Bearing non-contact fault diagnosis method based on quadratic convolutional neural network

Abstract

The invention relates to the technical field of bearing fault diagnosis, and provides a bearing non-contact fault diagnosis method based on a quadratic convolutional neural network, comprising the following steps: constructing a deep learning model guided by bearing physical degradation information; training the deep learning model; key characteristic parameters of the bearing vibration signals are extracted; through a maximum mean difference method, kurtosis optimization sparsity and multi-dimensional logarithmic deviation, a physical heuristic network loss function is jointly optimized, and neural network efficient convergence under physical information constraint is achieved; optimizing the deep learning model to obtain an optimal diagnosis model; and acquiring real-time vibration data of operation of the target bearing, and judging the fault type and the fault degree of the target bearing. According to the method, the problem of precision reduction caused by insufficient utilization of physical characteristics in a traditional fault diagnosis method is solved, meanwhile, the limitation that a contact sensor is difficult to install and easy to damage is solved, and the method is suitable for non-contact bearing fault diagnosis tasks in a complex industrial environment.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a bearing non-contact fault diagnosis method based on a secondary convolutional neural network. Background Art

[0002] With the development of automation and intelligence in industrial equipment, bearings are one of the most critical rotating parts in mechanical equipment, and their operating status is directly related to the overall performance and safety of the equipment. Traditional bearing fault diagnosis methods mostly rely on contact sensors, such as vibration sensors and temperature sensors, to determine the health status of bearings by detecting parameters such as vibration signals and temperature changes during operation. However, these methods usually require the sensor to directly contact the bearing surface, which is not only complicated to install and maintain, but also easily damaged in high-speed operation environments.

[0003] In recent years, non-contact fault diagnosis technology has gradually received attention, especially in high-speed and complex environments, where non-contact detection methods can effectively overcome the shortcomings of contact sensors. At the same time, with the rapid development of artificial intelligence technology, the advantages of neural networks in processing complex data pattern recognition have gradually emerged, and their application in the field of fault diagnosis has also shown significant results.

[0004] In the field of fault diagnosis, the combination of physical models and neural networks has become a hot topic of research. Among them, the secondary physical neural network provides a new solution for bearing fault diagnosis by introducing physical property constraints and nonlinear feature extraction capabilities. The network can extract multi-dimensional features in combination with the physical properties of vibration signals under non-contact conditions, especially for the sparsity, peak characteristics and periodic impact signals in bearing fault signals, which significantly improves the accuracy and robustness of fault diagnosis.

[0005] This paper proposes a non-contact bearing fault diagnosis method based on a secondary convolutional neural network. By introducing peak sparsity optimization, multi-dimensional logarithmic deviation constraints and maximum mean difference method, this method realizes the key feature extraction and fault severity assessment of bearing vibration data under physical constraints, while effectively overcoming the limitations of traditional contact diagnosis methods, and is suitable for bearing health status monitoring and fault diagnosis under complex working conditions. Summary of the invention

[0006] Based on this, in order to solve the problems of complex installation, easy damage and interference to the bearing structure of contact sensors in traditional bearing fault diagnosis methods, the present invention provides a non-contact bearing fault diagnosis method based on a secondary convolutional neural network. By combining the physical degradation model with deep learning technology, the present invention realizes the non-contact acquisition and feature extraction of multi-dimensional vibration signals, and improves the accuracy and robustness of fault diagnosis by optimizing the designed physical information embedded loss function. The specific technical solution is as follows:

[0007] A non-contact bearing fault diagnosis method based on a secondary convolutional neural network comprises the following steps:

[0008] Collect bearing vibration data based on non-contact method;

[0009] Dividing the bearing vibration data into samples to obtain a training set and a test set;

[0010] Acquire a bearing physical degradation model according to peak sparsity, acquire a secondary physical neural network according to the bearing physical degradation model and secondary convolution neurons, and construct a deep learning model of bearing physical information according to the secondary physical neural network;

[0011] Iteratively training the deep learning model using the training set to extract key characteristic parameters of the bearing vibration data;

[0012] Based on the key feature parameters, the network loss function of the deep learning model is optimized in combination with the maximum mean difference method, the kurtosis optimization sparsity constraint and the multi-dimensional logarithmic deviation;

[0013] The deep learning model is verified by using the optimized network loss function and test set to obtain the optimal diagnostic model;

[0014] Real-time vibration data of the target bearing operation is obtained and input into the diagnostic model, the fault type and fault degree of the target bearing are determined, and non-contact fault diagnosis of the target bearing is achieved.

[0015] The training set is used to iteratively train the secondary physical neural network, and the key characteristic parameters of the vibration signal are obtained and optimized by combining multiple loss functions inspired by physics, including:

[0016] Peak sparsity optimization loss: used to maximize the sparsity of periodic shocks in vibration signals;

[0017] Multi-dimensional logarithmic deviation loss: By quantifying the characteristic differences of multi-dimensional bearing real-time data such as vibration signals, infrared temperature signals and acoustic signals, the ratio of real-time data X to standard data in each dimension is calculated, the constant logarithmic value is taken, and the weighted average value is calculated after assigning weight coefficients to each dimension. In this way, a loss function is constructed to accurately measure the differences between multi-dimensional data, thereby improving the comprehensive diagnostic ability of the model for multi-modal data and the accuracy of fault severity assessment;

[0018] Maximum mean difference loss: ensures data distribution consistency between training and test sets.

