Method for predicting remaining service life of gearbox of wind turbine generator based on space-time diagram network

Abstract

The invention relates to the technical field of life prediction of wind turbine gearboxes, in particular to a wind turbine gearbox remaining service life prediction method based on a space-time diagram network, which comprises the following steps of: installing sensors at key monitoring points of a wind turbine gearbox; performing position coding on data acquired by each sensor, and loading the position code to the acquired data to obtain a sample data tensor; taking the sample data tensor of each sensor as a node, and taking the cosine similarity between two nodes as an edge to construct graph data; and constructing a space-time diagram network, training the space-time diagram network, and inputting wind turbine generator gearbox diagram data of which the residual service life needs to be predicted into the space-time diagram network to obtain a prediction result. The method can effectively improve the prediction precision and reliability of the remaining service life of the gearbox of the wind turbine generator under the complex working condition, and provides powerful technical guarantee for intelligent operation and maintenance in the wind power industry.

Description

Technical Field

[0001] This invention relates to the field of wind turbine gearbox life prediction technology, and in particular to a method for predicting the remaining service life of wind turbine gearboxes based on spatiotemporal graph networks. Background Technology

[0002] As a cornerstone of the global clean energy transition, the reliability and economy of wind energy are of paramount importance. Wind turbines, as large and complex pieces of equipment operating under harsh conditions for extended periods, have transmission systems, especially gearboxes, that are undoubtedly among the most critical and vulnerable components. Wind turbine gearboxes are constantly subjected to complex loads caused by the randomness, impact, and wide-ranging variations resulting from wind condition changes. A serious failure can lead to substantial losses in power generation due to downtime. Therefore, ensuring the healthy operation of wind turbine gearboxes is of great practical and economic significance for improving wind farm operational efficiency and reducing levelized cost of energy (LCOE). Traditional scheduled maintenance and reactive repair strategies are no longer sufficient to meet the demands of modern wind farm lean operation. Real-time monitoring and data analysis of equipment status are crucial for accurately predicting its remaining service life, allowing for the rational planning of maintenance timing and resources before failures occur. Developing high-precision prediction of the remaining service life of wind turbine gearboxes is key to transforming the operation and maintenance model from "passive response" to "proactive intervention," and is a core technological support for ensuring the safe, stable, and efficient operation of wind turbines.

[0003] Modern wind turbines are typically equipped with a multi-sensor monitoring system distributed throughout key components of the gearbox, capable of continuously collecting multi-dimensional, high-frequency status data. These sensors include, but are not limited to, vibration acceleration sensors, acoustic emission sensors, oil analysis sensors, temperature sensors, and torque and speed sensors. This multi-source, heterogeneous sensor data collectively constitutes a "multi-dimensional portrait" of the gearbox's health status, providing far richer and more comprehensive information than any single sensor's data. However, the sheer volume of multi-sensor data itself is not directly equivalent to value. Therefore, efficiently fusing this multi-sensor data and extracting robust and reliable degradation characteristics has become a key task in improving the accuracy and robustness of remaining service life predictions.

[0004] In recent years, with the development of artificial intelligence technology, data-driven methods for predicting remaining useful life (UPS) have become the mainstream research paradigm. Deep learning models, such as Long Short-Term Memory (LSTM) networks and their variants, in particular, have shown great potential in the field of UPS prediction due to their powerful automatic feature extraction and sequence modeling capabilities. These methods attempt to learn the end-to-end mapping relationship from raw data to UPS directly from historical monitoring data, avoiding reliance on complex physical models. However, despite significant progress, existing, especially purely data-driven methods, still face two interconnected core bottlenecks when addressing the multi-sensor UPS prediction problem of wind turbine gearboxes:

[0005] First, the depth of multi-sensor fusion is insufficient. Many existing methods for fusing multi-sensor data remain at a shallow stage, either in the early or late stages. These methods fail to explicitly model the spatiotemporal topological dependencies between sensors. The sensor network of a gearbox is essentially a graph structure, with nodes being sensors and edges representing physical connections or functional associations. Ignoring this inherent graph topology makes it impossible to achieve truly deep, physically interpretable information fusion, limiting the model's ability to capture the dynamics of system-level fault propagation.

[0006] Secondly, the lack of physical mechanisms. Purely data-driven remaining useful life prediction models are like a "black box," their learning process relying entirely on statistical patterns in the data, making it difficult to effectively incorporate prior physical knowledge. This leads to two problems: first, the model's generalization ability drops sharply under unknown operating conditions outside the training data distribution, because the features it learns may only be statistical correlations under specific operating conditions, rather than universal physical degradation laws; second, the prediction results lack physical consistency, sometimes yielding predictions that violate basic physical principles, greatly reducing their credibility and usability in engineering practice.

