System and method for detecting abnormality of time sequence data of wind generating set

Abstract

The invention belongs to the technical field of wind generating set operation parameter monitoring, and aims to solve the problem that information propagation time is different due to the fact that the distance of the relation between devices is different. The invention discloses a wind generating set time sequence data anomaly detection system and method. The system comprises a multi-scale generation module, a graph structure self-learning module, a scale adaptive fusion module and a decoding module. The method comprises the steps of generating data representations of different time scales, representing a relationship between sensors in a graph structure form, self-learning a weight coefficient of each scale layer through a neural network training mode, mapping features of a potential space back to an original space through a three-layer full-connection layer, generating reconstruction data, and obtaining the reconstructed data. The loss is calculated through the mean square error of the reconstructed data and the original data, and the model parameters are optimized by using a back propagation algorithm. According to the invention, diversified interaction modes between sensors can be effectively captured, so that the accuracy and robustness of anomaly detection are improved.

Description

Technical Field

[0001] The present application belongs to the technical field of wind turbine operating parameter monitoring, and in particular relates to a system and method for detecting abnormalities in time series data of a wind turbine. Background Art

[0002] Sensors used in smart industry generate large amounts of data with time series characteristics. This time series data can reflect the operating status of industrial equipment. For example, during the operation of a wind turbine, various sensors can collect operational data such as temperature, voltage, and current. By analyzing this sensor data and performing anomaly diagnosis, personnel can quickly locate anomalies and take intervention measures, thereby reducing safety risks and economic losses in the factory. However, relying on manual analysis of time series data generated by sensors not only makes it difficult to accurately detect anomalies, but also consumes a large amount of manpower. Time series data anomaly detection, a key research area in data mining, can apply anomaly detection algorithms to the status monitoring of industrial equipment, enabling intelligent operation and maintenance of equipment. This not only improves factory operating efficiency but also effectively prevents industrial accidents caused by undetected equipment failures.

[0003] In recent years, research on anomaly detection algorithms for time series data has made significant progress. However, when these algorithms are applied to complex industrial environments, the performance of the models often falls short of expectations. The fundamental reason for this performance gap lies in the unique characteristics of industrial time series data. Industrial time series data not only exhibits correlations in the temporal dimension but also interacts in the spatial dimension.

[0004] Currently, deep learning-based time series data anomaly detection algorithms can be divided into two categories: reconstruction-based and prediction-based methods.

[0005] Reconstruction-based methods identify anomalies by reconstructing the original input data and comparing the differences between the reconstructed data and the original input data. This method is based on the assumption that there are significant differences in the data distribution between normal and abnormal samples. By training the model on reconstruction of normal samples, it can reconstruct normal data well. However, when the model is used to process abnormal samples, it often fails to accurately reconstruct them, resulting in large reconstruction errors, which in turn lead to anomaly detection.

[0006] Deep learning models such as autoencoders (AEs), variational autoencoders (VAEs), and generative adversarial networks (GANs) are widely used in reconstruction-based anomaly detection methods. Autoencoders map data into a latent space to capture its intrinsic feature representation and reconstruct the data from this representation. Many derivative models are based on autoencoders, such as the denoising autoencoder (DAE), which trains the autoencoder through local denoising to generate better high-level feature representations. The sparse autoencoder (SAE) introduces a sparsity penalty term to make the activations of hidden layer neurons more sparse, thereby better capturing the important features of the input data. However, autoencoders are susceptible to overfitting problems, especially when the training data contains noise, and their performance is poor.

[0007] Unlike autoencoders (AEs), VAEs and GANs focus not only on the compressed representation of data but also on the distribution of data in the latent space. VAEs improve generalization performance by modeling the underlying probability distribution of data and introducing regularization terms. Although VAEs excel in capturing the underlying distribution of data, their lack of analysis of the time dimension limits their application in anomaly detection in time series data. To this end, subsequent studies have proposed a variety of improved methods, such as LSTM-VAE, which uses long short-term memory networks (LSTMs) in the inference and generation networks of VAEs to capture local features and long-term dependencies of sequences. However, these methods fail to fully consider the differences in the operating cycles of wind turbine equipment when analyzing the change patterns in the time dimension, resulting in insufficient learning of the change patterns in the time dimension.

[0008] Generative Adversarial Networks (GANs), first used in computer vision, have quickly become a research hotspot in deep learning due to their unique architecture and powerful generative capabilities. Using an adversarial structure consisting of a generator and a discriminator, GANs randomly sample from a latent space and generate samples that are highly similar to real data. The generator aims to produce samples that are as close to real data as possible, while the discriminator attempts to distinguish generated samples from real data as closely as possible. BeatGAN uses an autoencoder as the GAN generator, effectively identifying anomalies by comparing the differences between original and generated data.

[0009] Prediction-based methods attempt to analyze and predict the normal development trends of time series based on historical data, calculating anomaly scores by comparing the predicted results with the actual values. These methods primarily rely on recurrent neural networks and their variants, such as LSTM and GRU, to analyze and predict the normal development trends of time series. For example, the LSTM-NDT model uses LSTM to capture changing patterns in the temporal dimension and proposes a dynamic threshold setting method to reduce missed anomalies. However, due to the cyclical characteristics of historical data, which are susceptible to noise, and the unpredictable nature of multidimensional time series, prediction-based methods often suffer from high false positive rates.

