CN-122020307-A - Mamba multi-mode fusion-based rolling bearing diagnosis system and method

CN122020307ACN 122020307 ACN122020307 ACN 122020307ACN-122020307-A

Abstract

The invention discloses a rolling bearing diagnosis system and a rolling bearing diagnosis method based on Mamba multi-mode fusion, wherein the rolling bearing diagnosis system comprises the following steps of S1, obtaining rolling bearing vibration data, preprocessing, outputting a data set, S2, establishing a rolling bearing diagnosis system based on Mamba multi-mode fusion, S3, processing a long-sequence vibration signal by using a time sequence branch, extracting nonlinear dynamic characteristics and global time sequence dependency, outputting time sequence characteristics, S4, processing a time-frequency image by using an image branch, extracting multi-dimensional time-frequency characteristics through bidirectional state space modeling of a time axis and a frequency axis, and S5, fusing the time sequence characteristics output by the time sequence branch and the time-frequency characteristics output by the image branch through a characteristic fusion module, and outputting a rolling bearing fault classification result. By adopting the system and the method, the efficiency and the accuracy of the fault diagnosis of the rolling bearing can be effectively improved, and a practical solution is provided for the real-time fault diagnosis of the bearing in an industrial environment.

Inventors

MA SHAOJUAN
SUN GUANGPENG
DING SHAOHU
CHANG PEIJU
XU CHANGLIN
XU XINYI

Assignees

北方民族大学

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (10)

