Fault diagnosis method and system based on multi-scale dual feature reinforcement

Abstract

The application provides a fault diagnosis method and system based on multi-scale double feature reinforcement, and relates to the technical field of fault diagnosis.The application classifies the fault defects of rolling bearings through a fault recognition model, wherein the fault recognition model comprises a feature selection module, a multi-scale feature extraction module and a feature fusion and output module; a discriminative feature reinforcement mechanism is seamlessly embedded into the feature selection module and the multi-scale feature extraction module to reinforce the fault features, and the discriminative feature reinforcement mechanism comprises two modules, namely a multi-spectrum channel attention mechanism and a double-path excitation attention mechanism.The application extracts rich fault feature information of different scales, and through the discriminative feature reinforcement mechanism, not only the fine-grained attention between channels is realized, but also the important information of the pulse excitation segment in the fault signal on the time axis is emphasized, so that the influence of irrelevant information such as noise and working condition change on the final task is effectively reduced, and the fault diagnosis precision of the model is improved.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and specifically to a fault diagnosis method and system based on multi-scale dual feature enhancement. Background Technology

[0002] Rolling bearings, as core components of rotating machinery, are widely used in various industrial equipment in transportation, energy, and aviation, and are hailed as the "joints of industry." However, the long-term operation of rolling bearings in harsh environments undoubtedly greatly increases their likelihood of failure. In mechanical equipment failure cases, 30% to 40% of failures are caused by bearing failures. Therefore, accurate diagnosis of rolling bearing failures has always been an important engineering challenge.

[0003] Traditional fault diagnosis methods rely too heavily on expert experience and have low feature extraction efficiency, making it difficult to meet the requirements of high efficiency and high accuracy in fault diagnosis. Deep learning methods, on the other hand, effectively address the shortcomings of traditional fault diagnosis. Therefore, deep learning-based rolling bearing fault diagnosis methods, such as Long Short-Term Memory Network (LSTM), Deep Belief Network (DBN), and Convolutional Neural Network (CNN), are widely used and studied in bearing diagnosis.

[0004] Among them, CNN and its variants dominate the deep learning-based bearing fault diagnosis methods. However, existing fault diagnosis models based on convolutional neural networks still have shortcomings: (1) lack of discriminative feature learning mechanism. Different features in CNN have different importance to the fault diagnosis task. Traditional CNN cannot adaptively focus on learning more discriminative features and ignore useless features. (2) lack of rich feature extraction for reliable fault diagnosis. Since vibration signals have nonlinear and non-stationary characteristics, and traditional CNN extracts features at a single scale, some local information will be lost when the convolution kernel size is too large, while global features are easily ignored when it is too small, resulting in low reliability of fault diagnosis. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a fault diagnosis method and system based on multi-scale dual feature enhancement, which solves the technical problem of low reliability in fault diagnosis based on convolutional neural network-based fault diagnosis models.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] In a first aspect, the present invention provides a fault diagnosis based on multi-scale dual-feature enhancement, comprising:

[0010] Obtain vibration signals of rolling bearings with several different faults and defects over a certain period of time;

[0011] Several vibration signals were preprocessed to obtain a dataset;

[0012] A fault identification model is obtained by training and optimizing a pre-built multi-scale convolutional neural network with a discriminative feature enhancement mechanism using a dataset. This fault identification model is used to classify faults and defects in rolling bearings.

[0013] The pre-constructed discriminative feature enhancement mechanism in the multi-scale convolutional neural network includes a feature selection module, a multi-scale feature extraction module, and a feature fusion and output module. The feature selection module and the multi-scale feature extraction module embed a discriminative feature enhancement mechanism to enhance fault features. The discriminative feature enhancement mechanism includes a multi-spectral channel attention mechanism and a dual-path activation attention mechanism.

[0014] Preferably, the feature selection module includes a wide convolution module, which is used to perform feature filtering on the vibration signal, expressed as:

[0015] x′ i =P(f(BN(Conv(x)) i ))))

[0016] Where P(·) is the max pooling operation, BN(·) is the batch normalization operation; f(·) is the ReLU activation function, Conv(·) is the convolution operation, and x i It is the i-th vibration signal, x′ i This indicates the features selected after the above operations.

[0017] Preferably, the multi-scale feature extraction module includes convolutional modules of different scales with a three-branch parallel structure.

[0018] Preferably, the feature fusion and output module includes a fusion channel, a global average pooling layer, and a fully connected layer. In the feature fusion and output module, a multi-spectral channel attention mechanism is used to adaptively adjust the features fused along the channels. The adjusted fused features are processed by a global average pooling layer to reduce feature parameters. The output of the global average pooling layer is input to the fully connected layer, and the output of the fully connected layer is processed by Softmatrix. x The function outputs the fault classification results.