[0010] Despite significant progress in anomaly detection in time series data, most research still focuses on analyzing the temporal dimension, neglecting the exploration of spatial relationships. However, in complex industrial systems, monitoring equipment sensors are often closely interconnected, suggesting that the spatial relationships of time series data are equally important. Graph neural networks, due to their outstanding performance in capturing spatial relationships between data points, have attracted widespread attention.

[0011] Graph Neural Network (GNN) has attracted widespread attention due to its outstanding performance in modeling the relationship between entities in complex networks, and has provided a new direction for solving the spatial relationship of sensor data. In 2021, the GDN model proposed by Deng et al. marked an important attempt in this field. The model constructs a graph structure through the cosine similarity between sensor embedding vectors, and uses graph neural network technology for data representation and predictive analysis. Although GDN has achieved certain results, it is only a preliminary exploration of the application of graph neural networks in the field of time series data anomaly detection, and there is still room for improvement. The single graph structure it constructs fails to fully reveal the complex interactions between sensors, and also ignores the importance of the change pattern of the time dimension. Therefore, based on graph neural networks, continuing to explore models that can simultaneously capture and integrate the relationship between the time dimension and spatial dimension of time series data is crucial to improving the detection accuracy and maintenance efficiency of wind power systems.

[0012] In wind power systems, numerous sensors are deployed to monitor the operating status of equipment. Due to factors such as functional dependencies, sensors often have interdependencies. When an anomaly occurs in the system, the network of relationships between sensors can help map and track the propagation path of the anomaly information. Therefore, understanding and leveraging the intersensor correlations is crucial for improving the accuracy of anomaly detection. However, most existing anomaly detection research on time series data often neglects to learn and utilize the complex relationships between sensors, significantly limiting the effectiveness of the models in real-world industrial settings. Summary of the Invention

[0013] The purpose of this application is to provide a system and method for detecting anomalies in time series data of a wind turbine generator set, so as to solve the problem of differences in information propagation time caused by the different degrees of distance between devices, the difficulty of unified modeling and analysis by traditional methods, resulting in insufficient anomaly detection accuracy, and the difficulty in extracting features from time series data generated during the operation of a wind turbine generator set.

[0014] In order to achieve the above objectives, this application provides the following technical solutions:

[0015] In a first aspect, the present application provides a wind turbine generator set time series data anomaly detection system, comprising:

[0016] Multi-scale generation module, which generates data representations at different time scales to support subsequent multi-scale learning;

[0017] The graph structure self-learning module represents the relationship between sensors in the form of a graph structure, with sensors as nodes and the relationships between sensors as edges in the graph;

[0018] The scale-adaptive fusion module self-learns the weight coefficients of each scale layer through neural network training;

[0019] The decoding module maps the features of the latent space back to the original space through three fully connected layers to generate reconstructed data. The loss is calculated by the mean square error between the reconstructed data and the original data, and the back propagation algorithm is used to optimize the model parameters.

[0020] In some embodiments, the multi-scale generation module uses a one-dimensional convolutional network to perform convolution operations on the original data in the time dimension to generate a multi-scale feature representation of the original data. Each scale layer takes the output of the previous scale layer as input and generates a feature representation of a larger scale.

[0021] In some embodiments, the original time series data X∈R N×T , N represents the number of sensors, T represents the length of the time dimension; a one-dimensional convolutional network is used to process the data of each scale layer, different scale layers use different convolution kernel sizes, the convolution step is set to 2, and the data of the kth scale layer The calculation formula is as follows:

[0022]

[0023] Where k represents the index of the scale layer, represents the convolution operator, W k and b k They represent the convolution kernel and bias vector of the kth layer respectively.

[0024] In some embodiments, the graph structure self-learning module measures the closeness of the relationship between sensors by using the cosine similarity between sensor embedding vectors, and represents their topological relationship in the form of a graph structure.

[0025] In some embodiments, in the graph structure self-learning module, the graph structure is represented by an adjacency matrix A. ij Indicates whether there is an edge between sensor node i and sensor node j;

[0026] In the initial stage, the graph structure is set to a fully connected form. For sensor i, its neighbor node set is

[0027] In the graph structure self-learning module, an embedding vector is set for the sensor. The sensor embedding vector represents V i ∈R d ,i∈{1,2,...,N}, N represents the number of sensors, d represents the dimension of the embedding vector, and i represents the index of the sensor;

[0028] If the relationship between sensor i and sensor j is close, there is an edge between node i and node j in the graph structure, otherwise, there is no edge between them;

[0029] The degree of relationship between sensors is calculated by the cosine similarity between sensor embedding vectors. The cosine similarity between sensor i and sensor j can be expressed as e ij , and its calculation formula is as follows:

[0030]

[0031] Where V i and V j They represent the embedding vector of sensor i and the embedding vector of sensor j respectively.

[0032] In some embodiments, for sensor node i, the nodes with the top k cosine similarities are selected as its neighbor nodes, A ij The value of is shown as follows:

[0033]

[0034] The learning of graph structure is converted into the learning of sensor embedding vectors, which are integrated into the graph attention network for training. Then, the sensor embedding vectors and the time series data generated by the sensor are concatenated to obtain g i , as shown below:

[0035]

[0036] Where, represents the sequence fragment of sensor i at time t, Indicates series operation, V i represents the embedding vector of sensor i, and W represents the trainable weight matrix.

