Bearing fault diagnosis method and fault self-monitoring device based on pulse neural network

Abstract

The application belongs to the field of mechanical fault detection, and particularly relates to a bearing fault diagnosis method and a fault self-monitoring device based on a pulse neural network. The method comprises constructing a pulse neural network for predicting the fault type of a bearing according to an input bearing vibration signal; the pulse neural network comprises a feature extraction module, a multi-scale attention module and a classifier in sequence; the feature extraction module comprises a convolution layer, a max-pooling layer and a pulse coding layer, and the multi-scale attention module comprises three attention branches; the classifier comprises an average-pooling layer and a full connection layer. The original vibration signals of different fault types are acquired as sample data and used for supervised training of the pulse neural network; the bearing vibration signals in the running process of a target bearing are collected in real time, and fault prediction is realized through the trained pulse neural network. The application solves the problems of poor real-time performance, insufficient robustness and generalization of the existing bearing fault diagnosis method based on a neural network.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of mechanical fault detection, and particularly relates to a bearing fault diagnosis method based on a spiking neural network, a corresponding computer program product, a storage medium and a bearing fault self-monitoring device. BACKGROUND

[0002] Rolling bearings, as the core components of rotating machinery, are widely used in manufacturing, power, transportation, aerospace and other industrial fields. They usually operate in high-temperature and high-pressure environments and are prone to inner ring failure, outer ring failure, rolling element failure, etc. If not diagnosed in time, it may lead to serious safety accidents.

[0003] Existing bearing fault diagnosis methods are mainly divided into physical model-based methods and data-driven methods. The physical model-based method (such as short-time Fourier transform, local mean decomposition, etc.) relies on a large number of hyperparameters and expert experience, and has limited generalization ability; in the data-driven method, traditional machine learning (such as SVM, KNN) is limited by the effectiveness of feature extraction, and deep learning (such as CNN, ResNet) can automatically extract features, but has high delay and high energy consumption problems, and is difficult to adapt to edge computing scenarios.

[0004] As the third generation of neural networks, spiking neural networks (SNN) simulate the pulse transmission mechanism of biological neurons, have event-driven characteristics and high energy efficiency advantages, and show potential in the field of fault diagnosis. However, the existing SNN diagnosis method has the following defects: the original signal needs to be encoded into a two-dimensional image or feature conversion, resulting in information loss and additional computational overhead; the deep network has gradient instability problem, and the feature extraction ability is insufficient; the noise resistance performance needs to be improved, and it is difficult to adapt to high-noise industrial environment.

[0005] Therefore, how to develop a bearing fault diagnosis method without complex data preprocessing, high efficiency and robustness has become a technical problem to be solved by the technical personnel in the field. SUMMARY

[0006] In order to solve the problems of poor real-time performance, insufficient robustness and generalization of the existing bearing fault diagnosis method based on neural network, the present application provides a bearing fault diagnosis method based on a spiking neural network, a corresponding computer program product, a storage medium and a bearing fault self-monitoring device.

[0007] The present application adopts the following technical solutions:

[0008] A bearing fault diagnosis method based on a spiking neural network, which is used for online monitoring of bearing faults during equipment operation, the method comprising the following steps:

[0009] A spiking neural network (SNN) is constructed to predict bearing fault types based on input bearing vibration signals. The SNN comprises a feature extraction module, a multi-scale attention module, and a classifier. The feature extraction module consists of convolutional layers, max-pooling layers, and a pulse coding layer based on a LIF neuron model. This module extracts, downsamples, and pulse-triggered propagation of features from the input bearing vibration signal to obtain state features containing fault information. The multi-scale attention module has three attention branches, each consisting of a dual-spiking attention module with residual connections and a deep separable pulse feedforward network. The convolutional kernels of the dual-spiking attention modules in the three branches are 16, 8, and 4, respectively. The attention-weighted enhanced features extracted from the three branches are then concatenated to form the fused features. The classifier generates classification results based on the fused features using average pooling layers and fully connected layers.

[0010] Raw vibration signals of different fault types were acquired, preprocessed, and labeled with faults to serve as sample data. The sample dataset, containing a large number of samples, was divided into training, testing, and validation sets. Cross-entropy loss was used as the loss function, and a supervised training algorithm was employed to train the spiking neural network. The weight parameters of the model that met the performance requirements after training were retained.

[0011] The bearing vibration signal during the operation of the target bearing is collected in real time and input into the pre-trained pulse neural network to generate the prediction result of the fault type.

[0012] As a further improvement of the present invention, the film potential update formula of the pulse coding layer is as follows:

[0013] ;

[0014] In the above formula, v l ( t -1) and v l ( t ) are respectively the first l The membrane potentials of layer neurons at time t-1 and time t; W l Represented as the first l The weights of interlayer synapses in layer neurons. s l-1 ( t () represents the pulse output of the front layer neurons; θ l-1 This represents the threshold of neurons in the front layer; This represents the membrane time constant.

