Mechanical fault detection method based on WFEConvformer model
By enhancing features through wavelet packet transform and fuzzy entropy algorithm, and combining a lightweight WFEConvformer model with separable multi-scale convolution and broadcast self-attention module, the problems of high computational complexity and weak feature capture ability of existing mechanical fault detection methods are solved, and efficient and accurate fault detection is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LUZHOU LAOJIAO CO LTD
- Filing Date
- 2026-01-20
- Publication Date
- 2026-05-08
AI Technical Summary
Existing mechanical fault detection methods suffer from high model computational complexity, weak multi-scale and global feature capture capabilities, and insufficient feature discrimination, making it difficult to meet the real-time and accuracy requirements of industrial equipment for fault diagnosis.
We employ wavelet packet transform and fuzzy entropy algorithms for feature enhancement, and combine them with lightweight separable multi-scale convolutional modules and broadcast self-attention modules to construct a WFEConvformer model. This enables multi-scale feature extraction and global information fusion, reducing computational complexity and improving feature discriminative power.
It achieves high-precision and robust mechanical fault detection with low resource consumption, can be deployed in real time on edge devices, and can adapt to fault identification under complex working conditions.
Smart Images
Figure CN121994479A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mechanical fault detection, and specifically to a mechanical fault detection method based on the WFEConvformer model. Background Technology
[0002] Mechanical fault detection is a crucial part of industrial production. By identifying potential equipment faults in a timely and accurate manner, production interruptions and equipment damage can be effectively avoided, maintenance costs can be reduced, and production efficiency can be improved. Traditional mechanical fault detection methods mostly rely on empirical rules and signal processing-based techniques. These methods require manual extraction of fault features, which is not only inefficient but also difficult to adapt to complex and ever-changing fault modes.
[0003] With the development of the Industrial Internet and intelligent manufacturing, data-driven machine learning methods have gradually become a research hotspot in the field of fault detection. These methods, by learning features from equipment operating data, can better adapt to different operating conditions and fault modes, exhibiting strong generalization capabilities and automation characteristics. Deep learning technology, in particular, has achieved significant results in many fault detection applications due to its powerful feature extraction and nonlinear modeling capabilities.
[0004] However, existing deep learning methods still have some problems:
[0005] First, most deep learning models have complex structures, large parameter scales, and high computational complexity, making them difficult to deploy on edge devices with limited resources in industrial scenarios and unable to meet the needs of real-time fault detection.
[0006] Secondly, mechanical equipment fault signals have multi-scale characteristics, including both high-frequency vibration and impact features and low-frequency trend change information. Traditional convolutional neural networks focus more on local feature extraction and lack effective fusion of multi-scale features, resulting in insufficient ability to identify complex fault modes.
[0007] Third, the feature distribution of some fault modes is global, and the changes in local features may not be significant. Traditional fault detection methods rely on learning local features, making it difficult to capture the essential laws of such faults.
[0008] The WFEConvformer model, as a novel fault detection model, integrates feature enhancement techniques such as wavelet packet transform and fuzzy entropy with a lightweight network structure. It has the potential for multi-scale feature extraction and fusion of global and local information. However, it still suffers from problems such as low training efficiency and insufficient feature adaptability in ultra-large-scale datasets or extremely complex fault modes. Further optimization is needed to meet the high requirements of industrial scenarios. Summary of the Invention
[0009] The technical problem to be solved by this invention is to provide a mechanical fault detection method based on the WFEConvformer model, which solves the problems of high model computational complexity, weak multi-scale and global feature capture ability, and insufficient feature discrimination in existing mechanical fault detection methods, and meets the real-time and accuracy requirements of industrial equipment for fault diagnosis.
[0010] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:
[0011] The mechanical fault detection method based on the WFEConvformer model includes the following steps:
[0012] S1. Collect multi-channel vibration signals from the mechanical system, use a sliding window to extract the vibration signals, and generate a sample dataset;
[0013] S2. Perform feature enhancement processing on the sample dataset: convert the multi-channel time-domain signal of the sample into a time-frequency feature matrix through wavelet packet transform, calculate the complexity feature matrix of the sample through fuzzy entropy algorithm, and concatenate the time-frequency feature matrix and the complexity feature matrix along the channel dimension to obtain enhanced features;
[0014] S3. Normalize the enhanced features to obtain training data;
[0015] S4. Construct the WFEConvformer model, which includes an input layer, a lightweight feature extraction layer, and an output layer;
[0016] The input layer is used to receive input features, complete temporal dimension optimization, channel adaptation and preliminary integration of local information, and output standardized features adapted to subsequent module processing.
