CN-121859201-B - Bearing fault diagnosis method based on time-frequency double-flow complementation and self-adaptive gating fusion

CN121859201BCN 121859201 BCN121859201 BCN 121859201BCN-121859201-B

Abstract

The invention discloses a bearing fault diagnosis method based on time-frequency double-flow complementation and self-adaptive gating fusion, which comprises the steps of collecting an original vibration signal of a bearing, segmenting an original vibration signal sequence into a plurality of sample fragments with fixed length to form an original data set, dividing the original data set into a training set and a testing set by adopting a physical segmentation strategy based on a time axis, constructing a time-frequency double-flow characteristic extraction module, extracting a time domain characteristic vector and a frequency domain characteristic vector, constructing a self-adaptive gating fusion module to dynamically generate gating weight, carrying out weighted complementation fusion on the time domain characteristic vector and the frequency domain characteristic vector to obtain a fused fault characteristic vector, constructing a classification decision network to map the fused fault characteristic vector into a decision space, optimizing model parameters by utilizing a label smoothing regularization loss function and a cosine annealing strategy, inputting the testing set into a trained fault diagnosis model, calculating posterior probability distribution of samples belonging to each fault class, and outputting a bearing fault diagnosis result.

Inventors

LI XU
LUAN FENG
LIU SHUHAO
WU YAN
HAN YUEJIAO
ZHANG CHI
SUN DI
XIONG WEI

Assignees

东北大学

Dates

Publication Date: 20260512
Application Date: 20260317

Claims (7)