[0015] spiking output of neurons in the current layer sl t ) is:

[0016]

[0017] In the above formula, H (·) represents the Heaviside step function, θ l represents the threshold value of the current layer neuron.

[0018] As a further improvement of the present application, during the training phase, the surrogate gradient algorithm calculates the pulse output of the neurons in the pulse coding layer through a continuously differentiable surrogate gradient function; and the gradient derived from the surrogate activation function is used to update the weights. Wherein, the surrogate gradient function is as follows:

[0019]

[0020] In the above formula, represents the partial derivative symbol; represents the derivative of the surrogate activation function , which adopts sigmoid function or smooth function of soft threshold.

[0021] The weight update rule is:

[0022]

[0023] In the above formula, Δ W l represents the weight of the interlayer synapse of the updated neuron in the l layer, L represents the loss function, η represents the learning rate.

[0024] As a further improvement of the present application, in the double pulse attention module, the input features are divided into two paths, one of which is processed through convolution layer, pulse coding layer and batch normalization layer in turn, and then matrix multiplied with the input of the other path. The feature map features obtained after matrix operation are processed through the pulse coding layer to obtain the attention map. The attention map and the output of the batch normalization layer are matrix multiplied again, and the obtained features are processed through the pulse coding layer again to obtain the output of the double pulse attention module.

[0025] Wherein, the standard deviation of the matrix after each matrix multiplication operation is also taken as a scaling factor to adjust the amplitude of the attention score.

[0026] ​​​​​As a further improvement of the present application, the deep separable pulse feedforward network is composed of 3 pulse coding layers, 2 point convolution layers and 1 deep separable convolution module. The input features are input to the deep separable convolution module after being processed by the pulse coding layers, the point convolution layers and the pulse coding layers in turn, and the input and output of the deep separable convolution are connected in residual connection, and then processed by the pulse coding layers and the point convolution layers respectively to obtain the output features of the deep separable pulse feedforward network.

[0027] As a further improvement of the present application, the bearing fault types predicted by the pulse neural network include bearing defects of different sizes such as inner ring fault, outer ring fault, rolling element fault, etc., and normal state.

[0028] As a further improvement of the present application, the pre-processing method of the sample data comprises the following steps:

[0029] (i) The Z-score normalization method is used to unify the distribution of all original signal data.

[0030] (ii) The original signal is segmented and processed according to a fixed length to serve as a sample signal.

[0031] (iii) Gaussian white noise with a signal-to-noise ratio range of -4dB to 4dB is added to the sample signal to achieve data increase; wherein the added Gaussian white noise conforms to the normal distribution characteristics with a mean of 0 and a variance of a specific value.

[0032] The present application also includes a storage medium having a computer program stored therein, wherein the computer program, when executed by a processor, creates a pulse neural network trained in the bearing fault diagnosis method based on the pulse neural network as described above, and is used for fault prediction of the input bearing vibration signal.

[0033] The present application also includes a computer program product comprising a computer program, wherein the computer program, when executed by a processor, creates a pulse neural network trained in the bearing fault diagnosis method based on the pulse neural network as described above, and is used for fault prediction of the input bearing vibration signal.

[0034] The present application also includes a bearing fault self-monitoring device, which comprises an acceleration sensor and an analyzer. The acceleration sensor is installed at the bearing on the device and is used to collect the original vibration signal of the target bearing.

[0035] The analyzer comprises a memory, a processor and a computer program stored on the memory and running on the processor, wherein the processor, when executing the computer program, creates a pulse neural network trained in the bearing fault diagnosis method based on the pulse neural network as described above; the analyzer is used to receive the original vibration signal collected by the acceleration sensor in real time, convert it into a bearing vibration signal and input it into the pulse neural network to realize fault prediction.

[0036] The technical scheme provided by the present application has the following beneficial effects:

[0037] In view of the characteristics of the vibration signal containing bearing fault, a new pulse neural network is designed in the embodiment. The network model is a self-organizing neural network architecture, which can capture the dependency and interaction in the input signal by using the self-attention mechanism, so that the model can focus on the most relevant area to achieve accurate diagnosis. The deep separable convolution layer based on neural network is used in the model, which can improve the computing efficiency and extract rich feature representation from the input signal. By integrating the two components, the pulse neural network effectively combines the advantages of context-aware feature extraction and efficient multi-scale feature learning, significantly improving the performance of the bearing fault diagnosis task. The mechanism in the network model can also significantly enhance the model's ability to adapt to changing environments, especially in noisy interference scenarios.

[0038] Unlike existing bearing fault diagnosis neural network models, the scheme of the present application directly trains based on the original one-dimensional vibration signal, without the need for data conversion or preprocessing to achieve excellent classification accuracy and noise resistance. The model has fewer parameters, lower computing power requirements, and better real-time performance, making it more suitable for embedded deployment on devices and enabling online monitoring of bearing faults. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 A step flowchart of the bearing fault diagnosis method based on the pulse neural network provided in Embodiment 1 of the present application.

[0040] Figure 2 A model architecture diagram of the pulse neural network constructed in Embodiment 1 of the present application.