[0037] In some embodiments, the attention coefficients α of sensor i and sensor j are i,j The calculation is as follows:

[0038]

[0039] Where a represents the learning coefficient vector of the attention mechanism.

[0040] In some embodiments, the scale-adaptive fusion module fuses information of each scale layer by weighted fusion, so that the model pays different attention to different scale layers.

[0041] In some embodiments, the scale adaptive fusion module concatenates the feature vectors output by each scale layer and performs average pooling processing on the scale dimension. The output of each scale layer is recorded as h 1 ,...,h k ,...,h K , where h k ∈R N ×dim , dim represents the dimension of the latent vector, as shown in the following formula:

[0042] h cat =Concate(h 1 ,h 2 ...h K )∈R K×N×dim

[0043]

[0044] Use two fully connected layers to learn the weight factors α corresponding to each scale layer, α∈R K , which is calculated as follows:

[0045] f1=ReLU(W1h pool +b1)

[0046] α=Sigmoid(W2f1+b2)

[0047] Where W1 and W2 are the learnable weight matrices in the two fully connected layers, and b1 and b2 are the corresponding bias vectors;

[0048] Perform weighted sum operation on each scale layer to obtain the final scale fusion feature vector H, H∈R N×dim , which is calculated as shown in the formula:

[0049]

[0050] In a second aspect, the present application provides a method for detecting anomalies in time series data of a wind turbine generator set, comprising:

[0051] A one-dimensional convolutional network is used to perform convolution operations on the original data in the time dimension to generate multi-scale feature representations of the original data. Each scale layer takes the output of the previous scale layer as input and generates a larger scale feature representation.

[0052] The relationship between sensors is represented in the form of a graph structure, with sensors as nodes and relationships between sensors as edges. The closeness of the relationship between sensors is measured by the cosine similarity between sensor embedding vectors, and their topological relationship is represented in the form of a graph structure.

[0053] Through neural network training, the weight coefficients of each scale layer are self-learned; through weighted fusion, the information of each scale layer is integrated, so that the model pays different attention to different scale layers.

[0054] The features of the latent space are mapped back to the original space through three fully connected layers to generate reconstructed data. The loss is calculated by the mean square error between the reconstructed data and the original data, and the back propagation algorithm is used to optimize the model parameters.

[0055] Compared with the existing technology, the wind turbine generator set time series data anomaly detection system and method provided by this application has the following beneficial effects:

[0056] This application analyzes data through multi-scale processing in the time dimension. Processing data at a larger time scale helps eliminate redundant information introduced by time delays, thereby making the abnormal propagation path clearer and helping the model better capture and understand the relatively distant relationships between devices. In contrast, a smaller time scale can accurately capture spatial associations with short time delays and close relationships between devices. Therefore, learning spatial relationships at different time scales can effectively capture diverse interaction patterns between sensors, thereby improving the accuracy and robustness of anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solution of this application, the following is a brief introduction to the drawings required for the technical description.

[0058] Figure 1 A schematic diagram of the structure of the wind turbine generator set time series data anomaly detection system provided in this application;

[0059] Figure 2 The overall framework diagram of the wind turbine generator set time series data anomaly detection system provided in this application;

[0060] Figure 3 This is a flow chart of the method for detecting abnormality in time series data of a wind turbine generator set provided in this application. DETAILED DESCRIPTION

[0061] The following is further explained in detail through specific implementation methods.

[0062] This application sets embedding vectors for sensors to represent the diversity of device sensors. It calculates the cosine similarity between sensor embedding vectors to represent the closeness of the relationship between sensors. It also uses multi-scale processing operations to generate multi-scale representations of the raw data, thereby learning the spatial relationships of time series data at different time scales. Furthermore, considering the different contributions of each scale layer to the detection results, a method is proposed to determine the weight factors of each scale layer using neural network training.

[0063] This application proposes a Multi-Scale Graph Structure Self-Learning Time Series Anomaly Detection (MGSS) model. The model aims to learn the relationship between sensors at different time scales and aggregate information between itself and neighboring nodes through a graph attention network. It deeply explores the rich and implicit spatial relationships at multiple time scales, thereby improving the accuracy and applicability of anomaly detection.

[0064] The multi-scale graph structure self-learning time series anomaly detection model is a reconstruction-based unsupervised time series anomaly detection model. By reconstructing normal data, the model learns the encoding representation of normal data in the latent space. When encountering abnormal data, the model cannot effectively reconstruct the abnormal data, and detects the anomaly based on the deviation between the reconstructed data and the input data. Considering the complex relationship between sensors in the operating environment of wind turbines, which plays a certain auxiliary role in improving the detection rate of abnormal events, MGSS uses a graph attention network to learn the relationship between sensors at different time scales. The multi-scale learning method enables the model to more flexibly capture the interaction relationship between sensors. It consists of four modules: a multi-scale generation module, a graph structure self-learning module, a scale-adaptive fusion module, and a decoding module.