1. A bearing fault diagnosis method based on time-frequency double-flow complementation and self-adaptive gating fusion is characterized by comprising the following steps: step 1, collecting an original vibration signal of a bearing, and dividing the original vibration signal sequence into a plurality of sample fragments with fixed lengths to form an original data set; Step 2, dividing an original data set into a training set and a testing set by adopting a physical segmentation strategy based on a time axis; step 3, constructing a time-frequency double-flow feature extraction module to extract time domain feature vectors and frequency domain feature vectors, wherein the time-frequency double-flow feature extraction module comprises heterogeneous parallel time domain feature extraction branches based on a bidirectional recursion scanning mechanism and frequency domain feature extraction branches based on transfer learning; Step 4, constructing a self-adaptive gating fusion module, dynamically generating gating weight according to the characteristic distribution of the input sample, and carrying out weighted complementary fusion on the time domain characteristic vector and the frequency domain characteristic vector to obtain a fusion fault characteristic vector; step 5, constructing a classification decision network, forming a fault diagnosis model with a time-frequency double-flow feature extraction module and a self-adaptive gating fusion module, mapping a fusion fault feature vector to a decision space through the classification decision network, and optimizing model parameters by using a label smoothing regularization loss function and a cosine annealing strategy until the model converges; Step 6, inputting the test set into a trained fault diagnosis model, calculating posterior probability distribution of the samples belonging to each fault category, and outputting a final bearing fault diagnosis result according to a maximum posterior probability criterion; The extracting the time domain feature vector and the frequency domain feature vector in the step 3 specifically comprises the following steps: Step 3.1, normalized sample fragment After being input into a time domain feature extraction branch based on a bidirectional recursion scanning mechanism, the following processing is carried out: Firstly, carrying out feature projection on an input signal through a one-dimensional convolution layer, mapping a time domain sampling point of a single channel into a high-dimensional potential feature vector sequence, and setting a vector of a t-th time step of the projected high-dimensional potential feature vector sequence as a vector of a t-th time step of the projected high-dimensional potential feature vector sequence ; Subsequently, the high-dimensional sequence of potential feature vectors is input into a stacked plurality of bi-directional recursive scanning units, in each of which the initial hidden state of the recursive process is to be determined Set to all zero vector for the first A time step, based on the hidden state of the previous moment And input features at the current time Calculating hidden state at current moment And output the characteristics The recursive equation of state transition and output is defined as: Wherein, the Represent the first Historical state information captured at a moment; the method comprises the steps of controlling the retention degree of historical information for a state transition parameter matrix capable of being learned; For inputting a control parameter matrix, controlling the blending proportion of the current input information; mapping the hidden state to an output feature for outputting the projection matrix; the bidirectional recursive scanning unit comprises a forward path and a backward path, and the forward path is indexed according to time To the point of Sequentially executing the above recursive equation and calculating the output characteristics at each moment Arranged in time sequence to obtain forward characteristic sequence Reverse order flipping of a high-dimensional latent feature vector sequence on a time axis in a backward path according to To the point of And (2) sequentially executing a recursive equation, and collecting output characteristics at each moment to obtain a backward characteristic sequence A hysteresis correlation for capturing faults; To fuse timing contexts in different directions, forward feature sequences are used And backward feature sequences Splicing the characteristic channel dimensions to form a dimension doubled combined characteristic sequence, compressing the combined characteristic sequence back to the original characteristic dimension by utilizing a linear projection network, introducing a residual error connection mechanism, and carrying out initial input sequence of the bidirectional recursion scanning unit Adding the two elements with the sequence subjected to linear projection to obtain an output sequence of the bidirectional recursion scanning unit The calculation process is expressed as follows: Wherein, the Representing a stitching operation along the channel dimension, And Respectively a linear projection matrix and a bias vector; a plurality of bidirectional recursion scanning units are stacked in series, and the total stacking layer number is set as In the first layer of the network, the initial input sequence is to perform characteristic projection on the input signal through a one-dimensional convolution layer, and the time domain sampling points of a single channel are mapped into a high-dimensional potential characteristic vector sequence, and in the subsequent hierarchical transmission, the first layer is The output sequence of the layer bidirectional recursion scanning unit is directly used as the first after being combined with root mean square normalization operation The input sequence of the layer bidirectional recursion scanning unit is processed by the end-to-end progressive mode Deep cascade extraction of layers, and finally the output sequence of one layer successfully aggregates the time sequence dependence of the global long distance and is recorded as a deep time domain sequence ; Finally, for converting sequence features into fixed length vectors which can be fused with frequency domain feature extraction branches based on transfer learning, deep time domain sequences Performing global maximum pooling operation on a time axis, performing dimension alignment through a nonlinear mapping network after redundant time sequence dimensions are eliminated by the pooling operation, and finally outputting a high-dimensional time domain global feature vector The calculation formula of the process is as follows: Wherein, the Represents a global maximum pooling operation performed along the time dimension, for extracting the maximum response value of the whole sequence over the various characteristic channels, Representing a composite function for feature mapping and dimension lifting, including linear mapping, layer normalization and activation operations; step 3.2 normalized vibration Signal After being input into a frequency domain feature extraction branch based on transfer learning, the following processing is carried out: First, short-time Fourier transform is utilized Transform to the frequency domain To enhance the contrast of weak fault characteristics, the background noise and fault impact texture are more obviously distinguished, and the transformed image is processed Carrying out logarithmic enhancement treatment, wherein the calculation formula is as follows: subsequently, the log