[0007] Given the research gaps and technical challenges mentioned above, there is an urgent need to propose a novel method for predicting the remaining useful life of wind turbine gearboxes that integrates multi-sensor data and physical mechanisms. This method should not only deeply explore the spatiotemporal correlation characteristics within the multi-sensor data, but also embed physical prior knowledge into the data-driven model to guide the learning process, thereby achieving a remaining useful life prediction that combines high accuracy, strong robustness, and good interpretability. Summary of the Invention

[0008] To address the problem of inaccurate prediction of the remaining service life of wind turbine gearboxes due to insufficient depth and lack of physical mechanisms in existing multi-sensor fusion models, this invention proposes a method for predicting the remaining service life of wind turbine gearboxes based on spatiotemporal graph networks, specifically including the following steps:

[0009] Sensors are installed at various key monitoring points in the wind turbine gearbox. The data collected by each sensor is encoded at a location, and the location code is loaded onto the collected data to obtain a sample data tensor.

[0010] The sample data tensor of each sensor is regarded as a node, and the cosine similarity between two nodes is used as the edge to construct the graph data;

[0011] A spatiotemporal graph network is constructed and trained. The gearbox diagram data of the wind turbine that needs to be predicted for its remaining service life is input into the spatiotemporal graph network to obtain the prediction results.

[0012] Compared with existing technologies, the spatiotemporal graph network in this invention integrates the spatial topology modeling capabilities of graph convolutional networks with the advantages of multi-head self-attention mechanisms in capturing long-range temporal features. It achieves adaptive updates of node dependencies through a dynamic spatiotemporal attention mechanism, constructing a deep spatiotemporal fusion framework for multi-dimensional sensor data. Secondly, it embeds a trainable Laplacian wavelet kernel convolution module to explicitly encode physical degradation trends into the feature extraction process, strengthening the fine-grained table of degradation features under mechanistic constraints through the synergistic effect of global-local temporal convolution. This invention can provide strong technical support for intelligent operation and maintenance in the wind power industry. Attached Figure Description

[0013] Figure 1 This is a spatiotemporal graph network structure diagram in the spatiotemporal graph network prediction method for wind turbine gearboxes of the present invention;

[0014] Figure 2 This is a diagram of the neural network structure of the present invention;

[0015] Figure 3 This is a diagram of the multi-head attention network structure of the present invention;

[0016] Figure 4 This is a network structure diagram of the trainable Laplacian wavelet kernel convolution module of the present invention;

[0017] Figure 5 This is the spatiotemporal aggregation network application process of the present invention;

[0018] Figure 6 This is the source of the data used in the verification experiments of this invention;

[0019] Figure 7 The data used for training and validation;

[0020] Figure 8 This is a schematic diagram of the network prediction results. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] This invention proposes a method for predicting the remaining service life of wind turbine gearboxes based on spatiotemporal graph networks, specifically including the following steps:

[0023] Sensors are installed at various key monitoring points in the wind turbine gearbox. The data collected by each sensor is encoded at a location, and the location code is loaded onto the collected data to obtain a sample data tensor.

[0024] The sample data tensor of each sensor is regarded as a node, and the cosine similarity between two nodes is used as the edge to construct the graph data;

[0025] A spatiotemporal graph network is constructed and trained. The gearbox diagram data of the wind turbine that needs to be predicted for its remaining service life is input into the spatiotemporal graph network to obtain the prediction results.

[0026] In this embodiment, multiple types of sensors deployed at key monitoring points of the gearbox acquire multi-channel raw monitoring data through synchronous sampling, forming a multi-dimensional time series signal. In the formula, M represents M sensors and L represents the signal length; for the raw data A sliding window segmentation is performed to construct a regular temporal sample set suitable for deep learning model input, and the training set and validation set are divided in a 9:1 ratio to retain effective dynamic information during the degradation process.

[0027] Preferably, a sliding window with length W and step size S is used for multi-dimensional time series signals. Perform sliding sampling to construct the sample dataset. ,in This represents a multi-sensor sample data set with a time series length of W and M sensors, where N represents the number of sample data sets.

