Construction method of target domain data set, rolling bearing fault diagnosis method and computer program product

Abstract

The invention relates to the technical field of rolling bearing fault diagnosis, and discloses a target domain data set construction method, a rolling bearing fault diagnosis method and a computer program product, and the target domain data set construction method comprises the steps: selecting an amplitude spectrum of an actual measurement sample as a fitness function of a CS-GWO algorithm; taking the minimum value of the amplitude spectrum entropy as a fitness function, and searching an optimal parameter combination in the VMD algorithm; decomposing the actually measured samples in the original data set by using the VMD algorithm of the optimal parameter combination to obtain K modal components; and calculating the correlation between each modal component and the actually measured sample, retaining the high-correlation modal components, and carrying out summation reconstruction to obtain target domain data in the target domain data set. According to the method, the fault features are accurately extracted by using the CS-GWO optimized VMD algorithm, the fault feature coefficient and the signal-to-noise ratio are remarkably improved, noise interference in an actual measurement sample is effectively inhibited, and thus the accuracy and reliability of target domain data are improved.

Description

Technical Field

[0001] This invention relates to the field of rolling bearing fault diagnosis technology, and in particular to a method for constructing a target domain dataset, a method for diagnosing rolling bearing faults, and a computer program product. Background Technology

[0002] As a core component of rotating machinery, the operating condition of rolling bearings directly affects the reliability and safety of the equipment. Therefore, rolling bearing fault diagnosis is a key technology for ensuring the stable operation of mechanical systems. Existing rolling bearing fault diagnosis methods largely rely on training diagnostic models with large amounts of labeled laboratory test data. However, this approach has several drawbacks in practical applications: firstly, acquiring laboratory test fault data is costly and time-consuming, especially since it's difficult to comprehensively cover fault data for different operating conditions and bearing models; secondly, there are domain distribution differences between simulation mechanism data and laboratory test data, leading to a significant decrease in diagnostic accuracy when models trained on simulation data are directly applied to test data.

[0003] While existing transfer learning methods attempt to address the issue of domain distribution differences, they still have shortcomings: some methods only use a single loss function to measure domain differences, making it difficult to fully align the feature distributions of simulation and measured data; for noisy measured data, there is a lack of effective feature extraction and noise reduction preprocessing mechanisms, which causes fault features to be masked by noise, further reducing diagnostic accuracy; at the same time, in complex scenarios such as bearing models and small samples, the generalization ability and diagnostic robustness of existing methods are insufficient, failing to meet the needs of actual engineering.

[0004] Therefore, there is an urgent need in this field for a method that can integrate mechanism simulation data and laboratory measured data, and achieve high-precision fault diagnosis of rolling bearings in complex scenarios through accurate feature extraction and efficient domain adaptive transfer learning. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing a target domain dataset, a method for diagnosing rolling bearing faults, and a computer program product, so as to solve or at least partially solve the technical problems mentioned in the background art.

[0006] To achieve this objective, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for constructing a target domain dataset, comprising: Vibration signals measured under four operating conditions of rolling bearings—normal condition, inner ring fault, outer ring fault, and rolling element fault—were collected as measured samples to construct the original dataset. The amplitude spectrum of the measured samples is selected as the fitness function of the CS-GWO algorithm, and the minimum value of the amplitude spectrum entropy is used as the fitness function to search for the optimal parameter combination in the VMD algorithm. Where K is the modality number K, It is a secondary penalty factor; The VMD algorithm with optimal parameter combination is used to decompose the measured samples in the original dataset to obtain K modal components; Calculate the correlation between each modal component and the measured sample, retain the highly correlated modal components and sum them to reconstruct the target domain data in the target domain dataset.

[0007] Optionally, the VMD algorithm using the optimal parameter combination decomposes the measured samples in the original dataset to obtain K modal components, specifically including: The measured vibration signal Decomposed into K modal components; set up This represents the K modal components obtained after decomposition. The center frequency of each component; For each modal component, the analytic signal is calculated using the Hilbert transform, thus obtaining the one-sided spectrum: j is the imaginary unit. It is a unit impulse function. This is a convolution operation; The spectrum of the modal function is modulated by aliasing the exponential term corresponding to the center frequency of each modal function: ; Gaussian smoothing is used to smooth the spectrum of each mode function, the bandwidth of the mode function is estimated, and a constrained variational problem is solved to minimize the sum of the estimated bandwidths of each mode function; the constrained variational expression is: ; This represents the derivative of the function with respect to time t.

[0008] Optionally, solving the constrained variational problem specifically involves: Using a secondary penalty factor and Lagrange multiplication operators Solving the constrained variational expression transforms the constrained variational problem into an unconstrained variational problem, i.e.: ; Solving unconstrained variational problems using the alternating direction multiplier method, through alternating updates... , as well as Find the saddle point of the augmented Lagrange expression, and solve for it: ;as well as, The method for updating the center frequency is as follows: .

[0009] Optionally, the step of calculating the correlation between each modal component and the measured sample, retaining the highly correlated modal components, summing and reconstructing them to obtain the target domain data in the target domain dataset includes: Correlation analysis was performed on each modal component and the original vibration signal to assess their correlation degree; the correlation calculation formula is as follows: ; In the formula, The signal of the modal component at sampling point i, This is the signal mean of the modal signal. The original vibration signal is the signal at sampling point i. The mean value of the original vibration signal is denoted as N, and the number of sampling points is N. Modal components with correlation greater than or equal to a predetermined correlation threshold are retained and summed to reconstruct the original vibration signal.

