A multivariate time series anomaly detection method, system, medium, device and terminal

Abstract

The application belongs to the technical field of wind turbine abnormality detection, and discloses a multivariate time sequence abnormality detection method, system, medium, equipment and terminal, a fusion framework of abnormality detection composed of a spatial feature reconstruction module and a time sequence prediction module is constructed; in the spatial feature reconstruction module, a DSAE deep stack autoencoder is used to extract spatial correlation features among multivariate, output low-dimensional feature representation, and calculate reconstruction error; in the time sequence prediction module, a self-attention-based Transformer prediction subnetwork is used to extract complex global time sequence dependency; two subnetworks in the fusion framework are simultaneously optimized by using a joint training method, reconstruction error and prediction error are minimized, and multivariate time sequence abnormality detection for wind turbine SCADA data is realized. The application combines the advantages of reconstruction methods and prediction methods, extracts spatial features and time sequence dependencies in multivariate time sequence data, and realizes reliable wind power SCADA abnormality detection.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine anomaly detection technology, and particularly relates to a multivariable time-series anomaly detection method, system, medium, equipment and terminal. Background Technology

[0002] In recent decades, new energy sources have become a key focus for countries worldwide, gradually replacing fossil fuels as major sources of air pollution. Driven by the global shift towards renewable energy production, wind power, as a promising alternative, has received substantial investment. Globally, over 90GW of new wind power installations have been installed, demonstrating strong growth momentum. Besides environmentally friendly power supply, the profitability of the wind power industry must be guaranteed. The operation and maintenance costs of wind turbines are a significant factor. A major component of wind turbine maintenance costs is premature failure due to excessive wear and tear in harsh and remote environments. Therefore, early identification of these failures is crucial. To reduce wind turbine costs and unplanned downtime, an efficient intelligent condition monitoring and anomaly detection method can be designed to provide high availability for wind turbines, improve wind farm efficiency, and enhance economic benefits. Current anomaly detection methods for wind turbines are mainly divided into two categories: mechanistic model-based methods and data-driven methods. Mechanistic model-based methods rely on establishing accurate mathematical models. However, the components in wind turbines are highly coupled, exhibiting complex characteristics and nonlinear relationships. Therefore, accurately establishing explicit mathematical models of wind turbines is extremely difficult. In contrast, data-driven methods are more suitable for complex systems with high nonlinearity. Currently, wind farms have widely installed Supervisory Control and Data Acquisition (SCADA) systems to collect operational status data of wind turbines, which includes dozens of different time-series data sets containing a wealth of valuable information. Therefore, this invention takes multivariate wind turbine SCADA time-series data as its research object. Most existing anomaly detection methods are designed for point anomalies and cannot be used for conditional anomalies and group anomalies because they exhibit completely different behaviors from point anomalies. For multidimensional data, a major challenge is that some anomalies can only be detected when considering two or more data dimensions. Wind power SCADA data is a type of multivariate time-series data, and in addition to temporal features, it is also necessary to extract spatial correlations between variables. Therefore, traditional anomaly detection techniques, such as OCSVM, PCA, and GMM, cannot effectively capture the spatial and temporal dependencies in multivariate time series. However, multivariate time-series anomaly detection based on wind power SCADA data relies on contextual feature extraction, which requires the use of deep learning methods. To address this key challenge—extracting the spatial correlation and temporal dependency of multivariate time series data—it is urgent to design a new method and system for detecting multivariate time series anomalies.

[0003] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0004] (1) Currently, among the methods for detecting anomalies in wind turbines, those based on mechanistic models require the establishment of accurate mathematical models. However, the components in wind turbines are highly coupled and have complex characteristics and nonlinear relationships, making it very difficult to accurately establish a display mathematical model of wind turbines.

[0005] (2) Most existing anomaly detection methods are designed for point anomalies and are difficult to apply to conditional anomalies and group anomalies because they exhibit completely different behaviors from point anomalies. For multidimensional data, some anomalies can only be detected when more than two data dimensions are considered at the same time.

[0006] (3) Traditional anomaly detection techniques cannot effectively capture the spatial and temporal dependencies in multivariate time series. However, multivariate time series anomaly detection based on wind power SCADA data requires contextual feature extraction, which necessitates the use of deep learning methods. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a multivariate time-series anomaly detection method, system, medium, device, and terminal, and particularly relates to a multivariate time-series anomaly detection method based on wind power SCADA data, aiming to solve the anomaly detection problem of wind power SCADA data.