[0019] Preferably, the expression for the discriminative feature enhancement mechanism is:

[0020] X DFEM=F DPEAM (F MSCAM (X))

[0021] X DFEM F represents the enhanced features following the discriminative feature mechanism. MSCAM (·) indicates that the data has been processed by a multispectral channel attention mechanism, F DPEAM (·) indicates that the data has been processed through a dual-path attention mechanism.

[0022] Preferably, the data processing procedure in the multispectral channel attention mechanism includes:

[0023] The Fourier transform converts the input time-series signal into the frequency domain, specifically including:

[0024] The input feature X is represented as X = [X 0 X 1 , ..., X c-1 ],in i∈{0,1,…,c-1}, N is the signal length, convert the timing input to the frequency domain:

[0025]

[0026] Where FFT(·) is the FFT function, It is the j-th frequency component on the i-th channel;

[0027] The frequency domain components are divided into blocks and weighted, specifically including:

[0028] The frequency components of each channel are divided into three parts—low-frequency, mid-frequency, and high-frequency—based on their magnitude, and the weights of each frequency band are calculated.

[0029]

[0030]

[0031] Where σ is the sigmoid function, f(·) is the ReLU activation function, and Conv(·) is the convolution operation; This represents the weight of the k-th frequency band on the i-th channel, where k = 1, 2, 3, representing the low-frequency band, mid-frequency band, and high-frequency band, respectively.

[0032] The weights of each frequency band are combined to achieve channel recalibration, specifically including:

[0033]

[0034]

[0035]

[0036] in, It is the final weight of each channel. It is element-wise multiplication, with weight W ci Multiplying by the corresponding frequency component yields the recalibrated frequency characteristic X(F). i The temporal recalibration feature vector is obtained by performing an inverse Fourier transform (IFFT) (·).

[0037] By utilizing the concept of residual learning, residual connections are introduced to obtain multispectral channel attention features X. MSCAM :

[0038]

[0039] Preferably, the data processing procedure in the dual-path stimulated attention mechanism includes:

[0040] Features X adaptively weighted by the channel MSCAM First, perform convolution operations along the channel domain to obtain the convolutional feature X′:

[0041] X′=f(Conv 1×1 (X MSCAM ))

[0042] Among them, X MSCAM =[X 0 X 1 , ..., X n-1 ],in This represents the j-th position on the timing signal;

[0043] Then, a bi-branch convolution operation is performed on feature X′ to obtain bi-convolution feature X″:

[0044] X″=Conv 1×1 (concat[f(Conv 3×1 (X′)), f(Conv) 7×1 (X′))])

[0045] Temporal recalibration is performed on the biconvolutional feature X″ to obtain the temporal recalibrated feature vector X. N :

[0046] W s =σ(X″)

[0047]

[0048] Where s∈{0,1,…,N-1},

[0049] The temporal recalibration feature vector and the input feature are residually concatenated to obtain the dual-path excitation attention X. DPEAM :

[0050] X DPEAM =X MSCAM +X N .

[0051] Secondly, the present invention provides a fault diagnosis system based on multi-scale dual-feature enhancement, comprising:

[0052] The data acquisition module is used to acquire vibration signals of rolling bearings with several different faults and defects over a certain period of time.

[0053] The preprocessing module is used to preprocess several vibration signals to obtain a dataset;

[0054] The model training module is used to train and optimize a pre-built multi-scale convolutional neural network with a discriminative feature enhancement mechanism using a dataset to obtain a fault identification model, which is used to classify faults and defects in rolling bearings.

[0055] The pre-constructed discriminative feature enhancement mechanism in the multi-scale convolutional neural network includes a feature selection module, a multi-scale feature extraction module, and a feature fusion and output module. The feature selection module and the multi-scale feature extraction module embed a discriminative feature enhancement mechanism to enhance fault features. The discriminative feature enhancement mechanism includes a multi-spectral channel attention mechanism and a dual-path activation attention mechanism.

[0056] Thirdly, the present invention provides a computer-readable storage medium storing a computer program for fault diagnosis based on multi-scale dual-feature enhancement, wherein the computer program causes a computer to execute the fault diagnosis method based on multi-scale dual-feature enhancement as described above.

[0057] Fourthly, the present invention provides an electronic device, comprising:

[0058] One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing fault diagnosis based on multi-scale dual-feature enhancement as described above.