[0010] Optionally, the amplitude spectrum of the selected measured sample is used as the fitness function of the CS-GWO algorithm, and the minimum value of the amplitude spectrum entropy is used as the fitness value to search for the optimal parameter combination in the VMD algorithm. Specifically, it includes: Initialize the parameters of the CS-GWO algorithm and VMD; Using the minimum amplitude spectral entropy as the fitness function, the optimal parameter combination in the VMD algorithm is searched using the CS-GWO algorithm. The formula for calculating the amplitude spectral entropy is: ; In the above formula, L i For modal components u i The amplitude spectrum, H i For modal components u i The amplitude spectral entropy, where N is the length of the modal component.

[0011] Secondly, the present invention provides a method for diagnosing rolling bearing faults, characterized in that it includes: Based on the nonlinear dynamic equations of rolling bearings, simulation data are generated under four types of working conditions: normal state, inner ring fault, outer ring fault, and rolling element fault, and a source domain dataset with complete fault labels is constructed. Furthermore, the target domain dataset is constructed using a method for constructing a target domain dataset as described in any one of claims 1-5; A portion of the target domain dataset and the source domain dataset are randomly divided into a training set, and the remaining portion of the target domain dataset is divided into a test set to train a pre-built improved alternating transfer learning model. A trained improved alternating transfer learning model is used for full-life rolling bearing fault diagnosis, and the diagnosis results are output.

[0012] Optionally, the method for constructing the improved alternating transfer learning model is as follows: Construct a CNN model containing five convolutional layers, two pooling layers, and two fully connected layers. Add a batch normalization layer after each convolutional layer. Use ReLU as the activation function for the hidden layers and Softmax as the activation function for the output layer. The first pooling layer is located between the second and third convolutional layers, and the second pooling layer is located after the fifth convolutional layer and before the first fully connected layer. The CORAL loss function is calculated after the first convolutional layer of the CNN model to reduce the difference in second-order statistics between the source and target domain datasets. The sum of the MMD loss function and the classification loss is calculated in the fully connected layer, and the network weights and bias parameters are updated by alternating backpropagation.

[0013] Optionally, the CORAL loss function can be set as follows: The distance between the second-order statistics of the data features in the source domain dataset and the target domain dataset is defined as the CORAL loss function L1: , is the Frobenius norm of the mean square matrix; in, Let D be the number of dimensions of this sample, and D be the source domain dataset. S , and the target domain dataset D T The characteristic covariance matrices are C S With C T ; MMD is used to measure the feature set D. S and D T The distributional differences in the reproducing kernel Hilbert space are used as the sum of the MMD loss and the classification loss as the L2 loss function: , , , For the regenerating nucleus Hilbert space; In the above formula, L C The classification loss is the difference between the true label and the predicted label. , Mapping function , n is the balance coefficient. S Let y be the number of samples, and y be the true label of the sample data. The label predicted by the classifier.

[0014] Optionally, the nonlinear dynamic equations based on rolling bearings generate simulation data for four types of operating conditions: normal state, inner ring fault, outer ring fault, and rolling element fault, constructing a source domain dataset with complete fault labels, including: Based on Hertz contact theory, a nonlinear dynamic model of the rolling bearing is established by modifying the radial displacement excitation function, equivalent contact stiffness, and impact force of the rolling element. The radial displacement excitation function is modified according to the geometric position of the rolling element in the defect region, and the impact force is modified according to the size of the fault defect and the bearing speed. The nonlinear dynamic model of the rolling bearing is solved by the fourth-order Runge-Kutta method, and the simulated vibration signals under the four working conditions are generated to form a labeled source domain dataset.

[0015] Thirdly, the present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements a method for constructing a target domain dataset as described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This application utilizes the CS-GWO optimized VMD algorithm to accurately extract fault features, significantly improving the fault feature coefficient and signal-to-noise ratio, effectively suppressing noise interference in the measured samples, thereby improving the accuracy and reliability of the target domain data. This, in turn, helps the model learn the real patterns in the data, improving its prediction accuracy and generalization ability, and enabling high-precision fault diagnosis of rolling bearings in complex scenarios. It is especially suitable for application scenarios with variable working conditions and has broad engineering application value. Attached Figure Description

[0017] 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.

[0018] Figure 1 This is a flowchart of a method for diagnosing rolling bearing faults, provided in an embodiment of the present invention.

[0019] Figure 2 This invention provides a method for constructing a target domain dataset.

[0020] Figure 3 This is a schematic diagram of a nonlinear dynamic model of a rolling bearing provided in an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram illustrating the dimensional relationship of a rolling element entering a defect region, provided as an embodiment of the present invention.

[0022] Figure 5 This is a schematic diagram of a rolling body overcoming an obstacle, provided in an embodiment of the present invention.

[0023] Figure 6 This is a schematic diagram of the impact force at a contact defect of a rolling element, provided as an embodiment of the present invention.

[0024] Figure 7 This is a schematic diagram of a rolling bearing fault diagnosis principle based on improved alternating transfer learning, provided for an embodiment of the present invention.

[0025] Figure 8a This invention provides a source domain time-domain waveform and grayscale image.

