Wind turbine unit variable pitch reducer state online monitoring method and system

Abstract

The present application relates to the field of wind turbine variable pitch reducer state monitoring, and provides a wind turbine variable pitch reducer state online monitoring method and system, which synchronizes the high-frequency vibration data and low-frequency SCADA working condition data of the wind turbine variable pitch reducer in time, constructs a double-branch network model to extract deep features from the converted vibration time-frequency diagram and working condition vector sequence, then, a dynamic fusion strategy based on attention mechanism is introduced innovatively, the extracted macro working condition features are used as dynamic guides to adaptively focus on and filter the micro vibration feature components with the most information value under the current specific working condition, in this way, it is understood that under what operation what vibration is produced, so as to effectively distinguish the violent vibration caused by normal operation from the abnormal vibration caused by real fault, solve the diagnosis problem caused by the inability to effectively fuse working condition information, and improve the accuracy and reliability of monitoring.

Description

Technical Field

[0001] This invention relates to the field of wind turbine pitch reducer condition monitoring technology, and in particular to an online monitoring method and system for wind turbine pitch reducer condition. Background Technology

[0002] The pitch reducer is a core component of a wind turbine, responsible for adjusting blade angles, controlling aerodynamic energy capture, and managing operating loads. Its health directly impacts the overall power generation efficiency and operational safety of the wind turbine. Due to the randomness and fluctuations in natural conditions such as wind speed and direction, the pitch reducer operates under frequent start-stop cycles and variable speed / load conditions, making it a high-risk component for failure. A pitch reducer malfunction can lead to unplanned shutdowns of the wind turbine, resulting in significant power generation losses, or even more serious safety incidents.

[0003] To monitor the condition of pitch reducers, existing technologies typically employ vibration analysis, using sensors mounted on the reducer to collect vibration signals and determine its health status. Meanwhile, Supervisory Control and Data Acquisition (SCADA) systems for wind turbines record macroscopic operating condition data, including blade angle, wind speed, output power, and generator speed. However, existing technologies face significant challenges in fusing these two heterogeneous data sources. One such problem lies in the complex conditional dependence between the microscopic physical response of the equipment, as represented by vibration signals, and its macroscopic operating conditions. For example, during an emergency rapid pitch change caused by strong gusts, the severe impact vibration signal generated by the pitch reducer, if analyzed independently without considering the operating context, may exhibit characteristics highly similar to signals from severe faults such as broken gear teeth, leading to false alarms. Current solutions either analyze vibration signals in isolation or simply splice and fuse vibration characteristics with SCADA operating condition characteristics, failing to effectively establish a deep understanding and guidance relationship between macroscopic operating conditions and microscopic vibration characteristics. The shortcomings of this fusion method result in the model being unable to distinguish the true physical meaning of similar vibration signals under different operating conditions, thereby reducing the accuracy and reliability of the diagnosis and making it difficult to meet the actual needs of accurate condition assessment of pitch reducers under complex and variable operating conditions. 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 method and system for online monitoring of the status of wind turbine pitch reducer.

[0005] One aspect of the present invention provides a method for online monitoring of the status of a wind turbine pitch reducer, the method comprising: The time-series data of vibration data from wind turbine pitch reducer and the time-series data of SCADA are aligned based on timestamps to obtain synchronized data blocks. The synchronous data blocks are preprocessed in a dual-path parallel manner to obtain a sequence of vibration time-frequency diagrams and operating condition context vectors; The vibration time-frequency diagram and the sequence of operating condition context vector are passed through a feature capture device based on a dual-branch network to obtain vibration feature embedding and operating condition feature embedding; The working condition feature embedding and vibration feature embedding are dynamically fused based on an attention mechanism to obtain the working condition-vibration information fusion feature vector; The feature vector fused from the operating condition and vibration information is input into a classification network to obtain health status prediction data.

[0006] Optionally, the SCADA data includes blade angle, wind speed, output power, and generator speed, and the synchronization data block includes aligned vibration data segment and aligned SCADA data segment.

[0007] Optionally, the synchronous data blocks are preprocessed in dual parallel streams to obtain a sequence of vibration time-frequency diagrams and operating condition context vectors, including: The aligned vibration data segments are subjected to bandpass filtering and short-time Fourier transform to generate a vibration time-frequency diagram; Based on the aligned SCADA data segments, the blade angle change rate and power standard deviation are calculated to obtain the blade angle change rate time segment and the power standard deviation time segment. The blade angle change rate time segment, power standard deviation time segment, and aligned SCADA data segment are constructed into a sequence of operating condition context vectors.

[0008] Optionally, the vibration time-frequency diagram and the sequence of the operating condition context vector are passed through a feature capture device based on a dual-branch network to obtain vibration feature embedding and operating condition feature embedding, including: The vibration time-frequency map is passed through a dilated convolutional neural network in the vibration time-frequency sensing branch to obtain vibration feature embedding; Each working context vector in the sequence of working context vectors is passed through a multilayer perceptron in the working condition perception branch to obtain a sequence of working context feature vectors. The sequence of aggregated working condition context feature vectors is used to obtain the working condition feature embedding.

[0009] Optionally, aggregating the sequence of operating condition context feature vectors to obtain operating condition feature embedding includes: passing the sequence of operating condition context feature vectors through an LSTM encoder in the operating condition awareness branch to obtain operating condition feature embedding.

[0010] Optionally, the working condition feature embedding and vibration feature embedding are dynamically fused based on an attention mechanism to obtain a working condition-vibration information fusion feature vector, including: According to the following formula, the local temporal features of the working condition and the local temporal features of the joint vibration are subjected to feature value-level interaction to obtain the working condition-joint vibration feature value-level interactive encoding vector: ; in, This refers to the local temporal features of the operating conditions embedded in the operating condition features. This refers to the local temporal features of joint vibration embedded in vibration features. This represents element-wise multiplication. To sum by position, For interactive encoding of the weight matrix, For the interactive bias vector, It is a ReLU nonlinear activation function. This is an interactive encoding vector at the level of working condition-joint vibration feature value; According to the following formula, an explicit interaction analysis is performed on the eigenvalues ​​at each position in the cross-coding vector of load condition-joint vibration eigenvalue levels to obtain a set of explicit factors for the cross-interaction intensity of load condition-joint vibration: ; in, and These are the gated transformation weight matrix and the gated bias vector, respectively. for Activation function For normalization function, The mean of the cross-coding vector at the working condition-joint vibration eigenvalue level. The standard deviation of the cross-coding vector for the eigenvalues ​​of the working condition and joint vibration is given. Let e ​​be the value of the logarithmic function with the natural constant e as the base. It is a set of explicit factors of the interaction intensity of working conditions and joint vibrations; Based on the following formula, and using the set of explicit factors for the interaction intensity of load condition-joint vibration, the interaction encoding vectors of load condition-joint vibration eigenvalue levels are filtered to obtain the fusion feature vector of load condition-vibration information: ; in, The noise suppression coefficient and ; This is a random drop operation used to suppress noise signals; A vector containing only 1s; This is the feature vector for the fusion of operating condition and vibration information.