[0059] (III) Beneficial Effects

[0060] This invention provides a fault diagnosis method and system based on multi-scale dual feature enhancement. Compared with existing technologies, it has the following advantages:

[0061] The fault diagnosis method and system based on multi-scale dual feature enhancement of this invention classifies faults and defects in rolling bearings using a pre-trained fault identification model. The fault identification model includes a feature selection module, a multi-scale feature extraction module, and a feature fusion and output module. A discriminative feature enhancement mechanism is seamlessly embedded into the feature selection module and the multi-scale feature extraction module to enhance fault features. This discriminative feature enhancement mechanism comprises two modules: a multi-spectral channel attention mechanism and a dual-path excitation attention mechanism. By extracting rich fault feature information at different scales and employing the discriminative feature enhancement mechanism, this invention achieves both fine-grained attention to the interdependencies between channels and emphasizes the important information of the pulse excitation segment in the fault signal on the time axis. This effectively reduces the impact of irrelevant information such as noise and changes in operating conditions on the final task, thereby improving the model's fault diagnosis accuracy. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a block diagram of a fault diagnosis method based on multi-scale dual feature enhancement according to an embodiment of the present invention;

[0064] Figure 2 This is a detailed structural diagram of the multi-scale convolutional neural network for the discriminative feature enhancement mechanism in this embodiment of the invention;

[0065] Figure 3 A detailed structural diagram of the multi-spectral channel attention mechanism;

[0066] Figure 4 This is a structural diagram of the dual-path attention mechanism. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] This application provides a fault diagnosis method and system based on multi-scale dual feature enhancement, which solves the technical problem of low reliability of current fault diagnosis models based on convolutional neural networks. It improves the reliability of model fault diagnosis by extracting rich fault feature information at different scales.

[0069] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows:

[0070] Current deep learning-based methods for diagnosing rolling bearing faults mainly suffer from the following drawbacks:

[0071] (1) Existing fault diagnosis models based on convolutional neural networks lack sufficient feature richness in their feature extraction capabilities. Their ability to extract features from a single scale is limited. Due to the nonlinear and non-stationary characteristics of vibration signals, single-scale convolutional neural networks are insufficient to extract rich features for reliable fault diagnosis.

[0072] (2) Since the information value contained in features at different channels and scales is not equal during feature extraction, the fault diagnosis model based on convolutional neural networks treats all features equally due to the lack of a discriminative feature enhancement mechanism, which makes the model difficult to train due to feature redundancy, thus affecting the improvement of fault diagnosis performance. Furthermore, the compression operation of the traditional channel attention mechanism may lead to the loss of a large amount of information, which will have a negative impact on the prediction of channel weights.

[0073] (3) In practical industrial applications, strong noise pollution and variable operating conditions are unavoidable. Although there are many algorithms for noise immunity, most are only applicable to noisy environments with a signal-to-noise ratio greater than -4dB, and not to more severe noise environments. Secondly, due to the significant differences in signal distribution under different operating conditions, the diagnostic methods have poor domain adaptability and are difficult to apply in practical engineering environments. Finally, these two scenarios often do not occur independently in actual industry, but models that can simultaneously achieve robust noise immunity and cross-condition diagnosis are very limited.

[0074] (4) Deep learning-based fault diagnosis models are considered “black box” models due to their lack of interpretability. This “black box” nature leads to a lack of explanation of the internal decision-making process, which limits their application in industrial applications that require high transparency and interpretability.

[0075] To address the aforementioned issues and enhance the model's feature extraction capabilities, robustness to strong noise, domain adaptability, and interpretability, this invention proposes a fault diagnosis and system based on a discriminative feature enhancement mechanism and multi-scale dual feature enhancement. The improvements of this invention are mainly reflected in the following aspects:

[0076] (1) A multi-scale feature learning strategy was established. First, a wide convolutional module was introduced to filter the features of the original signal, obtaining long-term features from the vibration signal, thereby capturing low-frequency information as much as possible while reducing the impact of noise. Then, a three-branch multi-scale structure was set up to extract multi-scale features of different fine granularities from the non-stationary signal. Convolutional kernels of different sizes learn fault features at different scales in parallel. The features extracted from multiple receptive fields have the ability to capture global and local information, thus obtaining rich fault feature representation information.

[0077] (2) A discriminative feature enhancement mechanism (DFEM) is proposed. Since the collected vibration signals often contain irrelevant information such as environmental noise and rotational speed, in order to enable the model to focus on learning features that are strongly correlated with diagnostic and identification information and improve the fault classification effect, a discriminative feature enhancement mechanism is introduced and embedded in the backbone network to improve the accuracy of bearing fault diagnosis. This mechanism consists of a cascaded multi-spectral channel attention mechanism (MSCAM) and a dual-path excitation attention mechanism (DPEAM).

