Wind turbine variable pitch system fault prediction method and system based on neural network

Abstract

The application discloses a neural network-based wind turbine variable pitch system fault prediction method and system. Based on deep time sequence feature coding of original SCADA data, an innovative space-time dual saliency mechanism is used to enable the model to self-adaptively focus on and amplify the weak features of the most relevant multi-dimensional data in the time and space dimensions of fault evolution, while suppressing the interference of irrelevant working condition information. In this way, the accuracy and reliability of wind turbine variable pitch system fault prediction are significantly improved, and the false positive rate is effectively reduced, thereby providing key technical support for realizing precise predictive maintenance.

Description

Technical Field

[0001] This application relates to the field of intelligent prediction, specifically to a method and system for predicting faults in wind turbine pitch systems based on neural networks. Background Technology

[0002] Wind power, as a clean energy source, occupies an increasingly important position in the global energy structure. The pitch system of a wind turbine is a core component for regulating power output and ensuring the safe operation of the unit; its operating status directly affects the efficiency and reliability of wind power generation. Because the pitch system operates in harsh environments for extended periods, malfunctions are frequent. Failure to detect and address these malfunctions promptly can lead to serious equipment damage and economic losses. Therefore, developing an accurate and reliable fault prediction scheme for wind turbine pitch systems is crucial for implementing predictive maintenance, reducing operation and maintenance costs, and ensuring the stable operation of wind turbines.

[0003] Currently, utilizing the massive amounts of data generated by wind turbine monitoring and data acquisition (SCADA) systems, combined with artificial intelligence technologies such as neural networks, for fault prediction has become a mainstream research direction in this field. However, existing prediction methods generally face the technical bottleneck of high false alarm rates in practical applications. The fundamental reason is that information representing changes in the normal operating conditions of the unit and fault characteristic information representing the degradation of the system's health status are often deeply coupled and entangled in the raw sensor data. On the one hand, the parameter fluctuations caused by the wind turbine control system's adjustments to cope with normal operating conditions such as drastic wind speed changes are very similar in data form to the abnormal parameter changes caused by early faults, making it difficult for the model to distinguish them. On the other hand, traditional neural network models are prone to misjudging normal drastic fluctuations as faults when facing complex operating conditions. This limitation, which lacks an effective mechanism for separating operating condition information from health status information, seriously restricts the accuracy and practical value of existing prediction technologies. Summary of the Invention

[0004] This application provides a neural network-based method and system for predicting faults in wind turbine pitch systems. Based on deep temporal feature encoding of raw SCADA data, it employs an innovative spatiotemporal dual saliency mechanism. This mechanism enables the model to adaptively focus on and amplify the weak temporal and spatial features of the multidimensional data most relevant to fault evolution, while suppressing interference from irrelevant operating condition information. This significantly improves the accuracy and reliability of fault prediction for wind turbine pitch systems, effectively reducing the false alarm rate and providing key technical support for achieving accurate predictive maintenance.

[0005] On one hand, embodiments of this application provide a method for fault prediction of wind turbine pitch system based on neural networks, including: Obtain raw SCADA data; Multidimensional time-series data preprocessing and windowing are performed on the raw SCADA data to obtain the time series of the multimodal data co-coding vector of the pitch system; The time series of the multimodal data co-coding vector of the pitch system is subjected to basic feature encoding to obtain the time series of the basic feature encoding vector of the multimodal data of the pitch system; Spatiotemporal doubly saliency is performed on the time series of the multimodal basic feature encoding vector of the pitch system to obtain the spatiotemporal saliency feature vector of the multimodal pitch system; Fault classification is performed on the multimodal spatiotemporal salient feature vectors of the pitch system to obtain the predicted fault probability.