[0017] The lightweight feature extraction layer is composed of a separable multi-scale convolutional module and a broadcast self-attention module connected in series.
[0018] Among them, the separable multi-scale convolution module first performs cross-channel convolution on the standardized features through a 1×1 convolution kernel, then uses multiple separable convolution kernels of different scales to process the features after cross-channel convolution separately, and finally concatenates the convolution outputs of all scales to obtain multi-scale local features.
[0019] The broadcast self-attention module takes the multi-scale local features as input, first applies three sets of linear mappings to them to generate a time score vector, a key matrix and a value matrix, then uses the time score vector to weight and aggregate the features of the key matrix to obtain a channel feature weight vector, broadcasts the channel feature weight vector back to the value matrix and weights the temporal features, and obtains the enhanced feature matrix through linear mapping;
[0020] The enhanced feature matrix is then normalized and enhanced through layer normalization and a lightweight feedforward network to obtain a fused feature that integrates local and global information, which serves as the output of the lightweight feature extraction layer.
[0021] The output layer is used to map the fused features output by the lightweight feature extraction layer to the probability distribution of each health state category of the mechanical system;
[0022] S5. Train the WFEConvformer model based on the training data to obtain the trained WFEConvformer model;
[0023] S6. In practical applications, multi-channel vibration signals of the mechanical system are acquired in real time, and after sliding window truncation, feature enhancement processing and normalization processing, they are input into the trained WFEConvformer model to obtain fault detection results.
[0024] Furthermore, in step S1, the method of using a sliding window to extract the vibration signal includes: dividing the acquired original multi-channel vibration signal into independent, equal-length sample segments according to a fixed window length and step size.
[0025] Furthermore, in step S2, the conversion of the multi-channel time-domain signal of the sample into a time-frequency feature matrix through wavelet packet transform includes:
[0026] The multi-channel time-domain signal of the sample is decomposed into two-level wavelet packet decomposition for each channel. Each channel signal is decomposed into four sub-band signals with different frequency scales. The coefficients of each sub-band signal are extracted as feature channels. The sub-band feature channels corresponding to all channels are concatenated along the channel dimension to form a time-frequency feature matrix. The number of channels of the time-frequency feature matrix is the product of the number of channels of the original signal and the number of sub-band signals, and the length is consistent with that of the original time-domain signal.
[0027] Furthermore, in step S2, calculating the complexity feature matrix of the sample using the fuzzy entropy algorithm includes:
[0028] The first channel of the original time-domain signal is selected as the input for fuzzy entropy calculation. A delay vector with an embedding dimension of 2 is constructed. The maximum distance between any two delay vectors is calculated and their similarity is judged. The fuzzy entropy value is calculated based on the similarity statistics. The fuzzy entropy value is expanded into a one-dimensional feature vector with the same length as the original time-domain signal through a broadcast operation to obtain the complexity feature matrix.
[0029] The fuzzy entropy value is calculated as follows:
[0030] ;
[0031] ;
[0032] in, Represents the fuzzy entropy value; Indicates the signal length; Indicates the embedding dimension; This represents the maximum distance between delay vectors; Indicates similarity tolerance; It represents the standard deviation of the original time-domain signal.
[0033] Furthermore, in step S3, the enhanced features are normalized, including:
[0034] ;
[0035] in, This represents one of the enhanced features; Represents the normalized features; Representation of features The minimum value; Representation of features The maximum value.
[0036] Furthermore, in step S4, the separable multi-scale convolution module first performs cross-channel convolution on the input features using a 1×1 convolution kernel, including:
[0037] ;
[0038] in, Features obtained after cross-channel convolution; Input features; These are the weights of the cross-channel convolution kernel; The channel index of the input feature, with a value range of 1. , This is the original number of channels for the input features; This is the output channel index, with a value range of [value range missing]. , It is the number of output channels after the convolution operation; This is a time-series location index, with a value range of [value range missing]. , The length of the signal is the input feature.
[0039] Multiple separate convolutional kernels of different scales are used to process the features after cross-channel convolution separately, including:
[0040] ;
[0041] in, For the first The output signal of the separated convolution at each scale in the channel Timing position eigenvalues; , This indicates the length of the signal after convolution; Indicates the first The weights of the convolutional kernel; For the signal after cross-channel convolution in the channel Timing position eigenvalues.
