A rolling bearing fault diagnosis method based on digital twinning

Abstract

A kind of rolling bearing fault diagnosis method based on digital twinning includes establishing the digital twin model of rolling bearing;Get the simulation vibration signal of digital twin model running in normal state, the simulation vibration signal of running in fault state;Wavelet transform is carried out to the simulation vibration signal of digital twin model running in normal state, the simulation vibration signal of running in fault state, corresponding two-dimensional time-frequency diagram is generated;ResNet-50 network is improved, and the two-dimensional time-frequency diagram is used to train the improved ResNet-50 network, and first fault diagnosis neural network is obtained;The first fault diagnosis neural network is migrated training, and second fault diagnosis neural network is obtained;The fault of actual running rolling bearing is detected using second fault diagnosis neural network;The design generates a large number of highly realistic simulation data to train the improved neural network, so that the improved neural network can accurately diagnose the fault.

Description

Technical Field

[0001] This invention relates to the fields of electronic engineering and computer science, and in particular to a method for diagnosing rolling bearing faults based on digital twins, specifically applicable to diagnosing bearing faults through bearing vibration signals. Background Technology

[0002] The rapid development of big data and artificial intelligence technologies has provided tremendous opportunities for intelligent manufacturing in industrial systems, while also posing greater challenges to the functionality, diversity, and safety of various mechanical equipment. Modern mechanical equipment needs to operate stably for extended periods, so even a minor malfunction can have serious consequences, ranging from equipment downtime to severe safety accidents. Existing research and applications show that early detection of malfunctions allows for pre-failure inspection and maintenance, preventing accidents, improving equipment efficiency, and significantly reducing maintenance costs. Rotating machinery accounts for approximately 80% of all mechanical equipment, and bearings are an irreplaceable component within rotating machinery. Statistics show that bearing failure accounts for 30% of all mechanical equipment failures.

[0003] In recent years, deep learning methods have been introduced into intelligent fault diagnosis and have achieved excellent results in bearing fault diagnosis. However, training neural networks in deep learning methods requires a large amount of sample data, which is difficult to obtain in actual bearing monitoring. At the same time, the sample data obtained in actual monitoring usually shows that the data of the equipment is operating normally is much larger than the data of the equipment is operating in faulty conditions, which can easily cause data imbalance and thus affect the diagnostic results of the neural network. Summary of the Invention

[0004] The purpose of this invention is to overcome the problem in the prior art that it is difficult to obtain sufficiently rich and balanced sample data for training neural network models in actual bearing monitoring. It provides a rolling bearing fault diagnosis method based on digital twins, which can establish a rolling bearing digital twin model that closely matches the physical entity, run the twin model to obtain a large amount of simulation data that closely matches reality to train the improved neural network, and enable the improved neural network to accurately diagnose faults.

[0005] To achieve the above objectives, the technical solution of the present invention is:

[0006] A rolling bearing fault diagnosis method based on digital twins, the fault diagnosis method comprising the following steps:

[0007] Step 1: Establish a digital twin model of the rolling bearing;

[0008] Step 2: Obtain the simulated vibration signal of the digital twin model under normal operating conditions, and obtain the simulated vibration signal of the digital twin model under fault conditions.

[0009] Step 3: Perform wavelet transform on the simulated vibration signals of the digital twin model operating under normal conditions and the simulated vibration signals of the digital twin model operating under fault conditions to generate corresponding two-dimensional time-frequency diagrams;

[0010] Step 4: Improve the ResNet-50 network to obtain the improved ResNet-50 network, and train the improved ResNet-50 network using the two-dimensional time-frequency graph to obtain the first fault diagnosis neural network.

[0011] Step 5: Perform transfer learning training on the first fault diagnosis neural network to obtain the second fault diagnosis neural network;

[0012] Step 6: Use the second fault diagnosis neural network to detect faults in the actual operating rolling bearing.

[0013] In step four, the ResNet-50 network is improved to obtain the improved ResNet-50 network, specifically including:

[0014] In the ResNet-50 network, three feature extraction branches are set up. The parameters of the three feature extraction branches are the same, and each feature extraction branch contains 16 sequentially connected residual blocks.

[0015] The fully connected layers in the ResNet-50 network are replaced with improved layers. The outputs of the three feature extraction branches are each connected to the input of the improved layer through an average pooling layer (AvgPOOL). The inputs of the three feature extraction branches are simultaneously connected to the output of the maximum pooling layer (MAXPOOL) in the ResNet-50 network.

[0016] The improved layer includes a Dropout layer, a Dense layer, and a Softmax layer connected in sequence. The input of the Dropout layer is connected to the output of the three average pooling layers (AvgPOOL). The output of the Dropout layer is connected to the Softmax layer through the Dense layer.