[0028] As an optional implementation method, this embodiment performs quality management and data cleaning on the collected sample data, specifically including:

[0029] ① Data cleaning: The interquartile range method is used in combination with operating condition information to identify and remove abnormal values ​​caused by instantaneous sensor failures or communication interference;

[0030] ② Noise reduction processing: Adaptive wavelet threshold noise reduction is applied to the vibration acceleration signal to suppress environmental noise while retaining the impact fault components;

[0031] ③ Resampling and Alignment: Signals with different sampling rates are averaged using a sliding window and aligned with timestamps to be unified to the same time granularity, forming a synchronized multivariate time series;

[0032] ④ Standardization: Perform Z-score standardization on the data of each sensor channel to eliminate the influence of units, making model training more stable and efficient.

[0033] like Figure 1 This invention includes an encoder and a decoder, wherein the encoder mainly includes a multi-head attention layer and a graph convolutional layer, and the encoder mainly includes a Laplacian wavelet kernel convolution module and a multi-layer gated recurrent unit matched to the number of sensors. Figure 5 In this embodiment, vibration signals are first collected at key locations of the device based on the installed sensors. The key locations include at least the transmission connection locations of each device in the wind turbine. The collected data is processed, including cleaning, adding position encoding, and constructing graph data. Based on the constructed data, the feature extraction module of the encoder-decoder architecture extracts features for life prediction.

[0034] For Transformer-like spatiotemporal graph network models, the model itself lacks the ability to process positional information in sequential data. Therefore, positional encoding is needed to provide the model with positional information for time-series data. Positional encoding is generated using sine and cosine wave functions, ensuring the uniqueness of positions and the learning of relative positional relationships. Positional encoding is learned along with other model parameters, enabling the spatiotemporal graph network model to better handle time-series data. To provide the model with positional information for each time step, the dimensions of the temporal positional encoding matrix are generated. :

[0035]

[0036]

[0037] in, The value represents the temporal sequence order, and its value is the length W of the constructed sample data. i represents the input temporal dimension, and d represents the dimension of the location encoding. For multi-sensor data, only the temporal information is considered, and the value can be 1.

[0038] To provide the model with location information from temporal data, temporal location encoding is added to the original multi-sensor sample data to obtain a location-encoded sample data tensor, represented as follows: ; The location of the sample data is encoded as follows: .

[0039] To model the complex interdependencies between sensor signals, cosine similarity is used to evaluate feature associations by measuring the directional consistency between feature vectors derived from the sensors. A graph adjacency matrix of location-encoded multisensor data is constructed to capture node-level dependencies in a computationally efficient and interpretable manner. Location-encoded multisensor data is treated as vertices of a graph based on sensor type, and edges are established based on the signal correlations between multisensor time-series data, thus constructing a graph-location-encoded multisensor data set.

[0040] Use this data tensor Each sensor is considered a node in the graph, and the node's characteristic is the time-series data of the position encoding of the m-th sensor. Furthermore, a graph edge is constructed based on the cosine similarity of sensor signals to capture the spatial dependencies between sensors. For any two sensor nodes, the cosine similarity between them is calculated to evaluate the consistency of signal direction.

[0041]

[0042] in, Represents the dot product. express The modulus length. Further generation. Weighted adjacency matrix :

[0043]

[0044]

[0045] The expanded dimension adds temporal positional encoding to the positional encoding sample data. The above yields the graph-location encoded sample data tensor. .

[0046] This invention employs graph convolutional networks and multi-head self-attention mechanisms to capture the spatial and temporal dependencies of multi-sensor data. In this embodiment, the encoder performs spatial feature extraction sequentially through a combination of multi-head self-attention mechanisms and graph convolutional units, including the following steps:

[0047] At the first layer of the unit, the encoder first uses a multi-head self-attention mechanism to globally model the graph-position encoded multi-sensor data in the time dimension. By calculating the attention weights of all time steps, it achieves remote perception of key long-range degradation events.

[0048] Then, the encoder uses graph convolution operations to extract spatial dependencies between sensor nodes in graph-position encoded multi-sensor data and capture spatially related residual lifetime degradation propagation features.

[0049] After the spatiotemporal features are extracted at each layer of the unit, the features are fused by residual connection and layer normalization to avoid gradient vanishing.

[0050] The encoder serially processes graph-position encoded multi-sensor data and outputs a spatial representation of the remaining lifetime features that includes the spatiotemporal dependencies of the multi-sensor data.