[0078] ① Multi-channel attention mechanism (MSCAM)

[0079] Compared to traditional channel attention modules, the proposed module is designed from a spectral perspective, avoiding the significant information loss caused by compression operations in traditional channel feature representations, which leads to a lack of diversity in channel feature representations. It utilizes time-frequency analysis to convert the time-series signal into the frequency domain, using multiple frequency components to enhance channel feature representations, while also addressing the problem of time-series signal features being easily masked by noise. Secondly, it captures different frequency components of different channels through an attention mechanism, resulting in more refined channel attention inference. Compared to traditional channel attention mechanisms, this method extends the traditional channel feature representation from compressing channels into a single scalar using Global Average Pooling (GAP) to multiple frequency components, thereby significantly enhancing the channel fault characterization capability and contributing to improved model diagnostic capabilities.

[0080] ② Dual-path stimulation attention mechanism (DPEAM)

[0081] Since the pulse excitation segments exhibited by time-series signals in the time dimension reflect fault characteristics, these pulse excitation segments are crucial for fault diagnosis. Therefore, it is necessary to enhance the signal segments containing fault characteristics to improve the accuracy and efficiency of diagnosis. The Dual-Path Excitation Attention (DPEAM) mechanism proposed in this embodiment of the invention aims to help the model focus on learning pulse excitation segments closely related to the fault through a dual-branch structure, accurately locating the fault impact segment, and also improving the interpretability of the network feature learning process.

[0082] In summary, the Discriminative Feature Enhancement Mechanism (DFEM) uses time-series signals as input to fully utilize the time and frequency domain information of fault signals. It can focus on the interdependencies between channels and extract important information from the signal pulse excitation portion on the time axis, thereby effectively reducing the impact of irrelevant information such as noise and operating condition changes on the target task. Furthermore, the introduction of the attention mechanism improves the interpretability of the neural network feature learning mechanism.

[0083] (3) Considering the plug-and-play nature of the attention mechanism, the discriminative feature enhancement mechanism of the present invention can be seamlessly embedded into other existing neural network architectures to enhance the network’s extraction and learning of discriminative features. By extracting feature weights, the key activation regions in which the model plays a role in the final task can be visualized, thereby enhancing the interpretability of the model.

[0084] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0085] This invention proposes a fault diagnosis method based on multi-scale dual feature enhancement, such as... Figure 1 As shown, the method includes:

[0086] S1. Obtain vibration signals of rolling bearings with several different faults and defects over a certain period of time;

[0087] S2. Preprocess several vibration signals to obtain a dataset;

[0088] S3. The pre-constructed multi-scale convolutional neural network based on the discriminative feature enhancement mechanism is trained and optimized using the dataset to obtain a fault identification model, which is used to classify the faults and defects of rolling bearings.

[0089] The pre-constructed discriminative feature enhancement mechanism in the multi-scale convolutional neural network includes a feature selection module, a multi-scale feature extraction module, and a feature fusion and output module. The feature selection module and the multi-scale feature extraction module embed a discriminative feature enhancement mechanism to enhance fault features. The discriminative feature enhancement mechanism includes a multi-spectral channel attention mechanism and a dual-path activation attention mechanism.

[0090] The following is a detailed description of each step:

[0091] In one embodiment, S1, acquiring vibration signals of a rolling bearing with several different faults over a certain period of time. Specifically, this includes:

[0092] By manually introducing faults into different locations of the rolling bearing (outer ring, inner ring, rolling elements, and other common easily damaged locations), and installing an acceleration sensor at the equipment's signal acquisition point, vibration signals are collected over a certain period of time and uploaded to a computer.

[0093] In one embodiment, S2, several vibration signals are preprocessed to obtain a dataset.

[0094] Specifically, it includes:

[0095] S201. Preprocessing and data augmentation of several vibration signals, including:

[0096] Considering that the collected signals are from a short period of time and the amount of data is relatively small, and that deep learning networks often require massive amounts of data for model training, this embodiment of the invention uses a random starting point continuous sampling technique to expand the data.

[0097] S202 Data Normalization. To accelerate network convergence, z-score is used to normalize the augmented data.

[0098] S202. Split the dataset. Randomly divide the normalized total sample set into a training set, a validation set, and a test set in a 7:2:1 ratio. The training and validation sets are used to prepare for training the neural network, while the test set is used to test the model's performance.