[0019] Through the joint optimization of the above loss functions, a deep learning model guided by physical information is constructed, and the model performance is verified using a test set to obtain the optimal diagnostic model;

[0020] In summary, the non-contact fault diagnosis method for bearings based on physically guided secondary convolutional neural networks collects bearing vibration data in a non-contact manner, obtains a bearing physical degradation model based on peak sparsity, and obtains a secondary physical neural network based on secondary convolutional neurons, thereby constructing a deep learning model of bearing physical information. Based on the key feature parameters, combined with the maximum mean difference method, kurtosis optimization sparsity constraints and multi-dimensional logarithmic deviation, the network loss function of the deep learning model is optimized to obtain the optimal diagnostic model, and the fault type and degree of the target bearing are determined based on the optimal diagnostic model. It does not require the installation of contact sensors, which not only overcomes the problems of difficult installation and easy damage of contact sensors in traditional bearing fault diagnosis methods, but also avoids the interference of traditional bearing fault diagnosis methods with the bearing structure due to the need to install contact sensors, thereby improving measurement accuracy and stability.

[0021] Preferably, the specific method for collecting bearing vibration data in a non-contact manner comprises the following steps:

[0022] Based on the laser self-mixing interference technology, the laser beam is irradiated on the surface of the bearing base, the laser information reflected from the surface of the bearing base is detected, and the interference signal is obtained;

[0023] Bearing vibration data is collected according to the phase change of the interference signal.

[0024] Preferably, the bearing physical degradation model includes peak sparsity, bearing degradation law and bearing fault vibration characteristics under different working conditions; the method for obtaining the bearing physical degradation model includes: constructing a bearing physical degradation model according to the peak sparsity of the bearing vibration signal and the bearing degradation law, which is used to characterize the physical degradation characteristics and fault characteristics of the bearing under different operating conditions.

[0025] Preferably, the bearing physical degradation model is obtained according to the peak sparsity, and the specific method of obtaining the secondary physical neural network according to the bearing physical degradation model and the secondary convolutional neuron includes the following steps:

[0026] Obtaining bearing physical degradation models based on peak sparsity

[0027] According to the bearing physical degradation model and the secondary convolutional neuron

[0028] Get the secondary physical neural network σ(f(x))=σ((w r L vib-kurt +b r )(w g L vib-kurt +b g )+w b (L vib-kurt ⊙L vib-kurt )+m);

[0029] in, vib represents the peak value of the vibration signal, stiff represents the equivalent contact stiffness level between the bearing ball and the inner and outer ring raceways, m represents the equivalent mass of the system, ε represents the related imbalance in the system load, V represents the speed, and w r 、w g 、w b represents the weight, b r 、b g and c denote bias parameters, σ(·) is the activation function, ⊙ denotes the Hadamard product, x i represents the i-th sample data, x j represents the jth sample data, P represents the total number of samples, Represents weight.

[0030] Preferably, the secondary physical neural network comprises:

[0031] An input layer, used for receiving the bearing vibration data after sample division;

[0032] A convolution layer, constructed based on the secondary physical convolution neuron, for performing nonlinear mapping and feature fusion on key features of bearing vibration data;

[0033] The fully connected layer further extracts the fused high-dimensional features;

[0034] The output layer is used to classify and quantitatively evaluate the type and degree of the bearing fault.

[0035] Preferably, the loss function of the quadratic physical neural network includes a kurtosis optimization sparsity constraint Multi-dimensional logarithmic deviation loss and maximum mean difference loss The total loss function is L total =αL s-MMD +βL log-dev +γL s-kurt ;

[0036] Among them, y(n) is the time series, Δ norm is the weighted average of the normalized logarithmic deviation, where the logarithmic deviation is X i , Y i are the real-time data and standard data in the i-th dimension, φ(·) represents the reproducing kernel Hilbert space (RKHS), D s and D t Represent the sample sets of original working condition and target working condition respectively, n s and n t represents the number of samples in the original working condition and the target working condition respectively, and optimizes the network loss function L of the deep learning model total =αL s-MMD +βL log-dev +γL s-kurt , α, β, and γ are the weights of different loss terms respectively.

[0037] Preferably, the vibration data includes vibration displacement, vibration velocity, vibration acceleration and rotation speed, and the fault type includes bearing wear, crack and fatigue.

[0038] Preferably, before dividing the bearing vibration data into samples, the bearing vibration data is preprocessed.

[0039] Preferably, the preprocessing includes wavelet transform denoising and bandpass filtering.

