Search

CN-121980372-A - Bearing fault identification method based on multiscale space-time synchronization attention and space-time alignment

CN121980372ACN 121980372 ACN121980372 ACN 121980372ACN-121980372-A

Abstract

The invention discloses a bearing fault identification method based on multi-scale space-time synchronization attention and space-time alignment, which carries out sliding window segmentation and normalization pretreatment on an original vibration signal; the method comprises the steps of constructing an adjacent matrix by adopting a K neighbor algorithm, extracting local spatial features by using a graph convolution neural network, extracting bidirectional sequence features by using a bidirectional gating circulation unit network, carrying out weighted fusion by combining a global attention mechanism, and realizing fault prediction by self-adaptive average pooling and a fully-connected classifier after space and time sequence features are spliced. The method integrates graph modeling, GCN, biGRU and attention mechanism, realizes the collaborative extraction of space-time dual-path characteristics, effectively solves the problems of weak space modeling and time sequence dependency deficiency in the traditional method, and remarkably improves the bearing fault recognition precision and generalization capability under the environment of multiple working conditions and strong noise.

Inventors

  • TAO TAO
  • FEI YUE
  • ZHOU HAO

Assignees

  • 安徽工业大学

Dates

Publication Date
20260505
Application Date
20260211

Claims (5)

  1. 1. A bearing fault identification method based on multi-scale space-time synchronization attention and space-time alignment is characterized by comprising the following steps: S1, preprocessing data, namely acquiring original vibration signals of a bearing through a sensor, constructing a multi-scale sliding window sampling mechanism, and generating signal fragment sets with different time resolutions; S2, constructing a graph structure and extracting spatial features, namely taking each scale signal segment as a graph node, constructing an adjacent matrix by using a K nearest neighbor algorithm, and extracting spatial topological feature streams of each scale through a graph convolution network; s3, time sequence modeling, namely inputting the normalized signal sequence into a bidirectional gating circulating unit, capturing bidirectional context information generated and propagated by fault characteristics, and forming a time characteristic stream; S4, space-time synchronous interaction, namely introducing a space-time synchronous attention module, and enabling the space topological feature stream and the time feature stream to interact in real time through a cross attention mechanism so as to realize mutual information enhancement of the features in an extraction stage; s5, space-time alignment and fusion, namely introducing a space-time alignment loss function, restraining response consistency of space features and time features at the occurrence time of faults, and carrying out deep fusion on the aligned features; s6, classification decision, namely outputting a bearing fault identification result through the fusion characteristic and the Softmax function through the full-connection layer.
  2. 2. The bearing fault identification method based on multi-scale spatiotemporal synchronous attention and spatiotemporal alignment according to claim 1, wherein the multi-scale sliding window in S1 comprises a long window and a short window, wherein the long window is used for capturing steady state fault evolution characteristics, and the short window is used for capturing transient impact characteristics.
  3. 3. The method for recognizing bearing faults based on multi-scale time-space synchronization attention and time-space alignment according to claim 1, wherein the time-space synchronization attention module automatically learns the mapping relationship between time segments and space topology nodes by constructing a time-space alignment matrix in S4.
  4. 4. The bearing fault identification method based on multi-scale spatiotemporal synchronous attention and spatiotemporal alignment according to claim 1, wherein the spatiotemporal alignment loss function in S5 is constructed based on cosine similarity or euclidean distance and is used for measuring the distribution consistency of the spatial feature stream and the temporal feature stream on a time axis.
  5. 5. The bearing fault identification method based on multi-scale space-time synchronization attention and space-time alignment according to claim 1, wherein the super-parameter K of the K nearest neighbor algorithm in S2 supports adaptive selection, and sparsity of a graph structure is adjusted in real time according to change of signal energy entropy.

Description

Bearing fault identification method based on multiscale space-time synchronization attention and space-time alignment Technical Field The invention relates to the technical field of mechanical equipment state monitoring and intelligent fault diagnosis, in particular to a bearing fault identification method based on multiscale spatio-temporal synchronization attention and spatio-temporal alignment. Background Bearings are used as key basic components of rotary mechanical equipment, and the running state of the bearings directly influences the safety and stability of the system. The traditional bearing fault diagnosis method is mostly dependent on manual extraction of time domain or frequency domain characteristics, and is difficult to adapt to strong noise and variable working conditions. With the development of deep learning, the prior art starts to employ Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs) for automatic feature extraction. For example, chinese patent application publication No. CN116861343a discloses a bearing fault diagnosis method that uses ACmix hybrid network model, extracts features through dual paths and performs simple fusion. However, the prior art still has the defects that firstly, local similarity and space dependence relationship which are implied between vibration signal fragments and are determined by physical mechanisms are difficult to capture effectively through regular convolution, so that space topological structure information is lost, secondly, the prior timing model is mostly independent branches, deep modeling of fault evolution dual-context information is lacked, key abnormal fragments are difficult to highlight, finally, the prior characteristic fusion mode is mostly simple weighting or splicing, depth interaction and alignment mechanism is lacked between the timing branches and space branch characteristics, so that when a model processes composite fault or high noise signals, the resonant point of a fault source in space-time dimension is difficult to lock accurately, and the diagnosis precision is limited. Disclosure of Invention 1. Technical problem to be solved by the invention The invention aims to solve the following key technical problems in the current bearing fault diagnosis: (1) The problem of missing modeling of the spatial topological relation is that the existing method cannot effectively model the local similarity and the spatial dependency relation which are implied in the one-dimensional vibration signal and are determined by a physical mechanism into a graph structure, so that precious spatial structure information is ignored. (2) The problem of insufficient extraction of time sequence dynamic and key information is that the existing time sequence model is difficult to synchronously capture the two-way context information of fault evolution, and an automatic discovery and reinforcement mechanism for key discriminant fragments in a sequence is generally lacking, so that the model is easily interfered by redundant information and noise. (3) The key feature identification is insufficient, and the key signal fragments are difficult to accurately identify by the existing method under the multi-working condition or high noise condition. (4) The feature fusion mode is single, and the lack of a mechanism for effectively fusing spatial and temporal features leads to insufficient diagnosis performance. Therefore, the invention provides a fault diagnosis network model, which is a multi-channel bearing fault recognition method based on the fusion of KNN-GCN and BiGRU-GATT, and the method is used for integrating graph modeling, GCN, biGRU and global attention mechanisms to fundamentally improve the characteristic modeling capacity and classification recognition precision of multi-channel vibration signals. 2. Technical proposal In order to achieve the above purpose, the technical scheme provided by the invention is as follows: the invention discloses a bearing fault identification method based on multi-scale space-time synchronization attention and space-time alignment, which comprises the following steps: S1, preprocessing data, namely acquiring original vibration signals of a bearing through a sensor, constructing a multi-scale sliding window sampling mechanism, and generating signal fragment sets with different time resolutions; S2, constructing a graph structure and extracting spatial features, namely taking each scale signal segment as a graph node, constructing an adjacent matrix by using a K nearest neighbor algorithm, and extracting spatial topological feature streams of each scale through a graph convolution network; s3, time sequence modeling, namely inputting the normalized signal sequence into a bidirectional gating circulating unit, capturing bidirectional context information generated and propagated by fault characteristics, and forming a time characteristic stream; S4, space-time synchronous interaction, namely introducing a