[0042] Furthermore, in step S4, the multi-scale local features are used as input, and three sets of linear mappings are applied to them to generate a time score vector, a key matrix, and a value matrix, including:
[0043] ; ;
[0044] ;
[0045] in, , representing the time score vector; , represents the bond matrix; Represents a value matrix;
[0046] , , This is the weight matrix; , which are the multi-scale local features of the input. For sequence length, For feature dimensions.
[0047] Furthermore, in step S4, the channel feature weight vector is obtained by weighting and aggregating the features of the key matrix using the time score vector, including:
[0048] ;
[0049] in, This represents the channel feature weight vector, reflecting the importance of each channel in the global perspective; Indicates the first Time score for each sequence position; Indicates the first Key vectors at each sequence position; This indicates a broadcast operation.
[0050] Furthermore, in step S4, the channel feature weight vector is broadcast to the return value matrix and the temporal features are weighted. An enhanced feature matrix is obtained through linear mapping, including:
[0051] ;
[0052] in, , representing the enhanced feature matrix; , indicating the first A value matrix for each channel; Indicates the first Channel feature weights for each channel; , indicating the output linear mapping weights.
[0053] Furthermore, in step S4, the fused features output by the lightweight feature extraction layer are mapped to the probability distribution of each health state category of the mechanical system, including:
[0054] First, perform global average pooling on the sequence dimension of the fused features to obtain the global feature vector:
[0055] ;
[0056] in, , representing the global feature vector; The temporal length of the fused features; Indicates fusion characteristics;
[0057] Then, the global feature vector is projected onto the preset fault category space through a linear mapping:
[0058] ;
[0059] in, This represents the output vector after linear mapping; , represents the output layer weight matrix; , representing the bias vector. Number of fault categories;
[0060] Finally, the linear mapping result is normalized to a probability distribution using the SoftMax function:
[0061] ;
[0062] in, Indicates that the device belongs to the first The probability of a fault type; Indicates the first The score for each type of fault; Indicates the first The score for the type of fault, This represents the number of fault categories.
[0063] The beneficial effects of this invention are:
[0064] (1) Strong ability to express features
[0065] This invention extracts multi-scale time-frequency features through wavelet packet transform and combines fuzzy entropy to quantify signal complexity, thereby achieving synergistic enhancement of time-frequency information and complexity features. This significantly improves the distinguishability of fault features and can accurately capture the essential characteristics of faults even in scenarios with noise interference and fluctuating operating conditions.
[0066] (2) The model is lightweight and efficient
[0067] The separable multi-scale convolution module (SMC module) of the lightweight feature extraction layer in the WFEConvformer model of this invention achieves effective extraction of multi-scale features while reducing the model parameter size and computational complexity through 1×1 cross-channel convolution and multi-scale separable convolution; the broadcast self-attention module (BSA module) does not require the construction of a high-dimensional attention matrix and achieves global feature modeling in a lightweight manner, enabling the model to be deployed on edge devices and meet real-time detection requirements.
[0068] (3) High detection accuracy and robustness
[0069] The lightweight feature extraction layer in the WFEConvformer model constructed in this invention achieves deep fusion of local features and global information through the synergistic effect of the SMC module and the BSA module. It captures both local detailed features of the signal and takes into account the global distribution law. Combined with normalization processing to eliminate the influence of data distribution offset, the model has excellent recognition accuracy and robustness for mechanical faults of different types and under different working conditions. Attached Figure Description
[0070] Figure 1 This is a flowchart of the mechanical fault detection method based on the WFEConvformer model in Embodiment 1 of the present invention.
[0071] Figure 2 This is a schematic diagram of the WFEConvformer model framework structure of the present invention.
[0072] Figure 3 This is a schematic diagram of the lightweight feature extraction layer in the WFEConvformer model of the present invention.