[0040] Preferably, the non-contact bearing fault diagnosis method based on physically guided secondary convolutional neural network also includes the following steps: acquiring bearing data of multiple different dimensions, and determining the bearing fault location and fault degree according to the bearing data of multiple different dimensions. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The present invention can be further understood from the following description in conjunction with the accompanying drawings. The components in the figures are not necessarily drawn to scale, but the emphasis is placed on illustrating the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0042] Figure 1 It is a schematic diagram of the overall process of a non-contact bearing fault diagnosis method based on a secondary convolutional neural network in one embodiment of the present invention;

[0043] Figure 2 It is a flowchart of a specific method for obtaining a secondary physical neural network in one embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with its embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the protection scope of the present invention.

[0045] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be a central element. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be a central element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only and do not represent the only implementation method.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0047] The “first” and “second” mentioned in the present invention do not represent specific quantities and orders, but are merely used to distinguish names.

[0048] like Figure 1 As shown, a non-contact bearing fault diagnosis method based on a secondary physical neural network in one embodiment of the present invention includes the following steps:

[0049] S1, collecting bearing vibration data based on non-contact method.

[0050] Preferably, in step S1, the specific method for collecting bearing vibration data based on a non-contact manner includes the following steps: based on laser self-mixing interference technology, a laser beam is irradiated on the surface of the bearing base, laser information reflected from the surface of the bearing base is detected, and an interference signal is obtained; and according to the phase change of the interference signal, bearing vibration data is collected.

[0051] Specifically, the laser self-mixing interferometry (SMI) technology is used to collect bearing vibration signals in a non-contact manner. This technology irradiates the surface of the bearing and / or the bearing base with a laser, and the reflected laser signal enters the laser cavity to interfere with the original laser to form an interference signal. The phase of the interference signal changes with the frequency and amplitude of the vibration, thereby obtaining high-precision bearing vibration data. SMI technology has the advantages of high sensitivity and non-contact measurement. It is particularly suitable for obtaining the vibration state of the bearing without interfering with its normal operation, providing basic data for subsequent fault diagnosis.

[0052] Through laser self-mixing interference technology, the tiny vibration data of the bearing during operation can be obtained. This non-contact measurement method based on laser self-mixing interference technology avoids the complexity of contact sensor installation and potential damage to the bearing surface in traditional bearing fault diagnosis methods, and it shows high efficiency and stability in high-speed operation environments.

[0053] S2, dividing the bearing vibration data into samples to obtain a training set and a test set.

[0054] The vibration data includes but is not limited to vibration displacement, vibration velocity, vibration acceleration and rotation speed.

[0055] Before the bearing vibration data is sampled, the bearing vibration data may be preprocessed, including wavelet transform denoising and bandpass filtering to remove environmental noise and interference signals, thereby ensuring a high signal-to-noise ratio and accuracy of the bearing vibration data input into the secondary physical neural network.

[0056] S3, obtaining a bearing physical degradation model according to peak sparsity, obtaining a secondary physical neural network according to the bearing physical degradation model and secondary convolution neurons, and constructing a deep learning model of bearing physical information according to the secondary physical neural network.

[0057] Preferably, the bearing physical degradation model includes the degradation law and operating conditions of the bearing, which is used to obtain the fault vibration characteristics of the bearing under different operating conditions.

[0058] In step S3, if Figure 2 As shown, the specific method for obtaining a secondary physical neural network according to the bearing physical degradation model and the secondary convolutional neuron includes the following steps:

[0059] S30, construct secondary convolutional neurons.

[0060] S31, embedding the bearing physical degradation model into the secondary convolution neuron to obtain a secondary physical convolution neuron.

[0061] S32, obtaining the secondary physical neural network according to the secondary physical convolution neuron.

[0062] The secondary physical neural network includes an input layer, a convolutional layer, a fully connected layer and an output layer.

[0063] The input layer is used to receive the bearing vibration data after sample division; the convolution layer is constructed based on the secondary physical convolution neuron, and is used to perform nonlinear mapping and feature fusion on the key features of the bearing vibration data; the fully connected layer further extracts the fused high-dimensional features; the output layer is used to classify the fault type.

[0064] By embedding the bearing physical degradation model that includes the bearing degradation law and operating conditions into the secondary convolutional neuron to construct a secondary physical convolutional neuron, it is possible to simulate the fault vibration characteristics of the bearing under different operating conditions and effectively reflect the abnormal operating state of the bearing, so that the secondary physical neural network can extract nonlinear features related to bearing faults.

[0065] The secondary physical neural network combines the bearing physical degradation model and the secondary physical convolutional neuron, which can not only capture the complex patterns in the bearing vibration signal, but also enhance the understanding and extraction of bearing fault characteristics with the help of the prior knowledge of the bearing physical degradation model.

[0066] S4, iteratively training the deep learning model using the training set to extract key characteristic parameters of the bearing vibration data.

[0067] The deep learning model is trained using the training set and continuously iterated.

[0068] S5, obtaining key characteristic parameters of the bearing vibration data, combining the key characteristic parameters, and jointly optimizing the network loss function of the deep learning model through the Maximum Mean Discrepancy (MMD) method, peak sparsity optimization, and multi-dimensional logarithmic deviation constraints.