[0026] Figure 8b This invention provides a time-domain waveform and grayscale image of a target domain.

[0027] Figure 9 This is a migration task result confusion matrix provided in an embodiment of the present invention.

[0028] Figure 10 This is a t-SNE visualization result diagram of a migration task provided in an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0030] Example 1: Please refer to Figure 1 , Figure 1 A flowchart of a method for diagnosing rolling bearing faults provided in an embodiment of the present invention; the method includes: Step 110: Based on the nonlinear dynamic equations of rolling bearings, generate simulation data for four types of working conditions: normal state, inner ring fault, outer ring fault, and rolling element fault, and construct a source domain dataset with complete fault labels.

[0031] Specifically, for rolling bearings used in full-life testing (such as Rexnord ZA-2115 bearings), the rolling bearing system is simplified into a mass-spring-damped model consisting of an outer ring, inner ring, rolling elements, and a high-frequency resonator, and the model is assumed to satisfy the following conditions: The outer ring is fixed to a rigid structure, while the inner ring maintains a constant rotational speed. The rolling elements are evenly distributed with no relative slippage; The effect of the rolling element's rotation is not considered; The inner and outer rings and the rolling elements move in the same plane.

[0032] like Figure 3 As shown, Figure 3 A schematic diagram of a nonlinear dynamic model of a rolling bearing provided in an embodiment of the present invention; in the figure, For the quality of the inner ring, This is the damping coefficient between the inner ring and the bearing housing. This refers to the support stiffness between the inner ring and the bearing housing. For the quality of the outer ring, , These represent the support stiffness and damping coefficient between the outer ring and the bearing housing, respectively. For the mass of a single rolling element, This indicates the angular position of the j-th rolling element.

[0033] When a rolling element enters a defect area, it rotates around the leading edge of the defect, causing a sudden change in its radial displacement. This application calculates the relative position of the rolling element in the defect area using geometric relationships, thus accurately reconstructing the obstacle-crossing process of the rolling element.

[0034] according to Figure 4 The dimensional relationships shown indicate the angle at which the rolling element enters the defect region. The expressions for sine and cosine are: , ; cage speed Rotational angular velocity of the inner ring The relationship between them is: ; In the above formula, d is the rolling element diameter; D is the bearing pitch diameter; β is the contact angle; ; In the above formula, Z represents the number of rolling elements, j = 1, 2, 3, ...; The angular position of the first rolling element relative to the positive y-axis at time zero.

[0035] According to Hertzian contact theory, the contact between rolling elements and raceways leads to elastic deformation, thus creating Hertzian contact force. This contact force is one of the important forces during bearing operation. When a bearing has defects, the contact deformation is divided into two cases: contact deformation in the non-defective area and contact deformation in the defective area. These deformations have a significant impact on the normal operation and fault diagnosis of the bearing.

[0036] like Figure 5 As shown, the inner and outer rings of the bearing are given two degrees of freedom in the horizontal and vertical directions, respectively. When the inner and outer rings of the bearing undergo relative displacement, , Let represent the relative displacement vectors of the j-th rolling element with respect to the inner and outer raceways of the bearing, respectively. Their specific coordinate relationships are as follows: , , In the formula, Indicates the extreme diameter of the j-th rolling element; , These represent the displacements of the shaft and the inner ring in the x (horizontal displacement) and y (vertical displacement) directions, respectively. , These represent the horizontal and vertical displacements of the shaft and the outer ring, respectively.

[0037] Under fault-free conditions, the contact deformations of the j-th rolling element with the inner and outer raceways of the bearing are as follows: ;in, This refers to the radial clearance of the bearing; ; When a local defect occurs in the outer raceway of a bearing, this application simplifies the defect area into a rectangle with a sufficiently large depth h, satisfying 4dh ≥ 4h. 2 +L 2 Where d is the diameter of the rolling element; L is the width of the defect. The situation where the rolling element touches the bottom of the defect during obstacle crossing is not discussed.

[0038] When the rolling element enters the defect area, it rotates around the leading edge of the defect. The radial displacement of the rolling element undergoes a sudden change. When the rolling element simultaneously contacts the leading and trailing edges of the defect, that is... At that time, the radial displacement reaches its maximum value due to the sudden change. The value of the radial displacement during this abrupt change is a process of "from small to large" followed by "from large to small".

[0039] To address the above situation, traditional research uses a half-sine function to describe this process. However, the instantaneous displacement excitation of the rolling element over an obstacle is related to factors such as the relative position of the rolling element and the size of the fault. A single half-sine function is insufficient to accurately describe the changes in displacement excitation, especially when describing the relative position of the rolling element in the defect area. Therefore, to more accurately reconstruct the rolling element obstacle-crossing process, this method uses geometric relationships to calculate the relative position of the rolling element in the defect area. This method can more accurately capture the motion of the rolling element in the defect area.

[0040] Figure 5 In this context, R represents the outer ring raceway radius of the bearing; r represents the rolling element radius. The angle at which the rolling element enters the defect area; , These are the angular positions of the front and rear edges of the defect area, respectively. The expression is: ; The radial displacement excitation function s for the abrupt change in the rolling element entering the defect region is expressed as: ; Then the radial displacement of the contact deformation between the j-th rolling element and the outer ring raceway of the bearing Revised to: .

