Time sequence diagram attention icing detection model and method based on multi-scale wavelet decomposition

Abstract

The invention discloses a time sequence diagram attention icing detection model and method based on multi-scale wavelet decomposition, and relates to the technical field of online monitoring of fan blade icing. The problem that various dynamic changes of blade icing cannot be captured and adapted due to the fact that an existing deep learning algorithm for detecting wind turbine blade icing does not consider correlation between variables specific to the scales between different scales is solved. The detection model adopts a multilayer wavelet decomposition network module to extract detail coefficients and approximation coefficients of different scales of data to obtain multi-scale wavelet input; adopting a scale correlation learning module to construct an adjacent matrix according to the detail coefficients of different scales; a sequence diagram attention module is adopted to convert a diagram structure formed by the multi-scale wavelet input and the adjacent matrix into scale specific representation; a scale attention module is adopted to convert the input of the scale specific representation into scale attention representation and send the scale attention representation to an icing classifier for detection; the invention further provides a corresponding icing detection method. The method is suitable for on-line monitoring of icing of the fan blade.

Description

Technical Field

[0001] The invention belongs to the technical field of fan blade fault detection, and in particular relates to the technical field of online monitoring of fan blade icing. Background Art

[0002] With the continuous development of my country's economy and society, the consumption of electricity by residents has been rising year by year, people's concerns about environmental pollution and global warming have been increasing year by year, and the demand for the adoption of renewable energy and green energy has been growing rapidly. Therefore, wind energy, as one of the cleanest energy sources, has attracted more and more attention. Install more wind turbines in areas with high wind speed and dense cold air, such as high-altitude mountainous areas. The wind resources in these areas are conducive to increasing wind power generation. However, due to the cold environment, especially in winter, blade icing is common in these places. During the icing process, supercooled water droplets in the air collide with the incoming flow and hit the surface of the blade. When the temperature and liquid water content are low, the water droplets freeze immediately. However, when the temperature and liquid water content are high, only a part of the droplets freeze when they collide, and the rest of the droplets form a water film on the blade surface and eventually freeze, leading to safety and efficiency problems of wind turbines. Therefore, it is of great significance to detect icing on wind turbine blades.

[0003] Traditionally, applied physics researchers have addressed this challenge by designing and installing new physical detectors based on the physical properties of ice, such as ultrasonic damping, thermal infrared radiation measurement, ultrasonic guided waves, etc. These methods are quite reliable when used directly to measure the physical properties of ice. However, they may have a negative impact on the structure of the blade, causing expensive human and material resources, and they are insensitive to small amounts of ice accumulation. Therefore, in practice, engineers use indirect methods to assess blade icing conditions, such as monitoring power curves, anemometers, vibration time waveforms, dew points, and temperatures to determine whether to trigger deicing procedures. However, these methods may produce many incorrect estimates due to internal instabilities in the sensors. In addition, these methods are sometimes ineffective in detecting blade icing and are also insensitive to small levels of ice accumulation.

[0004] Methods based on wind turbine monitoring signals can be divided into two types: model-driven and data-driven. Model-based methods rely on high-level domain knowledge to build mathematical models from multiple perspectives to reflect the correlation between signals and blade icing conditions, such as rotor angular velocity, power curve, and thermodynamics. The limitations of model-based methods include over-reliance on prior domain knowledge and lack of generality. In addition, expensive specialized tools such as wind tunnels are required to fine-tune model parameters, and the accuracy of the model will be significantly reduced when the operating environment changes. Data-driven machine learning methods, including shallow machine learning and deep learning-based methods, appear to be a more powerful means of blade icing detection because of the large amount of available historical data collected by the Supervisory Control and Data Acquisition (SCADA) system. SCADA is a powerful data acquisition and monitoring system and is the most commonly used system in fault diagnosis of wind turbine components. The SCADA system can not only collect operating parameters and signals of wind turbines, but also environmental parameters such as temperature, wind speed, and wind direction, provide connection interfaces for various sensors across wind turbines, perform various monitoring and control operations, and fully understand the status of wind turbines. Since using SCADA data for fault detection is a potential low-cost option that does not require additional sensors, many methods based on this have been developed in recent years. Therefore, the SCADA system can manage the installed common sensors that monitor the operation of wind turbines and provide multivariate time series data of the real-time operating environment for data-driven machine learning models, such as linear discriminant, KNN, SVM, and some hybrid methods. However, shallow machine learning methods heavily rely on hand-crafted feature design, which is time-consuming and expensive, while deep learning methods do not require so much manual feature extraction, which has attracted great interest from researchers.