[0069] S6, optimizing the deep learning model through the optimized network loss function of the deep learning model and the test set to obtain the optimal diagnostic model.

[0070] Based on the acquired bearing vibration data, the key feature parameters are extracted, and the obtained key feature parameters are combined with the maximum mean difference method, multi-dimensional logarithmic deviation and peak sparsity optimization to jointly optimize the network loss function, and further optimize the deep learning model with the participation of the test set to obtain the optimal diagnostic model. Specifically, by continuously iteratively optimizing the deep learning model, the key features of the vibration signal and multiple constraints inspired by physics are comprehensively utilized in each iteration to gradually improve the model's sensitivity and classification capabilities to fault characteristics. The model is iteratively optimized until the optimal solution is reached or the set maximum number of iterations is reached. The final optimized deep learning model is used as the optimal diagnostic model to achieve accurate diagnosis and status assessment of bearing faults under complex working conditions.

[0071] S7, acquiring real-time vibration data of the target bearing and inputting the real-time vibration data into the diagnostic model, determining the fault type and fault degree of the target bearing, and realizing non-contact fault diagnosis of the target bearing.

[0072] The fault types include but are not limited to bearing wear, cracks and fatigue.

[0073] In the non-contact bearing fault diagnosis method based on physics-guided secondary convolutional neural network, after training, the secondary physical neural network can automatically identify different bearing failure modes and accurately classify the health status of the bearing, including common bearing wear, cracks, fatigue and other fault types.

[0074] Preferably, the non-contact fault diagnosis method for bearings based on physically guided secondary convolutional neural networks determines the fault type and fault severity of the target bearing and realizes non-contact fault diagnosis of the target bearing. It also automatically generates a corresponding diagnostic report based on the diagnosis results, indicating the fault type, severity and possible causes of the bearing. At the same time, it issues an alarm in real time when a fault occurs, helping equipment maintenance personnel to take timely measures.

[0075] In summary, the non-contact fault diagnosis method for bearings based on physically guided secondary convolutional neural networks collects bearing vibration data in a non-contact manner, obtains a bearing physical degradation model based on peak sparsity, and obtains a secondary physical neural network based on secondary convolutional neurons, thereby constructing a deep learning model of bearing physical information. Based on the key feature parameters, combined with the maximum mean difference method, kurtosis optimization sparsity constraints and multi-dimensional logarithmic deviation, the network loss function of the deep learning model is optimized to obtain the optimal diagnostic model, and the fault type and degree of the target bearing are determined based on the optimal diagnostic model. It does not require the installation of contact sensors, which not only overcomes the problems of difficult installation and easy damage of contact sensors in traditional bearing fault diagnosis methods, but also avoids the interference of traditional bearing fault diagnosis methods with the bearing structure due to the need to install contact sensors, thereby improving measurement accuracy and stability.

[0076] As a preferred technical solution, the bearing physical degradation model includes peak sparsity, bearing degradation law and bearing fault vibration characteristics under different working conditions; the method for obtaining the bearing physical degradation model includes: constructing a bearing physical degradation model according to the peak sparsity of the bearing vibration signal and the bearing degradation law, which is used to characterize the physical degradation characteristics and fault characteristics of the bearing under different operating conditions.

[0077] A specific method for obtaining a bearing physical degradation model according to peak sparsity and obtaining a secondary physical neural network according to the bearing physical degradation model and secondary convolution neurons includes the following steps:

[0078] Obtaining bearing physical degradation models based on peak sparsity

[0079] According to the bearing physical degradation model and the secondary convolutional neuron

[0080] Get the secondary physical neural network σ(f(x))=σ((w r L vib-kurt +b r )(w g L vib-kurt +b g )+w b (L vib-kurt ⊙L vib-kurt )+m);

[0081] in, vib represents the peak value of the vibration signal, stiff represents the equivalent contact stiffness level between the bearing ball and the inner and outer ring raceways, m represents the equivalent mass of the system, ε represents the related imbalance in the system load, V represents the speed, and w r 、w g 、wb represents the weight, b r 、b g and c denote bias parameters, σ(·) is the activation function, ⊙ denotes the Hadamard product, x i represents the i-th sample data, x j represents the jth sample data, P represents the total number of samples, Represents weight.

[0082] The loss function of the quadratic physical neural network includes kurtosis optimization and sparsity constraints Multi-dimensional logarithmic deviation loss and maximum mean difference loss The total loss function is L total =αL s-MMD +βL log-dev +γL s-kurt ;

[0083] Among them, y(n) is the time series, Δ norm is the weighted average of the normalized logarithmic deviation, where the logarithmic deviation is X i , Y i are the real-time data and standard data in the i-th dimension, φ(·) represents the reproducing kernel Hilbert space (RKHS), D s and D t Represent the sample sets of original working condition and target working condition respectively, n s and n t represents the number of samples in the original working condition and the target working condition respectively, and optimizes the network loss function L of the deep learning model total =αL s-MMD +βL log-dev +γL s-kurt , α, β, and γ are the weights of different loss terms respectively.

