Wind turbine generator variable pitch system fault diagnosis method and system based on data driving

Abstract

The embodiment of the invention relates to the technical field of fault diagnosis of a variable pitch system of a wind turbine generator, and provides a fault diagnosis method and system for the variable pitch system of the wind turbine generator based on data driving. Strong correlation influences of external operation working conditions such as wind speed and power on diagnosis parameters such as motor current are accurately learned and stripped, so that a residual time sequence which can more purely reflect the internal health state of a wind turbine generator pitch system is extracted, then a time sequence dynamic mode prediction model is constructed for the residual time sequence, and a prediction result is obtained. According to the method, a dynamic evolution rule of a residual error in a normal state is deeply learned, when a wind turbine generator pitch system has an early failure, an actual dynamic behavior deviates from a learned normal mode, so that a predicted residual error is obviously increased, a sensitive abnormal score is formed, and finally, a health index is generated by smoothing the abnormal score. And stable and accurate early warning of early faults of the variable pitch system is realized.

Description

Data-driven fault diagnosis method and system for wind turbine pitch systems Technical Field

[0001] This invention relates to the field of fault diagnosis technology for wind turbine pitch systems, and in particular to a data-driven fault diagnosis method and system for wind turbine pitch systems. Background Technology

[0002] Wind power, as a clean and renewable energy source, occupies an increasingly important position in the global energy structure. The stable and reliable operation of wind turbines is crucial to ensuring the economic benefits of wind farms and the security of the power grid. Among the many subsystems of a wind turbine, the pitch system plays a vital role. It optimizes energy capture efficiency and controls the load on the wind turbine by adjusting the blade pitch angle in real time, and is the core actuator ensuring the efficient and safe operation of the wind turbine under different wind conditions. However, the pitch system is exposed to harsh and variable natural environments for extended periods, enduring complex alternating loads, making it one of the components most prone to failure. A failure in the pitch system can lead to anything from minor power generation losses and increased maintenance costs to potentially catastrophic accidents such as overspeeding.

[0003] Existing wind turbine fault diagnosis methods largely rely on operational data acquired from Supervisory Control and Data Acquisition (SCADA) systems. However, these methods generally face significant challenges in detecting early, subtle faults in pitch systems. Specifically, the physical characteristics of early-stage faults are extremely weak. For example, when a pitch bearing experiences slight wear due to insufficient early lubrication, the data may only show a small increase in drive motor current or torque. These weak fault characteristics are completely drowned out by the large fluctuations in normal data caused by drastic changes in operating conditions such as wind speed and airflow turbulence. Traditional alarm methods based on fixed thresholds require extremely sensitive thresholds to capture such subtle changes, inevitably leading to numerous false alarms during normal fluctuations in operating conditions. Furthermore, directly using conventional data-driven models to predict and detect anomalies in raw sensor signals proves ineffective in distinguishing between normal fluctuations caused by changes in operating conditions and abnormal fluctuations caused by system health degradation, thus failing to effectively identify early-stage faults hidden in a background of strong noise. Summary of the Invention

[0004] The present invention aims to solve at least one of the problems existing in the prior art, and provides a data-driven method and system for fault diagnosis of wind turbine pitch system.

[0005] One aspect of the present invention provides a data-driven fault diagnosis method for wind turbine pitch systems. The method includes: inputting a time-series stream of acquired SCADA data into a trained operating condition baseline model to obtain current residual time-series data; inputting the current residual time-series data into a trained time-series dynamic pattern prediction model to obtain predicted residual values; comparing the predicted residual values ​​with the actual residual values ​​at corresponding times to obtain anomaly scores; smoothing the time-series stream of anomaly scores in the time dimension to obtain health indicators; comparing the health indicators with a preset warning threshold, and generating an early fault warning signal when the health indicators exceed the preset warning threshold.

[0006] Optionally, SCADA data includes operating data and target diagnostic parameters of the wind turbine pitch system. The operating data includes wind speed, generator power, and pitch angle, while the target diagnostic parameters include pitch motor current and pitch motor torque.

[0007] Optionally, the time-series stream of the acquired SCADA data is input into the trained operating condition baseline model to obtain current residual time-series data, including: training the operating condition baseline model based on the operating condition data to obtain the trained operating condition baseline model; inputting real-time operating condition data into the trained operating condition baseline model to obtain the predicted pitch motor current value; and subtracting the predicted pitch motor current value from the synchronously acquired actual pitch motor current value to obtain the current residual.

[0008] Optionally, the baseline model for operating conditions is a gradient boosting regression tree model.

[0009] Optionally, inputting the current residual time series data into the trained time-series dynamic pattern prediction model to obtain predicted residual values ​​includes: performing Z-score standardization on the current residual time series data to obtain standardized current residual time series data; dividing the standardized current residual time series data into multiple standardized current residual local time series data based on a sliding time window; encoding and sorting the multiple standardized current residual local time series data based on the time dimension to obtain a sequence of current residual encoding vectors; and inputting the sequence of current residual encoding vectors into the trained time-series dynamic pattern prediction model for information-passing prediction to obtain predicted residual values.

[0010] Optionally, the sequence of current residual encoded vectors is input into the trained time-series dynamic pattern prediction model for information transfer prediction to obtain the predicted residual value, including: using the trained time-series dynamic pattern prediction model to perform time-series dynamic pattern prediction on the sequence of current residual encoded vectors to obtain the current residual time-series prediction vector; and decoding the current residual time-series prediction vector to obtain the predicted residual value.

[0011] Optionally, the trained time-series dynamic pattern prediction model is used to perform time-series dynamic pattern prediction on the sequence of current residual encoded vectors to obtain a time-series predicted vector of current residuals. This includes: inputting the sequence of the current residual encoded vectors into an initial long short-term memory network to obtain a sequence of implicit semantic representations of current residuals; and calculating a semantic deviation quantification index between the sequence of implicit semantic representations of current residuals and the corresponding vectors at each time step in the sequence of current residual encoded vectors according to the following formula to obtain a set of current residual semantic deviation quantification indices: ; ; ; ;in, The first in the sequence of current residual encoding vectors A current residual encoding vector The characteristic projection function of the current residual is... For current residual characteristics, a multilayer sensor is used. for The corresponding current residual encoding implicit semantic representation. The projection function for encoding semantic hidden features of current residuals. A multilayer perceptron is used to encode semantic hidden features for current residuals. Let L2 norm be the vector. For smoothing parameters, A set of quantitative indicators for semantic deviation of current residuals. for and The semantic deviation quantification index of the current residual between the two They are respectively The value is The semantic deviation quantification index of current residual is used; based on the set of semantic deviation quantification indexes of current residual, a set of semantic fidelity correction factors for current residual is generated according to the following formula: ; ;in, This represents the mean of all semantic deviation quantification indices for current residuals within the set of semantic deviation quantification indices. The standard deviation of each semantic deviation quantification index of current residuals in the set of semantic deviation quantification indices is given. and All are adjustable parameters. As a preset constant, Let e ​​be the value of the logarithmic function with the natural constant e as the base. for Activation function for The corresponding current residual semantic fidelity correction factor, This is a set of semantic fidelity correction factors for current residuals. They are respectively for The value is The semantic fidelity correction factor for the current residual is determined. Based on the set of current residual semantic fidelity correction factors, a semantic bias constraint is applied to the sequence of implicit semantic representations encoded by the current residual, to obtain a sequence of semantically enhanced implicit representations of the current residual: ; ;in, for inverse transform, for The corresponding semantically enhanced implicit representation of the current residual. This is a sequence of implicit representations with semantic enhancement for current residuals. They are respectively The value is The current residual semantic enhancement implicit representation is used; the sequence of the current residual semantic enhancement implicit representation is input into the terminal long short-term memory network to generate the current residual time series prediction vector.