[0065] First, as Figure 1 and Figure 2 As shown, the present application provides a wind turbine generator set time series data anomaly detection system, which is a wind turbine generator set time series data anomaly detection system based on multi-scale graph structure self-learning, including:

[0066] Multi-scale generation module, which generates data representations at different time scales to support subsequent multi-scale learning;

[0067] The graph structure self-learning module represents the relationship between sensors in the form of a graph structure, with sensors as nodes and the relationships between sensors as edges. The cosine similarity between sensor embedding vectors is used to measure the closeness of the relationship between sensors, and their complex topological relationships are represented in the form of a graph structure.

[0068] The scale-adaptive fusion module self-learns the weight coefficients of each scale layer through neural network training. It fuses the information of each scale layer through weighted fusion, so that the model pays different attention to different scale layers.

[0069] The decoding module maps the features of the latent space back to the original space through three fully connected layers to generate reconstructed data. The loss is calculated by the mean square error between the reconstructed data and the original data, and the back propagation algorithm is used to optimize the model parameters.

[0070] The multi-scale generation module generates data representations at different time scales from raw time series data, providing support for subsequent multi-scale learning. The multi-scale generation module processes raw time series data and generates data representations at different time scales. This multi-scale data representation helps capture the operational characteristics of wind turbines over different time periods, providing more comprehensive and in-depth data support for subsequent multi-scale learning.

[0071] Specifically, wind turbine operating data is often complex and nonlinear, and often exhibits different characteristics at different time scales. For example, on short time scales, the data may exhibit high-frequency fluctuations, while on long time scales, the data may show more stable trends. The multi-scale generation module generates data representations that reflect these characteristics at different time scales by filtering and downsampling the raw data. This multi-scale data processing and learning approach improves the accuracy and efficiency of anomaly detection and ensures stable operation and fault prediction of wind turbines.

[0072] The graph-based self-learning module introduces sensor embedding vectors to represent the diversity of sensor devices in the wind power system. The cosine similarity between sensor embedding vectors is used to measure the closeness of relationships between sensors, representing their complex topological relationships in the form of a graph. Furthermore, a graph attention network is used to weightedly aggregate information about the node itself and its neighbors.

[0073] In view of the different contributions of features at different scales to the results, the model adaptively learns the weight factors of each scale through a neural network, and performs weighted fusion of feature representations at different scales, thereby dynamically adjusting the impact of each scale on the results.

[0074] The decoding module maps the features of the latent space back to the original space through three fully connected layers to generate reconstructed data. Finally, the loss is calculated by the mean squared error between the reconstructed data and the original data, and the backpropagation algorithm is used to optimize the model parameters.

[0075] Specifically, the multi-scale generation module uses a one-dimensional convolutional network to perform convolution operations on the raw data in the time dimension, generating a multi-scale feature representation of the original data. Each scale layer takes the output of the previous scale layer as input and generates a feature representation at a larger scale. This pyramid-level structure facilitates the subsequent extraction of spatial features from time series data at different time scales.

[0076] Original time series data X∈R N×T , N represents the number of sensors, and T represents the length of the time dimension. A one-dimensional convolutional network is used to process the data of each scale layer. Different scale layers use different convolution kernel sizes, and the convolution step size is uniformly set to 2 to increase the time scale. The data of the kth scale layer The calculation is shown in the formula:

[0077]

[0078] Among them, k represents the index of the scale layer, represents the convolution operator, W k and b k They represent the convolution kernel and bias vector of the kth layer respectively.

[0079] By stacking multiple scale layers and gradually reducing the convolution kernel size of each scale layer, a pyramid structure can be constructed to generate a multi-scale representation of the original data.

[0080] This application uses the spatial relationship between sensors. The graph structure self-learning module represents the relationship between sensors in the form of a graph structure, with sensors as nodes and the relationship between sensors as edges in the graph. The graph structure is represented by the adjacency matrix A. ij Indicates whether there is an edge between sensor node i and sensor node j. In the initial stage, without prior knowledge of sensor relationships, the graph structure is set to a fully connected form. For sensor i, its neighbor node set That is, the neighbor nodes of sensor i are all the remaining nodes except itself.

[0081] In view of the diversity of sensors, embedding vectors are set for sensors in the graph structure self-learning module to fully demonstrate their diverse characteristics. The sensor embedding vector represents V i ∈R d,i∈{1,2,...,N}, N represents the number of sensors, d represents the dimension of the embedding vector, and i represents the index of the sensor.

[0082] If the relationship between sensor i and sensor j is close, there is an edge between nodes i and j in the graph structure, otherwise there is no edge between them. The closeness of the relationship between sensors is calculated by the cosine similarity between the sensor embedding vectors. The cosine similarity between sensor i and sensor j can be expressed as e ij , its calculation formula is as follows:

[0083]

[0084] Among them, V i and V j Denote the embedding vector of sensor i and the embedding vector of sensor j respectively. For sensor node i, the nodes with the top k similarities to its cosine are selected as its neighbor nodes, i.e. A ij The value of is shown in the formula:

[0085]

[0086] This converts the learning of graph structure into the learning of sensor embedding vectors. Next, the sensor embedding vectors are integrated into the graph attention network for training. First, the sensor embedding vectors and the time series data generated by the sensor are concatenated to obtain g i , as shown in the formula:

[0087]

[0088] in, represents the sequence fragment of sensor i at time t, Indicates series operation, V i represents the embedding vector of sensor i, W represents the trainable weight matrix, and the attention coefficient α of sensor i and sensor j i,j The calculation is shown in the formula:

[0089]

[0090] Where a represents the learning coefficient vector of the attention mechanism, π(i, j) represents the similarity score between sensors i and j, LeakyReLU is the activation function, and α i,j represents the normalized attention coefficient, exp() represents the exponential function, and k∈C(i)Y{i} represents the set of neighbor nodes of sensor i.

