Rolling bearing fault diagnosis monitoring method based on vibration and current dual modules

Abstract

The application provides a vibration and current double-module rolling bearing fault diagnosis monitoring method, which comprises an offline modeling stage and an online diagnosis stage; first, multivariate bearing data are acquired, including five different variable signal types of current, torque, vibration, force and speed; second, a double-module neural network model is constructed, training samples are decomposed into multiple modules, a vibration and current double-module rolling bearing fault diagnosis model is constructed by analyzing the dissimilarity between the modules, the activity degree of the training samples in the modules and the difference between the data of the bearing variables, and iterative training is carried out; finally, test samples are taken as the input of the optimized vibration and current double-module rolling bearing fault diagnosis model, fault diagnosis results under different models and output forms are given through experiments, and the fault diagnosis task is completed. The fault diagnosis model composed of the double-module structure and three-layer convolution blocks can effectively reduce the model complexity and improve the feature learning ability of the diagnosis network.

Description

Technical Field

[0001] This invention relates to the field of multivariable complex signal rolling bearing fault diagnosis technology, and in particular to a rolling bearing fault diagnosis and monitoring method based on vibration and current dual modules. Background Technology

[0002] As a core component of rotating machinery, rolling bearings pose a significant risk of failure, immediately impacting equipment and making safe and normal operation under complex conditions extremely difficult. Because the operating conditions of rotating machinery are affected by external factors such as environment, temperature, and noise, rolling bearings have a high risk of damage. Unexpected damage to bearing components can lead to potential risks to personnel safety, industrial operation, and costs. Therefore, research into complex, multivariable bearing signals and the development of an intelligent and efficient fault diagnosis method have become crucial in the field of fault diagnosis.

[0003] In recent years, fault diagnosis technology has failed to demonstrate its advantages in real-time monitoring despite the rapid development of mechanical equipment, exhibiting drawbacks such as low efficiency and misjudgment. Traditional analysis methods have made some progress in bearing fault diagnosis, but still suffer from limitations such as the inability to extract multivariate feature information, difficulties in processing complex information, and interference from environmental factors on diagnostic results. To achieve the rapid and efficient application of neural networks, this paper introduces the modular approach to problem processing into neural networks, leveraging the characteristics of multivariate and multi-module information processing. The modular approach based on deep learning, as a highly efficient pattern recognition network, offers broad prospects for the development of intelligent fault diagnosis methods. Summary of the Invention

[0004] To address the challenges of multivariable bearing signal feature extraction, model complexity analysis, and optimal offline modeling, this invention proposes a rolling bearing fault diagnosis and monitoring method based on a dual-module approach using vibration and current. A convolutional neural network is employed as the dual-module structure for corresponding variable fault diagnosis to ensure both the simplicity of the overall model structure and the accuracy of the diagnostic results.

[0005] The technical solution of this invention is implemented as follows:

[0006] A method for fault diagnosis and monitoring of rolling bearings based on a dual-module approach of vibration and current includes the following steps:

[0007] S1: Offline Modeling Stage

[0008] Multivariate bearing data was acquired as a dataset, which was then divided into training samples and test samples. The dataset included five different variable signal types: current, torque, vibration, force, and velocity. Each signal type included normal bearing data and faulty bearing data.

[0009] A dual-module neural network model is constructed. The training samples are decomposed into M subsets and assigned to M modules. By analyzing the dissimilarity between modules, the activity of training samples in modules, and the differences in data between bearing variables, a dual-module rolling bearing fault diagnosis model for vibration and current is constructed and iteratively trained.

[0010] S2: Online Diagnosis Phase

[0011] The test samples are used as input to the optimized vibration and current dual-module rolling bearing fault diagnosis model. The fault diagnosis results under different models and output forms are given through experiments to complete the fault diagnosis task.

