Search

CN-121996959-A - Wind turbine generator system gear box residual service life prediction method based on space-time diagram network

CN121996959ACN 121996959 ACN121996959 ACN 121996959ACN-121996959-A

Abstract

The invention relates to the technical field of life prediction of wind turbine gearboxes, in particular to a method for predicting the residual service life of a wind turbine gearbox based on a space-time diagram network, which comprises the steps of installing sensors at key monitoring points of the wind turbine gearbox, carrying out position coding on collected data of each sensor, and loading the position coding on the collected data to obtain a sample data tensor; and constructing a space-time diagram network, training the space-time diagram network, and inputting the wind turbine generator gearbox diagram data needing to predict the residual service life into the space-time diagram network to obtain a prediction result. The method can effectively improve the prediction precision and reliability of the residual service life of the gearbox of the wind turbine under the complex working condition, and provides powerful technical guarantee for intelligent operation and maintenance of the wind power industry.

Inventors

  • YONG BIN

Assignees

  • 重庆城市管理职业学院

Dates

Publication Date
20260508
Application Date
20260119

Claims (10)

  1. 1. The method for predicting the residual service life of the wind turbine gearbox based on the space-time diagram network is characterized by comprising the following steps of: Installing sensors at each key monitoring point of a wind turbine generator gearbox, performing position coding on data acquired by each sensor, and loading the position coding on the acquired data to obtain a sample data tensor; Taking the sample data tensor of each sensor as a node, and taking cosine similarity between two nodes as edge construction graph data; And (3) constructing a space-time diagram network, training the space-time diagram network, and inputting the data of the gearbox diagram of the wind turbine generator, which needs to predict the residual service life, into the space-time diagram network to obtain a prediction result.
  2. 2. The method for predicting the residual service life of the wind turbine generator gear box based on the space-time diagram network is characterized in that the space-time diagram network comprises an encoder-decoder framework, in the encoder, the characteristics extracted from sample data by utilizing a multi-head self-attention mechanism are subjected to residual connection with sample data tensors and then subjected to layer normalization to obtain attention characteristics, then the characteristics extracted from the attention characteristics by utilizing diagram convolution are subjected to residual connection with the attention characteristics and then subjected to layer normalization to obtain residual service life characteristic space characterization, in the decoder, laplace wavelet kernel convolution is utilized to carry out convolution along the time sequence dimension of the sample data to obtain a multiple Laplace wavelet kernel convolution intermediate layer characteristic, the intermediate characteristic is input into a time characteristic extraction module formed by cascading multiple-gate-control circulating unit layers to obtain multiple-gate-control circulating unit intermediate layer characteristics, the space-time diagram network adopts the multi-gate-control circulating unit intermediate layer characteristics as query input, the residual service life characteristic space characterization is used as key and value input, the residual service life characteristic characterization containing high-time-space dependence is calculated, and the residual service life characteristic is firstly subjected to linear combination through a flat layer dimension, and then a full-level prediction result is obtained through a full-level function.
  3. 3. The method for predicting the remaining service life of a gearbox of a wind turbine generator based on a space-time diagram network according to claim 2, wherein the processing procedure of the encoder data comprises the following steps: Wherein, the Features extracted from sample data using a multi-headed self-attention mechanism; representing the multi-headed attentiveness mechanism, Representing sample data in a multi-head attention mechanism The mapping is to a query vector that is, Representing sample data in a multi-head attention mechanism The mapping is performed as a key vector, Representing sample data in a multi-head attention mechanism Mapping into a value vector; for adding position-coded sample data, i.e. a sample data tensor; Representing a layer normalization function; Is a attention feature; to convolve features extracted from the attention features using a graph; Is adjacency information; the weight matrix is a weight matrix of the learnable parameters and is used for carrying out linear transformation on node characteristics; Is a ReLU activation function; Characterizing the residual life characteristic space.