[0005] Although many deep learning algorithms have been developed to detect ice on wind turbine blades, they still face many problems. First, blade ice formation produces various dynamic changes depending on weather conditions, and the generated signal is highly dependent on the frequency, duration, severity, and intensity of the icing conditions, which means that abnormal changes in patterns embedded in the time and frequency domains must be considered and exploited. Second, although several attempts have been made to extract multi-scale features from the time and frequency domains using wavelet decomposition, scale-specific inter-variable correlations between different scales have not been exploited. Instead, existing models usually learn salient and shared correlations across all scales. Summary of the invention The present invention solves the problem of deep learning algorithms for detecting icing on wind turbine blades. In order to consider and utilize abnormal changes in patterns embedded in the time and frequency domains, prominent and shared correlations of all scales are usually learned, but scale-specific correlations between variables at different scales are not utilized, which results in the inability to capture and adapt to various dynamic changes of blade icing.

[0006] To achieve the above object, the present invention provides the following solutions: The present invention provides a time-series graph attention icing detection model based on multi-scale wavelet decomposition, wherein the icing detection model comprises a data preprocessing module, a multi-layer wavelet decomposition network module, a scale-related learning module, a time-series graph attention module, a scale attention module and an icing classifier; The data preprocessing module is used to preprocess the collected data and send it to the multi-layer wavelet decomposition network module; The multi-layer wavelet decomposition network module is used to extract detail coefficients and approximate coefficients of different scales of the preprocessed data, obtain multi-scale wavelet input, and send them to the scale-related learning module and the time series graph attention module respectively; The scale-dependent learning module is used to construct an adjacency matrix using two types of embedding vectors according to detail coefficients of different scales, and send the matrix to the timing graph attention module; The temporal graph attention module is used to convert the graph structure formed by the multi-scale wavelet input and the adjacency matrix into a scale-specific representation; The scale attention module is used to convert the input of the scale-specific representation into a scale attention representation and send it to the icing classifier for detection.

[0007] Furthermore, in a preferred embodiment, the multi-layer wavelet decomposition network module includes two functions for performing hierarchical decomposition on the time series.

[0008] Furthermore, there is another preferred embodiment, the above two functions are respectively: ; ; in, is the activation function, and Represent the weight matrices of low frequency and high frequency respectively, and Respectively represent -1 layer generated Low frequency subsequence and high frequency subsequence, and Both represent the bias vector.

[0009] Furthermore, in a preferred embodiment, the above-mentioned temporal graph attention module includes a graph attention network GAT and a temporal convolutional network TCN; The graph attention network GAT is used to linearly transform the features of each scale; The temporal convolutional network TCN is used to obtain scale features.

[0010] Furthermore, there is a preferred embodiment in which the ice classifier is composed of The structure of the network is composed of fully connected layers, and ReLU and Sigmoid activation functions are applied to each layer.

[0011] The present invention also provides an icing detection method implemented by the time series graph attention icing detection model based on multi-scale wavelet decomposition described in any one of the above, wherein the icing detection method is: S1: The preprocessed data is decomposed through a multi-layer wavelet decomposition network module to extract detail coefficients and approximate coefficients of different scales and obtain multi-scale wavelet input; S2: Constructing adjacency matrices using two types of embedding vectors according to detail coefficients at different scales through a scale-dependent learning module; S3: Input the graph structure formed by multi-scale wavelet input and adjacency matrix into the temporal graph attention convolutional network to form scale feature representation; S4: Input the scale feature representation into the scale attention module to generate the scale attention representation; S5: Input the scaled attention representation into the ice classifier for detection.

