Continuous anomaly detection method and system for wind power equipment state monitoring system

Abstract

The invention provides a continuous anomaly detection method and system for a wind power equipment state monitoring system, and the method comprises the following steps: S1, collecting original multi-dimensional time series data related to the equipment state in the operation process of wind power equipment, and dividing the original multi-dimensional time series data into n tasks according to a time window; performing standardization and random mask preprocessing on the data of each task to generate n tasks to be trained; s2, constructing a continuous learning model based on a Gaussian mixture variational auto-encoder, wherein the continuous learning model comprises an encoder, a decoder and an attention priori network; s3, performing local training on the continuous learning model by using the (1-n) th to-be-trained task; s4, performing global training on the initial global decoder by using the (2-n) th to-be-trained task, and outputting a trained continuous learning model; and S5, inputting a to-be-detected sample into the trained continuous learning model, and outputting a detection result. The problem of disastrous forgetting in an existing anomaly detection method is solved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a continuous abnormality detection method and system for a wind power equipment status monitoring system. Background Art

[0002] Wind power generation, as a vital component of clean energy, has been widely deployed worldwide. Wind turbine equipment, as the linchpin of wind power systems, has a direct impact on the stability and economic viability of the entire system. As wind farms continue to expand, wind turbines are often located in remote, complex environments. These equipment are susceptible to factors such as fluctuating wind speeds, changing climate conditions, and equipment aging, leading to anomalies and even failures. Therefore, developing an efficient and accurate wind turbine equipment status monitoring and anomaly detection system is crucial for ensuring wind power system safety and reducing operation and maintenance costs.

[0003] Traditional wind turbine monitoring methods often rely on physical modeling or expert rule-of-thumbnail analysis. While effective in some scenarios, these methods, due to their strong reliance on prior knowledge, struggle to adapt to the dynamic changes in data distribution during wind turbine operation. This is particularly true when dealing with high-dimensional, nonlinear, and multi-source sensor data, where generalization and flexibility are limited. Recent advances in data acquisition and storage technologies have enabled the acquisition of vast amounts of wind turbine operating data in real time, paving the way for data-driven intelligent monitoring. In particular, deep learning methods, such as autoencoders and variational autoencoders, have garnered widespread attention in unsupervised anomaly detection tasks due to their powerful feature extraction and reconstruction capabilities. These models, by learning the underlying distribution of normal data, can identify unknown anomalies and demonstrate significant advantages in industrial equipment condition monitoring.

[0004] Despite this, current deep learning-based anomaly detection methods still face the following challenges: 1) As wind turbine equipment operates over time, the distribution of system data continues to evolve, making it difficult for a single model to maintain detection performance in long-term operation. This is especially true in multi-task or multi-time period training scenarios, where the model is prone to catastrophic forgetting, resulting in a decrease in the ability to recognize historical anomaly patterns. 2) Most existing methods use a static training model, lacking continuous learning capabilities and struggling to cope with the ever-changing operating conditions and environments of actual operation. 3) Wind turbine equipment anomalies are diverse, ranging from sudden failures to slow degradation. Fixed-threshold detection methods struggle to adapt to multiple anomaly patterns simultaneously, impacting detection accuracy and robustness. Summary of the Invention

[0005] In view of this, an object of the present invention is to provide a continuous anomaly detection method and system for a wind power equipment condition monitoring system, so as to solve the catastrophic forgetting problem existing in existing anomaly detection methods.

[0006] The technical means adopted in the present invention are as follows: A continuous abnormality detection method for a wind power equipment condition monitoring system, comprising: S1. Collect original multidimensional time series data related to the equipment status during the operation of wind power equipment, divide the original multidimensional time series data into n tasks according to time windows; standardize and perform random mask preprocessing on the data of each task to generate n tasks to be trained; S2. Construct a continuous learning model based on a Gaussian mixture variational autoencoder, wherein the continuous learning model includes an encoder, a decoder, and an attention prior network; S3, use the 1st to nth tasks to be trained to perform local training on the continuous learning model; S4, use the 2nd to nth tasks to be trained to perform global training on the initial global decoder, and output the trained continuous learning model; S5. Input the sample to be tested into the trained continuous learning model and output the test results.