[0041] Figure 3 A schematic diagram of the double-pulse attention module in the pulse neural network.

[0042] Figure 4 A schematic diagram of the deep separable pulse feedforward network in the pulse neural network.

[0043] Figure 5 Performance convergence curves of the present application on the CWRU and JNU data sets in the simulation test.

[0044] Figure 6 t-SNE visualization images of the present application and the control group in the simulation test. DETAILED DESCRIPTION

[0045] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present application and not to limit the present application.

[0046] Embodiment 1

[0047] The embodiment provides a bearing fault diagnosis method based on a pulse neural network. The method directly processes one-dimensional vibration signals collected through a constructed lightweight pulse neural network, and efficiently diagnoses bearing faults according to the timing characteristics of the vibration signals through pulse calculation. In the new pulse neural network constructed in the embodiment, the original input vibration signals are extracted for features in combination with a pulse coding layer; a self-attention mechanism is used to capture the dependency and interaction in the input signals, so that the model can focus on the most relevant area to achieve accurate diagnosis; and a deep separable convolution layer based on a neural network extracts rich feature representations from the input signals while improving the calculation efficiency. The model of the embodiment effectively combines the advantages of context-aware feature extraction and efficient multi-scale feature learning, improves the accuracy and efficiency of bearing fault diagnosis, and performs excellently in noise resistance, so that the real-time performance and robustness of the scheme can meet the practical requirements.

[0048] The bearing fault diagnosis method based on the pulse neural network provided in the embodiment can realize online monitoring of bearing faults in the running process of equipment according to the vibration signals of the bearing collected in real time. Specifically, as shown in Figure 1 the method comprises the following steps:

[0049] I. Constructing a pulse neural network for predicting the fault type of a bearing according to an input bearing vibration signal.

[0050] As shown in Figure 2 the pulse neural network in the embodiment sequentially comprises a feature extraction module, a multi-scale attention module and a classifier. The feature extraction module sequentially comprises a convolution layer, a max-pooling layer and a pulse coding layer based on a LIF neuron model; the feature extraction module is used for feature extraction, down-sampling and pulse trigger type transmission of the input bearing vibration signal, and then state features containing fault information are obtained. The multi-scale attention module comprises three attention branches of different scales, each branch sequentially comprises a double-pulse attention module with residual connection and a deep separable pulse feedforward network; the enhanced features weighted by the attention of the three branches are used as fusion features after feature concatenation. The classifier is used for generating a classification result from the fusion features through an average pooling layer and a fully connected layer.

[0051] To make the working principle and performance advantages of the spiking neural network constructed in this embodiment more apparent, the following sections provide a detailed explanation of the three main functional components of the spiking neural network: the feature extraction module, the multi-scale attention module, and the classifier.

[0052] 1.1 Feature Extraction Module

[0053] In this embodiment, the input bearing vibration signal s ( t ( ) is a discrete pulse signal of fixed length. In practical applications, the bearing vibration signal can be obtained from the raw vibration signal acquired by the accelerometer installed on the equipment. x ( t It is obtained by performing discrete encoding based on signal amplitude.

[0054] In this embodiment, the preset signal length in the signal encoding stage is 1024, therefore the bearing vibration signal is a one-dimensional signal containing 1024 points. The convolutional layer in the feature extraction module uses a 4×1 convolutional kernel to perform preliminary feature extraction on the 1×1024 input signal, and then performs downsampling through a 3×1 max pooling layer to obtain a 32×256 feature map. This signal processing stage can both enhance the network's ability to extract local features and reduce the computational complexity of the network model in the subsequent stages.

[0055] In practical applications, a normalization layer can be added before the max pooling layer to stabilize the feature distribution and help optimize the feature extraction effect of the pooling layer. The output of the max pooling layer is input to the pulse coding layer for processing. In this embodiment, the pulse coding layer adopts the leaky integral discharge (LIF) neuron model in the spiking neural network, which can simulate the behavior of neurons and process the input discrete pulse signals. When the membrane potential of any neuron in the spiking neural network exceeds a threshold, a pulse is triggered, and the timing of the pulse can be used to encode information.

[0056] In the pulse coding layer of this embodiment, the membrane potential update formula for neurons is:

[0057] ;

[0058] In the above formula, v l ( t -1) and v l ( t ) are respectively the first l The membrane potentials of layer neurons at time t-1 and time t; W l Represented as the first l The weights of interlayer synapses in layer neurons.s l-1 ( t ) is the pulse output of the previous layer neuron; θ l-1 represents the threshold value of the previous layer neuron; represents the decay caused by the membrane potential leakage; represents the membrane time constant, which is used to control the change of the decay rate of the membrane potential over time. This leakage term simulates the natural decay of the membrane potential in the neuron, reflecting the behavioral characteristics of the neuron.