[0099] In one embodiment, S3, the pre-constructed multi-scale convolutional neural network with a discriminative feature enhancement mechanism is trained and optimized using a dataset to obtain a fault identification model, which is used to classify faults and defects in rolling bearings. Specifically, this includes:

[0100] In this embodiment of the invention, the specific structure of the pre-constructed multi-scale convolutional neural network with discriminative feature enhancement mechanism is as follows: Figure 2 As shown, it includes a feature selection module, a multi-scale feature extraction module, and a feature fusion and output module. Furthermore, a discriminative feature enhancement mechanism is seamlessly embedded into the feature selection module and the multi-scale feature extraction module to enhance fault features. This discriminative feature enhancement mechanism consists of a cascaded multi-spectral channel attention mechanism (MSCAM) and a dual-path excitation attention mechanism (DPEAM). It can be represented as:

[0101] X DFEM =F DPEAM (F MSCAM (X))

[0102] In this embodiment of the invention, a multi-spectral channel attention mechanism (MSCAM) is used to improve the traditional attention mechanism's lack of diversity in channel feature representation and its limitation on the model's ability to perceive subtle local differences in fault features by starting from the global receptive field and ignoring the local receptive field. The specific structure of the multi-spectral channel attention mechanism is shown in Figure 3. First, the time-series signal input is converted to the frequency domain using Fourier transform, and then the frequency domain components are divided into blocks and weighted. Specifically, it includes:

[0103] Suppose the input feature X is represented as X = [X 0 X 1 , ..., X c-1 ],in i∈{0, 1, ..., c-1}, N is the signal length. First, convert the timing input to the frequency domain:

[0104]

[0105] Where FFT(·) is the FFT function, It is the j-th frequency component on the i-th channel.

[0106] Then, based on the magnitude of the frequency components, the frequency components of each channel are divided into three parts: low-frequency band, mid-frequency band, and high-frequency band, and weighted accordingly:

[0107]

[0108]

[0109] Here, σ is the sigmoid function, f(·) is the ReLU activation function, and Conv(·) is the convolution operation. By weighting different frequency bands of each channel, fine-grained adjustment and capture of frequency domain features in different channels are achieved. The contribution of each frequency band to the final task is dynamically learned from the local receptive field, resulting in richer and more accurate channel feature representations.

[0110] Then, by combining the weights of each frequency band, channel recalibration is achieved.

[0111]

[0112]

[0113]

[0114] in, This is the final weight of each channel. It is element-wise multiplication, with weight W ci Multiplying by the corresponding frequency component yields the recalibrated frequency characteristic X(F). iThe temporal recalibration feature vector is obtained by performing an inverse Fourier transform (IFFT) (·). This step integrates the weights of each frequency band to adjust the importance of each channel, enabling the model to more effectively process and emphasize important frequency components, thereby improving signal processing performance. The weighted frequency domain representation is then converted to the time domain via an inverse Fourier transform layer, adaptively enhancing or suppressing specific feature channels and improving the network's responsiveness to key features.

[0115] Finally, by utilizing the concept of residual learning, residual connections are introduced to enhance the training stability of the model while preserving the original information.

[0116]

[0117] In MSCAM, fine-grained adjustment of cross-channel frequency characteristics is achieved through processes such as time-frequency transformation, frequency division weighting, and weight synthesis, so as to adaptively recalibrate the characteristic response of the channel.

[0118] The dual-path excitation attention mechanism of this invention can help the network accurately locate fault segments and focus on learning pulse excitation segments closely related to the fault, effectively improving the efficiency and reliability of feature learning, while also enhancing the interpretability of the network feature learning process.

[0119] The specific structure of the dual-path stimulation attention mechanism is as follows: Figure 4 As shown, firstly, a 1*1 convolution is used to obtain the spatial features of channel compression. Secondly, a dual-branch convolution is set up using convolution kernels of different sizes to capture fault features at different scales, enabling the model to simultaneously focus on local details and global structure, enhancing the feature representation of impulse excitation segments, and improving the model's expressive power and the localization accuracy of excitation segments. Then, the weights are multiplied by the corresponding positions of the original input to obtain the weighted excitation impulse signal. Finally, a residual connection is used to add the original input and the weighted excitation impulse signal, preserving the relevant information between the time-series data. The specific steps are as follows:

[0120] Assume input features X MSCAM =[X 0 X 1 , ..., X n-1 ],in This represents the j-th position on the timing signal. The basic steps are as follows:

[0121] Step 1: Perform convolution operation along the channel domain

[0122] X′=f(Conv 1×1 (X MSCAM ))

[0123] Step 2: Two-branch convolution operation

[0124] X″=Conv 1×1 (concat[f(Conv 3×1 (X′)), f(Conv) 7×1 (X′))])

[0125] Step 3: Timing Recalibration

[0126] W s =σ(X″)

[0127]

