Icing situation awareness method and system based on time-space fusion

Abstract

The invention provides an icing situation sensing method and system based on time-space fusion, and the method comprises the steps: obtaining sensing data which comprehensively reflects the formation and spatial distribution of icing from different angles through a multi-point monitoring mode, and extracting key parameters, namely multi-source characteristic parameters, representing the icing situation from the sensing data, the method comprises the following steps: carrying out weighted fusion on icing data change rate, temperature deviation value and maximum icing difference of icing data on a blade based on an attention mechanism, adaptively highlighting key features, weakening the weight of secondary features, and obtaining fusion features; and then, on the basis of a related topological graph constructed by fusion features, a graph neural network and a Transform network are utilized to capture spatial distribution features and time evolution laws of icing, so that the limitation that spatial correlation is difficult to describe only depending on a time sequence model traditionally is overcome, spatial features and time sequence features are effectively extracted, final spatial-temporal features are obtained, and the spatial distribution features and the time evolution laws of icing are extracted. And prediction is carried out based on the spatial-temporal characteristics so as to improve the precision of icing situation perception.

Description

Technical Field

[0001] This invention relates to the fields of intelligent sensing and artificial intelligence, specifically to the field of wind power equipment operation monitoring, and more specifically, to a spatiotemporal fusion-based method and system for icing situation perception. Background Technology

[0002] Wind power generation uses wind energy to drive the blades of wind turbines to rotate, which in turn drives a generator to produce electricity. It is a clean and renewable power generation method.

[0003] Wind turbine blades are prone to icing (ice formation) in cold and humid environments at high latitudes or altitudes. Blade icing leads to deterioration of the turbine's aerodynamic performance, inaccurate readings from wind and temperature measuring devices, reduced power output, and may cause safety accidents such as blade vibration imbalance, structural damage, and ice fall. Therefore, it is necessary to consider relevant measures to ensure the safe and efficient operation of wind turbines and guide anti-icing and de-icing procedures. Currently, the industry has proposed various technical solutions for wind turbine icing monitoring, but significant shortcomings still exist.

[0004] Traditional sensor monitoring solutions utilize dedicated sensors for single physical quantities to detect icing, such as single fiber optic sensors, vibration acceleration sensors, and infrared thermometers. These solutions cannot fully utilize multi-source information and are therefore susceptible to noise interference and misjudgments in complex environments.

[0005] Manual inspection and image detection solutions: Blade icing is inspected manually (using binoculars) or by drone aerial photography. These offline methods cannot be used continuously during wind turbine operation, suffer from monitoring lag, and require shutdown or manual intervention, resulting in high maintenance costs. Image-based detection is also susceptible to weather and lighting conditions, making it difficult to detect icing in low visibility conditions.

[0006] Software-based detection solutions using SCADA data: These solutions utilize existing environmental, electrical, and operational parameter data from the wind turbine's own SCADA (Supervisory and Data Acquisition) system to detect icing faults through data-driven methods. Existing research has attempted to use machine learning and deep learning models to analyze signals such as power curves and rotational speed vibrations to determine icing, such as the method of fusing support vector machines with random forests and the method of modeling time-series features using recurrent neural networks (LSTM, IndRNN, etc.). These data-driven models have achieved some success in offline historical data analysis, but they mostly rely on powerful computing capabilities and large amounts of labeled data, and are typically deployed on high-performance equipment in cloud devices (such as cloud servers or server clusters) or wind farm central control rooms, which is not conducive to real-time applications at the front-end. Furthermore, due to the differences in data distribution across different wind farms and turbine models, these models have poor generalization ability after training in a single wind farm or under specific conditions. Their accuracy often drops significantly when switching to new wind farms, meteorological conditions, or turbine models, lacking model transfer and environmental adaptation capabilities. In addition, most existing solutions can only provide detection alarms for the presence or absence of icing, and lack the ability to quantify and classify the degree of icing and predict the trend of icing (whether it intensifies or eases), thus failing to provide sufficient early warning guidance.

[0007] When researching wind turbine blade icing detection technology, the inventors discovered that the shortcomings of existing technologies were mainly due to the following reasons:

[0008] Insufficient feature and spatial awareness: Relying solely on a single sensor signal for determination fails to effectively characterize the coupling relationship and spatial correlation between multi-source observations, resulting in insufficient capture of key information on the formation and expansion mechanism of icing.

[0009] Insufficient temporal modeling: The scheme using general temporal networks does not adequately characterize the multi-timescale features of short-term disturbances and long-term evolution, and it is difficult to simultaneously take into account the joint modeling and discrimination of transient changes and slow trends.

[0010] It should be noted that the background information presented here is only for illustrating relevant information about the present invention to aid in understanding the technical solution of the present invention, and does not imply that the relevant information is necessarily prior art. The relevant information was submitted and disclosed together with the present invention, and should not be considered prior art unless there is evidence that the relevant information was disclosed before the filing date of the present invention. Summary of the Invention

[0011] Therefore, the purpose of this invention is to overcome the shortcomings of the prior art and provide a spatiotemporal fusion-based method and system for icing situational awareness.

[0012] The objective of this invention is achieved through the following technical solution:

[0013] According to a first aspect of the present invention, a method for sensing the icing situation of wind turbine blades is provided, comprising: acquiring sensing data collected by multiple sensing nodes dispersed on the leading edge and / or windward side of the blade, each sensing node including an icing sensor and a temperature sensor; constructing multi-source feature parameters for each sensing node based on the sensing data, including: icing data change rate, maximum icing difference on the blade, and temperature deviation value; acquiring a trained icing situation prediction model, including a fusion feature extraction module, a graph neural network, a Transformer network, and a prediction layer; for each sensing node, using the fusion feature extraction module to perform self-attention weighted fusion of each parameter in its multi-source feature parameters to obtain fusion features; acquiring a topology graph corresponding to multiple sensing nodes, setting the features of the graph nodes of the topology graph as the fusion features of the corresponding sensing nodes, extracting spatial features from the topology graph using the graph neural network, stacking the spatial features at multiple time points in temporal order to obtain a multi-source spatial feature sequence; processing the multi-source spatial feature sequence using the Transformer network to obtain spatiotemporal features; and determining the icing situation sensing result of the blade using the prediction layer based on the spatiotemporal features. This scheme achieves at least the following beneficial technical effects: First, by setting up multiple sensor nodes at key locations on wind turbine blades prone to icing (such as the leading edge, the windward side, or both), it collects crucial data on areas prone to icing. Furthermore, by distributing these sensor nodes and simultaneously installing icing and temperature sensors at each node, it obtains comprehensive sensor data reflecting icing formation and spatial distribution from different angles through multi-point monitoring. Second, the scheme extracts key parameters characterizing the icing situation, namely multi-source feature parameters, including: the rate of change of icing data, temperature deviation, and the maximum icing difference on the blade. These are then weighted and fused based on an attention mechanism, adaptively highlighting key features and weakening the weight of secondary features to obtain fused features. Finally, based on the relevant topology map constructed from the fused features, graph neural networks and Transformer networks are used to simultaneously capture the spatial distribution characteristics and temporal evolution patterns of icing. This overcomes the limitation of traditional time-series models, which struggle to characterize spatial relationships, effectively extracting both spatial and temporal features and improving the accuracy of icing situation perception.