[0059] Based on the updated membrane potential, the pulse output of the current layer neuron s l ( t ) is:

[0060] ;

[0061] In the above formula, H (·) represents the Heaviside step function, θ l represents the threshold value of the current layer neuron; geometrically, when the membrane potential v l ( t ) exceeds the threshold value θ l , the neuron will be triggered to produce a pulse, resulting in s l ( t )=1. When the neuron completes the pulse release, the membrane potential is reset to prepare for the accumulation of the next pulse, and the pulse output s l ( t ) is transmitted to the next layer to promote the propagation and communication of interlayer signals.

[0062] 1.2, Multi-scale attention module

[0063] In this embodiment, the multi-scale attention module includes three attention branches of different convolution kernels, each branch includes a double-pulse attention module with a residual connection and a depth separable pulse feedforward network in turn, that is: in each attention branch, the input features are first processed by the double-pulse attention module, and the input and output of the double-pulse attention module are connected in residual connection, and are jointly used as the input of the depth separable pulse feedforward network; after the input and output of the depth separable pulse feedforward network are connected in residual connection, the enhanced features extracted by the attention branch can be obtained.

[0064] The network structures of the three attention branches of the embodiment are the same, and the only difference is that the convolution kernel sizes of the convolution layers of the double-pulse attention modules are different. Specifically, in actual application, the convolution kernels of the double-pulse attention modules in the three attention branches of the embodiment are 16, 8 and 4 respectively. Among them, the fine convolution with a smaller scale focuses on capturing sudden vibration changes, the medium-scale convolution can extract local structure patterns, and the coarse convolution with a larger scale is better at aggregating long-range dependencies to maintain the key contextual associations required for fault pattern recognition under noise conditions. Therefore, after the state features output by the feature extraction module enter the multi-scale attention module, the three attention branches can extract attention information at different scales from the input state features and generate enhanced features containing different levels of feature information. The multi-scale attention module performs feature concatenation on the three enhanced features obtained, and the fusion features containing multi-scale feature information are obtained.

[0065] The double-pulse attention module in the embodiment realizes the attention mechanism through pulse-driven calculation, performs double-pulse transformation on the input features and calculates the attention score. The attention model uses convolution instead of traditional matrix multiplication to reduce the calculation overhead. The attention map is generated through the pulse neuron layer, and a scaling factor is introduced to adjust the amplitude of the attention score. Finally, the input features are weighted based on the attention map, and the feature matrix enhanced by attention is output.

[0066] Specifically, as shown in Figure 3 The double-pulse attention module of the embodiment is composed of one convolution layer (conv), three pulse coding layers (SN) and one batch normalization layer (BN). In the double-pulse attention module, the input features are divided into two paths, one of which is processed through the convolution layer, the pulse coding layer and the batch normalization layer in turn and then multiplied with the input of the other path. The feature map obtained after matrix operation is processed through the pulse coding layer to obtain the attention map. The attention map and the output of the batch normalization layer are multiplied again to obtain the feature, which is processed through the pulse coding layer again to obtain the output of the double-pulse attention module. Among them, the standard deviation of the matrix after each matrix multiplication operation is also taken as a scaling factor to adjust the amplitude of the attention score.

[0067] Specifically, the data processing process of the double-pulse attention module is as follows:

[0068]

[0069] In the above formula, F 1 represents the input feature of the double-pulse attention module; F 2 represents the output feature of the batch normalization layer; F 3 represents the attention map output by the pulse coding layer; F 4 represents the output feature of the double-pulse attention module;Conv (·) represents a convolution operation; BN(·) represents a batch normalization operation; SN(·) represents a data processing operation of a spike encoding layer; represents matrix multiplication.

[0070] In each attention branch, the features output by the double-spike attention module are input into a deep separable spike feedforward network for processing. As shown in FIG. 3, the deep separable spike feedforward network in this embodiment is composed of 3 spike encoding layers, 2 point convolution layers and 1 deep separable convolution module. The input features are sequentially input into the deep separable convolution module after being processed by the spike encoding layers, the point convolution layers and the spike encoding layers. After the input and output of the deep separable convolution are connected in residual, the output features of the deep separable spike feedforward network are obtained after being processed by the spike encoding layers and the point convolution layers, respectively. Figure 4

[0071] Specifically, the data processing process of the deep separable spike feedforward network is as follows:

[0072]

[0073] In the above formula, Conv1x1(·) represents a point convolution operation; DSC(·) represents a deep separable convolution operation; F 5 represents the output features of the second spike encoding layer in the deep separable spike feedforward network; F 6 represents the output features of the deep separable spike feedforward network.

[0074] In this embodiment, the enhanced feature matrices finally output by the three attention branches are denoted as , and respectively. Then, the fusion features containing multi-scale feature information extracted by the multi-scale attention module can be obtained by concatenating the features in the three channels. Due to the sparse and asynchronous characteristics of spike signals, directly applying the self-attention mechanism which relies on dense activation may lead to instability. By selectively enhancing information activation signals and suppressing redundant spikes, the calculation resources are dynamically allocated to relevant events, thereby optimizing the spike representation. The outputs of the three paths are fused by using a channel-level feature concatenation strategy, which not only strengthens the shared discriminative features but also reduces noise interference.

