Bearing fault detection method, system, device and storage medium based on pulse neural network

Abstract

The present application relates to the technical field of bearing fault detection, and in particular to a bearing fault detection method, system, device and storage medium based on a pulse neural network. The present application aims to overcome the defects of the existing methods for diagnosing rolling bearing faults, which cannot meet various types and various operating conditions, such as bearing detection under different rotating speeds and loads, or require a large amount of sample data to train a fault diagnosis model, and cannot be applied to small sample situations. The present application provides a bearing fault detection method based on a pulse neural network. The method uses variational mode decomposition as a sample data decomposition method, and selects 10 characteristic indicators to describe the characteristic information of the corresponding sample data. At the same time, the Tempotron learning algorithm is used to learn the pulse neural network model in the present application, thereby solving the above problems.

Description

Technical Field

[0001] This invention relates to the field of bearing fault detection technology, and specifically to a bearing fault detection method, system, device, and storage medium based on a spiking neural network. Background Technology

[0002] As mechanical equipment rapidly develops towards larger scale, integration, precision, and intelligence, the connectivity and coupling between different components within the same equipment and between different devices are becoming increasingly complex. When one part of a mechanical device malfunctions, it can lead to serious problems such as downtime and damage to the entire device. Therefore, maintaining the safe and stable operation of mechanical equipment has become an important research issue in the scientific and industrial communities.

[0003] Rolling bearings are crucial components in rotating machinery, and their reliable and stable operation plays a vital role in ensuring the normal functioning of the entire machine. Failures in rotating machinery caused by rolling bearing malfunctions are commonplace in manufacturing. Therefore, effective health monitoring and fault diagnosis of key components, including rolling bearings, can provide a strong basis for the daily maintenance and management of machinery, thereby effectively reducing economic losses and even personal injury consequences caused by equipment failures.

[0004] Currently, technologies in the field of rolling bearing fault diagnosis can be mainly divided into three aspects: physical model-based diagnostic methods, expert experience-based diagnostic methods, and data-driven diagnostic methods. Physical model-based diagnostic techniques have high accuracy, but the modeling process is complex, and a single model is only effective for a few specific types of bearings, unable to be extended to other bearings. Expert experience-based methods face the same problem. Data-driven fault detection methods effectively solve the problems of the first two types of methods, but because the original fault signals contain a lot of noise, they are often not directly usable. Therefore, data-driven fault diagnosis methods mainly extract relevant feature information from the bearing fault signals and then design a reasonable fault mode classifier for fault diagnosis. Thus, this type of method requires a large amount of sample data to train the fault diagnosis model. However, in actual mechanical equipment, it is often difficult to obtain a large amount of sample data, especially fault data. This makes it difficult for many data-driven fault detection methods to be truly applied to practical fault diagnosis. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing rolling bearing fault diagnosis methods, which cannot meet the needs of various models and operating conditions, such as bearing testing at different speeds and loads, or the need for a large amount of sample data to train the fault diagnosis model, making them unsuitable for small sample situations.

[0006] To achieve the above objectives, the technical solution adopted by the present invention to solve its technical problem is as follows:

[0007] This paper presents a bearing fault detection method based on a spiking neural network. This method uses a spiking neural network to detect faults in rolling bearings, employs variational mode decomposition as the decomposition method for sample data, and selects 10 feature indicators to describe the feature information of the corresponding sample data. Simultaneously, it uses the Tempotron learning algorithm to learn the spiking neural network model described in this application. This solves the problem that existing rolling bearing fault diagnosis methods cannot meet the needs of various bearing models and operating conditions, such as bearings with different speeds and loads, or require a large amount of sample data to train the fault diagnosis model, making them unsuitable for small sample situations. Details are as follows:

[0008] The bearing fault detection method based on spiking neural networks is characterized by comprising the following steps:

[0009] S1. Vibration signal feature extraction

[0010] Feature extraction is performed on the original vibration signal to obtain feature indices, thereby reducing the noise and dimensionality of the original vibration signal;

[0011] S2. Fault Detection

[0012] A spiking neural network is constructed, comprising two layers. The first layer acts as an encoding neuron, which encodes feature indicators with pulses. The second layer acts as a spiking neuron, which determines the bearing fault type based on the pulses transmitted from the neurons in the first layer.