[0051] This invention employs a multi-head self-attention mechanism to globally model graph-location encoded multi-sensor data in the temporal dimension. By calculating the attention weights at all time steps, it achieves remote sensing of key long-range degradation events. The structure of the multi-head attention mechanism layer in this embodiment is as follows: Figure 3 As shown, the multi-head attention layer, at the core of the multi-head attention mechanism, enables the model to dynamically weight the importance of different segments of the input sequence relative to the prediction task. This is achieved through a scaled dot-product attention operation, where the input is projected onto three distinct representations—a query (Q), a keyword (K), and a value (V) matrix—through a learnable parameter matrix. Each head performs this operation in parallel in a lower-dimensional space, allowing the model to collectively process information from different representation subspaces. The outputs of all heads are then concatenated and linearly transformed to produce the final participating representation. This design not only enhances the model's performance in recognizing significant temporal and cross-sensor features, but also... (The sentence is incomplete and requires further context to translate accurately.) The encoder first uses a multi-head attention mechanism layer along the sample data. Mapping the temporal dimension to output the multi-head attention intermediate layer features of the encoder. :

[0052]

[0053]

[0054]

[0055] in, Indicates a multi-head attention layer; Indicates the join function; This represents the output of the i-th head attention layer; This indicates the number of head attention layers, and its value is the number of sensors. , and These are the learnable weight matrices used to obtain the corresponding Q, K, and V matrices, respectively.

[0056] This embodiment utilizes graph convolution operations to extract spatial dependencies between sensor nodes in graph-location encoded multi-sensor data, capturing spatially correlated remaining lifetime degradation propagation features. For example... Figure 4As shown, a graph convolutional network (GCNN) is a neural network that takes graph data as input. Therefore, GCNNs are used to capture the inherent spatial topological dependencies in sensor networks, modeling the interaction patterns of different sensor nodes based on their physical layout. For the graph-position encoded sample data tensor input to the encoder... The encoder then uses a spectrum-based graph convolutional network along the relational dimension of the graph edges and neighbors in the sample data. Mapping is performed by transferring and learning features through node neighborhood aggregation operations based on the Laplacian matrix. While preserving the topological structure information of the graph, the final output is the intermediate layer features of the encoder graph. :

[0057]

[0058]

[0059]

[0060] Among them, by Add the identity matrix get This ensures that nodes retain and aggregate their own characteristics during information propagation, rather than relying solely on neighboring nodes to obtain information through a normalized adjacency matrix to prevent numerical inflation or deflation. ; It is the degree matrix representing the nodes; It is a weight matrix with learnable parameters that performs a linear transformation on the node features; It is the ReLU activation function.

[0061] In this embodiment, after the spatiotemporal feature extraction is completed at each layer of the unit, the features are fused through residual connections and layer normalization to avoid gradient vanishing. After serial processing, the encoder outputs a spatial representation of the remaining lifetime features containing multi-sensor spatiotemporal dependencies in the form of key K and value V. Specifically, as shown... Figure 2 The encoder structure shown first uses a multi-head self-attention mechanism to extract features from the intermediate multi-head attention layer. Then, residual connections and layer normalization are introduced; furthermore, graph convolution is used to extract features from intermediate layers of the graph. Then, by introducing residual connections and layer normalization, the final output is a spatial representation of the remaining lifetime features that includes multi-sensor spatiotemporal dependencies. :

[0062]

[0063]

[0064]

[0065]

[0066] in, This represents the layer normalization function. Assume... For the encoder in the spatiotemporal graph network, the process of extracting the remaining lifetime feature spatial representation containing multi-sensor spatiotemporal dependencies by the encoder can be further simplified as follows:

[0067]

[0068] The decoder inputs multi-sensor data and first uses Laplacian wavelet kernel convolution and multi-gated recurrent units to mine the temporal dependencies of the multi-sensor data to generate a temporal representation of the remaining lifetime features. Then, it uses a multi-head self-attention mechanism to fuse the spatial representation captured by the encoder and finally outputs a high-order spatiotemporal representation of the remaining lifetime features.

[0069] This embodiment employs a trainable Laplacian wavelet kernel convolution module in its decoder. This module combines the physical priors of gearbox degradation with data-driven feature extraction to achieve accurate fault component extraction from the input position-encoded multi-sensor data. The Laplacian wavelet kernel convolution module uses parameterized Laplacian wavelet basis functions, whose asymmetric shape highly matches the impact-response degradation features. Through backpropagation, it adaptively optimizes the wavelet scale (controlling waveform width to match fault components of different frequencies) and translation parameters (controlling waveform position to match the time point of fault occurrence), achieving accurate extraction of fault-related components such as transient impacts and modulation phenomena from vibration signals. The decoder also introduces a multi-gated recurrent unit layer, which simultaneously captures time dependence and inter-sequence correlation through its edge memory GRU module and joint memory GRU module, outputting the temporal representation of remaining lifetime features in the form of a query Q. The decoder further utilizes a multi-head self-attention mechanism, using temporal representation Q as the query and spatial representation output by the encoder as key K and value V, to calculate the similarity between dynamic physical information and degradation features, guiding the model to focus on the most relevant degradation patterns. Finally, it achieves accurate fusion of spatiotemporal information through weighted summation, constructs a high-order spatiotemporal representation of remaining lifetime features containing high-order spatiotemporal dependencies, and realizes accurate tracking of gearbox degradation patterns.