[0128] Where s∈{0,1,…,N-1},

[0129] Step 4: Residual Connection

[0130] X DPEAM =X MSCAM +X N

[0131] In the fault identification model, the process of processing the vibration signal includes: first, using a wide convolution module in the feature selection module to filter features from the original vibration signal, which can be represented as:

[0132] x′ i =P(f(BN(Conv(x)) i ))))

[0133] Then, feature extraction is performed in the multi-scale feature extraction module using a three-branch parallel multi-scale convolutional module. A discriminative feature enhancement mechanism is seamlessly embedded after the wide convolutional module in the feature selection module and after each multi-scale convolutional module in the multi-scale feature extraction module to adaptively enhance fault features. Secondly, the purpose of feature fusion is to learn and fuse complementary fault features at different scales. However, simple concatenation cannot fully fuse these complementary features, and redundant features are detrimental to model learning. Therefore, in the feature fusion and output module, MSCAM is used to adaptively adjust the features fused along the fusion channel, followed by a GAP (Global Average Pooling) layer to reduce the feature parameters after adaptive fusion, thus preventing model overfitting to some extent. Finally, the output of the fully connected layer is connected to the Softmax function to achieve fault classification.

[0134] In this embodiment of the invention, the fault identification model was trained under a Windows 11 operating system, with an i7-137000k CPU at 3.40GHz and an NVIDIA RTX 4060 GPU. The code was written in Python 3.8 and PyTorch. To enhance the reliability of the experimental results and reduce the impact of randomness, all experiments were run 10 times. (During training, the cross-entropy loss function was used, the Aadm optimizer was selected to update the model parameters, the number of training epochs was set to 100, the batch size to 32, and the learning rate to 0.001. The learning rate training strategy was set so that if the loss on the validation set did not decrease in 10 consecutive iterations, the learning rate would be reduced to 0.1 of the current value. The training and validation set data from step S2 were input into the model for training. During training, the model parameters were updated using the gradient backpropagation principle. The accuracy of the model on the validation and training sets was evaluated, and the model with the highest accuracy on the validation set was saved.)

[0135] This invention also provides a fault diagnosis system based on multi-scale dual feature enhancement, comprising:

[0136] The data acquisition module is used to acquire vibration signals of rolling bearings with several different faults and defects over a certain period of time.

[0137] The preprocessing module is used to preprocess several vibration signals to obtain a dataset;

[0138] The model training module is used to train and optimize a pre-built multi-scale convolutional neural network with a discriminative feature enhancement mechanism using a dataset to obtain a fault identification model, which is used to classify faults and defects in rolling bearings.

[0139] The pre-constructed discriminative feature enhancement mechanism in the multi-scale convolutional neural network includes a feature selection module, a multi-scale feature extraction module, and a feature fusion and output module. The feature selection module and the multi-scale feature extraction module embed a discriminative feature enhancement mechanism to enhance fault features. The discriminative feature enhancement mechanism includes a multi-spectral channel attention mechanism and a dual-path activation attention mechanism.

[0140] It is understood that the fault diagnosis system based on multi-scale dual feature enhancement provided in this embodiment of the invention corresponds to the fault diagnosis method based on multi-scale dual feature enhancement described above. The explanation, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the fault diagnosis based on multi-scale dual feature enhancement, and will not be repeated here.

[0141] This invention also provides a computer-readable storage medium storing a computer program for fault diagnosis based on multi-scale dual-feature enhancement, wherein the computer program causes a computer to execute the fault diagnosis method based on multi-scale dual-feature enhancement as described above.

[0142] Embodiments of the present invention also provide an electronic device, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including functions for performing fault diagnosis based on multi-scale dual-feature enhancement as described above.

[0143] In summary, compared with existing technologies, it has the following beneficial effects:

[0144] 1. This invention proposes a multi-scale learning strategy to improve the model's fault diagnosis accuracy by extracting rich fault feature information at different scales. The specific strategy is as follows: First, a wide convolutional module is used to filter features from the original vibration signal. Compared to traditional small convolutional kernels, this approach captures features over a longer time scale, thus better suppressing noise interference. Second, filters with three branches and different kernel sizes are used to extract feature information of different granularities at different spatial scales in parallel, thereby improving the network model's ability to perceive features of different sizes.