[0007] Furthermore, S1 specifically includes the following steps: S11, collecting original multi-dimensional time series data of wind power equipment, said original multi-dimensional time series data including vibration, temperature and speed data, dividing the original multi-dimensional time series data into n tasks according to the window size ; S12. Perform Z-score standardization on the data of each task; S13. Each task is divided into a training set containing only normal samples and a test set containing normal and abnormal samples. A 50% random feature mask is applied to the training set to simulate the data missing scenario.

[0008] Furthermore, in S2, the continuous learning model based on Gaussian mixture variational autoencoder includes: Encoder: Receives the input multi-dimensional time series data of wind power equipment status, extracts features through a multi-layer fully connected neural network, and outputs the Gaussian distribution parameters of the latent variables, including the mean of the latent variable x and log variance , and the mean μ of the context variable w used for prior modeling w and log variance ; The encoder also outputs the probability q of the sample belonging to each Gaussian mixture component through Softmax z , to achieve mixed distribution modeling; Decoder: Gaussian distribution output from encoder In the latent variable z, we sample it through reparameterization technique. , and then perform nonlinear mapping through a three-layer fully connected neural network, and finally reconstruct it into a sample with the same dimension as the original input ; Attention prior network: First, the contextual latent variables Perform a linear transformation and map it into a fixed-dimensional vector, and enhance the context representation capability through a multi-head attention mechanism. This representation is then input into multiple parallel fully connected layers (each corresponding to a Gaussian mixture component), which output the prior mean and logarithmic variance μ of each component. px and , used to construct K prior Gaussian distributions.

[0009] Furthermore, S3 specifically includes the following steps: Local training is performed to minimize the reconstruction error and KL divergence as the loss function; The encoder, decoder, and attention prior network are jointly trained using the data of the first task to be trained to obtain the initialized encoder, decoder, and attention prior network. Freeze the initialized decoder as the initial global decoder; Use the data of subsequent tasks to be trained in turn to train the current encoder and attention prior network, and update the encoder and attention prior network; Until all tasks to be trained are trained, the final encoder and the final attention prior network are output.

[0010] Furthermore, S4 specifically includes the following steps: For the current task, the previous task is defined as the old task, and the old task pseudo samples are generated through the initial global decoder; Merge the old task pseudo samples with the current task real samples to construct a loss function, and train the initial global decoder based on the loss function to obtain an updated global decoder; Use the data of subsequent tasks to be trained in sequence to train the current global decoder and update the global decoder; Until all tasks to be trained are trained, the final global decoder is output; The trained continual learning model consists of the final encoder, the final attention prior network, and the final global decoder.

[0011] Further, further: Use noise Reconstruct old task pseudo samples through the initial global decoder , the samples of the new task and Merge to get the merged sample ; Pass the sample of the new task through the final encoder to get the mean and standard deviation , generating latent variables ,Will and noise Merge to merge noise , using the combined noise Generate reconstructed samples through the current global decoder ; calculate and The reconstruction error between them is used as the loss function of the global training stage, keeping the local encoder unchanged, retraining the parameters of the global decoder, and obtaining a new global decoder.

[0012] Furthermore, S5 specifically includes the following steps: S51, calculating a dynamic threshold according to a threshold sliding window; S52. Calculate the reconstruction error between the input sample of the test set and its reconstructed sample. If the reconstruction error is greater than a threshold, it is determined to be abnormal, otherwise it is normal.

[0013] Furthermore, S51 specifically includes the following steps: S511. Calculate the reconstruction error between the input sample of the training set and its reconstructed sample, take the 95th percentile as the global threshold, and for each sample in the test set, take all samples in the threshold window to calculate the mean of the reconstruction error. and standard deviation ,calculate , as the local threshold corresponding to the sample, and the weighted average of the two is taken as the final threshold for each sample in the test set; S512: Setting global threshold weight coefficient , the local threshold weight coefficient is , for each sample in the test set, its final dynamic threshold It is given by the following formula:

[0014] in, is the global reconstruction error threshold, The local threshold is determined by the mean error of samples in the current time window.