spectrum is interpolated using bilinear interpolation Resampling to be a logarithmic spectrogram which accords with the standard size input by the visual model, copying the logarithmic spectrogram with the standard size into three parts, and stacking the logarithmic spectrogram into a three-channel pseudo-color image according to the channel dimension, so that the three-channel pseudo-color image simulates a natural image on a data format; Introducing EFFICIENTNET pre-trained on an ImageNet large-scale image dataset as a backbone network, inputting EFFICIENTNET three-channel pseudo-color images, utilizing the general texture extraction capability already learned in a deep convolution structure of the three-channel pseudo-color images, migrating the feature expression capability of the natural image field into a bearing fault diagnosis task, efficiently extracting periodic texture features reflecting fault impact in a spectrogram by freezing part of shallow layer parameters and fine-tuning deep layer parameters EFFICIENTNET, and finally outputting high-dimensional frequency domain feature vectors through a full-connection mapping layer 。
2. The method for diagnosing bearing faults based on time-frequency double-flow complementation and adaptive gating fusion according to claim 1, wherein the step1 is specifically: Step 1.1, collecting continuous vibration signals through an acceleration sensor arranged on a bearing seat of a rotary machine; step 1.2, setting the collected vibration signal sequence as follows: Wherein, the Setting the length of the sliding window as the total length of the vibration signal Step size of Splitting an original one-dimensional vibration signal sequence into a plurality of sample fragments with fixed lengths; step 1.3, constructing the original data set by a plurality of sample fragments with fixed lengths.
3. The method for diagnosing bearing faults based on time-frequency double-flow complementation and adaptive gating fusion according to claim 1, wherein the step 2 is specifically: step 2.1, dividing a plurality of sample fragments into a training set and a testing set by adopting a physical segmentation strategy based on a time axis; setting the training set dividing ratio as The time slicing points are as follows: The training set contains only historical data before the cut point: the test set contains only future data after the cut point: Wherein, the A t data point which is a vibration signal sequence; step 2.2, independent instance normalization processing is carried out on each sample segment after segmentation in the training set and the test set; For sample fragments Calculate the average value thereof And standard deviation : Wherein, the For sample fragments An ith data point of (b); Normalized sample fragment The method comprises the following steps: Wherein, the To prevent a small constant with zero denominator.
4. The bearing fault diagnosis method based on time-frequency double-flow complementation and self-adaptive gating fusion according to claim 3 is characterized in that the time domain feature extraction branch based on a bidirectional recursion scanning mechanism is used for capturing long-time sequence dynamic evolution features of signals and outputting time domain feature vectors rich in full life cycle dynamic information, and the frequency domain feature extraction branch based on migration learning is used for capturing high-dimensional frequency domain texture features of the signals and outputting high-dimensional frequency domain feature vectors.
5. The method for diagnosing bearing faults based on the fusion of time-frequency double-flow complementation and self-adaptive gating according to claim 1, wherein the step 4 is specifically: step 4.1, time domain feature vector And frequency domain feature vectors Splicing in channel dimension to construct joint feature space representation ; Step 4.2, constructing a gating sensing network comprising dimension reduction projection, layer normalization and random inactivation, which is used for learning the cross-modal interaction relation from the joint characteristics, and gating coefficients The generation process of (1) is defined as: Wherein, the Representing Sigmoid activation function, constraining output to Intervals to represent gating probabilities; And As a matrix of weights that can be learned, And Is that And pair of The corresponding bias is respectively responsible for feature compression and scalar mapping; the method is a layer normalization operation and is used for stabilizing gradient distribution; Is that A nonlinear activation function; representing bernoulli random inactivation operations for enhancing the generalized robustness of the gating network; step 4.3 based on the generated gating coefficients Performing weighted complementary fusion on the time domain feature vector and the frequency domain feature vector to obtain a final fused fault feature vector : Wherein, the Representing a term-wise multiplication.
6. The method for diagnosing bearing faults based on the fusion of time-frequency double-flow complementation and self-adaptive gating according to claim 1, wherein the step 5 is specifically: Step 5.1, constructing a classification decision network comprising a layer normalization, random inactivation and full connection mapping layer; step 5.2, inputting the fused fault feature vector into a classification decision network, and calculating to obtain the prediction probability distribution of the samples belonging to each fault class k after linear projection and Softmax normalization processing ; Step 5.3 setting The original single-hot coded real label is 1 if the sample belongs to the category k, otherwise, the sample is 0; As a total number of fault categories, As a smoothing factor, the reconstructed smoothed label distribution The definition is as follows: Based on this, a final optimized objective function is constructed I.e. minimizing predictive distribution And smooth the true distribution Kullback-Leibler divergence between: In the parameter updating stage of the fault diagnosis model, a AdamW optimizer with decoupling weight attenuation characteristic is adopted to solve the problem that the traditional Adam optimizer may converge to a local suboptimal solution on a non-convex optimization surface, and meanwhile, a cosine annealing learning rate scheduling strategy is introduced to dynamically adjust the learning rate in the training process The strategy enables the learning rate to be reduced along with the training round d according to a cosine function curve, and the formula is as follows: Wherein, the And The periodic learning rate adjustment mechanism can help the model to quickly pass through a flat area in the initial stage of training and finely converge to a global minimum point of a loss function in the later stage, thereby remarkably improving the convergence stability and the test precision of the final model.
7. The method for diagnosing bearing faults based on the fusion of time-frequency double-flow complementation and adaptive gating according to claim 1, wherein the step 6 is specifically: Step 6.1, inputting unknown vibration signal samples in the test set into a trained fault diagnosis model, and outputting posterior probability distribution vectors of the samples belonging to each fault category by the model through forward propagation calculation ; Step 6.2, following the maximum posterior probability criterion by And determining a final fault type judgment result by operation, thereby realizing end-to-end intelligent diagnosis.