[0145] 2. This invention addresses the problem that traditional convolutional neural networks, by treating features across different channels and scales equally, fail to emphasize the crucial role of key features in the final classification task, leading to a large number of redundant features affecting model training and thus impacting fault diagnosis performance. It proposes a discriminative feature enhancement mechanism to adaptively strengthen fault features and weaken meaningless features, thereby endowing deep learning models with intelligent fault feature perception capabilities. First, from a frequency domain perspective, channel attention is expanded from the original GAP method of compressing channels into a single scalar to multiple frequency components. This enriches feature representation while enhancing the model's ability to perceive subtle local differences in fault features through local receptive fields, achieving adaptive adjustment of the importance of different channels. Second, by accurately locating the fault pulse segment of the time-series signal, the feature learning ability of the fault pulse is enhanced. Finally, although this module uses the original time-series signal as input, it more comprehensively achieves the fusion and enhancement of different fault features in both the time and frequency domains, enabling the model to adaptively focus on learning more discriminative features while ignoring useless features.

[0146] 3. This invention addresses the problem that the lack of interpretability in deep learning models not only reduces the credibility of intelligent fault results but also limits their application in fault diagnosis with high reliability requirements. By visualizing the attention weight distribution of the proposed discriminative feature mechanism in different channels and time series, the feature extraction process of the proposed model is successfully decoded, resulting in an intuitive and physically interpretable result, thereby improving the learning ability and interpretability of neural networks.

[0147] 4. The discriminative feature enhancement mechanism proposed in this embodiment of the invention is a plug-and-play module that can be seamlessly embedded into existing deep learning models to improve performance and provides an effort to explore the intrinsic mechanisms of models in feature learning and classification. Experiments have also demonstrated that embedding this module into existing models is worthwhile because it requires only a small number of parameters and computational overhead, enabling the model with this module embedded to capture the key features needed to achieve the target task while ignoring irrelevant features, thus stably and effectively improving the model's diagnostic capabilities.

[0148] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0149] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fault diagnosis method based on multi-scale dual feature enhancement, characterized in that, include: Obtain vibration signals of rolling bearings with several different faults and defects over a certain period of time; Several vibration signals were preprocessed to obtain a dataset; A fault identification model is obtained by training and optimizing a pre-built multi-scale convolutional neural network with a discriminative feature enhancement mechanism using a dataset. This fault identification model is used to classify faults and defects in rolling bearings. The pre-constructed discriminative feature enhancement mechanism in the multi-scale convolutional neural network includes a feature selection module, a multi-scale feature extraction module, and a feature fusion and output module. The feature selection module and the multi-scale feature extraction module embed a discriminative feature enhancement mechanism to enhance fault features. The discriminative feature enhancement mechanism includes a multi-spectral channel attention mechanism and a dual-path activation attention mechanism. The feature fusion and output module includes a fusion channel, a global average pooling layer, and a fully connected layer. In the feature fusion and output module, a multi-spectral channel attention mechanism is used to adaptively adjust the features fused along the channel. The adjusted fused features are processed by reducing feature parameters through a global average pooling layer. The output of the global average pooling layer is input to the fully connected layer, and the output of the fully connected layer is output as a fault classification result through a Softmax function. The data processing steps in the multispectral channel attention mechanism include: The Fourier transform converts the input time-series signal into the frequency domain, specifically including: The input feature X is represented as X = [X 0 X 1 , ..., X c-1 ],in i∈{0,1,…,c-1}, N is the signal length, convert the timing input to the frequency domain: Where FFT(·) is the FFT function, It is the j-th frequency component on the i-th channel; The frequency domain components are divided into blocks and weighted, specifically including: The frequency components of each channel are divided into three parts—low-frequency, mid-frequency, and high-frequency—based on their magnitude, and the weights of each frequency band are calculated. Where σ is the sigmoid function, f(·) is the ReLU activation function, and Conv(·) is the convolution operation; This represents the weight of the k-th frequency band on the i-th channel, where k = 1, 2, 3, representing the low-frequency band, mid-frequency band, and high-frequency band, respectively. The weights of each frequency band are combined to achieve channel recalibration, specifically including: in, It is the final weight of each channel. It is element-wise multiplication, with weight W ci Multiplying by the corresponding frequency component yields the recalibrated frequency characteristic X(F). i The temporal recalibration feature vector is obtained by performing an inverse Fourier transform (IFFT) (·). By utilizing the concept of residual learning, residual connections are introduced to obtain multispectral channel attention features X. MSCAM :

2. The fault diagnosis based on multi-scale dual feature enhancement as described in claim 1, characterized in that, The feature selection module includes a wide convolution module, which is used to perform feature filtering on the vibration signal, and is expressed as follows: x′ i =P(f(BN(Conv(x i )))) Where P(·) is the max pooling operation, BN(·) is the batch normalization operation; f(·) is the ReLU activation function, Conv(·) is the convolution operation, and x i It is the i-th vibration signal, x′ i This indicates the features selected after the above operations.