[0012] Furthermore, in another preferred embodiment, the above step S2 is specifically as follows: S21: Use the shared feature embedding vector and the unique embedding vector at each scale to calculate the scale-specific embedding matrix; S22: According to A scale-specific embedding matrix with two trainable parameters and Generate matrix decomposition vector; S23: Generate the first vector by matrix decomposition The original adjacency matrix of the scale is resized and sparsely processed.

[0013] Furthermore, in another preferred embodiment, the above step S3 is specifically as follows: S31: Use the graph attention network GAT to linearly transform the features of each scale to obtain the first The characteristics of the time step t of the scale; S32: Calculation scale Its adjacent scale The attention coefficient between them is normalized; S33: Use the weighted adjacency matrix to adjust the normalized attention coefficient; S34: According to the adjusted attention coefficient The features of the scaled time step t are weighted summed to obtain the scale The final feature of S35: The scale The final feature is input into the temporal convolutional network TCN to obtain the Scale characteristics of wavelet scale.

[0014] The method for detecting attention icing in a time series graph based on multi-scale wavelet decomposition described in the present invention can be fully implemented by computer software. Therefore, correspondingly, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, any one of the methods for detecting attention icing in a time series graph based on multi-scale wavelet decomposition described above is executed.

[0015] The present invention also provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes any one of the above-mentioned time series graph attention icing detection methods based on multi-scale wavelet decomposition.

[0016] The beneficial effects of the present invention are: 1. The present invention provides a time-series graph attention icing detection model based on multi-scale wavelet decomposition, which extracts multi-scale features in the time domain and frequency domain through a multi-layer wavelet decomposition network; then a scale-related learning module is applied to automatically infer a scale-specific graph, that is, a graph structure formed by an adjacency matrix and a multi-scale wavelet input, so that the model can utilize the implicit information of the relationship between variables in the time domain and the frequency domain. Multi-scale information is also obtained by adopting a time-series graph attention convolutional network instead of using the traditional RNN method. In order to further improve the model, the scale attention module is adopted to enable the importance of temporal patterns at different scales to be weighted and the correlation information between scales to be captured. Finally, the icing classifier is entered for classification.

[0017] The present invention is an invention that overcomes the prejudices of the prior art.

[0018] Compared with the prior art, the present invention does not learn the prominent and shared correlations of all scales, but instead utilizes the scale-specific correlations between variables at different scales, so as to better capture and adapt to various dynamic changes of blade icing. Therefore, the present invention is an invention that overcomes the bias of the prior art.

[0019] The present invention provides a time-series graph attention icing detection model based on multi-scale wavelet decomposition, which uses a multi-scale wavelet decomposition network to simulate the relationship between multi-scale wavelets in the time and frequency domains, and applies the time-series graph attention module to wind turbine blade icing detection for the first time. The proposed model is experimentally evaluated using real SCADA data from a wind farm, and the results verify the effectiveness of the model.

[0020] The present invention is suitable for online monitoring of icing of fan blades. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is the overall framework of a wind turbine blade icing detection model based on a time series graph attention convolution network with multi-scale wavelet decomposition as described in the first embodiment; Figure 2 It is a structural diagram of the multi-layer wavelet decomposition network module described in Implementation Mode 1. DETAILED DESCRIPTION

[0022] The specific embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can be made without departing from the concept of the present invention.

[0023] Implementation method 1, see Figure 1 This embodiment is described. This embodiment provides a time-series graph attention icing detection model based on multi-scale wavelet decomposition, the model comprising a data preprocessing module, a multi-layer wavelet decomposition network module, a scale-dependent learning module, a time-series graph attention module, a scale attention module and an icing classifier; The data preprocessing module is used to preprocess the collected data and send it to the multi-layer wavelet decomposition network module; The multi-layer wavelet decomposition network module is used to extract detail coefficients and approximate coefficients of different scales of the preprocessed data, obtain multi-scale wavelet input, and send them to the scale-related learning module and the time series graph attention module respectively; The scale-dependent learning module is used to construct an adjacency matrix using two types of embedding vectors according to detail coefficients of different scales, and send it to the time series graph attention module; The temporal graph attention module is used to convert the graph structure formed by the multi-scale wavelet input and the adjacency matrix into a scale-specific representation; The scale attention module is used to convert the input of scale-specific representation into scale attention representation and send it to the icing classifier for detection.

