Bearing fault detection method based on multi-channel deep pulse coupled neural network

Abstract

The present invention discloses a bearing fault detection method based on a multi-channel deep pulse coupled neural network. By introducing a deep network structure, the feature extraction capability of complex fault signals is greatly improved, and the shortcomings of traditional PCNN in high-dimensional signal processing are overcome through optimized network structure and algorithm design. MC-DPCN can not only process complex vibration signals, but also adaptively adjust network parameters to improve the generalization ability and robustness of the model. In addition, the method can still maintain high fault detection accuracy and real-time performance under noise interference and dynamic working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault detection, and in particular to a bearing fault detection method based on a multi-channel deep pulse coupled neural network. Background Art

[0002] With the large-scale and improved automation level of industrial equipment, the working status monitoring and fault diagnosis of bearings in rotating machinery and equipment have become an important part of ensuring production safety and stable operation of equipment. As an important mechanical component in rotating machinery, the performance of bearings is directly related to the overall operating status of the equipment. Once a failure occurs, it may cause vibration and noise at the least, or even cause equipment shutdown, or even cause serious economic losses and safety accidents. Therefore, the development of efficient bearing fault detection technology is of great significance to improving equipment reliability and reducing operating costs.

[0003] At present, bearing fault detection technology mainly relies on the analysis and processing of vibration signals. Common technical means include time domain analysis, frequency domain analysis (such as fast Fourier transform, FFT) and time-frequency domain analysis (such as wavelet transform). These methods have achieved remarkable results in the early stages, but with the diversification of fault modes and the increase in the complexity of working conditions, the limitations of traditional methods have gradually become apparent;

[0004] In order to overcome the shortcomings of traditional methods, bearing fault detection methods based on artificial intelligence (AI) and deep learning have gradually attracted widespread attention in recent years. In particular, deep neural networks (DNN) and convolutional neural networks (CNN) have shown good application prospects in the field of fault detection due to their powerful feature extraction and adaptive capabilities. Neural networks can automatically learn and extract fault-related features from large amounts of data, reducing dependence on manual experience. However, despite the good performance of deep learning methods, the following problems still exist:

[0005] 1. Large data requirements: Deep learning methods usually rely on a large amount of labeled data for training. However, in actual applications, especially during normal operation of equipment, fault data is relatively scarce, which limits the practical application of these methods.

[0006] 2. Long training time and high computing resource requirements: Deep neural networks have complex structures and numerous parameters. The training process requires a lot of computing resources and time, making it difficult to use them efficiently in scenarios with limited resources.

[0007] 3. Poor model interpretability: Although deep learning models perform well in detection accuracy, their internal working mechanisms are complex and it is difficult to provide users with intuitive fault diagnosis basis, resulting in low interpretability and acceptability of the model.

[0008] As a biologically inspired neural network model, pulse coupled neural network (PCNN) has been widely used in image processing, feature extraction and other fields in recent years. PCNN has a unique synchronous oscillation mechanism that can adaptively adjust the input signal, especially when processing data with spatial complexity. Compared with traditional neural networks, PCNN has significant advantages in extracting edge features and mutation points of signals. However, the classic PCNN model still has the following limitations when processing high-dimensional complex signals (such as vibration signals):

[0009] ①. Shallow structure limits feature extraction capabilities: Traditional PCNN has a shallow structure and is difficult to effectively capture deep features in complex signals.

[0010] ②. Lack of global information processing capabilities: PCNN relies more on the extraction of local information and has limited modeling capabilities for global signal features, resulting in poor performance when dealing with complex fault modes.

[0011] In order to solve the above problems, inspired by brain-like computing, a bearing fault detection method based on multi-channel deep pulse coupled neural network (MC-DPCN) is proposed. By introducing a deep network structure, this method greatly improves the feature extraction capability of complex fault signals, and overcomes the shortcomings of traditional PCNN in high-dimensional signal processing through optimized network structure and algorithm design. MC-DPCN can not only process complex vibration signals, but also adaptively adjust network parameters to improve the generalization ability and robustness of the model. In addition, this method can still maintain high fault detection accuracy and real-time performance under noise interference and dynamic conditions. Summary of the invention

[0012] The purpose of the present invention is to provide a bearing fault detection method based on a multi-channel deep pulse coupled neural network to solve the problems raised in the above background technology.