[0017] Step five specifically includes:

[0018] Collect actual operating data of the rolling bearing, including:

[0019] Vibration signals of actual rolling bearings operating under normal conditions, vibration signals of actual rolling bearings operating under outer ring failure conditions, vibration signals of actual rolling bearings operating under inner ring failure conditions, and vibration signals of actual rolling bearings operating under rolling element failure conditions.

[0020] The improved layer in the first fault diagnosis neural network is trained using actual rolling bearing operating data to obtain the second fault diagnosis neural network.

[0021] In step two, obtaining the simulated vibration signal of the digital twin model operating under normal conditions includes: simulating the normal operation of the rolling bearing through the digital twin model and generating the simulated vibration signal of the rolling bearing operating under normal conditions.

[0022] Step two, specifically obtaining the simulated vibration signal of the digital twin model operating under fault conditions, includes:

[0023] A1. Set a fault in the actual rolling bearing and make it run in the fault state, and collect the actual vibration signal of the actual rolling bearing running in the fault state through sensors.

[0024] A2. Set the same faults as those in the actual rolling bearings on the digital twin model so that the digital twin model can run under fault conditions;

[0025] A3. Collect fault simulation vibration signals generated by the digital twin model during operation;

[0026] A4. Calculate the difference between the vibration acceleration of the actual vibration signal and the vibration acceleration of the fault simulation vibration signal to obtain the first difference; calculate the difference between the rate of change of the amplitude of the actual vibration signal and the rate of change of the amplitude of the fault simulation vibration signal to obtain the second difference.

[0027] A5. Determine whether the first difference and the second difference are within the preset range:

[0028] If at least one of the first difference and the second difference is not within the preset range, adjust the parameters of the dynamic model of the digital twin model, and then return to step A3;

[0029] If both the first difference and the second difference are within the preset range, the parameters of the current dynamic model of the digital twin model remain unchanged, and the digital twin model continues to run, while generating a simulated vibration signal of the digital twin model running under fault conditions.

[0030] In step A5, adjusting the dynamic model parameters of the digital twin model specifically includes:

[0031] A5-1. Calculate the first correction difference f according to the following formula. Xi Second correction difference f Yi :

[0032]

[0033] In the formula, i = 1, 2, or r; when the fault set on the actual rolling bearing is an outer ring fault, i = 1; when the fault set on the actual rolling bearing is an inner ring fault, i = 2; when the fault set on the actual rolling bearing is a rolling element ring fault, i = r; when i = 1, m i The mass of the bearing outer ring; when i = 2, m i The mass of the bearing inner ring; when i = r, m i The mass of the rolling elements of the bearing; X i Y represents the abscissa of the fault simulation signal on the vibration time-domain waveform diagram; i X represents the vertical coordinate of the fault simulation signal on the vibration time-domain waveform diagram; X represents the position of the actual vibration signal on the vibration time-domain waveform diagram; k and c are both weighting coefficients, with k ranging from [0-1] and c ranging from [0-1].

[0034] A5-2, using the first correction difference f Xi Second correction difference f Yi Adjusting the parameters in the dynamic model of the digital twin model:

[0035]

[0036]

[0037]

[0038]

[0039]

[0040]

[0041] In the formula, F X1 F represents the force exerted on the outer ring of the rolling element along the bearing axis. Y1 F represents the force exerted on the outer ring of the bearing along the bearing's meridian direction. X2 F represents the force exerted on the inner ring of the bearing along the bearing axis. Y2 F represents the force exerted on the inner ring of the bearing along the bearing's meridian direction. Xr F represents the force exerted on the rolling element along the bearing axis. Yr m1 represents the force exerted on the rolling element along the bearing's warp direction; m2 represents the mass of the bearing's outer ring, m3 represents the mass of the bearing's inner ring, and m4 represents the mass of the rolling element. rLet g be the mass of the rolling elements of the bearing, and g be the acceleration due to gravity.

[0042] Step one specifically includes:

[0043] C1. Measure the structure and dimensions of the rolling bearings actually used;

[0044] C2. Then, use MATLAB software to build a geometric model of the rolling bearing based on the measured structure and dimensions.

[0045] C3. Use simulation toolkits to add corresponding physical properties to various parts of the rolling bearing's geometric model;

[0046] C4. Establish the operating rules for the geometric model of the rolling bearing and build the dynamic model for the operation of the rolling bearing;

[0047] C5. Based on the actual operating environment of the rolling bearing, set its operating state to obtain a dynamic digital twin model of the rolling bearing.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] 1. This invention discloses a rolling bearing fault diagnosis method based on digital twins. The method acquires simulated vibration signals of the digital twin model operating under normal conditions and simulated vibration signals of the digital twin model operating under fault conditions. By using the digital twin model of the rolling bearing, a large number of simulated vibration signals under fault conditions can be obtained. Compared to training the fault diagnosis neural network using actual collected data, this method can obtain a larger training sample, and the data in the sample is more balanced. Therefore, this design can generate a large amount of training sample data through the digital twin model of the rolling bearing, and the data in the training sample is more balanced, which is beneficial to improving the diagnostic accuracy of the trained fault diagnosis neural network.