[0091] Aggregate representation of the i-th sensor at time t The calculation is shown in the formula:

[0092]

[0093] Where ReLU is the activation function.

[0094] By integrating sensor embedding vectors into the graph attention network, the sensor embedding vectors are continuously adjusted during the network training process, thereby achieving self-learning of the graph structure.

[0095] This application takes into account the different contributions of different scale layers to the final result. The scale adaptive fusion module self-learns the weight coefficients of each scale layer through neural network training. By fusing the information of each scale layer through weighted fusion, the model pays different attention to different scale layers. First, the feature vectors output by each scale layer are concatenated together and average pooled in the scale dimension. The output of each scale layer is recorded as h 1 ,...,h k ,...,h K , where h k ∈R N×dim , dim represents the dimension of the latent vector, as shown in the formula:

[0096] h cat =Concate(h 1 ,h 2 ...h K )∈R K×N×dim

[0097]

[0098] Next, two fully connected layers are used to learn the weight factors α, α∈R corresponding to each scale layer K , which is calculated as shown in the formula:

[0099] f1=ReLU(W1h pool +b1)

[0100] α=Sigmoid(W2f1+b2)

[0101] Among them, W1 and W2 are the learnable weight matrices in the two fully connected layers, and b1 and b2 are the corresponding bias vectors. Finally, the weighted sum operation of each scale layer is performed to obtain the final scale fusion feature vector H, H∈R N×dim , which is calculated as shown in the formula:

[0102]

[0103] The decoding module in this application sends the scale fusion feature vector H to the three-layer fully connected layer to reconstruct The calculation is shown in the formula:

[0104] out1=ReLU(W2(ReLU(W1H+b1))+b2)

[0105]

[0106] in, w represents the length of the sliding window, W1, W2, and W3 are the learnable weight matrices in the three fully connected layers, and b1, b2, and b3 are the corresponding bias vectors.

[0107] The model optimizes the model parameters by minimizing the mean square error between the reconstructed data and the true observed data. MSE As shown in the formula:

[0108]

[0109] Where T′=T-w+1, which represents the number of time series segments obtained after the sliding window operation. S={s1,s2,...,s T′}∈R T′×w×n represents a set of input time series segments, Represents the set of time series segments after model reconstruction.

[0110] The anomaly score is determined by the reconstruction error at a certain moment. The anomaly time point is determined by comparing the anomaly score with the threshold. The anomaly score is calculated as shown in the formula:

[0111]

[0112] Among them, x t represents the real data at time t, Represents the data corresponding to time t reconstructed by the model.

[0113] In order to select a suitable threshold, the present invention adopts a grid search method to automatically select the threshold to ensure the model

[0114] Optimal anomaly detection performance is achieved. Specifically, the anomaly scores calculated at all time points in the test dataset are used as a potential threshold candidate set. Then, within this candidate set, the F1 score of the model is calculated for each candidate threshold. Finally, the F1 scores of all candidate thresholds are compared, and the threshold that maximizes the F1 score is selected as the optimal anomaly detection threshold, denoted as ε. Thereafter, for each time point t in the test dataset, if its corresponding anomaly score exceeds ε, the moment is considered an anomaly; otherwise, it is considered normal. This is shown in the formula:

[0115]

[0116] Among them, yt Represents the abnormal label value at time t output by the model.

[0117] In the second aspect, based on the above-mentioned wind turbine generator set time series data anomaly detection system, the present application also provides a wind turbine generator set time series data anomaly detection method, such as Figure 3 Shown, including:

[0118] Generate data representations at different time scales to support subsequent multi-scale learning; use a one-dimensional convolutional network to perform convolution operations on the original data in the time dimension to generate multi-scale feature representations of the original data. Each scale layer takes the output of the previous scale layer as input and generates feature representations at a larger scale.

[0119] The relationship between sensors is represented in the form of a graph structure, with sensors as nodes and relationships between sensors as edges. The cosine similarity between sensor embedding vectors is used to measure the closeness of the relationship between sensors, and their complex topological relationships are represented in the form of a graph structure.

[0120] Through neural network training, the weight coefficients of each scale layer are self-learned; through weighted fusion, the information of each scale layer is integrated, so that the model pays different attention to different scale layers.

[0121] The features of the latent space are mapped back to the original space through three fully connected layers to generate reconstructed data. The loss is calculated by the mean square error between the reconstructed data and the original data, and the back propagation algorithm is used to optimize the model parameters.

[0122] It should be noted that the specific implementation principles or specific implementation details of each step of the above-mentioned wind turbine generator set time series data anomaly detection method can be found in the corresponding content of the above-mentioned wind turbine generator set time series data anomaly detection system, which will not be repeated here.