[0013] To achieve the above object, the present invention provides the following technical solutions:

[0014] The bearing fault detection method based on multi-channel deep pulse coupled neural network is characterized by comprising the following steps:

[0015] The following steps are involved:

[0016] S1. Vibration signal acquisition and preprocessing

[0017] Vibration signals during equipment operation are collected in real time by installing acceleration sensors at the motor drive end, fan end, base end, etc.

[0018] The collected vibration signal is subjected to noise reduction processing, a low-pass filter is used to remove high-frequency noise in the signal, and the signal amplitude is normalized to a range between 0 and 1 through normalization processing to facilitate subsequent feature extraction;

[0019] S2. Feature extraction

[0020] Wavelet transform is used to transform the signal from time domain to time-frequency domain. According to the window size of 256 and the step size of 64, the acceleration data of multiple sites are converted into frequency spectra as multiple channel inputs of deep pulse coupled neural network;

[0021] Assume that the vibration signal collected from the acceleration sensor is ,in is the time variable, is the acceleration signal of the bearing vibration, which is a discrete sampling signal ,in is the number of sampling points, assuming the sampling frequency is , then each time step is .

[0022]

[0023] S3. Model construction

[0024] According to the characteristics of the vibration signal, a suitable network structure is selected. The model consists of multiple layers of PCNN, which can extract multi-scale features in the signal layer by layer.

[0025] The spectrum of the bearing vibration signal first enters the convolution unit; each convolution unit consists of a convolution layer and an average pooling layer, a convolution block, a batch normalization block, and a PCNN layer; then it enters the fully connected unit; the fully connected block consists of a fully connected layer, a batch normalization block, and a PCNN layer; at the final moment, the output of T=N is used for classification;

[0026] S3.1 Batch Normalization

[0027] BN first calculates the mean and variance of a batch of samples, then standardizes the samples in this batch, and finally scales and translates them through learnable parameters;

[0028]

[0029] BN (Batch Normalize) stands for batch normalization, X is one of the input samples, and the subscript represents a batch of samples, represents element-by-element multiplication, γ and β are respectively learnable scaling and translation coefficients, is the standard deviation of the batch samples, and are the mean and variance of the batch samples, respectively, and are calculated by the following formula:

[0030]

[0031]

[0032] in is the translation coefficient, which obtains the mean and variance of the entire data set through moving average, ensuring that the model can maintain stable performance when processing new, unseen data;

[0033] S3.2 PCNN neurons

[0034] The PCNN layer with convolutional feed-forward connections can be described as:

[0035]

[0036]

[0037]

[0038]

[0039]

[0040] in:

[0041]

[0042] Step function:

[0043]

[0044] S4. Model training

[0045] The model is trained using historical bearing fault data. The data set includes normal operating conditions and various fault types. The supervised learning method is used to train the model through labeled data.

[0046] After training is completed, the model is evaluated using the cross-validation method to ensure that it has good generalization ability. The validation process requires testing the model's classification accuracy, error rate and other indicators under different fault types. The accuracy rate must be no less than 95%;

[0047] S5. Real-time fault detection: Deploy the trained multi-channel deep pulse coupled neural network model to the bearing monitoring system, process the vibration signals collected by the sensor in real time, and monitor and diagnose the operating status of the equipment. The system automatically determines whether there is a fault and identifies the type of fault;

[0048] S6. Fault warning: When a fault is detected, the system will issue a corresponding warning signal based on the fault type and severity to guide maintenance personnel to perform equipment maintenance to avoid further expansion of the fault.

