Search

CN-121997178-A - Bearing fault diagnosis method based on cross-flow characteristic enhancement and graph topology interaction

CN121997178ACN 121997178 ACN121997178 ACN 121997178ACN-121997178-A

Abstract

The invention discloses a bearing fault diagnosis method based on cross-flow characteristic enhancement and graph topology interaction. The method comprises the steps of constructing a dynamic kernel evolution network in a time domain flow, guiding convolutional kernel weight self-adaptive evolution by utilizing a time context, capturing transient impact and constructing a multi-resolution graph topology, constructing a frequency spectrum sparse purification network in a frequency domain flow, establishing a saliency assessment mechanism based on a learnable quantization matrix, utilizing a soft threshold switch to realize high-fidelity purification and denoising of fault key information, providing a heterogeneity perception cross-flow interaction mechanism for fully utilizing the flow characteristics, recovering time domain energy heterogeneity through a latent feature recalibration module, fusing the frequency domain characteristics to form cross-flow enhancement characterization, introducing the enhancement characterization into the self-adaptive graph Transformer, and driving node characteristics to perform cross-flow space dynamic interaction in the multi-scale topology by taking the graph topology as a framework. Experiments show that the invention can effectively overcome the problems of strong noise interference and characteristic homogenization and remarkably improve fault diagnosis precision.

Inventors

  • SU SHUZHI
  • YANG YANRAN
  • ZHU YANMIN

Assignees

  • 安徽理工大学

Dates

Publication Date
20260508
Application Date
20260129

Claims (4)