[0015] The present invention further provides a continuous abnormality detection system for a wind power equipment status monitoring system, which is used to implement any of the above-mentioned continuous abnormality detection methods for a wind power equipment status monitoring system, comprising: The task flow construction layer is used to divide the multi-dimensional time series data of wind power equipment into serialized task flows according to time windows. The task division module performs the time window division operation, the data preprocessing module performs standardization, random masking, and training / test set division on each task data, and the model initialization module initializes the encoder, decoder, and attention prior network of the Gaussian mixture variational autoencoder. Local training layer: used to train the local encoder for the normal samples of the current task, so that the model can learn the feature representation of the current device state from the task and maintain the stability of the local representation ability. For the first task, the parameters of the encoder and decoder are jointly trained, and the obtained decoder is used as the global decoder. After that, only the local encoder parameters are updated to adapt to the data distribution. Global update layer: used to incrementally update the global decoder based on the acquisition of new task samples and historical task samples, thereby improving the decoder's adaptability to cross-task representations. It includes a pseudo sample generation module and a multi-task fusion module. The pseudo sample generation module uses random noise to reconstruct historical task pseudo samples using a frozen global decoder; The multi-task fusion module combines new task samples with historical samples and merges their potential representations; The decoder incremental optimization module is jointly trained based on the merged latent representation. The loss function is calculated using the reconstructed samples generated by the decoder and the samples merged by the multi-task fusion module to incrementally update the global decoder weights. Abnormal decision layer: Anomaly judgment is achieved through dynamic threshold detection, the reconstruction error of the test sample is calculated as the abnormality score, and an adaptive threshold that integrates the global threshold and the local threshold is generated based on the sliding window mechanism, and the abnormality detection result is output.

[0016] Compared with the prior art, the present invention has the following advantages: The present invention provides a method and system for continuous anomaly detection in a wind power equipment status monitoring system, which solves the problem of reduced detection performance of existing anomaly detection models when faced with data distribution changes and catastrophic forgetting problems that occur during the long-term operation of wind power equipment. In response to the problem that wind power equipment status data exhibits dynamic distribution characteristics under different tasks and different time periods, a continuous anomaly detection model structure combining local training and a global shared decoder is proposed; by introducing a pseudo-sample generation mechanism to retain historical distribution information, and combining an adaptive threshold mechanism to achieve more accurate anomaly discrimination in multi-stage tasks; at the same time, in order to adapt to the possible concept drift in wind power operation data, a flexible model incremental update and parameter sharing strategy is designed, which effectively improves the stability and generalization ability of detection. This method can realize the timely identification of abnormal states in new tasks while ensuring the model's memory of historical knowledge, thereby enhancing the intelligence and practicality of the wind power equipment status monitoring system. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0018] Figure 1 Schematic diagram of the data anomaly detection system structure of the present invention.

[0019] Figure 2 This is a diagram of the dynamic threshold calculation process in the anomaly detection method of the present invention.

[0020] Figure 3 This is the continuous anomaly detection diagram of the present invention. DETAILED DESCRIPTION

[0021] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0022] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0023] like Figure 1 As shown, the present invention provides a continuous abnormality detection method for a wind power equipment status monitoring system, comprising the following steps: S1. Collect original multidimensional time series data related to the equipment status during the operation of wind power equipment, divide the original multidimensional time series data into n tasks according to time windows; standardize and perform random mask preprocessing on the data of each task to generate n tasks to be trained; S11, collecting original multi-dimensional time series data of wind power equipment, said original multi-dimensional time series data including vibration, temperature and speed data, dividing the original multi-dimensional time series data into n tasks according to the window size ; S12. Perform Z-score standardization on the data of each task; S13. Each task is divided into a training set containing only normal samples and a test set containing normal and abnormal samples. A 50% random feature mask is applied to the training set to simulate the data missing scenario.