1. A rolling bearing diagnosis system based on Mamba multi-mode fusion is characterized by comprising a time sequence branch, an image branch and a feature fusion module; The time sequence branches are used for processing the preprocessed long-sequence vibration signals, extracting nonlinear dynamic characteristics and global time sequence dependency relations in the signals and outputting time sequence characteristics; The image branches are used for processing the preprocessed time-frequency image converted by the vibration signal, and extracting multi-dimensional time-frequency characteristics through bidirectional state space modeling of a time axis and a frequency axis; the feature fusion module is used for receiving and adaptively fusing the time sequence feature output by the time sequence branch and the time frequency feature output by the image branch, and outputting a rolling bearing fault classification result.
2. The Mamba multi-modal fusion-based rolling bearing diagnosis system as claimed in claim 1, wherein the time sequence branch comprises an input normalization layer, a multi-scale convolution layer, a characteristic enhancement module, a state space scanning module and an output layer which are sequentially connected in series, wherein the input normalization layer receives a preprocessed long-sequence vibration signal and performs normalization processing to obtain initial characteristics and outputs the initial characteristics to the multi-scale convolution layer; The image branch comprises a downsampling layer, a biaxial state space block module and a global pooling layer; the system comprises a down sampling layer, a dual-axis state space block module, a global pooling layer, a multi-dimensional time-frequency module and a multi-dimensional time-frequency module, wherein the down sampling layer receives a preprocessed time-frequency image and performs down sampling processing and outputs processed features to the dual-axis state space block module; The dual-axis state space block module comprises TFSSM Block a, wherein the TFSSM Block a is connected with TFSSM Block a, and the TFSSM Block a is connected with TFSSM Block a 3 after performing depth separable downsampling, wherein the TFSSM Block a 3 is used for depth feature refinement, and the depth separable downsampling is used for reducing space dimension and reserving channel information; The feature fusion module comprises a low-rank projection layer, a cross attention calculation layer and an output layer, wherein the low-rank projection layer receives the time sequence features and the multi-dimensional time-frequency features, projection calculation complexity is reduced through low-rank parameterization processing, the processed features are output to the cross attention calculation layer, the cross attention calculation layer takes the multi-dimensional time-frequency features as inquiry, the time sequence features are keys and values, bimodal feature interaction calculation is completed, and finally rolling bearing fault category distribution is output through a classification head of the output layer after feature weighted fusion, normalization, amplitude cutting and random inactivation processing.
3. A rolling bearing diagnosis method based on Mamba multi-mode fusion is characterized by comprising the following steps: s1, acquiring vibration data of a rolling bearing, preprocessing the vibration data, and outputting a data set containing a long-sequence vibration signal and a time-frequency image; S2, establishing the Mamba multi-mode fusion-based rolling bearing diagnosis system according to any one of claims 1-2; s3, using a time sequence branch to process the preprocessed long-sequence vibration signal, extracting nonlinear dynamic characteristics and global time sequence dependency in the signal, and outputting time sequence characteristics; s4, extracting multi-dimensional time-frequency characteristics by using a time-frequency image converted by the vibration signal after image branch processing pretreatment through bidirectional state space modeling of a time axis and a frequency axis; S5, receiving and adaptively fusing the time sequence characteristics output by the time sequence branches and the time frequency characteristics output by the image branches through a characteristic fusion module, and outputting a rolling bearing fault classification result.
4. A method of diagnosing a rolling bearing based on Mamba multi-modal fusion as claimed in claim 3 wherein preprocessing the vibration data in S1 includes: s11, dividing a data set according to a preset rule on the original vibration signal level, and distributing vibration signals corresponding to different working conditions or different acquisition sources to a training set, a verification set and a test set which are not overlapped with each other; S12, performing length normalization processing on each original vibration signal, and performing normalization processing on each data subset; s13, after signal level division is completed, segmenting the normalized long-sequence vibration signal by adopting a segmentation strategy, and constructing a time sequence sample with fixed length; s14, generating a time-frequency image based on the time sequence sample through a time-frequency conversion algorithm, and taking the time-frequency image as input data of a time-frequency image branch; and S15, respectively using a training set, a verification set and a test set which comprise long-sequence vibration signals and time-frequency images to train the system learnable parameters in the S2, verify the validity of the super-parameter tuning and model in the training process, and evaluate the final fault classification precision and generalization capability.
5. The Mamba multi-modal fusion-based rolling bearing diagnostic method as claimed in claim 3, wherein S3 specifically includes: S31, performing normalization processing on the preprocessed long-sequence vibration signals through an input normalization layer of the time sequence branch, and transmitting the processed signals into a multi-scale convolution layer; S32, performing multi-scale parallel convolution on the output result of the S31 through a multi-scale convolution layer, and capturing a signal local transient mode; s33, performing feature enhancement processing on the output result of the S32 by using a feature enhancement module, and extracting nonlinear dynamic features of the signals; s34, executing state space scanning on the output result of the S33 through a state space scanning module, realizing global time sequence dependency modeling, and finally outputting time sequence characteristics through an output layer.
6. The method for diagnosing a rolling bearing based on Mamba multi-modal fusion as set forth in claim 5, wherein S32 specifically includes: the multi-scale convolution layer adopts a mode of combining parallel multi-scale depth convolution and 1 multiplied by 1 point convolution, a plurality of convolution kernels with different kernel lengths are selected to execute the depth convolution in parallel, and the depth convolution output characteristics corresponding to different scales are obtained, wherein the formula is as follows: ; ; Wherein, the For the length of the convolution kernel, The normalized long-sequence vibration signal characteristics output by the step S31, For a single scale depth convolution output feature, In order for the deep convolution operation to be performed, Channel stitching operations for different scale output features, A1 x 1 point convolution operation is shown, Is a multi-scale convolution fusion feature, as input to S33.