[0012] Preferably, the method for constructing the dual-module neural network model is as follows:

[0013] Obtain the bearing dataset X = {(x1,z1),(x2,z2),…,(x e ,z e )},x e Let x represent the e-th sample of the signal variable. e ∈R e ;z e This represents the probability of the existence of the e-th sample of the signal variable, 0 ≤ z e ≤1;

[0014] Calculate the difference value u of the bearing signal d :

[0015]

[0016] Among them, u d For the p-th dataset (x) p ,z p ) and the qth dataset (x) q ,z q The difference between the signals;

[0017] Obtain the optimal nearest neighbor dataset and the optimal nearest neighbor difference value:

[0018] Optimal nearest neighbor dataset D(x) i ,z i ) is the i-th dataset (x) i ,z i With ) as the core, satisfying |(u dn -u de ) / u de For datasets |≤s, where u dn Indicates the current signal difference value; u de This represents the difference value of the previous signal; s represents the threshold parameter, which takes the value (0,1);

[0019] Optimal nearest neighbor difference value O(u)d ) is based on the dataset (x) i ,z i The core and the optimal nearest neighbor dataset D(x) i ,z i The average value of the difference values ​​of the signals in the given data is calculated using the following formula:

[0020]

[0021] Among them, u dij This represents the signal difference value between the i-th data group and the j-th data group;

[0022] If different signal datasets (x) p ,z p ) and (x q ,z q If the inclusion relation is satisfied, then:

[0023]

[0024] According to the dataset (x) i ,z i The optimal nearest neighbor dataset D(x) i ,z i ) and the maximum difference value Max(u d ) Calculate the dissimilarity F of the sub-models f :

[0025]

[0026] Based on the dissimilarity F of the sub-model f and the difference value of the best nearest neighbor O(u) d Determine the degree of correlation between sub-models (FG):

[0027]

[0028] The optimal sub-model is determined based on the value of the correlation degree between sub-models, FG.

[0029] Sub-model weight assignment:

[0030] Calculate the activity level of the training samples: When the Rth training sample enters the module, the activity level of the training sample is:

[0031] G(R) = [G1(R), ..., G m (R),…,G M (R)];

[0032] The sample with the highest activity level is:

[0033] R max =arg max[G(R)];

[0034] When the training samples enter the sample with the highest activity level R max When assigning a module, assign R max The module's corresponding weights complete the sample training; the weight coefficients are shown below:

[0035]

[0036] The diagnostic results are integrated and output through the output layer, as shown in the formula below:

[0037]

[0038] in, This represents the output of the d-th module network; This represents the weight coefficients from the d-th network to the k-th output; This indicates the output fault diagnosis result value.

[0039] Preferably, the dual-module neural network model includes two neural network sub-models. The network structure of the neural network sub-model is: convolutional layer I - pooling layer I - convolutional layer II - pooling layer II - convolutional layer III - pooling layer III - fully connected layer. The sub-model extracts feature information using feature extraction, global connection, and weight sharing operations. The feature information is integrated and output through weighted summation.

[0040] Preferably, the specific implementation method of the online diagnosis stage is as follows:

[0041] The data acquisition module is activated to record the sampling frequency, number of samplings, and variable type of the bearing signal in real time to obtain bearing data.

[0042] For the vibration signal X1 and current signal X2 of the bearing data, the vibration signal... and current signal After inputting the two sub-models from the dual-module neural network model respectively, the diagnostic results are output by integration.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] 1) This invention uses multivariate original bearing data as model input, and the dual-module fault diagnosis concept of neural network has advantages in vibration and current multivariate signal analysis.

[0045] 2) This invention utilizes a fault diagnosis model composed of a multivariate dual-module structure and three-layer convolutional blocks, which effectively reduces model complexity and improves the diagnostic network's ability to learn features.

[0046] 3) This invention has advantages in improving diagnostic accuracy and reducing training time by constructing a dual-module system to learn multivariate signal feature information and integrating the output fault diagnosis results. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of the fault diagnosis process of the present invention;

[0049] Figure 2 This is a diagram of the diagnostic network structure of the present invention;

[0050] Figure 3 The diagram shows the results of the integrated output and direct output confusion matrix of this invention; where (a) is the integrated output and (b) is the direct output.