[0123] Example

[0124] The multi-scale graph structure self-learning time series anomaly detection model belongs to the reconstruction-based unsupervised time series anomaly detection model. By reconstructing normal data, the model learns the encoding representation of normal data in the latent space. When encountering abnormal data, the model cannot effectively reconstruct the abnormal data, and detects the anomaly based on the deviation between the reconstructed data and the input data. Taking into account that the intricate relationship between sensors in the operating environment of wind turbines has a certain auxiliary effect on improving the detection rate of abnormal events, MGSS uses graph attention networks to learn the relationship between sensors at different time scales. The multi-scale learning method enables the model to more flexibly capture the interaction relationship between sensors. It consists of four major modules, namely multi-scale generation module, graph structure self-learning module, scale adaptive fusion module and decoding module. The schematic diagram of the MGSS model structure is shown below. Figure 2 shown.

[0125] First, the multi-scale generation module generates data representations at different time scales from raw time series data, supporting subsequent multi-scale learning. Next, the graph-structured self-learning module introduces sensor embedding vectors to represent the diversity of sensor devices in the wind power system. The cosine similarity between sensor embedding vectors is used to measure the closeness of relationships between sensors, representing their complex topological relationships in a graph structure. Simultaneously, a graph attention network is used to weightedly aggregate information from the node itself and its neighbors. Then, based on the varying contributions of features at different scales to the results, the model adaptively learns weight factors for each scale through a neural network, combining feature representations at different scales to dynamically adjust the impact of each scale on the results. Next, the decoding module maps the features from the latent space back to the original space through three fully connected layers, generating reconstructed data. Finally, the loss is calculated by comparing the mean squared error between the reconstructed and original data, and the model parameters are optimized using a backpropagation algorithm.

[0126] The experiment is as follows:

[0127] The following experiments are conducted to verify and analyze the effectiveness of the wind turbine generator set time series data anomaly detection system and method based on multi-scale graph structure self-learning proposed in the present invention.

[0128] (1) Experimental Dataset

[0129] This paper uses four public datasets to evaluate the performance of the model: the SWaT dataset, the WADI dataset, the GECCO dataset, and the MSL dataset. The detailed information of these four datasets is shown in the table.

[0130]

[0131]

[0132] SWaT (Secure Water Treatment) dataset. The SWaT dataset, published by the Public Utilities Board of Singapore, records data from a tap water treatment testbed. Water treatment operations consist of six steps: raw water storage (P1), pretreatment (P2), ultrafiltration (P3), UV dechlorination (P4), reverse osmosis (P5), and treated water storage (P6). Each step has different sensors, such as pressure sensors, temperature sensors, and liquid level sensors. The testbed uses a total of 51 sensors to record data values ​​at each time point. The dataset covers 11 consecutive days of time series data from the water treatment process. The first seven days recorded data under normal operation, while the next four days simulated 36 attack scenarios, including abnormal data from attacks at different time intervals. Therefore, the first seven days of data were used as the training set, and the last four days of data were used as the test set.

[0133] WADI (Water Distribution) dataset. The WADI dataset is also released by the Public Utilities Board of Singapore. WADI simulates a water supply process consisting of a large number of water supply pipes. Compared with SWaT, WADI is larger in scale and the relationship between sensors is more complex. It integrates water treatment, storage, and distribution functions. WADI consists of three main stages: water supply (P1), distribution network (P2), and return water system (P3). Each stage uses 123 sensors to record data. This dataset collects time series data generated by the water supply process over 16 days. The system operated normally in the first 14 days, while 15 attacks at different time intervals were simulated in the last 2 days. Therefore, the data from the first 14 days are used as the training set, and the data from the last 2 days are used as the test set data.

[0134] The GECCO dataset is a drinking water quality test dataset provided by the GECCO industrial challenge organized by the Computer Association. It contains water quality data and operational data. The water quality data includes chlorine dioxide content, pH value, and conductivity in water, and the operational data includes water flow rate and temperature. A total of nine sensor data are collected.

[0135] The MSL (Mars Science Laboratory) dataset is annotated and publicly available by NASA. This dataset consists of scientific data collected by the Mars rover during its exploration mission on the Martian surface. It includes various sensor data such as temperature, air pressure, and light intensity. It covers a total of 27 entities, each containing 55 monitoring indicators.

[0136] (2) Data preprocessing

[0137] Since different sensors have different numerical ranges, in order to fairly evaluate the data of different sensors, the time series data needs to be normalized before the experiment. The present invention uses the maximum and minimum values ​​of each sensor data in the training data set to normalize the time series data to eliminate the deviation that may be introduced due to dimensional differences between different sensor data. Through the normalization operation, all sensor data will be converted to the same numerical range, making the model training process more efficient and fair. The normalization operation is shown in the formula:

[0138]

[0139] Among them, X (d) represents the normalized value of the data generated by the dth sensor, X train Represents the training data set, min(X train (d)) represents the minimum value of the data generated by the d-th sensor in the training set, max(X train (d) ) represents the maximum value of the data generated by the d-th sensor in the training set. In particular, if the maximum and minimum values ​​of the data generated by the d-th sensor in the training data set are equal, in order to avoid the denominator being zero, X (d) Set to 0.5.

[0140] Since time series have dependencies in the time dimension, when performing anomaly detection on time series data, it is necessary to use a sliding window to process the original data into time-windowed sequence segment data.