[0006] On the other hand, this application also provides a fault prediction system for wind turbine pitch system based on neural networks, including: The data acquisition module is used to acquire raw SCADA data; The data processing module is used to perform multi-dimensional time-series data preprocessing and windowing on the raw SCADA data to obtain the time series of the multi-modal data co-coding vector of the pitch system; The basic feature encoding module is used to perform basic feature encoding on the time series of the multimodal data co-encoding vector of the pitch system to obtain the time series of the basic feature encoding vector of the multimodal data of the pitch system. The spatiotemporal doubly saliency module is used to perform spatiotemporal doubly saliency on the time series of the multimodal basic feature encoding vector of the pitch system to obtain the spatiotemporal saliency feature vector of the multimodal pitch system; The fault classification module is used to classify faults in the multimodal spatiotemporal salient feature vectors of the pitch system to obtain the predicted fault probability.

[0007] Compared with existing technologies, this application provides a neural network-based fault prediction method and system for wind turbine pitch systems. Based on deep temporal feature encoding of the original SCADA data, it employs an innovative spatiotemporal dual saliency mechanism. This allows the model to adaptively focus on and amplify the weak temporal and spatial features of the multidimensional data most relevant to fault evolution, while suppressing interference from irrelevant operating condition information. This significantly improves the accuracy and reliability of fault prediction for wind turbine pitch systems, effectively reducing the false alarm rate and providing key technical support for achieving accurate predictive maintenance. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a flowchart of a neural network-based fault prediction method for wind turbine pitch systems according to an embodiment of this application; Figure 2 This is a data flow diagram of a neural network-based fault prediction method for wind turbine pitch systems according to an embodiment of this application. Figure 3 This is a block diagram of a neural network-based wind turbine pitch system fault prediction system according to an embodiment of this application. Detailed Implementation

[0010] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0011] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0012] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0013] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0014] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0015] In the technical solution of this application, a fault prediction method for wind turbine pitch system based on neural network is proposed. Figure 1 This is a flowchart of a neural network-based fault prediction method for wind turbine pitch system according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow for a neural network-based fault prediction method for wind turbine pitch systems according to an embodiment of this application. Figure 1 and Figure 2 As shown, the neural network-based fault prediction method for wind turbine pitch system according to an embodiment of this application includes the following steps: S1, acquiring raw SCADA data; S2, performing multi-dimensional time-series data preprocessing and windowing on the raw SCADA data to obtain a time series of multi-modal data co-coding vectors for the pitch system; S3, performing basic feature encoding on the time series of multi-modal data co-coding vectors for the pitch system to obtain a time series of multi-modal basic feature encoding vectors for the pitch system; S4, performing spatiotemporal doubly saliency on the time series of multi-modal basic feature encoding vectors for the pitch system to obtain a spatiotemporal saliency vector for the pitch system; S5, performing fault classification on the spatiotemporal saliency vectors for the pitch system to obtain a predicted fault probability.

[0016] Specifically, S1 involves acquiring raw SCADA data. It should be understood that during wind turbine operation, the monitoring and data acquisition (SCADA) system continuously records massive amounts of sensor data and status parameters, which constitute a complete description of the unit's behavior. Crucially, this raw data simultaneously contains two intertwined key pieces of information: first, the normal adjustment behavior of the unit in response to external environmental factors such as wind speed changes, i.e., operating condition information; and second, performance degradation or early fault symptoms caused by internal factors such as equipment wear and aging, i.e., health status information. The main bottleneck of existing technologies lies in the inability to effectively separate these two deeply coupled pieces of information, leading to a large number of false alarms in predictive models. Therefore, acquiring unprocessed raw SCADA data is a fundamental prerequisite for preserving the integrity of all information and providing the possibility for subsequent steps to apply advanced algorithms to decouple these two pieces of information. It is worth mentioning that raw SCADA data usually exists in the form of time series. Each record contains a precise timestamp and the measured values ​​of multiple monitored variables at that moment. Furthermore, raw SCADA data is multi-dimensional, covering multiple variables that affect the state of the pitch system, such as wind speed, generator power, blade angle, pitch motor current, pitch motor temperature, and pitch drive status.