[0008] This invention is implemented as follows: a multivariate time series anomaly detection method, the multivariate time series anomaly detection method comprising:

[0009] A fusion framework for anomaly detection is constructed, consisting of a spatial feature reconstruction module and a temporal prediction module. In the spatial feature reconstruction module, a DSAE deep stacked autoencoder is used to extract spatial correlation features between multiple variables, output low-dimensional feature representations, and calculate reconstruction errors. In the temporal prediction module, a Transformer prediction sub-network based on self-attention is used to extract complex global temporal dependencies. A joint training method is adopted to simultaneously optimize the two sub-networks in the fusion framework, minimizing reconstruction and prediction errors, thereby achieving multivariate temporal anomaly detection for wind turbine SCADA data.

[0010] Furthermore, the multivariate time series anomaly detection method also includes:

[0011] After basic data processing of the wind turbine SCADA data, the spatial correlation features between multiple variables are extracted from the multidimensional time-series input DSAE deep stacked autoencoder. Through encoder compression extraction, a low-dimensional feature representation z of the multidimensional variables is obtained. The multidimensional time-series data r for model reconstruction is obtained through a multi-layer decoder, which is used to calculate the distance-metric-based reconstruction error E. r ;

[0012] The low-dimensional feature representation z and the reconstruction error E r The input is fed into the time series prediction module, where temporal dependencies are extracted through a Transformer prediction sub-network based on self-attention, the future values ​​of the temporal features are predicted, and the prediction error E is calculated. p After joint training of the framework, the anomaly threshold T is calculated. AD As the decision threshold; during the online detection phase, the input test data is fed into the trained model to obtain the loss value E(x). test If E(x) test ) greater than T AD If the condition is met, the test sample will be marked as an abnormal state; otherwise, it will be marked as a normal state.

[0013] Furthermore, the multivariate time series anomaly detection method includes the following steps:

[0014] Step 1: Preprocess the SCADA data using variational mode decomposition;

[0015] Step 2: Construct a spatial feature reconstruction module and extract spatially relevant features;

[0016] Step 3: Construct a time series prediction module and extract global time series dependencies;

[0017] Step 4: Joint training to minimize reconstruction error and prediction error;

[0018] Step 5: Perform anomaly analysis and online monitoring respectively.

[0019] Furthermore, the SCADA data preprocessing in step one includes:

[0020] SCADA data was used as experimental data, which was collected by acceleration sensors installed at various parts of the wind turbine generator. To address data noise, variational mode decomposition (VMD) was used to process the non-stationary SCADA signal. The intrinsic mode functions (IMFs) in the frequency domain signal were selected by using the kurtosis index to filter out the noise components in the SCADA data.

[0021] The second step of constructing the spatial feature reconstruction module and extracting spatially relevant features includes:

[0022] Multidimensional SCADA data is input into the DSAE depth stack autoencoder to extract spatial correlation features among multiple variables.

[0023] (1) Constructing a spatial feature reconstruction module

[0024] The autoencoder (AE) obtains a compressed representation of the input vector X by encoding it through hidden layers, and then decodes it back to the original vector. The compressed vector in the intermediate layer is the low-dimensional feature representation z of the original input vector. Using the low-dimensional feature representation z as the input vector, the deep encoder is built by stacking layers to obtain a deeper stacked autoencoder network. The compressed representation is extracted three times to obtain a seven-layer symmetric structure DSAE encoder-decoder network.

[0025] The principle by which the encoder Encode(x) obtains the compressed representation is shown in the following formula:

[0026] z i =σ(W i ×x i +b i );

[0027] Where σ(·) is the activation function ReLU, W i x is the weight parameter of the i-th layer. i b is the input vector of the i-th layer. i It is the bias term of the i-th layer.

[0028] The principle by which the decoder Decode(z) recovers the reconstructed vector from the compressed vector is shown in the following formula:

[0029]

[0030] Where σ(·) is the activation function ReLU, z is the weight parameter of the decoder corresponding to the compressed vector of layer i. i b is the compression vector of the i-th layer. i It is the bias term of the decoder corresponding to the i-th layer compression vector.

[0031] (2) Extracting low-dimensional feature representation z of spatial correlation

[0032] Multidimensional SCADA data is input into a DSAE deep stacked autoencoder to extract spatial correlation features between multiple variables. Through compression extraction by the deep autoencoder, a low-dimensional feature representation z of the multidimensional variables is finally obtained.