[0051] Figure 4 This is a diagram showing the fault diagnosis results for different training samples in this invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] like Figure 1 As shown in the figure, this invention provides a method for fault diagnosis and monitoring of rolling bearings based on a dual-module approach of vibration and current. The specific steps are as follows:

[0054] S1: Offline modeling stage: Use the rolling bearing data acquisition platform to acquire multivariable bearing signals, and construct a rolling bearing fault diagnosis model based on vibration and current dual modules by analyzing the dissimilarity between modules, the activity level of samples in modules and the differences in signals between bearing variables, and perform iterative training.

[0055] S2: Online Diagnosis Phase: The test samples are used as input to the optimized vibration and current dual-module rolling bearing fault diagnosis model. Fault diagnosis results under different models and output forms are given through experiments to complete the fault diagnosis task.

[0056] In step S1, multivariate bearing signals were first acquired: the rolling bearing dataset provided by the University of Paderborn, Germany, was used for the experiment. The experimental data consisted of five different variable signal types: current, torque, vibration, force, and velocity. Each variable type included normal bearing data and faulty bearing data. Faulty bearings were further divided into artificial damage and actual damage based on the type of damage. Artificial damage mainly included cracks, spalling, and pitting, while actual damage was obtained through an accelerated life testing bench. From the obtained raw data, the experiment selected 3 healthy bearings, 6 bearings with artificial damage, and 6 bearings with actual damage under operating conditions of 900 rpm, 0.7 Nm torque load, and 1000 N radial force, for a total of 15 different types of bearing data.

[0057] The dataset description is shown in Table 1. Normal represents normal bearing data. OR1 and OR2 represent outer ring fault data due to artificial and actual bearing damage, respectively. IR1 and IR2 represent inner ring fault data due to artificial and actual bearing damage, respectively. Normal bearing data is further subdivided into three different types based on different break-in periods, radial loads, and speeds. Outer and inner ring faults caused by different types of damage are also categorized into 12 different types of fault data based on different damage degrees and types: OR1_1, OR1_2, OR1_3, OR2_1, OR2_2, OR2_3, IR1_1, IR1_2, IR1_3, IR2_1, IR2_2, and IR2_3. The damage degree in the fault data is determined by the damage level: 0-2% fault percentage and damage size less than 2mm are classified as Level 1 damage; 2-5% fault percentage and damage size greater than 2mm are classified as Level 2 damage; and 5-15% fault percentage and damage size greater than 4.5mm are classified as Level 3 damage. Labels 0-15 represent 15 different types of bearing data.

[0058] Table 1 Description of the experimental dataset

[0059]

[0060] Secondly, a dual-module neural network model is constructed. First, based on the different multivariate data of the rolling bearing, the model is decomposed into multiple modules. If the bearing dataset is X={(x1,z1),(x2,z2),…,(x…z2),…,(x…z2)…z2 ... e ,z e )},x e Let x represent the e-th sample of the signal variable. e ∈R e ;z e This represents the probability of the existence of the e-th sample of the signal variable, 0 ≤ z e≤1. The multivariate bearing data is decomposed into M subsets and assigned to M modules for fault diagnosis. Secondly, model construction is the core step in ensuring network performance and the diagnostic task. Selecting a model with a concise structure, good generalization, and robustness is crucial for model training and testing results, achieving a non-linear mapping from input to output. Finally, the output uses different weights assigned to each module to integrate the fault diagnosis results.

[0061] The key to model construction is determining the optimal modules. The module division is based on the difference value u of the bearing signals. d Analyze the differences in signals among different bearing variables, for the p-th dataset (x p ,z p ) and the qth dataset (x) q ,z q The formula for the difference between signals is shown below:

[0062]

[0063] Among them, u d For the p-th dataset (x) p ,z p ) and the qth dataset (x) q ,z q The difference between signals.

[0064] Obtain the optimal nearest neighbor dataset and the optimal nearest neighbor difference value: Optimal nearest neighbor dataset D(x) i ,z i ) and the optimal nearest neighbor difference value O(u d Determining the optimal nearest neighbor dataset is a key indicator for differentiating data type building blocks. The optimal nearest neighbor dataset is generated through dataset optimization, resulting in strong differences in data among signal variables.