[0011] Optionally, the online monitoring method for the status of the wind turbine pitch reducer further includes: generating a status warning signal for the wind turbine pitch reducer when the health status prediction data is less than or equal to a preset warning threshold.

[0012] Another aspect of the present invention provides an online monitoring system for the status of a wind turbine pitch reducer, the online monitoring system comprising: The data alignment processing module is used to perform time-stamp-based alignment processing on the time-series vibration data of the wind turbine pitch reducer and the time-series SCADA data to obtain synchronized data blocks. The data preprocessing module is used to perform dual-channel parallel preprocessing on the synchronous data blocks to obtain a sequence of vibration time-frequency diagrams and operating condition context vectors; The feature embedding module is used to obtain vibration feature embedding and operating condition feature embedding by passing the sequence of vibration time-frequency diagram and operating condition context vector through a feature capture device based on a dual-branch network. The dynamic fusion module is used to dynamically fuse the embedded operating conditions and vibration features based on an attention mechanism to obtain a fused feature vector of operating conditions and vibration information. The health status prediction data generation module is used to input the fused feature vector of working condition-vibration information into the classification network to obtain health status prediction data.

[0013] Optionally, the synchronization data block includes aligned vibration data segments and aligned SCADA data segments; the data preprocessing module includes: The vibration time-frequency diagram generation unit is used to perform bandpass filtering and short-time Fourier transform on the aligned vibration data segments to generate a vibration time-frequency diagram. The blade angle change rate time series and power standard deviation time series calculation unit is used to calculate the blade angle change rate and power standard deviation based on the aligned SCADA data segments to obtain the blade angle change rate time series and power standard deviation time series. The operating condition context vector sequence construction unit is used to construct a sequence of operating condition context vectors from the blade angle change rate time segment, the power standard deviation time segment, and the aligned SCADA data segment.

[0014] Optionally, the feature embedding module includes: The vibration feature embedding acquisition unit is used to obtain vibration feature embedding by passing the vibration time-frequency map through the dilated convolutional neural network in the vibration time-frequency sensing branch; The working condition context feature vector sequence acquisition unit is used to obtain the sequence of working condition context feature vectors by passing each working condition context vector in the working condition perception branch through the multilayer perceptron. The working condition context feature vector sequence aggregation unit is used to aggregate the sequence of working condition context feature vectors to obtain the working condition feature embedding.

[0015] Optionally, the online monitoring system for the wind turbine pitch reducer also includes: The early warning module is used to generate a status early warning signal for the wind turbine pitch reducer when the predicted health status data is less than or equal to a preset early warning threshold.

[0016] Compared to existing technologies, this invention synchronizes high-frequency vibration data and low-frequency SCADA operating condition data of wind turbine pitch reducers in time, and constructs a dual-branch network model to extract deep features from the converted vibration time-frequency map and operating condition vector sequence to capture the microscopic physical response and macroscopic operating status of the equipment. Specifically, this invention does not simply splice features from high-frequency vibration data and low-frequency SCADA operating condition data, but innovatively introduces a dynamic fusion strategy based on an attention mechanism. It uses the extracted macroscopic operating condition features as a dynamic guide to adaptively focus on and filter the most informative microscopic vibration feature components under the current specific operating condition. In this way, it is possible to understand what kind of vibration was generated under what kind of operation, thereby effectively distinguishing between severe vibration caused by normal operation and abnormal vibration caused by real faults. This solves the diagnostic problem caused by the inability to effectively fuse operating condition information and improves the accuracy and reliability of wind turbine pitch reducer condition monitoring. Attached Figure Description

[0017] 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.

[0018] Figure 1 A flowchart of an online monitoring method for the status of a wind turbine pitch reducer provided by the present invention; Figure 2 This is a data flow diagram of another method for online monitoring of the status of a wind turbine pitch reducer provided by the present invention; Figure 3 The flowchart of another method for online monitoring of the state of a wind turbine pitch reducer provided by the present invention is as follows: the sequence of vibration time-frequency diagram and operating condition context vector is passed through a feature capture device based on a dual-branch network to obtain vibration feature embedding and operating condition feature embedding. Figure 4 The flowchart of another method for online monitoring of the condition of a wind turbine pitch reducer provided by the present invention is as follows: dynamic fusion of embedded operating conditions and embedded vibration features based on an attention mechanism to obtain a fused feature vector of operating conditions and vibration information. Figure 5 The present invention also provides a block diagram of an online monitoring system for the status of a wind turbine pitch reducer. Detailed Implementation

[0019] 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.

[0020] 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.

[0021] 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.

[0022] 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.

[0023] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0024] Currently, online monitoring technologies for wind turbine pitch reducers generally suffer from diagnostic challenges due to the inability to effectively integrate operating condition information. This makes it difficult for monitoring systems to distinguish between genuine fault signals and normal response impacts under varying operating conditions, resulting in a high false alarm rate. Therefore, this invention proposes an online monitoring method for the condition of wind turbine pitch reducers. Specifically, this method first performs precise timestamp-based alignment of the high-frequency vibration data time-series stream and the low-frequency SCADA data time-series stream of the wind turbine pitch reducer to form a synchronized data block. Next, through dual-path parallel processing, the vibration data is converted into a vibration time-frequency map, and the SCADA data is constructed into a sequence of operating condition context vectors that characterize the system's dynamic operation. Subsequently, a feature capturer based on a dual-branch network automatically extracts deep vibration feature embeddings and operating condition feature embeddings from the vibration time-frequency map and operating condition context, respectively. Then, this invention innovatively employs a dynamic fusion module based on an attention mechanism, using the operating condition feature embedding as a dynamic guide to adaptively enhance or suppress different information components in the vibration feature embedding, thereby generating a feature vector that deeply fuses operating condition and vibration information. Finally, this fused feature vector is input into a classification network to obtain predictive data that accurately reflects the true health status of the equipment under specific operating conditions, thus solving the diagnostic ambiguity problem caused by changes in operating conditions.