[0012] Optionally, comparing the predicted residual value with the actual residual value at the corresponding time to obtain an anomaly score includes: calculating the squared error between the predicted residual value and the actual residual value at the corresponding time to obtain the anomaly score.

[0013] Optionally, the time series of outlier scores in the time dimension can be smoothed to obtain health indicators, including: calculating an exponentially weighted moving average of the time series of outlier scores to obtain health indicators.

[0014] Another aspect of the present invention provides a data-driven fault diagnosis system for wind turbine pitch systems. The data-driven wind turbine pitch system fault diagnosis system includes: a time-series data acquisition module, used to input the time-series stream of acquired SCADA data into a trained operating condition baseline model to obtain current residual time-series data; a residual value prediction module, used to input the current residual time-series data into a trained time-series dynamic pattern prediction model to obtain predicted residual values; a residual value comparison module, used to compare the predicted residual values ​​with the actual residual values ​​at corresponding times to obtain anomaly scores; a time-series stream smoothing module, used to smooth the time-series stream of anomaly scores in the time dimension to obtain health indicators; and a fault early warning signal generation module, used to compare the health indicators with a preset early warning threshold, and generate an early fault early warning signal when the health indicators exceed the preset early warning threshold.

[0015] Compared with existing technologies, this invention first constructs a baseline model of operating conditions to accurately learn and isolate the strong correlation between external operating conditions such as wind speed and power and diagnostic parameters such as motor current. This allows for the extraction of a residual time series that more purely reflects the inherent health status of the wind turbine pitch system. Subsequently, a time-series dynamic pattern prediction model is constructed based on this residual time series, enabling it to deeply learn the dynamic evolution law of the residual under normal conditions. When an early fault occurs in the wind turbine pitch system, its actual dynamic behavior deviates from the learned normal pattern, resulting in a significant increase in the predicted residual and the formation of a sensitive anomaly score. Finally, by smoothing the anomaly score, a health index is generated, achieving stable and accurate early warning of early faults in the pitch system. Attached Figure Description

[0016] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0017] Figure 1 is a flowchart of a data-driven wind turbine pitch system fault diagnosis method according to an embodiment of the present invention; Figure 2 is a data flow diagram of a data-driven wind turbine pitch system fault diagnosis method according to another embodiment of the present invention; Figure 3 is a flowchart of a data-driven wind turbine pitch system fault diagnosis method according to another embodiment of the present invention, in which current residual time series data is input into a trained time-series dynamic pattern prediction model to obtain predicted residual values; Figure 4 is a flowchart of a data-driven wind turbine pitch system fault diagnosis method according to another embodiment of the present invention, in which the trained time-series dynamic pattern prediction model is used to perform time-series dynamic pattern prediction on the sequence of current residual encoding vectors to obtain current residual time-series prediction vectors; Figure 5 is a block diagram of a data-driven wind turbine pitch system fault diagnosis system according to another embodiment of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] Unless otherwise specifically stated, the technical or scientific terms used in the embodiments of this invention should be understood in their ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms "comprising" or "including," as used in the embodiments of this invention, do not limit the shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof mentioned, nor do they exclude the appearance or addition of one or more other different shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof, or the inclusion of these.

[0020] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale, and techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail; however, where appropriate, the illustrated techniques, methods, and apparatus should be considered part of the specification. In all the examples shown and discussed herein, any other specific example may have different values. It should be noted that similar symbols and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0021] In the description of the embodiments of the present invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In the embodiments of the present invention, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of the present invention, as well as the features of different embodiments or examples.

[0022] Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. It is obvious that the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention, and it should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0023] Existing data-driven fault diagnosis technologies for wind turbine pitch systems generally face the problem of insufficient sensitivity in detecting early, weak fault signals: the weak characteristics at the initial stage of a fault are completely submerged in the strong data fluctuations caused by drastic changes in operating conditions such as wind speed and turbulence, making them difficult to identify effectively. Therefore, this invention proposes a data-driven fault diagnosis method for wind turbine pitch systems. Specifically, the process first inputs the time-series SCADA data into a trained baseline model. This baseline model accurately depicts the strong correlation between operating conditions such as wind speed and power and target diagnostic parameters like pitch and motor current under normal operating conditions. By subtracting the actual current value from the theoretically normal value predicted by the model, a current residual time-series data stripped of the main operating conditions is obtained. Subsequently, this current residual time-series data is input into a trained time-series dynamic pattern prediction model. When an early fault occurs in the system, the actual residual dynamics deviate from the learned normal pattern, resulting in a significant deviation between the predicted residual value and the actual residual value at the corresponding time. This deviation is quantified into a highly sensitive anomaly score by calculating the squared error. Finally, to eliminate misjudgments caused by instantaneous disturbances, the time-series stream of anomaly scores in the time dimension is smoothed by an exponentially weighted moving average, generating a health index that stably reflects the long-term degradation trend of the wind turbine pitch system's health status. Based on the comparison between this health index and a preset warning threshold, a reliable early fault warning signal is generated.

[0024] Figure 1 is a flowchart of a data-driven fault diagnosis method for a wind turbine pitch system according to an embodiment of the present invention. Figure 2 is a data flow diagram of the data-driven fault diagnosis method for a wind turbine pitch system according to an embodiment of the present invention. Referring to Figures 1 and 2 together, the data-driven fault diagnosis method for a wind turbine pitch system according to an embodiment of the present invention includes the following steps: S100, inputting the time-series stream of acquired SCADA data into a trained operating condition baseline model to obtain current residual time-series data; S200, inputting the current residual time-series data into a trained time-series dynamic pattern prediction model to obtain predicted residual values; S300, comparing the predicted residual values ​​with the actual residual values ​​at the corresponding time points to obtain anomaly scores; S400, smoothing the time-series stream of anomaly scores in the time dimension to obtain health indicators; S500, comparing the health indicators with a preset warning threshold, and generating an early fault warning signal when the health indicators are greater than the preset warning threshold.