[0024] S2. Construct a continuous learning model based on a Gaussian mixture variational autoencoder, wherein the continuous learning model includes an encoder, a decoder, and an attention prior network; Encoder: Receives the input multi-dimensional time series data of wind power equipment status, extracts features through a multi-layer fully connected neural network, and outputs the Gaussian distribution parameters of the latent variables, including the mean of the latent variable x and log variance , and the mean μ of the context variable w used for prior modeling w and log variance In addition, the encoder also outputs the probability q of the sample belonging to each Gaussian mixture component through Softmax z , to achieve mixed distribution modeling; Decoder: Gaussian distribution output from encoder In the latent variable z, we sample it through reparameterization technique. , and then perform nonlinear mapping through a three-layer fully connected neural network, and finally reconstruct it into a sample with the same dimension as the original input ; Attention prior network: First, the contextual latent variables Perform a linear transformation and map it into a fixed-dimensional vector, and enhance the context representation capability through a multi-head attention mechanism. This representation is then input into multiple parallel fully connected layers (each corresponding to a Gaussian mixture component), which output the prior mean and logarithmic variance μ of each component. px and , used to construct K prior Gaussian distributions.

[0025] S3, use the 1st to nth tasks to be trained to perform local training on the continuous learning model; use the data of the first task to train the encoder With decoder Perform joint training to obtain the initial model parameters by minimizing the loss. The trained decoder is retained as a global shared module. In subsequent tasks, only the encoder weights are updated to adapt to the new data distribution. The specific steps are as follows: Local training is performed to minimize the reconstruction error and KL divergence as the loss function; The encoder, decoder, and attention prior network are jointly trained using the data of the first task to be trained to obtain the initialized encoder, decoder, and attention prior network. Freeze the initialized decoder as the initial global decoder; Use the data of subsequent tasks to be trained in turn to train the current encoder and attention prior network, and update the encoder and attention prior network; Until all tasks to be trained are trained, the final encoder and the final attention prior network are output.

[0026] S4. Use the 2nd to nth tasks to be trained to globally train the initial global decoder and output the trained continuous learning model. Generate pseudo samples of the old task through the frozen global decoder and merge them with the real samples of the current task. Use the joint optimization strategy to retrain the global decoder to achieve knowledge fusion between different tasks, thereby alleviating the catastrophic forgetting problem of the model. The specific steps are as follows: For the current task, the previous task is defined as the old task, and the old task pseudo samples are generated through the initial global decoder; Merge the old task pseudo samples with the current task real samples to construct a loss function, and train the initial global decoder based on the loss function to obtain an updated global decoder; Use the data of subsequent tasks to be trained in sequence to train the current global decoder and update the global decoder; Until all tasks to be trained are trained, the final global decoder is output; The trained continual learning model consists of the final encoder, the final attention prior network, and the final global decoder.

[0027] Further: Use noise Reconstruct old task pseudo samples through the initial global decoder , the samples of the new task and Merge to get the merged sample ; Pass the sample of the new task through the final encoder to get the mean and standard deviation , generating latent variables ,Will and noise Merge to merge noise , using the combined noise Generate reconstructed samples through the current global decoder ; calculate and The reconstruction error between them is used as the loss function of the global training stage, keeping the local encoder unchanged, retraining the parameters of the global decoder, and obtaining a new global decoder.

[0028] S5, input the sample to be tested into the trained continuous learning model and output the test results. Specifically, for each test sample , the reconstructed sample is obtained through the global decoder , and calculate the reconstruction error: (1) It is used as an anomaly scoring indicator and introduces an adaptive threshold mechanism to dynamically adjust the detection threshold to achieve accurate identification of sudden and gradual anomalies.

[0029] S51, calculating a dynamic threshold according to a threshold sliding window; S511. Calculate the reconstruction error between the input sample of the training set and its reconstructed sample, take the 95th percentile as the global threshold, and for each sample in the test set, take all samples in the threshold window to calculate the mean of the reconstruction error. and standard deviation ,calculate , as the local threshold corresponding to the sample, and the weighted average of the two is taken as the final threshold for each sample in the test set; S512: Setting global threshold weight coefficient , the local threshold weight coefficient is , for each sample in the test set, its final dynamic threshold It is given by the following formula:

[0030] in, is the global reconstruction error threshold, The local threshold is determined by the mean error of samples in the current time window.

[0031] S52. Calculate the reconstruction error between the input sample of the test set and its reconstructed sample. If the reconstruction error is greater than a threshold, it is determined to be abnormal, otherwise it is normal.

