Bearing fault detection method and device

Abstract

The invention provides a bearing fault detection method and device. The method comprises the following steps: acquiring a to-be-detected bearing signal; inputting the to-be-detected bearing signal into a pre-trained fault detection model to obtain a fault detection result; wherein the fault detection model is obtained by training a pre-constructed training data set based on a neural network, and the training data set is generated by inputting an actual measurement signal sample and a simulation signal sample into a pre-trained frequency domain bidirectional cyclic generative adversarial network. According to the invention, the pre-trained frequency domain bidirectional cyclic generative adversarial network is used to generate the fault sample with the fault features and the actually measured noise features, so that the training data set is obtained, and the training data set is used to train the fault detection model. The problems of fault sample scarcity, simulation-actual measurement data difference and insufficient feature fidelity in the prior art are solved, and the precision and generalization ability of bearing fault detection are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical fault diagnosis, and in particular to a bearing fault detection method and device. BACKGROUND

[0002] As an important component of rotating machinery, bearings play a key role in the fields of aerospace, automotive industry, wind power generation, etc. The running state of the bearing system directly affects the safety and reliability of the entire mechanical system. Once a fault occurs, the maintenance cost is extremely high. If the fault cannot be diagnosed in time, it may cause a chain of equipment damage, leading to safety accidents and causing significant economic losses. Through timely fault diagnosis, the fault can be detected at an early stage, thereby avoiding potential risks caused by further expansion of the fault.

[0003] There are two mainstream paradigms in the field of fault diagnosis: data-driven methods and model-based diagnosis techniques. Data-driven methods, represented by deep learning (such as vibration signal analysis based on convolutional neural networks), achieve fault recognition by mining hidden features in massive data, and exhibit strong adaptability in complex working conditions. However, its implementation is highly dependent on sufficient and balanced fault samples, and it is prone to diagnosis failure in early fault detection or extreme working conditions. Complementary to this, the model-based method avoids data dependence through physical modeling, has the advantages of strong explainability and clear mechanism, but the modeling process requires a lot of expert experience, and it is difficult to accurately depict the dynamic characteristics of complex nonlinear systems.

[0004] To break through the limitations of traditional methods, the "digital and model dual driving" method combining physical models and data-driven methods has emerged. However, the current digital and model dual driving method has some problems, for example, the difference in frequency domain characteristics between simulation signals and measured signals limits the generalization ability of cross-domain mapping models; the full life cycle dynamic model relies on measured data for iterative updating, and early fault weak signals are prone to introduce parameter identification errors; high-fidelity modeling requires multidisciplinary collaboration, and engineering complexity restricts technology popularization; noise interference in the cross-domain migration process easily destroys feature consistency, affecting the robustness of diagnosis. That is, due to the lack of fault samples, the large difference between simulation and measured data, and the lack of feature fidelity, the model trained according to the fault samples has low diagnosis precision and generalization ability.

[0005] In summary, the existing technology has the problem of low precision and generalization ability of the fault detection model caused by bearing fault samples. SUMMARY

[0006] The present application provides a bearing fault detection method and device to solve the problem of low precision and generalization ability of the fault detection model caused by bearing fault samples in the prior art, and to realize high-precision and high-generalization bearing fault detection.

[0007] The application provides a bearing fault detection method, comprising the following steps:

[0008] Obtaining a bearing signal to be detected;

[0009] Inputting the bearing signal to be detected into a pre-trained fault detection model to obtain a fault detection result;

[0010] The fault detection model is based on a neural network and is obtained by training a pre-constructed training data set, and the training data set is obtained by inputting measured signal samples and simulation signal samples into a pre-trained frequency domain bidirectional cyclic generative adversarial network.

[0011] According to the bearing fault detection method provided by the application, the measured signal samples and the simulation signal samples are input into the pre-trained frequency domain bidirectional cyclic generative adversarial network to generate the training data set, comprising:

[0012] Obtaining a plurality of measured signal samples and a plurality of simulation signal samples;

[0013] Converting the measured signal samples into measured domain frequency domain tensors and converting the simulation signal samples into simulation domain frequency domain tensors;

[0014] Based on the bidirectional long short-term memory network, a generator is constructed, and a discriminator is constructed based on PatchGAN to obtain an initial cyclic generative adversarial network;

[0015] According to the measured domain frequency domain tensor and the simulation domain frequency domain tensor, the initial cyclic generative adversarial network is trained by using a pre-constructed total loss function to obtain the frequency domain bidirectional cyclic generative adversarial network;

[0016] The measured domain frequency domain tensor and the simulation domain frequency domain tensor are input into the frequency domain bidirectional cyclic generative adversarial network to obtain a measured domain-simulation domain enhanced data tensor pair;

[0017] The measured domain-simulation domain enhanced data tensor pair is subjected to inverse fast Fourier transform to obtain a measured domain-simulation domain training sample pair, and then a training data set is obtained.