[0024] In practical applications, the time series graph attention icing detection model based on multi-scale wavelet decomposition is as follows: Figure 1 As shown, the detection model includes data collection, data preprocessing module, multi-layer wavelet decomposition network (mWDN) module, scale-dependent learning module, temporal graph attention module, scale attention module and icing classifier; Among them, data collection is used to send the collected data to the data preprocessing module; the data preprocessing module preprocesses the received data and sends it to the multi-layer wavelet decomposition network (mWDN) module; the multi-layer wavelet decomposition network (mWDN) module is used to extract the detail coefficients and approximate coefficients of different scales of the preprocessed data to obtain multi-scale wavelet input. According to the detail coefficients and approximate coefficients of different scales, multi-scale feature extraction in the time domain and frequency domain can be completed. The scale-related learning module uses two types of embedding vectors to construct an adjacency matrix according to the detail coefficients of different scales, and forms a graph structure with the adjacency matrix and the multi-scale wavelet input, so that the model can use the implicit information of the relationship between variables in the time domain and frequency domain. The formed graph structure is converted into a scale-specific representation through the time-series graph attention module, so that more scale information can be obtained, and the scale-specific representation is converted into a scale-specific representation and sent to the scale attention module. The scale attention module is used to generate a scale attention representation from the scale-specific representation input, so that the importance of the time pattern can be weighted at different scales and the correlation information between scales can be captured. Finally, the scale attention representation is input into the ice classifier to obtain the detection result.

[0025] This embodiment provides a time-series graph attention icing detection model based on multi-scale wavelet decomposition, which extracts multi-scale features in the time domain and frequency domain through a multi-layer wavelet decomposition network; then a scale-related learning module is applied to automatically infer a scale-specific graph, that is, a graph structure formed by an adjacency matrix and a multi-scale wavelet input, so that the model can utilize the implicit information of the relationship between variables in the time domain and the frequency domain. Multi-scale information is also obtained by adopting a time-series graph attention convolutional network instead of using the traditional RNN method. In order to further improve the model, the scale attention module is adopted to enable the importance of temporal patterns at different scales to be weighted and the correlation information between scales to be captured. Finally, the icing classifier is entered for classification.

[0026] Implementation method 2: This implementation method is to illustrate the multi-layer wavelet decomposition network module in the time series graph attention icing detection model based on multi-scale wavelet decomposition described in implementation method 1 by way of example; The multi-layer wavelet decomposition network module includes two functions for performing hierarchical decomposition on time series.

[0027] Implementation method 3: This implementation method is to illustrate two functions in a time series graph attention icing detection model based on multi-scale wavelet decomposition described in implementation method 2 by way of example; The two functions are: ; ; in, is the activation function, and Represent the weight matrices of low frequency and high frequency respectively, and Respectively represent -1 layer generated Low frequency subsequence and high frequency subsequence, and Both represent the bias vector.

[0028] Implementation 4: This implementation is to illustrate the time series graph attention module in the time series graph attention icing detection model based on multi-scale wavelet decomposition described in Implementation 1 by way of example; The temporal graph attention module includes a graph attention network GAT and a temporal convolutional network TCN; The graph attention network GAT is used to perform linear transformation on features of each scale; The temporal convolutional network TCN is used to obtain scale features.

[0029] In practical applications, the temporal graph attention module includes a graph attention network GAT and a temporal convolutional network TCN, wherein the graph attention network GAT is used to perform linear transformation on the features of each scale to obtain the first The features of the time step t of the scale are input into the time convolution network TCN to obtain the Scale characteristics of wavelet scale.