[0065] Optimal nearest neighbor dataset D(x) i ,z i ) is the i-th dataset (x) i ,z i With ) as the core, satisfying |(u dn -u de ) / u de For datasets |≤s, where u dn Indicates the current signal difference value; u de This represents the difference value of the previous signal; s represents the threshold parameter, which takes the value (0,1).

[0066] Optimal nearest neighbor difference value O(u) d ) is based on the dataset (x) i ,z i The core and the optimal nearest neighbor dataset D(x) i ,zi The average value of the difference values ​​of the signals in the given data is calculated using the following formula:

[0067]

[0068] in, This represents the signal difference between the i-th data set and the j-th data set.

[0069] By analyzing the inclusion of signals between datasets, we can determine the true differences between signals, that is, whether the optimal neighbor difference value of any dataset contains signals of different categories, and construct network models for different tasks. If different signal datasets (x...) p ,z p ) and (x q ,z q If the inclusion relation is satisfied, then:

[0070]

[0071] According to the dataset (x) i ,z i The optimal nearest neighbor dataset D(x) i ,z i ) and the maximum difference value Max(u d ) Calculate the dissimilarity F of the sub-models f :

[0072]

[0073] Dissimilarity F of sub-models f This will serve as an important indicator for judging the differences between sub-models constructed with different variables. Based on the dissimilarity Fi of the sub-models... f and the difference value of the best nearest neighbor O(u) d Determine the degree of correlation between sub-models (FG):

[0074]

[0075] The correlation degree (FG) between sub-models is key to determining the optimal sub-model. Quantifying the correlation degree ensures the optimal sub-network model. The number of optimal sub-models is determined by the compactness and sparsity between sub-networks. The larger the values ​​of compactness and sparsity, the larger the value of FG, indicating a higher degree of sub-network optimization.

[0076] The network outputs fault diagnosis results through the output layer, which then integrates the results. Signals are assigned to corresponding modules based on the different bearing data, and the weights of these modules are determined by the activity values ​​of the samples.

[0077] When the Rth training sample enters the module, the activity level of the training sample is:

[0078] G(R) = [G1(R), ..., G m (R),…,G M (R)];

[0079] The sample with the highest activity level is:

[0080] R max =arg max[G(R)];

[0081] When the training samples enter the sample with the highest activity level R max When assigning a module, assign R max The module's corresponding weights complete sample training; the output layer integrates and outputs the diagnostic results, as shown in the following formula:

[0082]

[0083] in, This represents the output of the d-th module network; This represents the weight coefficients from the d-th network to the k-th output; This indicates the output fault diagnosis result value.

[0084] In the fault diagnosis and monitoring model, network decomposition and integrated output are closely related. Since each sub-network represents the bearing fault diagnosis result, weight parameters are used to set the integrated output diagnosis result. The weight coefficients are set as follows:

[0085]

[0086] Finally, the parameters of the neural network were determined: the dual-module neural network model consists of two neural network sub-models. The network structure of the neural network sub-model is: convolutional layer I - pooling layer I - convolutional layer II - pooling layer II - convolutional layer III - pooling layer III - fully connected layer. The sub-model uses the ideas of feature extraction, global connection and weight sharing to extract feature information. The extracted data features are integrated and output through the setting of weight parameters.

[0087] The parameter settings for the vibration and current dual-module rolling bearing fault diagnosis and monitoring model are shown in Table 2. The model mainly consists of three two-dimensional convolutional blocks. The kernel size of each convolutional block is 5×5, and the number of convolutional kernels is 32, 64, and 128 respectively. Compared with traditional convolutional neural network models, the complexity of this model is significantly reduced, effectively reducing the model training and testing time. Through multiple iterations of training, the decay rate, number of iterations, learning rate, and training batches were determined to be 0.9, 100, 0.0003, and 64, respectively.

[0088] Table 2. Parameter settings for the rolling bearing fault diagnosis and monitoring model based on vibration and current dual modules.