[0017] In practice, a communication connection can be established with the SCADA database of the wind farm or the local controller of the wind turbine, and data can be requested and received through preset interface protocols (such as OPC, Modbus, etc.). During this process, the system will initiate data queries according to the set acquisition frequency (e.g., once per second, per minute, or once every 10 minutes) to obtain a continuous time-series data stream. This process ensures that all subsequent analyses are based on the same true and complete historical and real-time operation records.

[0018] Specifically, S2 involves performing multi-dimensional time-series data preprocessing and windowing on the raw SCADA data to obtain the time series of the multimodal data co-encoding vector for the pitch system. The raw SCADA data is collected directly from the industrial field and may contain missing values ​​(due to sensor malfunctions or communication interruptions), outliers, or isolated points (due to measurement errors or extreme environmental events). This noise can severely interfere with the model's learning process. Secondly, the SCADA system collects multimodal data, i.e., signals from different sensors (such as blade angle, motor current, hydraulic oil temperature, vibration, etc.). These signals have vastly different physical units, numerical ranges (dimensions), and amplitudes of variation. Without normalization, this can lead to unstable model training, causing features with larger numerical ranges to dominate the model. Furthermore, neural networks, especially recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) that process time series data, require fixed-length sequences as input samples. Since the original SCADA data is a continuous data stream, the technical solution of this application further windowes the original SCADA data to divide it into segments with consistent time lengths. Each segment represents an independent event or process that can be analyzed by the model, so that the model can learn the dynamic evolution of the data over a specific time span.

[0019] In practice, the first step is multidimensional time-series data preprocessing, which can be further divided into data cleaning, feature selection, and data normalization. During data cleaning, missing values ​​in the original data need to be addressed. Common methods include filling with specific values ​​(such as 0 or the mean), filling with the previous valid observation (forward filling), or filling with the next valid observation (backward filling), or using linear interpolation for data with a clear linear trend. Simultaneously, outliers need to be identified and processed. For example, statistical methods (such as the 3-sigma rule) or more complex algorithms (such as isolation forests) can be used to remove or correct outliers that significantly deviate from the normal data distribution. During feature selection, signals most relevant to the pitch system's health status need to be selected from numerous SCADA variables as input features for the model. This can be done based on expert experience or using feature importance assessment algorithms. Data normalization aims to scale all selected feature data to a uniform, smaller interval, typically [-1, 1]. A commonly used normalization method is min-max normalization. Secondly, there's windowing. That is, after data preprocessing, the continuous time series data is divided into a series of fixed-length segments that can be used as individual training samples. This process can be implemented using a sliding window. In this process, the module sets a window length W and a sliding step size S. Assuming the preprocessed multidimensional time series has a total of L time points, the windowing operation will start from the beginning of the sequence, extract a data segment of length W as a sample, then slide the window forward by S time points, extract the next sample of length W, and so on, until the window slides to the end of the sequence. In this way, a long sequence is transformed into a batch of two-dimensional matrix samples of shape [W, D] that can be used for model training, where D is the dimension of the SCADA data (i.e., the number of variables).

[0020] Specifically, in S3, the basic feature encoding of the time series of the multimodal data co-encoding vector of the pitch system is performed to obtain the time series of the multimodal basic feature encoding vector of the pitch system. It should be understood that although the time series of the multimodal data co-encoding vector of the pitch system obtained in the previous stage integrates information from multiple sensors at the same time, each vector in the sequence is isolated and does not explicitly contain its temporal correlation information. The operating state of the wind turbine pitch system is a continuous dynamic process; the occurrence of faults is often not an instantaneous event, but rather a gradual accumulation and evolution of a series of minute state changes. Therefore, analyzing only a snapshot of the state at a single time point is insufficient to capture these early fault signs. Deep encoding of the entire time series is necessary to effectively extract the patterns, trends, and long-term and short-term dependencies of the data over time. This deep temporal feature extraction is crucial for distinguishing normal operating condition fluctuations from abnormal fault precursors and is a necessary prerequisite for improving the accuracy and lead time of fault prediction.