[0041] Furthermore, based on Hertzian contact theory, the equivalent contact stiffness between the rigid rolling element and the raceways of the inner and outer rings of the bearing is... It can be represented as: ; In the above formula, , These are the stiffness coefficients of the inner and outer raceways of the bearing, respectively; n is the load-deformation coefficient. In this embodiment, n is taken as 1.5 for the rolling element bearing. After obtaining the contact deformation of the rolling elements and the inner and outer raceways of the bearing, the nonlinear contact force F generated in the directions perpendicular to the inner and outer raceways is calculated, and its magnitude is: , ; and The correction angle for the Hertzian contact force on the outer and inner rings is calculated using the following formula. , , The formula for the Hertzian contact force between the inner and outer rings of the bearing on the rolling element in the radial direction is as follows: , ; In rolling bearings, when the rolling element moves to the edge of the defect area, the elastic deformation contact surface becomes half of its original size. The angular range of this contact surface is defined as the band range. At this time, the above assumption does not hold, so the equivalent contact stiffness between the rolling element and the raceway of the inner and outer rings of the bearing needs to be corrected.

[0042] The formula for calculating time-varying contact stiffness is as follows: ; In the formula, This is the corrected time-varying contact stiffness coefficient. This is a reduction factor, with a range of values. >0.5; S is the shape factor, which is set to 1 here; to simplify the calculation, the band range is used here. Take a constant value, that is: .

[0043] Furthermore, such as Figure 6 As shown, during the process of the rolling element over the obstacle, in addition to the sudden change in the radial displacement of the rolling element mentioned above, an instantaneous impact force will also be generated at the moment when the rolling element contacts the edge line of the defect. ; In the above formula, The radial velocity of the rolling element; The value is related to the angular position of the rolling element, and the expression is as follows: ; The formula for the radial velocity of the rolling element can be simplified to: ; In the formula, Let be the angle between the impact force and the radial velocity of the rolling element. Based on geometric relationships, its expression is: ; Impact force duration The time from when the rolling element impacts the edge of the defect to when it leaves the defect area is expressed as: ; The impact force generated at the edge of the rolling element after impacting the defect can be calculated using the impulse theorem. Substituting the above parameters into the equation, we get: ; In the above formula, L is the defect width: m b For the mass of the rolling element; Will Orthogonal decomposition yields horizontal and vertical components: ; When the rolling element passes through the defect area of ​​the outer ring, the magnitude of the total Hertzian contact force on the inner and outer rings of the bearing can be calculated using the following formula: , ; In the formula, The discriminant function for determining the presence of impact force is defined as follows: ; After correction, the total Hertzian contact force on the inner and outer rings of the bearing model with outer ring defects is shown below: , .

[0044] In summary, the nonlinear dynamic equation of the rolling bearing is: ; The first two lines are the equations of motion for the inner circle, where, For the quality of the inner ring, This is the damping coefficient between the inner ring and the bearing housing. This refers to the support stiffness between the inner ring and the bearing housing. , These represent the displacements of the inner ring in the x and y directions, respectively. , For speed, , For acceleration, , F represents the x and y components of the contact force between the inner ring and the rolling element, respectively. r Where is the bearing load, and e is the eccentricity of the inner ring. ω is the angular velocity of the inner ring. The third and fourth lines are the equations of motion for the outer circle, where... For the quality of the outer ring, , These represent the support stiffness and damping coefficient between the outer ring and the bearing housing, respectively. , These represent the horizontal and vertical displacements of the shaft and the outer ring, respectively. , These represent the displacements of the center of the rolling element in the x and y directions, respectively. , These are the outer ring fault excitation forces in the x and y directions, respectively; The fifth and sixth lines are the equations of motion for the rolling element, in which... For the mass of a single rolling element; The last line is the equation of motion for the cage, where, To maintain rack quality, Let J be the cage angular displacement corresponding to the j-th rolling element. This represents the extreme diameter of the j-th rolling element. To maintain the angular velocity of the frame, This indicates the angular position of the j-th rolling element. , The Hertzian contact force between the inner and outer rings of the bearing on the rolling elements in the radial direction.

[0045] This set of nonlinear dynamic equations has no analytical solution and needs to be solved numerically to obtain the vibration response of the system in the time / frequency domain.

[0046] Furthermore, the nonlinear dynamic model of the rolling bearing is solved by the fourth-order Runge-Kutta method to generate simulated vibration signals under the four types of working conditions, forming a labeled source domain dataset; based on the generated simulation data, fault simulation is performed to construct a source domain dataset with complete fault labels.

[0047] It should be noted that the fourth-order Runge-Kutta method (RK4) is a commonly used high-precision single-step numerical solution method in engineering and numerical computation. It is used to solve initial value problems of first-order ordinary differential equations. Its core advantages are high accuracy and good stability. It does not require the calculation of higher-order derivatives and can complete the solution in one step through only four function value iterations. Since it is a commonly used mathematical method, its solution process will not be described in detail in this embodiment.

[0048] For example, for the SKF6205-2RS deep groove ball bearing, the inner ring radius is set to 25mm, the outer ring radius to 52mm, the number of rolling elements to 9, the radial clearance to 14μm, the bearing load to 1000N, the speed to 1797r / min, and the sampling frequency to 12kHz. A nonlinear dynamic model is established based on Hertz contact theory, and the displacement excitation function, equivalent contact stiffness, and impact force are corrected. The model is solved to generate simulation signals for four states, with 300 samples for each state, forming a source domain dataset.