[0014] Optionally, edges and their weights are established in the topology graph based on the physical distance between graph nodes and / or historical icing correlation. A graph convolutional network or a graph attention network is used as the graph neural network to extract spatial features from the topology graph, obtaining a multi-source spatial feature sequence. The Transformer network processing the multi-source spatial feature sequence is lightweight, with its self-attention encoding module having only one or two encoding layers. This scheme achieves at least the following beneficial technical effects: First, by establishing edges and their weights in the topology graph based on the physical distance between graph nodes and / or historical icing correlation, additional information can be provided by leveraging the spatial distance relationships between nodes and / or historical icing correlations, thereby improving the accuracy of icing situation perception. Second, graph convolutional networks or graph attention networks are used to better aggregate the features of nodes at different locations to obtain spatial features. Third, lightweight Transformers are used to extract spatiotemporal features, ultimately capturing the spatial distribution features and temporal evolution of ice cover with less computation. This overcomes the limitation of traditional time-series models, which are difficult to characterize spatial relationships. It is easier to deploy and run efficiently in scenarios with limited computing power (such as wind turbines), and helps to achieve the unification of high-precision detection and embedded real-time deployment.

[0015] Optionally, the icing situation prediction model needs to be trained as follows before use: Obtain the icing situation prediction model to be trained; obtain a training set, including multiple samples, each sample consisting of a sensor data sequence composed of sensor data collected periodically at multiple times and a label, where the label is the ground truth of the icing situation, including an icing state category, or an icing state category and an icing trend category; construct multi-source feature parameters based on the sensor data in the samples of the training set, input the multi-source feature parameters into the icing situation prediction model, output the icing situation perception confidence of the blade, and use the situation perception loss value calculated based on the difference between the blade's icing situation perception confidence and the label to guide the model parameter update, in order to minimize the loss value. This scheme can achieve at least the following beneficial technical effects: This scheme can set an icing state category in the label, or set an icing state category and an icing trend category, to guide the model to learn relevant knowledge and improve its ability to perceive the icing situation.

[0016] Optionally, the labels also include the intraday time of sample collection and the season. The method further includes: setting up a period prediction module to assist training during training. This period prediction module is configured to analyze the time confidence and seasonal confidence of the corresponding samples at each time of collection based on spatiotemporal characteristics; calculate the situational awareness loss based on the difference between the leaf icing situational awareness confidence and the labels; calculate the period prediction loss based on the differences between the time confidence, seasonal confidence, and labels; and guide the parameter updates of the model and the period prediction module based on the weighted loss value of the situational awareness loss and the period prediction loss to minimize the weighted loss value. This scheme can achieve at least the following beneficial technical effects: To improve the model's sensitivity to seasonal cyclical patterns, cyclical information is specifically added to the labels: the intraday time and current season are used as additional labels, combined with the additionally set period prediction module, to help the model understand the location of the current time point within a day or year based on data from icing sensors and temperature sensors, thus improving the model's ability to perceive icing situations.

[0017] Optionally, the processing of constructing multi-source feature parameters for each sensor node based on sensor data includes: sequentially performing sliding window processing on the icing data and temperature data in the sensor data to calculate the average value of the local data captured by each sliding window, thus obtaining the average values ​​of icing and temperature; performing wavelet transform denoising and trend separation on the average values ​​of icing and temperature respectively, thus obtaining the denoised results of icing and temperature; calculating the rate of change of icing data using the denoised results of the previous time point and the current time point; subtracting the minimum value from the maximum value of the icing data at the current time point to obtain the maximum icing difference on the blade; and subtracting the minimum value from the denoised temperature results from the preset baseline temperature to obtain the temperature deviation value. The scheme can achieve at least the following beneficial technical effects: the sliding window processing can obtain the average value in the short term as a representative, so as to weaken the influence of abnormal noise; the wavelet transform domain denoising processing can further reduce the influence of noise and obtain the denoising results of icing and temperature. Based on the denoising results of icing and temperature, the key parameters (iceing data change rate, maximum icing difference and temperature deviation value) used for analyzing the icing situation can be extracted more stably.

[0018] Optionally, the multi-source feature parameters also include: temperature data, icing data, and / or temperature gradient, where the temperature gradient is the rate of temperature change calculated based on the temperature data or the denoising result of the temperature. This scheme can achieve at least the following beneficial technical effects: adding raw temperature data and / or icing data to the multi-source feature parameters can diversify the analysis perspective and improve the accuracy of situational awareness; adding a temperature gradient allows the model to intuitively obtain the rate of temperature rise or fall, so as to better analyze the icing situation.

[0019] According to a second aspect of the present invention, an edge-cloud collaborative wind turbine blade icing situation awareness system is provided. The system includes: a terminal device comprising multiple sensing nodes for collecting sensing data from the blade and a terminal processor for executing the method described in the first aspect; an edge device configured to: communicate with the terminal device, collect the sensing data collected by the terminal device and its corresponding icing situation awareness results, and notify a local monitoring platform and a cloud device when a sensing anomaly is detected based on the sensing data and its corresponding icing situation awareness results; and a cloud device configured to: communicate with the edge device, periodically acquire the sensing data, video monitoring data, and human feedback data collected by the edge device to update the sample database of each wind turbine, and, when a sensing anomaly occurs on a terminal device, update the icing situation awareness result of the icing situation prediction model for the blade using the prediction result of a cloud model that is more accurate than the icing situation prediction model, and remind human intervention to label or clean abnormal samples so as to periodically guide the training of the icing situation prediction model based on the collected samples using the cloud model as a teacher model. This solution achieves at least the following beneficial technical effects: Firstly, the cloud-edge-device collaborative architecture combines the advantages of real-time on-site response and global optimization, giving the system the ability to "get smarter with use," effectively solving the problem in existing technologies where models struggle to adapt to new environments after a single offline training session. With this architecture, the terminal processor can quickly perform initial icing assessments; cloud devices perform more complex and detailed analyses, providing a powerful supplement to the terminal devices' predictions, further improving overall detection accuracy. Through online transfer learning, the model can continuously update itself based on different wind field climates, maintaining high accuracy and a low false alarm rate even after years of operation. Secondly, because the terminal devices can independently complete detection tasks without relying on a continuous network connection, the system can still monitor and operate locally even when the network is interrupted, improving reliability. In summary, the layered collaborative design of this solution not only improves the real-time accuracy of icing situation awareness for a single wind turbine but also enables model sharing and continuous optimization across wind fields, significantly enhancing the adaptability and long-term effectiveness of icing detection.