[0025] Figure 1 This is a flowchart of an online monitoring method for the status of a wind turbine pitch reducer according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the data flow in the online monitoring method for the state of a wind turbine pitch reducer according to an embodiment of the present invention. (In conjunction with...) Figure 1 and Figure 2According to an embodiment of the present invention, the online monitoring method for the status of a wind turbine pitch reducer includes the following steps: S100, performing time-stamp-based alignment processing on the time-series stream of vibration data and SCADA data of the wind turbine pitch reducer to obtain a synchronous data block; S200, performing dual-path parallel preprocessing on the synchronous data block to obtain a sequence of vibration time-frequency maps and operating condition context vectors; S300, passing the sequence of vibration time-frequency maps and operating condition context vectors through a feature capture device based on a dual-branch network to obtain vibration feature embedding and operating condition feature embedding; S400, performing dynamic fusion of the operating condition feature embedding and vibration feature embedding based on an attention mechanism to obtain an operating condition-vibration information fusion feature vector; S500, inputting the operating condition-vibration information fusion feature vector into a classification network to obtain health status prediction data.

[0026] Specifically, in step S100, the time-series data streams of the vibration data and SCADA data of the wind turbine pitch reducer are aligned based on timestamps to obtain synchronized data blocks. It should be understood that the time-series data streams of the vibration data and SCADA data of the wind turbine pitch reducer differ significantly in sampling frequency and physical meaning. The former reflects microscopic physical responses at the kilohertz level, while the latter records macroscopic operating states at the hertz level. Without processing, the misalignment between the two in the time dimension will prevent the establishment of a precise causal relationship between operating conditions and physical vibrations. Therefore, in the technical solution of this invention, the time-series data streams of the vibration data and SCADA data of the wind turbine pitch reducer are aligned based on timestamps to construct data units that simultaneously contain macroscopic operating condition information and microscopic vibration responses under a unified time reference. This provides a data foundation with precise causal relationships for subsequent feature fusion analysis, ensuring that each vibration data segment has its exact corresponding operating condition context, thereby avoiding diagnostic errors caused by data mismatch.

[0027] More specifically, in a specific example of the present invention, the SCADA data includes blade angle, wind speed, output power, and generator speed, and the synchronization data block includes aligned vibration data segments and aligned SCADA data segments. More specifically, the timestamp-based alignment process first uses low-frequency SCADA data time points as a reference to extract corresponding time segments from the continuous high-frequency vibration data stream to form a synchronization data block. In implementation, a first-in-first-out vibration data buffer is established to receive and temporarily store the high-frequency vibration data stream in real time. Whenever a SCADA data point carrying a precise timestamp is received, such as data containing values ​​for blade angle, wind speed, output power, and generator speed at a specific moment, a vibration data segment of 1 second in duration is extracted from the vibration data buffer, centered on the timestamp of that SCADA data point, according to a preset time window width, for example, 0.5 seconds before and after. This vibration data segment is the aligned vibration data segment. Subsequently, this aligned vibration data segment is combined with the SCADA data point that triggered this extraction operation and encapsulated into a synchronization data block. This process repeats continuously as SCADA data is input, eventually generating a series of time-aligned and information-complete synchronized data block sequences for subsequent processing.

[0028] Specifically, in step S200, the synchronous data block undergoes dual-path parallel preprocessing to obtain a sequence of vibration time-frequency maps and operating condition context vectors. It should be understood that the synchronous data block obtained in the previous step contains a one-dimensional vibration time-series signal in its original form and discrete SCADA operating condition snapshots. The inherent characteristics of these two types of data are not explicitly expressed, making them difficult to directly use for effective extraction of deep features. Therefore, in the technical solution of this invention, the synchronous data block is further preprocessed in dual-path parallel to obtain a sequence of vibration time-frequency maps and operating condition context vectors. This explicitly transforms the non-stationary, time-varying features contained in the vibration signal into a two-dimensional vibration time-frequency map through time-frequency transformation. Simultaneously, discrete SCADA data points are constructed into operating condition context vectors that reflect the dynamic trends of the system through derived feature calculation and serialization. In this way, two heterogeneous original data sources can be converted into structured representations most suitable for feature learning by deep network models, providing high-quality, information-dense inputs for subsequent capture of microscopic physical response features and macroscopic operating state features, which is a necessary prerequisite for achieving high-precision dynamic fusion diagnosis.

[0029] More specifically, in a specific example of the present invention, the synchronous data block is subjected to dual-path parallel preprocessing to obtain a sequence of vibration time-frequency diagrams and operating condition context vectors, including: performing bandpass filtering and short-time Fourier transform on the aligned vibration data segment to generate a vibration time-frequency diagram; calculating the blade angle change rate and power standard deviation based on the aligned SCADA data segment to obtain a blade angle change rate time sequence and a power standard deviation time sequence; and constructing the blade angle change rate time sequence, the power standard deviation time sequence, and the aligned SCADA data segment into a sequence of operating condition context vectors.

[0030] Accordingly, the aligned vibration data segment is bandpass filtered and subjected to short-time Fourier transform to generate a vibration time-frequency diagram. It should be understood that the original aligned vibration data segment is a one-dimensional time-series signal, which not only mixes the target characteristic frequency with background noise from irrelevant frequency bands, but also, under varying operating conditions, its fault characteristic frequency changes constantly, making it difficult for traditional frequency domain analysis to capture such dynamic characteristics. Therefore, in the technical solution of this invention, the aligned vibration data segment is further subjected to bandpass filtering and short-time Fourier transform to generate a vibration time-frequency diagram, thereby filtering out noise interference to improve the signal-to-noise ratio, and mapping the one-dimensional time-series signal to a two-dimensional time-frequency plane, thus intuitively revealing the evolution of signal frequency components over time. In this way, the non-stationary, time-varying fault features contained in the signal can be transformed into a structured image representation, making it suitable for subsequent feature extraction by convolutional neural networks and enhancing the identifiability of weak fault features.