[0084] Among them, the peak sparsity optimization loss is used to maximize the sparsity of periodic shocks in the vibration signal;

[0085] The role of multi-dimensional logarithmic deviation loss is to quantify the characteristic differences of multi-dimensional bearing real-time data such as vibration signals, infrared temperature signals and acoustic signals, specifically calculate the ratio of real-time data X in each dimension to standard data, take the constant logarithmic value, and calculate the weighted average value after assigning weight coefficients to each dimension. In this way, the loss function is constructed to accurately measure the differences between multi-dimensional data, thereby improving the comprehensive diagnostic ability of the model for multi-modal data and the accuracy of fault severity assessment;

[0086] By setting the maximum mean difference loss, we can ensure the consistency of data distribution between the training set and the test set.

[0087] Through the joint optimization of the above loss functions, a deep learning model guided by physical information is constructed, and the model performance is verified using a test set to obtain the optimal diagnostic model.

[0088] In order to make the content of the non-contact bearing fault diagnosis method based on the physics-guided secondary convolutional neural network more specific, the technical details such as specific data size and network parameters will be further described in detail below.

[0089] In the present invention, the bearing vibration data is acquired contactlessly based on the laser self-mixing interference technology, and the sampling frequency of the data is set to 50kHz to ensure that the slight vibration changes of the bearing can be captured under high speed conditions. The sampling frequency can effectively capture high-frequency vibration signals and adapt to a variety of speeds.

[0090] Each sampling time window is set to 10 seconds to ensure that the data collected in each time window is sufficient to reflect the changes in the bearing operating status.

[0091] In addition to bearing vibration signals, this method can also simultaneously collect bearing data in multiple different dimensions such as infrared temperature and acoustic signals, providing strong support for bearing fault diagnosis. Specifically, by collecting infrared temperature data, the temperature distribution information of the bearing base surface, bearing outer ring, bearing inner ring and rolling element is obtained, and the characteristics of the acoustic signal and the characteristics of the vibration signal are extracted. Using these multimodal data, combined with acoustic characteristics and / or vibration characteristics, the fault type and degree of the bearing are determined; after determining the fault type, the fault location is further determined through temperature distribution information, including bearing outer ring fault, bearing inner ring fault, bearing rolling element fault and bearing base fault. Here, a method for obtaining the degree of bearing fault is provided: real-time bearing data of multiple different dimensions such as infrared temperature, acoustic signal, vibration signal, etc. are collected. For each dimension of real-time bearing data, the ratio X / Y between the real-time bearing data X and the standard bearing data Y in a preset time period is first calculated, and then the constant logarithmic value lg (X / Y) of the ratio is assigned a corresponding weight coefficient to the constant logarithmic value in each dimension, and the weighted average value of the constant logarithmic value and the weight coefficient in multiple dimensions is calculated; after determining the type of bearing fault, the degree of bearing fault is obtained according to the weighted average value of the constant logarithmic value and the weight coefficient in multiple dimensions. Through the above method, the degree of fault can be obtained more objectively and accurately in combination with multi-dimensional data. Preferably, in the process of deep learning model training, combined with multi-dimensional logarithmic deviation constraints, the normalized fault degree index is integrated into the loss function of the model, and the classification performance and fault severity assessment ability of the model are optimized through multi-modal data, so as to realize the feature extraction of multi-modal data and the linkage optimization of loss function. Preferably, the logarithmic values ​​of the constants in multiple dimensions and the weighted average values ​​of the weight coefficients are normalized, and the normalized value of the logarithmic values ​​of the constants in multiple dimensions and the weighted average values ​​of the weight coefficients after the normalization is used as the degree of bearing fault, so as to facilitate technicians to understand the degree of fault more clearly.

[0092] Wavelet transform is used for signal denoising, and three-level wavelet decomposition is selected to filter high-frequency noise. The processed bearing vibration data retains the characteristics related to bearing faults. A bandpass filter is used to limit the signal frequency range to within 20kHz to remove unnecessary ultra-low frequency and ultra-high frequency interference signals.

[0093] The processed bearing vibration data is sampled by sliding window, and the size of the sliding window is 2048. In order to ensure that the deep learning model has strong generalization ability, the training samples cover different fault types and operating states. The bearing vibration data set is divided into 80% training set, 10% validation set, and 10% test set.

[0094] The bearing physical degradation model including bearing degradation law, operating conditions and peak sparsity is embedded into the secondary convolutional neuron to construct the secondary physical convolutional neuron. This neuron can simulate the fault vibration characteristics of the bearing under different working conditions, especially it is highly sensitive to the key features under abnormal operating conditions, which helps the secondary physical neural network to extract nonlinear characteristics and physical information related to bearing faults.