[0070] Unlike symmetric wavelets, Laplace wavelets exhibit inherent asymmetry, making them particularly well-suited for matching transient and asymmetric degradation patterns commonly found in real vibration or torque signals from wind turbine gearboxes. Through continuous scaling (adjusting frequency support) and shifting (temporally localizing), this module extracts multi-resolution temporal features, enabling precise detection of sudden fault pulses and trends caused by gradual wear. Therefore, this invention employs a trainable Laplace wavelet kernel convolution to enhance feature extraction, characterizing non-stationary and nonlinear degradation behavior in multi-sensor signals. Based on the mathematical principles of continuous wavelet transform, the trainable Laplace wavelet kernel convolution combines a trainable Laplace wavelet kernel within a deep learning framework to adaptively capture discriminative time-frequency features indicative of mechanical degradation. The trainable Laplace wavelet kernel convolution performs convolution operations using a learnable Laplace wavelet kernel optimized during training, directly operating on embedded and position-encoded multi-sensor time-series data, decomposing it into rich multi-scale representations. This allows the model to highlight physically meaningful, dynamically sensitive features associated with system degradation. For the position-encoded sample data input to the decoder The decoder first uses a Laplacian wavelet kernel convolution along the sample data. Convolution is performed on the temporal dimension, and the output decoder is convolved with intermediate layer features using multiple Laplacian wavelet kernels. :

[0071]

[0072]

[0073]

[0074] in, Features of the intermediate layer of the multi-Laplace wavelet kernel convolution; Indicates the max pooling layer. Indicates the batch normalization layer; Let W be the time series length of the i-th sample data. express Data with a time series of t; This is the wavelet kernel convolution operation function; This represents the wavelet kernel operation under the nth channel. This is the scaling factor for the wavelet kernel operation in the nth channel. The shift factor for wavelet kernel operation in the nth channel is: C represents the number of channels.

[0075] To more effectively analyze multi-sensor time-series data, this invention introduces a multi-gated recurrent unit (mGRU) layer. The mGRU simultaneously captures temporal dependencies and inter-sequence correlations through its built-in edge-memory GRU modules and joint-memory GRU modules. Specifically, the mGRU layer is designed with M edge-memory GRUs, each dedicated to extracting temporal dynamics from a single sensor variable. These modules independently model the autotemporal evolution of each sensor stream. Simultaneously, a joint-memory GRU is introduced to integrate and fuse interaction information between different sensor groups at each time step, thereby explicitly capturing cross-sensor dependencies and cooperative behaviors. By combining dedicated and shared memory mechanisms, the mGRU layer enhances representation learning of multi-sensor time data, improving the model's ability to identify degradation patterns in complex systems under dynamic operating conditions. This is achieved by using the multi-Laplacian wavelet kernel convolutional intermediate layer features output from the decoder front end. Data from the m-th sensor and the w-th time series The forward propagation of the m-th edge memory GRU in the multi-gated cyclic unit layer can be expressed as:

[0076]

[0077]

[0078]

[0079]

[0080] in, , , and These represent the reset gate, final memory, update gate, and edge candidate memory of the w-th time-series m-th edge storage module, respectively; the other parameters represent the parameters to be trained for the m-th edge storage module. Let represent the Tanh activation function. Then, the multi-gated recurrent unit layer uses a joint memory module to combine M marginal candidate memories and analyzes their inter-sequence correlation on the w-th time series. The recurrent forward propagation of the joint memory GRU can then be expressed as:

[0081]

[0082]

[0083]

[0084] in, , and These represent the joint candidate memory, joint update gate, and final complete output memory of the w-th temporal joint memory module, respectively; the other parameters represent the parameters to be trained for the m-th edge memory module. Through the cooperative recurrent forward propagation of the above multi-gated recurrent unit layers, the final output decoder multi-gated recurrent unit intermediate layer features are obtained. :

[0085] mGRU( ).

[0086] This invention utilizes a multi-head self-attention mechanism, with temporal representation Q as the query and spatial representation output by the encoder as key K and value V. It calculates the similarity between dynamic physical information and degradation features, guides the model to focus on the most relevant degradation mode, and finally achieves accurate fusion of spatiotemporal information through weighted summation. It constructs a high-order spatiotemporal representation of remaining lifetime features containing high-order spatiotemporal dependencies, thereby achieving accurate tracking of gearbox degradation modes.