  1. 1. A bearing fault diagnosis method based on cross-flow characteristic enhancement and graph topology interaction comprises the following steps: (1) Collecting original vibration signals of the bearing in different health states on a rolling bearing fault experiment platform; (2) Constructing a dynamic nuclear evolution network, extracting time domain dynamic characteristics from signals, and constructing a multi-resolution graph topology; (3) Constructing a frequency spectrum sparse purification network, and realizing self-adaptive screening and noise resistance enhancement of a key frequency band; (4) Designing a heterogeneity perception cross-flow interaction mechanism to realize the depth alignment of the time domain latent feature recalibration module and the time-frequency space semantics; (5) The dynamic nuclear evolution network, the frequency spectrum sparse purification network and the heterogeneity perception cross-flow interaction mechanism are fused to form a bearing fault diagnosis method model based on cross-flow characteristic enhancement and graph topology interaction, and intelligent diagnosis of rolling bearing faults is achieved.
  2. 2. The method for diagnosing a bearing fault based on cross-flow feature enhancement and graph topology interaction of claim 1, wherein the constructing a dynamic core evolution network in step (2) extracts time domain dynamic features from signals and constructs a multi-resolution graph topology, comprising the steps of: (2a) Generating dynamic calibration coefficients For input feature sequences Wherein B is the batch size, C is the number of characteristic channels, T is the sequence length, global and local context descriptors are extracted respectively, and first, global average pooling is performed on input features in the time dimension to obtain a global context descriptor g: Where T represents the length of time of the feature sequence, The feature vector representing the t-th time step, then projecting and fusing the global context with the local features using a linear mapping function, Wherein the method comprises the steps of For the input sequence X of local feature vectors at time t, Representing the projection of the global descriptor g to a linear fully connected layer of the same dimension as the local feature, Representing the fused t-th time context characteristics, generating dynamic calibration coefficients based on the fused characteristics , Wherein the method comprises the steps of Is that The function is activated and the function is activated, A convolution layer for up-scaling the number of recovery channels is shown, Representing the function of the ReLU activation, A convolution layer for dimension reduction is shown, The dynamic calibration coefficient is generated for the t moment; (2b) Performing dynamic convolution operations The convolution kernel at the t-th moment Is decomposed into tensor products of reference weights and dynamic calibration coefficients, Wherein the method comprises the steps of Representing the convolution weights of the static reference, A broadcast multiplication operation is indicated and, An adaptive dynamic convolution kernel generated for input at time t, and then convolving the input signal with a dynamic convolution kernel, Wherein the method comprises the steps of A convolution operation is represented and is performed, Is the output characteristic after dynamic calibration; (2c) Construction of multi-resolution graph topology After multi-resolution time series embedding through dynamic nuclear evolution network, a set of resolution-dependent graph structures is constructed for each resolution level Is provided with For a feature representation output at this resolution, Wherein Representing the time length under the mth resolution, and performing dimension rearrangement on the embedded tensor to obtain a node characteristic matrix in order to construct a node characteristic representation based on the channel Constructing a corresponding graph based on a node feature matrix The connection between nodes is formed by a learnable adjacency matrix An implicit modeling is performed such that, Wherein the method comprises the steps of Representing a dimension permutation operation, i.e. adjusting the order of the tensors in dimension, Corresponds to the timing characteristics of a channel at that resolution.
  3. 3. The method for diagnosing bearing faults based on cross-flow characteristic enhancement and graph topology interaction according to claim 1, wherein the step (3) is characterized in that a frequency spectrum sparse purification network is constructed to realize adaptive screening and noise immunity enhancement of a key frequency band, and the steps are carried out as follows: (3a) Prefiltering and frequency domain transformation Firstly, introducing a gating convolution structure as a pre-filter, and primarily purifying a feature Z extracted by a dynamic nuclear evolution network by using a gating linear unit to obtain a pre-feature , Wherein the method comprises the steps of A linear projection is represented and is shown, Representing a gated convolution operation with a 1 x 1 convolution kernel, followed by a frequency domain transformation of the features, mapping the time domain features to a frequency domain space, Wherein the method comprises the steps of Representing a feature expansion operation for constructing a local image block, Representing a fast fourier transform, for transforming the unfolded features from the time domain to the frequency domain, Generating a frequency domain characteristic spectrum; (3b) Quantization matrix screening and frequency domain reconstruction Constructing a leachable quantization matrix Q adaptively aligned with the spectrum dimension, performing element-by-element weighted screening on the frequency domain characteristics, wherein each element in the quantization matrix serves as a soft threshold switch, adaptively enhancing the critical frequency band associated with the fault and suppressing the background noise, Where Q is a matrix of leachable quantization weights, Representing the multiplication by element, Finally, restoring the characteristic dimension through inverse folding operation, introducing residual connection and GEGLU activation function, and outputting the frequency domain enhancement characteristic with high signal-to-noise ratio , Wherein the method comprises the steps of Representing an inverse folding operation.
  4. 4. The bearing fault diagnosis method based on cross-flow feature enhancement and graph topology interaction of claim 1, wherein the design heterogeneity perception cross-flow interaction mechanism of step (4) realizes the depth alignment of a time domain latent feature recalibration module and time-frequency space semantics, and the steps are as follows: (4a) Time domain latent feature recalibration In order to solve the problem that low-frequency trend covers weak fault impact, a first-order finite difference idea is introduced into the front end of attention calculation, and firstly, proper filling operation is introduced into a time embedding dimension, so that a characteristic representation output under the Mth resolution is obtained Subsequently, a first-order finite difference between adjacent time steps is calculated, Wherein the method comprises the steps of Representing the sequence first order difference at time step t, , For the eigenvalues of the corresponding time positions, this operation is physically equivalent to an adaptive high-pass filter, eliminating low-frequency trend terms between adjacent time steps, then calculating the attention distribution based on the differential features, defining the query, key and value matrix, calculating the attention weighted result using the scaled dot product attention mechanism , Where d is the scale factor and where, Linear mapping the attention output to obtain attention characteristic , Wherein the method comprises the steps of Finally, carrying out residual connection on the differential feature and the attention feature, and inputting the residual connection to a latent feature recalibration function Obtaining final time domain branch output , Wherein, the Is a nonlinear recalibration function, which is specifically defined as, Wherein the method comprises the steps of The scaling vector is perceived for the channel, , For the affine transformation parameters, Is a learnable bias term; (4b) Cross-flow fusion and graph topology interaction Recalibrating the time domain And frequency domain quantization features Fusion is carried out to obtain comprehensive characterization containing multi-view information , Then, a local attention mechanism is introduced in graph structure sequence modeling, in order to accurately describe the dependency relationship between a node and the neighborhood thereof, firstly, the input node is subjected to self-adaptive transformation through a nonlinear recalibration strategy of a characteristic dimension level, and specifically, the representation of a certain node p is given The result of the latent feature recalibration is that, Wherein the method comprises the steps of The non-linear recalibration function defined in the step (4 a) is used for enhancing the discrimination of the node characteristics before the space interaction, and the node p and the neighbors thereof are based on the discrimination Is a local attention weight of (2) Expressed as a group of compounds, which are, Wherein the method comprises the steps of , To recalibrate the latent feature to the node feature, For a set of neighborhoods of node p, The function of the similarity is represented by a function of similarity, Learning bias for local attention, spatial attention features derived therefrom Each node feature it contains The formula is obtained by weighting aggregation calculation of the neighborhood nodes as follows, Wherein the method comprises the steps of Representation of And finally, carrying out feature aggregation on the multi-resolution graph structure constructed in the step (2 c) by utilizing a graph rolling network, Wherein the method comprises the steps of Representing a non-linear activation function, Is a normalized adjacency matrix, Is a learner-able weight for the stacking of the volumes, A graph convolution output representation at an mth resolution; (4c) Multi-scale aggregation and fault classification Introducing an average pooling strategy for all resolution levels Unified integration of the graph convolution outputs to obtain a global multi-scale representation , Wherein the method comprises the steps of Representing the fused global multi-scale feature representation, wherein m is the total layer level of the resolution map structure, and finally, the fused representation And inputting the full-connection layer and the Softmax classifier, and outputting a fault class classification result of the rolling bearing.