[0073] Figure 4 This is a schematic diagram comparing the detection results of the present invention using the solution of the present invention in Embodiment 2 with the detection results of the prior art. Detailed Implementation
[0074] This invention aims to provide a mechanical fault detection method based on the WFEConvformer model, addressing the problems of high computational complexity, weak multi-scale and global feature capture capabilities, and insufficient feature discrimination in existing mechanical fault detection methods, thus meeting the real-time and accuracy requirements of industrial equipment for fault diagnosis. Its core idea is as follows: First, considering the characteristics of fuzzy and multi-scale distribution of fault features, wavelet packet transform is used to decompose multi-channel vibration signals to obtain time-frequency features at multiple frequency scales. Simultaneously, fuzzy entropy is used to quantify signal complexity, and feature concatenation is used to achieve multi-dimensional information synergistic enhancement, improving fault feature discrimination. Second, a lightweight feature extraction layer is constructed, consisting of a separable multi-scale convolutional module (SMC) and a broadcast self-attention module (BSA). The SMC module captures multi-scale local features with low computational cost, while the BSA module achieves global feature aggregation in a lightweight manner. The two work together to achieve deep fusion of local and global features. Finally, through global average pooling, linear mapping, and probability normalization, high-dimensional features are accurately mapped to the fault category space, ultimately achieving high-precision and robust mechanical fault detection with low resource consumption.
[0075] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0076] Example 1
[0077] This embodiment provides a mechanical fault detection method based on the WFEConvformer model. (See also...) Figure 1 It includes the following implementation process: S1. Multi-channel vibration signal acquisition
[0078] In this step, multi-channel vibration signals of the mechanical system are collected, and the vibration signals are extracted using a sliding window to generate a sample dataset. Since vibration signals from different parts can reflect the fault characteristics at different locations, multi-channel data can provide more comprehensive fault information, avoid the limitations of single-channel signals, and improve the completeness of features.
[0079] For the acquired raw signals, a sliding window with a fixed window length and step size is used to extract the vibration signals without overlap, generating a sample dataset. Each sample contains the vibration signals within a certain time window.
[0080] S2. Feature enhancement is performed using wavelet packet transform and fuzzy entropy algorithm.
[0081] In this step, feature enhancement processing is performed on the samples in the sample dataset: the multi-channel time-domain signals of the samples are converted into time-frequency feature matrices through wavelet packet transform, the complexity feature matrix of the samples is calculated through fuzzy entropy algorithm, and the time-frequency feature matrix and the complexity feature matrix are concatenated along the channel dimension to obtain enhanced features.
[0082] In this process, wavelet packet transform decomposes the multi-channel time-domain signal of the sample into wavelet packets one by one. Specifically, through two-level decomposition, each channel signal is decomposed into four sub-band signals with different frequency scales. The coefficients of each sub-band signal are extracted as feature channels. Then, the sub-band feature channels corresponding to all channels are concatenated along the channel dimension to form a wavelet packet time-frequency feature matrix. The number of channels in the feature matrix is the product of the number of channels in the original signal and the number of sub-band signals, and the length is consistent with that of the original time-domain signal.
[0083] The fuzzy entropy algorithm selects the first channel of the original time-domain signal as the input for fuzzy entropy calculation, constructs a delay vector with an embedding dimension of 2, calculates the maximum distance between any two delay vectors and determines their similarity, and calculates the fuzzy entropy value based on the similarity statistics. The formula is:
[0084] ;
[0085] ;
[0086] in, Represents the fuzzy entropy value; Indicates the signal length; Indicates the embedding dimension; This represents the maximum distance between delay vectors; Indicates similarity tolerance; It represents the standard deviation of the original time-domain signal.
[0087] Next, the fuzzy entropy value is expanded into a one-dimensional feature vector with the same length as the original time-domain signal through a broadcast operation to obtain a complexity feature matrix. Then, the time-frequency feature matrix and the complexity feature matrix are concatenated along the channel dimension to obtain the enhanced feature.
[0088] Since mechanical fault signals have multi-scale characteristics, this embodiment can accurately divide the signal into four key frequency scales through two-layer wavelet packet decomposition, covering the typical characteristic frequency range of mechanical faults. At the same time, multi-channel sub-band feature splicing can preserve the frequency information differences of each channel and improve the richness of features.
[0089] Fuzzy entropy can effectively quantify the complexity and irregularity of signals. The complexity of vibration signals under fault conditions is significantly higher than that under normal conditions. Fuzzy entropy can accurately characterize this difference in order to precisely capture the essential characteristics of the fault.
[0090] S3. Enhanced Feature Normalization Processing
[0091] In this step, the enhanced features are normalized to obtain training data.
[0092] Specifically, this embodiment employs a min-max normalization method for the normalization of enhanced features, scaling the value of each feature to the range of [0,1] to eliminate dimensional differences. The formula is as follows:
[0093] ;
[0094] in, This represents one of the enhanced features; Represents the normalized features; Representation of features The minimum value; Representation of features The maximum value.