[0033] (3) Calculate the reconstructed multidimensional time series data r and the reconstruction error E. r

[0034] The low-dimensional feature representation z extracted from the intermediate layer is input into the decoder network of the second half of DSAE to obtain the reconstructed multi-dimensional temporal data r, and the reconstruction error E based on the Euclidean distance metric is calculated. r The formula is as follows:

[0035] E r =||xr||2;

[0036] Where x is the input multidimensional time series data, and r is the reconstructed data.

[0037] Furthermore, step three, which involves constructing a time-series prediction module and extracting global time-series dependencies, includes:

[0038] Train the Transformer prediction subnetwork to extract global temporal dependencies from the time series data.

[0039] (1) Constructing a Transformer-based prediction subnetwork

[0040] Construct a Transformer model with an encoder-decoder structure. The encoder part consists of a set of encoders with a stack of 6 layers. Each encoder has normalization, multi-head attention, residual summation, normalization, and fully connected units. The decoder consists of a set of decoders with the same number as the encoder. By using a self-attention mechanism, the Transformer obtains global feature extraction capabilities.

[0041] (2) Transformer prediction subnetwork output prediction

[0042] The low-dimensional feature representation z of the DSAE output and the reconstruction error E r The input is fed into the time series prediction module, where the encoder extracts the global time series dependencies; the decoder outputs the predicted value z. p , with the predicted value z p Compared with the true value z t The mean squared error (MSE) is used as a loss metric E. p The calculation formula is as follows:

[0043] E p =||z p -z t ||2.

[0044] Furthermore, the joint training in step four includes:

[0045] For training the overall fusion framework, a joint training approach is used, proposing an end-to-end minimized joint training method to obtain better spatially compressed feature representations and global temporal dependency information. The joint training consists of two parts: reconstruction error E r and prediction error E p The objective function is:

[0046] J(w)=E r +λE p =||xr||2+λ||z p -z t ||2;

[0047] Among them, E r For reconstruction error, E p Let λ be the prediction error, and λ be a hyperparameter that controls the importance of the two errors.

[0048] The anomaly analysis in step five includes:

[0049] To distinguish between normal and abnormal states, an abnormality threshold T needs to be determined. AD The corresponding anomaly threshold T is calculated based on the training results of the model using the training set data. AD :

[0050]

[0051] Where E(x) is the loss value of the model's objective function. This represents the mean of the model's loss values.

[0052] The online monitoring includes:

[0053] Test data x test After preprocessing according to step one, the data is input into the already trained model, and the corresponding reconstruction error E is calculated. r Prediction error E p and the objective function loss value E(x) test According to the abnormal decision-making rule, when E(x) test )>t AD When the test result is positive, the sample is marked as abnormal; otherwise, it is marked as normal.

[0054] Another object of the present invention is to provide a multivariate temporal anomaly detection system applying the aforementioned multivariate temporal anomaly detection method, the multivariate temporal anomaly detection system comprising:

[0055] The data preprocessing module is used to preprocess SCADA data using variational mode decomposition.

[0056] The spatial feature reconstruction module is used to input multidimensional SCADA data into the DSAE depth stack autoencoder and extract spatial correlation features between multiple variables.

[0057] The temporal prediction module is used to train the Transformer prediction sub-network and extract global temporal dependencies from temporal data.

[0058] The joint training module is used to minimize reconstruction error and prediction error through joint training.

[0059] The anomaly analysis module is used to calculate anomaly thresholds based on the training results of the model using training set data.

[0060] The online monitoring module is used to determine the test results of samples based on anomaly decision rules.

[0061] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the multivariate timing anomaly detection method.

[0062] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the multivariate timing anomaly detection method.

[0063] Another objective of this invention is to provide an information data processing terminal for implementing the aforementioned multivariable time-series anomaly detection system.

[0064] Based on the above technical solutions and the technical problems solved, please analyze the advantages and positive effects of the technical solution to be protected by this invention from the following aspects:

[0065] First, addressing the technical problems existing in the prior art and the difficulty in solving them, this paper closely analyzes, in conjunction with the technical solution to be protected by this invention and the results and data obtained during the research and development process, how the technical solution of this invention solves the technical problems, and the inventive technical effects brought about by solving these problems. The specific description is as follows:

[0066] This invention aims to solve the problem of monitoring the operational status of wind turbine generators. Primarily using SCADA data from wind turbine generators as the analysis object, it proposes a multivariate time-series anomaly detection method based on wind power SCADA data to address the status monitoring issue of wind turbine generators. This invention utilizes a fusion framework of reconstruction and prediction models to solve the anomaly detection problem in wind power SCADA data.