[0025] Specifically, in step S100, the time-series stream of acquired SCADA data is input into the trained operating condition baseline model to obtain current residual time-series data. It should be understood that because the target diagnostic parameters of the wind turbine pitch system, such as the pitch motor current value, exhibit a strong nonlinear coupling relationship with external operating conditions such as wind speed and generator power, weak abnormal signals caused by early equipment health degradation are completely submerged in the background noise of severe operating condition fluctuations, making them difficult to detect directly. Therefore, in the technical solution of this invention, the time-series stream of acquired SCADA data is input into the trained operating condition baseline model to obtain current residual time-series data. This establishes a dynamic benchmark that accurately reflects the strong correlation between operating condition changes and target diagnostic parameters under normal operating conditions, thereby effectively stripping away the influence of operating conditions. This generates a residual data stream that only reflects the internal state changes of the system. This residual data stream significantly amplifies the signal-to-noise ratio of early faults, providing an ideal analytical object for subsequent high-sensitivity fault diagnosis.

[0026] Specifically, SCADA data includes operating condition data and target diagnostic parameters for the wind turbine pitch system. The operating condition data includes wind speed, generator power, and pitch angle, while the target diagnostic parameters include pitch motor current and pitch motor torque. The operating condition baseline model is a gradient boosting regression tree model.

[0027] More specifically, in a specific example of the present invention, the time-series stream of acquired SCADA data is input into the trained operating condition baseline model to obtain current residual time-series data, including: training the operating condition baseline model based on the operating condition data to obtain the trained operating condition baseline model; inputting real-time operating condition data into the trained operating condition baseline model to obtain the predicted pitch motor current value; and subtracting the predicted pitch motor current value from the synchronously acquired actual pitch motor current value to obtain the current residual.

[0028] More specifically, the process of acquiring current residual time series data first involves the offline training and construction of the operating condition baseline model. From massive amounts of historical SCADA data, a dataset representing the long-term healthy operating state of the wind turbine pitch system is selected. In this dataset, operating condition data such as wind speed, generator power, and pitch angle are used as input features to the operating condition baseline model, and the corresponding pitch and motor current values ​​at the same time are used as the learning target. Based on this selected data, a gradient boosting regression tree model is trained as the operating condition baseline model, allowing it to fully learn and fit the complex nonlinear mapping relationship between operating condition features and pitch and motor current. After training, a fixed operating condition baseline model is constructed. In the subsequent online monitoring phase, the real-time operating condition data stream of the wind turbine pitch system, i.e., the real-time operating condition data, is continuously input into the trained operating condition baseline model. The operating condition baseline model synchronously outputs a predicted pitch and motor current value sequence based on the current real-time operating conditions. Finally, by subtracting the predicted pitch motor current value sequence output by the baseline model from the actual pitch motor current value sequence synchronously collected from the SCADA system at each time point, a stable current residual time series data that eliminates the influence of the main operating conditions can be obtained. Specifically, for each time point, the current residual at the corresponding time point can be obtained by subtracting the predicted pitch motor current value output by the baseline model from the synchronously collected actual pitch motor current value.

[0029] Specifically, in step S200, the current residual time series data is input into the trained time-series dynamic pattern prediction model to obtain the predicted residual value. It should be understood that the current residual time series data obtained after operating condition stripping, during healthy operation of the wind turbine, is not pure white noise, but contains inherent, time-dependent dynamic behavioral characteristics of the system. If only a simple threshold judgment is made on the instantaneous amplitude of the residual, it is difficult to effectively distinguish between harmless random disturbances and systemic pattern shifts representing early fault buds, leading to missed or false alarms. Therefore, in the technical solution of this invention, the current residual time series data is further input into the trained time-series dynamic pattern prediction model to obtain the predicted residual value, thereby enabling deep learning and accurate capture of the dynamic evolution law and time-dependent pattern of the residual sequence under healthy conditions. In this way, an accurate prediction of the residual value at the next moment based on historical data patterns can be generated as a predicted value. This predicted value can serve as a dynamic, high-fidelity benchmark of normal behavior, providing a reference for subsequent identification of weak abnormal pattern deviations.

[0030] Figure 3 is a flowchart illustrating the process of inputting current residual time series data into a trained time-series dynamic pattern prediction model to obtain predicted residual values ​​in a data-driven wind turbine pitch system fault diagnosis method according to an embodiment of the present invention. As shown in Figure 3, step S200 includes: S210, performing Z-score normalization on the current residual time series data to obtain normalized current residual time series data; S220, dividing the normalized current residual time series data into multiple normalized current residual local time series data based on a sliding time window; S230, encoding and sorting the multiple normalized current residual local time series data based on the time dimension to obtain a sequence of current residual encoding vectors; S240, inputting the sequence of current residual encoding vectors into the trained time-series dynamic pattern prediction model for information-transfer prediction to obtain predicted residual values.

[0031] In step S210, the current residual time series data is Z-score standardized to obtain standardized current residual time series data. It should be understood that, since the dimensions and numerical range of the current residual time series data generated in the aforementioned steps are related to the electrical characteristics of the specific wind turbine and the accuracy of the sensors, directly using it as input to the subsequent deep learning model may lead to slow gradient descent, difficulty in convergence, or even unstable model performance during training due to differences in data scale. Therefore, in the technical solution of this invention, the current residual time series data is further Z-score standardized to obtain standardized current residual time series data, thereby eliminating the influence of data dimensions and mapping the original residual data to a standard normal distribution with a mean of 0 and a standard deviation of 1. This provides a uniformly scaled and well-distributed input for the subsequent time-series dynamic pattern prediction model, thereby improving the model's training efficiency and prediction accuracy.

[0032] More specifically, in a specific example of the present invention, the current residual time series data under a long-term healthy operating state used for training the time-series dynamic pattern prediction model is selected, and its overall mean is calculated. with standard deviation These two statistics, the mean and standard deviation, will be saved as fixed transformation parameters. During the model training phase, each data point in the current residual time series data will be... By transforming the formula The transformation yields a complete, standardized historical dataset of current residuals that follows a standard normal distribution, which is used for training subsequent models. That is, data points The corresponding standardized current residual. After entering the online real-time monitoring phase, for each newly generated real-time current residual value, the saved average is retrieved. and standard deviation The same transformation formula is used to process the data to ensure that the online input data and the offline training data are consistent in scale, thereby ensuring the effectiveness and reliability of the prediction model.