[0018] According to the bearing fault detection method provided by the application, the simulation signal samples are obtained by inputting simulation parameters into a pre-constructed rigid-flexible coupling dynamics model;

[0019] The rigid-flexible coupling dynamics model is a finite element model constructed based on the multi-body dynamics theory, simulating the vibration response characteristics of bearing inner and outer ring faults by parameterized modeling, and introducing a noise injection mechanism to simulate the measured environmental interference.

[0020] The bearing fault detection method provided by the application, the finite element model is a model obtained by discretizing the bearing through a free tetrahedron mesh division algorithm and locally refining the contact area of the bearing through a second-order gradient driven adaptive mesh division strategy;

[0021] The trigger condition of the adaptive mesh division strategy is:

[0022]

[0023] When η e > 0.7, the size of the region unit is ensured to meet:

[0024]

[0025] In the formula, η e is the local change rate of the stress field; is the second-order gradient of the von Mises stress; ‖·‖ e is the Euclidean norm of the variable in the unit, indicating the amplitude of the stress gradient in the unit; h e is the size of the region unit; D eq is the equivalent diameter (unit: m), which is calculated by the projected area A e of the unit.

[0026] The bearing fault detection method provided by the application, the total loss function comprises:

[0027]

[0028] In the formula, is the total loss function, G A→B is a generator from a simulation domain to a measured domain; G B→A is a generator from a measured domain to a simulation domain; D A is a discriminator of the simulation domain; D B is a discriminator of the measured domain; is a first adversarial loss function; is a second adversarial loss function; is a cycle consistency loss function; λ cyc is a cycle consistency loss weight parameter; λ identity is an identity mapping loss weight parameter; is an identity mapping loss function; λ sp is a spectral constraint loss weight parameter; L sp (G A→B ) is a first spectral constraint loss function, L sp (G B→A ) is a second spectral constraint loss function.

[0029] According to the bearing fault detection method provided by the application, the fault detection model is trained based on a neural network and by using a pre-constructed training data set, and the method comprises the following steps:

[0030] An initial classification model is constructed based on a neural network.

[0031] The initial classification model is trained according to the measured domain-simulation domain training sample pair in the training data set, so as to learn the time domain features of the training samples in the measured domain and the simulation domain, until the confusion matrix and / or the classification accuracy rate meet a preset condition, and the training result is output as the fault detection model.

[0032] The application further provides a bearing fault detection device, comprising the following modules:

[0033] A data acquisition unit is configured to acquire a bearing signal to be detected.

[0034] A fault detection unit is configured to input the bearing signal to be detected into a pre-trained fault detection model, so as to obtain a fault detection result.

[0035] The fault detection model is trained based on a neural network and by using a pre-constructed training data set, and the training data set is generated by inputting measured signal samples and simulation signal samples into a pre-trained frequency domain bidirectional cyclic generative adversarial network.

[0036] The application further provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the bearing fault detection method according to any one of the above when executing the program.

[0037] The application further provides a non-transitory computer readable storage medium having a computer program stored thereon, and the computer program is executable on a processor to implement the bearing fault detection method according to any one of the above.

[0038] The application further provides a computer program product comprising a computer program, and the computer program is executable on a processor to implement the bearing fault detection method according to any one of the above.

[0039] The bearing fault detection method and device provided by the application generate fault samples with fault features and measured noise characteristics by using a pre-trained frequency domain bidirectional cyclic generative adversarial network, so as to obtain a training data set, and train a fault detection model by using the training data set, thereby solving the problems of a lack of fault samples, a simulation-measured data difference, and a lack of feature fidelity in the prior art, and improving the precision and generalization ability of bearing fault detection. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to make the technical solutions in the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those of ordinary skill in the art without any creative effort.