[0067] This invention employs a deep learning-based anomaly detection method and proposes a fusion framework for anomaly detection, mainly composed of a spatial feature reconstruction module and a temporal prediction module. In the spatial feature reconstruction module, this invention uses a DSAE deep stacked autoencoder to extract spatial correlation features between multiple variables, outputting low-dimensional feature representations and calculating reconstruction errors. In the temporal prediction module, this invention uses a self-attention-based Transformer prediction sub-network to extract complex global temporal dependencies. Finally, this invention employs a joint training method to simultaneously optimize the two sub-networks in the fusion framework, minimizing reconstruction and prediction errors, thereby achieving multivariate temporal anomaly detection for wind turbine SCADA data.

[0068] This invention utilizes variational mode decomposition to preprocess SCADA data, eliminating significant noise in wind power SCADA data and improving the accuracy of anomaly detection. It solves the challenge of high-dimensional data correlation extraction by extracting spatial correlations between multidimensional data through a DSAE-based reconstruction method. A Transformer-based prediction module extracts temporal dependencies between multivariate time-series data, and the global extraction capability of self-attention addresses the low computational efficiency caused by the recursive structure of RNNs. End-to-end joint training promotes the optimization of the fusion-based model, preventing independent optimization by the two modules from getting trapped in local optima. After training the anomaly detection model, the data to be detected can be fed into the model in real time for online anomaly detection, meeting the requirements of engineering applications.

[0069] Second, considering the technical solution as a whole or from a product perspective, the technical effects and advantages of the technical solution to be protected by this invention are specifically described as follows:

[0070] To address the complex noise interference present in wind power data, this invention employs the Variational Mode Decomposition (VMD) method to remove the noise component from wind power SCADA data. This method can better solve the problems of suppressing mode aliasing and endpoint effects, thereby efficiently processing non-stationary data.

[0071] This invention proposes a multivariate temporal anomaly detection method for wind power SCADA based on a fusion framework. It combines the advantages of reconstruction and prediction methods, makes up for the shortcomings of each method, and can more effectively extract spatial features and temporal dependencies from multivariate time series data. Furthermore, it ensures the global optimization effect of the multi-step model through joint training, thereby achieving reliable wind power SCADA anomaly detection.

[0072] This invention employs a Transformer prediction subnetwork based on an autonomous force mechanism, which is not hindered by the serial computation and local memory issues of current mainstream RNN-based methods. The Transformer prediction subnetwork has the ability to extract global temporal dependencies and can compute self-attention in parallel, thus significantly improving prediction efficiency.

[0073] Third, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:

[0074] (1) The expected benefits and commercial value of the technical solution of the present invention after transformation are as follows: The wind power industry is currently booming, but the maintenance cost of equipment will become a key factor in the profitability of wind farms within a few years. However, by using the data of the wind power SCADA system already equipped in the wind farm for abnormal status detection, abnormal events can be detected in advance without additional costs, so as to avoid failures or arrange maintenance and repair work in advance, reduce unplanned downtime, and bring cost benefits to the wind farm.

[0075] (2) Does the technical solution of this invention solve a technical problem that people have long desired to solve but have never been able to successfully solve? This invention proposes a multivariate time series anomaly detection method that combines reconstruction and prediction methods for wind power SCADA data. This breaks through the problem that existing methods in the industry are difficult to effectively extract the contextual spatial correlation and temporal dependency of multivariate time series data. The fusion structure effectively improves the anomaly detection performance and robustness of the model, which is beneficial to the application of the algorithm in engineering. Attached Figure Description

[0076] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0077] Figure 1 This is a flowchart of the multivariate temporal anomaly detection method provided in the embodiments of the present invention;

[0078] Figure 2 This is a flowchart of the wind power anomaly detection process provided in an embodiment of the present invention;

[0079] Figure 3 This is a block diagram of the multivariable temporal anomaly detection system provided in this embodiment of the invention;

[0080] In the diagram: 1. Data preprocessing module; 2. Spatial feature reconstruction module; 3. Temporal prediction module; 4. Joint training module; 5. Anomaly analysis module; 6. Online monitoring module. Detailed Implementation

[0081] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0082] To address the problems existing in the prior art, the present invention provides a method, system, medium, device and terminal for detecting multivariable temporal anomalies. The present invention will be described in detail below with reference to the accompanying drawings.