[0050] 2. The fault diagnosis neural network in the rolling bearing fault diagnosis method based on digital twins of this invention is obtained by training an improved ResNet-50 network twice. The first training uses a large amount of simulation data, and the second training uses real vibration data for transfer learning. Finally, a trained second fault diagnosis neural network is obtained, which has higher fault diagnosis accuracy. Therefore, in this design, the improved ResNet-50 network is trained twice to obtain the second fault diagnosis neural network, resulting in higher fault diagnosis accuracy.

[0051] 3. In the rolling bearing fault diagnosis method based on digital twins of this invention, when using real vibration data to train the first fault diagnosis neural network through transfer learning, the three feature extraction branches before the improved layer are frozen, and training is performed only on the improved layer. Compared to training the entire first diagnostic neural network, the number of samples required for transfer training of the improved layer is very small, which facilitates the collection of actual samples and allows the trained neural network to converge faster. Simultaneously, training in a real data environment enables the diagnostic neural network to fit the real environment, improving the accuracy of fault diagnosis. Therefore, this design uses real vibration data to train the improved layer through transfer learning, which reduces the required actual data samples, allows the trained neural network to converge faster, and improves the accuracy of fault diagnosis.

[0052] 4. In this invention, a rolling bearing fault diagnosis method based on digital twins improves the ResNet-50 network by dividing the original single residual path into three feature extraction branches. Each feature extraction branch uses an average pooling layer (AvgPOOL) to preserve background information. When processing the two-dimensional time-frequency graph, each feature extraction branch obtains a corresponding feature map. Batch standardization of these feature maps accelerates network training and convergence. Simultaneously, an improved layer including Dropout, Dense, and Softmax layers replaces the original ResNet-50 network. The fully connected layers in the network, and the improved layers fuse the feature maps output from the three feature extraction branches, and perform residual learning on the neural network model. Based on the Dropout layer in the improved layers, only a portion of the neurons in the neural network are activated during each training iteration, which significantly reduces overfitting. Simultaneously, the Dropout layer ensures that two neurons do not necessarily appear in the same Dropout network each time, so weight updates no longer rely on the combined effect of hidden nodes with fixed relationships, avoiding the situation where certain features are only effective under other specific features. The trained diagnostic network can effectively reduce overfitting, accelerate convergence, and more accurately derive fault classification results. Therefore, this design improves the ResNet-50 network, enabling the trained diagnostic network to effectively reduce fitting, accelerate convergence, and more accurately derive fault classification results.

[0053] 5. In the rolling bearing fault diagnosis method based on digital twin of the present invention, the normal operation of the rolling bearing can be directly simulated through the digital twin model to generate simulated vibration signals under normal operation. When generating simulated vibration signals under fault conditions, a fault is set on the actual rolling bearing and it is made to operate under fault conditions. At the same time, the actual vibration signals generated by the actual rolling bearing are collected by sensors. The same operating process as the actual rolling bearing is simulated through the digital twin model to generate fault simulation vibration signals. The differences between the actual vibration signals and the fault simulation vibration signals are combined to correct the dynamic model parameters of the digital twin model, so that the simulation data generated by the digital twin model under fault conditions is more in line with reality. Then, a large number of simulated vibration signals of the digital twin model under fault conditions are obtained by adjusting the dynamic parameters of the digital twin model. In this method, the dynamic model parameters of the digital twin model simulating fault conditions can be corrected with a small sample size of actual vibration signals, so that the simulated vibration signals obtained by the digital twin model are more in line with reality. Therefore, in this design, when obtaining the simulated vibration signal of the digital twin model operating under fault conditions, the dynamic model parameters of the digital twin model are modified in conjunction with the actual vibration signal to make the obtained simulated vibration signal more consistent with reality. Attached Figure Description

[0054] Figure 1 It is a flowchart for obtaining the simulated vibration signals of a digital twin model operating under fault conditions.

[0055] Figure 2 It is a time-domain waveform diagram of the vibration signal.

[0056] Figure 3 It is a two-dimensional time-frequency diagram generated by wavelet transform of the vibration signal.

[0057] Figure 4 This is a schematic diagram of the existing ResNet-50 network and the improved ResNet-50 network.

[0058] Figure 5 This is a schematic diagram of the improved layer. Detailed Implementation

[0059] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] See Figures 1 to 5 A method for diagnosing rolling bearing faults based on digital twins, the method comprising the following steps:

[0061] Step 1: Establish a digital twin model of the rolling bearing;

[0062] Step 2: Obtain the simulated vibration signal of the digital twin model under normal operating conditions, and obtain the simulated vibration signal of the digital twin model under fault conditions.