[0087] Specifically, the intermediate layer features of the decoder's multi-gated recurrent unit... Treated as a query, it incorporates physical prior features extracted by the Laplacian wavelet kernel convolution module; and spatially represents the remaining lifetime features that are spatiotemporally dependent on multiple sensors. Treated as keys and values, it carries global spatial topology and temporal location information of multiple sensors, ensuring that query Q can accurately match the spatial degradation features of relevant sensors; a multi-head self-attention mechanism performs a weighted summation of the temporal similarity between Q and K, and further integrates the spatial features of V to generate a spatiotemporal representation of remaining lifetime features containing higher-order spatiotemporal dependencies. :

[0088]

[0089]

[0090] in, This indicates that in the multi-head attention mechanism, Mapped to a query vector This indicates that in the multi-head attention mechanism, Mapped to key vectors This indicates that in the multi-head attention mechanism, Mapped to a value vector for transpose; Indicates a splicing operation; This represents the output of the h-th head in multi-head attention, where h is the number of heads in the multi-head attention process; This indicates attention calculation; is the dimension of the key vector.

[0091] This invention obtains a continuous predicted value of the final remaining lifetime by linear combination and activation function mapping. For the high-order spatiotemporal representation of the input remaining lifetime features, a flattening layer is first used to flatten the dimensions, followed by a linear combination using two fully connected layers. The output dimension is 1, and the activation function is sigmoid to ensure the output value is non-negative, ultimately outputting a continuous predicted value of the remaining lifetime. :

[0092]

[0093] Where y is the final prediction result; For the spatiotemporal representation of remaining lifetime features containing higher-order spatiotemporal dependencies; This represents a two-layer fully connected layer function.

[0094] As an optional embodiment, this invention constructs a spatiotemporal graph network based on a training set, using the MSE loss function and the Adam optimizer. The spatiotemporal graph network is trained on the constructed dataset, using the MSE loss function, expressed as:

[0095]

[0096] Where N is the number of training samples in each training batch. To predict remaining lifespan; This represents the actual remaining lifespan.

[0097] Parameter updates based on the Adam optimizer:

[0098]

[0099] in, These are the network parameters for the (t+1)th iteration; For bias correction, For second-order moment estimation, the first two values ​​are automatically calculated by the Adam optimizer based on gradient calculation. This is the learning rate.

[0100] To verify the effectiveness of the proposed method for predicting the remaining service life of wind turbine gearboxes based on spatiotemporal graph networks, a comparative experiment was conducted using multi-sensor monitoring data from the gearbox of a 2 MW doubly-fed wind turbine generator at a wind farm in Xinjiang. This type of wind turbine gearbox employs a single-stage planetary gear and a two-stage parallel shaft structure. Figure 6The wind farm is equipped with a Condition Monitoring System (CMS). Eight acceleration vibration sensors are placed at key mechanical interfaces of the gearbox (meshing points, bearing housings) to capture physical signals reflecting mechanical wear and degradation characteristics, including RMS values ​​and spectral peak values ​​that directly characterize the wear process. The CMS collects vibration data four times daily at a sampling frequency of 25.6 kHz, each lasting 1.28 seconds, for a cumulative daily sampling of 5.12 seconds, generating a total of 131,072 samples. Monitoring data is saved in CSV format, supplemented by daily maintenance records and fault reports, recording abnormal observations, alarm times, and fault types. Based on wind farm fault reports and on-site expert opinions, this paper selects gearbox monitoring data (denoted as Dataset_WT1 and Dataset_WT2) from two units of the same model in the same wind farm to evaluate model performance. For each unit, data from 180 days (WT1) and 108 days (WT2) before the fault are selected, and it is assumed that the remaining service life decreases linearly. Figure 7 The original vibration data of the sensor located at measuring point 5 are shown, where (a) is Dataset_WT1 and (b) is Dataset_WT2.

[0101] The proposed spatiotemporal graph network was used to generate remaining lifetime prediction results for experiments with Dataset_WT1 and Dataset_WT2, as shown below. Figure 8 As shown in Figure (a), the remaining lifetime prediction results for Dataset_WT1 are shown, and Figure (b) shows the remaining lifetime prediction results for Dataset_WT2. The projected curves are closely aligned with the actual curves, indicating a slight prediction error. The prediction curves fluctuate in the early and middle stages due to insufficient spatial dependencies between network nodes at the beginning. These dependencies require further adaptive training. However, the prediction results in later stages show considerable stability.

