. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0184] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. 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 unsupervised anomaly detection in the progressive distillation process of industrial equipment, characterized in that, Include: S1. Acquire monitoring time series data of multi-dimensional sensors of industrial equipment under various operating conditions and perform preprocessing; S2. Based on the continuity characteristics of time series, calculate the time nearest neighbor set of each sample in the time dimension; S3. Perform spatial similarity measurement in the spatial dimension, and construct a Ball Tree index structure through a hierarchical spatial partitioning strategy, and then search to obtain the set of spatial nearest neighbors; S4. Based on the time nearest neighbor set and the spatial nearest neighbor set, perform spatiotemporal joint normalization processing on the current sample using a spatiotemporal weighted fusion strategy. S5. Input the normalized feature samples into the adaptive threshold unsupervised anomaly detection teacher model based on GAT-DiffusionVAE for training; S6. A progressive distillation strategy is adopted to train the student network in stages. The intermediate denoised distribution of the teacher model output is used to guide the student model learning. Knowledge transfer and feature alignment are achieved by calculating the Wasserstein distance between the student network and the teacher network output distribution, and a lightweight distilled student model as an anomaly detection model is obtained. S7. Input the preprocessed data to be detected into the anomaly detection model, calculate the reconstruction error between the input sample and the reconstructed sample, and obtain the anomaly detection result by combining the adaptive threshold judgment mechanism. S2 specifically includes: S21. Model the preprocessed multidimensional time series data of industrial equipment to construct time features; ； In the formula For feature set; For time steps ; The length of the time series; Indicates the time step as Samples at that time; Represent real numbers; The number of sensors; S22. Based on the continuity characteristic of time series, on the time axis for each... Establish a time neighborhood window Candidate samples are selected from within the window to form a time nearest neighbor set. ;in, This represents the half-width or time offset of the time neighborhood window; S23. Calculate the temporal similarity between the current sample and the candidate samples using weighted Euclidean distance. And select the nearest one according to the distance from smallest to largest. These samples form a time nearest neighbor set. ; ； ； In the formula For the sample Its time neighbor Temporal similarity between them; This is the time offset step size; This is the sensor's serial number; These are the weighting coefficients for different channels; For time steps The The monitored values ​​of each sensor; For time steps The The monitored values ​​of each sensor; S24. Calculate the weighted mean of the time nearest neighbor set. With variance ; S25. Integrate the statistical features of the current sample and the time neighborhood to obtain the time-dependent feature representation. ; ； In the formula It is a time feature mapping function; S26. Set the time nearest neighbors of each sample and... and This serves as the input basis for subsequent spatial nearest neighbor search and dynamic spatiotemporal joint standardization.

2. The method for unsupervised anomaly detection in progressive distillation of industrial equipment according to claim 1, characterized in that, S3 specifically includes: S31. Construct a high-dimensional feature space; assume the industrial equipment is in time step... The running status is determined by High-dimensional sensor feature vector representation; within a time window, all samples constitute a multi-condition high-dimensional feature set. ; ； ； In the formula, For feature set; For time steps ; For the first Samples at each time step; for The The monitored values ​​of each sensor; express The Middle The monitored values ​​of each sensor; Represent real numbers; S32. Use Euclidean distance to measure the spatial similarity between samples; ； In the formula Indicates sample and samples Spatial distance between them; For the first The first sample The monitored values ​​of each sensor; For the first The first sample The monitored values ​​of each sensor; S33. The feature space is organized using a hierarchical partitioning method based on spheres to construct a Ball Tree index structure; wherein each node in the index structure... Represents a spherical region, which is defined by a central point. and radius definition; The steps for constructing a Ball Tree index structure are as follows: Calculate the mean of the sample set of the current node as the center of the ball. ; Calculate radius In the formula Indicates sample and center point The distance between the two points is calculated; the two points with the greatest distance are selected as the partition centers to divide the left and right subsets; the left and right child nodes are then recursively constructed until the number of node samples is less than the threshold. Get the Ball Tree index structure; S34, Spatial Nearest Neighbor Search; in the query sample When considering spatial nearest neighbors, the Ball Tree employs a heuristic pruning strategy: starting from the root node, it calculates the distance between the query sample and the center of the ball. ,like And it is known that the nearest neighbor distance is less than If the node is found to be true, then prune the node; otherwise, recursively search the child nodes. As the distance threshold; ultimately, we obtain A set of spatial nearest neighbors; S35. Convert the spatial nearest neighbor set obtained by the search into a similarity matrix; ； In the formula, Indicates sample With sample Spatial correlation; The natural exponential function is τ; τ is the scale parameter. For the sample The spatial nearest neighbor set.