[0031] More specifically, in a specific example of the present invention, the processing of the aligned vibration data segment first involves applying a digital bandpass filter to the aligned vibration data segment to obtain a filtered vibration data segment. The passband range of this digital bandpass filter is set based on prior mechanical knowledge such as the gear meshing frequency of the pitch reducer and the characteristic frequency of bearing failure, in order to retain effective frequency information and suppress low-frequency interference and high-frequency electromagnetic noise caused by tower sway. After obtaining the filtered vibration data segment, a short-time Fourier transform is performed on it. This process selects a window function, such as a Hanning window, and sets a fixed window length and overlap rate. The window slides along the entire filtered vibration data segment, and a fast Fourier transform is performed on the data within each window. The transformation results of all windows are finally arranged into a two-dimensional matrix, which is the vibration time-frequency diagram. The horizontal axis of this vibration time-frequency diagram represents time, and the vertical axis represents frequency. The value of each element in the two-dimensional matrix represents the energy intensity of the signal at a specific frequency at a specific time.

[0032] Accordingly, based on the aligned SCADA data segments, the blade angle change rate and power standard deviation are calculated to obtain the blade angle change rate time series and power standard deviation time series. It should be understood that the original aligned SCADA data segments are merely a series of discrete instantaneous state snapshots, which cannot directly reflect the dynamic characteristics and stability of the pitch system during operation, such as whether the blade angle is rotating at high speed or remaining stationary, or whether the output power is stable or fluctuates violently. Therefore, in the technical solution of this invention, the blade angle change rate and power standard deviation are further calculated based on the aligned SCADA data segments to obtain the blade angle change rate time series and power standard deviation time series, thereby extracting and quantifying derived features that characterize the dynamic behavior and load stability of the pitch system from the original state data. This transforms implicit operating condition change trends into explicit feature data, providing richer and more informative input for subsequently constructing an operating condition context vector that comprehensively describes macroscopic operating dynamics.

[0033] More specifically, in a concrete example of the present invention, a fixed-length time window is set to store the latest aligned SCADA data segment sequence. For calculating the blade angle change rate, at each new time step, the blade angle value at the current moment is taken and differs from the blade angle value at the previous moment. The difference result is then divided by the time interval to obtain the instantaneous change rate at that moment, and this is stored as a new data point in the blade angle change rate time series. For calculating the power standard deviation, all output power values ​​within the current sliding window are taken, and the standard deviation of these values ​​is calculated. This standard deviation reflects the recent power fluctuation amplitude and is stored as a data point in the power standard deviation time series. This process continues as new SCADA data segments are continuously added, thereby generating in real-time derived feature time series that are synchronized with the original data stream and reflect the operational dynamics.

[0034] Accordingly, the blade angle change rate time series, power standard deviation time series, and aligned SCADA data segments are constructed into a sequence of operating condition context vectors. It should be understood that the preceding steps have already obtained the aligned SCADA data segments representing the instantaneous state and the derived feature time series segments representing the dynamic trend, respectively. This information is scattered across multiple independent data streams and has not yet formed a unified and comprehensive operating condition description that can be directly processed by the feature extraction network. Therefore, in the technical solution of this invention, the blade angle change rate time series, power standard deviation time series, and aligned SCADA data segments are further constructed into a sequence of operating condition context vectors. This integrates the static and dynamic information describing the operating state of the pitch system into a high-dimensional feature vector at each time step. This generates a structured, highly condensed sequenced data that fully represents the macroscopic operating dynamics of the pitch system, providing a standardized data format that can be directly input and used for deep feature learning by the subsequent operating condition perception branch network.

[0035] More specifically, in a concrete example of the present invention, firstly, at any given time point, the numerical values ​​of each item in the aligned SCADA data segment at that moment are extracted, namely, blade angle, wind speed, output power, and generator speed, and simultaneously the values ​​of the blade angle change rate and power standard deviation calculated from the previous steps are extracted. Subsequently, these extracted values ​​from different data sources but completely corresponding in time are concatenated in a preset fixed order to form an original operating condition context vector containing all original and derived operating condition information. Considering that each feature value has different physical units and numerical ranges, to eliminate dimensional differences, the original operating condition context vector also needs to be normalized, for example, by using a max-min normalization method to linearly map all feature values ​​to the interval between 0 and 1. This process is repeated at each time point, ultimately forming a time series composed of the normalized operating condition context vector, i.e., a sequence of operating condition context vectors.

[0036] Specifically, in step S300, the vibration time-frequency diagram and the sequence of operating condition context vectors are processed by a feature capturer based on a dual-branch network to obtain vibration feature embedding and operating condition feature embedding. It should be understood that although the vibration time-frequency diagram and the sequence of operating condition context vectors generated in the preceding steps are rich in information, their features are still in a high-dimensional original representation space, and their data structures are quite different; the former is two-dimensional image-like data, and the latter is time-series vector data, making direct and effective association and fusion impossible. Therefore, in the technical solution of this invention, the vibration time-frequency diagram and the sequence of operating condition context vectors are further processed by a feature capturer based on a dual-branch network to obtain vibration feature embedding and operating condition feature embedding. This utilizes the unique advantages of different network structures, automatically learning image-level deep patterns representing the microscopic physical response of the reducer through the vibration time-frequency sensing branch, while simultaneously capturing the temporal dependencies representing the macroscopic operating state and dynamic trends of the wind turbine through the operating condition sensing branch. In this way, two heterogeneous high-dimensional input data can be mapped to low-dimensional, compact and highly information-condensed feature embedding spaces, forming equivalent vibration feature embeddings and working condition feature embeddings that can be deeply interactive, providing a high-quality feature foundation for subsequent dynamic fusion based on attention mechanisms.

[0037] Figure 3 This is a flowchart illustrating the online monitoring method for the condition of a wind turbine pitch reducer according to an embodiment of the present invention. It describes how a sequence of vibration time-frequency maps and operating condition context vectors is processed by a feature capture device based on a dual-branch network to obtain vibration feature embedding and operating condition feature embedding. (See flowchart for example.) Figure 3 As shown, step S300 includes: S310, passing the vibration time-frequency map through a dilated convolutional neural network in the vibration time-frequency sensing branch to obtain vibration feature embedding; S320, passing each working condition context vector in the sequence of working condition context vectors through a multilayer perceptron in the working condition sensing branch to obtain a sequence of working condition context feature vectors; S330, aggregating the sequence of working condition context feature vectors to obtain working condition feature embedding.