[0041] Figure 1 is one of the flowcharts of the bearing fault detection method provided by the present application;

[0042] Figure 2 is another flowchart of the bearing fault detection method provided by the present application;

[0043] Figure 3 is the diagnostic precision confusion matrix of one embodiment of the bearing fault detection method provided by the present application, wherein, Figure 3 (a) is the diagnostic precision confusion matrix of the simulation signal, Figure 3 (b) is the diagnostic precision confusion matrix of the CycleGAN generated signal, Figure 3 (c) is the diagnostic precision confusion matrix of the FB-CycleGAN generated signal, Figure 3 (d) is the diagnostic precision confusion matrix of the real experimental signal;

[0044] Figure 4 is a structural schematic diagram of the bearing fault detection device provided by the present application;

[0045] Figure 5 is a structural schematic diagram of the electronic device provided by the present application. DETAILED DESCRIPTION

[0046] In order to make the technical solutions in the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those of ordinary skill in the art without any creative effort.

[0047] The bearing fault detection method of the present application will be described below in combination with Figures 1-3 The bearing fault detection method of the present application will be described below in combination with Figure 1 is one of the flowcharts of the bearing fault detection method provided by the present application, as shown in Figure 1 The method comprises the following steps.

[0048] Step 110: acquiring a bearing signal to be detected.

[0049] The bearing signal to be detected is a signal of a bearing to be detected in actual measurement. The application does not limit the acquisition mode of the signal, and the acquisition mode can be selected according to the actual situation.

[0050] Step 120: inputting the bearing signal to be detected into a pre-trained fault detection model to obtain a fault detection result.

[0051] The fault detection model is based on a neural network and is obtained by training a pre-constructed training data set. The training data set is input into a pre-trained frequency domain bidirectional cycle generative adversarial network.

[0052] The application adopts a frequency domain bidirectional cycle generative adversarial network to construct a training data set of a fault detection model. Specifically, the frequency domain bidirectional cycle generative adversarial network (FB-CycleGAN) realizes the mapping from the simulation domain to the actual measurement domain by learning the actual measurement signal sample and the simulation signal sample, so as to generate a large amount of enhanced data with fault characteristics and actual measurement noise characteristics on the basis of limited actual measurement signal samples and simulation signal samples, thereby providing a solid foundation for the training of the fault detection model.

[0053] The fault detection model is a classification model. In actual application, the bearing signal to be detected is input into the fault detection model to obtain a fault detection result of whether there is a fault or not.

[0054] The bearing fault detection method provided by the application generates fault samples with fault characteristics and actual measurement noise characteristics through a pre-trained frequency domain bidirectional cycle generative adversarial network, thereby obtaining a training data set, training a fault detection model by using the training data set, and solving the problems of lack of fault samples, simulation-actual measurement data difference and insufficient feature fidelity in the prior art, thereby improving the precision and generalization ability of bearing fault detection.

[0055] The specific construction method of the training data set is further described below. In some embodiments, the actual measurement signal sample and the simulation signal sample are input into a pre-trained frequency domain bidirectional cycle generative adversarial network to generate the training data set, comprising:

[0056] Step 210: obtaining a plurality of actual measurement signal samples and a plurality of simulation signal samples.

[0057] It should be noted that the actual measurement signal sample is a signal data sample obtained by data acquisition in the actual application scene of the bearing. The application does not limit the specific data acquisition mode, and the data acquisition mode can be selected according to the actual situation.

[0058] Further, the measured signal samples can be real health data, i.e., signal data samples of the bearing in a state without faults, can be real fault data, i.e., signal data samples of the bearing in a state with faults, or can be real health data + real fault data, and the present application does not make specific limitations thereon.

[0059] The simulation signal samples are signals of the bearing simulated by a mathematical model or an algorithm, and the specific mathematical model or algorithm can be selected according to actual conditions.

[0060] In some embodiments, a rigid-flexible coupling dynamics model of the bearing-rotor system is constructed to simulate the vibration response characteristics of the bearing components, and simulation health data and / or simulation fault data are generated as simulation signal samples, which will be further described in subsequent embodiments.

[0061] It should be emphasized that the measured signal samples and the simulation signal samples do not necessarily have a strict pairing relationship, and their quantities are not necessarily the same. The core of this step is to obtain independent sets of measured signal samples and simulation signal samples to provide source domain and target domain data for training of the frequency domain bidirectional cycle generative adversarial network.