[0095] A secondary physical neural network is constructed based on secondary physical convolutional neurons. The network structure includes input layer, convolution layer, fully connected layer and output layer. The input layer is used to receive the divided samples. The convolution layer performs nonlinear mapping and feature fusion on the features. The fully connected layer further extracts the fused deep features through the fully connected network and maps the high-dimensional features to the classification space. At the same time, the multi-task structure of fault location and type is combined to improve the diagnosis accuracy. The output layer classifies according to the fault type.

[0096] The data set is input into the input layer of the secondary physical neural network. The middle layer contains 6 convolutional layers. The 6 convolutional layers can perform deep feature extraction on the bearing vibration signal, especially the capture of high-order features, which enhances the model's ability to identify different fault types.

[0097] Each layer extracts the features of the bearing vibration signal through convolution. The first convolution layer: the convolution kernel size is (3,1), the number of convolution kernels is 64, the step size is 1, and the ReLU activation function is used. This layer mainly extracts the local features of the signal. The second convolution layer: the convolution kernel size is (3,64), the number of convolution kernels is 128, the step size is 1, and the ReLU activation function is used to further extract the mid-frequency band features of the vibration signal. The third convolution layer: the convolution kernel size is (3,128), the number of convolution kernels is 256, the step size is 1, and the ReLU activation function is used to capture higher-order features. The fourth convolution layer: the convolution kernel size is (3,256), the number of convolution kernels is 512, the step size is 1, and the ReLU activation function is used to extract deep features. The fifth convolution layer: the convolution kernel size is (3,512), the number of convolution kernels is 1024, the step size is 1, and the ReLU activation function is used. The sixth convolution layer: The convolution kernel size is (3,1024), the number of convolution kernels is 2048, the step size is 1, and the ReLU activation function is used to map high-order features to a lower-dimensional space.

[0098] The output of the convolutional layer is globally averaged pooled or flattened into a one-dimensional vector and then input into the fully connected layer for fault classification. Assuming there are four types of faults, the fully connected layer has four nodes and outputs the probability distribution of the fault classification.

[0099] The Softmax function of the output layer outputs the probability of each type of fault. Assume that the output of a bearing is [0.05, 0.2, 0.1, 0.65, 0.0], among which the fourth type has the highest probability, indicating that fatigue cracks appear in the inner ring of the bearing.

[0100] Preferably, the fully connected layer is provided with NxM nodes, N represents the fault location, including the inner ring, outer ring and rolling element of the bearing, and M represents the fault type, including wear, crack, fatigue, deformation, overheating, etc. By setting NxM nodes, the location and type of the bearing fault can be output more accurately to improve the accuracy of bearing fault diagnosis.

[0101] The loss function adopts a physical heuristic design based on the characteristic parameters of bearing operation, combining the maximum mean difference method, multi-dimensional logarithmic deviation constraint, kurtosis optimization sparsity and classification error. This loss function can effectively measure the difference in feature distribution between normal working conditions and fault conditions, maximize the mean difference between the two, and multi-dimensional logarithmic deviation constraint: Combining the multi-modal characteristics of bearing vibration signals, infrared temperature and acoustic signals, using the logarithmic deviation weighted constraint method, the model's comprehensive diagnostic ability for multi-modal data is improved.

[0102] The Adam optimizer is used to update parameters to adaptively adjust the learning rate and quickly converge to the global optimal solution.

[0103] In the above-mentioned physics-guided secondary convolutional neural network, a multi-channel structure is introduced to further enhance the processing capability of multi-modal data. The vibration signal, infrared temperature signal and acoustic signal of the bearing are used as multiple input channels, and each channel represents the data of one mode.

[0104] In the above-mentioned non-contact bearing fault diagnosis method based on physics-guided secondary convolutional neural network, after training, the secondary physical neural network can automatically identify different fault modes and accurately classify the health status of the bearing. The fault modes include common bearing wear, cracks, fatigue and other fault types.

[0105] In the above-mentioned non-contact bearing fault diagnosis method based on physics-guided secondary convolutional neural network, the system automatically generates a corresponding diagnostic report according to the diagnosis results, indicating the fault type, severity and possible causes of the bearing. At the same time, it issues an alarm in real time when a fault occurs, helping equipment maintenance personnel to take timely measures.

[0106] As the system continues to run, the non-contact bearing fault diagnosis method based on the physical guided secondary convolutional neural network also has self-learning capabilities. Its working principle is: by continuously accumulating historical vibration data of the bearing, especially data at different fault stages. Using these data, the secondary physical neural network model is regularly updated so that it can adapt to new fault modes under different working conditions, thereby improving the accuracy and stability of the model.

[0107] In summary, the present invention provides a non-contact fault diagnosis method for bearings based on a secondary convolutional neural network. The bearing vibration data is collected with high precision through laser self-mixing interference technology, and the bearing physical degradation model and deep learning model are combined. The physical-inspired loss function is optimized to achieve real-time monitoring, efficient identification and precise positioning of bearing faults, which significantly improves the accuracy and robustness of fault diagnosis and has broad industrial application prospects.