[0030] Implementation mode 5. This implementation mode is to illustrate an icing classifier in a time series graph attention icing detection model based on multi-scale wavelet decomposition described in implementation mode 1 by way of example; The icing classifier is composed of The structure of the network is composed of fully connected layers, and ReLU and Sigmoid activation functions are applied to each layer.

[0031] Implementation Method 6: See Figure 2 This embodiment describes a method for detecting icing in a time series graph based on multi-scale wavelet decomposition. The detection method is implemented based on a time series graph attention icing detection model based on multi-scale wavelet decomposition described in any one of Embodiments 1 to 5. The detection method is: Step S1: The preprocessed data is decomposed through a multi-layer wavelet decomposition network module to extract detail coefficients and approximate coefficients of different scales to obtain multi-scale wavelet input; Specifically: Step S11: Input data is ; In the formula, N represents the number of signals and T represents the length of the input segment.

[0032] Step S12: Use Figure 2The multi-layer wavelet decomposition network module (mWDN) shown processes the input data, specifically: the multi-layer wavelet decomposition network module (mWDN) uses the following two functions to perform hierarchical decomposition on the time series of the input data; The two functions are: ; ; in, is the activation function, and Represent the weight matrices of low frequency and high frequency respectively, and Respectively represent -1 layer generated Low frequency subsequence and high frequency subsequence, and Both represent the bias vector.

[0033] in, and The average pooling layer is used to extract the intermediate variables and The downsampling is as follows: ; ; In the formula, and They all represent the intermediate variables after average pooling.

[0034] Step S13: obtaining the wavelet detail coefficients of the input signal at multiple scales and the wavelet approximation coefficients of the input signal at multiple scales, that is, multi-scale wavelet input, which is specifically ; Step S2: construct an adjacency matrix using two types of embedding vectors according to detail coefficients of different scales through a scale-dependent learning module; Step S21: Embedding vector using shared features and a unique embedding vector at each scale Calculate the Scale-specific embedding matrix, where is its embedding dimension, N is the number of signals; No. The scale-specific embedding matrix is: ; In the formula, Represents element-wise multiplication.

[0035] Step S22: According to A scale-specific embedding matrix with two trainable parameters and Generate matrix decomposition vector and ; ; ; Step S23: Decompose the matrix vector generated by step S21 above and Generate The original adjacency matrix of the scale ; ; Step S24: In order to increase the robustness of the model, reduce the amount of calculation and reduce the impact of noise, the original adjacency matrix generated in step S23 is Perform sparse processing; ; Among them, sparse is defined as: ; In the formula, is the threshold number of neighbors for each node in the critical matrix. represents the connection weight between node r and node k, represents the connection weight between node r and node i, represents the sum of the connection weights between node r and all other nodes i.

[0036] Step S3: Input the graph structure formed by multi-scale wavelet input and adjacency matrix into the temporal graph attention convolutional network to form scale feature representation; Specifically: Step S31: Use the graph attention network (GAT) submodule to perform linear transformation on the features of each scale to obtain the first The characteristics of the time step t of the scale ; ; In the formula, It is The time step of the scale The wavelet input is is the weight matrix.

[0037] Step S32: Calculate the scale Its adjacent scale The attention coefficient between : ; In the formula, is the attention weight, Indicates that and Make stitching.

[0038] Step S33: Use the softmax function to calculate the attention coefficient calculated in step S32 above Perform normalization; ; In the formula, Representation Node The set of neighbor nodes of Represents the attention coefficient between node i and node k.

[0039] Step S34: Using weighted adjacency matrix Adjust the normalized attention coefficient in step S33 above : ; According to the adjusted attention coefficient , the first The characteristics of the time step t of the scale Perform weighted summation to obtain the scale The final feature : ; In the formula, Represents a non-linear activation function, such as the ReLU function.

[0040] Step S35: Add the graph attention network (GAT) submodule to the Output on scale Output to the temporal convolutional network (TCN) submodule to obtain the Scale characteristics of wavelet scale: ; In the formula, It is TCN of scaled wavelet detail coefficients, The size is set to .