[0013] Furthermore, step S1 specifically includes:

[0014] S11. The original vibration signal is segmented and each segment of the original vibration signal is taken as a sample data.

[0015] S12. After obtaining the sample data obtained in step S11, the particle swarm optimization algorithm is applied to the sample data, where the average minimum envelope entropy is used as the fitness function. Finally, the two important parameters required for the variational mode decomposition of the sample are obtained, including the number of decompositions and the penalty factor.

[0016] S13. After obtaining the number of decompositions and the penalty factor obtained in step S12, perform variational mode decomposition on the sample data to obtain the components;

[0017] S14. For the original vibration signal and each component, calculate its corresponding 10 characteristic indices. For each sample data, obtain (k+1)*10 characteristic indices, where k is the number of variational mode decompositions corresponding to each sample data.

[0018] Furthermore, the 10 feature indicators in step S14 are: skewness factor, peak factor, waveform factor, margin factor, impulse factor, kurtosis factor, approximate entropy, sample entropy, permutation entropy, and fuzzy entropy.

[0019] Furthermore, the encoding process of the first layer of encoding neurons in step S2 includes: pulse encoding the (k+1)*10 feature indicators obtained in step S14 using a pulse encoding method based on Gaussian harmonic curves.

[0020] Furthermore, in step S2, the second layer of spiking neurons uses the Tempotron learning algorithm to learn the parameters of the neural network. After learning, each neuron can distinguish two different spiking modes. When one mode is input, the maximum membrane potential of the neuron will exceed the threshold, causing the neuron to fire a spiking signal. When the other mode is input, the maximum membrane potential of the neuron will not exceed the threshold, causing the neuron to not fire a spiking signal.

[0021] A bearing fault detection system based on a spiking neural network, characterized in that the system comprises:

[0022] The feature extraction module is used to extract features from the original vibration signal, obtain feature indicators, and reduce the noise and dimensionality of the original vibration signal.

[0023] The fault detection module is used to perform pulse coding and fault type detection on the extracted feature indicators through a spiking neural network.

[0024] Furthermore, the feature extraction module includes:

[0025] The signal interception unit is used to segment the original vibration signal and treat each segment of the original vibration signal as a sample data.

[0026] The particle swarm optimization (PSO) unit is used to apply the PSO algorithm to the sample data obtained by the signal interception unit. The average minimum envelope entropy is used as the fitness function. Finally, two important parameters required for variational mode decomposition of the sample are obtained, including the number of decompositions and the penalty factor.

[0027] The variational mode decomposition unit is used to perform variational mode decomposition on the sample data after the particle swarm algorithm unit obtains the number of decompositions and the penalty factor, so as to obtain the components.

[0028] The feature index acquisition unit is used to calculate 10 feature indices for each component of the original vibration signal. For each sample data, (k+1)*10 feature indices are obtained, where k is the number of variational mode decompositions corresponding to each sample data.

[0029] The 10 feature indicators in the feature indicator acquisition unit are: skewness factor, peak factor, waveform factor, margin factor, impulse factor, kurtosis factor, approximate entropy, sample entropy, permutation entropy, and fuzzy entropy.

[0030] Furthermore, the spiking neural network in the fault detection module includes an encoding layer and a spiking layer;

[0031] The pulse coding process of the coding layer includes: pulse coding the (k+1)*10 feature indicators obtained by the feature indicator acquisition unit using a pulse coding method based on Gaussian harmonic curves;

[0032] The pulse layer uses the Tempotron learning algorithm to learn the parameters of the neural network. After learning, each neuron can distinguish two different pulse modes. When one mode is input, the maximum membrane potential of the neuron will exceed the threshold, causing the neuron to fire a pulse. When the other mode is input, the maximum membrane potential of the neuron will not exceed the threshold, causing the neuron not to fire a pulse.