[0062] Step 220: converting the measured signal samples into measured domain frequency domain tensors and converting the simulation signal samples into simulation domain frequency domain tensors.

[0063] Before training the frequency domain bidirectional cycle generative adversarial network, the data needs to be preprocessed first, i.e., converting the time series data (measured signal samples and simulation signal samples) into frequency domain tensors.

[0064] In some embodiments, the measured signal samples / simulation signal samples are converted into measured domain frequency domain tensors / simulation domain frequency domain tensors by fast Fourier transform (FFT).

[0065] The present application does not make specific limitations on the input format of the frequency domain tensors, which can be set according to actual conditions.

[0066] In one specific embodiment, the input format of the frequency domain tensors is three-channel data of 64x64x3, including real part, imaginary part and amplitude spectrum.

[0067] Step 230: constructing a generator based on a bidirectional long short-term memory network and constructing a discriminator based on PatchGAN to obtain an initial cycle generative adversarial network.

[0068] It needs to be explained that the cycle generative adversarial network (CycleGans) includes a generator and a discriminator, and the core goal is to realize the style or content conversion between different domains without paired training data. The generator is based on a convolutional neural network (CNN), which captures complex patterns in data through residual blocks for learning cross-domain feature mapping. The discriminator improves the physical rationality of the generated data by local frequency domain feature discrimination, and is used to distinguish between real data and data generated by the generator.

[0069] The application takes the cycle generative adversarial network as the basis of the frequency domain bidirectional cycle generative adversarial network, enhances the feature representation ability of the vibration signal through the bidirectional extraction mechanism of the time sequence feature, constructs the correlation model of the simulation signal and the measured signal in the frequency spectrum space based on the frequency domain coupling mapping framework (such as fast Fourier transform (FFT)), and realizes the multi-scale fusion and cross-domain mapping of the frequency domain features.

[0070] The embodiment adopts a bidirectional long short-term memory network (Bi-LSTM) to construct the generator, and adopts a PatchGAN to construct the discriminator.

[0071] Further, the generator includes a down-sampling convolution module, a residual block, and an up-sampling transposed convolution module.

[0072] In some embodiments, a first generator G A→B and a second generator G B→A are constructed based on a bidirectional long short-term memory network (Bi-LSTM). A→B G B→A is used to convert the data of domain A (simulation domain) to domain B (measured domain), and G A→B then converts domain B (measured domain) back to domain A (simulation domain).

[0073] Based on the above embodiment, a Bi-LSTM module with 128 hidden units is used to construct the first generator G A→B and the second generator G B→A .

[0074] Based on the above embodiment, the frequency domain bidirectional cycle generative adversarial network captures the harmonic evolution law of the fault frequency through the bidirectional extraction mechanism of the time sequence feature, that is:

[0075]

[0076] In the formula, X[k] is a complex frequency domain sequence, the real part Re(X[k]) and the imaginary part Im(X[k]) of which correspond to the orthogonal components of the signal at the frequency f k =k·f s / N respectively (f s is the sampling frequency).

[0077] In some embodiments, the discriminator comprises a multi-layer convolutional down-sampling module for extracting local frequency domain features to finely discriminate the frequency domain local features, and the last layer outputs a reality probability.

[0078] In some embodiments, the discriminator comprises a four-layer convolutional down-sampling module.

[0079] It should be emphasized that each generator corresponds to a discriminator. Based on the above embodiments, the first generator G A→B corresponds to the discriminator D B , the second generator G B→A corresponds to the discriminator D A .

[0080] The generator and its corresponding discriminator are combined to construct an initial cycle generative adversarial network as a basic network for training the frequency domain bidirectional cycle generative adversarial network.

[0081] Step 240: training the initial cycle generative adversarial network by using a pre-constructed total loss function according to the measured domain frequency domain tensor and the simulation domain frequency domain tensor, to obtain the frequency domain bidirectional cycle generative adversarial network.

[0082] The optimization objective is constructed by the total loss function, and the initial cycle generative adversarial network is trained by using the measured domain frequency domain tensor and the simulation domain frequency domain tensor until a preset training target is reached, and the training result is output as the frequency domain bidirectional cycle generative adversarial network.