[0049] Furthermore, in order to obtain information in the frequency domain in S2, it is necessary to perform a discrete time domain signal Perform fast Fourier transform, the mathematical expression is:

[0050]

[0051] in:

[0052] is the Fourier coefficient of the frequency domain signal, representing the frequency The weight of is an imaginary unit; is the number of sampling points of FFT; k represents the number of discrete frequencies;

[0053] The spectrum diagram represents the amplitude of the frequency component. The spectrum amplitude of the vibration signal can be obtained by calculating the modulus of the Fourier coefficient:

[0054]

[0055] in, and are the real and imaginary parts of the Fourier transform results. In order to more intuitively represent the energy distribution of the signal, the amplitude is processed using a logarithmic scale (dB):

[0056]

[0057] So obtained It is the frequency domain amplitude displayed in the spectrum diagram, which represents the energy of each frequency component;

[0058] The result of FFT is symmetrical. The dot represents the positive frequency component. The dots represent negative frequency components; in practical applications, only the positive frequency components are retained. The frequency axis is defined as:

[0059] .

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. Preprocess the vibration acceleration data of the bearing through wavelet transform, and then use fast Fourier transform to obtain time domain data, which solves the problem that traditional visual models cannot process time series acceleration data. No complex vibration analysis is required.

[0062] 2. The multi-channel pulse coupled neural network model (MC-DPCN) developed based on brain-like computing is adopted. This model is inspired by brain-like computing and achieves high accuracy and high robustness by simulating the signal processing mechanism of the brain's visual cortex. Most of the existing patents / articles use traditional deep learning or machine learning models. This is the first time that brain-like computing has been introduced into the field of fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a schematic flow chart of a bearing fault detection method based on a multi-channel deep pulse coupled neural network of the present invention;

[0064] Figure 2 This is a spectrum diagram after the acceleration data at a sampling frequency of 12kHz is processed by the present invention;

[0065] Figure 3 This is a schematic diagram of data training performance using data collected at a sampling frequency of 12 kHz in the present invention. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] This paper proposes a bearing fault detection method based on multi-channel deep pulse coupled neural network (MC-DPCN), aiming to overcome the problems of insufficient extraction of complex fault signals, poor noise resistance, and low real-time performance of traditional detection methods. This method can accurately and real-timely diagnose the operating status of the bearing and determine the fault type by extracting features and pattern recognition of bearing vibration signals.

[0068] Schematic diagram of the bearing fault detection method based on multi-channel deep pulse coupled neural network Figure 1 As shown, the technical solution is mainly divided into the following steps:

[0069] 1. Vibration signal collection

[0070] Vibration signals during equipment operation are collected in real time by installing acceleration sensors at the motor drive end, fan end, base end, etc. The sensor installation should ensure reliable fastening to avoid external interference.

[0071] Conditions: sensor sensitivity is 10 mV / g, sampling frequency is 10 kHz.

[0072] Collection environment: Carry out under normal temperature and pressure to avoid high temperature or strong vibration environment affecting data quality.

[0073] 2. Signal preprocessing:

[0074] The collected vibration signal is subjected to noise reduction processing, a low-pass filter is used to remove high-frequency noise in the signal, and the signal amplitude is standardized to a range between 0 and 1 through normalization processing to facilitate subsequent feature extraction.

[0075] 3. Feature extraction

[0076] In order to more effectively extract fault features from vibration signals, wavelet transform is used to transform the signal from time domain to time-frequency domain. According to the window size of 256 and the step size of 64, the acceleration data of multiple locations are converted into frequency spectra as multiple channel inputs of deep pulse coupled neural network.

[0077] Assume that the vibration signal collected from the acceleration sensor is ,in is the time variable, is the acceleration signal of the bearing vibration. This signal is a discrete sampling signal ,in is the number of sampling points. Assume the sampling frequency is , then each time step is .

[0078]

[0079] In order to obtain information in the frequency domain, it is necessary to perform discrete time domain signal Perform fast Fourier transform, the mathematical expression is:

[0080]

[0081] in:

[0082] is the Fourier coefficient of the frequency domain signal, representing the frequency The weight of

[0083] is an imaginary unit;

[0084] is the number of FFT points (number of sampling points); k represents the number of discrete frequencies.