  4. 4. The method for predicting the remaining service life of a gearbox of a wind turbine generator based on a space-time diagram network according to claim 2, wherein the process of convolving along the time sequence dimension of the sample data by using Laplace wavelet kernel convolution to obtain the characteristics of a middle layer of the multiple Laplace wavelet kernel convolution comprises the following steps: Wherein, the Convolving the intermediate layer feature with a multiple laplace wavelet kernel; Representing the maximum pooling level of the layer, Representing a batch normalization layer; For the ith sample data, the time sequence length is W, Representation of Data with a medium time sequence of t; A wavelet kernel convolution operation function; representing a wavelet kernel operation under the nth channel, Scaling factors for wavelet kernel operations under the nth channel, The shift factor for wavelet kernel operation under the nth channel is C is the number of channels.
  5. 5. The method for predicting the remaining service life of a gearbox of a wind turbine generator based on a space-time diagram network according to claim 2, wherein in a time feature extraction module formed by cascading multiple gating circulating unit layers, circulating forward propagation of an mth gating circulating unit layer comprises: Wherein, the 、 、 And Respectively representing the outputs of a reset gate, a final memory, an update gate and an edge candidate memory of the mth gating cycle unit layer of the w time sequence; representing a Tanh activation function; Is a ReLU activation function; 、 、 、 、 、 The weight matrix is a weight matrix which needs to be trained; 、 、 Is a bias vector to be trained; the multi-gated loop cell layer loop forward propagation can be expressed as: Wherein, the 、 And The method comprises the steps of respectively representing a joint candidate memory, a joint update gate and a final complete output memory of a w time sequence multi-gate control circulating unit layer; 、 、 The weight matrix is a weight matrix which needs to be trained; 、 Is a bias vector to be trained; The intermediate layer features are convolved with a multiple laplace wavelet kernel.
  6. 6. The method for predicting the residual service life of the wind turbine gearbox based on the space-time diagram network according to claim 2, wherein the calculating the residual service life characteristic space-time representation comprising the high-order space-time dependence by taking the characteristics of the middle layer of the multi-gate control circulating unit as query input and the residual service life characteristic space representation as key and value input comprises the following steps: Wherein, the Space-time characterization for residual life characteristics including high-order space-time dependence; Representing a multi-headed attention mechanism; Representing multiple-head attention mechanisms The mapping is to a query vector that is, Representing multiple-head attention mechanisms The mapping is performed as a key vector, Representing multiple-head attention mechanisms The mapping is performed to a vector of values, Is that Is a transpose of (2); Representing a splicing operation; the output of the h head in the multi-head attention is represented, and h is the number of heads of the multi-head attention; representing an attention calculation; Is the dimension of the key vector.
  7. 7. The method for predicting the residual service life of a gearbox of a wind turbine generator based on a space-time diagram network according to claim 2, wherein the final prediction result is expressed as: wherein y is the final prediction result; space-time characterization for residual life characteristics including high-order space-time dependence; representing a double layer fully connected layer function.
  8. 8. The method for predicting the residual service life of the wind turbine gearbox based on the space-time diagram network according to claim 1, wherein when the space-time diagram network is trained, MSE of predicted values and true values in a training set is used as a loss function, and an Adam optimizer is used for updating network parameters.
  9. 9. The method for predicting residual service life of wind turbine gearbox based on space-time diagram network according to any one of claims 1-8, wherein the sliding window pair multidimensional time series signal with W as length and S as step length Sliding sampling is carried out to construct a sample data set Wherein Representing a multi-sensor sample data having a time sequence length W including M sensors, N representing the number of sample data.
  10. 10. The method for predicting remaining service life of wind turbine gearbox based on space-time diagram network as claimed in claim 9, wherein adding time sequence position codes to original multi-sensor sample data to obtain position code sample data tensors is expressed as ; Position coding of sample data, expressed as Wherein The value of pos takes the time sequence length of the sample data, i is the dimension of each time sequence in the sample data, d is the dimension of the position code, ; The expansion dimension adds the adjacency matrix in the graph data to Obtaining the tensor of the picture-position coding sample data Wherein the adjacency matrix in the graph data comprises: Wherein, the Representing sample data Is a graph weighted adjacency matrix; representing the connection relation between the mth sensor and the ith data sample of the nth sensor; Representing cosine similarity between the mth sensor and the ith data sample of the nth sensor; represents the ith data sample representing the mth sensor, M is the number of sensors.