[0020] Optionally, the perceived anomaly includes situations where the lowest temperature data in the sensing data is greater than a preset temperature reference value, but the icing situation perception result shows that the blades are iced. This solution can achieve at least the following beneficial technical effects: it can detect abnormal situations judged as icing under conditions where icing is not present, and guide human intervention in a timely manner to more efficiently improve and enhance model performance.

[0021] According to a third aspect of the present invention, an electronic device is provided, comprising: one or more processors; and a memory for storing executable instructions; wherein the one or more processors are configured to implement the steps of the method described in the first aspect by executing the executable instructions. Attached Figure Description

[0022] The embodiments of the present invention will be further described below with reference to the accompanying drawings, wherein:

[0023] Figure 1 This is a flowchart illustrating the icing situation perception method for wind turbine blades according to an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of deploying sensing nodes on a blade according to an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of a wind turbine blade icing situational awareness system based on edge-cloud collaboration according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the invention.

[0027] As mentioned in the background section, existing technologies suffer from insufficient feature and spatial awareness, and inadequate temporal modeling. To address this, this application employs multiple sensor nodes positioned at key icing locations on wind turbine blades (such as the blade leading edge, the windward side, or both) to collect crucial data from areas prone to icing. Furthermore, by distributing sensor nodes and simultaneously installing icing and temperature sensors at each node, comprehensive sensor data reflecting icing formation and spatial distribution can be acquired from different angles through multi-point monitoring. Secondly, this scheme extracts key parameters characterizing the icing situation, namely multi-source feature parameters, including: the rate of change of icing data, temperature deviation, and the maximum icing difference on the blade. These are then weighted and fused based on an attention mechanism, adaptively highlighting key features and weakening the weight of secondary features to obtain fused features. Finally, based on the relevant topology map constructed from the fused features, graph neural networks and Transformer networks are used to simultaneously capture the spatial distribution features and temporal evolution patterns of icing. This overcomes the limitation that traditional models that rely solely on time series models cannot effectively characterize spatial relationships, effectively extracting both spatial and temporal features, and improving the accuracy of icing situation awareness.

[0028] The leading edge of a blade is the arc-shaped part that is responsible for cutting through the airflow first when the blade rotates. The leading edge is the very front of the airfoil and is the thickest arc segment on the blade.

[0029] The windward side of a blade is the entire surface of the blade facing the relative wind direction when it rotates; it is also called the pressure side.

[0030] According to one embodiment of the present invention, see Figure 1 This invention provides a method for sensing the icing situation of wind turbine blades, including steps S1-S7. To better understand this invention, each step is described in detail below with reference to specific embodiments.

[0031] Step S1: Acquire sensor data from multiple sensor nodes distributed on the leading edge and / or windward side of the blade. Each sensor node includes an icing sensor and a temperature sensor.

[0032] According to one embodiment of the present invention, each sensing node is a combined sensing unit comprising an icing sensor for detecting icing and a temperature sensor for detecting temperature. The icing sensor can be a capacitive or fiber optic icing sensor. The temperature sensor can be a thermistor or thermocouple temperature sensor. The leading edge and windward side of the blade are key detection areas. Sensing nodes can be deployed on the leading edge or windward side of the blade, or simultaneously on both, depending on user needs. Since blades are typically large, multiple sensing nodes are distributed to better utilize data from different spatial locations. For example, sensing nodes can be distributed at various distances from the blade root on the leading edge and / or windward side. This distributed icing / temperature sensor multi-point monitoring method comprehensively reflects the formation and spatial distribution of icing, improving detection sensitivity and reliability. Multiple sensing nodes on each blade also reduce the impact of single-point sensor failure or noise on judgment. Furthermore, the combination of capacitance changes at different locations with temperature information helps to distinguish between local anomalies and overall icing trends. Temperature parameters provide environmental background references, which can compensate for the shortcomings of simple capacitance detection (such as distinguishing between ice layers and water accumulation).

[0033] According to one example of the invention, the sensing nodes on the wind turbine blades can be in the form of an array or a linear array. See also Figure 2This paper presents schematic diagrams of sensor nodes installed on a blade from two different perspectives. To monitor icing conditions at different parts of the blade, multiple sensor nodes are uniformly distributed longitudinally (radially, perpendicular to the rotation axis) in easily icing areas such as the leading edge and windward side of the blade. Each sensor node integrates a capacitive icing sensor for sensing local dielectric constant changes and a temperature sensor for sensing local temperature changes. The sensor array covers the key area from the blade root to the tip, enabling timely capture of the icing distribution at different locations, thus providing comprehensive detection data for subsequent models. The capacitive icing sensor employs a differential capacitance electrode design, improving its sensitivity to dielectric changes caused by icing. The dedicated capacitance measurement chip integrated within the sensor node amplifies and digitizes minute capacitance changes, achieving a typical capacitance resolution of 0.1 fF, and incorporates a built-in temperature compensation function, simultaneously outputting capacitance and temperature values. Multiple such sensor nodes are distributed along the blade from root to tip, preferably evenly distributed at easily icing areas such as the leading edge and windward side, to capture the entire process of icing from its inception to its spread. All sensor nodes can form a sensor network via a wired bus (e.g., RS485 differential bus) or a low-power wireless network, and connect to edge devices at the blade roots or within the nacelle. The terminal processor (MCU) of the terminal device polls each sensor node at a set period to obtain its current capacitance and temperature values, and packages the data from all nodes at the same timestamp into a sensor data vector, recording it as a continuous time-series data stream. The sampling frequency can be configured as needed (e.g., 1 to 5 times per second) to balance monitoring timeliness and power consumption.

[0034] Step S2: Construct multi-source feature parameters for each sensor node based on the sensor data, including: the rate of change of icing data, the maximum icing difference on the blade, and the temperature deviation value.

[0035] According to one embodiment of the present invention, the multi-source feature parameters include the rate of change of icing data, the maximum icing difference in icing data on the blade, the temperature deviation value, temperature data, icing data, and / or the temperature gradient. The temperature gradient is the rate of temperature change calculated based on the temperature data or the denoising result of the temperature.