[0108] The working principle of the present invention is as follows:

[0109] 1. Vibration data collection: Laser self-mixing interference technology is used to collect bearing vibration data. This technology can capture its tiny vibration changes without contacting the bearing, ensuring the accuracy of data collection. The sampling frequency is set to 50kHz, which can accurately obtain high-frequency vibration information and is suitable for detection requirements under various bearing speeds. Each sampling window is set to a data length of 10 seconds to ensure sufficient time to reflect the operating status of the bearing.

[0110] 2. Data preprocessing: The collected vibration signal is denoised by wavelet transform, and a bandpass filter is used to limit the signal frequency range to less than 20kHz to filter out irrelevant high-frequency or low-frequency noise. Then, the signal is divided into 2048 data points by sliding window technology and input into the model. In order to improve the generalization ability of the model, the vibration data set is reasonably divided, 80% as a training set, 10% as a validation set, and 10% as a test set.

[0111] 3. The secondary physical neural network introduces a bearing physical degradation model, which can simulate the operating characteristics of bearings under different working conditions. Through the structure of 6 convolutional layers, the network extracts high-order features of vibration signals layer by layer. Each layer of convolution kernel is designed for different frequency band characteristics, and the ReLU activation function is used to capture nonlinear fault characteristics in vibration signals. The 1st to 3rd layer of convolution kernels are mainly responsible for extracting local and medium-frequency features of vibration signals. The 4th to 6th layers gradually extract high-order features, capture the deep characteristics hidden in the bearing during the degradation process, and map high-dimensional features to low-dimensional space to provide more representative feature vectors for classification tasks.

[0112] 4. Fault classification and output: The features processed by the convolution layer are flattened into a one-dimensional vector through the global average pooling layer and input into the fully connected layer. The output of the fully connected layer contains 4 nodes, corresponding to 4 types of faults. The probability distribution of each type of fault is calculated through the Softmax function, and the most likely fault type is output.

[0113] 5. Physically guided loss function: The loss function is combined with the maximum mean difference method (MMD) to measure the difference between the normal working condition and the fault condition feature distribution, ensuring that the model can effectively distinguish between normal and faulty states; the kurtosis is combined to optimize sparsity, enhancing the model's sensitivity to periodic shocks and local abnormal features in vibration signals; the multi-dimensional logarithmic deviation is used to constrain the feature differences of different data dimensions to ensure the accurate fusion and optimization of multimodal features. The joint optimization of the three objectives not only improves the diagnostic accuracy and robustness of the model, but also effectively accelerates the convergence speed of the network under physical constraints. The Adam optimizer is used to adaptively adjust the learning rate to ensure rapid convergence to the global optimal solution.

[0114] 6. Self-learning and report generation: In long-term operation, the present invention has the ability of self-learning. Through the accumulated historical data, especially the data of the fault development process, the model is regularly updated to better adapt to the emerging fault mode. Based on the diagnosis results, the system automatically generates a report containing the fault type, severity, cause analysis, and issues an alarm when necessary to help equipment maintenance personnel take timely action.

[0115] In summary, the innovation and technical advantages of the bearing non-contact fault diagnosis method based on secondary physical neural network include:

[0116] 1. Non-contact bearing vibration data collection based on laser self-mixing interference technology not only realizes high-precision non-contact collection of vibration data, but also can maintain data stability under extreme working conditions, and is suitable for various complex industrial environments.

[0117] 2. The quadratic physical neural network architecture combined with the bearing physical degradation model greatly improves the efficiency of fault feature extraction and the accuracy of fault diagnosis, especially in the early stages of a fault, and can detect subtle anomalies that are difficult to detect with traditional methods.

[0118] 3. Based on the bearing physical degradation model, the maximum mean difference method, kurtosis optimization sparsity and multi-dimensional logarithmic deviation jointly optimize the network loss function, and further optimize the deep learning model with the participation of the test set, it can accurately simulate the fault vibration characteristics of the bearing under different working conditions, and then effectively reflect the abnormal operating state of the bearing, so that the secondary physical neural network can more efficiently extract the nonlinear characteristics related to the fault.

[0119] 4. The fault type and degree of the target bearing are determined based on the optimal diagnostic model. There is no need to install a contact sensor, which not only overcomes the problems of difficult installation and easy damage of contact sensors in traditional bearing fault diagnosis methods, but also avoids the interference of traditional bearing fault diagnosis methods to the bearing structure due to the need to install contact sensors, thereby improving the measurement accuracy and stability.

[0120] 5. Combining the bearing physical degradation model with the secondary convolutional neuron to build a deep learning model can achieve efficient and accurate diagnosis of bearing faults, and effectively solve the problems of low recognition efficiency, low accuracy and low physical credibility of traditional neural network fault models.