[0083] I. Explanatory and Illustrative Embodiments. To enable those skilled in the art to fully understand how the present invention is specifically implemented, this section provides an explanatory and illustrative description of the embodiments described in the claims.

[0084] like Figure 1 As shown, the multivariate temporal anomaly detection method provided in this embodiment of the invention includes the following steps:

[0085] S101, variational mode decomposition is used to preprocess SCADA data;

[0086] S102, Construct a spatial feature reconstruction module and extract spatially relevant features;

[0087] S103, Construct the time series prediction module and extract global time series dependencies;

[0088] S104, joint training, minimizes reconstruction error and prediction error;

[0089] S105 performs anomaly analysis and online monitoring respectively.

[0090] The principle diagram of the multivariate temporal anomaly detection method provided in this embodiment of the invention is as follows: Figure 2 As shown.

[0091] like Figure 3 As shown, the multivariate temporal anomaly detection system provided in this embodiment of the invention includes:

[0092] Data preprocessing module 1 is used to preprocess SCADA data using variational mode decomposition;

[0093] Spatial feature reconstruction module 2 is used to input multidimensional SCADA data into the DSAE depth stack autoencoder and extract spatial correlation features between multiple variables.

[0094] Temporal prediction module 3 is used to train the Transformer prediction sub-network and extract global temporal dependencies from temporal data;

[0095] Joint training module 4 is used to minimize reconstruction error and prediction error by employing joint training.

[0096] Anomaly analysis module 5 is used to calculate anomaly thresholds based on the training results of the model using training set data;

[0097] The online monitoring module 6 is used to determine the detection results of samples based on anomaly decision rules.

[0098] The technical solution of the present invention will be further described below with reference to specific embodiments.

[0099] The purpose of this invention is to solve the problem of operational status monitoring and anomaly detection in complex wind turbine systems. It proposes a multivariate temporal anomaly detection method based on a fusion framework. Taking wind turbine SCADA data as the analysis object, after basic data processing, the spatial correlation features between multiple variables are extracted from the multidimensional time-series input DSAE deep stacked autoencoder. Through encoder compression extraction, low-dimensional feature representations z of the multidimensional variables are obtained. Then, multi-layer decoders are used to obtain multidimensional time-series data r for model reconstruction, which is used to calculate the distance-metric-based reconstruction error E. r Then, the low-dimensional feature representation z and the reconstruction error E are... r The input is fed into the time series prediction module, where a self-attention-based Transformer prediction sub-network is used to model and extract temporal dependencies, predict future values ​​of temporal features, and then calculate the prediction error E. p Compared to RNN-based prediction networks, the Transformer prediction sub-network exhibits parallel computation and global feature extraction capabilities, resulting in a significant improvement in prediction performance. After joint training of the framework, the anomaly threshold T is calculated. AD This serves as the decision threshold. During the online detection phase, the input test data is fed into the trained model to obtain the loss value E(x). test If E(x) test ) greater than T AD If the condition is met, the test sample will be marked as an abnormal state; otherwise, it will be marked as a normal state.

[0100] The specific technical solutions adopted to achieve the objectives of this invention are as follows:

[0101] The multivariate time-series anomaly detection method based on wind power SCADA data provided in this invention has the following framework: Figure 2 As shown, the specific steps include the following:

[0102] (1) SCADA data preprocessing

[0103] SCADA data from a wind farm in northern China was used as the experimental data. The data was collected by acceleration sensors installed at various parts of the wind turbine generators. Because the operating environment of wind turbine generators is typically very harsh, data processing is necessary to eliminate noise pollution. Erroneous data due to various objective reasons, such as communication failures or missing data, needed to be removed. Variational Mode Decomposition (VMD) was used to process the non-stationary SCADA signals to address data noise. The dominant intrinsic mode functions (IMFs) in the frequency domain signal were selected using a kurtosis index to filter out noise components in the SCADA data.

[0104] (2) Spatial feature reconstruction module, extracting spatially relevant features

[0105] Multidimensional SCADA data is input into the DSAE depth stack autoencoder to extract spatial correlation features among multiple variables.

[0106] 1) Constructing a spatial feature reconstruction module

[0107] The principle of the autoencoder (AE) is to obtain a compressed representation of the input vector X by encoding it through a hidden layer, and then decode it back to the original vector. The compressed vector of the intermediate layer is the low-dimensional feature representation z of the original input vector. Then, using the low-dimensional feature representation z as the input vector, we continue to stack and build a deep encoder, which can obtain a deeper stacked autoencoder network. This method repeats the extraction of compressed representations three times to obtain a seven-layer symmetric structure DSAE encoder-decoder network.