3. The fault diagnosis based on multi-scale dual feature enhancement as described in claim 1, characterized in that, The multi-scale feature extraction module includes convolutional modules of different scales with a three-branch parallel structure.

4. The fault diagnosis based on multi-scale dual feature enhancement as described in claim 1, characterized in that, The expression for the discriminative feature enhancement mechanism is: X DFEM =F DPEAM (F MSCAM (X)) X DFEM F represents the enhanced features following the discriminative feature mechanism. MSCAM (·) indicates that the data has been processed by a multispectral channel attention mechanism, F DPEAM (·) indicates that the data has been processed through a dual-path attention mechanism.

5. The fault diagnosis based on multi-scale dual feature enhancement as described in any one of claims 1 to 4, characterized in that, The data processing procedure in the dual-path attention mechanism includes: Features X adaptively weighted by the channel MSCAM First, perform convolution operations along the channel domain to obtain the convolutional feature X′: X′=f(Conv 1×1 (X MSCAM )) Among them, X MSCAM =[X 0 X 1 , ..., X n-1 ],in This represents the j-th position on the timing signal; Then, a bi-branch convolution operation is performed on feature X′ to obtain bi-convolution feature X″: X″=Conv 1×1 (concat[f(Conv 3×1 (X′)),f(Conv 7×1 (X′))]) Temporal recalibration is performed on the biconvolutional feature X″ to obtain the temporal recalibrated feature vector X. N : Ws=σ(X″) Where s∈{0,1,…,N-1}, The temporal recalibration feature vector and the input feature are residually concatenated to obtain the dual-path excitation attention X. DPEAM : X DPEAM =X MSCAM +X N 。 6. A fault diagnosis system based on multi-scale dual feature enhancement, characterized in that, include: The data acquisition module is used to acquire vibration signals of rolling bearings with several different faults and defects over a certain period of time. The preprocessing module is used to preprocess several vibration signals to obtain a dataset; The model training module is used to train and optimize a pre-built multi-scale convolutional neural network with a discriminative feature enhancement mechanism using a dataset to obtain a fault identification model, which is used to classify faults and defects in rolling bearings. The pre-constructed discriminative feature enhancement mechanism in the multi-scale convolutional neural network includes a feature selection module, a multi-scale feature extraction module, and a feature fusion and output module. The feature selection module and the multi-scale feature extraction module embed a discriminative feature enhancement mechanism to enhance fault features. The discriminative feature enhancement mechanism includes a multi-spectral channel attention mechanism and a dual-path activation attention mechanism. The feature fusion and output module includes a fusion channel, a global average pooling layer, and a fully connected layer. In the feature fusion and output module, a multi-spectral channel attention mechanism is used to adaptively adjust the features fused along the channel. The adjusted fused features are processed by reducing feature parameters through a global average pooling layer. The output of the global average pooling layer is input to the fully connected layer, and the output of the fully connected layer is output as a fault classification result through a Softmax function. The data processing steps in the multispectral channel attention mechanism include: The Fourier transform converts the input time-series signal into the frequency domain, specifically including: The input feature X is represented as X = [X 0 X 1 , ..., X c-1 ],in i∈{0,1,…,c-1}, N is the signal length, convert the timing input to the frequency domain: Where FFT(·) is the FFT function, It is the j-th frequency component on the i-th channel; The frequency domain components are divided into blocks and weighted, specifically including: The frequency components of each channel are divided into three parts—low-frequency, mid-frequency, and high-frequency—based on their magnitude, and the weights of each frequency band are calculated. Where σ is the sigmoid function, f(·) is the ReLU activation function, and Conv(·) is the convolution operation; This represents the weight of the k-th frequency band on the i-th channel, where k = 1, 2, 3, representing the low-frequency band, mid-frequency band, and high-frequency band, respectively. The weights of each frequency band are combined to achieve channel recalibration, specifically including: in, It is the final weight of each channel. It is element-wise multiplication, with weight W ci Multiplying by the corresponding frequency component yields the recalibrated frequency characteristic X(F). i The temporal recalibration feature vector is obtained by performing an inverse Fourier transform (IFFT) (·). By utilizing the concept of residual learning, residual connections are introduced to obtain multispectral channel attention features X. MSCAM :

7. A computer-readable storage medium, characterized in that, It stores a computer program for fault diagnosis based on multi-scale dual-feature enhancement, wherein the computer program causes a computer to execute the fault diagnosis method based on multi-scale dual-feature enhancement as described in any one of claims 1 to 5.

8. An electronic device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the fault diagnosis method based on multi-scale dual feature enhancement as described in any one of claims 1 to 5.

Images