[0121] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0122] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A non-contact bearing fault diagnosis method based on a secondary convolutional neural network, characterized in that: The bearing non-contact fault diagnosis method based on physically guided secondary convolutional neural network comprises the following steps: Collect bearing vibration data based on non-contact method; Dividing the bearing vibration data into samples to obtain a training set and a test set; Acquire a bearing physical degradation model according to peak sparsity, acquire a secondary physical neural network according to the bearing physical degradation model and secondary convolution neurons, and construct a deep learning model of bearing physical information according to the secondary physical neural network; Iteratively training the deep learning model using the training set to extract key characteristic parameters of the bearing vibration data; Based on the key feature parameters, the network loss function of the deep learning model is optimized in combination with the maximum mean difference method, the kurtosis optimization sparsity constraint and the multi-dimensional logarithmic deviation; The deep learning model is verified by using the optimized network loss function and test set to obtain the optimal diagnostic model; Real-time vibration data of the target bearing operation is obtained and input into the diagnostic model, the fault type and fault degree of the target bearing are determined, and non-contact fault diagnosis of the target bearing is achieved.

2. A non-contact bearing fault diagnosis method based on a secondary convolutional neural network as claimed in claim 1, characterized in that: The specific method for collecting bearing vibration data based on a non-contact method comprises the following steps: Based on the laser self-mixing interference technology, the laser beam is irradiated on the surface of the bearing base, the laser information reflected from the surface of the bearing base is detected, and the interference signal is obtained; Bearing vibration data is collected according to the phase change of the interference signal.

3. A non-contact bearing fault diagnosis method based on a secondary convolutional neural network as claimed in claim 2, characterized in that: The bearing physical degradation model includes peak sparsity, bearing degradation law and bearing fault vibration characteristics under different working conditions; the method for obtaining the bearing physical degradation model includes: constructing a bearing physical degradation model according to the peak sparsity of the bearing vibration signal and the bearing degradation law, which is used to characterize the physical degradation characteristics and fault characteristics of the bearing under different operating conditions.

4. A non-contact bearing fault diagnosis method based on a secondary convolutional neural network as claimed in claim 3, characterized in that: A specific method for obtaining a bearing physical degradation model according to peak sparsity and obtaining a secondary physical neural network according to the bearing physical degradation model and secondary convolution neurons includes the following steps: Obtaining bearing physical degradation models based on peak sparsity According to the bearing physical degradation model and the secondary convolutional neuron Get the secondary physical neural network σ(f(x))=σ((w r L vib-kurt +b r )(w g L vib-kurt +b g )+w b (L vib-kurt ⊙L vib-kurt )+m); in, vib represents the peak value of the vibration signal, stiff represents the equivalent contact stiffness level between the bearing ball and the inner and outer ring raceways, m represents the equivalent mass of the system, ε represents the related imbalance in the system load, V represents the speed, and w r 、w g 、w b represents the weight, b r 、b g and c denote bias parameters, σ(·) is the activation function, ⊙ denotes the Hadamard product, x i represents the i-th sample data, x j represents the jth sample data, P represents the total number of samples, Represents weight.

5. A non-contact bearing fault diagnosis method based on a secondary convolutional neural network as claimed in claim 4, characterized in that: The secondary physical neural network comprises: An input layer, used for receiving the bearing vibration data after sample division; A convolution layer, constructed based on the secondary physical convolution neuron, for performing nonlinear mapping and feature fusion on key features of bearing vibration data; The fully connected layer further extracts the fused high-dimensional features; The output layer is used to classify and quantitatively evaluate the type and degree of the bearing fault.

6. A non-contact bearing fault diagnosis method based on a secondary convolutional neural network as claimed in claim 5, characterized in that: The loss function of the quadratic physical neural network includes kurtosis optimization and sparsity constraints Multi-dimensional logarithmic deviation loss and maximum mean difference loss The total loss function is L total =αL s-MMD +βL log-dev +γL s-kurt ; Among them, y(n) is the time series, Δ norm is the weighted average of the normalized logarithmic deviation, where the logarithmic deviation is X i , Y i are the real-time data and standard data in the i-th dimension, φ(·) represents the Hilbert space of the reproducing kernel, D s and D t Represent the sample sets of original working condition and target working condition respectively, n s and n t represents the number of samples in the original working condition and the target working condition respectively, and optimizes the network loss function L of the deep learning model total =αL s-MMD +βL log-dev +γL s-kurt , α, β, and γ are the weights of different loss terms respectively.

7. A non-contact bearing fault diagnosis method based on a secondary convolutional neural network as claimed in claim 6, characterized in that: The vibration data includes vibration displacement, vibration velocity, vibration acceleration and rotation speed, and the fault types include bearing wear, cracks and fatigue.

8. A non-contact bearing fault diagnosis method based on a secondary convolutional neural network as claimed in claim 7, characterized in that: Before dividing the bearing vibration data into samples, the bearing vibration data is preprocessed.

9. A non-contact bearing fault diagnosis method based on a secondary convolutional neural network as claimed in claim 8, characterized in that: The preprocessing includes wavelet transform denoising and bandpass filtering.

10. A non-contact bearing fault diagnosis method based on a secondary convolutional neural network as claimed in claim 9, characterized in that: The method also includes the following steps: obtaining bearing data of multiple different dimensions, and determining the bearing fault location and fault degree according to the bearing data of multiple different dimensions.

Images