[0075] 1.3, Classifier

[0076] ​Based on the fusion features extracted by the multi-scale attention module, the embodiment realizes fault classification through a classifier. Specifically, the classifier in the embodiment is composed of an average pooling layer and a fully connected network. The average pooling layer is used to reduce the size of the feature map, reduce the computational complexity of the subsequent convolution layer, and highlight the global statistical features. The full connection is used to map the feature image to the final classification result. In the embodiment, the fault types of the bearing output by the classifier in the pulse neural network include: different sizes (such as 0.18mm, 0.36mm, 0.54mm) of inner ring fault (IR), outer ring fault (OR), rolling element fault (B) and other bearing defects, and normal state.

[0077] II. Obtain original vibration signals of different fault types, pre-process them and add fault labels as sample data; and use them to supervise the training of the constructed pulse neural network.

[0078] In the embodiment, the corresponding original vibration signals can be collected under a plurality of different working conditions, and the sampling rate of the signals is set as needed; the collected original vibration signals contain a plurality of fault types. According to the collected sample data, the following pre-processing is performed respectively:

[0079] (i) The Z-score normalization method is used to unify the distribution of all original signal data. That is, the original signal is adjusted by the signal mean and standard deviation, so that the processed data has a uniform distribution feature.

[0080] (ii) The original signal is segmented according to a fixed length to obtain a sample signal. In the embodiment, the segmented sample signal includes 1024 data points. The discrete signal is used as sample data for model training and inference.

[0081] (iii) Gaussian white noise with a signal-to-noise ratio range of -4dB to 4dB is added to the sample signal to increase the data; wherein the added Gaussian white noise conforms to the normal distribution characteristics with a mean of 0 and a variance of a specific value. In order to improve the robustness and generalization of the model, the original signal is added with white noise to realize signal enhancement, and the enhanced sample data and the original sample data have the same fault label.

[0082] Next, the sample data set containing a large amount of sample data is divided into a training set, a test set and a validation set; the cross-entropy loss is used as the loss function, and the surrogate gradient algorithm is used to supervise the training of the pulse neural network; during the training process, the average result is obtained by independent training for multiple times (10 times) to reduce the randomness. After the training is completed, the weight parameters of the model that meets the performance requirements are retained.

[0083] Particularly, during the training phase of the spiking neural network, the spiking generation process in the spiking coding layer is non-differentiable, which hinders the training process. The embodiment introduces a surrogate gradient method to approximate the gradient of the spiking activation function during the backpropagation phase, so that the gradient-based learning method can be effectively applied. Specifically, during the training phase, the surrogate gradient algorithm calculates the spiking output of the neurons in the spiking coding layer through a continuously differentiable surrogate gradient function; and the gradient derived by the surrogate activation function is used to update the weights.

[0084] wherein the surrogate gradient function is as follows:

[0085] ;

[0086] In the above formula, denotes the partial derivative symbol; denotes the derivative of the surrogate activation function , which adopts a sigmoid function or a smooth function of a soft threshold.

[0087] The weight update rule is:

[0088]

[0089] In the above formula, Δ W l denotes the updated weight of the inter-layer synapse of the neurons in the l layer, L denotes the loss function, η denotes the learning rate. is the gradient of the loss function with respect to the synaptic weight, which is obtained by surrogate gradient calculation. The introduction of surrogate gradient bridges the gap between the spiking characteristics of the spiking neural network and the gradient optimization used in traditional artificial neural networks; it can make the training of the spiking neural network more efficient and can be applied to practical tasks.

[0090] III. Real-time acquisition of bearing vibration signals during the running process of the target bearing, which are input into the pre-trained spiking neural network to generate a prediction result of the fault type.

[0091] In the actual application phase of the embodiment, the trained spiking neural network model can be deployed to edge computing devices, including microcontrollers, single-chip microcomputers, ARM chips, embedded Linux systems, or FPGA hardware platforms. Thanks to the lightweight characteristics of the model (parameters are only 0.67M), the scheme of the embodiment can adapt to resource-constrained platforms such as microcontrollers and FPGAs, and its event-driven characteristics can reduce energy consumption, making it suitable for real-time industrial scenarios.

[0092] Embodiment 2

[0093] ​The bearing fault diagnosis method based on the pulse neural network provided in Embodiment 1 is essentially a data processing method. In order to better apply the scheme, the embodiment further provides a storage medium, a computer program product and a bearing fault self-monitoring device.

[0094] Specifically, the storage medium provided in the embodiment stores a computer program. When the computer program is executed by a processor, a pulse neural network trained in the bearing fault diagnosis method based on the pulse neural network in Embodiment 1 is created, and is used for fault prediction on an input bearing vibration signal.

[0095] The computer program product provided in the embodiment includes a computer program. When the computer program is executed by a processor, a pulse neural network trained in the bearing fault diagnosis method based on the pulse neural network in Embodiment 1 is created, and is used for fault prediction on an input bearing vibration signal.