[0095] After the above processing, the training data is obtained and then divided into training, validation, and test sets according to a preset ratio (e.g., 7:2:1). The training set is large enough to support model parameter learning, the validation set can effectively monitor model overfitting and optimize hyperparameters, and the test set has no overlap with the training and validation sets, allowing for an objective evaluation of the model's generalization ability on unseen data.
[0096] S4. Construct the WFEConvformer model
[0097] In this step, the constructed WFEConvformer model architecture can be found in [reference needed]. Figure 2 The system comprises an input layer, a lightweight feature extraction layer, and an output layer. The input layer receives multi-dimensional features after feature enhancement and normalization. It first compresses the temporal dimension and smooths noise interference through an average pooling layer to retain core temporal information. Then, convolutional blocks adapt the feature channel dimensions and perform preliminary aggregation of local information, outputting standardized features adapted for subsequent modules. The lightweight feature extraction layer receives the standardized features output from the input layer. It extracts multi-scale local fault features through a separable multi-scale convolutional module, achieves global feature modeling and local-global information fusion through a broadcast self-attention module, and optimizes feature representation through layer normalization and a lightweight feedforward network, outputting fused features with both discriminative power and robustness. The output layer receives the fused features output from the lightweight feature extraction layer. It compresses temporal information through global average pooling to obtain a global feature vector, projects it onto the fault category space through linear mapping, and then outputs the probability distribution of each health state category through probability normalization, providing a basis for fault detection result determination.
[0098] The lightweight feature extraction layer is the core layer of the WFEConvformer model. (See [link to relevant documentation]). Figure 3 It includes a separable multi-scale convolutional module (SMC) and a broadcast self-attention module (BSA).
[0099] Separable multi-scale convolutional modules (SMCs) are used to reduce model complexity while maintaining good feature extraction capabilities; for input signals... ,in, The number of channels for the input signal. The length of the signal; the Separable Multi-Scale Convolutional Module (SMC) uses a convolution kernel of size 1 to perform cross-channel convolution, obtaining a new signal by convolving each channel of the input signal. ,in, It is the number of output channels after the convolution operation, and the formula is:
[0100] ; in, , , This indicates a cross-channel convolution kernel.
[0101] After cross-channel convolution, the separable multi-scale convolution module (SMC) processes the signal using separable convolution kernels of multiple scales. The size of the separable convolution kernels is... ,in , It refers to the number of convolution kernels. Each scale of the convolution kernel processes the signal independently, and each scale of convolution produces an output. The formula is:
[0102] ;
[0103] in, , This indicates the length of the signal after convolution. Indicates the first The weights of the convolutional kernels.
[0104] Finally, the convolutional outputs of all scales are concatenated to form a new feature matrix. The final feature representation is obtained by performing nonlinear mapping through batch normalization and activation functions, as shown in the formula:
[0105] ;
[0106] in, This means concatenating the outputs of convolutions at multiple scales together. This indicates a batch normalization operation. This represents the activation function, used to increase the nonlinearity of the model.
[0107] The Broadcast Self-Attention (BSA) module is used to achieve global feature modeling while keeping computational complexity under control; its input is a feature sequence. ,in For sequence length, The feature dimension is used to achieve global information aggregation without constructing a high-dimensional attention matrix. The Broadcast Self-Attention (BSA) module first applies three sets of linear mappings to the input features, respectively through a weight matrix. , and The formula for generating the time score vector, key matrix, and value matrix is as follows:
[0108] ; ;
[0109] ;
[0110] in, , representing the time score vector, reflects the contribution of each sequence position to the global information, key matrix Value matrix This reflects the characteristic representation of each sequence position.
[0111] After obtaining the above three features, the Broadcast Self-Attention (BSA) module performs weighted aggregation of the key features using time scores to obtain the feature weights for each channel, as shown in the formula:
[0112] ;
[0113] in, This represents the channel feature weight vector, reflecting the importance of each channel in the global perspective. Indicates the first Time score for each sequence position Indicates the first A key vector at each sequence position. This indicates a broadcast operation.
[0114] Then, the Broadcast Self-Attention Module (BSA) will use the channel feature weight vector... Broadcast return value matrix The temporal features of each channel are weighted to obtain the enhanced output feature matrix, as shown in the formula:
[0115] ;
[0116] in, , represents the enhanced feature matrix containing global features. , indicating the first A matrix of values for each channel. Indicates the first Channel feature weights for each channel. , indicating the output linear mapping weights.