[0033] An electronic device, including at least one processor;

[0034] and a memory communicatively connected to the at least one processor;

[0035] The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method described above.

[0036] A non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause the computer to perform the method described above.

[0037] The beneficial effects of this invention are:

[0038] This invention provides a bearing fault detection method based on a spiking neural network (SNN). The SNN is used to detect faults in rolling bearings. During information processing, the information is encoded into a sequence of pulses, and only a small portion of these pulses are stimulus states, significantly reducing the computational complexity of the SNN and giving it low energy consumption. Therefore, using SNNs for rolling bearing fault detection not only provides a novel technical method for the field of bearing fault detection but also offers more references and insights for the transition from theoretical research to practical application of SNNs.

[0039] Compared with other rolling bearing fault diagnosis methods, the bearing fault detection method based on spiking neural networks provided by this invention firstly improves the accuracy of fault diagnosis; secondly, this invention has good generalization ability, and can effectively identify faults with different speeds, loads and damage widths, and can be applied to different rolling bearings; finally, this invention also has good recognition effect in experimental verification for small sample data. Attached Figure Description

[0040] Figure 1 This is a fault diagnosis framework diagram based on spiking neural networks;

[0041] Figure 2 This is a flowchart of vibration signal feature extraction;

[0042] Figure 3 This is a graph showing the recognition accuracy under different rotation speeds and training sets;

[0043] Figure 4 This is a comparison chart of the recognition accuracy of the technical solution presented in this paper with other methods;

[0044] Figure 5 This is a box plot comparing the identification results of the technical solution presented in this paper with those of other methods;

[0045] Figure 6 These are images showing the recognition results under different rotation speeds and training sets;

[0046] Figure 7 This is a graph showing the recognition results under different loads and training sets;

[0047] Figure 8 This is a comparison chart of the recognition accuracy of the technical solution presented in this paper with other methods;

[0048] In the figure: (a) PSO-VMD-SNN method, (b) the technical solution of this application. Detailed Implementation

[0049] The following will refer to the appendix. Figure 1-8 Exemplary embodiments of this disclosure will be described in more detail below. While exemplary embodiments of this disclosure are shown in the accompanying drawings, it should be understood that this disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.

[0050] Example:

[0051] I. Specific method of the present invention:

[0052] like Figure 1As shown, the technical solution of this application is divided into two parts. The first part is feature extraction for vibration signals. Because the original vibration signals contain a lot of noise, directly using them would first reduce the effect of model training, and secondly, due to the large amount of original data, directly using them would increase the computational complexity. Therefore, this application chooses to extract features from the original data to reduce the noise and dimensionality of the original data. This technology chooses to use variational mode decomposition (VMD) to decompose the bearing vibration signals. VMD is a novel adaptive signal processing method. The decomposition process of VMD is the solution process of a variational problem. This method determines the frequency center and bandwidth of each component by iteratively searching for the optimal solution of the variational model, thereby realizing the mode decomposition of the signal. VMD does not require specifying the basis functions to be decomposed like wavelet decomposition and short-time Fourier transform, and it also makes up for the problems of mode aliasing and unclear discontinuities caused by empirical mode decomposition. Since VMD requires specifying the number of decompositions and penalty parameters, this application uses particle swarm optimization (PSO) to find the number of decompositions and penalty parameters corresponding to each signal. Particle swarm optimization (PSO) is a swarm intelligence optimization algorithm. Compared to other intelligent optimization algorithms, it features fast search speed, fewer parameters, and a simple algorithm structure, making it very convenient to apply. Since envelope entropy can describe the sparsity of a signal, this application selects the average minimum envelope entropy as its fitness function to find a suitable number of variational mode decompositions and a penalty factor.

