CN-121994479-A - Mechanical fault detection method based on WFEConvformer model

CN121994479ACN 121994479 ACN121994479 ACN 121994479ACN-121994479-A

Abstract

The invention relates to the field of mechanical fault detection, and discloses a mechanical fault detection method based on WFEConvformer models, which solves the problems of high model calculation complexity, weak multi-scale and global feature capturing capability and insufficient feature differentiation in the existing mechanical fault detection method and meets the real-time and accuracy requirements of industrial equipment on fault diagnosis. The method comprises the steps of firstly collecting multichannel vibration signals of a mechanical system, dividing the multichannel vibration signals into a sample set through a sliding window, converting a time domain vibration signal into a multi-scale time-frequency characteristic matrix through wavelet packet transformation, combining fuzzy entropy quantization characteristic complexity to realize cooperative enhancement of time-frequency information and complexity characteristics, inputting the enhancement characteristics into a WFEConvformer model, outputting fault types of the mechanical system to realize accurate detection of the mechanical system, and balancing detection precision and calculation efficiency through enhanced fault characteristic distinction and combining a lightweight frame, so that high robustness is maintained under a noise environment, and real-time and reliability requirements of mechanical system fault detection can be effectively met.

Inventors

DAI HUAXIANG
HU TAO
FU HONGTAO
LIN ZIJI
XU CHAOXIA
XU QIANJING
ZHAO BINGKUN

Assignees

泸州老窖股份有限公司

Dates

Publication Date: 20260508
Application Date: 20260120

Claims (10)