[0036] According to an embodiment of the present invention, the process of constructing multi-source feature parameters for each sensing node based on sensing data includes steps S21-S25. Wherein:

[0037] Step S21: Perform sliding window processing on the icing data and temperature data in the sensing data in turn, so as to calculate the average value of the local data captured by each sliding window and obtain the average value of icing and temperature.

[0038] Since the raw sensor signal may contain noise interference and slow drift components, a sliding window process is used to smooth noise and eliminate transient spikes in the acquired multi-source sensor data. Illustratively, assuming the sliding window captures data from 10 sampling points at a time, it can be represented as follows:

[0039] Sliding window for icing data: C_windowed = sliding_window(C_raw, window_size=10)

[0040] Temperature data sliding window: T_windowed = sliding_window(T_raw, window_size=10)

[0041] Where C_windowed represents the average icing of the current window, sliding_window represents the sliding window function, C_raw represents the icing data, T_windowed represents the average temperature of the current window, T_raw represents the temperature data, and window_size represents the window size.

[0042] Step S22: Perform wavelet transform denoising and trend separation on the average values ​​of icing and temperature respectively to obtain the denoising results of icing and temperature.

[0043] According to one embodiment of the present invention, the sampling wavelet transform algorithm performs wavelet transform denoising and trend separation on the average value of ice accretion, and on the average value of temperature, as follows:

[0044] C_denoised = wavelet_denoise(C_windowed, wavelet='db4', level=3)

[0045] T_denoised = wavelet_denoise(T_windowed, wavelet='db4', level=3)

[0046] Here, C_denoised represents the denoising result of icing; wavelet_denoise represents the wavelet denoising function, which uses wavelet transform to decompose the signal at multiple scales, removes high-frequency noise components, and reconstructs the signal; C_windowed represents the average icing value of the current window; wavelet='db4' indicates that a Daubechies 4th-order wavelet is used as the wavelet basis function for wavelet decomposition and reconstruction; level=3 indicates that the wavelet decomposition has 3 levels, meaning the signal is decomposed and processed at 3 different scales; T_denoised represents the denoising result of temperature; and T_windowed represents the average temperature of the current window. This scheme decomposes sensor data into components at different time scales using time-frequency analysis methods such as wavelet transform, separating rapidly changing detail signals from slowly changing trend signals. For example, when icing begins on a certain part of the blade, the local capacitance value will suddenly increase due to the change in dielectric constant, while the temperature may drop sharply. Through preprocessing, key feature parameters reflecting these signs can be extracted. This allows abnormal changes during icing to stand out from background trends, helping to improve the accuracy of the final icing situational awareness.

[0047] Step S23: Calculate the rate of change of icing data using the denoising results of the previous and current time points.

[0048] According to one embodiment of the present invention, the rate of change of icing data can be calculated as follows:

[0049] capacitance_rate = (C_denoised[-1] - C_denoised[0]) / dt

[0050] Wherein, capacity_rate represents the rate of change of icing data (the rate of change of capacitance calculated using a capacitive icing sensor as an example), C_denoised[-1] represents the denoising result of icing at the previous time point, C_denoised[0] represents the denoising result of icing at the current time point, and dt=Δt represents the time interval between adjacent sampling.

[0051] Step S24: Subtract the minimum value from the maximum value of the icing data at the current time point to obtain the maximum icing difference on the blade.

[0052] According to one embodiment of the present invention, the maximum icing difference can be calculated as follows:

[0053] icing difference = max(C_current) - min(C_current)

[0054] Wherein, icing difference represents the maximum icing difference, max(C_current) represents the maximum value of the icing data at the current time point, and min(C_current) represents the minimum value of the icing data at the current time point.

[0055] Step S25: Subtract the minimum value in the denoising results of the temperature from the preset baseline temperature to obtain the temperature deviation value.

[0056] According to one embodiment of the present invention, the temperature deviation value is a deviation from a preset baseline temperature, and can be calculated as follows:

[0057] temperature_drop = T_baseline - min(T_denoised)

[0058] Where temperature_drop represents the temperature deviation (or the magnitude of the temperature drop relative to the baseline), T_baseline represents the baseline temperature, and min(T_denoised) represents the minimum value in the denoised temperature result.

[0059] After the above processing, the terminal processor can calculate several key feature parameters in real time, which can sensitively reflect the occurrence and development of icing events. Subsequent feature extraction can be completed with lower computational and storage overhead, providing refined input information for subsequent AI models. The original massive data is compressed into a small number of physically meaningful feature vectors, reducing noise interference while retaining the significant signs of icing events.

[0060] Step S3: Obtain a trained icing situation prediction model, which includes a fusion feature extraction module, a graph neural network, a Transformer network, and a prediction layer.

[0061] According to an embodiment of the present invention, the icing situation prediction model is described below from the perspectives of training samples, model structure, and model training.

[0062] (1) Training samples

[0063] The purpose of icing situational awareness includes identifying the current icing state and / or predicting future icing trends. Therefore, training samples need to be designed to meet user needs.

[0064] According to one embodiment of the present invention, the dataset for icing situation awareness includes multiple samples. Each sample includes a sequence of sensor data consisting of sensor data collected periodically at multiple times and a label, where the label is the ground truth of the icing situation. For example, the label includes an icing state category. Alternatively, the label includes an icing state category and an icing trend category. Illustratively, the icing state category is, for example: 0 = "No ice", 1 = "Light icing", 2 = "Moderate icing", 3 = "Heavy icing". The icing trend category (such as an estimate of the ice layer growth rate or a trend classification) is used to indicate whether the icing is intensifying, stabilizing, or slowing down. The icing trend category is, for example: 0 = "Intensifying", 1 = "Stable", 2 = "Slowing down". It should be understood that, regardless of the icing state category or the icing trend category, those skilled in the art can adjust the granularity (decrease or increase the number of levels) according to actual needs to obtain other implementations, and the present invention does not impose any limitations on this.

[0065] According to one embodiment of the present invention, since icing conditions are typically affected by time cycles such as intraday variations and seasonal changes, the samples can be modified by adding additional labels to allow the model to perceive the impact of time cycles on icing. Optionally, the labels also include the intraday time at which the sample was collected and the season in which it was collected. For intraday time, hours (e.g., 10, 15, 20) can be used; for season, spring, summer, autumn, and winter can be used. For example, Sample 1: collected at 10:00 AM, spring; Sample 2: collected at 3:00 PM, summer; Sample 3: collected at 8:00 PM, autumn; during training, each label can be in the form of a one-hot code.

[0066] (2) Model structure

[0067] According to one embodiment of the present invention, the icing situation prediction model includes a fusion feature extraction module, a graph neural network, a Transformer network, and a prediction layer. The various sub-modules in the model are described below.