[0102] Comparative experiments were conducted on the Dataset_WT1 and Dataset_WT2 datasets using four baseline methods to demonstrate the progress of spatiotemporal graph networks. The models selected for comparison were HAGCN, GGCN, LOGO, and DVGTformer. The results of the comparative experiments are shown in Table 1. Statistical analysis shows that spatiotemporal graph networks achieved the lowest RMSE and Score values ​​across all datasets, resulting in the most accurate RUL predictions. This can be attributed to the overall integration of spatiotemporal dependencies across multiple sensors by spatiotemporal graph networks, and their physically informed approach for dynamically adapting graph structures.

[0103] Table 1. Results of the spatiotemporal graph network and baseline algorithm based on the Dataset_WTs dataset.

[0104]

[0105] This invention proposes a method for predicting the remaining service life of gearboxes based on multi-sensor data fusion, namely, a spatiotemporal graph network. In the proposed spatiotemporal graph network model, an attention mechanism is used to capture the spatiotemporal dependence of multi-dimensional degradation feature sequences. Through random neighbor sampling and dynamic attention aggregation of a graph convolutional network, cross-channel temporal correlations related to equipment health status are adaptively fused, dynamically adapting to changes in the spatial correlation of sensor nodes during the degradation process. A trainable Laplacian wavelet kernel convolution module extracts the time-frequency features of physical signals based on a trainable Laplacian wavelet kernel, while a cascaded multivariate gated memory unit focuses on core degradation patterns. Through feature refinement, long-term degradation trends and short-term transient signals related to equipment deterioration are preserved, ensuring that physical features always anchor to the inherent degradation patterns of the equipment. Finally, performance evaluation and comparison were conducted based on a wind turbine gearbox dataset. Experimental results show that the proposed spatiotemporal graph network outperforms the baseline model, exhibiting excellent robustness and generalization ability in predicting the remaining service life of wind turbine gearboxes.

[0106] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the remaining service life of wind turbine gearboxes based on spatiotemporal graph networks, characterized in that, Specifically, the following steps are included: Sensors are installed at various key monitoring points in the wind turbine gearbox. The data collected by each sensor is encoded at a location, and the location code is loaded onto the collected data to obtain a sample data tensor. The sample data tensor of each sensor is regarded as a node, and the cosine similarity between two nodes is used as the edge to construct the graph data; A spatiotemporal graph network is constructed and trained. The gearbox diagram data of the wind turbine that needs to be predicted for its remaining service life is input into the spatiotemporal graph network to obtain the prediction results.

2. The method for predicting the remaining service life of wind turbine gearboxes based on spatiotemporal graph networks according to claim 1, characterized in that, The spatiotemporal graph network comprises an encoder-decoder architecture. In the encoder, features extracted from the sample data using a multi-head self-attention mechanism are residually concatenated with the sample data tensor and then normalized to obtain attention features. Next, features extracted from the attention features are residually concatenated with the attention features using graph convolution and then normalized to obtain the remaining lifetime feature space representation. In the decoder, Laplacian wavelet kernel convolution is performed along the temporal dimension of the sample data to obtain multi-Laplacian wavelet kernel convolution intermediate layer features. These intermediate features are input to a time feature extraction module composed of cascaded multi-gated recurrent units, resulting in multi-gated recurrent unit intermediate layer features. The spatiotemporal graph network employs a multi-head self-attention mechanism, using the multi-gated recurrent unit intermediate layer features as query input and the remaining lifetime feature space representation as key-value input to calculate a remaining lifetime feature spatiotemporal representation containing higher-order spatiotemporal dependencies. This representation is first flattened through a flatten layer, then linearly combined using two fully connected layers, and finally activated using a Sigmoid function to obtain the final prediction result.

3. The method for predicting the remaining service life of wind turbine gearboxes based on spatiotemporal graph networks according to claim 2, characterized in that, The encoder data processing procedure includes: in, To utilize the multi-head self-attention mechanism to extract features from sample data; This indicates a multi-head attention mechanism. This indicates that the multi-head attention mechanism will use sample data. Mapped to a query vector This indicates that the multi-head attention mechanism will use sample data. Mapped to key vectors This indicates that the multi-head attention mechanism will use sample data. Mapped to a value vector; For adding position-encoded sample data, i.e., sample data tensors; Indicates the layer normalization function; Attention features; Features extracted from attention features using graph convolution; For adjacency information; The weight matrix contains learnable parameters and is used to perform linear transformations on the node features. It is the ReLU activation function; It represents the feature space of remaining lifetime.