[0063] Step 3: Perform wavelet transform on the simulated vibration signals of the digital twin model operating under normal conditions and the simulated vibration signals of the digital twin model operating under fault conditions to generate corresponding two-dimensional time-frequency diagrams;

[0064] Step 4: Improve the ResNet-50 network to obtain the improved ResNet-50 network, and train the improved ResNet-50 network using the two-dimensional time-frequency graph to obtain the first fault diagnosis neural network.

[0065] Step 5: Perform transfer learning training on the first fault diagnosis neural network to obtain the second fault diagnosis neural network;

[0066] Step 6: Use the second fault diagnosis neural network to detect faults in the actual operating rolling bearing.

[0067] In step four, the ResNet-50 network is improved to obtain the improved ResNet-50 network, specifically including:

[0068] In the ResNet-50 network, three feature extraction branches are set up. The parameters of the three feature extraction branches are the same. Each feature extraction branch contains 16 sequentially connected residual blocks, which are called Bottleneck blocks.

[0069] The fully connected layers in the ResNet-50 network are replaced with improved layers. The outputs of the three feature extraction branches are each connected to the input of the improved layer through an average pooling layer (AvgPOOL). The inputs of the three feature extraction branches are simultaneously connected to the output of the maximum pooling layer (MAXPOOL) in the ResNet-50 network.

[0070] like Figure 4 As shown, the existing ResNet-50 network contains only one feature extraction path; the improved ResNet-50 network contains three feature extraction branches, and the improved ResNet-50 network uses an improved layer to replace the fully connected layer in the original network.

[0071] The improved layer comprises a Dropout layer, a Dense layer, and a Softmax layer connected in sequence. The input of the Dropout layer is simultaneously connected to the outputs of the three average pooling layers (AvgPOOL). The output of the Dropout layer is connected to the Softmax layer through the Dense layer. The Dropout layer is used to reduce overfitting during training. The Dense layer performs one-dimensional processing on the information output by the Dropout layer, extracts effective feature information, and then maps the effective feature information back to the output of the Softmax layer after nonlinear transformation. The Softmax layer is used to calculate and output the predicted probability of bearing fault categories.

[0072] The bearing failure categories include: bearing outer ring failure, bearing inner ring failure, and rolling element failure.

[0073] When training the improved ResNet-50 network, the Dropout layer controls the number of neurons activated in each training session to reduce overfitting. Setting the parameter of the Dropout layer to 0.5 means that the total number of neurons activated in each training session is half of the total number of neurons. In the ResNet-50 network, the MAXPOOL layer is used to preserve texture information and remove noise information, while the AVGPOOL layer is used to preserve the main relevant features.

[0074] Step five specifically includes:

[0075] Collect actual rolling bearing operating data, including: vibration signals of the actual rolling bearing operating under normal conditions, vibration signals of the actual rolling bearing operating under outer ring failure conditions, vibration signals of the actual rolling bearing operating under inner ring failure conditions, and vibration signals of the actual rolling bearing operating under rolling element failure conditions.

[0076] Wavelet transform was performed on the actual rolling bearing operating data, and the resulting two-dimensional time-frequency graph was used to train the improved layer in the first fault diagnosis neural network, thus obtaining the second fault diagnosis neural network. When training the improved ResNet-50 network, the two-dimensional time-frequency graph was used as input, and the bearing health status corresponding to the two-dimensional time-frequency graph was used as output. The bearing health status was categorized as: normal, outer race fault, inner race fault, or rolling element fault.

[0077] Since training does not change the connection structure of the neural network, both the first fault diagnosis neural network and the second fault diagnosis neural network have the same structure as the improved ResNet-50 network. Both the first fault diagnosis neural network and the second fault diagnosis neural network also contain three feature extraction branches and an improved layer.

[0078] The method for training the improved layer in the first fault diagnosis neural network is as follows: freeze all parameters in the three feature extraction branches of the first fault diagnosis neural network to obtain a partially frozen first fault diagnosis neural network, and train the partially frozen first fault diagnosis neural network, that is, train only the parameters of the improved layer.

[0079] In step two, obtaining the simulated vibration signal of the digital twin model operating under normal conditions includes: simulating the normal operation of the rolling bearing through the digital twin model and generating the simulated vibration signal under normal operation.

[0080] The simulated vibration signals of the digital twin model under normal operating conditions and the simulated vibration signals of the digital twin model under fault conditions are both one-dimensional vibration signals. The vibration waveforms of the one-dimensional vibration signals are shown in the figure below. Figure 2 As shown, vibration signals are sampled at a certain frequency to generate vibration waveforms. The horizontal axis (X-axis) of the vibration waveform graph represents the sampling time / the total number of sampling points, and the vertical axis (Y-axis) of the vibration waveform graph represents the vibration amplitude.

[0081] Step two, specifically obtaining the simulated vibration signal of the digital twin model operating under fault conditions, includes:

[0082] A1. Set a fault in the actual rolling bearing and make it run in the fault state, and collect the actual vibration signal of the actual rolling bearing in the fault state through an acoustic sensor.