[0038] In step S310, the vibration time-frequency map is processed through a dilated convolutional neural network in the vibration time-frequency sensing branch to obtain vibration feature embedding. It should be understood that fault features in the vibration time-frequency map, such as sidebands caused by gear faults or harmonics from bearing faults, may exhibit a large-span but sparsely structured texture pattern in the spectrum. Traditional convolutional neural networks, when expanding the receptive field through pooling operations, lose positional and detailed information, potentially leading to insufficient capture of such weak and scattered features. Therefore, in the technical solution of this invention, the vibration time-frequency map is further processed through a dilated convolutional neural network in the vibration time-frequency sensing branch to obtain vibration feature embedding. This exponentially increases the effective receptive field of the convolutional kernel without reducing the feature map resolution, thereby capturing a wider range of contextual dependencies in a single convolution operation. This allows the network to perceive both local fine features and global structural information, thus more effectively extracting and characterizing deep fault modes with multi-scale characteristics that reflect the microscopic physical response of the reducer.

[0039] More specifically, in a concrete example of the present invention, the vibration time-frequency map is fed into a network structure consisting of multiple stacked dilated convolutional layers, namely a dilated convolutional neural network. The dilated convolutional layers at different levels in this network are configured with increasing dilation rates, for example, from 1 to 8, enabling the network to analyze input features layer by layer with an exponentially growing receptive field, thereby capturing multi-scale information from local details to global contours. Each dilated convolutional layer is followed by a non-linear activation function to enhance the model's expressive power. After processing by all dilated convolutional layers, a set of high-dimensional feature maps is obtained. Subsequently, global average pooling is applied to this set of feature maps, reducing the dimensionality of each two-dimensional feature map to a single numerical value. Finally, all pooled values ​​are concatenated and input into a fully connected layer, the output of which is the fixed-dimensional vibration feature embedding obtained by deep encoding the original vibration time-frequency map.

[0040] In step S320, each working context vector in the sequence of working context vectors is passed through a multilayer perceptron in the working condition perception branch to obtain a sequence of working context feature vectors. It should be understood that although the working context vectors constructed in the preceding steps integrate multi-source working condition information, the deep nonlinear coupling relationships between the various feature dimensions have not yet been revealed; it is still a relatively shallow feature representation. Therefore, in the technical solution of this invention, each working context vector in the sequence of working context vectors is further passed through a multilayer perceptron in the working condition perception branch to obtain a sequence of working context feature vectors. This allows for nonlinear mapping of the working context vectors at each time step, thereby learning and extracting higher-order abstract features that reflect the complex intrinsic relationships between various working condition parameters. In this way, the original working condition representation can be elevated to a more discriminative feature space, providing higher-quality feature input for subsequently accurately capturing the temporal dependencies of the entire sequence.

[0041] More specifically, in a concrete example of the present invention, firstly, a multilayer perceptron comprising an input layer, several hidden layers, and an output layer is constructed. Then, one of the operating condition context vectors in the sequence is input to the input layer of the multilayer perceptron. After passing through the input layer, the data is propagated forward layer by layer between the hidden layers. Each hidden layer applies a nonlinear activation function, such as a linear rectifier unit, to the weighted sum of the inputs from the previous layer and passes the calculation result to the next layer. Through this progressive nonlinear transformation, the multilayer perceptron can automatically learn and combine the original operating condition features. Finally, the output layer of the multilayer perceptron generates a new feature vector with adjusted dimensions; this feature vector is the operating condition context feature vector for that time step. This process is repeated for all operating condition context vectors in the input sequence, ultimately generating a new sequence of the same length as the original sequence, i.e., the sequence of operating condition context feature vectors.

[0042] In step S330, the sequence of operating condition context feature vectors is aggregated to obtain the operating condition feature embedding. It should be understood that the preceding steps yield a sequence of operating condition context feature vectors, which only describes the abstract state at each discrete time point and has not yet modeled the continuous changing trend and temporal dependencies over the entire time period. Therefore, in the technical solution of this invention, the sequence of operating condition context feature vectors is further aggregated to obtain the operating condition feature embedding. This allows for learning the dynamic evolution process of the entire sequence, capturing and summarizing the overall temporal pattern that can characterize the macroscopic operating state and dynamic trend of the wind turbine over a period of time. In this way, a series of discrete state snapshots can be condensed into a single, fixed-dimensional feature vector that comprehensively reflects the process dynamics, providing a global and trend-based operating condition context representation for subsequent fusion with vibration features.

[0043] More specifically, in a specific example of the present invention, aggregating the sequence of operating condition context feature vectors to obtain operating condition feature embedding includes: passing the sequence of operating condition context feature vectors through an LSTM encoder in the operating condition awareness branch to obtain operating condition feature embedding.

[0044] More specifically, firstly, the sequence of operating condition context feature vectors generated by the multilayer perceptron is input into the LSTM encoder sequentially. At each time step, the LSTM unit combines the input vector of the current time step with the internal state passed from the previous time step, selectively updating and passing information through its unique input gate, forget gate, and output gate structure, thereby effectively learning long-term dependencies in the sequence. After the last operating condition context feature vector in the sequence has been processed, the final hidden state vector of the LSTM unit is extracted. This final hidden state vector, because it encodes the temporal dynamic information of the entire input sequence, is directly embedded as an operating condition feature representing the macroscopic operating state and dynamic trend of the wind turbine.

[0045] In step S400, the condition feature embedding and vibration feature embedding are dynamically fused based on an attention mechanism to obtain a condition-vibration information fusion feature vector. It should be understood that after the initial feature value granularity interaction between the condition feature embedding and vibration feature embedding, the resulting initial encoding vector treats all dimensions of interaction information equally, failing to distinguish which interactions are truly key signals that contribute significantly to the diagnostic task, and which are irrelevant or even harmful condition noise. Therefore, in the technical solution of this invention, the condition feature embedding and vibration feature embedding are further dynamically fused based on an attention mechanism to obtain a condition-vibration information fusion feature vector. This explicitly models the relative importance of feature values ​​at each position in the initial interaction encoding vector, learns and generates a set of interaction intensity explicit factors that can quantify the interaction intensity, and uses these factors to perform interaction discrimination enhancement operations on the original interaction encoding vector. In this way, it is possible to achieve targeted amplification of key operating condition-vibration interaction signals and precise suppression of redundant noise signals, ultimately generating an operating condition-vibration information fusion feature vector that has been intelligently screened and purified, with a higher signal-to-noise ratio and stronger discriminative power, thus providing a high-quality decision basis for accurate diagnosis.