[0033] In step S220, based on a sliding time window, the standardized current residual time series data is divided into multiple standardized current residual local time series data. It should be understood that subsequent time-series dynamic pattern prediction models, especially those based on recurrent neural network structures, cannot directly handle continuous and infinitely long time-series data streams. Their algorithmic structure requires fixed-length sequence samples as input to learn and infer temporal dependencies. Therefore, in the technical solution of this invention, the standardized current residual time series data is further divided into multiple standardized current residual local time series data based on a sliding time window. This reconstructs the continuous, unstructured time series data into a set of sample pairs consisting of input sequences and target values, suitable for supervised learning. This generates a formatted dataset that retains the temporal dynamic information of the original data while meeting the input requirements of the prediction model, laying a data foundation for subsequent model training and online prediction.

[0034] More specifically, in a concrete example of this invention, the normalized current residual time series data is segmented. First, the parameters of the time window need to be set, determining a fixed sliding window length, such as 60 time points. This means the model will make predictions based on data from the past 60 seconds. During the offline model training phase, this sliding window traverses the complete historical normalized current residual data. Starting from the first point of the normalized current residual time series, a 60-point segment is extracted as the first input sample, and the 61st data point is used as the target label for this sample. Subsequently, the window slides forward one time step, extracting a 60-point segment starting from the second point of the normalized current residual time series as the second input sample, and using the 62nd data point as its target label. This process is repeated until the window slides to the end of the normalized current residual time series, thereby transforming a long sequence into a large number of overlapping local time series samples that can be used for model training. In online monitoring applications, a first-in-first-out queue of length 60 is maintained in real time. Whenever a new standardized residual data point is generated, the data point is enqueued, and the oldest data point is dequeued. The 60 data points in the queue always constitute the unique input sequence for prediction at the current moment.

[0035] In step S230, multiple standardized current residual local time series data are encoded and sorted based on the time dimension to obtain a sequence of current residual encoded vectors. It should be understood that although the standardized current residual local time series data generated in the previous step is a fixed-length sequence, it is still a one-dimensional arrangement of the original values. Directly inputting it into a complex time series model may not fully highlight the implicit local patterns and deep dynamic features. To improve the subsequent model's ability to perceive changes in system state, these original sequence data need to be transformed into an expression with higher information density and more significant features. Therefore, in the technical solution of this invention, multiple standardized current residual local time series data are further encoded and sorted based on the time dimension to obtain a sequence of current residual encoded vectors. This allows for feature extraction and abstract representation of each local time series segment, compressing its inherent dynamic characteristics into a fixed-dimensional feature vector. In this way, the original numerical sequence can be transformed into a higher-level feature vector sequence. Each vector in this feature vector sequence can be regarded as a compact fingerprint of the system's dynamic behavior within the corresponding time window, providing a higher-quality input for subsequent models to accurately capture state evolution and concept drift.

[0036] More specifically, in a concrete example of this invention, the encoding and sorting process first requires constructing an encoder network for feature extraction. A one-dimensional convolutional neural network (1D-CNN) is selected as the encoder, its structure designed to receive a local time series of length 60 and extract local patterns, trends, and shapes from the data layer by layer through a series of convolutional and pooling layers. Finally, a fixed-dimensional feature vector, for example, 16-dimensional, is output through a fully connected layer. During encoding, a large amount of standardized current residual local time series data generated in the previous step is input into the 1D-CNN encoder one by one according to its original time order. Each input sequence of length 60 is mapped and converted into a 16-dimensional current residual encoding vector after forward propagation computation by the encoder. Finally, these generated current residual encoding vectors are arranged strictly according to the time order of their corresponding input sequences to form a completely new sequence, namely the sequence of current residual encoding vectors. Each element in this sequence is no longer a single scalar value, but a 16-dimensional vector containing rich dynamic information, constituting the final sequence of current residual encoding vectors.

[0037] In step S240, the sequence of current residual encoded vectors is input into the trained time-series dynamic pattern prediction model for information-passive prediction to obtain the prediction residual value. It should be understood that although the sequence of current residual encoded vectors obtained in the previous steps is a higher-order representation after feature extraction, directly inputting it into a traditional single-sequence model for prediction may treat the feature transformation as a black box during the learning process, lacking effective monitoring of the semantic consistency of the intermediate representation layer. This phenomenon, known as concept drift, may cause the dynamic pattern learned by the model to deviate significantly from its original physical meaning, thereby reducing the robustness and generalization ability of the prediction, especially when facing the complex and variable micro-state changes of wind turbines. Therefore, in the technical solution of this invention, the sequence of current residual encoded vectors is further input into the trained time-series dynamic pattern prediction model for information-passive prediction to obtain the prediction residual value, thereby introducing an explicit, closed-loop self-calibration mechanism to actively monitor and dynamically correct the semantic fidelity during the representation learning process. First, a first Long Short-Term Memory (LSTM) network is used to perform preliminary semantic encoding on the current residual encoding vector sequence. Then, the conceptual drift between the preliminary semantic representation and the original encoding vector is explicitly quantified, and an adaptive anti-drift correction factor is generated accordingly. Finally, this anti-drift correction factor is applied to the preliminary semantic representation to generate a semantically cleansed enhanced feature sequence, which is then used by a second LSTM network for final prediction. This ensures that the prediction process not only learns the deep temporal dependencies of the data, but also guarantees the dynamic fidelity of the representation through closed-loop management of internal feedback and control. Ultimately, it outputs a more reliable prediction residual value that has undergone rigorous self-checking and calibration, providing an extremely stable and accurate dynamic benchmark for subsequent anomaly detection.

[0038] More specifically, in a specific example of the present invention, inputting the sequence of current residual encoded vectors into a trained time-series dynamic pattern prediction model for information-transfer prediction to obtain predicted residual values ​​includes: using the trained time-series dynamic pattern prediction model to perform time-series dynamic pattern prediction on the sequence of current residual encoded vectors to obtain a current residual time-series prediction vector; and decoding the current residual time-series prediction vector to obtain predicted residual values.

[0039] Figure 4 is a flowchart illustrating how a data-driven fault diagnosis method for wind turbine pitch systems, based on an embodiment of the present invention, uses a trained time-series dynamic pattern prediction model to perform time-series dynamic pattern prediction on the sequence of current residual encoded vectors to obtain a time-series predicted vector for the current residuals. As shown in Figure 4, step S240 involves using the trained temporal dynamic pattern prediction model to perform temporal dynamic pattern prediction on the sequence of current residual encoded vectors to obtain a current residual temporal prediction vector. This includes the following steps: S241, inputting the sequence of the current residual encoded vectors into an initial long short-term memory network to obtain a sequence of implicit semantic representations of the current residual encoding; S242, calculating the semantic deviation quantization index between the sequence of implicit semantic representations of the current residual encoding and the corresponding vectors at each time step in the sequence of current residual encoded vectors to obtain a set of current residual semantic deviation quantization indices; S243, generating a set of current residual semantic fidelity correction factors based on the set of current residual semantic deviation quantization indices; S244, applying semantic deviation constraints to the sequence of implicit semantic representations of the current residual encoding based on the set of current residual semantic fidelity correction factors to obtain a sequence of semantically enhanced implicit representations of the current residuals; and S245, inputting the sequence of semantically enhanced implicit representations of the current residuals into a terminal long short-term memory network to generate a current residual temporal prediction vector.