[0032] The present invention also provides an abnormality detection system in a wind power equipment status monitoring system, comprising: The task flow construction layer is used to divide the multi-dimensional time series data of wind power equipment into serialized task flows according to time windows. The task division module performs the time window division operation, the data preprocessing module performs standardization, random masking, and training / test set division on each task data, and the model initialization module initializes the encoder, decoder, and attention prior network of the Gaussian mixture variational autoencoder. Local training layer: used to train the local encoder for the normal samples of the current task, so that the model can learn the feature representation of the current device state from the task and maintain the stability of the local representation ability. For the first task, the parameters of the encoder and decoder are jointly trained, and the obtained decoder is used as the global decoder. After that, only the local encoder parameters are updated to adapt to the data distribution. Global update layer: used to incrementally update the global decoder based on the acquisition of new task samples and historical task samples, thereby improving the decoder's adaptability to cross-task representations. It includes a pseudo sample generation module and a multi-task fusion module. The pseudo sample generation module uses random noise to reconstruct historical task pseudo samples using a frozen global decoder; The multi-task fusion module combines new task samples with historical samples and merges their potential representations; The decoder incremental optimization module is jointly trained based on the merged latent representation. The loss function is calculated using the reconstructed samples generated by the decoder and the samples merged by the multi-task fusion module to incrementally update the global decoder weights. Abnormal decision layer: Anomaly judgment is achieved through dynamic threshold detection, the reconstruction error of the test sample is calculated as the abnormality score, and an adaptive threshold that integrates the global threshold and the local threshold is generated based on the sliding window mechanism, and the abnormality detection result is output.

[0033] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A continuous abnormality detection method for a wind power equipment status monitoring system, characterized in that: The steps include: S1. Collect original multidimensional time series data related to the equipment status during the operation of wind power equipment and divide it into n tasks according to time windows; standardize and preprocess the data of each task with random masks to generate n tasks to be trained; S2. Construct a continuous learning model based on a Gaussian mixture variational autoencoder, wherein the continuous learning model includes an encoder, a decoder, and an attention prior network; S3, use the 1st to nth tasks to be trained to perform local training on the continuous learning model; S4, use the 2nd to nth tasks to be trained to perform global training on the initial global decoder, and output the trained continuous learning model; S5. Input the sample to be tested into the trained continuous learning model and output the test results.

2. The continuous abnormality detection method of the wind power equipment status monitoring system according to claim 1, characterized in that: S1 specifically includes the following steps: S11, collecting original multi-dimensional time series data of wind power equipment, said original multi-dimensional time series data including vibration, temperature and speed data, dividing the original multi-dimensional time series data into n tasks according to the window size ; S12. Perform Z-score standardization on the data of each task; S13. Each task is divided into a training set containing only normal samples and a test set containing normal and abnormal samples. A 50% random feature mask is applied to the training set to simulate the data missing scenario.

3. The continuous abnormality detection method of the wind power equipment status monitoring system according to claim 1, characterized in that: In S2, the continuous learning model based on Gaussian mixture variational autoencoder includes: Encoder: Receives the input multi-dimensional time series data of wind power equipment status, extracts features through a multi-layer fully connected neural network, and outputs the Gaussian distribution parameters of the latent variables, including the mean of the latent variable x and log variance , and the mean μ of the context variable w used for prior modeling w and log variance ; The encoder also outputs the probability q of the sample belonging to each Gaussian mixture component through Softmax z , to achieve mixed distribution modeling; Decoder: Gaussian distribution output from encoder In the latent variable z, we sample it through reparameterization technique. , and then perform nonlinear mapping through a three-layer fully connected neural network, and finally reconstruct it into a sample with the same dimension as the original input ; Attention prior network: First, the contextual latent variables Perform a linear transformation and map it into a fixed-dimensional vector, and enhance the context representation capability through a multi-head attention mechanism. This representation is then input into multiple parallel fully connected layers (each corresponding to a Gaussian mixture component), which output the prior mean and logarithmic variance μ of each component. px and , used to construct K prior Gaussian distributions.

4. The continuous abnormality detection method of the wind power equipment status monitoring system according to claim 1, characterized in that: S3 specifically includes the following steps: Local training is performed to minimize the reconstruction error and KL divergence as the loss function; The encoder, decoder, and attention prior network are jointly trained using the data of the first task to be trained to obtain the initialized encoder, decoder, and attention prior network. Freeze the initialized decoder as the initial global decoder; Use the data of subsequent tasks to be trained in turn to train the current encoder and attention prior network, and update the encoder and attention prior network; Until all tasks to be trained are trained, the final encoder and the final attention prior network are output.