[0049] Step 120: Construct the target domain dataset.

[0050] Please refer to Figure 2 , Figure 2 A flowchart of a method for constructing a target domain dataset provided in an embodiment of the present invention; step 120 specifically includes: Step 121: Collect the measured vibration signals of the rolling bearing under four working conditions: normal state, inner ring failure, outer ring failure, and rolling element failure, as measured samples to construct the original dataset.

[0051] For example, CWRU laboratory bearing data 6205-2RS and 6203-2RS were selected, covering fault data with speeds of 1730 r / min to 1797 r / min and loads of 0 W to 2205 W. Similarly, a target domain dataset of 300 samples for each state was constructed. All samples were converted into 32×32 two-dimensional grayscale images, normalized to the [0,1] interval, and the training set and test set were divided in an 8:2 ratio.

[0052] Step 122: Select the amplitude spectrum of the measured sample as the fitness function of the CS-GWO algorithm, and use the minimum value of the amplitude spectrum entropy as the fitness function to search for the optimal parameter combination in the VMD algorithm. .

[0053] The specific implementation method of step 122 is as follows: Step 1221: Initialize the parameters of the CS-GWO algorithm and VMD.

[0054] The Grey Wolf (GWO) algorithm is widely used in function optimization and other problems due to its strong search capabilities and high efficiency. However, its position update method makes its global search ability relatively weak, and it is prone to getting trapped in local minima. The Cuckoo Search (CS) algorithm, on the other hand, requires fewer parameters and can easily jump out of the current region, achieving a global search. The CS algorithm is inspired by Levi's flight and the parasitic feeding mechanism of the cuckoo. The location of the host's nest, parasitized by the cuckoo, is iteratively updated according to a formula based on the cuckoo's unique search method: ; In the above formula, i = 1, 2, ..., N, where N represents the number of host bird nests; This is an inner product operation; This represents the coordinates of the i-th host nest during the t-th iteration; Let these be the coordinates of the i-th host nest after iteration; =1 is the step size factor; It follows a Lévy distribution with a step size of .

[0055] Combining the advantages of the CS algorithm, such as its ability to easily jump from the computational region to the uncomputed region during global search, the GWO algorithm, which is improved upon, effectively overcomes the problem of the GWO algorithm easily getting trapped in local optima during the optimization process. This improvement allows the algorithm to better explore the search space, enhances its global search capability, and thus makes it more effective in solving various optimization problems.

[0056] Step 1222: Using the minimum value of amplitude spectral entropy as the fitness function, search for the optimal parameter combination in the VMD algorithm using the CS-GWO algorithm. .

[0057] Specifically, the fitness function is the minimum value of the amplitude spectral entropy of the modal components. The optimal parameter combination for VMD is obtained through iterative optimization using the CS-GWO algorithm, thus avoiding mode aliasing. The amplitude spectra are obtained by performing FFT transformation on the modal components obtained from VMD decomposition. The amplitude spectrum is then combined with the information entropy to calculate the amplitude spectral entropy, using the following formula: ; In the above formula, L i For modal components u i The amplitude spectrum, H i For modal components u i The amplitude spectral entropy, where N is the length of the modal component.

[0058] Among the many parameters of VMD, the number of decomposition modes K and the second-order penalty factor have the greatest impact on the results. Choose the VMD parameter combination. Using it as the fitness function optimization objective of the CS-GWO algorithm can effectively avoid obtaining suboptimal solutions due to improper parameter combinations.

[0059] In this step, the change in amplitude spectral entropy can reflect whether the signal contains a lot of fault characteristic information. When the modal components obtained by VMD decomposition contain a lot of noise components and fewer fault impact signal components, the amplitude spectral entropy of the modal component will be large, as can be seen from the formula calculation results. Conversely, when the modal component contains less noise signal, the regular periodic fault impact signal response is obvious, that is, the modal component is rich in fault characteristic information and has a small amplitude spectral entropy.

[0060] In summary, the signal decomposition process of the VMD algorithm includes five steps: analytic signal calculation, spectral modulation, solving the constrained variational problem, iterative update using the alternating direction multiplier method, and center frequency optimization. By using a quadratic penalty factor and the Lagrange multiplication operator, the constrained variational problem is transformed into an unconstrained problem, ensuring the stability and accuracy of the decomposition process.

[0061] Step 123: Use the VMD algorithm with optimal parameter combination to decompose the measured samples in the original dataset to obtain K modal components.

[0062] The measured vibration signal Decomposed into K modal components; set up This represents the K modal components obtained after decomposition. The center frequency of each component; For each modal component, the analytic signal is calculated using the Hilbert transform, thus obtaining the one-sided spectrum: j is the imaginary unit. It is a unit impulse function. This is a convolution operation; The spectrum of the modal function is modulated by aliasing the exponential term corresponding to the center frequency of each modal function: ; Gaussian smoothing is used to smooth the spectrum of each mode function, the bandwidth of the mode function is estimated, and a constrained variational problem is solved to minimize the sum of the estimated bandwidths of each mode function; the constrained variational expression is: ; This represents the derivative of the function with respect to time t.