[0040] In step S241, the sequence of the current residual encoding vector is input into the initial long short-term memory network to obtain the sequence of implicit semantic representations of the current residual encoding. Specifically, firstly, the sequence of the current residual encoding vector is input into the initial long short-term memory network to obtain the sequence of implicit semantic representations of the current residual encoding. Current residual encoding vector The sequence that makes up the current residual encoding vector Then the sequence of current residual encoding vectors Input the initial Long Short-Term Memory network, i.e., the first LSTM neural network You can get Implicit semantic representation of the corresponding current residual codes The sequence of implicit semantic representations of current residual encoding ,in, The total number of current residual code vectors in the sequence of current residual code vectors.

[0041] It should be understood that, since the current residual encoding vector sequence obtained from the preceding steps is essentially a series of high-level but independent feature snapshots, in order to accurately predict the future state of the wind turbine pitch system, it is necessary to deeply model the dynamic evolution patterns and time dependencies implicit in these snapshots. The operating data of wind turbines themselves have extremely strong temporal characteristics; the degradation of their health state often manifests as a gradual accumulation of weak pattern changes over a long period, rather than isolated events. Therefore, in the technical solution of this invention, the sequence of the current residual encoding vector is further input into an initial long short-term memory network to obtain a sequence of implicit semantic representations of the current residual encoding, thereby performing a preliminary semantic encoding process for the current residual encoding vector. This aims to utilize the unique gating mechanism of the long short-term memory network, enabling it to adaptively remember or forget historical information in the sequence, thereby effectively learning and capturing long-range dependencies in the data, mapping the current residual encoding vector to a higher-dimensional, more abstract semantic space. In this way, a sequence of implicit semantic representations of the current residual encoding can be generated, where each representation vector contains its dynamic evolution information within the entire sequence context. This is not just a simple transfer of the original features, but a re-creation that deeply integrates temporal logic, providing a high-quality input that has been initially refined, is richer in information, and is more concentrated for subsequent high-precision semantic deviation quantification and correction.

[0042] In step S242, a semantic deviation quantization index is calculated between the sequence of implicit semantic representations of the current residual code and the sequence of current residual coding vectors at each time step to obtain a set of current residual semantic deviation quantization indexes. Specifically, the semantic deviation quantization index between the sequence of implicit semantic representations of the current residual code and the sequence of current residual coding vectors at each time step is calculated according to the following formula to obtain a set of current residual semantic deviation quantization indexes: ; ; ; ;in, The first in the sequence of current residual encoding vectors A current residual encoding vector The characteristic projection function of the current residual is... For current residual characteristics, a multilayer sensor is used. for The corresponding current residual encoding implicit semantic representation. The projection function for encoding semantic hidden features of current residuals. A multilayer perceptron is used to encode semantic hidden features for current residuals. Let L2 norm be the vector. For smoothing parameters, A set of quantitative indicators for semantic deviation of current residuals. for and The semantic deviation quantification index of the current residual between the two They are respectively The value is The semantic deviation quantification index of current residual at that time.

[0043] It is understandable that, due to the deep nonlinear transformation performed by the initial Long Short-Term Memory network in extracting temporal dependencies, this process inherently carries a risk: the generated implicit semantic representation may deviate from the intrinsic meaning of its original input, the current residual encoding vector, at the semantic level. This phenomenon is known as semantic bias or conceptual drift. If this feature transformation process is treated as an uncontrollable black box without effective monitoring of the semantic consistency of the intermediate representation layer, the model's understanding of the normal dynamics of the wind turbine may be distorted, ultimately impairing its ability to generalize and identify early, minor faults. Therefore, in the technical solution of this invention, a semantic bias quantification index is further calculated between the sequence of implicit semantic representations of the current residual encoding and the corresponding vectors at each time step in the sequence of current residual encoding vectors. This is used to construct an active monitoring link in an explicit, closed-loop self-calibration mechanism. The core purpose of this is to transform the abstract risk of semantic bias into a concrete numerical index that can be accurately measured at each time step. Essentially, it measures the semantic difference between the current residual after temporal context encoding and the original domain before transformation. In this way, a precise quantitative index can be generated for the feature transformation process at each time point. The corresponding set of quantitative indices can clearly depict where and to what extent semantic information is lost or distorted during the representation learning process of the entire sequence, thus providing objective and detailed data basis for the subsequent adaptive correction stage.

[0044] In step S243, a set of current residual semantic fidelity correction factors is generated based on the current residual semantic deviation quantization index set. Specifically, the set of current residual semantic fidelity correction factors is generated based on the current residual semantic deviation quantization index set according to the following formula: ; ;in, This represents the mean of all semantic deviation quantification indices for current residuals within the set of semantic deviation quantification indices. The standard deviation of each semantic deviation quantification index of current residuals in the set of semantic deviation quantification indices is given. and All are adjustable parameters. As a preset constant, Let e ​​be the value of the logarithmic function with the natural constant e as the base. for Activation function for The corresponding current residual semantic fidelity correction factor, This is a set of semantic fidelity correction factors for current residuals. They are respectively for The value is The semantic fidelity correction factor for the current residual at that time.

[0045] It should be understood that the semantic deviation quantification index set obtained in the aforementioned steps is only an objective measure of the degree of distortion of the representation layer information and cannot be directly used as a correction signal to adjust the feature vector. If the original deviation value is used directly for intervention, the correction process may be unstable due to scale or distribution issues. In order to achieve stable and effective closed-loop control, the diagnostic quantification index must be transformed into a strategic and directly operable adjustment parameter. Therefore, in the technical solution of this invention, a current residual semantic fidelity correction factor set is further generated based on the current residual semantic deviation quantification index set to execute an adaptive adjustment strategy. The core of this strategy is to transform the quantified current residual semantic deviation degree into a specific and operable correction signal. The generation process of this correction signal is not a simple linear mapping, but rather a dynamic generation of a matching correction weight by comprehensively considering the deviation at each time point and its statistical position in the global sequence through a nonlinear function. This provides a precise formula for subsequent calibration, ensuring that the calibration intensity applied to each implicit semantic representation is precisely matched with its own semantic drift, thereby effectively avoiding the introduction of new noise due to over-calibration or under-calibration, and guaranteeing the stability and accuracy of the entire self-calibration mechanism.