[0083] A2. Set the same faults as those in the actual rolling bearings on the digital twin model, so that the digital twin model operates under fault conditions; setting faults on the digital twin model can be achieved by adjusting the dynamic model of the digital twin model;

[0084] A3. Collect the fault simulation vibration signal generated by the digital twin model;

[0085] The actual rolling bearing or digital twin model completes one rotation per motion cycle. When the actual rolling bearing or digital twin model runs for multiple motion cycles, a fluctuating vibration signal curve is generated on the vibration time-domain waveform diagram. The actual vibration signal and the fault simulation vibration signal are the vibration signals generated by the actual rolling bearing and the digital twin model starting from the same initial position and running for the same number of motion cycles, respectively.

[0086] A4. Calculate the difference between the vibration acceleration of the actual vibration signal and the vibration acceleration of the fault simulation vibration signal to obtain the first difference; calculate the difference between the rate of change of the amplitude of the actual vibration signal and the rate of change of the amplitude of the fault simulation vibration signal to obtain the second difference.

[0087] A5. Determine whether the first difference and the second difference are within the preset range:

[0088] If at least one of the first difference and the second difference is not within the preset range, it indicates that there is a significant difference between the actual vibration signal and the fault simulation vibration signal. Then, adjust the parameters of the dynamic model of the digital twin model and return to step A3.

[0089] If both the first and second differences are within the preset range, it indicates that the fault simulation vibration signal matches the actual vibration signal. In this case, the parameters of the current dynamic model of the digital twin model remain unchanged, and the digital twin model continues to run. The simulated vibration signal obtained at this time is the simulated vibration signal of the digital twin model running under fault conditions.

[0090] In step A1, the fault set on the rolling bearing in actual use can be a fault in the outer ring of the bearing, a fault in the inner ring of the bearing, or a fault in the rolling element.

[0091] When a bearing outer ring fault is set on a rolling bearing in actual use, step A5 acquires the simulated vibration signal of the digital twin model operating under the bearing outer ring fault condition; when a bearing inner ring fault is set on a rolling bearing in actual use, step A5 acquires the simulated vibration signal of the digital twin model operating under the bearing inner ring fault condition; when a rolling element fault is set on a rolling bearing in actual use, step A5 acquires the simulated vibration signal of the digital twin model operating under the rolling element fault condition. Through steps A1 to A5, simulated vibration signals of the digital twin model operating under various different fault conditions can be generated.

[0092] Wavelet transforms are performed on the simulated vibration signals of the digital twin model under normal operation, the simulated vibration signals of the digital twin model under the bearing outer ring fault state, the simulated vibration signals of the digital twin model under the bearing inner ring fault state, and the simulated vibration signals of the digital twin model under the rolling element fault state, respectively, to obtain two-dimensional time-frequency diagrams corresponding to four different bearing health states. The improved ResNet-50 network is trained using the above two-dimensional time-frequency diagrams corresponding to the four different bearing health states to obtain the first fault diagnosis neural network.

[0093] When training the improved ResNet-50 network, the two-dimensional time-frequency graph is used as the input of the improved ResNet-50 network, and the bearing health state corresponding to the two-dimensional time-frequency graph is used as the output of the improved ResNet-50 network.

[0094] In step A5, adjusting the dynamic model parameters of the digital twin model specifically includes:

[0095] A5-1. Calculate the first correction difference f according to the following formula. Xi Second correction difference f Yi :

[0096]

[0097] In the formula, i = 1, 2, or r; when the fault set on the actual rolling bearing is an outer ring fault, i = 1; when the fault set on the actual rolling bearing is an inner ring fault, i = 2; when the fault set on the actual rolling bearing is a rolling element ring fault, i = r; when i = 1, m i The mass of the bearing outer ring; when i = 2, m i The mass of the bearing inner ring; when i = r, m i The mass of the rolling elements of the bearing; X i Y represents the abscissa of the simulated vibration signal on the vibration time-domain waveform diagram; i X represents the vertical coordinate of the fault simulation signal on the vibration time-domain waveform diagram; X represents the position of the actual vibration signal on the vibration time-domain waveform diagram; k and c are both weighting coefficients, with k ranging from [0-1] and c ranging from [0-1].

[0098] A5-2, using the first correction difference f Xi Second correction difference f Yi The parameters in the dynamic model of the digital twin model are adjusted to correct the discrepancy between the analog and actual signals:

[0099]

[0100]

[0101]

[0102]

[0103]

[0104]

[0105] In the formula, F X1 F represents the force exerted on the outer ring of the rolling element along the bearing axis. Y1 F represents the force exerted on the outer ring of the bearing along the bearing's meridian direction. X2 F represents the force exerted on the inner ring of the bearing along the bearing axis. Y2 F represents the force exerted on the inner ring of the bearing along the bearing's meridian direction. XrF represents the force exerted on the rolling element along the bearing axis. Yr m1 represents the force exerted on the rolling element along the bearing's warp direction; m2 represents the mass of the bearing's outer ring, m3 represents the mass of the bearing's inner ring, and m4 represents the mass of the rolling element. r Let g be the mass of the rolling elements of the bearing, and g be the acceleration due to gravity.