Description

Bearing fault diagnosis method based on cross-flow characteristic enhancement and graph topology interaction Technical Field The invention relates to the technical field of industrial equipment fault diagnosis, in particular to a bearing fault diagnosis method based on cross-flow characteristic enhancement and graph topology interaction, which is particularly suitable for a rolling bearing fault diagnosis scene. According to the invention, dynamic calibration of time-varying impact characteristics and topology construction of a multi-resolution graph are realized by constructing a dynamic nuclear evolution network, self-adaptive screening of a frequency spectrum sparse purification network to a key frequency band is designed, deep time-frequency space semantic alignment is realized by a heterogeneous perception cross-flow interaction mechanism, and the problems of difficult extraction of weak fault characteristics and low fault diagnosis accuracy caused by strong background noise interference, non-stable geometric distortion and characteristic heterogeneity easy loss characteristics of a rolling bearing vibration signal under a complex working condition are solved. Background With the development of the rotary machinery to large-scale and high-speed, the fault diagnosis technology of the rolling bearing has become a core means for guaranteeing the safe operation of equipment, but the traditional method has the remarkable limitations that the traditional fault diagnosis method of the rolling bearing usually relies on manual feature extraction and has the remarkable defects of time consumption and high dependence on expert experience, meanwhile, the traditional convolutional neural network uses static convolutional weights to lack the self-adaptive capturing capability of time domain geometric distortion in non-stationary signals, widely used statistical normalization strategies tend to smooth energy heterogeneity among feature channels, so that early weak fault features are extremely easy to be covered by low-frequency trends, and in addition, the traditional frequency domain analysis method lacks an active discrimination mechanism for spectrum significance, is difficult to accurately screen key fault frequency bands under strong background noise coupling, and restricts the discrimination capability of bearing fault diagnosis under complex working conditions. Aiming at the problems, the invention provides a bearing fault diagnosis method based on cross-flow characteristic enhancement and graph topology interaction, the invention realizes dynamic calibration of time-varying impact characteristics and multi-resolution graph topology construction by constructing a dynamic nuclear evolution network, designs a frequency spectrum sparse purification network to carry out self-adaptive screening and noise-resistant enhancement on key frequency bands, and finally provides a heterogeneous perception cross-flow interaction mechanism, and realizes the deep alignment of time-frequency space semantics through latent characteristic recalibration and graph topology reasoning, thereby remarkably improving fault diagnosis capability. Disclosure of Invention Aiming at the problems that the existing rolling bearing fault diagnosis method has the problems of inundation of weak features and insufficient frequency domain noise resistance caused by static feature extraction rigidification and normalization under the conditions of strong noise and non-stable conditions, the invention provides a bearing fault diagnosis method based on cross-flow feature enhancement and graph topology interaction. The specific implementation steps of the invention are as follows: 1. A rolling bearing fault experiment platform consisting of a servo motor, a driver and a manual hydraulic cylinder is built, 100N radial force is applied to a bearing through the manual hydraulic cylinder to simulate the load in an actual production environment, a vibration sensor is used for collecting original vibration signals of the bearing in inner ring faults, rolling body faults, compound faults and health states, the collected vibration signals are subjected to standardization processing by calculating mean values and standard deviations, and a continuous time domain signal is divided into sample sequences with fixed lengths by utilizing a sliding window technology. 2. The dynamic nuclear evolution network is constructed, the dynamic calibration of the time domain features and the topology construction of the multi-resolution map are realized, and the steps are carried out as follows: (2a) Generating dynamic calibration coefficients For input feature sequencesWherein B is the batch size, C is the number of characteristic channels, T is the sequence length, global and local context descriptors are extracted respectively, and first, global average pooling is performed on input features in the time dimension to obtain a global context descriptor g: Where T represents t