[0046] In step S244, based on the current residual semantic fidelity correction factor set, a semantic bias constraint is applied to the sequence of implicit semantic representations encoded by the current residuals to obtain a sequence of implicit semantic representations enhanced by the current residuals. Specifically, according to the following formula, a semantic bias constraint is applied to the sequence of implicit semantic representations encoded by the current residuals based on the current residual semantic fidelity correction factor set to obtain a sequence of implicit semantic representations enhanced by the current residuals: ; ;in, for The inverse transformation ensures semantic consistency. for The corresponding semantically enhanced implicit representation of the current residual. This is a sequence of implicit representations with semantic enhancement for current residuals. They are respectively The value is The semantic enhancement of the implicit representation of the current residual at that time.

[0047] It should be understood that, since the aforementioned steps have generated a set of current residual semantic fidelity correction factors to guide the correction, in order to complete the entire active intervention loop of quantization drift-strategy generation-correction execution, this strategy must be applied to implicit semantic representations with semantic biases to achieve closed-loop management of representation quality. Without executing this crucial constraint step, the preceding bias quantization and strategy generation will lose their fundamental meaning, and the model will still make subsequent inferences based on potentially distorted internal representations. Therefore, in the technical solution of this invention, semantic bias constraints are further applied to the sequence of current residual encoded implicit semantic representations based on the aforementioned set of current residual semantic fidelity correction factors to obtain a sequence of current residual semantically enhanced implicit representations, thereby performing real-time semantic purification of the current residual semantic feature stream. This constraint process, through feature modulation technology, uses correction factors to impose stronger constraints on implicit semantic representations that have experienced severe semantic drift, pulling them back to a representation space closer to the original physical meaning, thereby dynamically correcting the representation learning process. In this way, a semantically enhanced implicit representation sequence can be generated. The significant advantage of this sequence is that it not only fully preserves the deep temporal dependencies on the dynamics of the pitch system captured by the initial long short-term memory network, but also effectively filters out the semantic noise and bias generated during the feature transformation process through the anti-drift mechanism, ultimately achieving a delicate balance between powerful representation capabilities and high semantic fidelity.

[0048] In step S245, the sequence of the semantically enhanced implicit representation of the current residual is input into the terminal long short-term memory network to generate a current residual time-series prediction vector. Specifically, the sequence of the semantically enhanced implicit representation of the current residual is... The input terminal long short-term memory network, also known as the second LSTM neural network, is a second LSTM neural network. The current residual time series prediction vector can then be obtained. .

[0049] It should be understood that since the aforementioned anti-drift constraint step has already produced a purified and enhanced implicit representation sequence, which is more stable at the semantic level and has higher information quality, a final integration step is needed to refine and condense the dynamic information distributed throughout the high-quality sequence to form a unified prediction of the future state in order to complete the final prediction task. Therefore, in the technical solution of this invention, the semantically enhanced implicit representation sequence is further input into the terminal long short-term memory network to generate the final current residual timing prediction vector, thereby performing the final information aggregation and encoding on the basis of a high-quality and semantically stable representation. The function of the terminal long short-term memory network is to further refine and integrate the dynamic information about the normal operation of the pitch system in the entire enhanced sequence and combine it into a single, highly condensed prediction output. In this way, a final current residual timing prediction vector can be generated. This vector is not only a global representation of the entire input sequence, but also a robust and reliable representation that has undergone rigorous self-examination and internal calibration, providing the most reliable dynamic benchmark for subsequent calculation of anomaly scores.

[0050] Accordingly, the current residual time-series prediction vector is decoded to obtain the predicted residual value. It should be understood that since the current residual time-series prediction vector output in the aforementioned steps is the final hidden state of the second LSTM neural network, it is essentially a high-dimensional, condensed representation existing in the abstract feature space within the model. The dimension and numerical domain of this vector are inconsistent with the actual, single-scalar residual value, making direct comparison impossible. Therefore, in the technical solution of this invention, the current residual time-series prediction vector is further decoded to obtain the predicted residual value, thereby mapping this high-dimensional abstract feature representation back to the one-dimensional physical space of the original residual data, completing the conversion from the model's internal representation to the specific predicted value. In this way, a single-scalar predicted residual value with the same physical meaning and data dimension as the actual residual value can be generated, providing a predictive benchmark that can be directly mathematically calculated for subsequent calculations of prediction errors, i.e., anomaly scores.

[0051] More specifically, in a concrete example of the present invention, the process of decoding the current residual timing prediction vector is implemented through a decoder structure. This decoder consists of a fully connected layer whose input dimension is set to be the same as the dimension of the current residual timing prediction vector output by the second LSTM neural network. The output dimension of the fully connected layer is set to 1, meaning it contains only one neuron, and this neuron uses a linear activation function. During decoding, each high-dimensional prediction vector output by the second LSTM neural network, i.e., the current residual timing prediction vector, is input to this fully connected layer. After weighted summation and linear transformation, the fully connected layer finally outputs a single scalar value, which is the predicted residual value obtained after decoding and exists in the same numerical space as the standardized actual residual value.

[0052] Specifically, in step S300, the predicted residual value is compared with the actual residual value at the corresponding time to obtain an anomaly score. It should be understood that the predicted residual value output by the aforementioned time-series dynamic pattern prediction model is a precise extrapolation of the normal dynamic behavior pattern learned from historical health data. The occurrence of any early fault will inevitably cause the actual dynamic behavior of the wind turbine pitch system to deviate from this normal pattern. To capture this deviation in a timely manner, a mechanism is needed to quantify the difference between the prediction and reality in real time. Therefore, in the technical solution of this invention, the predicted residual value is further compared with the actual residual value at the corresponding time to obtain an anomaly score, thereby transforming the degree of deviation of the system's health state into an instantaneous, quantifiable numerical indicator. This generates a time-series stream of anomaly scores, which characterize not the fluctuation of the original signal, but the deviation between the actual dynamic behavior of the system and the learned normal dynamic behavior pattern. Therefore, it has higher sensitivity to subtle pattern changes caused by early faults.

[0053] More specifically, in a specific example of the present invention, comparing the predicted residual value with the actual residual value at the corresponding time to obtain an anomaly score includes: calculating the squared error between the predicted residual value and the actual residual value at the corresponding time to obtain the anomaly score.

[0054] More specifically, at any time point t, the predicted residual value for the current time point t is first obtained from the prediction process of the previous time point t-1. Simultaneously, the actual residual value for the current time point t is calculated using the real-time SCADA data stream and the operating condition baseline model. After ensuring that the timestamps of these two values—the predicted residual value and the actual residual value at the current time point t—are aligned, a comparison operation is performed. This involves calculating the squared error between the predicted residual value and the actual residual value at the corresponding time to obtain the anomaly score. For example, if the predicted residual value at time t is 0.05, while the synchronously obtained actual residual value at the corresponding time is 0.15, then the anomaly score at that time is the square of the difference between the two, 0.05 and 0.15. The calculation of the anomaly score is repeated second-by-second as the SCADA data is updated, thus producing a continuous, non-negative time series of anomaly scores.