[0068] According to one embodiment of the present invention, a fusion feature extraction module is used to perform self-attention weighted fusion of the parameters in its multi-source feature parameters to obtain fused features. Let the set of parameters extracted from the multi-source feature parameters be F = {f1, f2, ..., f...} m}, where f1~ f m This represents the first to m parameters in the set. For example, assuming the set has only 3 parameters, then f1 is the rate of change of icing data, f2 is the maximum icing difference on the leaf, and f... m This represents the temperature deviation value. A fusion feature extraction module is defined, which calculates the weight of each feature through self-attention, expressed as:

[0069] αᵢ = Att(fᵢ), and Σᵢ αᵢ = 1.

[0070] The fusion feature is then represented as:

[0071] F_fusion = Σᵢ (αᵢ · fᵢ).

[0072] Here, αᵢ represents the i-th attention weight, Att(·) represents the self-attention calculation function, fᵢ represents the i-th parameter among the multi-source feature parameters, and F_fusion represents the fused feature. αᵢ is adaptively adjusted by the model based on the current input, making the fused feature F_fusion most discriminative of the icing state. This scheme designs a dynamic feature fusion strategy based on a self-attention mechanism: multi-dimensional features from different sources are input into a lightweight attention network, the current weight coefficients of each feature are calculated, and the features are fused according to their weights to form a fused feature, adaptively highlighting key features and weakening secondary features. For example, when the ambient temperature changes drastically, the weight of temperature-related features is automatically increased; conversely, when the temperature is stable but the capacitance value corresponding to icing is abnormal, the weight of icing-related features is emphasized. This allows the model to automatically adjust feature weights according to environmental changes, achieving adaptive feature selection, effectively suppressing noise interference, improving the accuracy of icing identification under complex meteorological conditions, and reducing false alarms and missed alarms.

[0073] According to one embodiment of the present invention, a graph neural network is used to extract spatial features from the topology graph. Stacking spatial features from multiple time points in temporal order yields a multi-source spatial feature sequence. The graph neural network employs a graph convolutional network or a graph attention network. Since multiple sensor nodes are deployed on the wind turbine blades, icing conditions at different locations may exhibit correlated spatiotemporal patterns, such as icing starting at one location and spreading to other locations. This invention utilizes a graph neural network to extract features based on the spatial correlation of the sensor array. First, a topology graph of the sensor nodes on the blade is constructed based on the geometric layout of the sensors on the blade: each sensor node is abstracted as a node in the graph, and edges are established between nodes based on physical distance or historical correlation. For example, an edge can be connected to a pair of sensors that are adjacent in distance and have a past icing event sequence relationship; the weight of the edge can depend on the distance between the sensors (the closer the distance, the greater the weight) or other relevant indicators. Then, at each discrete time step, the fused feature of all sensor nodes at that moment (i.e., F_fusion, one for each sensor node) is used as the initial input representation of the graph node. The topology graph is computed using a Graph Convolutional Network (GCN) or a Graph Attention Network (GAT), with the model integrating information from neighboring nodes to update the representation of each node. After several layers of graph neural network computation, the spatial feature representation of the entire blade at the current moment is extracted. In short, the graph neural network outputs a new feature vector that encodes the overall distribution characteristics of the icing state at each monitoring point on the current blade as spatial features. The spatial features implicitly contain the correlations between the readings of each sensor node, such as whether there are locally concentrated icing areas and the location of the area where icing first occurred. By utilizing the spatial feature analysis capabilities of graph neural networks, this invention can fully explore the collective information of multiple sensor nodes, which is more comprehensive and accurate than point-by-point independent analysis. This spatial feature analysis also provides high-level input features for subsequent time-series models (Transformer networks), providing a more reliable basis for judging the icing state and trend.

[0074] According to one embodiment of the present invention, a Transformer network is used to process multi-source spatial feature sequences to obtain spatiotemporal features. The Transformer network is lightweight. Preferably, the self-attention encoding module of the Transformer network has only one or two encoding layers. The spatial features of the graph neural network are stacked with time series and fed into the Transformer network to complete the discrimination of icing state and trend. The multi-head self-attention mechanism of the Transformer network is good at capturing long-distance temporal dependencies, enabling the model to simultaneously focus on short-term drastic changes and long-term cumulative trends in the leaf icing process. In implementation, if computing power allows, a standard Transformer model can be used. In some cases, to adapt to the computing power limitations of edge devices, the standard Transformer model can be pruned and optimized: for example, only one or two self-attention encoder layers can be retained, redundant feedforward fully connected layers can be removed, and sparse self-attention techniques such as ProbSparse can be used to reduce computational complexity. In addition, knowledge distillation pre-training of a small model on a large model can be performed on a cloud device to ensure that the small model, after significantly reducing the number of parameters, still maintains high accuracy and has a good learning ability for icing patterns. Based on the model's output, the system can determine whether the wind turbine blades are iced and the severity of the icing, and trigger corresponding alarms or subsequent processing accordingly. Compared to the traditional LSTM model, the Transformer network of this invention, as a temporal model, performs better in capturing long-period dependencies and complex periodic patterns, improving the accuracy and stability of icing situation awareness. If the Transformer network's self-attention encoding module has only one or two encoding layers, the inference speed can be faster to meet the field requirements of real-time output from edge devices.

[0075] According to one embodiment of the present invention, a prediction layer is used to determine the icing situation perception result of the blades based on spatiotemporal features. The prediction layer can be implemented using a fully connected layer and a softmax function. Specifically, the fully connected layer of the prediction layer converts the spatiotemporal features into predicted logistic values ​​for each category, and the softmax function determines the confidence level of each icing state category and icing trend category based on the predicted logistic values ​​for each category. The icing situation perception result is the icing state category with the highest confidence and / or the icing trend category with the highest confidence.

[0076] According to one embodiment of the present invention, in order to improve the model's sensitivity to seasonal cyclical patterns, if time series periodic embedding information (i.e., intraday time and season) is intentionally added to the labels as additional features to the model, it helps the model understand the position of the current time point within a day or year. Based on the aforementioned structure, a period prediction module is also provided to assist training. This period prediction module is configured to analyze the time confidence of the corresponding samples collected at each time point and the seasonal confidence of the samples collected in each season based on spatiotemporal characteristics. The trained model can accurately identify the relevant implicit information about the changes in icing conditions over time, such as whether the ice thickness is continuously increasing or decreasing, and can filter out false signals caused by transient anomalies such as sudden changes in wind speed.