[0041] Among them, the TCN submodule consists of The extended causal convolutional layer is composed of: ; In the formula, Indicates The input of the dilated causal convolutional layer, Represents the time step The convolution operation, represents element-by-element multiplication, is a convolution filter.

[0042] Step S4: Input the scale feature representation into the scale attention module to generate the scale attention representation; Specifically: Step S41: Perform an average pooling operation on the scale dimension to aggregate the scale-specific information of signal variables between different scales, thereby revealing the cross-scale correlation of the representation: ; In the formula, Is a signal variable The aggregate representation of .

[0043] Step S42: Use multiple fully connected layers to calculate the attention weight of each scale: ; ; in, Representing scale-specific features The weight of the fully connected layer is and , the bias is and .

[0044] Step S43: Scale-specific features Weighting to generate scaled attention representation : ; In the formula, Represents an element-wise Hadamard product operation.

[0045] Step S5: Input the scale attention representation into the icing classifier prediction to obtain the prediction result.

[0046] Specifically: Step S51: Input scale attention representation is passed to the icing classifier; Step S52: The icing classifier is composed of It consists of fully connected layers, and each layer uses ReLU and Sigmoid activation functions; The formula is: ; ; ; In the formula, is the output of the first fully connected layer, It is The output of the layer, It is the probability of ice detection normalized to the range of (0,1) by the Sigmoid function.

[0047] Step S53: In order to balance the data, a focal loss function is used, and its formula is: ; In the formula, is the true label probability of the ice layer, is a parameter that adjusts the weight ratio of positive and negative samples. is a parameter that reduces the loss value of negative samples.

[0048] When applying, the algorithm also needs to be learned. The algorithm learning process is: Input is the number of training rounds , the number of wavelet decomposition layers , input data matrix ,Number of nodes (number of signals at a specific wavelet scale), number of time steps , the true label

[0049] enter:

[0050] Output: predicted label

[0051] initialization: Randomly initialize trainable parameters

[0052]

[0053]

[0054]

[0055]

[0056]

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071] pass Update the model using gradient descent

[0072] Implementation method seven: This implementation method provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method for detecting icing in a time series graph based on multi-scale wavelet decomposition as described in implementation method six is ​​executed.

[0073] Implementation method eight: This implementation method provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the time series graph attention icing detection method based on multi-scale wavelet decomposition described in implementation method six.

[0074] Implementation method nine: This implementation method experimentally verifies the performance of a time series graph attention icing detection model based on multi-scale wavelet decomposition as described in any one of implementation methods one to five.

[0075] The software used in the experiment is PyCharm software and PyTorch framework.

[0076] The experimental data used are SCADA data from two wind turbines installed in China's Inner Mongolia Autonomous Region. The SCADA data is sampled every 7 seconds, with a total operating time of 696 hours and 306 hours respectively. There are hundreds of sensors inside the wind turbines providing readings, and 26 signals related to blade icing conditions are used as data sets.

[0077] The experimental results are shown in the following table:

[0078] It can be seen from the table that the time series graph attention icing detection model based on multi-scale wavelet decomposition verified in this embodiment has higher accuracy, recall rate and F1 score than LSTM, FCNN and RF-XGBoost, indicating that the time series graph attention icing detection model based on multi-scale wavelet decomposition verified in this embodiment is superior to the traditional algorithm.

[0079] In summary, this embodiment verifies a time-series graph attention icing detection model based on multi-scale wavelet decomposition, which uses a multi-scale wavelet decomposition network to simulate the relationship between multi-scale wavelets in the time and frequency domains, and for the first time applies the time-series graph attention module to wind turbine blade icing detection. The proposed model is experimentally evaluated using real SCADA data from a wind farm, and the results verify the effectiveness of the model.

[0080] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may be subject to various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of the claims of the present invention.