[0021] In practice, the time series of the co-encoded vector of the multimodal data of the pitch system is input into a bidirectional LSTM network to obtain the time series of the basic feature encoding vector of the multimodal data of the pitch system. The bidirectional LSTM network is the specific technical means to achieve basic feature encoding. It is a special type of recurrent neural network (RNN) that effectively solves the gradient vanishing or exploding problem that traditional RNNs easily encounter when processing long sequences through its internal gating units (input gate, forget gate, output gate). Its bidirectional structure gives the model the ability to simultaneously utilize historical and future information for encoding. Specifically, the execution process includes the following core operations: First, the time series of the co-encoded vector of the multimodal data of the pitch system generated in the previous stage is used as input to the bidirectional LSTM model, which consists of two independent LSTM layers, one for forward processing and the other for backward processing. The forward LSTM layer starts from the beginning of the time series (t=1) and processes it sequentially to the end (t=T). At each time point t, it generates a forward hidden state based on the current input and the hidden state of the previous time step. This state encodes historical information from the beginning of the sequence to the current time step. Simultaneously, the backward LSTM layer starts from the end of the time series (t=T) and processes it in reverse chronological order to the beginning (t=1). At each time point t, it generates a backward hidden state based on the current input and the hidden state of the next time step. This state encodes future information from the end of the sequence to the current time step. Finally, at each time point t, the model concatenates or merges the forward and backward hidden states to form the final basic feature encoding vector for that time step. Through this bidirectional processing mechanism, each output feature vector simultaneously contains its past and future contextual information, enabling a more comprehensive and profound understanding of the true state and dynamic relationships of the data at any given time, greatly enhancing the completeness and effectiveness of feature representation.

[0022] Specifically, in step S4, the time series of the multimodal basic feature encoding vector of the pitch system is subjected to spatiotemporal dual saliency to obtain the spatiotemporal salient feature vector of the multimodal pitch system. It should be understood that while the time series of the multimodal basic feature encoding vector of the pitch system output in the previous step captures temporal dependencies, it treats each feature dimension and each time point equally. However, in actual operating conditions, not all sensor data have equally important indicative significance for the occurrence of a fault at all times. For example, in the early stages of a fault, only one or two key sensors may show weak but persistent abnormal fluctuations, while at other times, most data may be within the normal range. Therefore, an attention mechanism is needed to dynamically evaluate the importance of different features and different times, assigning higher weights to those highly correlated with potential faults. In the technical solution of this application, by performing spatiotemporal dual saliency on the time series of the multimodal basic feature encoding vector of the pitch system, the model can more accurately locate the source of the fault signal and the key nodes of its development, thereby greatly improving the sensitivity and accuracy of prediction.

[0023] In practical implementation, firstly, spatial attention weighting and feature recalibration are performed on the time series of the multimodal basic feature encoding vector of the pitch system to obtain a recalibrated feature sequence and a spatial attention weight sequence. It should be understood that in the actual operation of wind turbine units, different types of faults are often highly correlated only with anomalies in a few specific physical quantities or sensor signals. For example, hydraulic system faults may first manifest in the oil temperature or oil pressure feature dimension, while electrical faults may be reflected in the current or voltage feature dimension. Therefore, a mechanism must be introduced to dynamically evaluate and distinguish the importance of different feature dimensions at each moment. In the technical solution of this application, by performing spatial attention weighting and feature recalibration on the time series of the multimodal basic feature encoding vector of the pitch system, the model can focus on the most important feature combinations at present, concentrating computational resources on the dimension most likely to contain fault signals, thereby effectively separating valuable fault precursor information from complex background noise and irrelevant fluctuations.

[0024] In this process, firstly, for each time series of the multimodal basic feature encoding vector of the pitch system, an attention network is used to calculate the importance score of each feature dimension within it. Specifically, for the multimodal basic feature encoding vector of the pitch system at time t, it is input into a neural network consisting of one or more fully connected layers to generate an unnormalized attention score vector with the same dimension as the multimodal basic feature encoding vector of the pitch system at time t.