7. The method for diagnosing a rolling bearing based on Mamba multi-modal fusion as recited in claim 5, wherein S33 specifically includes: S331, given delay order Sum step size Construction of delay stacks : ; Wherein, the For the moment of time Is ensured by forward zero padding; S332, calculating a variation measure, and constructing an attractor gating signal through adjacent frame difference The formula is: ; ; Wherein, the For the moment of time Is a measure of variation of (a) in (b), Is that The function is activated and the function is activated, In order for the coefficient to be a coefficient that can be learned, To learn the bias, initially set to a negative value, 、 A learnable parameter that is a feature enhancement module; s333 by embedding a function Will delay the coordinate vector Mapping to multi-scale convolution fusion features Then through the attractor gating signal Controlling the injection intensity of the mapping result, and combining the characteristics after injection with the original time Features of (2) Superposition to obtain enhanced features The formula is: 。
8. the method for diagnosing a rolling bearing based on Mamba multi-modal fusion as set forth in claim 5, wherein S34 specifically includes: selective state space scanning is realized by adopting stable parameterization and logarithmic domain centralization, diagonal stable parameterization is respectively carried out on each hidden channel, and diagonal stable state matrixes with characteristic values of which real parts are negative are constructed The formula is: ; Wherein, the Is a log domain stability parameter; Generating non-negative bounded step sizes The step size parameter Warp yarn Or (b) After the function transformation, the state is cut off to be within a section formed by 0 and a preset maximum step length, and the state update adopts a form of centralizing prefix sum of a logarithmic domain, adopts prefix sum operation to solve the original accumulated state parameters The formula is: ; for the original accumulated state parameters Performing a centering process to obtain a centralized accumulated state parameter The formula is: ; Based on Computing an log domain rescaling factor The formula is: ; In combination with input drive signals And state weight factor By calculating hidden states by element-by-element products The formula is: ; Wherein B is a read-write parameter; hidden state Reading status information via read-write parameter C, in combination with jump Supplementing input drive features, synchronizing input drive signals Hidden state State space scan results Amplitude clipping is carried out, and finally, a state space scanning result is output The formula is: ; state space scan results for all moments Performing time sequence dimension integration, and outputting robust time sequence characteristics after processing by an output layer 。
9. The method for diagnosing a rolling bearing based on Mamba multi-modal fusion as set forth in claim 3, wherein S4 specifically includes: S41, inputting a time-frequency image through a downsampling layer of an image branch Performing a downsampling Stem operation consisting of three-step convolution to output downsampled features ; S42, sequentially executing local axial coding, stable axial recursion, structured channel mixing, normalization and residual fusion through a biaxial state space block module of an image branch; The stable axial recursion independently develops state space recursion modeling along a frequency axis and a time axis respectively, and specifically comprises the following steps: For frequency axis recursion, a frequency axis state vector is defined The frequency axis state vector The initial value of (c) is set to the zero vector, the formula is: ; Wherein, the , Is the state vector dimension; Based on state updating rule of exponential decay mechanism, calculating current time frequency axis state vector by combining step length parameter, input characteristic and projection parameter The formula is: ; Wherein, the Downsampling the output characteristics of the Stem operation for S41; Obtaining frequency axis recursion initial output through inner product operation of current moment state vector and reading parameter The formula is: ; Performing numerical clipping on the initial output and superposing the product term of the input characteristic and the bias parameter to obtain the final frequency axis recursion characteristic The formula is: ; Wherein, the Is a bias parameter; For timeline recursion, a timeline state vector is defined The time axis state vector The initial value of (c) is set to the zero vector, the formula is: ; Wherein, the ; Based on state updating rule of exponential decay mechanism, calculating time axis state vector at current moment by combining step length parameter, input characteristic and projection parameter The formula is: ; obtaining initial output of time axis recursion through inner product operation of current moment state vector and reading parameter The formula is: ; The product item of the input characteristic and the bias parameter is overlapped after the numerical clipping operation is carried out on the initial output, and the final time axis recursion characteristic is obtained The formula is: ; Defining diagonal stability matrices respectively Step size parameter Inputting projection parameters Reading parameters The formula is: ; ; ; ; Wherein, the , Is a threshold value for the stability of the digital value, In order to scale the coefficient of the power consumption, In order to cut the upper and lower threshold values, As a minimum threshold value for the step size, As the step size maximum threshold value, For the log-domain step size parameter, 、、 Respectively corresponding to step length parameters Inputting projection parameters Reading parameters Is a convolution extraction operation of (1); S43, realizing channel interaction through approximate butterfly structured channel mixing, and executing global average pooling by adopting a global pooling layer of image branches to obtain multi-dimensional time-frequency characteristics of global space representation.
10. The method for diagnosing a rolling bearing based on Mamba multi-modal fusion as set forth in claim 3, wherein S5 specifically includes: S51, inquiring a matrix through a low-rank projection layer of the feature fusion module Key matrix Matrix of values Linear projection with low rank parameterization, maximum rank Representing the projection matrix as a decomposition base And Is introduced into the product form of the learning weights Treatment with softmax function Obtain the weight Based on Constructing diagonal matrix Dynamically weighting the decomposition base, wherein the formula is as follows: ; ; ; S52, for each attention head, respectively inquiring the matrix Key matrix Matrix of values Corresponding structure decomposition base 、、 Processing all decomposition bases based on the dynamic low-rank weighting rule described in S51 to obtain query projections Key projection Projection of values ; S53, the cross attention calculation layer through the feature fusion module uses time-frequency features As a query, timing characteristics As key and value, calculating multi-head cross attention to obtain attention output result The calculation formula is as follows: ; Wherein, the Hiding dimensions for attention; s54, based on time-frequency characteristics Time series characteristic mean value and attention output Calculating gating coefficients by a multi-layer perceptron And mapped by an activation function and then passed through a gating coefficient Time-frequency characteristics Output result of attention Weighting fusion is carried out to obtain fusion characteristics The formula is: ; ; Wherein, the Is a time sequence feature Calculating a time sequence characteristic mean value obtained by the global mean value; s55, sequentially performing normalization, amplitude clipping and random inactivation treatment on the fusion characteristics, and outputting fault category distribution through a classification head of an output layer.