Description

Bearing fault diagnosis method based on time-frequency double-flow complementation and self-adaptive gating fusion Technical Field The invention belongs to the technical field of intelligent fault diagnosis, and relates to a bearing fault diagnosis method based on time-frequency double-flow complementation and self-adaptive gating fusion. Background With the deep development of industrial internet and intelligent manufacturing, health status monitoring of rotating machinery (such as wind driven generators, aeroengines, industrial robots and the like) has become a core link for guaranteeing production safety and efficiency. The rolling bearing is used as a key part for bearing load and transmitting motion in rotary machinery, the running environment is extremely severe, and faults such as pitting corrosion, abrasion or fracture are extremely easy to occur. About 30% -40% of rotating machinery failures are statistically caused by bearing failure. Therefore, the method and the device accurately and efficiently diagnose and classify the bearing faults, and have important industrial values for reducing maintenance cost and preventing catastrophic accidents. Early bearing fault diagnosis mainly depends on signal processing technologies such as fast fourier transform, wavelet transform, empirical mode decomposition and the like, and although certain physical characteristics can be extracted by the methods, the methods depend on expert experience to carry out parameter selection seriously, and are difficult to adapt to non-stable and nonlinear complex industrial working conditions. In recent years, data driving methods typified by deep learning have become mainstream. The convolution neural network utilizes the convolution kernel to extract local characteristics, the long-term and short-term memory network utilizes the gating mechanism to process time sequence dependence, and the methods realize end-to-end intelligent diagnosis to a certain extent, thereby remarkably improving the automation level of fault identification. At present, a method based on a transducer architecture becomes a research hotspot in the field of bearing fault diagnosis due to strong global modeling capability. The transducer can capture long-distance dependence in signals by using a self-attention mechanism, and overcomes the defects of CNN receptive field limitation and low LSTM serial calculation efficiency. Meanwhile, partial researches are started to attempt to combine a time-frequency analysis technology, convert one-dimensional vibration signals into two-dimensional spectrograms, and utilize mature models in the field of computer vision (such as、) And performing transfer learning to solve the feature extraction problem under the condition of a small sample. However, the above-described methods still face many challenges in practical industrial applications. First, the mainstream model represented by the transducer is limited by the quadratic computational complexity #) When processing a long-sequence bearing signal with a high sampling rate, the calculation efficiency is low, the occupation of a video memory is large, and the real-time deployment requirement of an end side is difficult to meet. Secondly, the existing research is mostly limited to modeling in a single characteristic domain (only in a time domain or only in a frequency domain), and the phase evolution information of the time domain signal and the fine texture feature of the frequency domain spectrogram cannot be simultaneously considered, so that the information is not fully utilized. In addition, in the method related to multi-mode fusion, simple splicing or linear weighting is often adopted, and a gating mechanism capable of adaptively adjusting the mode weight according to sample characteristics is lacked, so that the diagnosis robustness and generalization capability of the model under the strong noise or complex mixed working condition are insufficient. Disclosure of Invention The invention aims to solve the problems of high computational load, incomplete single-mode feature extraction, lack of self-adaptability of a multi-mode fusion mechanism and the like of the conventional transducer architecture, and provides a bearing fault diagnosis method based on time-frequency double-flow complementation and self-adaptive gating fusion. The invention discloses a bearing fault diagnosis method based on time-frequency double-flow complementation and self-adaptive gating fusion, which comprises the following steps: step 1, collecting an original vibration signal of a bearing, and dividing the original vibration signal sequence into a plurality of sample fragments with fixed lengths to form an original data set; Step 2, dividing an original data set into a training set and a testing set by adopting a physical segmentation strategy based on a time axis; step 3, constructing a time-frequency double-flow feature extraction module, and extracting a time domain feature vector and a