[0025] Secondly, to make these scores comparable and form a standard weight distribution, the score vector is normalized, which is usually done using the Softmax function. The Softmax function can transform a vector of arbitrary real values ​​into a probability distribution where each element's value is between 0 and 1, and the sum of all elements is 1. After processing with the Softmax function, the spatial attention weight vector at time t can be obtained; Finally, feature recalibration is performed. This operation involves element-wise multiplication (also known as Hadamard product) of the calculated spatial attention weight vector with the original multimodal basic feature encoding vector of the pitch system at time t, resulting in a recalibrated feature vector. This operation preserves or amplifies the original values ​​of feature dimensions assigned higher attention, while significantly reducing the values ​​of those considered less important, thus achieving dynamic weighting and recalibration of the features. After performing this operation on the feature vectors at each time step in the sequence, the recalibrated feature sequence and spatial attention weight sequence are obtained.

[0026] Next, dual-stream gated attention encoding is performed on the recalibrated feature sequence and the spatial attention weight sequence to obtain the fused hidden state sequence. In neural network-based wind turbine fault prediction methods, while existing spatiotemporal attention mechanisms can enhance key sensor signals through spatial attention weights, they have an inherent design flaw. This flaw lies in the fact that after recalibrating the original input features, the calculated spatial attention weight vector sequence is directly discarded, and only the recalibrated feature sequence is used for subsequent temporal dimension analysis. This approach ignores a crucial information dimension: the temporal dynamics of the spatial attention weights themselves. Specifically, the spatial attention weight sequence is not a meaningless intermediate product; it constitutes a high-value meta-level time series. This sequence accurately depicts the trajectory of the model's internal focus dynamically shifting between different sensors in response to subtle changes in the input signal. In the technical solution of this application, when the wind turbine operates under healthy and stable conditions, although its various parameters fluctuate, the internal correlation patterns are relatively fixed. Therefore, the spatial attention distribution generated by the model should also exhibit a relatively stable pattern matching this operating condition. However, when early, subtle signs of failure appear in the system, such as initial pitting in bearings or minor anomalies in the lubrication system, they can trigger intermittent, localized abnormal fluctuations in related sensor data that are difficult to detect with the human eye. The model can keenly capture these precursors and respond by dynamically adjusting its spatial attention allocation; that is, at the precise moment the anomaly occurs, it assigns higher attention weights to the affected sensors. Therefore, the shift in spatial attention patterns from stable to drastic or continuous changes is itself a powerful and independent early warning signal of a fault, even more sensitive than the recalibrated original signal itself. Existing technologies fail to explicitly model this information, thus missing valuable opportunities to capture some of the earliest fault characteristics. To overcome these technical shortcomings, a dual-stream gated temporal attention mechanism is proposed. Its core lies in the deep modeling and intelligent fusion of two parallel information streams: signal dynamics and attention dynamics. This allows the calculation of the final temporal attention weights to comprehensively consider what the signal says and how the model's interpretation of the signal changes.

[0027] In this process, firstly, it should be understood that signal dynamics and attention dynamics are two different but complementary information sources, requiring independent time-series models to capture their inherent evolutionary patterns. Specifically, on the one hand, the recalibrated feature sequence is input into the bidirectional LSTM model to fully extract the deep patterns of the sensor values ​​themselves evolving over time, obtaining the hidden state sequence. On the other hand, the spatial attention weight sequence is input into the bidirectional LSTM model to capture the dynamic patterns of the model's focus shifting between different sensors, obtaining the attention hidden state sequence.