[0046] Figure 4 This is a flowchart illustrating the online monitoring method for the condition of a wind turbine pitch reducer according to an embodiment of the present invention, which dynamically fuses operating condition features and vibration features based on an attention mechanism to obtain a fused feature vector of operating condition and vibration information. Figure 4As shown, step S400 includes: S410, performing eigenvalue-level interaction on the local temporal features of the working condition and the local temporal features of the joint vibration to obtain the working condition-joint vibration eigenvalue-level interactive coding vector; S420, performing explicit interaction analysis on the eigenvalues ​​at each position in the working condition-joint vibration eigenvalue-level interactive coding vector to obtain a set of explicit factors of working condition-joint vibration interaction intensity; S430, based on the set of explicit factors of working condition-joint vibration interaction intensity, performing information discrimination and filtering on the working condition-joint vibration eigenvalue-level interactive coding vector to obtain the working condition-vibration information fusion feature vector.

[0047] In step S410, according to the following formula, the local temporal features of the working condition and the local temporal features of joint vibration are subjected to feature value level interaction to obtain the working condition-joint vibration feature value level cross-coding vector: ; in, This refers to the local temporal features of the operating conditions embedded in the operating condition features. This refers to the local temporal features of joint vibration embedded in vibration features. This represents element-wise multiplication. To sum by position, For interactive encoding of the weight matrix, For the interactive bias vector, It is a ReLU nonlinear activation function. It is an interactive encoding vector at the level of working condition-joint vibration feature value.

[0048] It should be understood that the local temporal features of the working condition and the local temporal features of joint vibration obtained in the preceding steps are two parallel feature vectors that have not yet established a direct correlation. The model needs an initial mechanism to explore the potential relationship between the two at the feature value granularity. Therefore, in the technical solution of this invention, the local temporal features of the working condition embedded in the working condition feature and the local temporal features of joint vibration embedded in the vibration feature are further subjected to feature value-level interaction to obtain a working condition-joint vibration feature value-level interactive encoding vector. In this way, through operations such as element-wise multiplication and positional summation, the additive and multiplicative relationship between the working condition and vibration is explicitly constructed in each dimension of the feature vector, and these diverse low-level interactive information are fused into a unified composite feature representation. In this way, a preliminary, unfiltered working condition-joint vibration feature value-level interactive encoding vector can be generated. This vector comprehensively contains the most direct correlation information between the two modal features, providing the necessary data foundation for subsequent screening and purification of key interaction modes through attention mechanisms.

[0049] In step S420, according to the following formula, an explicit interaction analysis is performed on the eigenvalues ​​at each position in the working condition-joint vibration eigenvalue level interactive coding vector to obtain a set of explicit factors for the working condition-joint vibration interactive intensity: ; in, and These are the gated transformation weight matrix and the gated bias vector, respectively. for Activation function For normalization function, The mean of the cross-coding vector at the working condition-joint vibration eigenvalue level. The standard deviation of the cross-coding vector for the eigenvalues ​​of the working condition and joint vibration is given. This represents the difference between each feature element and the mean in the working condition-joint vibration eigenvalue level cross-coding vector. Let be the value of the logarithmic function with the natural constant e as the base. The project constructs distance-aware decay, enhances interactions close to the mean (representing the normal mode), and suppresses outliers. It is a set of explicit factors of the interaction intensity of working conditions and joint vibration.

[0050] It should be understood that the condition-joint vibration eigenvalue level interaction encoding vector obtained in the preceding steps mixes valuable key interaction signals with redundant noise information, and all dimensions of features are treated equally, lacking an intrinsic mechanism to distinguish their relative importance. Therefore, in the technical solution of this invention, the interaction explicit analysis of the eigenvalues ​​at each position in the condition-joint vibration eigenvalue level interaction encoding vector is further performed to obtain a set of explicit factors of condition-joint vibration interaction intensity. This allows for the use of a composite mechanism combining gating transformation and distance-aware attenuation, not only learning the importance of interaction features from the data itself, but also constructing a distance-aware attenuation term using statistical principles, thereby strengthening the normal interaction pattern close to the mean and suppressing outliers. In this way, a set of quantifiable, explicit factors of condition-joint vibration interaction intensity can be generated. The set of these factors assigns dynamic, data-driven attention weights to each dimension of the interaction encoding vector, providing precise guidance for subsequent targeted information screening and enhancement.

[0051] In step S430, based on the set of explicit factors for the interaction intensity of working conditions and joint vibration, the working condition-joint vibration feature value level interaction coding vector is filtered to obtain the working condition-vibration information fusion feature vector according to the following formula: ; in, The noise suppression coefficient and . This is a random drop operation used to suppress noise signals. A vector whose elements are all 1s. This is the feature vector for the fusion of operating condition and vibration information.

[0052] It should be understood that the preceding steps have generated encoding vectors containing the original interaction information and a set of intensity factors for quantifying their importance, but these two types of information have not yet been combined to achieve substantial optimization of the original encoding vector. Therefore, in the technical solution of this invention, based on the set of explicit factors of the working condition-joint vibration interaction intensity, the working condition-joint vibration feature value level interaction encoding vector is further subjected to information discrimination and filtering to obtain a working condition-vibration information fusion feature vector. This set of explicit factors of the working condition-joint vibration interaction intensity is then used as dynamic weights to perform element-by-element weighting on the original working condition-joint vibration feature value level interaction encoding vector, thereby achieving targeted amplification of interaction signals judged to be important, and combining noise suppression coefficients and random discarding operations to accurately suppress secondary information. In this way, a working condition-vibration information fusion feature vector with higher signal-to-noise ratio and stronger discriminative power, which has undergone intelligent filtering and purification, can be finally output, providing a highly refined decision basis for the subsequent classification network, thus fundamentally improving the final performance of the diagnostic model.

[0053] Specifically, in step S500, the fused feature vector of operating condition and vibration information is input into a classification network to obtain health status prediction data. It should be understood that the fused feature vector of operating condition and vibration information generated in the preceding steps is an abstract numerical representation located in a high-dimensional feature space. Although it highly condenses equipment status information, it has not yet been mapped to a health status category with clear physical meaning that can be directly interpreted by maintenance personnel. Therefore, in the technical solution of this invention, the fused feature vector of operating condition and vibration information is further input into a classification network to obtain health status prediction data, thereby learning and establishing a nonlinear mapping relationship from the deeply fused feature space to the predefined health status label space. In this way, the complex feature vector can be transformed into an intuitive diagnostic conclusion representing the probability of various health states, thereby achieving end-to-end automatic assessment of the health status of the pitch reducer.