[0063] The solution to the constrained variational problem is specifically as follows: Using a secondary penalty factor and Lagrange multiplication operators Solving the constrained variational expression transforms the constrained variational problem into an unconstrained variational problem, i.e.: ; Solving unconstrained variational problems using the alternating direction multiplier method, through alternating updates... , as well as Find the saddle point of the augmented Lagrange expression, and solve for it: ;as well as, The method for updating the center frequency is as follows: .

[0064] Step 124: Calculate the correlation between each modal component and the measured sample, retain the highly correlated modal components and sum them to reconstruct the target domain data in the target domain dataset.

[0065] Specifically, correlation analysis is performed on each modal component and the original vibration signal to assess their correlation degree; the correlation calculation formula is as follows: ; In the formula, The signal of the modal component at sampling point i, This is the signal mean of the modal signal. The original vibration signal is the signal at sampling point i. The mean value of the original vibration signal is denoted as N, and the number of sampling points is N. Modal components with correlation greater than or equal to a predetermined correlation threshold are retained and summed to reconstruct the original vibration signal.

[0066] For example, the initial parameters of the CS-GWO algorithm are: The wolf pack size is 10, the maximum number of iterations is 30, the number of modalities K has a search range of [2,10], and the quadratic penalty factor α has a search range of [1000,20000]. Using the minimum value of amplitude spectral entropy as the fitness function, the optimal parameters (K=6, α=12000) are searched using the CS-GWO algorithm. The target domain test set signal is decomposed using the optimal parameter VMD, the correlation between each modal component and the original vibration signal is calculated, modal components with a correlation ≥ 0.8 are selected for reconstruction, envelope spectrum analysis is performed on the reconstructed signal, and fault characteristic frequencies are extracted.

[0067] Please refer to Figure 8a and Figure 8b This figure compares the time-domain waveforms and corresponding two-dimensional grayscale images of the source domain (simulation data) and the target domain (laboratory data). The source domain data exhibits a striped grayscale distribution due to the regularity of the dynamic model, while the target domain data has a more complex grayscale distribution due to noise. Both the source and target domain time-domain waveforms contain fault impact components, but the target domain signal noise interference is more obvious, verifying the necessity of the feature extraction algorithm.

[0068] Step 130: Randomly divide a portion of the target domain dataset and the source domain dataset into a training set, divide the remaining portion of the target domain dataset into a test set, and train the pre-built improved alternating transfer learning model.

[0069] like Figure 7 As shown, the improved alternating transfer learning model of this application is a CNN model containing five convolutional layers, two pooling layers, and two fully connected layers. A batch normalization layer is added after each convolutional layer. The ReLU function is used as the activation function of the hidden layer, and the Softmax function is used as the activation function of the output layer. The first pooling layer is located between the second and third convolutional layers, and the second pooling layer is located after the fifth convolutional layer and before the first fully connected layer. The CORAL loss function is calculated after the first convolutional layer of the CNN model to reduce the difference in second-order statistics between the source and target domain datasets. The sum of the MMD loss function and the classification loss is calculated in the fully connected layer, and the network weights and bias parameters are updated by alternating backpropagation.

[0070] Specifically, the method for setting the CORAL loss function is as follows: The distance between the second-order statistics of the data features in the source domain dataset and the target domain dataset is defined as the CORAL loss function L1: , is the Frobenius norm of the mean square matrix; in, Let D be the number of dimensions of this sample, and D be the source domain dataset.S , and the target domain dataset D T The characteristic covariance matrices are C S With C T ; MMD is used to measure the feature set D. S and D T The distributional differences in the reproducing kernel Hilbert space are used as the sum of the MMD loss and the classification loss as the L2 loss function: , , , For the regenerating nucleus Hilbert space; In the above formula, L C The classification loss is the difference between the true label and the predicted label. , Mapping function , n is the balance coefficient. S Let y be the number of samples, and y be the true label of the sample data. The label predicted by the classifier.

[0071] In this embodiment, the training of the improved alternating transfer learning model includes several steps: CNN pre-training, alternating transfer training, and testing and verification. Feature extraction is only applied to the target domain dataset. During the training process, CORAL loss update and MMD-cross-entropy loss update are performed alternately until the maximum number of iterations is reached.

[0072] For example, the CNN model parameters are set as follows: Convolutional layer 1 (5×5 kernels, 16 kernels), Convolutional layer 2 (3×3 kernels, 16 kernels), Pooling layer 1 (2×2 kernels, stride 2), Convolutional layers 3-5 (3×3 kernels, 32 / 32 / 64 kernels), Pooling layer 2 (2×2 kernels, stride 2), Fully connected layer 1 (1024 dimensions), Fully connected layer 2 (128 dimensions); Batch size 64, learning rate 0.001, number of iterations 100; The CNN is trained using source domain data during the pre-training phase; During the transfer training phase, the CORAL loss of convolutional layer 1 and the MMD-cross-entropy loss of fully connected layer are calculated alternately, and the parameters are updated by backpropagation. Test and verification: The processed target domain dataset was input into the learning model, and the fault diagnosis results were output. The results showed that the accuracy of the transfer diagnosis under different operating conditions was 92.89%, the accuracy of the transfer diagnosis across bearing models was 89.82%, and the accuracy of the transfer diagnosis for small samples was 84.65%.

[0073] Step 140: Use the trained improved alternating transfer learning model to perform full-life rolling bearing fault diagnosis and output the diagnosis results.