[0028] Thus, in the technical solution of this application, the cognitive process of the model itself is digitized and serialized, making it an analyzable object, and two sets of parallel temporal feature expressions containing rich contextual information are obtained. In other words, the system simultaneously obtains a deep temporal understanding of the objective physical world (sensor values) and the model's subjective world (attention allocation). Furthermore, the hidden state sequence and the attention hidden state sequence are dynamically gated and fused to obtain the fused hidden state sequence. It should be understood that simple information splicing cannot reflect the relative importance of the two information streams at different points in time; at some moments, drastic changes in the signal itself may be more critical, while at other moments, abnormal shifts in attention patterns may be the real fault signal. Therefore, an adaptive mechanism is needed to dynamically weigh these factors.

[0029] Specifically, for each time step, the signal hidden state and the attention hidden state are first concatenated, and then a gating value between 0 and 1 is calculated using a fully connected layer and a sigmoid activation function. This gating value acts like a soft switch, controlling the contribution ratio from the two information streams; this process is expressed by the formula:

[0030]

[0031] in, For each hidden state in the hidden state sequence, For each attention hidden state in the attention hidden state sequence, yes The threshold value at any given time. Represents the Sigmoid activation function. and These are the learnable weights and bias parameters of the gating network. yes The hidden state after the moment is spliced ​​together.

[0032] Subsequently, the hidden states of the two streams are weighted and summed using this gating value to generate the fused hidden state; this process is expressed by the formula:

[0033] in, yes The hidden state after the moment merges. Represents element-wise multiplication. and They are The hidden state of real-time signals and attention. Correspondingly, when the wind turbine encounters severe but normal operating conditions such as strong gusts, the signal flow changes drastically, but the attention flow pattern remains stable. The gating mechanism can learn to reduce the weight of the attention flow to avoid false alarms. Conversely, when the fault signal is weak but causes a continuous abnormal drift in the attention pattern, the gating can increase the weight of the attention flow to achieve early warning. In this way, a more robust and comprehensive fusion feature sequence is generated, allowing the system to obtain a more intelligent feature representation that can dynamically adjust the information emphasis according to real-time conditions.

[0034] Furthermore, temporal attention encoding is performed on the fused hidden state sequence to obtain the multimodal spatiotemporal salient feature vector of the pitch system. That is, within the more information-rich feature sequence after intelligent fusion, the most crucial moment for determining the final fault is further identified. Specifically, the attention score for each time step is first calculated; this process is expressed by the formula:

[0035] in, yes Unnormalized attention score at each moment , , These are the learnable parameters of a temporal attention network; Then, the normalized temporal attention weights are obtained through the Softmax function, and these weights are used to perform a weighted summation on the fused hidden state sequence to obtain the final context vector; this process is expressed by the formula:

[0036]

[0037] in, It is the spatiotemporal salient feature vector of the multimodal pitch system. It is based on Calculated The temporal attention weighting of each moment. That is, locating key moments in a feature space that has eliminated a lot of interference and incorporated metacognitive information, making the localization results more accurate and reliable. In this way, a highly condensed single vector that is extremely sensitive to fault information can be generated for the final classification decision. As a result, the model can focus on the real fault key points with higher confidence, thereby significantly improving the accuracy and timeliness of prediction.

[0038] Specifically, in step S5, fault classification is performed on the multimodal spatiotemporal salient feature vector of the pitch system to obtain the predicted fault probability. It should be understood that the multimodal spatiotemporal salient feature vector of the pitch system generated in the previous steps is an abstract mathematical representation located in a high-dimensional feature space, which cannot be directly understood or used for decision-making. Therefore, in the technical solution of this application, fault classification is performed on the multimodal spatiotemporal salient feature vector of the pitch system to map the complex internal feature representation to predefined fault categories with clear physical meaning. In this way, the deep understanding within the model is transformed into external, actionable predictive conclusions, such as the probability that the system is in a healthy state or the probability of a certain type of fault occurring, thereby completing a complete closed loop from raw data to final prediction.