[0077] (3) Model training

[0078] According to one embodiment of the present invention, the icing situation prediction model needs to be trained in the following manner before use: an icing situation prediction model to be trained is obtained; a training set is obtained, including multiple samples, each sample including a sensor data sequence composed of sensor data collected periodically at multiple times and a label, the label being the ground truth of the icing situation, including the icing state category and icing trend category to which the sensor data belongs; multi-source feature parameters are constructed based on the sensor data in the samples of the training set, and the multi-source feature parameters are input into the icing situation prediction model to output the icing situation perception confidence of the blade; the situation perception loss value calculated based on the difference between the icing situation perception confidence of the blade and the label guides the update of the model parameters to minimize the loss value. Optionally, the label may only contain the icing state category to which the sensor data belongs. The situation perception loss value can be calculated using the cross-entropy loss function.

[0079] According to one embodiment of the present invention, a period prediction module can also be set up during training to assist in training. The period prediction module is configured to analyze the time confidence and seasonal confidence of the corresponding samples collected at each time and in each season based on spatiotemporal characteristics; calculate the situational awareness loss based on the difference between the leaf icing situational awareness confidence and the label; calculate the period prediction loss based on the difference between the time confidence, seasonal confidence, and the label; and guide the parameter update of the model and the period prediction module based on the weighted loss value of the situational awareness loss and the period prediction loss to minimize the weighted loss value. The period prediction loss can also be calculated using the cross-entropy loss function.

[0080] According to one example of the present invention, this invention innovatively introduces a spatiotemporal fusion deep learning model at the algorithm level for deep pattern mining and icing state discrimination of sensor array data. Spatially, a sensor node topology map is constructed based on the distribution of sensors on the blades, and a graph neural network (GNN) is introduced to extract relevant features between different locations on the blades. Temporally, a pruned and optimized Transformer network is used to extract time-series features. Illustratively, the GNN uses graph convolution or graph attention mechanisms to aggregate features from sensor nodes at different locations at the same time, capturing spatially relevant information such as the starting location and propagation pattern of icing. The Transformer network models long-sequence dependencies based on a multi-head self-attention mechanism, enabling it to simultaneously focus on short-term drastic fluctuations and long-term cumulative trends during the icing process. To adapt to the limited computing power of terminal or edge devices, the Transformer network is lightweighted and optimized: only one or two self-attention encoder layers are retained, and redundant fully connected layers are removed. The ProbSparse self-attention technique is introduced to reduce the computational complexity from O(L...) to... 2 The computational cost is reduced to O(Llog L), and the simplified model is pre-trained on cloud devices by distilling smaller models from larger ones. This significantly reduces the number of parameters while maintaining accuracy, making it easier to deploy and run on MCUs. Through the above spatiotemporal fusion design, the model used in this invention can simultaneously capture the spatial distribution characteristics and temporal evolution of icing, overcoming the limitation of traditional LSTM time series models that rely solely on LSTM to characterize spatial relationships. In particular, the overall spatial features of the blade extracted by GNN condense the icing distribution state of the entire blade at the current moment, which can identify local icing hotspots and guide the optimal allocation of monitoring resources (such as increasing the sampling frequency of sensors in hotspot areas and reducing the frequency in non-critical areas to save energy). Transformer networks are good at capturing long-range dependencies and seasonal periodic effects, and have a significant improvement in the accuracy and stability of icing detection compared to traditional LSTM models. They can identify complex patterns and provide early warnings of icing trends (such as identifying signals of continuous ice thickening). In addition, the model structure can be simplified using the TinyML principle, resulting in a small parameter size and fast inference speed, enabling real-time output on terminal or edge devices to meet the needs of real-time on-site alarms. Compared to existing large-scale models that require cloud computing, the model used in this invention can run efficiently, achieving a unification of high-precision detection and embedded real-time deployment.

[0081] Step S4: For each sensing node, the fusion feature extraction module is used to perform self-attention weighted fusion of the parameters in its multi-source feature parameters to obtain the fused features.

[0082] According to one embodiment of the present invention, an attention score is calculated for each parameter among the multi-source feature parameters using a small neural network (such as a 2- to 3-layer network), representing the relative contribution of that parameter to the current icing situation perception task. Then, the parameters are weighted and summed according to their attention scores to obtain a unified feature representation after fusion, i.e., the fused feature. Thus, the model can adaptively highlight important features and weaken secondary features, fully utilizing multi-source information to improve discrimination performance. This mechanism achieves intelligent fusion of multimodal sensor data, improving the robustness of feature representation and reducing redundancy and noise interference compared to simple splicing. Compared to single-sensor data, it can more comprehensively extract features reflecting the icing status and / or trend of the blades.

[0083] Step S5: Obtain the topology graph corresponding to multiple sensor nodes, set the features of the graph nodes of the topology graph as the fusion features of the corresponding sensor nodes, use graph neural network to extract spatial features from the topology graph, stack the spatial features of multiple time steps in time sequence to obtain a multi-source spatial feature sequence.

[0084] According to one embodiment of the present invention, edges and their weights are established in the topology graph based on the physical distance between graph nodes and / or historical icing correlation. For example, edges are added between graph nodes corresponding to sensor nodes whose physical distance is less than or equal to a preset distance threshold, and the weight of the edge is set to the physical distance or a normalized value of the physical distance. No edges are set between other graph nodes. Alternatively, edges can be set between graph nodes corresponding to sensor nodes that have a historical icing sequence relationship, and the weight of the edge is set to the correlation degree of the icing sequence relationship. The correlation degree can be set based on historical statistical values ​​of the icing sequence relationship.

[0085] Step S6: Use the Transformer network to process the multi-source spatial feature sequence to obtain spatiotemporal features.

[0086] According to one embodiment of the present invention, if the spatial features of each time step in the multi-source spatial feature sequence are taken as an element of the sequence, the Transformer network can be used to mine the temporal features between spatial features in a time series processing manner, thereby mining spatiotemporal features that contain both spatial and temporal information.

[0087] Step S7: Based on the spatiotemporal characteristics, use the prediction layer to determine the icing situation perception results of the blades.

[0088] According to one embodiment of the present invention, spatiotemporal features can be processed by a fully connected layer and a Softmax function connected sequentially in the prediction layer to obtain the icing situation perception result.

[0089] According to an embodiment of the present invention, an edge-cloud collaborative wind turbine blade icing situation perception system is provided. The system includes: a terminal device, which includes multiple sensing nodes for collecting sensing data from the blade and a terminal processor for executing the icing situation perception method as described in the foregoing embodiment; an edge device, which is configured to: communicate with the terminal device, collect the sensing data collected by the terminal device and its corresponding icing situation perception results, and notify a local monitoring platform and a cloud device when a sensing anomaly is detected based on the sensing data and its corresponding icing situation perception results; and a cloud device, which is configured to: communicate with the edge device, periodically acquire the sensing data, video monitoring data and human feedback data collected by the edge device to update the sample database of each wind turbine, and when a sensing anomaly occurs in a terminal device, update the icing situation perception result of the icing situation prediction model for the blade using the prediction result of a cloud model that is more accurate than the icing situation prediction model, and remind human intervention to label or clean abnormal samples so as to periodically guide the training of the icing situation prediction model based on the collected samples using the cloud model as a teacher model. Among them, the sensing anomaly includes the lowest temperature data in the sensing data being greater than the preset temperature reference value, but the icing situation sensing results show that the blades are icy.