[0117] Through the above process, the channel feature weights are broadcast back to the sequence dimension, enabling the output of the Broadcast Self-Attention (BSA) module to simultaneously possess the ability of global feature weighting and local feature preservation, thus achieving a lightweight global attention mechanism; finally, the feature matrix is enhanced. After feature normalization and feature representation enhancement through layer normalization and a lightweight feedforward network, fused features with both global information and local details are obtained. This is the final output of the lightweight feature extraction layer.
[0118] The output layer of the mechanical fault detection model WFEConvformer takes the input feature matrix as input. Global average pooling is performed along the sequence dimension to compress temporal information into a global feature vector, as shown in the formula:
[0119] ;
[0120] in, , , representing the global feature vector.
[0121] Then, a linear mapping is used to project the global feature vector onto a preset fault category space, as shown in the formula:
[0122] ;
[0123] in, , represents the output layer weight matrix, , representing the bias vector. This represents the number of fault categories.
[0124] Finally, a function is applied to the linear output to normalize the output values of each category into a probability distribution, as shown in the formula:
[0125] ;
[0126] in, , Indicates that the device belongs to the first The probability of a class's running state or fault type;
[0127] By combining global pooling, linear mapping, and probability normalization, the model can effectively aggregate high-dimensional features and map them to an interpretable classification space, thus achieving the final determination of the type of mechanical fault.
[0128] S5. WFEConvformer Model Training
[0129] In this step, the constructed WFEConvformer model is trained on the training set partitioned in step S3, and the model overfitting is monitored and hyperparameters are optimized using the validation set. Finally, the generalization ability of the model is objectively evaluated on the test set.
[0130] S6. Perform fault detection
[0131] In this step, in practical applications, the trained WFEConvformer model is deployed as a fault detection model to collect multi-channel vibration signals of the mechanical system in real time. After sliding window truncation in step S1, feature enhancement processing in step S2, and normalization processing in step S3, the signals are input into the fault detection model to obtain the fault detection results.
[0132] Example 2
[0133] This embodiment takes the application of the present invention to gear fault detection on a mechanical platform as an example.
[0134] First, a brief introduction to the composition of the mechanical platform and the data acquisition process:
[0135] The mechanical platform consists of a drive motor, gearbox, coupling, load device, and multi-channel vibration acceleration sensors. Data is collected by multiple vibration acceleration sensors arranged in different parts of the gearbox to record the vibration response signals of the gears under different operating conditions in real time. The mechanical platform is equipped with various typical gear faults, including tooth root cracks, tooth surface wear, broken teeth, and normal conditions, and multi-condition datasets are generated by changing operating conditions such as speed and load. During the acquisition process, the vibration signals are continuously sampled at a high sampling rate to ensure that the local impact characteristics and high-frequency fault features during gear meshing can be captured.
[0136] The specific methods employed include:
[0137] Multi-channel vibration signals of the gear are collected by a vibration acceleration sensor. After the collected multi-channel vibration signals are processed by sliding window truncation, feature enhancement and normalization, they are input into the mechanical fault detection model to detect the fault category.
[0138] To verify the effectiveness of the method proposed in this invention, this embodiment uses the WFEConvformer model, CLFormer model, MCSwin-T model, and MobileNet-V2 model to perform detection on the same dataset. The detection results of the different models are compared, for example... Figure 4As shown in the figure, the WFEConvformer model's polygon line consistently lies above the other three polygon lines, and it converges the fastest, achieving high accuracy and stabilizing early in training. The CLFormer model has the lowest overall accuracy, and even after convergence, it still lags significantly behind other models. The MCSwin-T and MobileNet-V2 models have intermediate accuracy levels, with both their convergence speed and stabilized accuracy lower than the WFEConvformer model. This figure clearly demonstrates the advantages of the WFEConvformer model in fault detection tasks, namely faster convergence and higher detection accuracy.
[0139] To objectively measure the model's detection capability, this embodiment uses the average accuracy. and average output characteristic evaluation index The evaluation model's detection results are calculated using the following formula:
[0140] ;
[0141] ;
[0142] in, For the accuracy of a single experiment, It is the number of fault categories. It is the first The number of samples that are correctly predicted as belonging to that class. It is the total number of samples in the test set. It is the first The accuracy of this experiment. It represents the number of independent experiments.