[0085] The spectrum shows the amplitude of the frequency components. The spectrum amplitude of the vibration signal can be obtained by finding the modulus of the Fourier coefficients:

[0086]

[0087] in and are the real and imaginary parts of the Fourier transform results. In order to more intuitively represent the energy distribution of the signal, the amplitude is processed using a logarithmic scale (dB):

[0088]

[0089] So obtained It is the frequency domain amplitude displayed in the spectrum diagram, which represents the energy of each frequency component.

[0090] The result of FFT is symmetrical. The dot represents the positive frequency component. The dots represent negative frequency components. In practical applications, only the positive frequency components are retained. The frequency axis is defined as:

[0091]

[0092] 4. Model Construction

[0093] The appropriate network structure is selected according to the characteristics of the vibration signal. The model consists of multiple layers of PCNN, which can extract multi-scale features in the signal layer by layer.

[0094] The spectrogram of the bearing vibration signal first enters the convolution unit. Each convolution unit consists of a convolution layer and an average pooling layer. It consists of a convolution block (Conv), a batch normalization block (BN), and a PCNN layer. Then it enters the fully connected unit; the fully connected block consists of a fully connected layer, a batch normalization block (BN), and a PCNN layer (classifier). At the final moment, the output of T=N is used for classification.

[0095] 4.1 Batch Normalization

[0096] BN first calculates the mean and variance of a batch of samples, then standardizes the samples in this batch, and finally scales and translates them through learnable parameters.

[0097]

[0098] BN (Batch Normalize) stands for batch normalization, X is one of the input samples, and the subscript represents a batch of samples, represents element-by-element multiplication, γ and β are respectively learnable scaling and translation coefficients, is the standard deviation of the batch samples, and are the mean and variance of the batch samples, respectively, and are calculated by the following formula:

[0099]

[0100]

[0101] in is the translation coefficient, which obtains the mean and variance of the entire data set through moving average, ensuring that the model can maintain stable performance when processing new and unseen data.

[0102] 4.2 PCNN Neurons

[0103] The PCNN layer with convolutional feed-forward connections can be described as:

[0104]

[0105]

[0106]

[0107]

[0108]

[0109] in:

[0110]

[0111] Step function:

[0112]

[0113] 5. Model Training

[0114] The model is trained using historical bearing fault data, which includes normal operating conditions and various fault types (such as inner race fault, outer race fault, rolling element fault, etc.). The supervised learning method is used to train the model through labeled data.

[0115] Optimization algorithm: Adam optimizer is used, the learning rate is 0.001, and the number of training times is 500.

[0116] Loss function: The cross entropy loss function is used to optimize the model parameters by minimizing the loss function.

[0117] Hyperparameters:

[0118] Initialization parameters: Setting

[0119] After training is completed, the model is evaluated using the cross-validation method to ensure that it has good generalization ability. The validation process requires testing the model's classification accuracy, error rate and other indicators under different fault types.

[0120] Evaluation indicators: The accuracy rate must be no less than 95%.

[0121] 6. Real-time fault detection: Deploy the trained multi-channel deep pulse coupled neural network model to the bearing monitoring system to process the vibration signals collected by the sensor in real time and monitor and diagnose the operating status of the equipment. The system automatically determines whether there is a fault and identifies the type of fault.

[0122] Implementation conditions: The model processes newly collected data once a second and outputs the fault status.

[0123] 7. Fault warning: When a fault is detected, the system will issue a corresponding warning signal based on the type and severity of the fault to guide maintenance personnel to repair the equipment and avoid further expansion of the fault.

[0124] Figure 2 This is the spectrum diagram after processing. Figure 3 The training performance of the data collected using the 12kHz sampling frequency reached 99.89% accuracy on the test set and 100% accuracy on the training set after 20 training times.