1. The mechanical fault detection method based on WFEConvformer model is characterized by comprising the following steps: s1, collecting multichannel vibration signals of a mechanical system, and intercepting the vibration signals by adopting a sliding window to generate a sample data set; S2, performing feature enhancement processing on the sample data set, namely converting a multichannel time domain signal of a sample into a time-frequency feature matrix through wavelet packet transformation, calculating a complexity feature matrix of the sample through a fuzzy entropy algorithm, and splicing the time-frequency feature matrix and the complexity feature matrix along a channel dimension to obtain enhancement features; S3, carrying out normalization processing on the enhanced features to obtain training data; S4, constructing WFEConvformer a model, wherein the model comprises an input layer, a lightweight characteristic extraction layer and an output layer; the input layer is used for receiving input characteristics, completing time sequence dimension optimization, channel adaptation and local information preliminary integration, and outputting standardized characteristics processed by an adaptation follow-up module; The light-weight feature extraction layer is formed by connecting a separable multi-scale convolution module and a broadcast self-attention module in series; the separable multi-scale convolution module firstly executes cross-channel convolution through a 1X 1 convolution check standardized feature, then adopts a plurality of different scales of separation convolution checks to respectively process the cross-channel convolved features, and finally splices all scales of convolution output to obtain multi-scale local features; The broadcast self-attention module takes the multi-scale local feature as input, firstly applies three groups of linear mapping to generate a time score vector, a key matrix and a value matrix, then carries out weighted aggregation on the feature of the key matrix through the time score vector to obtain a channel feature weight vector, broadcasts the channel feature weight vector back to the value matrix, carries out weighted on the time feature, and obtains an enhanced feature matrix through linear mapping; The enhanced feature matrix performs feature normalization and expression enhancement through a layer normalization and lightweight feedforward network, and finally obtains fusion features fusing local and global information as the output of a lightweight feature extraction layer; the output layer is used for mapping the fusion characteristics output by the lightweight characteristic extraction layer into probability distribution of each health state category of the mechanical system; S5, training the WFEConvformer model based on the training data to obtain a WFEConvformer model after training; s6, in practical application, multichannel vibration signals of the mechanical system are collected in real time, and after the multichannel vibration signals are intercepted through a sliding window, subjected to characteristic enhancement processing and normalized processing, a WFEConvformer model after training is input, and a fault detection result is obtained.
2. The method for mechanical failure detection based on WFEConvformer model as claimed in claim 1, wherein, In step S1, the method for intercepting the vibration signal by adopting the sliding window comprises the step of dividing the acquired original multichannel vibration signal into independent sample fragments with equal length according to the fixed window length and the step length.
3. The method for mechanical failure detection based on WFEConvformer model as claimed in claim 1, wherein, In step S2, the converting the multichannel time domain signal of the sample into the time-frequency feature matrix through wavelet packet transformation includes: Carrying out 2-layer wavelet packet decomposition on multi-channel time domain signals of a sample one by one according to channels, decomposing each channel signal into 4 sub-band signals with different frequency scales, extracting coefficients of each sub-band signal as characteristic channels, and splicing sub-band characteristic channels corresponding to all channels along the channel dimension to form a time-frequency characteristic matrix, wherein the channel number of the time-frequency characteristic matrix is the product of the original signal channel number and the sub-band signal number, and the length of the time-frequency characteristic matrix is consistent with that of the original time domain signal.
4. The method for mechanical failure detection based on WFEConvformer model as claimed in claim 3, wherein, In step S2, the calculating the complexity characteristic matrix of the sample by the fuzzy entropy algorithm includes: Selecting a first channel of an original time domain signal as fuzzy entropy calculation input, constructing delay vectors with embedded dimensions of 2, calculating the maximum distance between any two delay vectors, judging similarity, calculating a fuzzy entropy value based on a similarity statistical result, expanding the fuzzy entropy value into a one-dimensional feature vector with the length consistent with that of the original time domain signal through broadcasting operation, and obtaining a complexity feature matrix; the fuzzy entropy value is calculated in the following way: ; ; Wherein, the Representing the fuzzy entropy value; Representing the signal length; Representing an embedding dimension; Representing the maximum distance between delay vectors; Representing similar tolerances; Representing the standard deviation of the original time domain signal.
5. The method for mechanical failure detection based on WFEConvformer model as claimed in claim 1, wherein, In step S3, normalizing the enhancement features includes: ; Wherein, the Representing one of the enhanced features; Representing the normalized features; Representing characteristics Is the minimum of (2); Representing characteristics Is the maximum value of (2); a separable multi-scale convolution module performs cross-channel convolution on input features by a1 x 1 convolution kernel, comprising: ; Wherein, the Features obtained after cross-channel convolution; is an input feature; weights that are cross-channel convolution kernels; The range of values for the channel index of the input feature is , The original channel number that is the input feature; for outputting channel index, the value range is , The number of output channels after convolution operation; the value range is that for time sequence position index , Signal length, which is the input feature.
6. The method for mechanical failure detection based on WFEConvformer model, as set forth in claim 5, The cross-channel convolved features are respectively processed by adopting a plurality of different-scale separate convolution cores, which comprises the following steps: ; Wherein, the Is the first Separate convolved output signals of individual scales are in the channel Time sequence position Is a characteristic value of (2); , Representing the length of the convolved signal; Represent the first The weights of the layer convolution kernels; In-channel for cross-channel convolved signals Time sequence position Is a characteristic value of (a).
7. The method for mechanical failure detection based on WFEConvformer model as claimed in claim 6, wherein, In step S4, taking the multi-scale local feature as an input, three sets of linear mappings are applied thereto to generate a time score vector, a key matrix, and a value matrix, including: ; ; ; Wherein, the Representing a time score vector; representing a key matrix; Representing a matrix of values; 、、 Is a weight matrix; For the input multi-scale local features, For the length of the sequence, Is a feature dimension.
8. The method for mechanical failure detection based on WFEConvformer model as claimed in claim 7, In step S4, the weighting and aggregation are performed on the features of the key matrix through the time score vector to obtain a channel feature weight vector, which includes: ; Wherein, the Representing channel characteristic weight vectors, reflecting the importance of each channel in the global view; Represent the first Time scores for individual sequence positions; Represent the first A key vector of individual sequence positions; Representing a broadcast operation.
9. The method for mechanical failure detection based on WFEConvformer model as claimed in claim 8, wherein, In step S4, broadcasting the channel feature weight vector back to the value matrix and weighting the time sequence feature, obtaining the enhanced feature matrix through linear mapping, including: ; Wherein, the Representing an enhanced feature matrix; Represents the first A matrix of values for the individual channels; Represent the first Channel feature weights for the individual channels; representing the output linear mapping weights.
10. The mechanical fault detection method based on WFEConvformer models as claimed in any one of claims 1 to 9, wherein in step S4, mapping the fusion features output by the lightweight feature extraction layer to probability distributions of each health state class of the mechanical system includes: Firstly, carrying out global average pooling on the fusion features in the sequence dimension to obtain a global feature vector: ; Wherein, the Representing a global feature vector; representing a time sequence length of the fusion feature; Representing a fusion feature; And projecting the global feature vector to a preset fault class space through linear mapping: ; Wherein, the Representing the linearly mapped output vector; Representing an output layer weight matrix; representing the offset vector, The number of fault categories; Finally, normalizing the linear mapping result into probability distribution through a softMax function: ; wherein, the Indicating that the device belongs to the first Probability of class fault type; Represent the first A class fault score; Represent the first A score for a class fault, Is the number of fault categories.