Description

Wind turbine generator system gear box residual service life prediction method based on space-time diagram network Technical Field The invention relates to the technical field of life prediction of wind power gearboxes, in particular to a method for predicting the residual service life of a wind turbine gearbox based on a space-time diagram network. Background Wind energy is used as a foundation stone for global clean energy transformation, and the reliability and economy of the wind energy are important. As a large-scale complex device which operates under severe working conditions for a long time, the transmission system, particularly the gearbox, of the wind turbine is definitely one of the most critical and fragile links in the whole wind turbine. Wind power gearboxes bear random, impact and widely-changed complex loads caused by wind condition changes for a long time, and once serious faults occur, the wind power gearboxes cause huge shutdown power generation loss. Therefore, the healthy operation of the wind power gear box is ensured, and the method has great practical and economic significance for improving the operation efficiency of the wind power plant and reducing the standardized energy cost. Traditional timing maintenance and post-maintenance strategies have difficulty in meeting the requirements of modern wind farm lean operation. Through real-time monitoring and data analysis to equipment state, the remaining service life of the equipment is accurately predicted, so that maintenance time and resources are reasonably planned before faults occur. The development of the prediction of the residual service life of the wind power gear box with high precision is a key for realizing the transformation of the operation and maintenance mode from passive response to active intervention, and is a core technical support for guaranteeing the safe, stable and efficient operation of the wind power generation set. Modern wind turbines are typically equipped with multi-sensor monitoring systems that are distributed throughout the critical parts of the gearbox, enabling continuous acquisition of multi-dimensional, high frequency status data. Such sensors include, but are not limited to, vibration acceleration sensors, acoustic emission sensors, oil analysis sensors, temperature sensors, torque, rotational speed sensors, and the like. Together, these multi-source heterogeneous sensor data form a "multi-dimensional image" of the health status of the gearbox, far more abundant and comprehensive than any single sensor data can provide. However, massive amounts of multi-sensor data are not directly equivalent in value per se. Therefore, how to efficiently fuse the multi-sensor data and extract robust and reliable degradation features from the multi-sensor data becomes a key task for improving the prediction accuracy and the robustness of the residual service life. In recent years, with the development of artificial intelligence technology, a data-driven residual service life prediction method has become a mainstream research paradigm. Particularly, deep learning models, such as long-term and short-term memory networks and variants thereof, have great potential in the field of residual service life prediction due to the strong feature automatic extraction and sequence modeling capabilities. These methods attempt to learn the end-to-end mapping from the original data to the remaining useful life directly from the historical monitoring data, avoiding reliance on complex physical models. However, despite significant advances, existing, and in particular purely data-driven approaches, face two interrelated core bottlenecks in dealing with the wind-powered gearbox multi-sensor residual life prediction problem: First, the depth of multi-sensor fusion is insufficient. Many existing methods stay in the shallow stages of early or late fusion of multisensor data. Such methods fail to explicitly model the spatio-temporal topological dependencies that exist between sensors. The sensor network of the gearbox is essentially a graph structure, the nodes are sensors, and the edges represent physical connections or functional associations. Ignoring this inherent graph topology, deep information fusion with physical interpretability cannot be truly achieved, limiting the ability of the model to capture system-level fault propagation dynamics. And secondly, the absence of a physical mechanism. The pure data-driven residual service life prediction model looks like a 'black box', and the learning process is completely dependent on the data statistics rule, so that priori physical knowledge is difficult to be effectively integrated. The two problems are that the generalization capability of the model is drastically reduced under the unknown working condition outside the training data distribution, because the learned characteristics of the model can be only statistical correlation under specific working conditions, but not universal physi