[0090] According to one example of the present invention, a layered system architecture design for cloud-edge-device collaboration is provided, see [link to relevant documentation]. Figure 3The edge-cloud collaborative wind turbine blade icing situation awareness system includes terminal devices, edge devices, and cloud devices. Optionally, the system may also include a monitoring platform. The terminal devices may include terminal processors (such as embedded MCU units) installed on each wind turbine blade, capable of independently performing real-time data acquisition, preprocessing, and local icing situation awareness. Edge devices may be industrial control computers or site servers located in the wind farm nacelle, or may be integrated with cloud devices. Cloud devices refer to remote central servers or cloud platforms, acting as the upper layer responsible for aggregating data from the entire field, performing deep analysis, and updating model training. When terminal devices encounter rare or complex operating conditions that lead to low model output confidence, relevant data can be uploaded to the cloud devices. The cloud devices can then use more complex large models or remote manual diagnostics for auxiliary analysis, thereby improving the accuracy of judgment in difficult situations. To ensure that the algorithm maintains high accuracy over a long period under different wind farm environments, this invention introduces an online transfer learning and cloud-edge-device collaborative model update mechanism. After system deployment, the cloud device periodically collects historical data and results uploaded by each edge device and compares and evaluates them with actual situation records (such as manual inspection reports or observations from other precision sensors in the nacelle). When it is found that the model discrimination error of some wind turbines is gradually increasing, or the distribution of environmental data changes significantly, the model update process is initiated. First, the cloud device retrains or fine-tunes the original model based on the large amount of new data collected. To avoid drastically changing the structure of the deployed model, the preferred method is to freeze most of the model's layers and only adjust the weight parameters of the fully connected layers of the prediction layer (or add a small number of parameters) to learn new environmental features, thereby achieving rapid migration. For extremely special new situations (such as the emergence of a completely new icing pattern or meteorological conditions), an improved new model can also be trained on the cloud device, and then a simplified version can be distributed to replace the old model on the terminal. At the same time, the terminal device or edge device can also perform simple incremental learning steps locally, such as dynamically updating the model parameters based on recent results and feedback, and continuously adapting to the current environment. By combining edge and cloud computing, the model used in this invention can continuously evolve: cloud devices possess powerful computing resources and a global data perspective, enabling them to periodically train new, more generalized models and send them to terminal devices for deployment; terminal devices, in turn, ensure real-time performance and provide feedback of actual operational data to the cloud devices. This iterative cycle ensures continuous optimization of model performance throughout its lifecycle, avoiding the performance stagnation and lack of improvement seen in traditional solutions after model training. Verification has shown that this invention effectively solves the problems of poor model generalization and aging failure through continuous learning, ensuring that the system maintains stable and reliable icing monitoring capabilities in different wind fields and years.

[0091] According to one example of the present invention, the cloud device continuously collects historical sensor data of each wind turbine blade for big data analysis, extracts universal patterns, optimizes model parameters, and periodically transfers relevant knowledge to the icing situation prediction model using knowledge distillation technology and online transfer learning mechanisms. This knowledge is then distributed to the terminal devices, enabling the icing situation prediction model to continuously evolve with environmental changes and maintain optimal performance. The cloud-edge-device collaborative architecture combines the advantages of real-time on-site response and global optimization, giving the system the ability to "get smarter with use," effectively solving the problem in existing technologies where models struggle to adapt to new environments after a single offline training session. Terminal devices can quickly make initial judgments about icing situations; the cloud device performs more complex and refined analysis, providing a powerful supplement to the terminal device's prediction results and further improving overall perception accuracy. Through the online transfer learning mechanism, the model can continuously update itself according to different wind field climates, maintaining high accuracy and a low false alarm rate even after years of operation. It is evident that the hierarchical collaborative design of the system of the present invention not only ensures the real-time accuracy of icing situation perception for a single wind turbine blade, but also allows cloud devices to collect sample data across wind fields. Through sharing and continuous optimization, the adaptability and long-term effectiveness of icing situation perception are significantly improved.

[0092] Optionally, under the cloud-edge-device collaborative architecture, the system of this invention also possesses flexible division of labor and fault tolerance capabilities: Normally, the terminal devices perform icing situation awareness and report the results to the edge devices and / or the wind farm monitoring platform; when the confidence level of the results is insufficient or an anomaly is encountered, the edge devices can request the intervention of the cloud devices. After receiving the reported anomaly and related data, the cloud devices can call more complex deep learning models or expert systems for in-depth analysis, providing more accurate and detailed results, and feeding the results back to the edge devices. For example, if the confidence level of the model output is lower than a set threshold (meaning the model is not confident enough), the sensor data is sent to the cloud devices for further judgment to ensure the accuracy of the results. Once icing is confirmed, or the icing status level exceeds the threshold, or the trend indicator exceeds the threshold, the results can be uploaded to the wind farm monitoring platform via the communication network, or a warning indicator light / buzzer alarm can be activated locally to remind maintenance personnel of the current icing level and trend. For severe icing with an aggravating trend, the system's early warning can prompt maintenance personnel to take timely measures such as power reduction operation and manual de-icing to prevent accidents. Finally, the edge device uploads the collected data and analysis results to the cloud device for archiving. When the cloud device detects enough new data or a decrease in model accuracy that triggers the update, it can automatically initiate the model update mechanism: incrementally train and fine-tune the model on the accumulated new data, compress the model using methods such as quantization and pruning, and then distribute the updated lightweight model to the terminal device to replace the old model, ensuring that the system's recognition capability is always in optimal condition.

[0093] In summary, some embodiments of the present invention can achieve at least one of the following beneficial effects:

[0094] (1) This invention organically combines multi-source data fusion, spatiotemporal modeling, and cloud-edge-device collaboration, which significantly improves the system's intelligence and adaptability while ensuring real-time and high-precision detection of icing on wind turbine blades. The technical solution provided by this invention can be widely applied in the field of icing monitoring of wind turbine generators, and has significant practical value and promotion prospects.