Description

Mamba multi-mode fusion-based rolling bearing diagnosis system and method Technical Field The invention relates to the technical field of rolling bearing fault diagnosis, in particular to a rolling bearing diagnosis system and method based on Mamba multi-mode fusion. Background The health status of the rolling bearing is directly related to the safety and usability of high-speed rotating equipment such as aircraft engines, wind generators, high-speed trains and the like. Once the fault detection is delayed or misdiagnosed, unplanned shutdown and high maintenance cost are easy to cause, so that real-time, accurate and robust diagnosis of long-term vibration signals is realized under complex working conditions, and the problem of leading-edge engineering to be broken through is still urgent. The existing diagnosis method is mainly divided into two types, namely, a method for extracting physical characteristics based on professional knowledge in the field of dependence and is easy to limit in a scene of variable working conditions or unknown fault modes, and a method for judging characteristics by means of automatic learning of machine learning, so that dependence on professional knowledge is reduced, and a remarkable short-board technology still exists. In the deep learning field, although models such as CNN, LSTM and the like represent inauguration under specific scenes, a transducer architecture has O (L 2) secondary complexity, calculation burden presents a multiplication effect in a multi-mode scene, millisecond-level real-time response requirements are difficult to meet, the problems of structural isomerism, asynchronous sampling, signal-to-noise ratio difference and the like of multi-mode data generally exist, the existing fixed weight fusion strategy is easy to be interfered by low-quality modes, fusion efficiency and reliability are insufficient, meanwhile, the weak nonlinear characteristics of early faults are difficult to capture by a traditional linear method, weak fault detection sensitivity is insufficient, and the diagnosis performance of the method in industrial scenes is severely restricted. Therefore, an innovative method is needed to break through the bottleneck of the prior art, synchronously improve the diagnosis precision and calculation efficiency of the model, and provide a practical and feasible solution for real-time fault diagnosis of the rolling bearing in an industrial scene. Disclosure of Invention The invention aims to provide a rolling bearing diagnosis system and a rolling bearing diagnosis method based on Mamba multi-mode fusion, which combine a Mamba model with linear complexity, self-adaptive multi-mode fusion and chaos enhancement, can effectively improve the efficiency and accuracy of rolling bearing fault diagnosis, realize millisecond-level real-time reasoning while maintaining high precision, and provide a practical solution for real-time bearing fault diagnosis in an industrial environment. In order to achieve the above object, the present invention provides a rolling bearing diagnosis system based on Mamba multi-mode fusion, including a time sequence branch, an image branch and a feature fusion module; The time sequence branches are used for processing the preprocessed long-sequence vibration signals, extracting nonlinear dynamic characteristics and global time sequence dependency relations in the signals and outputting time sequence characteristics; The image branches are used for processing the preprocessed time-frequency image converted by the vibration signal, and extracting multi-dimensional time-frequency characteristics through bidirectional state space modeling of a time axis and a frequency axis; the feature fusion module is used for receiving and adaptively fusing the time sequence feature output by the time sequence branch and the time frequency feature output by the image branch, and outputting a rolling bearing fault classification result. Also provided is a rolling bearing diagnosis method based on Mamba multi-mode fusion, comprising the steps of: s1, acquiring vibration data of a rolling bearing, preprocessing the vibration data, and outputting a data set containing a long-sequence vibration signal and a time-frequency image; S2, establishing a Mamba multi-mode fusion-based rolling bearing diagnosis system; s3, using a time sequence branch to process the preprocessed long-sequence vibration signal, extracting nonlinear dynamic characteristics and global time sequence dependency in the signal, and outputting time sequence characteristics; s4, extracting multi-dimensional time-frequency characteristics by using a time-frequency image converted by the vibration signal after image branch processing pretreatment through bidirectional state space modeling of a time axis and a frequency axis; S5, receiving and adaptively fusing the time sequence characteristics output by the time sequence branches and the time frequency characteristics output by the image bran