[0054] More specifically, in a concrete example of the present invention, the implementation process of the classification network begins with receiving a fused feature vector of operating condition-vibration information. This fused feature vector is first input to one or more fully connected layers. These fully connected layers further combine and abstract the fused features through weight matrix multiplication and transformation of nonlinear activation functions to learn the optimal classification decision boundary. The final layer of the classification network is an output layer, the number of which strictly corresponds to the number of preset health state categories, such as healthy, gear fault, and bearing fault. Finally, a normalization function, such as..., is applied to the output of this output layer. The function converts the output of the output layer into a probability distribution vector. Each element in this probability distribution vector corresponds to the predicted probability of a health state category, and the sum of all elements is 1. This probability distribution vector is the final output health state prediction data. Specifically, when the health state prediction data is less than or equal to a preset warning threshold, a status warning signal for the wind turbine pitch reducer is generated to achieve status warning.

[0055] In summary, the online monitoring method for the wind turbine pitch reducer according to embodiments of the present invention is explained. It synchronizes high-frequency vibration data and low-frequency SCADA operating condition data of the wind turbine pitch reducer in time, and constructs a dual-branch network model to extract deep features from the converted vibration time-frequency diagram and operating condition vector sequence, respectively, to capture the microscopic physical response and macroscopic operating state of the equipment. Specifically, the online monitoring method for the wind turbine pitch reducer according to embodiments of the present invention does not simply splice features from high-frequency vibration data and low-frequency SCADA operating condition data, but innovatively introduces a dynamic fusion strategy based on an attention mechanism. It uses the extracted macroscopic operating condition features as a dynamic guide to adaptively focus on and filter the most informative microscopic vibration feature components under the current specific operating condition. In this way, it is possible to understand what kind of vibration occurred under what operating conditions, thereby effectively distinguishing between severe vibration caused by normal operation and abnormal vibration caused by real faults. This solves the diagnostic problem caused by the inability to effectively fuse operating condition information, and improves the accuracy and reliability of wind turbine pitch reducer condition monitoring.

[0056] This invention also provides an online monitoring system for the status of a wind turbine pitch reducer.

[0057] Figure 5 This is a block diagram of an online monitoring system for the condition of a wind turbine pitch reducer according to an embodiment of the present invention. Figure 5As shown, the online monitoring system 100 for the condition of a wind turbine pitch reducer according to an embodiment of the present invention includes: a data alignment processing module 110, used to perform time-stamp-based alignment processing on the time-series stream of vibration data and SCADA data of the wind turbine pitch reducer to obtain a synchronous data block; a data preprocessing module 120, used to perform dual-path parallel preprocessing on the synchronous data block to obtain a sequence of vibration time-frequency maps and operating condition context vectors; a feature embedding module 130, used to pass the sequence of vibration time-frequency maps and operating condition context vectors through a feature capture device based on a dual-branch network to obtain vibration feature embedding and operating condition feature embedding; a dynamic fusion module 140, used to perform dynamic fusion of operating condition feature embedding and vibration feature embedding based on an attention mechanism to obtain an operating condition-vibration information fusion feature vector; and a health status prediction data generation module 150, used to input the operating condition-vibration information fusion feature vector into a classification network to obtain health status prediction data.

[0058] Specifically, the synchronization data block includes aligned vibration data segments and aligned SCADA data segments. The data preprocessing module 120 includes: a vibration time-frequency diagram generation unit, used to perform bandpass filtering and short-time Fourier transform on the aligned vibration data segments to generate a vibration time-frequency diagram; a blade angle change rate time series segment and power standard deviation time series segment calculation unit, used to calculate the blade angle change rate and power standard deviation based on the aligned SCADA data segments to obtain the blade angle change rate time series and power standard deviation time series; and a working condition context vector sequence construction unit, used to construct a sequence of working condition context vectors from the blade angle change rate time series, power standard deviation time series, and aligned SCADA data segments.

[0059] Specifically, the feature embedding module 130 includes: a vibration feature embedding acquisition unit, used to obtain vibration feature embedding by passing the vibration time-frequency map through a dilated convolutional neural network in the vibration time-frequency sensing branch; a working condition context feature vector sequence acquisition unit, used to obtain a sequence of working condition context feature vectors by passing each working condition context vector in the sequence of working condition context vectors through a multilayer perceptron in the working condition sensing branch; and a working condition context feature vector sequence aggregation unit, used to aggregate the sequence of working condition context feature vectors to obtain working condition feature embedding.

[0060] The specific implementation method of the online monitoring system for the status of the wind turbine pitch reducer provided in this embodiment of the invention can be found in the online monitoring method for the status of the wind turbine pitch reducer provided in this embodiment of the invention, and will not be repeated here.

[0061] The wind turbine pitch reducer online monitoring system 100 according to an embodiment of the present invention can be deployed in an edge computing unit at the wind turbine site, such as a dedicated industrial control computer deployed inside the wind turbine nacelle or tower base, and interact with vibration sensors installed on the pitch reducer and the wind turbine's SCADA monitoring system in real time. In one possible implementation, the wind turbine pitch reducer online monitoring system 100 according to an embodiment of the present invention can be integrated into the wind turbine's condition monitoring system as a separate software module or hardware module. For example, the core model used for condition diagnosis in the wind turbine pitch reducer online monitoring system 100, including a dual-branch feature capturer, a dynamic fusion network based on an attention mechanism, and a final classification network, can be trained, validated, and optimized offline using massive historical data on the backend server of the wind farm control center. The optimized model weight package can then be distributed to the front-end monitoring unit. Similarly, the complete diagnostic process used in the wind turbine pitch reducer online monitoring system 100 for real-time online monitoring, including multi-source data synchronization and alignment, dual-path parallel preprocessing, feature embedding and extraction, dynamic fusion, and health status prediction, can also be embedded in dedicated edge computing hardware, such as an embedded processor within the monitoring unit or an FPGA / GPU module with AI acceleration capabilities. This accelerates the processing of real-time data streams and the inference process of the diagnostic model, ensuring low-latency generation of the final condition warning.

[0062] 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 method for online monitoring of the status of a wind turbine pitch reducer, characterized in that, The online monitoring method for the status of the wind turbine pitch reducer includes: The time-series data of vibration data from wind turbine pitch reducer and the time-series data of SCADA are aligned based on timestamps to obtain synchronized data blocks. The synchronous data blocks are preprocessed in a dual-path parallel manner to obtain a sequence of vibration time-frequency diagrams and operating condition context vectors; The vibration time-frequency diagram and the sequence of operating condition context vector are passed through a feature capture device based on a dual-branch network to obtain vibration feature embedding and operating condition feature embedding; The working condition feature embedding and vibration feature embedding are dynamically fused based on an attention mechanism to obtain the working condition-vibration information fusion feature vector; The feature vector fused from the operating condition and vibration information is input into a classification network to obtain health status prediction data.