[0089]

[0090]

[0091] In step S2, the fault diagnosis module function is first verified: after the data acquisition module is started, the sampling frequency, number of samples, and variable type of the bearing signal are recorded in real time. After the sampling is completed, the bearing data is saved, and the diagnostic module is verified.

[0092] In the fault diagnosis module, the acquired multivariable bearing data undergoes feature analysis, and a diagnostic model is constructed based on these feature variables as input data for diagnosis. After the fault diagnosis command is activated, the fault diagnosis task begins. The system analyzes and identifies the multivariable bearing data according to the preset model, and trains the model using preset parameters, thereby achieving fault diagnosis of the multivariable bearing data. After the analysis is complete, the system outputs the diagnostic results, completing the functional test.

[0093] Upon completion of data acquisition, the bearing data is first read and variable feature analysis is performed. After feature analysis, the model enters the fault diagnosis state. The acquired raw vibration and current variable data are locally stored and visualized, and used as input to the dual-module analog signal channel. Bearing signal feature information of different variables is acquired and saved. Finally, the classifier of the fault diagnosis model performs fault classification and diagnosis, outputting the fault diagnosis results of the bivariate bearing signal of vibration and current.

[0094] like Figure 2 The diagram shows the network structure for rolling bearing fault diagnosis based on a dual-module system of vibration and current signals. A two-dimensional convolutional neural network is used as the dual-module network to diagnose rolling bearing faults based on the vibration and current signals X1 and X2 of the bearing. Figure 2 The sample set consists of decomposed signal 1 and decomposed signal 2. Decomposed signals X1 and X2 represent vibration signals respectively. and current signal As input to the diagnostic network; This output represents the integrated fault diagnosis results of vibration and current signals.

[0095] Secondly, an effectiveness experiment for online bearing fault diagnosis was conducted: training samples were used as input to the fault diagnosis model for diagnostic testing. The effectiveness of online bearing fault diagnosis was tested using different bearing states, output methods, model structures, and numbers of training samples. The experimental results are shown below.

[0096] Table 3 shows a comparison of the fault diagnosis results using integrated output and direct output in this invention, recording the precision, recall, and F1 score of 15 different tags. Figure 3This is a confusion matrix diagram showing the results of the integrated output and direct output methods. In the University of Paderborn dataset in Germany, the integrated output diagnostic accuracy is 99.76%. The fault diagnosis model, which uses a multivariate dual-module integrated output diagnostic method, achieves higher diagnostic accuracy than the direct output method.

[0097] Table 3 Comparison of Integrated Output and Direct Output Diagnostic Results

[0098]

[0099]

[0100] Table 4 shows a comparison of fault diagnosis results between the dual-module network of the present invention using different convolutional structures and the direct output. The comparison results show that the present invention has advantages in improving diagnostic accuracy and reducing training time. Using a three-layer convolutional module for integrated output improves diagnostic accuracy by 5.16% compared to direct output, while the three-layer convolutional structure reduces training time by 6.56 seconds compared to a four-layer structure.

[0101] Table 4. Fault diagnosis results for different numbers of modules under different output modes.

[0102]

[0103] Table 5 Fault diagnosis results of different methods

[0104]

[0105] Table 5 shows the fault diagnosis results of different diagnostic methods using the same dataset. Stacked Autoencoder (SAE), Convolutional Neural Network (CNN), ResNet, AResNet (Attention Mechanism Residual Network), CWTResNet (Continuous Wavelet Transform Residual Network), MSAResNet (Multi-Scale Attention Mechanism Residual Neural Network), and this invention all used the same simulation environment and model parameters for diagnostic experiments. This invention shows a significant advantage in fault diagnosis accuracy for multivariate bearing signals compared to other methods. This invention utilizes the multivariate dual-module approach to maximize the extraction of data features from different variables and integrate the results for output. It significantly reduces model complexity and improves fault diagnosis accuracy while simplifying the model structure.