[0055] In step S400, the time-series of abnormal scores in the time dimension is smoothed to obtain a health index. It should be understood that while the instantaneous abnormal score sequence generated in the aforementioned steps is highly sensitive to deviations in the behavior of the wind turbine pitch system, it may contain sharp fluctuations caused by sensor noise, communication jitter, or other non-fundamental instantaneous disturbances. Directly basing fault judgments on this glitch-laden sequence can easily lead to false alarms, thereby reducing the reliability and credibility of the entire diagnostic system. Therefore, in the technical solution of this invention, the time-series of abnormal scores in the time dimension is further smoothed to obtain a health index, thereby filtering out high-frequency noise and irrelevant instantaneous disturbances, and extracting the core trend reflecting the evolution of the system's health status from the fluctuating score sequence. This generates a smooth health index curve that clearly reflects the long-term degradation trend of the pitch system's health status, providing a solid foundation for setting stable and reliable early warning thresholds, thus significantly improving the accuracy of fault early warning.

[0056] More specifically, in a specific example of the present invention, smoothing the time-series stream of abnormal scores in the time dimension to obtain health indicators includes: calculating an exponentially weighted moving average of the time-series stream of abnormal scores to obtain health indicators.

[0057] In other words, more specifically, firstly, a smoothing coefficient is set. Its value ranges from 0 to 1, and this coefficient determines the decay rate of historical data weights. A smaller value... Values ​​that produce a smoother curve. For example, selecting... The initial value is 0.05. At the start of the monitoring process, the health indicator value at the first time point can be initialized to the first anomaly score. Thereafter, at each new time point... When a new abnormal score is calculated Then, based on the health indicators from the previous moment... To update current health indicators Its update formula is This calculation process iterates continuously over time, integrating a series of discrete and abrupt abnormal scores into a continuous and gently changing health indicator curve.

[0058] Specifically, in step S500, the health indicator is compared with a preset warning threshold. When the health indicator exceeds the preset warning threshold, an early fault warning signal is generated. It should be understood that since the health indicator generated in the aforementioned steps is a continuous curve characterizing the system's health status, in order to apply it to actual operation and maintenance decisions, this continuous quantitative information must be transformed into a clear, discrete judgment that can be executed by the automatic monitoring system. That is, an objective standard is needed to define the critical point at which the system transitions from a healthy state to an early fault state. Therefore, in the technical solution of this invention, the health indicator is further compared with a preset warning threshold, and an early fault warning signal is generated when the health indicator exceeds the preset warning threshold, thereby completing the closed loop from state quantification to fault decision-making and realizing the automatic output of diagnostic results. In this way, when the health status of the wind turbine pitch system deteriorates to a predetermined level, a clear, high-confidence warning signal can be automatically triggered, thereby providing timely and reliable decision support for the predictive maintenance of wind turbines and preventing further deterioration of potential faults.

[0059] More specifically, in a specific example of this invention, the generation process of the early warning signal first involves the offline setting of the early warning threshold. Using historical SCADA data in a confirmed healthy state that was not involved in model training, this data is completely input into the entire diagnostic process of the data-driven wind turbine pitch system fault diagnosis method described in this invention. This yields a time series of health indicators in a healthy state. The statistical distribution of this series is analyzed, and a fixed early warning threshold is selected as the preset early warning threshold. This preset early warning threshold is then deployed in the online monitoring system. After being put into online operation, the health indicators generated in real time are compared with the preset early warning threshold at each time step. When the health indicator exceeds the preset early warning threshold, an initial early fault early warning signal is generated. To increase the reliability of the early warning and avoid transient overshoot caused by a very small number of abnormal operating conditions, a duration condition is also set in the early warning logic. Only when the value of the health indicator continuously exceeds the preset early warning threshold for a preset duration, such as 10 minutes, will a formal early fault early warning signal be finally confirmed and generated. This signal is then transmitted to the wind farm central monitoring system and triggers the corresponding maintenance work order.

[0060] In summary, the data-driven fault diagnosis method for wind turbine pitch system according to embodiments of the present invention is explained. First, by constructing a baseline model of operating conditions, it accurately learns and isolates the strong correlation between external operating conditions such as wind speed and power and diagnostic parameters such as motor current, thereby extracting a residual time series that more purely reflects the inherent health state of the wind turbine pitch system. Then, for this residual time series, a time-series dynamic pattern prediction model is constructed, enabling it to deeply learn the dynamic evolution law of the residual under normal conditions. When an early fault occurs in the wind turbine pitch system, its actual dynamic behavior deviates from the learned normal pattern, leading to a significant increase in the predicted residual and forming a sensitive anomaly score. Finally, by smoothing the anomaly score, a health index is generated, achieving stable and accurate early warning of early faults in the pitch system.

[0061] This invention also provides a data-driven fault diagnosis system for wind turbine pitch systems.

[0062] Figure 5 is a block diagram of a data-driven wind turbine pitch system fault diagnosis system according to an embodiment of the present invention. As shown in Figure 5, the data-driven wind turbine pitch system fault diagnosis system 100 according to an embodiment of the present invention includes: a time series data acquisition module 110, used to input the time series stream of acquired SCADA data into a trained operating condition baseline model to obtain current residual time series data; a residual value prediction module 120, used to input the current residual time series data into a trained time series dynamic mode prediction model to obtain predicted residual values; a residual value comparison module 130, used to compare the predicted residual values ​​with the actual residual values ​​at the corresponding time to obtain anomaly scores; a time series stream smoothing processing module 140, used to smooth the time series stream of anomaly scores in the time dimension to obtain health indicators; and a fault early warning signal generation module 150, used to compare the health indicators with a preset early warning threshold, and generate an early fault early warning signal when the health indicators are greater than the preset early warning threshold.

[0063] The specific implementation method of the data-driven wind turbine pitch system fault diagnosis system provided in this embodiment of the invention can be found in the data-driven wind turbine pitch system fault diagnosis method provided in this embodiment of the invention, and will not be repeated here.

[0064] The data-driven wind turbine pitch system fault diagnosis system 100 according to embodiments of the present invention can be deployed in the edge computing unit at the wind turbine site, such as the main controller or dedicated industrial control computer deployed inside the wind turbine nacelle or tower base, and communicate with the wind turbine's SCADA monitoring system in real time. In one possible implementation, the data-driven wind turbine pitch system fault diagnosis system 100 according to embodiments of the present invention can be integrated into the main control system of the wind turbine as a software module or hardware module. For example, the core models used for fault diagnosis in the data-driven wind turbine pitch system fault diagnosis system 100, such as the operating condition baseline model, the time-series dynamic pattern prediction model, and the preset early warning threshold, can be trained, verified, and parameter-tuned offline on the back-end server of the wind farm control center. The trained and solidified model and parameter package can then be sent to the front-end diagnostic unit. Of course, the complete algorithms used for real-time fault diagnosis in the data-driven wind turbine pitch system fault diagnosis system 100, including residual calculation, time-series pattern prediction, anomaly score generation, and health index smoothing, can also be solidified in dedicated industrial computing hardware, such as an embedded processor or FPGA module inside the wind turbine main controller, to accelerate the processing of real-time data and the operation of diagnostic logic, and ensure low-latency output of the final fault warning signal.