[0096] The bearing fault self-monitoring device provided in the embodiment includes an acceleration sensor and an analyzer. The acceleration sensor is installed at a bearing on a device, and is used for collecting an original vibration signal of a target bearing. The analyzer includes a memory, a processor and a computer program stored in the memory and running on the processor. When the processor executes the computer program, a pulse neural network trained in the bearing fault diagnosis method based on the pulse neural network in Embodiment 1 is created. The analyzer is used for receiving the original vibration signal collected by the acceleration sensor in real time, converting the original vibration signal into a bearing vibration signal and inputting the bearing vibration signal into the pulse neural network to realize fault prediction.

[0097] The analyzer in the bearing fault self-monitoring device is essentially a computer device. In actual application, the analyzer can be an embedded device, and then directly realizes fault prediction according to the vibration signal collected in front end. The analyzer can also be an independent computer device, such as a notebook computer, a tablet computer, a desktop computer or a medium or large computer device capable of executing a computer program, such as a rack server, a blade server, a tower server or a cabinet server (including an independent server, or a server cluster composed of multiple servers). Then, the analyzer realizes processing on vibration signals of different sources output in front end respectively, and realizes fault analysis.

[0098] The computer device of the embodiment at least includes, but is not limited to, a memory and a processor which can be connected to each other in communication through a system bus. In the embodiment, the memory (i.e., a readable storage medium) includes a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the memory can be an internal storage unit of the computer device, such as a hard disk or a memory of the computer device. In other embodiments, the memory can also be an external storage device of the computer device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the computer device. Of course, the memory can include both the internal storage unit and the external storage device of the computer device. In the embodiment, the memory is generally used to store an operating system and various application software installed on the computer device, etc. In addition, the memory can also be used to temporarily store various data that have been output or will be output.

[0099] The processor can be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip in some embodiments. The processor is generally used to control the overall operation of the computer device.

[0100] Simulation test

[0101] In order to verify the performance of the bearing fault diagnosis method based on the pulse neural network provided by the application, the technical personnel simulate and train the related scheme, and test the performance of the trained related network model.

[0102] 1. Data set

[0103] This study uses two independent data sets to support the experimental results. Among them, data set 1 (CWRU) is derived from the bearing test bench of Case Western Reserve University to provide vibration signals required for fault diagnosis under controlled experimental conditions. The data is collected from motor bearings with a working load range of 0 to 3 horsepower and a rotational speed of 1720 to 1797 revolutions per minute. The vibration signals are recorded at a sampling frequency of 12 kHz and 48 kHz, covering normal and fault states. This experiment uses 12 kHz sampling data, including normal state and signal under fault states such as inner ring fault, outer ring fault and ball fault. Each fault type contains three defect sizes of 0.18 mm, 0.36 mm and 0.54 mm, forming ten different operating conditions.

[0104] Dataset 2 (JNU Dataset) contains vibration signals collected from faulty bearings and normal bearings under different operating conditions. The signal sampling frequency is 51.2 kHz, and it contains three types of faults: inner race fault, outer race fault, and ball fault, each with different damage severity. This dataset presents a variety of complex fault modes and is a valuable benchmark for evaluating fault diagnosis models.

[0105] 2. Experimental content

[0106] To comprehensively evaluate the effectiveness of the pulse neural network (referred to as DSRSNet) provided in this embodiment, the performance of this embodiment is compared with the state-of-the-art models commonly used in the field of fault diagnosis. The benchmark models include traditional deep learning architectures and similar architectures of pulse neural networks to achieve balanced evaluation. The control group schemes include:

[0107] SEW-ResNet: This model is a widely used SNN model that integrates alternative gradient learning techniques and is a strong representative of SNN based on direct training. By comparing it with the scheme of the present application, the effectiveness of the structural improvement of the scheme of the present application in the end-to-end spike training paradigm can be analyzed.

[0108] Spiking ResNet-18: This network model is a synaptic version of ResNet-18, which replaces the ReLU activation function with a leaky integrate-and-fire (LIF) neuron while retaining the overall structure. Using this model as a benchmark, the effectiveness of residual learning based on synaptic neurons can be evaluated.

[0109] MCNN-AM: This scheme is a converted pulse neural network that first trains an artificial neural network (ANN) and then converts it into a pulse neural network. This comparison helps to evaluate the differences between direct training and ANN-to-spike network in the fault diagnosis task.

[0110] LiconyFormer: LiconvFormer is a hybrid model that combines convolutional layers and Transformer attention mechanisms. It captures local features through convolutional modules and models global dependencies with the help of Transformers.

[0111] During the training phase of the network model, considering that in actual industrial environments, vibration signals are often contaminated by various noises. This embodiment enhances the sample data by filling in Gaussian white noise. Gaussian white noise is a Gaussian distribution random signal with a mean of zero, and the power spectral density of this signal is uniformly distributed at all frequencies.