Claims

1. A time-series graph attention icing detection model based on multi-scale wavelet decomposition, characterized by: The model includes a data preprocessing module, a multi-layer wavelet decomposition network module, a scale-dependent learning module, a temporal graph attention module, a scale attention module, and an icing classifier; The data preprocessing module is used to preprocess the collected data and send it to the multi-layer wavelet decomposition network module; The multi-layer wavelet decomposition network module is used to extract detail coefficients and approximate coefficients of different scales of the preprocessed data, obtain multi-scale wavelet input, and send them to the scale-related learning module and the time series graph attention module respectively; The scale-dependent learning module is used to construct an adjacency matrix using two types of embedding vectors according to detail coefficients of different scales, and send it to the time series graph attention module; The temporal graph attention module is used to convert the graph structure formed by the multi-scale wavelet input and the adjacency matrix into a scale-specific representation; The scale attention module is used to convert the input of scale-specific representation into scale attention representation and send it to the icing classifier for detection.

2. The time series graph attention icing detection model based on multi-scale wavelet decomposition according to claim 1 is characterized in that: The multi-layer wavelet decomposition network module includes two functions for performing hierarchical decomposition on time series.

3. The time series graph attention icing detection model based on multi-scale wavelet decomposition according to claim 2 is characterized in that: The two functions are: ； ； in, is the activation function, and Represent the weight matrices of low frequency and high frequency respectively, and Respectively represent -1 layer generated Low frequency subsequence and high frequency subsequence, and Both represent the bias vector.

4. The time series graph attention icing detection model based on multi-scale wavelet decomposition according to claim 1 is characterized in that: The temporal graph attention module includes a graph attention network GAT and a temporal convolutional network TCN; The graph attention network GAT is used to linearly transform the features of each scale; The temporal convolutional network (TCN) is used to obtain scale features.

5. The time series graph attention icing detection model based on multi-scale wavelet decomposition according to claim 1 is characterized in that: The icing classifier is composed of The structure of the network is composed of fully connected layers, and ReLU and Sigmoid activation functions are applied to each layer.

6. A time-series graph attention icing detection method based on multi-scale wavelet decomposition, characterized in that: The detection method is implemented based on the time series graph attention icing detection model based on multi-scale wavelet decomposition according to any one of claims 1 to 5, and the method is: S1: The preprocessed data is decomposed through a multi-layer wavelet decomposition network module to extract detail coefficients and approximate coefficients of different scales and obtain multi-scale wavelet input; S2: Constructing adjacency matrices using two types of embedding vectors according to detail coefficients at different scales through a scale-dependent learning module; S3: Input the graph structure formed by multi-scale wavelet input and adjacency matrix into the temporal graph attention convolutional network to form scale feature representation; S4: Input the scale feature representation into the scale attention module to generate the scale attention representation; S5: Input the scaled attention representation into the ice classifier for detection.

7. The method for detecting icing in a time series graph based on multi-scale wavelet decomposition according to claim 6, characterized in that: S2 is specifically: S21: Use the shared feature embedding vector and the unique embedding vector at each scale to calculate the scale-specific embedding matrix; S22: According to A scale-specific embedding matrix with two trainable parameters and Generate matrix decomposition vector; S23: Generate the first vector by matrix decomposition The original adjacency matrix of the scale is resized and sparsely processed.

8. The method for detecting icing in a time series graph based on multi-scale wavelet decomposition according to claim 6, characterized in that: S3 is specifically: S31: Use the graph attention network GAT to linearly transform the features of each scale to obtain the first The characteristics of the time step t of the scale; S32: Calculation scale Its adjacent scale The attention coefficient between them is normalized; S33: Use the weighted adjacency matrix to adjust the normalized attention coefficient; S34: According to the adjusted attention coefficient The features of the scaled time step t are weighted summed to obtain the scale The final feature of S35: The scale The final feature of is input into the temporal convolutional network TCN to obtain the Scale characteristics of wavelet scale.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, executes the time series graph attention icing detection method based on multi-scale wavelet decomposition as described in any one of claims 6 to 8.

10. A computer device, characterized in that: The device includes a memory and a processor, wherein a computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the time series graph attention icing detection method based on multi-scale wavelet decomposition as described in any one of claims 6 to 8.

Images