[0106] During model training, the number of training samples affects the fault diagnosis performance of different methods. The fault diagnosis results for different training samples under different fault diagnosis methods are shown below. Figure 4As shown, each method was tested using the same number of randomly selected training samples. The fault diagnosis accuracy increased with the increase in the number of training samples, and the present invention achieved the highest diagnostic accuracy with the same number of training samples. With 600 samples, the training accuracy of the present invention exceeded 90%. When the number of training samples was greater than 1500, over 99% of the training samples could be correctly diagnosed, demonstrating good fault diagnosis classification results. Compared with other methods, the present invention significantly improves diagnostic performance under limited training sample conditions and is robust to changes in the number of training samples.

[0107] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for fault diagnosis and monitoring of rolling bearings based on a dual-module approach of vibration and current, characterized in that, The steps are as follows: S1: Offline Modeling Stage Multivariate bearing data was acquired as a dataset, which was then divided into training samples and test samples. The dataset included five different variable signal types: current, torque, vibration, force, and speed. Each signal type included normal bearing data and faulty bearing data. Construct a two-module neural network model, decomposing the training samples into A subset and assigned to A dual-module rolling bearing fault diagnosis model for vibration and current is constructed and iteratively trained by analyzing the dissimilarity between modules, the activity level of training samples in modules, and the differences in data between bearing variables. The method for constructing the dual-module neural network model is as follows: Obtain bearing dataset , The first signal variable represents the signal variable. One sample, ; Represents the signal variable number 1 The probability of a sample existing. ; Calculate the difference value of bearing signals : ； in, For the first Group dataset and the Group dataset The difference between the signals; Obtain the optimal nearest neighbor dataset and the optimal nearest neighbor difference value: Optimal Nearest Neighbors Dataset Therefore, the first Group dataset With the core, to meet The dataset, in which, Indicates the current signal difference value; Indicates the difference value of the previous signal; This represents the threshold parameter, with a value of [value missing]. ; Optimal Neighbor Difference Value It is a dataset Core and best nearest neighbor dataset The average value of the difference in the signal is calculated using the following formula: ； in, Indicates the first Group data and the first Signal differences between data sets; If different signal datasets and If the inclusion relation is satisfied, then: ； According to the dataset The best nearest neighbor dataset and maximum difference value Calculate the dissimilarity of sub-models : ； Based on the dissimilarity of the sub-models Difference value with the best nearest neighbor Determine the degree of correlation between sub-models : ； Based on the degree of correlation between sub-models The value determines the optimal sub-model; Sub-model weight assignment: Calculate the activity of the training samples: when the th After a training sample enters the module, the activity level of the training sample is as follows: ； The sample with the highest activity level is: ； When the training samples enter the most active samples When assigning modules, The module's corresponding weights complete the sample training; the weight coefficients are shown below: ； The diagnostic results are integrated and output by the output layer, as shown in the formula below: ； in, Indicates the first The output of the module network; Indicates the first The network to the first The weight coefficients of each output; This indicates the output fault diagnosis result value; S2: Online Diagnosis Phase The test samples are used as input to the optimized vibration and current dual-module rolling bearing fault diagnosis model. The fault diagnosis results under different models and output forms are given through experiments to complete the fault diagnosis task.

2. The method for fault diagnosis and monitoring of rolling bearings based on a dual-module approach of vibration and current as described in claim 1, characterized in that, The dual-module neural network model includes two neural network sub-models. The network structure of the neural network sub-model is: convolutional layer I - pooling layer I - convolutional layer II - pooling layer II - convolutional layer III - pooling layer III - fully connected layer. The sub-model extracts feature information using feature extraction, global connectivity, and weight sharing operations. The feature information is integrated and output through weighted summation.

3. The method for fault diagnosis and monitoring of rolling bearings based on a dual-module approach of vibration and current as described in claim 1, characterized in that, The specific implementation method of the online diagnosis phase is as follows: The data acquisition module is activated to record the sampling frequency, number of samplings, and variable type of the bearing signal in real time to obtain bearing data. Vibration signals for bearing data and current signal , to vibration signal and current signal After inputting the two sub-models from the dual-module neural network model respectively, the diagnostic results are output by integration. .

Images