[0074] like Figure 9 As shown in the figure, the fault classification confusion matrix of TCA, CNN, CNN-MMD, CNN-CORAL, LATL and the algorithm of this invention is compared. The algorithm of this invention has the highest proportion of diagonal elements (94.32%), indicating that the classification accuracy is the best. Other algorithms have more misclassifications between rolling body faults and inner ring faults, while the algorithm of this invention effectively reduces inter-class confusion, proving that the inter-domain feature transfer adaptation effect is better. As can be seen, this algorithm has a high classification accuracy, which is basically 100%, effectively reducing the number of confused samples and correctly classifying samples that were misclassified in other algorithms.

[0075] like Figure 10 As shown in the figure, this diagram illustrates the t-SNE feature visualization results for 16 cross-condition migration tasks. Different colors represent different fault types. The features extracted by the algorithm of this invention exhibit more concentrated clustering and larger inter-class distances in the low-dimensional space. This verifies that the algorithm of this invention can effectively reduce the differences in feature distribution between domains, ensure the separability of different fault categories, and improve classification accuracy.

[0076] In summary, this application generates high-quality source domain data through a mechanistic model, accurately extracts fault features using a CS-GWO optimized VMD algorithm, and addresses the cross-domain data adaptation problem by combining an improved alternating transfer learning algorithm. During implementation, the dynamic model parameters can be adjusted according to the actual bearing model. Feature extraction performance can be optimized by adjusting parameters such as the wolf pack size and iteration count of the CS-GWO algorithm, and model training efficiency can be improved by adjusting parameters such as the CNN convolution kernel size and learning rate. This method is applicable to fault diagnosis of rolling bearings in various rotating machinery, especially suitable for laboratory scenarios where fault data is scarce and operating conditions are variable, and has broad engineering application value.

[0077] Based on the same concept, embodiments of the present invention also provide a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement a method for constructing a target domain dataset or a method for diagnosing rolling bearing faults provided in embodiments of the present invention.

[0078] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0079] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0080] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including—but not limited to—wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0081] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0082] Based on the same concept, embodiments of the present invention also provide a computer program product, including a computer program / instruction, which, when executed by a processor, implements a method for constructing a target domain dataset or a method for diagnosing rolling bearing faults provided in embodiments of the present invention.

[0083] Computer program products may be loaded onto computer devices, and the components of computer devices may include, but are not limited to: one or more processors or processing units, system memory, and buses connecting different system components (including system memory and processing units).

[0084] Computer devices typically include a variety of computer system-readable media. These media can be any available media that can be accessed by a computer device, including volatile and non-volatile media, and removable and non-removable media.

[0085] System memory may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The computer device may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the storage system may be used to read and write non-removable, non-volatile magnetic media. The computer program product has a set (e.g., at least one) of program modules configured to perform the functions of the various embodiments of the present invention.

[0086] A program / utility having a set (at least one) of program modules can be stored, for example, in memory. Such program modules include, but are not limited to, an operating system, one or more applications, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. The program modules typically perform the functions and / or methods described in the embodiments of this invention.

[0087] Computer devices can also communicate with one or more external devices (such as keyboards, pointing devices, monitors, etc.), one or more devices that enable users to interact with the computer device, and / or any device that enables the computer device to communicate with one or more other computing devices (such as network interface cards, modems, etc.). This communication can be performed through input / output (I / O) interfaces. Furthermore, computer devices can communicate with one or more networks (such as local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via network adapters. As shown in the figure, the network adapter communicates with other modules of the computer device via a bus. It should be understood that other hardware and / or software modules can be used in conjunction with the computer device, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0088] The processing unit executes various functional applications and data processing by running programs stored in the system memory, such as implementing a method for constructing a target domain dataset or a method for diagnosing rolling bearing faults provided in the embodiments of the present invention.

[0089] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware, or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0090] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a target domain dataset, characterized in that, include: Vibration signals measured under four operating conditions of rolling bearings—normal condition, inner ring fault, outer ring fault, and rolling element fault—were collected as measured samples to construct the original dataset. The amplitude spectrum of the measured samples is selected as the fitness function of the CS-GWO algorithm, and the minimum value of the amplitude spectrum entropy is used as the fitness function to search for the optimal parameter combination in the VMD algorithm. Where K is the modality number K, It is a secondary penalty factor; The VMD algorithm with optimal parameter combination is used to decompose the measured samples in the original dataset to obtain K modal components; Calculate the correlation between each modal component and the measured sample, retain the highly correlated modal components and sum them to reconstruct the target domain data in the target domain dataset.

2. The method for constructing a target domain dataset according to claim 1, characterized in that, The VMD algorithm, which utilizes optimal parameter combinations, decomposes the measured samples in the original dataset to obtain K modal components, specifically including: The measured vibration signal Decomposed into K modal components; set up This represents the K modal components obtained after decomposition. The center frequency of each component; For each modal component, the analytic signal is calculated using the Hilbert transform, thus obtaining the one-sided spectrum: j is the imaginary unit. It is a unit impulse function. This is a convolution operation; The spectrum of the modal function is modulated by aliasing the exponential term corresponding to the center frequency of each modal function: ； Gaussian smoothing is used to smooth the spectrum of each mode function, the bandwidth of the mode function is estimated, and a constrained variational problem is solved to minimize the sum of the estimated bandwidths of each mode function; the constrained variational expression is: ; This represents the derivative of the function with respect to time t.