[0112] During the training process, the hyperparameters were set as follows: batch size of 64, maximum of 100 epochs, and learning rate of 0.005. The experiment was conducted in a PyTorch 1.8 and Python 3.8 environment, with hardware configuration of Intel(R) Core i9-14900K CPU 6.0GHz paired with RTX3090 GPU.

[0113] 3. Experimental data and analysis

[0114] (3.1) Performance of the model of the present application

[0115] The present experiment first tested the convergence process of the performance of the present application on two data sets, and all results were based on the average value of ten independent runs, with 95% confidence interval. The experimental results are shown in Figure 5

[0116] From the data in the graph, it can be seen that on the CWRU and JNU data sets, the training accuracy and validation accuracy curves of the present application showed a clear convergence trend with the increase of the number of training epochs. The minimum difference between the training accuracy and the validation accuracy indicates that the model effectively avoids overfitting and maintains strong generalization ability under different working conditions and fault types. In addition, the consistent training dynamics across data sets highlight the robustness of the present application, which is due to the integration of its directly trained spike architecture and attention mechanism, enabling reliable and adaptive feature learning in various diagnostic scenarios.

[0117] (3.2) Performance comparison

[0118] The present experiment further compared the performance of the present application and the control group on two data sets in terms of accuracy (Accuracy) and model parameter quantity. The experimental results are shown in the following table;

[0119] Table 1: Performance comparison of the present application and the control group

[0120]

[0121] ​Analyzing the data in the above table can find that the network course model of the application achieves excellent diagnostic accuracy on the CWRU and JNU data sets, reaching 100% and 96.78% respectively, while maintaining a low parameter number of only 0.67M. Compared with SEW-ResNet and Spiking ResNet, which are directly trained synaptic neural networks, the application scheme not only has higher accuracy, but also shows better model compactness. Although MCNN-AM achieves perfect accuracy on the CWRU data set, its performance on the JNU data set decreases significantly, indicating that its generalization ability is limited. LiconvFormer shows competitive accuracy and the smallest model size, but its architecture is based on artificial neural networks rather than synaptic neural networks. These comparisons highlight the effective balance between diagnostic performance and model efficiency of DSRSNet, confirming its advantages in achieving accurate and lightweight fault diagnosis under diverse working conditions.

[0122] In addition, the application scheme shows excellent diagnostic performance on the CWRU and JNU data sets, with classification accuracy of 100% and 98.78% respectively. In contrast, the two representative directly trained synaptic neural network models, SEW-ResNet and Spiking ResNet, have an accuracy of 96.03% and 95.35% on the CWRU data set, and 94.77% and 92.83% on the JNU data set, respectively, which is slightly inferior. This shows that the application scheme not only maintains competitive accuracy in noise-free and complex environments, but also exhibits stronger generalization ability than its SNN counterparts. In terms of model complexity, the application scheme contains only 670,000 parameters, much less than SEW-ResNet (4.12 million) and Spiking ResNet (3.85 million), but outperforms them in accuracy. Although LiconvFormer achieves strong diagnostic accuracy with a very low parameter overhead (0.32M), the application scheme achieves comparable performance with a slightly larger model size (0.67M) and further leverages the inherent advantages of directly trained neural networks, including excellent adaptability for deployment on brain-like platforms.

[0123] Although MCNN-AM achieves 100% accuracy on the CWRU data set, its performance on the JNU data set decreases significantly to 90.42%, reflecting its limited adaptability to changes in data distribution. In addition, MCNN-AM is a conversion type SNN method, which usually needs longer time steps to match the performance of the original ANN, and lacks the inherent temporal coding advantages of directly trained SNNs.

[0124] (3.2) Visualization

[0125] The experiment further visualizes the two-dimensional t-SNE images of the CWRU dataset for the present application (DSRSNet) and the control group scheme (SEW-ResNet and Spiking ResNet), and the results are as shown in Figure 6

[0126] It can be found from the data in the analysis graph that the cluster distribution generated by the present application scheme is more dispersed and the boundary is clearer, which indicates that it forms highly compact and clearly separated clusters for each type of fault, and the scheme has excellent ability in extracting discriminative features and distinguishing different fault types. In contrast, SEW-ResNet and Spiking ResNet present overlapping and fragmented clusters, and the boundaries between classes are blurred. This dispersion phenomenon indicates that its feature extraction capability is limited, which may affect the classification performance in complex or noisy environments. These observations are consistent with the previous quantitative results, further confirming that the present application scheme not only achieves higher accuracy in bearing fault diagnosis tasks, but also has enhanced robustness and generalization ability.