[0141] For any given timestamp t and its corresponding data vector x t For example, we can define a time window sequence s with a length of w t ={x t-w+1 ,x t-w+2 ,...,x t}∈R w×N , s t Represents a continuous sequence segment from timestamp t-w+1 to timestamp t, where N is the number of sensors. By processing the original input data with a fixed window size w and moving the time window forward by one time unit each time, a series of time series segments S = {s1, s2, ..., s T′}∈R T ′×w×N , where T′=T-w+1 represents the total number of sequence segments generated by the sliding window method. The input of the model is the set of time series segments S, and the output is the abnormal label sequence Y={y1,y2,...,y T′}.

[0142] (3) Evaluation indicators

[0143] Anomaly detection in time series data aims to identify whether an anomaly has occurred at a specific point in time. This is a typical binary classification problem. In industrial applications, because anomalous data is difficult to collect, the number of anomalous samples is far less than that of normal samples, resulting in significant dataset imbalance. Therefore, anomalous samples are typically considered positive, while normal samples are considered negative. The confusion matrix for this binary classification problem is shown in the table below.

[0144]

[0145] TP (True Positive) represents samples that are actually positive and correctly detected as positive; TN (True Negative) represents samples that are actually negative and correctly detected as negative; FP (False Positive) represents samples that are actually negative but mistakenly detected as positive; FN (False Negative) represents samples that are actually positive but mistakenly detected as negative.

[0146] The present invention uses three indicators based on the confusion matrix: precision, recall and F1 score to evaluate the performance of the time series data anomaly detection model.

[0147] Precision, also known as the call accuracy rate, indicates the proportion of true anomalies among the anomalies detected by the model. The higher the precision, the lower the likelihood that the model will falsely report anomalies. Its calculation formula is as follows:

[0148]

[0149] Recall, also known as recall rate, represents the proportion of anomalies detected by the model among all true anomalies. The higher the recall rate, the lower the probability that the model will miss anomalies. Its calculation formula is as follows:

[0150]

[0151] The F1 score combines precision and recall, taking into account both missed and false positives, and is a key indicator for evaluating the performance of anomaly detection in time series data. Its calculation formula is as follows:

[0152]

[0153] (4) Experimental results

[0154] To evaluate the performance of the MGSS model in time series data anomaly detection, we compared it with six recently proposed unsupervised time series data anomaly detection models on four public datasets. These models include LSTM-VAE, MAD-GAN, MSCRED, MTAD-GAT, GDN, and MAD-SGCN. The experimental results are shown in the table, with the best precision, recall, and F1 score indicated in bold.

[0155] Experimental results show that the MGSS model achieves the highest F1 scores on all datasets. Specifically, the MGSS model achieves F1 scores of 0.853, 0.589, 0.577, and 0.885 on the SWaT, WADI, GECCO, and MSL datasets, respectively. Compared to the best F1 scores of the baseline model, these scores are 5.17%, 2.79%, 9.48%, and 0.24%, respectively. This fully demonstrates the effectiveness of the MGSS method in time series data anomaly detection tasks.

[0156] Four baseline models—MSCRED, MTAD-GAT, GDN, and MAD-SGCN—all consider learning spatial relationships in time series data. Experimental results show that the MGSS model achieves at least a 7.83% improvement in F1 scores over MSCRED and at least a 6.09% improvement over MTAD-GAT on all four datasets. MSCRED constructs a multi-scale feature matrix to represent the system state at different moments and calculates the feature matrix based on the similarity between sensor data to learn the spatial relationships in time series data. However, due to the diverse functions and data formats of sensors, relying solely on similarity cannot accurately represent the complex and diverse relationships between sensors. This results in the MSCRED model's inability to learn spatial relationships, and its performance still needs improvement. MTAD-GAT utilizes graph neural networks to learn the temporal and spatial relationships in time series data. However, its use of a fully connected graph structure to represent the topological relationships between sensors fails to reflect the true relationships between sensors. In reality, not all sensor devices in a wind power system are connected. Therefore, MTAD-GAT's fully connected graph structure is not suitable for the actual wind power system, resulting in poor performance in real-world wind turbine operating environments.

[0157] Across four datasets, MGSS's F1 score improved by at least 0.24% compared to GDN and by at least 2.97% compared to MAD-SGCN. GDN leverages graph neural networks to capture spatial relationships through self-learning of graph structures, while MAD-SGCN employs a self-supervised graph structure learning mechanism. However, both GDN and MAD-SGCN employ a single graph structure design strategy. While this simplifies model construction, it also limits their ability to capture complex sensor network topological relationships, resulting in poor performance on real-world industrial datasets. It is worth noting that MGSS's accuracy is slightly lower than GDN's on some datasets. Accuracy reflects the model's false alarm rate, and MGSS's false alarms may be due to its use of an autoencoding structure, which is susceptible to noise interference, leading to false alarm anomalies.

[0158] Furthermore, MGSS achieves a 9.48% improvement in F1 score compared to LSTM-VAE and MAD-GAN, two other reconstruction-based methods. This is because MGSS explicitly models spatial relationships in its model design, while LSTM-VAE and MAD-GAN fail to incorporate this key factor, resulting in their weaker performance.

[0159]

[0160]

[0161] The above description is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed in this application should be covered by the scope of protection of the present application.