5. The continuous abnormality detection method of the wind power equipment status monitoring system according to claim 1, characterized in that: S4 specifically includes the following steps: For the current task, the previous task is defined as the old task, and the old task pseudo samples are generated through the initial global decoder; Merge the old task pseudo samples with the current task real samples to construct a loss function, and train the initial global decoder based on the loss function to obtain an updated global decoder; Use the data of subsequent tasks to be trained in sequence to train the current global decoder and update the global decoder; Until all tasks to be trained are trained, the final global decoder is output; The trained continual learning model consists of the final encoder, the final attention prior network, and the final global decoder.

6. The continuous abnormality detection method of the wind power equipment status monitoring system according to claim 5, characterized in that: Further: Use noise Reconstruct old task pseudo samples through the initial global decoder , the samples of the new task and Merge to get the merged sample ; Pass the sample of the new task through the final encoder to get the mean and standard deviation , generating latent variables ,Will and noise Merge to merge noise , using the combined noise Generate reconstructed samples through the current global decoder ; calculate and The reconstruction error between them is used as the loss function of the global training stage, keeping the local encoder unchanged, retraining the parameters of the global decoder, and obtaining a new global decoder.

7. The continuous abnormality detection method of the wind power equipment status monitoring system according to claim 1, characterized in that: S5 specifically includes the following steps: S51, calculating a dynamic threshold according to a threshold sliding window; S52. Calculate the reconstruction error between the input sample of the test set and its reconstructed sample. If the reconstruction error is greater than a threshold, it is determined to be abnormal, otherwise it is normal.

8. The continuous abnormality detection method of the wind power equipment status monitoring system according to claim 7, characterized in that: S51 specifically includes the following steps: S511. Calculate the reconstruction error between the input sample of the training set and its reconstructed sample, take the 95th percentile as the global threshold, and for each sample in the test set, take all samples in the threshold window to calculate the mean of the reconstruction error. and standard deviation ,calculate , as the local threshold corresponding to the sample, and the weighted average of the two is taken as the final threshold for each sample in the test set; S512: Setting global threshold weight coefficient , the local threshold weight coefficient is , for each sample in the test set, its final dynamic threshold It is given by the following formula: in, is the global reconstruction error threshold, The local threshold is determined by the mean error of samples in the current time window.

9. A continuous abnormality detection system for a wind power equipment status monitoring system, used to implement the continuous abnormality detection method for a wind power equipment status monitoring system according to any one of claims 1 to 8, characterized in that: include: Task flow construction layer; It is used to divide the multi-dimensional time series data of wind power equipment into serialized task flows according to time windows. The task division module performs the time window division operation, the data preprocessing module performs standardization, random masking and training / test set division on each task data, and the model initialization module initializes the encoder, decoder and attention prior network of the Gaussian mixture variational autoencoder; Local training layer: used to train the local encoder for the normal samples of the current task, so that the model can learn the feature representation of the current device state from the task and maintain the stability of the local representation ability. For the first task, the parameters of the encoder and decoder are jointly trained, and the obtained decoder is used as the global decoder. After that, only the local encoder parameters are updated to adapt to the data distribution. Global update layer; It is used to incrementally update the global decoder based on the acquisition of new task samples and historical task samples, thereby improving the decoder's adaptability to cross-task representations. It includes a pseudo sample generation module and a multi-task fusion module. The pseudo sample generation module uses random noise to reconstruct historical task pseudo samples using a frozen global decoder; The multi-task fusion module combines new task samples with historical samples and merges their potential representations; The decoder incremental optimization module is jointly trained based on the merged latent representation. The loss function is calculated using the reconstructed samples generated by the decoder and the samples merged by the multi-task fusion module to incrementally update the global decoder weights. Abnormal decision-making layer; Anomaly determination is achieved through dynamic threshold detection. The reconstruction error of the test sample is calculated as the anomaly score. An adaptive threshold that fuses the global threshold and the local threshold is generated based on the sliding window mechanism, and the anomaly detection result is output.

Images