Description

Mechanical fault detection method based on WFEConvformer model Technical Field The invention relates to the field of mechanical fault detection, in particular to a mechanical fault detection method based on WFEConvformer models. Background The mechanical fault detection is a critical ring in industrial production, potential faults of equipment can be timely and accurately identified, production interruption and equipment damage can be effectively avoided, operation and maintenance cost is reduced, and production efficiency is improved. The traditional mechanical fault detection methods are mostly dependent on empirical rules and signal processing-based technologies, and the methods need to manually extract fault characteristics, so that the efficiency is low, and complex and changeable fault modes are difficult to adapt. With the development of industrial internet and intelligent manufacturing, a machine learning method based on data driving is gradually becoming a research hotspot in the field of fault detection. By learning the characteristics from the equipment operation data, the method can be better suitable for different working conditions and fault modes, and has stronger generalization capability and automation characteristic. In particular, deep learning techniques, by virtue of their powerful feature extraction capability and nonlinear modeling capability, have achieved significant results in many fault detection applications. However, some problems still remain with existing deep learning methods: Firstly, most deep learning models have complex structures, huge parameter scale and high calculation complexity, are difficult to be deployed in the edge equipment with limited resources in the industrial scene, and cannot meet the real-time fault detection requirement. Secondly, the mechanical equipment fault signal has multi-scale characteristics, not only comprises high-frequency vibration impact characteristics, but also covers low-frequency trend change information, the traditional convolutional neural network is focused on local feature extraction, and the multi-scale characteristics are not effectively fused, so that the recognition capability of a complex fault mode is insufficient. Third, the feature distribution of the partial fault mode has global property, the change of the local feature may not be obvious, but the traditional fault detection method relies on the learning of the local feature, and is difficult to capture the intrinsic law of the fault. The WFEConvformer model is used as a novel fault detection model, integrates the characteristic enhancement technologies such as wavelet packet transformation and fuzzy entropy and a lightweight network structure, has the potential of multi-scale characteristic extraction and global and local information fusion, but still has the problems of low training efficiency, insufficient characteristic suitability and the like under a super-large-scale data set or an extremely complex fault mode, and needs to be further optimized to meet the high requirements of industrial scenes. Disclosure of Invention The invention aims to solve the technical problems of high model calculation complexity, weak multi-scale and global feature capturing capability and insufficient feature differentiation in the existing mechanical fault detection method, and meets the requirements of industrial equipment on real-time performance and accuracy of fault diagnosis. The technical scheme adopted for solving the technical problems is as follows: The mechanical fault detection method based on WFEConvformer model comprises the following steps: s1, collecting multichannel vibration signals of a mechanical system, and intercepting the vibration signals by adopting a sliding window to generate a sample data set; S2, performing feature enhancement processing on the sample data set, namely converting a multichannel time domain signal of a sample into a time-frequency feature matrix through wavelet packet transformation, calculating a complexity feature matrix of the sample through a fuzzy entropy algorithm, and splicing the time-frequency feature matrix and the complexity feature matrix along a channel dimension to obtain enhancement features; S3, carrying out normalization processing on the enhanced features to obtain training data; S4, constructing WFEConvformer a model, wherein the model comprises an input layer, a lightweight characteristic extraction layer and an output layer; the input layer is used for receiving input characteristics, completing time sequence dimension optimization, channel adaptation and local information preliminary integration, and outputting standardized characteristics processed by an adaptation follow-up module; The light-weight feature extraction layer is formed by connecting a separable multi-scale convolution module and a broadcast self-attention module in series; the separable multi-scale convolution module firstly executes cross-channel convolutio