Claims

1. A wind turbine generator set time series data anomaly detection system, characterized in that: include: Multi-scale generation module, which generates data representations at different time scales to support subsequent multi-scale learning; The graph structure self-learning module represents the relationship between sensors in the form of a graph structure, with sensors as nodes and the relationships between sensors as edges in the graph; The scale-adaptive fusion module self-learns the weight coefficients of each scale layer through neural network training; The decoding module maps the features of the latent space back to the original space through three fully connected layers to generate reconstructed data. The loss is calculated by the mean square error between the reconstructed data and the original data, and the back propagation algorithm is used to optimize the model parameters.

2. The wind turbine generator set time series data anomaly detection system according to claim 1, characterized in that: The multi-scale generation module uses a one-dimensional convolutional network to perform convolution operations on the original data in the time dimension to generate multi-scale feature representations of the original data. Each scale layer takes the output of the previous scale layer as input and generates a feature representation of a larger scale.

3. The wind turbine generator set time series data anomaly detection system according to claim 2, characterized in that: Original time series data X∈R N×T , N represents the number of sensors, T represents the length of the time dimension; a one-dimensional convolutional network is used to process the data of each scale layer, different scale layers use different convolution kernel sizes, the convolution step is set to 2, and the data of the kth scale layer The calculation formula is as follows: Where k represents the index of the scale layer, represents the convolution operator, W k and b k They represent the convolution kernel and bias vector of the kth layer respectively.

4. The wind turbine generator set time series data anomaly detection system according to claim 1, characterized in that: The graph structure self-learning module measures the closeness of the relationship between sensors through the cosine similarity between sensor embedding vectors and represents their topological relationship in the form of a graph structure.

5. The wind turbine generator set time series data anomaly detection system according to claim 1 or 4, characterized in that: In the graph structure self-learning module, the graph structure is represented by the adjacency matrix A. ij Indicates whether there is an edge between sensor node i and sensor node j; In the initial stage, the graph structure is set to a fully connected form. For sensor i, its neighbor node set is In the graph structure self-learning module, an embedding vector is set for the sensor. The sensor embedding vector represents V i ∈R d ,i∈{1,2,...,N}, N represents the number of sensors, d represents the dimension of the embedding vector, and i represents the index of the sensor; If the relationship between sensor i and sensor j is close, there is an edge between node i and node j in the graph structure, otherwise, there is no edge between them; The degree of relationship between sensors is calculated by the cosine similarity between sensor embedding vectors. The cosine similarity between sensor i and sensor j can be expressed as e ij , and its calculation formula is as follows: Among them, V i and V j They represent the embedding vector of sensor i and the embedding vector of sensor j respectively.

6. The wind turbine generator set time series data anomaly detection system according to claim 5, characterized in that: For sensor node i, select the nodes with the top k similarities to its cosine as its neighbor nodes, A ij The value of is shown as follows: The learning of graph structure is converted into the learning of sensor embedding vectors, which are integrated into the graph attention network for training. Then, the sensor embedding vectors and the time series data generated by the sensor are concatenated to obtain g i , as shown below: Where, represents the sequence fragment of sensor i at time t, Indicates series operation, V i represents the embedding vector of sensor i, and W represents the trainable weight matrix.

7. The wind turbine generator set time series data anomaly detection system according to claim 6, characterized in that: The attention coefficient α of sensor i and sensor j i,j The calculation is as follows: Where a represents the learning coefficient vector of the attention mechanism.

8. The wind turbine generator set time series data anomaly detection system according to claim 1, characterized in that: The scale-adaptive fusion module fuses the information of each scale layer through weighted fusion, so that the model pays different attention to different scale layers.

9. The wind turbine generator set time series data anomaly detection system according to claim 1 or 8, characterized in that: The scale adaptive fusion module concatenates the feature vectors output by each scale layer and performs average pooling processing on the scale dimension. The output of each scale layer is recorded as h 1 ,...,h k ,...,h K , where h k ∈R N×dim , dim represents the dimension of the latent vector, as shown in the following formula: h cat =Concate(h 1 ,h 2 ...h K )∈R K×N×dim Use two fully connected layers to learn the weight factors α corresponding to each scale layer, α∈R K , which is calculated as follows: f1=ReLU(W1h pool +b1) α=Sigmoid(W2f1+b2) Where W1 and W2 are the learnable weight matrices in the two fully connected layers, and b1 and b2 are the corresponding bias vectors; Perform weighted sum operation on each scale layer to obtain the final scale fusion feature vector H, H∈R N×dim , which is calculated as shown in the formula:

10. A method for detecting abnormality in time series data of a wind turbine generator set, characterized in that: include: A one-dimensional convolutional network is used to perform convolution operations on the original data in the time dimension to generate multi-scale feature representations of the original data. Each scale layer takes the output of the previous scale layer as input and generates a larger scale feature representation. The relationship between sensors is represented in the form of a graph structure, with sensors as nodes and relationships between sensors as edges. The closeness of the relationship between sensors is measured by the cosine similarity between sensor embedding vectors, and their topological relationship is represented in the form of a graph structure. Through neural network training, the weight coefficients of each scale layer are self-learned; through weighted fusion, the information of each scale layer is integrated, so that the model pays different attention to different scale layers. The features of the latent space are mapped back to the original space through three fully connected layers to generate reconstructed data. The loss is calculated by the mean square error between the reconstructed data and the original data, and the back propagation algorithm is used to optimize the model parameters.

Images