3. The method for constructing a target domain dataset according to claim 2, characterized in that, The solution to the constrained variational problem is specifically as follows: Using a secondary penalty factor and Lagrange multiplication operators Solving the constrained variational expression transforms the constrained variational problem into an unconstrained variational problem, i.e.: ； Solving unconstrained variational problems using the alternating direction multiplier method, through alternating updates... , as well as Find the saddle point of the augmented Lagrange expression, and solve for it: ； as well as, The method for updating the center frequency is as follows: 。 4. The method for constructing a target domain dataset according to claim 3, characterized in that, The process involves calculating the correlation between each modal component and the measured sample, retaining highly correlated modal components, summing and reconstructing them to obtain the target domain data in the target domain dataset, including: Correlation analysis was performed on each modal component and the original vibration signal to assess their correlation degree; the correlation calculation formula is as follows: ； In the formula, The signal of the modal component at sampling point i, This is the signal mean of the modal signal. The original vibration signal is the signal at sampling point i. The mean value of the original vibration signal is denoted as N, and the number of sampling points is N. Modal components with correlation greater than or equal to a predetermined correlation threshold are retained and summed to reconstruct the original vibration signal.

5. The method for constructing a target domain dataset according to claim 1, characterized in that, The amplitude spectrum of the selected measured samples is used as the fitness function of the CS-GWO algorithm, and the minimum value of the amplitude spectrum entropy is used as the fitness value to search for the optimal parameter combination in the VMD algorithm. Specifically, it includes: Initialize the parameters of the CS-GWO algorithm and VMD; Using the minimum amplitude spectral entropy as the fitness function, the optimal parameter combination in the VMD algorithm is searched using the CS-GWO algorithm. The formula for calculating the amplitude spectral entropy is: ； In the above formula, L i For modal components u i The amplitude spectrum, H i For modal components u i The amplitude spectral entropy, where N is the length of the modal component.

6. A method for diagnosing rolling bearing faults, characterized in that, include: Based on the nonlinear dynamic equations of rolling bearings, simulation data are generated under four types of working conditions: normal state, inner ring fault, outer ring fault, and rolling element fault, and a source domain dataset with complete fault labels is constructed. Furthermore, the target domain dataset is constructed using a method for constructing a target domain dataset as described in any one of claims 1-5; A portion of the target domain dataset and the source domain dataset are randomly divided into a training set, and the remaining portion of the target domain dataset is divided into a test set to train a pre-built improved alternating transfer learning model. A trained improved alternating transfer learning model is used for full-life rolling bearing fault diagnosis, and the diagnosis results are output.

7. The method for diagnosing rolling bearing faults according to claim 6, characterized in that, The method for constructing the improved alternating transfer learning model is as follows: Construct a CNN model containing five convolutional layers, two pooling layers, and two fully connected layers. Add a batch normalization layer after each convolutional layer. Use ReLU as the activation function for the hidden layers and Softmax as the activation function for the output layer. The first pooling layer is located between the second and third convolutional layers, and the second pooling layer is located after the fifth convolutional layer and before the first fully connected layer. The CORAL loss function is calculated after the first convolutional layer of the CNN model to reduce the difference in second-order statistics between the source and target domain datasets. The sum of the MMD loss function and the classification loss is calculated in the fully connected layer, and the network weights and bias parameters are updated by alternating backpropagation.

8. The method for diagnosing rolling bearing faults according to claim 7, characterized in that, The method for setting the CORAL loss function is as follows: The distance between the second-order statistics of the data features in the source domain dataset and the target domain dataset is defined as the CORAL loss function L1: , is the Frobenius norm of the mean square matrix; in, Let D be the number of dimensions of this sample, and D be the source domain dataset. S , and the target domain dataset D T The characteristic covariance matrices are C S With C T ; MMD is used to measure the feature set D. S and D T The distributional differences in the reproducing kernel Hilbert space are used as the sum of the MMD loss and the classification loss as the L2 loss function: ， ， , For the regenerating nucleus Hilbert space; In the above formula, L C The classification loss is the difference between the true label and the predicted label. , Mapping function , n is the balance coefficient. S Let y be the number of samples, and y be the true label of the sample data. The label predicted by the classifier.

9. A method for diagnosing rolling bearing faults according to claim 8, characterized in that, The nonlinear dynamic equations based on rolling bearings generate simulation data for four operating conditions: normal state, inner ring fault, outer ring fault, and rolling element fault. A source domain dataset with complete fault labels is constructed, including: Based on Hertz contact theory, a nonlinear dynamic model of the rolling bearing is established by modifying the radial displacement excitation function, equivalent contact stiffness, and impact force of the rolling element. The radial displacement excitation function is modified according to the geometric position of the rolling element in the defect region, and the impact force is modified according to the size of the fault defect and the bearing speed. The nonlinear dynamic model of the rolling bearing is solved by the fourth-order Runge-Kutta method, and the simulated vibration signals under the four working conditions are generated to form a labeled source domain dataset.

10. A computer program product comprising a computer program / instructions, characterized in that, When a computer program / instruction is executed by a processor, it implements a method for constructing a target domain dataset as described in any one of claims 1-5.

Images