[0106] Formulas (2)-(7) above are some of the formulas in the dynamic model of the digital twin model, F X1 F Y1 F X2 F Y2 F Xr F Yr All of these are parameters that have been adjusted.

[0107] Step one specifically includes:

[0108] C1. Measure the structure and dimensions of the rolling bearings actually used;

[0109] C2. Then, use MATLAB software to build a geometric model of the rolling bearing based on the measured structure and dimensions.

[0110] The structure and geometry of the rolling bearing include: the number of balls, the ball diameter, the bearing raceway pitch diameter, the bearing contact angle, the inner ring raceway radius, and the outer ring raceway radius.

[0111] Including the physical characteristics of the metal materials used in each part; measuring the installation position and method of each part of the rolling bearing;

[0112] C3. Use simulation toolkits to add corresponding physical properties to various parts of the rolling bearing's geometric model;

[0113] C4. Establish the operating rules of the geometric model of the rolling bearing, set the inner ring rotation frequency, the relative rotation frequency between the inner and outer rings, the frequency of the rolling element passing through a point on the inner ring, the frequency of the rolling element passing through a point on the outer ring, the revolution frequency of the rolling element, and the rotation frequency of the cage, and establish the dynamic model of the bearing operation.

[0114] C5. Based on the actual operating environment of the rolling bearing, set its operating state to obtain a dynamic digital twin model of the rolling bearing.

[0115] The principle of this invention is explained as follows:

[0116] Digital twin technology, as an existing modeling method, can effectively reflect the physical properties of an entity. By establishing a digital twin model of a rolling bearing that closely matches the physical model, running the twin model yields simulation data that closely reflects reality. Combining actual vibration data with the simulation data generated by the twin model and inputting it into an improved neural network for training, and then using the trained neural network to diagnose faults, can effectively achieve the diagnosis and prediction of rolling bearing fault states.

[0117] Example 1:

[0118] A rolling bearing fault diagnosis method based on digital twins, the fault diagnosis method comprising the following steps:

[0119] Step 1: Establish a digital twin model of the rolling bearing;

[0120] Step 2: Obtain the simulated vibration signal of the digital twin model under normal operating conditions, and obtain the simulated vibration signal of the digital twin model under fault conditions.

[0121] Step 3: Perform wavelet transform on the simulated vibration signals of the digital twin model operating under normal conditions and the simulated vibration signals of the digital twin model operating under fault conditions to generate corresponding two-dimensional time-frequency diagrams;

[0122] Step 4: Improve the ResNet-50 network to obtain the improved ResNet-50 network, and train the improved ResNet-50 network using the two-dimensional time-frequency graph to obtain the first fault diagnosis neural network.

[0123] Step 5: Perform transfer learning training on the first fault diagnosis neural network to obtain the second fault diagnosis neural network;

[0124] Step 6: Use the second fault diagnosis neural network to detect faults in the actual operating rolling bearing.

[0125] Example 2:

[0126] Example 2 is basically the same as Example 1, except that:

[0127] In step four, the ResNet-50 network is improved to obtain the improved ResNet-50 network, specifically including:

[0128] In the ResNet-50 network, three feature extraction branches are set up. The parameters of the three feature extraction branches are the same, and each feature extraction branch contains 16 sequentially connected residual blocks.

[0129] The fully connected layer in the ResNet-50 network is improved into an improved layer. The outputs of the three feature extraction branches are each connected to the input of the improved layer through an average pooling layer (AvgPOOL). The inputs of the three feature extraction branches are simultaneously connected to the output of the maximum pooling layer (MAXPOOL) in the ResNet-50 network.

[0130] The improved layer includes a Dropout layer, a Dense layer, and a Softmax layer connected in sequence. The input of the Dropout layer is connected to the output of the three average pooling layers (AvgPOOL). The output of the Dropout layer is connected to the Softmax layer through the Dense layer. The Dropout layer is used to reduce overfitting during training. The Dense layer performs one-dimensional processing on the information, so that the extracted features are back-mapped to the output of the Softmax layer after nonlinear transformation. The Softmax layer is used to calculate and output the predicted probability of bearing fault categories.

[0131] Step five specifically includes:

[0132] Collect actual operating data of the rolling bearing, including:

[0133] Vibration signals of actual rolling bearings operating under normal conditions, vibration signals of actual rolling bearings operating under outer ring failure conditions, vibration signals of actual rolling bearings operating under inner ring failure conditions, and vibration signals of actual rolling bearings operating under rolling element failure conditions.

[0134] The improved layer in the first fault diagnosis neural network is trained using actual rolling bearing operating data to obtain the second fault diagnosis neural network.