[0108] The principle by which the encoder Encode(x) obtains the compressed representation is shown in formula (1):

[0109] z i =σ(W i ×x i +b i (I)

[0110] Where σ(·) is the activation function ReLU, W i x is the weight parameter of the i-th layer. i b is the input vector of the i-th layer. i It is the bias term of the i-th layer.

[0111] The principle by which the decoder Decode(z) recovers the reconstructed vector from the compressed vector is shown in Equation (2):

[0112]

[0113] Where σ(·) is the activation function ReLU, z is the weight parameter of the decoder corresponding to the compressed vector of layer i. ib is the compression vector of the i-th layer. i It is the bias term of the decoder corresponding to the i-th layer compression vector.

[0114] 2) Extracting low-dimensional feature representation z of spatial correlation

[0115] Multidimensional SCADA data is input into a DSAE deep stacked autoencoder to extract spatial correlation features between multiple variables. Through compression extraction by the deep autoencoder, a low-dimensional feature representation z of the multidimensional variables is finally obtained.

[0116] 3) Calculate the reconstructed multidimensional time series data r and the reconstruction error E. r

[0117] The low-dimensional feature representation z extracted from the intermediate layer is input into the decoder network of the second half of DSAE to finally obtain the reconstructed multi-dimensional temporal data r, and then the reconstruction error E based on the Euclidean distance metric is calculated. R Its formula (3) is as follows:

[0118] E r =||xr||2 (3)

[0119] Where x is the input multidimensional time series data, and r is the reconstructed data.

[0120] (3) Time series prediction module

[0121] Train the Transformer prediction subnetwork to extract global temporal dependencies from the time series data.

[0122] 1) Construct a Transformer-based prediction subnetwork

[0123] A Transformer model with an encoder-decoder structure is constructed. The encoder consists of a set of encoders, stacked with 6 layers. Each encoder has units for normalization, multi-head attention, residual summation, normalization, and full connection. The decoder also consists of a set of decoders, with the same number as the encoder. By using a self-attention mechanism, the Transformer gains global feature extraction capabilities because this mechanism can extract temporal dependencies from the entire time series. This means that the Transformer is not affected by the sequential order of computation, possessing the advantage of computational parallelism, which can significantly improve training efficiency. In contrast, traditional neural networks based on RNNs or LSTMs are affected by sequential computation, requiring the calculation of the next prediction based on the previous result. This not only prevents parallel training but also limits the extraction of only local temporal features.

[0124] 2) Transformer prediction subnetwork output prediction

[0125] The low-dimensional feature representation z of the DSAE output and the reconstruction error Er The input is fed into the time series prediction module, where the encoder extracts the global time series dependencies, and the decoder outputs the predicted value z. p , with the predicted value z p Compared with the true value z t The mean squared error (MSE) is used as a loss metric E. p The calculation formula is as follows:

[0126] E p =||z p -z t ||2 (4)

[0127] (4) Joint training

[0128] Since iterative training of individual modules is prone to getting trapped in local optima and cannot achieve collaborative optimization, this method employs a joint training approach for training the overall fusion framework. It proposes an end-to-end minimization joint training method to simultaneously obtain better spatially compressed feature representations and global temporal dependency information. This method mainly consists of two parts: reconstruction error E... r and prediction error E p Its objective function is:

[0129] J(w)=E r +λE p =||xr||2+λ||z p -z t ||2 (5)

[0130] Among them, E r For reconstruction error, E p Let λ be the prediction error, and λ be a hyperparameter that controls the importance of the two errors.

[0131] (5) Anomaly Analysis

[0132] To distinguish between normal and abnormal states, an abnormality threshold T needs to be determined. AD Based on the training results of the model using the training set data, this invention can calculate the corresponding anomaly threshold T. AD :

[0133]

[0134] Where E(x) is the loss value of the model's objective function. This represents the mean of the model's loss values.

[0135] (6) Online monitoring

[0136] Test data x test After preprocessing according to step (1) above, the data is input into the already trained model, and the corresponding reconstruction error E is calculated.r Prediction error E p and the objective function loss value E(x) test According to the abnormal decision-making rule, when E(x) test )>T AD When the result is positive, the sample is marked as abnormal; otherwise, it is marked as normal.

[0137] II. Application Examples. To demonstrate the inventiveness and technical value of the technical solution of this invention, this section provides application examples of the technical solution of the claims on specific products or related technologies.