[0143] ;
[0144] ;
[0145] in, It is the inter-class scatter matrix. It is the within-class scatter matrix. It is a norm 2. It is the first The output feature evaluation index of this experiment It represents the number of independent experiments.
[0146] After five rounds of experiments, the fault detection method based on the WFEConvformer model in this embodiment is compared with the detection results of existing technologies, as shown in Table 1.
[0147] Table 1. Comparison of test results data between the proposed solution and existing technologies.
[0148]
[0149] As can be seen from the table above, the average accuracy of the detection results obtained by the method proposed in this invention is... The average output characteristic evaluation index was 98.92%. The accuracy was 2.83; in contrast, the average accuracy of the detection results using the CLFormer model was... The average output characteristic evaluation index is 90.03%. The average accuracy of the detection results using the MCSwin-T model is 1.84. The average output characteristic evaluation index is 98.09%. The average accuracy of the detection results using the MobileNet-V2 model is 2.77. The average output characteristic evaluation index was 97.86%. The average accuracy and average output feature evaluation index of the fault detection model used in this embodiment are higher than those of the comparison model, thus proving that the fault detection capability of the fault detection model provided in this embodiment is more in line with the expected results.
[0150] Although embodiments of the present invention have been described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present invention, and all such changes and alterations shall not depart from the protection scope of the present invention.
Claims
1. A mechanical fault detection method based on the WFEConvformer model, characterized in that, Includes the following steps: S1. Collect multi-channel vibration signals from the mechanical system, use a sliding window to extract the vibration signals, and generate a sample dataset; S2. Perform feature enhancement processing on the sample dataset: convert the multi-channel time-domain signal of the sample into a time-frequency feature matrix through wavelet packet transform, calculate the complexity feature matrix of the sample through fuzzy entropy algorithm, and concatenate the time-frequency feature matrix and the complexity feature matrix along the channel dimension to obtain enhanced features; S3. Normalize the enhanced features to obtain training data; S4. Construct the WFEConvformer model, which includes an input layer, a lightweight feature extraction layer, and an output layer; The input layer is used to receive input features, complete temporal dimension optimization, channel adaptation and preliminary integration of local information, and output standardized features adapted to subsequent module processing. The lightweight feature extraction layer is composed of a separable multi-scale convolutional module and a broadcast self-attention module connected in series. Among them, the separable multi-scale convolution module first performs cross-channel convolution on the standardized features through a 1×1 convolution kernel, then uses multiple separable convolution kernels of different scales to process the features after cross-channel convolution separately, and finally concatenates the convolution outputs of all scales to obtain multi-scale local features. The broadcast self-attention module takes the multi-scale local features as input, first applies three sets of linear mappings to them to generate a time score vector, a key matrix and a value matrix, then uses the time score vector to weight and aggregate the features of the key matrix to obtain a channel feature weight vector, broadcasts the channel feature weight vector back to the value matrix and weights the temporal features, and obtains the enhanced feature matrix through linear mapping; The enhanced feature matrix is then normalized and enhanced through layer normalization and a lightweight feedforward network to obtain a fused feature that integrates local and global information, which serves as the output of the lightweight feature extraction layer. The output layer is used to map the fused features output by the lightweight feature extraction layer to the probability distribution of each health state category of the mechanical system; S5. Train the WFEConvformer model based on the training data to obtain the trained WFEConvformer model; S6. In practical applications, multi-channel vibration signals of the mechanical system are acquired in real time, and after sliding window truncation, feature enhancement processing and normalization processing, they are input into the trained WFEConvformer model to obtain fault detection results.
2. The mechanical fault detection method based on the WFEConvformer model as described in claim 1, characterized in that, In step S1, the step of using a sliding window to extract the vibration signal includes: dividing the acquired original multi-channel vibration signal into independent, equal-length sample segments according to a fixed window length and step size.
3. The mechanical fault detection method based on the WFEConvformer model as described in claim 1, characterized in that, In step S2, the conversion of the multi-channel time-domain signal of the sample into a time-frequency feature matrix through wavelet packet transform includes: The multi-channel time-domain signal of the sample is decomposed into two-level wavelet packet decomposition for each channel. Each channel signal is decomposed into four sub-band signals with different frequency scales. The coefficients of each sub-band signal are extracted as feature channels. The sub-band feature channels corresponding to all channels are concatenated along the channel dimension to form a time-frequency feature matrix. The number of channels of the time-frequency feature matrix is the product of the number of channels of the original signal and the number of sub-band signals, and the length is consistent with that of the original time-domain signal.