[0053] First, the vibration signal obtained in this application is segmented, and each segment of the vibration signal is used as a sample data. For example... Figure 2As shown, the particle swarm optimization algorithm is applied to the obtained sample data, using the average minimum envelope entropy as the fitness function to find a suitable number of decompositions and a penalty factor. After obtaining the two parameters of the number of decompositions and the penalty factor, this application performs variational mode decomposition on the sample data. Although the decomposed components solve the problem of a large amount of noise in the original signal, directly performing pulse coding on these components would greatly increase the computational complexity of the pulse neural network in this application. Therefore, in the technical solution of this application, for the original vibration signal and each component, 10 corresponding feature indices are calculated, including skewness factor, peak factor, waveform factor, margin factor, impulse factor, kurtosis factor, approximate entropy, sample entropy, permutation entropy, and fuzzy entropy. The first 6 indices are dimensionless feature values, often used to detect and measure whether there is impact in the signal and the wear of mechanical equipment. The latter 4 are entropy values ​​with different definitions, mainly used to measure the degree of disorder in the signal. For signals with different fault types, the frequencies and waveforms presented in the signals are different, and the entropy values ​​can reflect the differences between different signals. Therefore, choosing the aforementioned feature indices to replace the original vibration data and components can more clearly describe the features contained in the signal and greatly reduce the computational complexity of the spiking neural network. Based on this, for each sample data, this application can obtain (k+1)*10 feature indices, where k is the number of variational mode decompositions corresponding to each sample data. Since the number of decompositions varies for each sample data in actual decomposition, but in the experiments of this application, only the majority of sample data have the same number of decompositions k. To utilize as many sample data as possible, this application selects sample data with the same number of decompositions and the largest proportion of all decompositions. For sample data with a greater than 100 decompositions, this application selects the first k components of that sample data and the original signal; for sample data with a less than 100 decompositions, this application discards them directly. In reality, the proportion of the ultimately discarded sample data is very small, and there is no waste of data.

[0054] The second part is to perform fault detection on the vibration signal features extracted in the first part. After obtaining the feature information of the sample data, this application first performs pulse coding on (k+1)*10 feature indicators. The technical solution of this application selects the pulse coding method based on Gaussian harmonic curve. This is a group coding method. In the spiking neural network of this application, the neurons of the first layer are equivalent to a group of Gaussian functions. Each neuron covers a certain coding range in the form of a Gaussian function. Multiple overlapping Gaussian functions are used to encode the feature information, thereby encoding the feature information into a group of pulse sequences. In the technical solution of this application, the spiking neural network uses the Tempotron

[10] learning algorithm to learn the parameters of the neural network. The Tempotron algorithm is a simple and easy-to-implement supervised learning algorithm. It trains the weight parameters of the spiking neural network so that each neuron can distinguish two different pulse modes (here marked as "+" and "-" respectively). After the spiking neural network is trained, when one mode is input (assuming it is "+"), the maximum membrane potential of the neuron will exceed the threshold, causing the neuron to fire a pulse. When the other mode (assuming it is "-") is input, the maximum membrane potential of the neuron will not exceed the threshold, causing the neuron to not fire a pulse. This application uses a two-layer spiking neural network. The first layer acts as an encoding neuron, encoding the feature indicators of this application using pulses. The second layer acts as a spiking neuron, determining the fault type of the sample data based on the pulses transmitted from the previous layer. The first part of this technical solution is implemented using Matlab, and the second part is implemented using the PyTorch framework.

[0055] In summary, the method steps of this invention include:

[0056] S1. Vibration signal feature extraction

[0057] Feature extraction is performed on the original vibration signal to obtain feature indices, thereby reducing the noise and dimensionality of the original vibration signal;

[0058] S11. The original vibration signal is segmented and each segment of the original vibration signal is taken as a sample data.