[0138] The method was experimentally verified on the SCADA data of a 2.0MW wind turbine in a wind farm in northern China. It recorded the wind turbine operation data from January to December 2019, as well as the generator-related status and alarm fault records, totaling 427,020 records and 45 sensor measurement points, with a sampling interval of 10 minutes for each measurement point.

[0139] III. Evidence of the Relevant Effects of the Embodiments. The embodiments of the present invention have achieved some positive effects during research and development or use, and indeed possess significant advantages compared to existing technologies. The following description, in conjunction with data, charts, and other materials from the experimental process, illustrates these advantages.

[0140] The experiment used F1 score and recall rate to evaluate the anomaly detection performance, and the evaluation results are shown in Table 1.

[0141] Table 1 Evaluation Results

[0142]

[0143] Compared with baseline models OCSVM and IFOrest, the method proposed in this embodiment of the invention improves the F1 score and recall by an average of 25.91% and 44.78%, respectively. Experimental results show that the present invention can greatly improve the F1 score and recall of anomaly detection, which means that more early anomalies can be detected while maintaining high accuracy in anomaly detection.

[0144] It should be noted that embodiments of the present invention can be implemented using hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented using hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or using software executed by various types of processors, or using a combination of the above-described hardware circuitry and software, such as firmware.

[0145] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting multivariate temporal anomalies, characterized in that, The multivariate time series anomaly detection method includes: A fusion framework for anomaly detection is constructed, consisting of a spatial feature reconstruction module and a temporal prediction module. In the spatial feature reconstruction module, a DSAE deep stacked autoencoder is used to extract spatial correlation features between multiple variables, output low-dimensional feature representations, and calculate reconstruction errors. In the temporal prediction module, a Transformer prediction sub-network based on self-attention is used to extract complex global temporal dependencies. A joint training method is adopted to simultaneously optimize the two sub-networks in the fusion framework, minimizing reconstruction and prediction errors, thereby achieving multivariate temporal anomaly detection for wind turbine SCADA data. The multivariate time series anomaly detection method is specifically as follows: After basic data processing, the multidimensional time series data of the wind turbine SCADA system is input into the DSAE deep stacked autoencoder to extract the spatial correlation features between multiple variables. Through the compression extraction of the encoder, the low-dimensional feature representation z of the multidimensional variables is obtained. The multidimensional time series data r of the model reconstruction is obtained through the multi-layer decoder and used to calculate the reconstruction error Er based on the distance metric. The low-dimensional feature representation z and reconstruction error Er are input into the temporal prediction module. Temporal dependencies are extracted by modeling a Transformer prediction sub-network based on self-attention, and the future values ​​of temporal features are predicted to obtain the prediction error Ep. After joint training of the framework, the anomaly threshold TAD is calculated as the decision threshold. In the online detection stage, the test data is input into the trained model to obtain the loss value E(xtest). If E(xtest) is greater than TAD, the test sample is marked as an abnormal state; otherwise, it is a normal state.

2. The multivariate temporal anomaly detection method as described in claim 1, characterized in that, The multivariate temporal anomaly detection method includes the following steps: Step 1: Preprocess the SCADA data using variational mode decomposition; Step 2: Construct a spatial feature reconstruction module and extract spatially relevant features; Step 3: Construct a time series prediction module and extract global time series dependencies; Step 4: Joint training to minimize reconstruction error and prediction error; Step 5: Perform anomaly analysis and online monitoring respectively.