[0095] (2) By deploying multiple sensor nodes along the longitudinal direction on the surface of the wind turbine blade to collect multi-source sensing data, and using an optimized graph neural network and Transformer network to preprocess, fuse features and model the spatiotemporal data, the accurate real-time detection and trend prediction of the blade icing state can be achieved, while supporting online updates and adaptation.

[0096] (3) To efficiently extract icing features on resource-constrained equipment, an intelligent data preprocessing algorithm was designed, such as sliding window calculation of short-term changes, wavelet transform for denoising and trend separation, etc., to extract key feature parameters representing the icing situation in real time while strictly controlling the computational complexity.

[0097] (4) In order to construct an AI model that takes into account both spatial and temporal dimensions, capture the correlation of icing at multiple sensor nodes in different locations, and the long-term and short-term patterns of icing evolution over time, the inventors innovatively combined graph neural networks and Transformer networks to extract spatiotemporal features. While ensuring recognition accuracy, the model size was significantly compressed to adapt to the deployment requirements of devices with insufficient computing power.

[0098] (5) To improve the system’s adaptability and evolution capability to different environments, a cloud-edge-device collaborative mechanism is introduced, which can realize online model updates and continuous learning of new data distributions according to different wind field environments, maintain high accuracy and extend the model life cycle.

[0099] It should be noted that although the steps are described in a specific order above, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required function can be achieved.

[0100] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0101] Computer-readable storage media can be tangible devices that hold and store instructions for use by an instruction execution device. Computer-readable storage media can include, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof.

[0102] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for sensing the icing situation of wind turbine blades, comprising: Acquire sensor data from multiple sensor nodes distributed on the leading edge and / or windward side of the blade, each sensor node including an icing sensor and a temperature sensor. Multi-source feature parameters for each sensor node are constructed based on the sensor data, including: the rate of change of icing data, the maximum icing difference on the blade, and the temperature deviation value. Obtain a trained icing situation prediction model, which includes a fusion feature extraction module, a graph neural network, a Transformer network, and a prediction layer; For each sensing node, the fusion feature extraction module performs self-attention weighted fusion of the parameters in its multi-source feature parameters to obtain the fused features; Obtain the topology graph corresponding to multiple sensor nodes, set the features of the graph nodes of the topology graph as the fusion features of the corresponding sensor nodes, use graph neural network to extract spatial features from the topology graph, and stack the spatial features of multiple time steps in time sequence to obtain a multi-source spatial feature sequence. The Transformer network is used to process multi-source spatial feature sequences to obtain spatiotemporal features; Based on spatiotemporal characteristics, the icing situation perception results of the blades are determined using the prediction layer.

2. The method according to claim 1, characterized in that, In the topological graph, edges and their weights are established based on the physical distance between graph nodes and / or historical icing correlation. Graph convolutional networks or graph attention networks are used as graph neural networks to extract spatial features from the topological graph, resulting in a multi-source spatial feature sequence; The Transformer network, which processes multi-source spatial feature sequences, is lightweight, with its self-attention coding module having only 1 or 2 coding layers.

3. The method according to claim 1 or 2, characterized in that, The icing situation prediction model needs to be trained in the following way before use: Obtain the icing situation prediction model to be trained; Obtain a training set, which includes multiple samples. Each sample consists of a sequence of sensor data collected periodically at multiple times and a label. The label is the ground truth of the icing situation, which includes the icing state category, or includes the icing state category and the icing trend category. Multi-source feature parameters are constructed based on the sensor data in the training set samples. The multi-source feature parameters are input into the icing situation prediction model, and the icing situation perception confidence of the blade is output. The situation perception loss value calculated based on the difference between the icing situation perception confidence of the blade and the label guides the update of the model parameters to minimize the loss value.

4. The method according to claim 3, characterized in that, The label also includes the time of day and season in which the sample was collected, and the method further includes: During training, a period prediction module is also set up to assist training. The period prediction module is configured to analyze the time confidence of the corresponding sample at each time when it is collected and the seasonal confidence of it at each season based on the spatiotemporal characteristics. The situational awareness loss is calculated based on the difference between the confidence level of the icing situational awareness of the blades and the label. The period prediction loss is calculated based on the difference between the confidence level of time, the confidence level of season, and the label. The weighted loss value of the situational awareness loss and the period prediction loss guides the parameter update of the model and the period prediction module to minimize the weighted loss value.

5. The method according to claim 1 or 2, characterized in that, The process of constructing multi-source feature parameters for each sensor node based on sensor data includes: The icing data and temperature data in the sensor data are processed by sliding window in turn, and the average value of the local data captured by each sliding window is calculated to obtain the average value of icing and temperature. Wavelet transform was applied to the average values ​​of icing and temperature for denoising and trend separation, respectively, to obtain the denoising results for icing and temperature. The rate of change of icing data is calculated using the denoising results of the previous and current time points. The maximum icing difference on the blade is obtained by subtracting the minimum value from the maximum value of the icing data at the current time point. The minimum value in the denoised temperature result is subtracted from the preset baseline temperature to obtain the temperature deviation value.

6. The method according to claim 5, characterized in that, The multi-source feature parameters also include: temperature data, icing data, and / or temperature gradient, wherein the temperature gradient is the rate of temperature change calculated based on the temperature data or the denoising result of the temperature.

7. An edge-cloud collaborative wind turbine blade icing situational awareness system, the system comprising: A terminal device, comprising a plurality of sensing nodes configured to acquire sensing data from the blades and a terminal processor for performing the method as described in any one of claims 1-6; Edge devices are configured to: communicate with terminal devices, collect sensor data collected by terminal devices and their corresponding icing situation perception results, and notify the local monitoring platform and cloud devices when a perception anomaly is detected based on the sensor data and their corresponding icing situation perception results. The cloud device is configured to: communicate with edge devices, periodically acquire sensor data, video surveillance data, and human feedback data collected by the edge devices to update the sample database of each wind turbine, and when a terminal device experiences a perception anomaly, update the icing situation prediction model's perception of the blade's icing situation using the prediction results of the cloud model, which is more accurate than the icing situation prediction model, and remind human intervention to label or clean abnormal samples so as to periodically guide the training of the icing situation prediction model using the cloud model as a teacher model based on the collected samples.

8. The system according to claim 7, characterized in that, The sensing anomaly includes a situation where the lowest temperature data in the sensing data is greater than the preset temperature reference value, but the icing situation sensing results show that the blades are icy.

9. A computer-readable storage medium, characterized in that, It stores a computer program that can be executed by a processor to implement the steps of the method according to any one of claims 1-6.

10. An electronic device, characterized in that, include: One or more processors; as well as Memory, wherein the memory is used to store executable instructions; The one or more processors are configured to implement the steps of the method according to any one of claims 1-6 by executing the executable instructions.

Images