[0065] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A data-driven fault diagnosis method for wind turbine pitch systems, characterized in that, The data-driven fault diagnosis method for wind turbine pitch systems includes: inputting the time-series stream of acquired SCADA data into a trained operating condition baseline model to obtain current residual time-series data; inputting the current residual time-series data into a trained time-series dynamic pattern prediction model to obtain predicted residual values; comparing the predicted residual values ​​with the actual residual values ​​at the corresponding time points to obtain anomaly scores; smoothing the time-series stream of anomaly scores in the time dimension to obtain health indicators; comparing the health indicators with a preset warning threshold, and generating an early fault warning signal when the health indicators exceed the preset warning threshold.

2. The data-driven fault diagnosis method for wind turbine pitch system according to claim 1, characterized in that, SCADA data includes operating data and target diagnostic parameters of the wind turbine pitch system. Operating data includes wind speed, generator power and pitch angle, while target diagnostic parameters include pitch motor current and pitch motor torque.

3. The data-driven fault diagnosis method for wind turbine pitch system according to claim 2, characterized in that, The time-series stream of acquired SCADA data is input into the trained operating condition baseline model to obtain current residual time-series data, including: training the operating condition baseline model based on the operating condition data to obtain the trained operating condition baseline model; inputting real-time operating condition data into the trained operating condition baseline model to obtain the predicted pitch motor current value; and subtracting the predicted pitch motor current value from the synchronously acquired actual pitch motor current value to obtain the current residual.

4. The data-driven fault diagnosis method for wind turbine pitch system according to claim 3, characterized in that, The baseline model for operating conditions is a gradient boosting regression tree model.

5. The data-driven fault diagnosis method for wind turbine pitch system according to claim 1, characterized in that, The process of inputting current residual time series data into a trained time-series dynamic pattern prediction model to obtain predicted residual values ​​includes: performing Z-score standardization on the current residual time series data to obtain standardized current residual time series data; dividing the standardized current residual time series data into multiple standardized current residual local time series data based on a sliding time window; encoding and sorting the multiple standardized current residual local time series data based on the time dimension to obtain a sequence of current residual encoding vectors; and inputting the sequence of current residual encoding vectors into the trained time-series dynamic pattern prediction model for information-passing prediction to obtain predicted residual values.

6. The data-driven fault diagnosis method for wind turbine pitch system according to claim 5, characterized in that, The sequence of current residual encoded vectors is input into the trained time-series dynamic pattern prediction model for information transfer prediction to obtain the predicted residual value. This includes: using the trained time-series dynamic pattern prediction model to perform time-series dynamic pattern prediction on the sequence of current residual encoded vectors to obtain the current residual time-series prediction vector; and decoding the current residual time-series prediction vector to obtain the predicted residual value.

7. The data-driven fault diagnosis method for wind turbine pitch system according to claim 6, characterized in that, The trained temporal dynamic pattern prediction model is used to perform temporal dynamic pattern prediction on the sequence of current residual encoded vectors to obtain the current residual temporal prediction vector. This includes: inputting the sequence of the current residual encoded vectors into an initial long short-term memory network to obtain a sequence of implicit semantic representations of the current residuals; and calculating the semantic deviation quantification index between the sequence of implicit semantic representations of the current residuals and the corresponding vectors at each time step in the sequence of current residual encoded vectors according to the following formula, to obtain a set of current residual semantic deviation quantification indices: ； ； ； ;in, The first in the sequence of current residual encoding vectors A current residual encoding vector The characteristic projection function of the current residual is... For current residual characteristics, a multilayer sensor is used. for The corresponding current residual encoding implicit semantic representation. The projection function for encoding semantic hidden features of current residuals. A multilayer perceptron is used to encode semantic hidden features for current residuals. Let L2 norm be the vector. For smoothing parameters, A set of quantitative indicators for semantic deviation of current residuals. for and The semantic deviation quantification index of the current residual between the two They are respectively The value is The semantic deviation quantification index of current residual is used; based on the set of semantic deviation quantification indexes of current residual, a set of semantic fidelity correction factors for current residual is generated according to the following formula: ； ;in, This represents the mean of all semantic deviation quantification indices for current residuals within the set of semantic deviation quantification indices. The standard deviation of each semantic deviation quantification index of current residuals in the set of semantic deviation quantification indices is given. and All are adjustable parameters. As a preset constant, Let e ​​be the value of the logarithmic function with the natural constant e as the base. for Activation function for The corresponding current residual semantic fidelity correction factor, This is a set of semantic fidelity correction factors for current residuals. They are respectively for The value is The semantic fidelity correction factor for the current residual is determined. Based on the set of current residual semantic fidelity correction factors, a semantic bias constraint is applied to the sequence of implicit semantic representations encoded by the current residual, to obtain a sequence of semantically enhanced implicit representations of the current residual: ； ;in, for inverse transform, for The corresponding semantically enhanced implicit representation of the current residual. This is a sequence of implicit representations with semantic enhancement for current residuals. They are respectively The value is The current residual semantic enhancement implicit representation is used; the sequence of the current residual semantic enhancement implicit representation is input into the terminal long short-term memory network to generate the current residual time series prediction vector.

8. The data-driven fault diagnosis method for wind turbine pitch system according to claim 1, characterized in that, The predicted residual value is compared with the actual residual value at the corresponding time to obtain an anomaly score, including: calculating the squared error between the predicted residual value and the actual residual value at the corresponding time to obtain the anomaly score.

9. The data-driven fault diagnosis method for wind turbine pitch system according to claim 1, characterized in that, The time series of outlier scores is smoothed to obtain health indicators, including: calculating the health indicators by performing an exponentially weighted moving average on the time series of outlier scores.

10. A data-driven fault diagnosis system for wind turbine pitch systems, characterized in that, The data-driven wind turbine pitch system fault diagnosis system includes: a time-series data acquisition module, used to input the time-series stream of acquired SCADA data into a trained operating condition baseline model to obtain current residual time-series data; a residual value prediction module, used to input the current residual time-series data into a trained time-series dynamic pattern prediction model to obtain predicted residual values; a residual value comparison module, used to compare the predicted residual values ​​with the actual residual values ​​at the corresponding time to obtain anomaly scores; a time-series stream smoothing module, used to smooth the time-series stream of anomaly scores in the time dimension to obtain health indicators; and a fault early warning signal generation module, used to compare the health indicators with a preset early warning threshold, and generate an early fault early warning signal when the health indicators are greater than the preset early warning threshold.

Images