3. The multivariate temporal anomaly detection method as described in claim 2, characterized in that, The SCADA data preprocessing in step one includes: SCADA data was used as experimental data, which was collected by acceleration sensors installed at various parts of the wind turbine generator. Variational mode decomposition (VMD) was used to process the non-stationary SCADA signal to address data noise. The intrinsic mode functions (IMFs) in the frequency domain signal were selected by using the kurtosis index to filter out the noise components in the SCADA data. The second step of constructing the spatial feature reconstruction module and extracting spatially relevant features includes: Multidimensional SCADA data is input into the DSAE depth stacked autoencoder to extract spatial correlation features among multiple variables. (1) Constructing a spatial feature reconstruction module The autoencoder (AE) obtains a compressed representation of the input vector X by encoding it through a hidden layer, and then decodes it back to the original vector. The compressed vector in the intermediate layer is the low-dimensional feature representation z of the original input vector. Using the low-dimensional feature representation z as the input vector, the deep encoder is built by stacking layers to obtain a deeper stacked autoencoder network. The compressed representation is extracted three times to obtain a seven-layer symmetric structure DSAE encoder-decoder network. The principle by which the encoder Encode(x) obtains the compressed representation is shown in the following formula: zi = σ(Wi × xi + bi); Where σ(·) is the activation function ReLU, Wi is the weight parameter of the i-th layer, xi is the input vector of the i-th layer, and bi is the bias term of the i-th layer; The principle by which the decoder Decode(z) recovers the reconstructed vector ri from the compressed vector is shown in the following formula: ； Where σ(·) is the activation function ReLU, is the weight parameter of the decoder corresponding to the compressed vector of the i-th layer, zi is the compressed vector of the i-th layer, and bi is the bias term of the decoder corresponding to the compressed vector of the i-th layer; (2) Extracting low-dimensional feature representation z of spatial correlation Multidimensional SCADA data is input into a DSAE deep stacked autoencoder to extract spatial correlation features between multiple variables. Through compression extraction by the deep autoencoder, a low-dimensional feature representation z of the multidimensional variables is finally obtained. (3) Calculate the reconstructed multidimensional time series data r and the reconstruction error Er. The low-dimensional feature representation z extracted from the intermediate layer is input into the decoder network of the second half of DSAE to obtain the reconstructed multi-dimensional temporal data r. The reconstruction error Er based on the Euclidean distance metric is calculated as follows: ； Where x is the input multidimensional time series data, and r is the reconstructed data.

4. The multivariate temporal anomaly detection method as described in claim 2, characterized in that, Step three, which involves constructing the time series prediction module and extracting global time series dependencies, includes: Train the Transformer prediction subnetwork to extract global temporal dependencies from temporal data; (1) Constructing a Transformer-based prediction subnetwork A Transformer model with an encoder-decoder structure is constructed. The encoder part consists of a set of encoders with a stack of 6 layers. Each encoder has normalization, multi-head attention, residual summation, normalization, and fully connected units. The decoder consists of a set of decoders with the same number as the encoder. By using a self-attention mechanism, the Transformer obtains global feature extraction capabilities. (2) Transformer prediction subnetwork output prediction The low-dimensional feature representation z and reconstruction error Er from the DSAE output are input into the temporal prediction module, where the encoder extracts the global temporal dependency; the decoder outputs the predicted value zp, and the mean squared error (MSE) between the predicted value zp and the true value zt is used as the loss metric Ep, calculated as follows: 。 5. The multivariate temporal anomaly detection method as described in claim 2, characterized in that, The joint training in step four includes: For training the overall fusion framework, a joint training approach is proposed, employing an end-to-end minimization method to simultaneously obtain better spatially compressed feature representations and global temporal dependency information. The joint training consists of two parts: reconstruction error Er and prediction error Ep, with the objective function being: ； Where Er is the reconstruction error, Ep is the prediction error, and λ is a hyperparameter that controls the importance of the two errors; The anomaly analysis in step five includes: Distinguishing between normal and abnormal states requires determining an anomaly threshold (TAD). This threshold is calculated using the training results of the model on the training set data. ； Where E(x) is the loss value of the model's objective function. This represents the mean of the model's loss values; The online monitoring includes: After preprocessing the test data xtest according to step one, the data is input into the already trained model to calculate the corresponding reconstruction error Er, prediction error Ep, and objective function loss value E(xtest). According to the anomaly decision rule, when E(xtest)>TAD, the detection result of the sample is marked as an abnormal state, otherwise it is marked as a normal state.

6. A multivariate temporal anomaly detection system applying the multivariate temporal anomaly detection method as described in any one of claims 1 to 5, characterized in that, The multivariate time series anomaly detection system includes: The data preprocessing module is used to preprocess SCADA data using variational mode decomposition. The spatial feature reconstruction module is used to input multidimensional SCADA data into the DSAE depth stack autoencoder and extract spatial correlation features between multiple variables. The temporal prediction module is used to train the Transformer prediction sub-network and extract global temporal dependencies from temporal data. The joint training module is used to minimize reconstruction error and prediction error through joint training. The anomaly analysis module is used to calculate anomaly thresholds based on the training results of the model using training set data. The online monitoring module is used to determine the test results of samples based on anomaly decision rules.

7. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the multivariate timing anomaly detection method as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, the processor performs the multivariate timing anomaly detection method as described in any one of claims 1 to 5.

9. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the multivariate time series anomaly detection system as described in claim 6.

Images