[0135] Example 3:

[0136] Example 3 is basically the same as Example 2, except that:

[0137] In step two, obtaining the simulated vibration signal of the digital twin model operating under normal conditions includes: simulating the normal operation of the rolling bearing through the digital twin model and generating the simulated vibration signal of the rolling bearing operating under normal conditions.

[0138] Step two, specifically obtaining the simulated vibration signal of the digital twin model operating under fault conditions, includes:

[0139] A1. Set a fault in the actual rolling bearing and make it run in the fault state, and collect the actual vibration signal of the actual rolling bearing running in the fault state through sensors.

[0140] A2. Set the same faults as those in the actual rolling bearings on the digital twin model so that the digital twin model can run under fault conditions;

[0141] A3. Collect fault simulation vibration signals generated by the digital twin model during operation;

[0142] A4. Calculate the difference between the vibration acceleration of the actual vibration signal and the vibration acceleration of the fault simulation vibration signal to obtain the first difference; calculate the difference between the rate of change of the amplitude of the actual vibration signal and the rate of change of the amplitude of the fault simulation vibration signal to obtain the second difference.

[0143] A5. Determine whether the first difference and the second difference are within the preset range:

[0144] If at least one of the first difference and the second difference is not within the preset range, adjust the parameters of the dynamic model of the digital twin model, and then return to step A3;

[0145] If both the first difference and the second difference are within the preset range, the parameters of the current dynamic model of the digital twin model remain unchanged, and the digital twin model continues to run, while generating a simulated vibration signal of the digital twin model running under fault conditions.

[0146] In step A5, adjusting the dynamic model parameters of the digital twin model specifically includes:

[0147] A5-1. Calculate the first correction difference f according to the following formula. Xi Second correction difference f Yi :

[0148]

[0149] In the formula, i = 1, 2, or r; when the fault set on the actual rolling bearing is an outer ring fault, i = 1; when the fault set on the actual rolling bearing is an inner ring fault, i = 2; when the fault set on the actual rolling bearing is a rolling element ring fault, i = r; when i = 1, m i The mass of the bearing outer ring; when i = 2, m i The mass of the bearing inner ring; when i = r, m i The mass of the rolling elements of the bearing; X i Y represents the abscissa of the fault simulation signal on the vibration time-domain waveform diagram; i X represents the vertical coordinate of the fault simulation signal on the vibration time-domain waveform diagram; X represents the position of the actual vibration signal on the vibration time-domain waveform diagram; k and c are both weighting coefficients, with k ranging from [0-1] and c ranging from [0-1].

[0150] A5-2, using the first correction difference fXi Second correction difference f Yi Adjustments were made to the dynamic model of the digital twin model:

[0151]

[0152]

[0153]

[0154]

[0155]

[0156]

[0157] In the formula, F X1 F represents the force exerted on the outer ring of the rolling element along the bearing axis. Y1 F represents the force exerted on the outer ring of the bearing along the bearing's meridian direction. X2 F represents the force exerted on the inner ring of the bearing along the bearing axis. Y2 F represents the force exerted on the inner ring of the bearing along the bearing's meridian direction. Xr F represents the force exerted on the rolling element along the bearing axis. Yr m1 represents the force exerted on the rolling element along the bearing's warp direction; m2 represents the mass of the bearing's outer ring, m3 represents the mass of the bearing's inner ring, and m4 represents the mass of the rolling element. r Let g be the mass of the rolling elements of the bearing, and g be the acceleration due to gravity.

[0158] The above description is only a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. Any equivalent modifications or changes made by those skilled in the art based on the content disclosed in the present invention should be included within the scope of protection set forth in the claims.