3. The method for unsupervised anomaly detection in progressive distillation of industrial equipment according to claim 1, characterized in that, The teacher model for anomaly detection is built based on the GAT-DiffusionVAE model, which combines graph attention network and diffusion variational autoencoder. The specific construction steps are as follows; The multidimensional time series samples after dynamic spatiotemporal nearest neighbor standardization are regarded as the node features of the graph structure. The nodes represent samples at different time steps, and the edges are determined by the temporal and spatial nearest neighbor relationships, forming an adjacency matrix. ； In the formula For the sample With sample The adjacency matrix; For spatiotemporal tradeoffs; For time similarity; Spatial similarity; A graph attention network is applied to the graph structure to adaptively weight and aggregate node features. By fusing feature correlations at different neighborhood scales through a multi-head attention mechanism, a graph structure embedding representation of the nodes is obtained. ; ； ； In the formula For the updated node ; For nodes The nearest neighbor set; For nodes With nodes The spatiotemporal tradeoff coefficients are calculated from the attention function; It is a linear transformation matrix; It is a natural exponential function; For ReLU functions; This is the attention vector; for transpose; For nodes ; For nodes ; For nodes ; A diffusion process is introduced into the embedding space to model noise perturbation and inverse denoising for normally distributed samples; and the inverse mapping is learned through a neural network to achieve the generation and recovery of the distribution from noisy samples to normal samples. A variational encoder and decoder structure is introduced on the latent space representation output by the diffusion module, and the latent variable distribution is learned. With the generated distribution Minimize the reconstruction error and KL divergence loss; where, Represents conditional distribution; Indicates a given Under the conditions, Distribution situation; ； In the formula The loss function of the variational autoencoder; This is the expected value; These are the parameters of the encoder; As latent variables; The output of the decoder; in the reconstruction error Represents the L2 norm; For balance parameters; Let KL divergence be a metric. For standard normal priors; KL divergence is the standard notation for relative entropy, used to measure the degree of difference between two probability distributions.

4. The method for unsupervised anomaly detection in progressive distillation of industrial equipment according to claim 1, characterized in that, S6 specifically includes: To measure the difference in the generation distribution between the teacher model and the student model on the same noise layer, Wasserstein distance is introduced for distribution alignment; ； In the formula, For probability distribution and The distance; The infimum; This represents the set of all joint distributions with P and Q as marginal distributions; This is the expected value; For a pair of samples drawn from a joint distribution; Indicates sample pairs From the set Obtained by sampling from the middle; Indicates sampling from the specified set; The distance is Euclidean. In actual training, the sample mean is used as an approximation: ； In the formula This is an estimate of the Wasserstein distance; and The teacher model and student model respectively represent the first two models in the second stage. Output distribution across each noise layer; This represents the number of samples in the batch processing. This refers to the sample number from the distillation process; and These are the output results of the teacher and the student under the same noise input, respectively. Progressive staged distillation design: dividing the back diffusion process of the teacher model into... Distillation stages Each stage corresponds to a set of noise layers, progressing from high noise to low noise; the student model learns the feature mapping relationships of different noise layers at each stage; the student model in the... The stage uses the output of the noise layer corresponding to the teacher model as the "guidance label," and its learning objective is: By gradually reducing the number of diffusion steps, the reasoning complexity is decreased, thereby accelerating the distillation process; where, For the student model in the first The generation mapping of the noise layer corresponding to the stage learning teacher model; For the first The hidden variable states of each noise layer; Generative mapping for the teacher model; This represents the total number of steps in the reverse diffusion process. For the first One distillation stage; The objective of single-stage distillation is: In the formula, Weighted distillation loss; These are the noise layer weighting coefficients; The overall training objective is: ； ； ； In the formula, This is the total loss function; For reconstruction loss; The balancing coefficient for reconstruction loss; For potential regularization terms; The balance coefficient for the potential regularization term; This is the expected value; These are samples drawn from the joint distribution; Reconstructed samples output for the student model; Represents the L2 norm; Let KL divergence be a metric. The distribution of latent variables learned by the encoder; It is a standard normal prior.