[0127] The above merely describes preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.​

Claims

1. A bearing fault diagnosis method based on a spiking neural network, characterized by, It comprises: constructing a spiking neural network for predicting the fault type of a bearing according to an input bearing vibration signal; The spiking neural network sequentially comprises a feature extraction module, a multi-scale attention module and a classifier; the feature extraction module sequentially comprises a convolution layer, a max-pooling layer and a spiking coding layer based on a LIF neuron model, and is used for feature extraction, down-sampling and spiking trigger type transmission of the input bearing vibration signal to obtain state features; the multi-scale attention module comprises three attention branches, each branch sequentially comprises a double-spiking attention module with residual connection and a deep separable spiking feedforward network with residual connection, and the convolution kernels of the double-spiking attention module are 16, 8 and 4 respectively; The enhanced features weighted by the attention extracted through the three branches are spliced as fusion features after being spliced; the classifier is used for generating a classification result according to the fusion features through an average pooling layer and a fully connected layer; Obtain the original vibration signal of different fault types, and add a fault label after preprocessing to obtain sample data; divide the sample data set containing a large amount of sample data into a training set, a test set and a validation set; use cross-entropy loss as a loss function, and use a surrogate gradient algorithm to supervise the training of the spiking neural network, and retain the model parameters that meet the performance requirements; Real-time acquisition of bearing vibration signals during the operation of the target bearing, inputting the bearing vibration signals into the pre-trained spiking neural network to generate a prediction result of the fault type; In the double-spiking attention module, the input features are divided into two paths, one of which is processed by a convolution layer, a spiking coding layer and a batch normalization layer in sequence, and then multiplied by the input of the other path to obtain a feature map feature, which is processed by a spiking coding layer to obtain an attention map; the output of the batch normalization layer is multiplied by the attention map again to obtain a feature, which is processed by a spiking coding layer to obtain the output of the double-spiking attention module; wherein, the standard deviation of the matrix after each matrix multiplication operation is used as a scaling factor to adjust the amplitude of the attention score; The deep separable spiking feedforward network is composed of 3 spiking coding layers, 2 point convolution layers and 1 deep separable convolution module; the input features are processed by a spiking coding layer, a point convolution layer and a spiking coding layer in sequence, and then input into the deep separable convolution module; the input and output of the deep separable convolution are connected in residual, and then processed by a spiking coding layer and a point convolution layer respectively to obtain the output features of the deep separable spiking feedforward network.

2. The pulse neural network based bearing fault diagnosis method of claim 1, wherein: The membrane potential update formula of the spiking coding layer is: ； In the above formula, v l ( t -1) and v l ( t ) are respectively the first l The membrane potentials of layer neurons at time t-1 and time t; W l Represented as the first l The weights of interlayer synapses in layer neurons. s l-1 ( t () represents the pulse output of the front layer neurons; This represents the threshold of neurons in the front layer; Indicates the membrane time constant; a pulse output of a neuron of a current layer s l ( t ) is: In the above formula, H (·) denotes the Heaviside step function, denotes the threshold value of the current layer neuron.

3. The pulse neural network based bearing fault diagnosis method of claim 2, wherein: In the training stage, the surrogate gradient algorithm calculates the neuron pulse output in the spiking coding layer through the following continuously differentiable surrogate gradient function; and the gradient derived from the surrogate activation function is used to update the weight: ； In the above formula, denotes the partial derivative symbol; denotes the derivative of the alternative activation function sigmoid function or a soft thresholding smooth function;​ The weight update rule is: In the above formula, denotes the updated first l weights of the inter-layer synapses of the layer neurons, L denotes the loss function, denotes the learning rate.

4. The pulse neural network based bearing fault diagnosis method of claim 1, wherein: The fault type of the bearing predicted by the spiking neural network includes: inner ring fault, outer ring fault, rolling element fault and normal state of different sizes.

5. The pulse neural network based bearing fault diagnosis method of claim 1, wherein: The preprocessing method of the sample data comprises the following steps: (i) using Z-score normalization method to unify the distribution of all original signal data; (ii) segmenting the original signal according to a fixed length to obtain a sample signal; (iii) adding Gaussian white noise with a signal-to-noise ratio ranging from -4dB to 4dB to the sample signal to achieve data increase, wherein the added Gaussian white noise conforms to normal distribution characteristics with a mean of 0 and a variance of a specific value.

6. A storage medium having stored therein a computer program, characterized in that The computer program, when executed by a processor, creates the pulse neural network trained in the bearing fault diagnosis method based on the pulse neural network according to any one of claims 1-5, and is used for fault prediction on an input bearing vibration signal.

7. A computer program product comprising a computer program, characterised in that, The computer program, when executed by a processor, creates the pulse neural network trained in the bearing fault diagnosis method based on the pulse neural network according to any one of claims 1-5, and is used for fault prediction on an input bearing vibration signal.

8. A self-monitoring device for bearing failure, characterized by: The device comprises an acceleration sensor and an analyzer; the acceleration sensor is installed at a bearing on the device and is used to collect original vibration signals of the target bearing; the analyzer comprises a memory, a processor, and a computer program stored in the memory and running on the processor; when the processor executes the computer program, a pulse neural network trained in the bearing fault diagnosis method based on the pulse neural network according to any one of claims 1-5 is created; the analyzer is used to receive the original vibration signals collected by the acceleration sensor in real time, convert the original vibration signals into bearing vibration signals, and input the bearing vibration signals into the pulse neural network to achieve fault prediction.

Images