Claims

1. A method for diagnosing rolling bearing faults based on digital twins, characterized in that: The fault diagnosis method includes the following steps: Step 1: Establish a digital twin model of the rolling bearing; Step 2: Obtain the simulated vibration signal of the digital twin model under normal operating conditions, and obtain the simulated vibration signal of the digital twin model under fault conditions. Step 3: Perform wavelet transform on the simulated vibration signals of the digital twin model operating under normal conditions and the simulated vibration signals of the digital twin model operating under fault conditions to generate corresponding two-dimensional time-frequency diagrams; Step 4: Improve the ResNet-50 network to obtain the improved ResNet-50 network, and use the two-dimensional time-frequency graph to train the improved ResNet-50 network to obtain the first fault diagnosis neural network. Step 5: Perform transfer learning training on the first fault diagnosis neural network to obtain the second fault diagnosis neural network; Step Six: Detect faults in the actual operating rolling bearing using the second fault diagnosis neural network; In step four, the ResNet-50 network is improved to obtain the improved ResNet-50 network, specifically including: In the ResNet-50 network, three feature extraction branches are set up. The parameters of the three feature extraction branches are the same, and each feature extraction branch contains 16 sequentially connected residual blocks. The fully connected layers in the ResNet-50 network are replaced with improved layers. The outputs of the three feature extraction branches are each connected to the input of the improved layer through an average pooling layer (AvgPOOL). The inputs of the three feature extraction branches are simultaneously connected to the output of the maximum pooling layer (MAXPOOL) in the ResNet-50 network. The improved layer includes a Dropout layer, a Dense layer, and a Softmax layer connected in sequence. The input of the Dropout layer is connected to the output of the three average pooling layers (AvgPOOL). The output of the Dropout layer is connected to the Softmax layer through the Dense layer. Step two, specifically obtaining the simulated vibration signal of the digital twin model operating under fault conditions, includes: A1. Set a fault in the actual rolling bearing and make it run in the fault state, and collect the actual vibration signal of the actual rolling bearing running in the fault state through sensors. A2. Set the same faults as those in the actual rolling bearings on the digital twin model so that the digital twin model can run under fault conditions; A3. Collect fault simulation vibration signals generated by the digital twin model during operation; A4. Calculate the difference between the vibration acceleration of the actual vibration signal and the vibration acceleration of the fault simulation vibration signal to obtain the first difference; calculate the difference between the rate of change of the amplitude of the actual vibration signal and the rate of change of the amplitude of the fault simulation vibration signal to obtain the second difference. A5. Determine whether the first difference and the second difference are within the preset range: If at least one of the first difference and the second difference is not within the preset range, adjust the parameters of the dynamic model of the digital twin model, and then return to step A3; If both the first difference and the second difference are within the preset range, the parameters of the current dynamic model of the digital twin model remain unchanged, and the digital twin model continues to run to generate a simulated vibration signal of the digital twin model running under fault conditions. In step A5, adjusting the dynamic model parameters of the digital twin model specifically includes: A5-1. Calculate the first correction difference f according to the following formula. Xi Second correction difference f Yi : （1）； In the formula, i = 1, 2, or r; i = 1 when the fault set on the actual rolling bearing is an outer ring fault; i = 2 when the fault set on the actual rolling bearing is an inner ring fault; i = r when the fault set on the actual rolling bearing is a rolling element ring fault; when i = 1, m i The mass of the bearing outer ring; when i=2, m i For the mass of the bearing inner ring; when i=r, m i The mass of the rolling elements of the bearing; X i Y represents the abscissa of the fault simulation signal on the vibration time-domain waveform diagram; i X represents the vertical coordinate of the fault simulation signal on the vibration time-domain waveform diagram; X represents the position of the actual vibration signal on the vibration time-domain waveform diagram; k and c are both weighting coefficients, with k ranging from [0-1] and c ranging from [0-1]. A5-2, using the first correction difference f Xi Second correction difference f Yi Adjusting the parameters in the dynamic model of the digital twin model: （2）； （3）； （4）； （5）； （6）； （7）； In the formula, F X1 F represents the force exerted on the outer ring of the rolling element along the bearing axis. Y1 F represents the force exerted on the outer ring of the bearing along the bearing's meridian direction. X2 F represents the force exerted on the inner ring of the bearing along the bearing axis. Y2 F represents the force exerted on the inner ring of the bearing along the bearing's meridian direction. Xr F represents the force exerted on the rolling element along the bearing axis. Yr m1 represents the force exerted on the rolling element along the bearing's warp direction; m2 represents the mass of the bearing's outer ring, m3 represents the mass of the bearing's inner ring, and m4 represents the mass of the rolling element. r Let g be the mass of the rolling elements of the bearing, and g be the acceleration due to gravity.

2. The rolling bearing fault diagnosis method based on digital twin according to claim 1, characterized in that: Step five specifically includes: Collect actual operating data of the rolling bearing, including: Vibration signals of actual rolling bearings operating under normal conditions, vibration signals of actual rolling bearings operating under outer ring failure conditions, vibration signals of actual rolling bearings operating under inner ring failure conditions, and vibration signals of actual rolling bearings operating under rolling element failure conditions. The improved layer in the first fault diagnosis neural network is trained using actual rolling bearing operating data to obtain the second fault diagnosis neural network.

3. The rolling bearing fault diagnosis method based on digital twin according to claim 1, characterized in that: In step two, obtaining the simulated vibration signal of the digital twin model operating under normal conditions includes: simulating the normal operation of the rolling bearing through the digital twin model and generating the simulated vibration signal of the rolling bearing operating under normal conditions.

4. The rolling bearing fault diagnosis method based on digital twin according to claim 1, characterized in that: Step one specifically includes: C1. Measure the structure and dimensions of the rolling bearings actually used; C2. Then, use MATLAB software to build a geometric model of the rolling bearing based on the measured structure and dimensions. C3. Use simulation toolkits to add corresponding physical properties to various parts of the rolling bearing's geometric model; C4. Establish the operating rules for the geometric model of the rolling bearing and build the dynamic model for the operation of the rolling bearing; C5. Based on the actual operating environment of the rolling bearing, set its operating state to obtain a dynamic digital twin model of the rolling bearing.

Images