4. The method for predicting the remaining service life of wind turbine gearboxes based on spatiotemporal graph networks according to claim 2, characterized in that, The process of using Laplacian wavelet kernel convolution to convolve along the temporal dimension of the sample data to obtain the intermediate layer features of the multi-Laplacian wavelet kernel convolution includes: in, Features of the intermediate layer of the multi-Laplace wavelet kernel convolution; Indicates the max pooling layer. Indicates the batch normalization layer; Let W be the time series length of the i-th sample data. express Data with a time series of t; This is the wavelet kernel convolution operation function; This represents the wavelet kernel operation under the nth channel. This is the scaling factor for the wavelet kernel operation in the nth channel. The shift factor for wavelet kernel operation in the nth channel is: C represents the number of channels.

5. The method for predicting the remaining service life of wind turbine gearboxes based on spatiotemporal graph networks according to claim 2, characterized in that, In the time feature extraction module composed of cascaded gated loop unit layers, the cyclic forward propagation of the m-th gated loop unit layer includes: in, , , and These represent the outputs of the reset gate, final memory, update gate, and edge candidate memory of the w-th time-series m-th gated recurrent unit layer, respectively; the other parameters represent the parameters to be trained for the m-th edge memory module. This represents the Tanh activation function; It is the ReLU activation function; , , , , , This is the weight matrix that needs to be trained; , , The bias vector that needs to be trained; The forward propagation of a multi-gated cyclic unit layer can be represented as: in, , and These represent the joint candidate memory, joint update gate, and final complete output memory of the w-th time-series multi-gated cyclic unit layer, respectively. , , This is the weight matrix that needs to be trained; , The bias vector that needs to be trained; These are the features of the intermediate layer of the multi-Laplace wavelet kernel convolution.

6. The method for predicting the remaining service life of wind turbine gearboxes based on spatiotemporal graph networks according to claim 2, characterized in that, Using the intermediate layer features of the multi-gated recurrent unit as the query input and the remaining lifetime feature space representation as the key and value input, the spatiotemporal representation of the remaining lifetime feature containing higher-order spatiotemporal dependencies is calculated, including: in, For the spatiotemporal representation of remaining lifetime features containing higher-order spatiotemporal dependencies; This represents the multi-head attention mechanism; This indicates that in the multi-head attention mechanism, Mapped to a query vector This indicates that in the multi-head attention mechanism, Mapped to key vectors This indicates that in the multi-head attention mechanism, Mapped to a value vector for transpose; Indicates a splicing operation; This represents the output of the h-th head in multi-head attention, where h is the number of heads in the multi-head attention process; This indicates attention calculation; is the dimension of the key vector.

7. The method for predicting the remaining service life of wind turbine gearboxes based on spatiotemporal graph networks according to claim 2, characterized in that, The final prediction result is expressed as follows: Where y is the final prediction result; For the spatiotemporal representation of remaining lifetime features containing higher-order spatiotemporal dependencies; This represents a two-layer fully connected layer function.

8. The method for predicting the remaining service life of wind turbine gearboxes based on spatiotemporal graph networks according to claim 1, characterized in that, When training the spatiotemporal graph network, the mean squared error (MSE) of the predicted values ​​and the true values ​​in the training set is used as the loss function, and the Adam optimizer is used to update the network parameters.

9. The method for predicting the remaining service life of wind turbine gearboxes based on spatiotemporal graph networks according to any one of claims 1 to 8, characterized in that, A sliding window with length W and step size S for multi-dimensional time series signals Perform sliding sampling to construct the sample dataset. ,in This represents a multi-sensor sample data set with a time series length of W and including M sensors, where N represents the number of sample data sets.

10. The method for predicting the remaining service life of wind turbine gearboxes based on spatiotemporal graph networks according to claim 9, characterized in that, The temporal location code is added to the original multi-sensor sample data to obtain the location-coded sample data tensor, represented as follows: ; The location of the sample data is encoded as follows: ,in The value of pos is taken as the temporal length of the sample data, i is the dimension of each temporal sequence in the sample data, and d is the dimension of the location encoding. ; Expanding the dimension adds the adjacency matrix from the graph data to... The above yields the graph-location encoded sample data tensor. The adjacency matrix in the graph data includes: in, Representing sample data The graph weighted adjacency matrix; This represents the connection relationship between the m-th sensor and the i-th data sample of the n-th sensor; This represents the cosine similarity between the m-th sensor and the i-th data sample from the n-th sensor; Let represent the i-th data sample from the m-th sensor; M is the number of sensors.

Images