5. The method for unsupervised anomaly detection in progressive distillation of industrial equipment according to claim 1, characterized in that, S7 includes: Anomaly detection based on adaptive threshold using K-means; After the model training is complete, the input samples are... Perform encoding-decoding reconstruction and calculate the reconstruction error as an anomaly score; In the formula Indicates time step Reconstruction error at any given moment; To reconstruct the samples; Represents the L2 norm; Reconstruction error of all time steps Composition of anomaly score dataset The K-means clustering algorithm was used to... Perform unsupervised clustering with cluster size of 2 to obtain normal clusters. with abnormal clusters and its cluster centers are and ;in The length of the time series; The anomaly detection threshold is dynamically determined based on the mean of the two cluster centers. : ; When the anomaly score of the sample satisfies If the condition is met, the sample is determined to be in an abnormal state; otherwise, it is determined to be in a normal state.

6. A method for detecting unsupervised anomalies in the progressive distillation of industrial equipment according to any one of claims 1 to 5, characterized in that, S1 preprocessing includes: Data cleaning and outlier removal: The acquired time-series data from multi-dimensional sensor monitoring of industrial equipment undergoes denoising, missing value imputation, and invalid sample removal. Three methods are employed. The principle is to eliminate outliers; Global preliminary normalization processing is performed, and Z-Score standardization is applied to sensor signals of different dimensions; Multi-condition identification and partitioning: Based on equipment operation logs, control parameters, or environmental labels, historical data is divided into operating conditions to form stable operating condition ranges and transitional operating condition ranges. Time window slicing divides time series samples into segments using sliding time windows of adaptive length. Each window sample contains multidimensional features and their temporal order. Feature smoothing and noise suppression: noise is suppressed by using Savitzky-Golay smoothing inside the window; S4 specifically includes: Calculate the weighted statistics for the temporal nearest neighbor set and the spatial nearest neighbor set, respectively; where the weighted statistics include the mean and variance. The current sample is analyzed using a spatiotemporal weighted fusion strategy. Perform normalization processing; ； In the formula Normalized time step ; For spatiotemporal tradeoffs; It is the mean of the time nearest neighbor set; The standard deviation of the time nearest neighbor set; To prevent smoothing terms from being divided by zero; It is the mean of the set of spatial nearest neighbors; The standard deviation of the spatial nearest neighbor set; S6 further includes: testing and evaluating the anomaly detection model; when the test and evaluation results meet the set requirements, the anomaly detection model is deemed qualified and can be used for subsequent anomaly detection; otherwise, the model is retrained. The test evaluation results of the anomaly detection model are calculated using the following indicators; Accuracy: In the formula, Accuracy represents the accuracy rate. Accuracy: In the formula, Precision represents the accuracy rate. Recall rate: In the formula, Recall represents the recall rate. F1 score: In the formula, F1-score represents the F1 score; In the formula Indicates the actual number of cases; Indicates the number of false positives; Indicates the number of true counterexamples; This indicates the number of false counterexamples.

7. An unsupervised anomaly detection device for progressive distillation in industrial equipment, characterized in that, A method for detecting unsupervised anomalies in progressive distillation of an industrial equipment as described in any one of claims 1 to 6; The unsupervised anomaly detection device for progressive distillation includes: The training data acquisition module is used to acquire and preprocess the monitoring time series data of multi-dimensional sensors of industrial equipment under various operating conditions. The time dimension module is used to calculate the time nearest neighbor set of each sample in the time dimension based on the continuous characteristics of the time series. The spatial dimension module is used to measure spatial similarity in the spatial dimension and to build a Ball Tree index structure through a hierarchical spatial partitioning strategy, and then search to obtain the set of spatial nearest neighbors; The normalization module is used to perform spatiotemporal joint normalization processing on the current sample based on the temporal nearest neighbor set and the spatial nearest neighbor set, using a spatiotemporal weighted fusion strategy. The teacher model module is used to input normalized feature samples into the adaptive threshold unsupervised anomaly detection teacher model based on GAT-DiffusionVAE for training; The student model module is used to train the student network in stages using a progressive distillation strategy. It uses the intermediate denoised distribution of the teacher model output to guide the student model's learning. It achieves knowledge transfer and feature alignment by calculating the Wasserstein distance between the student network and the teacher network output distribution, and obtains a lightweight distilled student model as an anomaly detection model. The detection module is used to input the preprocessed data to be detected into the anomaly detection model, calculate the reconstruction error between the input sample and the reconstructed sample, and obtain the anomaly detection result by combining the adaptive threshold judgment mechanism.

8. An industrial device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory; the computer program can be executed by the processor to implement a progressive distillation unsupervised anomaly detection method for industrial equipment as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform a progressive distillation unsupervised anomaly detection method for industrial equipment as described in any one of claims 1 to 6.

Images