[0059] S12. After obtaining the sample data obtained in step S11, the particle swarm optimization algorithm is applied to the sample data, where the average minimum envelope entropy is used as the fitness function. Finally, the two important parameters required for the variational mode decomposition of the sample are obtained, including the number of decompositions and the penalty factor.

[0060] S13. After obtaining the number of decompositions and the penalty factor, perform variational mode decomposition on the sample data to obtain the components;

[0061] S14. For the original vibration signal and each component, calculate its corresponding 10 characteristic indices. For each sample data, obtain (k+1)*10 characteristic indices, where k is the number of variational mode decompositions corresponding to each sample data.

[0062] S2. Fault Detection

[0063] A spiking neural network is constructed, comprising two layers. The first layer acts as an encoding neuron, performing pulse encoding on the feature indicators. The construction process of the first layer includes pulse encoding the (k+1)*10 feature indicators obtained in step S14 using a pulse encoding method based on Gaussian harmonic curves. The second layer acts as a spiking neuron, determining the bearing fault type based on the pulses transmitted from the first layer neurons. The second layer spiking neurons use the Tempotron learning algorithm to learn the parameters of the neural network. After learning, each neuron can distinguish two different pulse modes. When one mode is input, the neuron's maximum membrane potential exceeds a threshold, causing the neuron to fire a pulse; when the other mode is input, the neuron's maximum membrane potential does not exceed the threshold, causing the neuron not to fire a pulse.

[0064] II. Comparative Experiment:

[0065] To verify the effectiveness of the technical solution of this application, experimental verification was conducted on two types of datasets. The first type is experimental data of the drive-end 6205 SKF deep groove ball bearing provided by the Bearing Data Center of Case Western Reserve University (CWRU). This data is bearing vibration acceleration data, including vibration signals of normal operation, inner ring, outer ring, and rolling element failure, collected by an accelerometer mounted on a magnetic base housing, with a sampling frequency of 12 kHz. The second type of data is the bearing normal operation, inner ring failure, and outer ring failure signals simulated by this application based on the HD-FD-H-03X rotor rolling bearing and gearbox integrated fault test bench. The signals are the bearing vibration acceleration signals in the horizontal direction.

[0066] Table 1. CWRU dataset.

[0067]

[0068] Table 1 shows the vibration signal types of CWRU bearings selected in this application. This application acquired four types of vibration signals at four different speeds, including normal vibration signals, outer ring, inner ring, and rolling element fault signals. The damage point width for each type of fault was 0.1778 mm. Each type of vibration signal had 60 training set samples and 20 test set samples. To verify the effectiveness of our proposed method, this application randomly selected a portion of samples (15, 30, 45, 60, 75) from each training set for training. The total number of test set samples in each training session was 80. For CWRU data, each sample contained 600 points. Due to the different number of decompositions for each sample, this application selected the first three components of each sample, and then calculated these three components and 10 features of the original signal. Finally, each sample contained 40 feature indices.

[0069] Figure 3 The figures show error bar charts under different rotational speeds and training sets of varying sizes. Experimental results demonstrate that, firstly, the proposed method can effectively diagnose and identify rolling bearing faults; the accuracy rate for each condition shown in the figures is an average of over 85%. Secondly, the proposed method exhibits good robustness; 15 experiments were conducted for the same condition. Figure 2 The error bars show that the deviation of the experimental results in this application is extremely small in each experiment. Finally, even with small sample data, the technology of this application can still effectively detect faults in vibration signals.

[0070] Figure 4 and Figure 5 This paper compares the technical solution of this application with the rolling bearing fault diagnosis method based on PSO-VMD-SNN

[11] . Under different training sets, the fault diagnosis technology of this application has a higher recognition accuracy. At the same time, in multiple repeated experiments, the technical solution of this application has more significant robustness.