[0039] In practice, the multimodal spatiotemporal salient feature vector of the pitch system is input into a classifier to obtain the predicted fault probability. Within a neural network framework, this classifier typically consists of one or more fully connected layers and a final activation function. Its execution process can be broken down into the following sub-steps: First, the fixed-length multimodal spatiotemporal salient feature vector of the pitch system output from the previous step is input into a fully connected layer. This layer calculates the score or logical value for each predefined fault category through a linear transformation, namely multiplying it by a learnable weight matrix and adding a bias vector. The resulting score or logical value is the unnormalized raw score. To convert it into an output conforming to the probability definition, it needs to be processed by an activation function. In multi-class classification tasks, the Softmax function is most commonly used. The Softmax function can transform an arbitrary real-valued vector into a probability distribution vector, where each element's value is between 0 and 1, and the sum of all elements is 1. Finally, we can obtain a probability vector P after processing by the Softmax function. Each element of P represents the probability that the model predicts the input sample belongs to the corresponding category, that is, the probability of predicting the fault.

[0040] In summary, the neural network-based fault prediction method for wind turbine pitch systems according to the embodiments of this application is explained. Based on deep temporal feature encoding of the original SCADA data, it employs an innovative spatiotemporal dual saliency mechanism to enable the model to adaptively focus on and amplify the weak features of multidimensional data most relevant to fault evolution in the temporal and spatial dimensions, while suppressing interference from irrelevant operating condition information. This significantly improves the accuracy and reliability of fault prediction for wind turbine pitch systems, effectively reducing the false alarm rate, thus providing key technical support for achieving accurate predictive maintenance.

[0041] Furthermore, a fault prediction system for wind turbine pitch system based on neural networks is also provided.

[0042] Figure 3 This is a block diagram of a neural network-based fault prediction system for wind turbine pitch systems according to an embodiment of this application. Figure 3 As shown, the neural network-based wind turbine pitch system fault prediction system 300 according to an embodiment of this application includes: a data acquisition module 310 for acquiring raw SCADA data; a data processing module 320 for performing multi-dimensional time-series data preprocessing and windowing on the raw SCADA data to obtain a time series of multi-modal data co-coding vectors of the pitch system; a basic feature encoding module 330 for performing basic feature encoding on the time series of multi-modal data co-coding vectors of the pitch system to obtain a time series of multi-modal basic feature encoding vectors of the pitch system; a spatiotemporal doubly saliency module 340 for performing spatiotemporal doubly saliency on the time series of multi-modal basic feature encoding vectors of the pitch system to obtain a spatiotemporal saliency vector of the pitch system; and a fault classification module 350 for performing fault classification on the spatiotemporal saliency vector of the pitch system to obtain a predicted fault probability.

[0043] As described above, the neural network-based wind turbine pitch system fault prediction system 300 according to the embodiments of this application can be implemented in various wireless terminals, such as servers with a neural network-based wind turbine pitch system fault prediction algorithm. In one possible implementation, the neural network-based wind turbine pitch system fault prediction system 300 according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the neural network-based wind turbine pitch system fault prediction system 300 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the neural network-based wind turbine pitch system fault prediction system 300 can also be one of many hardware modules of the wireless terminal.

[0044] Alternatively, in another example, the neural network-based wind turbine pitch system fault prediction system 300 and the wireless terminal can also be separate devices, and the neural network-based wind turbine pitch system fault prediction system 300 can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0045] The various embodiments of this disclosure 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 improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A fault prediction method for wind turbine pitch system based on neural networks, characterized in that, include: Obtain raw SCADA data; Multidimensional time-series data preprocessing and windowing are performed on the raw SCADA data to obtain the time series of the multimodal data co-coding vector of the pitch system; The time series of the multimodal data co-coding vector of the pitch system is subjected to basic feature encoding to obtain the time series of the basic feature encoding vector of the multimodal data of the pitch system; Spatiotemporal doubly saliency is performed on the time series of the multimodal basic feature encoding vector of the pitch system to obtain the spatiotemporal saliency feature vector of the multimodal pitch system; Fault classification is performed on the multimodal spatiotemporal salient feature vectors of the pitch system to obtain the predicted fault probability.