[0125] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting from any point of view, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0126] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A bearing fault detection method based on a multi-channel deep pulse coupled neural network, characterized in that: The following steps are involved: S1. Vibration signal acquisition and preprocessing The vibration signals during the operation of the equipment are collected in real time by installing acceleration sensors at the motor drive end, fan end and base end. The collected vibration signal is subjected to noise reduction processing, a low-pass filter is used to remove high-frequency noise in the signal, and the signal amplitude is normalized to a range between 0 and 1 through normalization processing to facilitate subsequent feature extraction; S2. Feature extraction Wavelet transform is used to transform the signal from time domain to time-frequency domain. Acceleration data of multiple sites are converted into frequency spectra with a window size of 256 and a step size of 64, which are used as multiple channel inputs of deep pulse coupled neural network. Assume that the vibration signal collected from the acceleration sensor is ,in is the time variable, It is the acceleration signal of bearing vibration, which is a discrete sampling signal. ,in is the number of sampling points, assuming the sampling frequency is , then each time step is ; S3. Model construction The network structure is selected according to the characteristics of the vibration signal. The multi-channel pulse coupled neural network model is composed of multiple layers of PCNN, which can extract multi-scale features in the signal layer by layer. The spectrum of the bearing vibration signal first enters the convolution unit; each convolution unit consists of a convolution layer and an average pooling layer, and the convolution layer contains a convolution block, a batch normalization block, and a PCNN layer; then it enters the fully connected unit; the fully connected block consists of a fully connected layer, a batch normalization block, and a PCNN layer; at the final moment, the output of T=N is used for classification; S3.1 Batch Normalization BN first calculates the mean and variance of a batch of samples, then standardizes the samples in this batch, and finally scales and translates them through learnable parameters; Where BN stands for batch normalization, X is one of the input samples, and the subscript represents a batch of samples, represents element-by-element multiplication, γ and β are respectively learnable scaling and translation coefficients, is the standard deviation of the batch samples, and are the mean and variance of the batch samples, respectively, and are calculated by the following formula: in is the translation coefficient, which obtains the mean and variance of the entire data set through moving average, ensuring that the model can maintain stable performance when processing new and unseen data; S3.2 PCNN neurons The PCNN layer with convolutional feed-forward connections can be described as: Where: L represents coupling connection; F represents feedforward input; U represents the internal activity value of the neuron; E represents the dynamic threshold of the neuron; Y represents continuous output; W, M represent trainable weight matrices; V E represents the amplitude gain; represents leakage factor; H represents step function; Step function: S4. Model training The model is trained using historical bearing fault data. The data set includes normal operating conditions and various fault types. The supervised learning method is used to train the model through labeled data. After training is completed, the model is evaluated using the cross-validation method to ensure that it has good generalization ability. The validation process requires testing the classification accuracy and error rate indicators of the model under different fault types, and the accuracy rate is required to be no less than 95%; S5. Real-time fault detection: Deploy the trained multi-channel deep pulse coupled neural network model to the bearing monitoring system to process the vibration signals collected by the sensor in real time, monitor and diagnose the operating status of the equipment; the system automatically determines whether there is a fault and identifies the fault type; S6. Fault warning: When a fault is detected, the system will issue a corresponding warning signal based on the fault type and severity to guide maintenance personnel to perform equipment maintenance to avoid further expansion of the fault.

2. The bearing fault detection method based on multi-channel deep pulse coupled neural network according to claim 1 is characterized in that: In order to obtain information in the frequency domain in S2, the discrete time domain signal sampled by the acceleration sensor needs to be Perform fast Fourier transform, the mathematical expression is: in: is the Fourier coefficient of the frequency domain signal, representing the frequency The amount of is an imaginary unit; is the number of sampling points of FFT; k represents the number of discrete frequencies; The spectrum diagram represents the amplitude of the frequency component. The spectrum amplitude of the vibration signal can be obtained by calculating the modulus of the Fourier coefficient: in, and are the real and imaginary parts of the Fourier transform results respectively. In order to more intuitively represent the energy distribution of the signal, the amplitude is processed using a logarithmic scale: So obtained It is the frequency domain amplitude displayed in the spectrum diagram, which represents the energy of each frequency component, and dB is a logarithmic scale; The result of FFT is symmetrical. The dot represents the positive frequency component. The dots represent negative frequency components; in practical applications, only the positive frequency components are retained; the frequency axis is defined as: 。

Images