[0071] Table 2. Simulation Dataset

[0072]

[0073] Table 2 shows the normal, inner ring, and outer ring fault signals of the bearing under different speeds, loads, and damage widths, simulated using the HD-FD-H-03X rotor rolling bearing and gearbox integrated fault test bench. To verify the effectiveness of this invention in small sample sizes, a subset of samples (50, 100, 150, 200, 250) was randomly selected from each training set for training, with 30 samples in each test set. The Gaussian receptive field of the encoded pulse sequence was 10, the time window of the pulse sequence was 100ms, the learning rate was 0.001, and the epochs were 10 (for training sets with a total number not exceeding 1350) and 20 (for training sets with a total number exceeding 1350). For the data in Table 2, each sample contains 500 points. Due to the different number of decompositions per sample, this paper selected the first two components of each signal, then calculated the 10 features of these two components and the original signal, resulting in 30 feature values ​​per sample.

[0074] Figure 6 The rolling bearing fault signal (including normal, outer ring and inner ring fault) with a load of 0 and a damage width of 1mm is used. This application takes different training sets (150 and 750) and the identification results at different speeds. Figure 7 This application identifies rolling bearing fault signals (including normal, outer ring, and inner ring faults) with a rotational speed of 3000 r / min and a damage width of 1 mm. Different training sets (150, 300, 450, 600, and 750) were selected, and the identification results were obtained under different loads. Experimental results show that the method of this application has good identification performance for different bearing loads, different shaft speeds, and different training sets, demonstrating good generalization ability.

[0075] Finally, this application conducted comparative tests with other methods for the HD-FD-H-03X rotor rolling bearing, such as... Figure 8 As shown, under a load of 0 and a bearing speed of 3000 r / min, this application compared the experimental results of its method with the LMD-SNN method using nine types of vibration signals, including normal signals and fault signals with different damage widths (0.4, 1, 2, and 3 mm), with different training sets (450, 900, 1350, 1800, and 2250). The experimental results show that the method of this application exhibits better recognition performance, and the repeated experiments under each state demonstrate good robustness.

[0076] Spiking neural networks (SNNs) are considered the third generation of artificial neural networks. While neural network models, including ANNs and later CNNs, have achieved successful applications in many fields, the information transmission between neurons in these networks is largely based on mathematical forms and cannot be described using biological mechanisms of inter-neuronal information transmission. Spiking neural networks simulate the propagation mechanisms of biological neural networks and possess far greater computational power. To describe the dynamics of neurons, researchers have established various neuron models, including the classic Hodgkin-Huxley model, the Izhikevich model, and the Leaky Integrate-Fire (LIF) model. Although the earliest research and applications of spiking neural networks can be traced back to the 1950s, due to the complex dynamics of neurons, applications for spiking neural networks are still relatively limited compared to ANNs. However, in the past few decades, thanks to the contributions of many researchers, various learning algorithms have been developed for spiking neural networks, bringing them back into the spotlight. In real-world environments, neurons only respond to stimuli, calculating neuronal membrane potentials; otherwise, they remain in a resting state. In processing information, spiking neural networks encode information into pulse sequences, where only a small portion represents stimulus states. This significantly reduces the computational complexity of spiking neural networks, resulting in their low-energy consumption. Therefore, using spiking neural networks for rolling bearing fault detection not only provides a novel technical method for the field of bearing fault detection but also offers more references and insights for the transition from theoretical research to practical applications of spiking neural networks.

Claims

1. A bearing fault detection method based on a spiking neural network, characterized in that, Includes the following steps: S1. Vibration signal feature extraction Feature extraction is performed on the original vibration signal to obtain feature indices, thereby reducing the noise and dimensionality of the original vibration signal; S2. Fault Detection A spiking neural network is constructed, comprising two layers. The first layer acts as an encoding neuron, which encodes feature indicators with pulses. The second layer acts as a spiking neuron, which determines the bearing fault type based on the pulses transmitted from the neurons in the first layer. In step S2, the second layer of spiking neurons uses the Tempotron learning algorithm to learn the parameters of the neural network. After learning, each neuron is used to distinguish between two different spiking modes. When one mode is input, the maximum membrane potential of the neuron will exceed the threshold, causing the neuron to fire a spiking signal. When the other mode is input, the maximum membrane potential of the neuron will not exceed the threshold, causing the neuron not to fire a spiking signal.