4. The mechanical fault detection method based on the WFEConvformer model as described in claim 3, characterized in that, In step S2, calculating the complexity feature matrix of the sample using the fuzzy entropy algorithm includes: The first channel of the original time-domain signal is selected as the input for fuzzy entropy calculation. A delay vector with an embedding dimension of 2 is constructed. The maximum distance between any two delay vectors is calculated and their similarity is judged. The fuzzy entropy value is calculated based on the similarity statistics. The fuzzy entropy value is expanded into a one-dimensional feature vector with the same length as the original time-domain signal through a broadcast operation to obtain the complexity feature matrix. The fuzzy entropy value is calculated as follows: ； ； in, Represents the fuzzy entropy value; Indicates the signal length; Indicates the embedding dimension; This represents the maximum distance between delay vectors; Indicates similarity tolerance; It represents the standard deviation of the original time-domain signal.
5. The mechanical fault detection method based on the WFEConvformer model as described in claim 1, characterized in that, In step S3, the enhanced features are normalized, including: ； in, This represents one of the enhanced features; Represents the normalized features; Representation of features The minimum value; Representation of features The maximum value; The separable multi-scale convolution module first performs cross-channel convolution on the input features using a 1×1 convolution kernel, including: ； in, Features obtained after cross-channel convolution; Input features; These are the weights of the cross-channel convolution kernel; The channel index of the input feature, with a value range of 1. , This is the original number of channels for the input features; This is the output channel index, with a value range of [value range missing]. , It is the number of output channels after the convolution operation; This is a time-series location index, with a value range of [value range missing]. , The length of the signal is the input feature.
6. The mechanical fault detection method based on the WFEConvformer model as described in claim 5, characterized in that, Multiple separate convolutional kernels of different scales are used to process the features after cross-channel convolution separately, including: ； in, For the first The output signal of the separated convolution at each scale in the channel Timing position eigenvalues; , This indicates the length of the signal after convolution; Indicates the first The weights of the convolutional kernel; For the signal after cross-channel convolution in the channel Timing position eigenvalues.
7. The mechanical fault detection method based on the WFEConvformer model as described in claim 6, characterized in that, In step S4, the multi-scale local features are used as input, and three sets of linear mappings are applied to them to generate a time score vector, a key matrix, and a value matrix, including: ； ； ； in, , representing the time score vector; , represents the bond matrix; Represents a value matrix; , , This is the weight matrix; , which are the multi-scale local features of the input. For sequence length, For feature dimensions.
8. The mechanical fault detection method based on the WFEConvformer model as described in claim 7, characterized in that, In step S4, the channel feature weight vector is obtained by weighting and aggregating the features of the key matrix using the time score vector, including: ； in, This represents the channel feature weight vector, reflecting the importance of each channel in the global perspective; Indicates the first Time score for each sequence position; Indicates the first A key vector for each sequence position; This indicates a broadcast operation.
9. The mechanical fault detection method based on the WFEConvformer model as described in claim 8, characterized in that, In step S4, the channel feature weight vector is broadcast to the return value matrix and the temporal features are weighted. An enhanced feature matrix is obtained through linear mapping, including: ； in, , representing the enhanced feature matrix; , indicating the first A value matrix for each channel; Indicates the first Channel feature weights for each channel; , indicating the output linear mapping weights.
10. The mechanical fault detection method based on the WFEConvformer model as described in any one of claims 1 to 9, characterized in that, In step S4, the fused features output by the lightweight feature extraction layer are mapped to the probability distribution of each health state category of the mechanical system, including: First, perform global average pooling on the sequence dimension of the fused features to obtain the global feature vector: ； in, , representing the global feature vector; The temporal length of the fused features; Indicates fusion characteristics; Then, the global feature vector is projected onto the preset fault category space through a linear mapping: ； in, This represents the output vector after linear mapping; , represents the output layer weight matrix; , representing the bias vector. Number of fault categories; Finally, the linear mapping result is normalized to a probability distribution using the SoftMax function: ； in, Indicates that the device belongs to the first The probability of a fault type; Indicates the first The score for each type of fault; Indicates the first The score for the type of fault, This represents the number of fault categories.