2. The method for fault prediction of wind turbine pitch system based on neural network according to claim 1, characterized in that, The basic feature encoding of the time series of the multimodal data co-encoding vector of the pitch system is performed to obtain the time series of the basic feature encoding vector of the pitch system, including: inputting the time series of the multimodal data co-encoding vector of the pitch system into a bidirectional LSTM network to obtain the time series of the basic feature encoding vector of the pitch system.

3. The method for fault prediction of wind turbine pitch system based on neural network according to claim 1, characterized in that, Spatiotemporal doubly saliency is applied to the time series of the multimodal fundamental feature encoding vector of the pitch system to obtain the spatiotemporal salient feature vector of the multimodal pitch system, including: Spatial attention weighting and feature recalibration are performed on the time series of the multimodal basic feature encoding vector of the pitch system to obtain the recalibrated feature sequence and the spatial attention weight sequence; Two-stream gated attention encoding is performed on the recalibrated feature sequence and the spatial attention weight sequence to obtain the fused hidden state sequence; Temporal attention encoding is performed on the fused hidden state sequence to obtain the multimodal spatiotemporal salient feature vector of the pitch system.

4. The method for fault prediction of wind turbine pitch system based on neural network according to claim 3, characterized in that, Two-stream gated attention encoding is performed on the recalibrated feature sequence and the spatial attention weight sequence to obtain the fused hidden state sequence, including: The recalibrated feature sequence is input into the bidirectional LSTM model to obtain the hidden state sequence; The spatial attention weight sequence is input into the bidirectional LSTM model to obtain the attention hidden state sequence; The hidden state sequence and the attention hidden state sequence are dynamically gated and fused to obtain the fused hidden state sequence.

5. The method for fault prediction of wind turbine pitch system based on neural network according to claim 4, characterized in that, The process of dynamically gating and fusing the hidden state sequence and the attention hidden state sequence to obtain the fused hidden state sequence includes: dynamically gating and fusing the hidden state sequence and the attention hidden state sequence using the following formula: in, For each hidden state in the hidden state sequence, For each attention hidden state in the attention hidden state sequence, yes The threshold value at any given time. Represents the Sigmoid activation function. and These are the learnable weights and bias parameters of the gating network. yes The hidden state after the merging of moments. This represents element-wise multiplication.

6. The method for fault prediction of wind turbine pitch system based on neural network according to claim 3, characterized in that, Temporal attention encoding is performed on the fused hidden state sequence to obtain the multimodal spatiotemporal salient feature vector of the pitch system, including: temporal attention encoding of the fused hidden state sequence using the following formula: in, yes Unnormalized attention score at each moment , , These are the learnable parameters of a temporal attention network. It is the spatiotemporal salient feature vector of the multimodal pitch system. It is based on Calculated Temporal attention weights at each moment.

7. The method for fault prediction of wind turbine pitch system based on neural network according to claim 1, characterized in that, To classify the multimodal spatiotemporal salient feature vector of the pitch system to obtain the predicted fault probability, the method includes: inputting the multimodal spatiotemporal salient feature vector of the pitch system into a classifier to obtain the predicted fault probability.

8. A fault prediction system for wind turbine pitch system based on neural networks, characterized in that, include: The data acquisition module is used to acquire raw SCADA data; The data processing module is used to perform multi-dimensional time-series data preprocessing and windowing on the raw SCADA data to obtain the time series of the multi-modal data co-coding vector of the pitch system; The basic feature encoding module is used to perform basic feature encoding on the time series of the multimodal data co-encoding vector of the pitch system to obtain the time series of the basic feature encoding vector of the multimodal data of the pitch system. The spatiotemporal doubly saliency module is used to perform spatiotemporal doubly saliency on the time series of the multimodal basic feature encoding vector of the pitch system to obtain the spatiotemporal saliency feature vector of the multimodal pitch system; The fault classification module is used to classify faults in the multimodal spatiotemporal salient feature vectors of the pitch system to obtain the predicted fault probability.

Images