2. The bearing fault detection method based on a spiking neural network according to claim 1, characterized in that, Step S1 specifically includes: S11. The original vibration signal is segmented and each segment of the original vibration signal is taken as a sample data. S12. After obtaining the sample data obtained in step S11, the particle swarm optimization algorithm is applied to the sample data, where the average minimum envelope entropy is used as the fitness function. Finally, the two important parameters required for the variational mode decomposition of the sample are obtained, including the number of decompositions and the penalty factor. S13. After obtaining the number of decompositions and the penalty factor obtained in step S12, perform variational mode decomposition on the sample data to obtain the components; S14. For the original vibration signal and each component, calculate its corresponding 10 characteristic indices. For each sample data, obtain... There are 10 feature indicators, where k is the number of variational mode decompositions corresponding to each sample data.

3. The bearing fault detection method based on a spiking neural network according to claim 2, characterized in that: The 10 feature indicators in step S14 are: skewness factor, peak factor, waveform factor, margin factor, impulse factor, kurtosis factor, approximate entropy, sample entropy, permutation entropy, and fuzzy entropy.

4. The bearing fault detection method based on a spiking neural network according to claim 3, characterized in that: The encoding process of the first layer of encoding neurons in step S2 includes: taking the result obtained in step S14... Each feature index is pulse-coded using a pulse coding method based on Gaussian harmonic curves.

5. A bearing fault detection system based on a spiking neural network, characterized in that, The system includes: The feature extraction module is used to extract features from the original vibration signal, obtain feature indicators, and reduce the noise and dimensionality of the original vibration signal. The fault detection module constructs a spiking neural network, which includes two layers. The first layer acts as an encoding neuron, which performs pulse encoding on the feature indicators. The second layer acts as a spiking neuron, which determines the bearing fault type based on the pulses transmitted from the neurons in the first layer. The spiking neural network in the fault detection module includes an encoding layer and a spiking layer; The pulse coding process of the coding layer includes: processing the pulse coding obtained by the feature index acquisition unit... Each feature index is pulse-coded using a pulse coding method based on Gaussian harmonic curves; The pulse layer uses the Tempotron learning algorithm to learn the parameters of the neural network. Each neuron after learning is used to distinguish between two different pulse modes. When one mode is input, the maximum membrane potential of the neuron will exceed the threshold, causing the neuron to fire a pulse. When the other mode is input, the maximum membrane potential of the neuron will not exceed the threshold, causing the neuron not to fire a pulse.

6. The system according to claim 5, characterized in that, The feature extraction module includes: The signal interception unit is used to segment the original vibration signal and treat each segment of the original vibration signal as a sample data. The particle swarm optimization (PSO) unit is used to apply the PSO algorithm to the sample data obtained by the signal interception unit. The average minimum envelope entropy is used as the fitness function. Finally, two important parameters required for variational mode decomposition of the sample are obtained, including the number of decompositions and the penalty factor. The variational mode decomposition unit is used to perform variational mode decomposition on the sample data after the particle swarm algorithm unit obtains the number of decompositions and the penalty factor, so as to obtain the components. The feature index acquisition unit is used to calculate 10 feature indices for each component of the original vibration signal. For each sample data, it obtains... There are 1 feature index, where k is the number of variational mode decompositions corresponding to each sample data; The 10 feature indicators in the feature indicator acquisition unit are: skewness factor, peak factor, waveform factor, margin factor, impulse factor, kurtosis factor, approximate entropy, sample entropy, permutation entropy, and fuzzy entropy.

7. An electronic device comprising at least one processor; and a memory communicatively connected to the at least one processor; in, The memory stores instructions that are executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-4.

8. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-4.

Images