2. The online monitoring method for the status of a wind turbine pitch reducer according to claim 1, characterized in that, SCADA data includes blade angle, wind speed, output power, and generator speed. The synchronization data block includes aligned vibration data segment and aligned SCADA data segment.

3. The online monitoring method for the status of a wind turbine pitch reducer according to claim 2, characterized in that, The synchronous data blocks are preprocessed in dual parallel streams to obtain a sequence of vibration time-frequency diagrams and operating condition context vectors, including: The aligned vibration data segments are subjected to bandpass filtering and short-time Fourier transform to generate a vibration time-frequency diagram; Based on the aligned SCADA data segments, the blade angle change rate and power standard deviation are calculated to obtain the blade angle change rate time segment and the power standard deviation time segment. The blade angle change rate time segment, power standard deviation time segment, and aligned SCADA data segment are constructed into a sequence of operating condition context vectors.

4. The online monitoring method for the status of a wind turbine pitch reducer according to claim 1, characterized in that, The vibration time-frequency plot and the sequence of operating condition context vectors are processed by a feature capture function based on a dual-branch network to obtain vibration feature embeddings and operating condition feature embeddings, including: The vibration time-frequency map is passed through a dilated convolutional neural network in the vibration time-frequency sensing branch to obtain vibration feature embedding; Each working context vector in the sequence of working context vectors is passed through a multilayer perceptron in the working condition perception branch to obtain a sequence of working context feature vectors. The sequence of aggregated working condition context feature vectors is used to obtain the working condition feature embedding.

5. The online monitoring method for the status of a wind turbine pitch reducer according to claim 4, characterized in that, The sequence of aggregated working condition context feature vectors to obtain working condition feature embedding includes: passing the sequence of working condition context feature vectors through an LSTM encoder in the working condition awareness branch to obtain working condition feature embedding.

6. The online monitoring method for the status of a wind turbine pitch reducer according to claim 1, characterized in that, The working condition feature embedding and vibration feature embedding are dynamically fused based on an attention mechanism to obtain a working condition-vibration information fusion feature vector, including: According to the following formula, the local temporal features of the working condition and the local temporal features of the joint vibration are subjected to feature value-level interaction to obtain the working condition-joint vibration feature value-level interactive encoding vector: ； in, This refers to the local temporal features of the operating conditions embedded in the operating condition features. This refers to the local temporal features of joint vibration embedded in vibration features. This represents element-wise multiplication. To sum by position, For interactive encoding of the weight matrix, For the interactive bias vector, It is a ReLU nonlinear activation function. This is an interactive encoding vector at the level of working condition-joint vibration feature value; According to the following formula, an explicit interaction analysis is performed on the eigenvalues ​​at each position in the cross-coding vector of load condition-joint vibration eigenvalue levels to obtain a set of explicit factors for the cross-interaction intensity of load condition-joint vibration: ； in, and These are the gated transformation weight matrix and the gated bias vector, respectively. for Activation function For normalization function, The mean of the cross-coding vector at the working condition-joint vibration eigenvalue level. The standard deviation of the cross-coding vector for the eigenvalues ​​of the working condition and joint vibration is given. Let e ​​be the value of the logarithmic function with the natural constant e as the base. It is a set of explicit factors of the interaction intensity of working conditions and joint vibrations; Based on the following formula, and using the set of explicit factors for the interaction intensity of load condition-joint vibration, the interaction encoding vectors of load condition-joint vibration eigenvalue levels are filtered to obtain the fusion feature vector of load condition-vibration information: ； in, The noise suppression coefficient and ; This is a random drop operation used to suppress noise signals; A vector containing only 1s; This is the feature vector for the fusion of operating condition and vibration information.

7. The online monitoring method for the status of a wind turbine pitch reducer according to claim 1, characterized in that, The online monitoring method for the status of the wind turbine pitch reducer further includes: generating a status warning signal for the wind turbine pitch reducer when the health status prediction data is less than or equal to a preset warning threshold.

8. An online monitoring system for the status of a wind turbine pitch reducer, characterized in that, The online monitoring system for the wind turbine pitch reducer includes: The data alignment processing module is used to perform time-stamp-based alignment processing on the time-series vibration data of the wind turbine pitch reducer and the time-series SCADA data to obtain synchronized data blocks. The data preprocessing module is used to perform dual-channel parallel preprocessing on the synchronous data blocks to obtain a sequence of vibration time-frequency diagrams and operating condition context vectors; The feature embedding module is used to obtain vibration feature embedding and operating condition feature embedding by passing the sequence of vibration time-frequency diagram and operating condition context vector through a feature capture device based on a dual-branch network. The dynamic fusion module is used to dynamically fuse the embedded operating conditions and vibration features based on an attention mechanism to obtain a fused feature vector of operating conditions and vibration information. The health status prediction data generation module is used to input the fused feature vector of working condition-vibration information into the classification network to obtain health status prediction data.

9. The online monitoring system for the status of the wind turbine pitch reducer according to claim 8, characterized in that, The synchronization data block includes the aligned vibration data segment and the aligned SCADA data segment; The data preprocessing module includes: The vibration time-frequency diagram generation unit is used to perform bandpass filtering and short-time Fourier transform on the aligned vibration data segments to generate a vibration time-frequency diagram. The blade angle change rate time series and power standard deviation time series calculation unit is used to calculate the blade angle change rate and power standard deviation based on the aligned SCADA data segments to obtain the blade angle change rate time series and power standard deviation time series. The operating condition context vector sequence construction unit is used to construct a sequence of operating condition context vectors from the blade angle change rate time segment, the power standard deviation time segment, and the aligned SCADA data segment.

10. The online monitoring system for the status of the wind turbine pitch reducer according to claim 8, characterized in that, The feature embedding module includes: The vibration feature embedding acquisition unit is used to obtain vibration feature embedding by passing the vibration time-frequency map through the dilated convolutional neural network in the vibration time-frequency sensing branch; The working condition context feature vector sequence acquisition unit is used to obtain the sequence of working condition context feature vectors by passing each working condition context vector in the working condition perception branch through the multilayer perceptron. The working condition context feature vector sequence aggregation unit is used to aggregate the sequence of working condition context feature vectors to obtain the working condition feature embedding.

Images