[0083] It should be noted that the total loss function at least includes an adversarial loss, a cycle consistency loss and an identity mapping loss. In some embodiments, in order to jointly optimize the frequency domain amplitude and phase matching degree, the present application further introduces a spectrum constraint loss function into the total loss function, which will be described in subsequent embodiments.

[0084] The preset training target can include a training round or a threshold of a preset index, and the present application does not limit this, which can be set according to actual conditions.

[0085] Step 250: inputting the measured domain frequency domain tensor and the simulation domain frequency domain tensor into the frequency domain bidirectional cycle generative adversarial network to obtain a measured domain-simulation domain enhanced data tensor pair.

[0086] After the frequency domain bidirectional cyclic generative adversarial network is trained, bidirectional mapping of the simulation domain and the measured domain can be realized. The frequency domain tensor of the measured domain / frequency domain tensor of the simulation domain is input into the frequency domain bidirectional cyclic generative adversarial network, and the frequency domain tensor data of the corresponding simulation domain / measured domain can be generated. Thus, the measured domain-simulation domain enhanced data tensor pair is formed with the input frequency domain tensor of the measured domain / frequency domain tensor of the simulation domain.

[0087] It can be understood that the generated measured domain-simulation domain enhanced data tensor pair has both fault characteristics and measured noise characteristics, and solves the problems of fault sample scarcity, large simulation-measured data difference and insufficient feature fidelity in the prior art.

[0088] Step 260: inverse fast Fourier transform is performed on the measured domain-simulation domain enhanced data tensor pair to obtain a measured domain-simulation domain training sample pair, and then a training data set is obtained.

[0089] The measured domain-simulation domain enhanced data tensor pair is converted into time series data, denoted as a measured domain-simulation domain training sample pair, and a plurality of groups of measured domain-simulation domain training sample pairs constitute a training data set.

[0090] Further, as shown in Figure 2 The generated training data set is used to train a fault detection model, and the obtained fault detection model is used to perform steps 110-120, i.e., to obtain a fault detection result according to the real fault data.

[0091] The acquisition of the simulation signal sample is further described below. In some embodiments, the simulation signal sample is obtained by inputting simulation parameters into a pre-constructed rigid-flexible coupling dynamics model.

[0092] The rigid-flexible coupling dynamics model is a finite element model constructed based on the multi-body dynamics theory, using parameterized modeling to simulate the vibration response characteristics of bearing inner and outer ring faults, and introducing a noise injection mechanism to simulate the measured environmental interference.

[0093] The rigid-flexible coupling dynamics model is a rigid-flexible coupling dynamics finite element model of the bearing-rotor system constructed. In the construction process, parameterized modeling is performed based on the multi-body dynamics theory to simulate the damage morphology of the inner ring, the outer ring and the rolling element, i.e., to obtain the vibration response characteristics of the bearing inner and outer ring faults.

[0094] In some embodiments, the dynamic coupling relationship between the rolling element and the inner and outer rings and the retainer is simulated by rigid body contact and hinge joint. The boundary conditions are smoothly loaded with step function for rotation speed and load, and fixed constraints are applied at the bottom of the bearing seat to simulate the actual installation environment.

[0095] It is to be noted that the material properties of the bearing components are set based on actual engineering parameters, including, for example, the modulus of elasticity, density, and Poisson's ratio.

[0096] Further, in one specific embodiment, the inner and outer rings of the bearing and the rolling elements are made of GCr15 material (a high-carbon chromium bearing steel), with a modulus of elasticity of 206 GPa, and the bearing seat is made of cast iron HT150 material (gray cast iron).

[0097] On this basis, the simulation parameters are input into the rigid-flexible coupling dynamics model, so as to generate multi-physical field simulation signals including time-domain waveforms, frequency-domain characteristics, and motion states, which are taken as simulation signal samples. By simulating the bearing vibration signals under different working conditions through different simulation parameters, simulation signal samples under different working conditions are obtained.

[0098] The simulation parameters can include dynamic load and rotating speed.

[0099] Further, based on the above embodiment, in some embodiments, the finite element model is a model in which the bearing is discretized through a free tetrahedron meshing algorithm, and the contact area of the bearing is locally refined through a second-order gradient-driven adaptive meshing strategy.

[0100] The triggering condition of the adaptive meshing strategy is that:

[0101]

[0102] When η e > 0.7, it is ensured that the size of the region element meets:

[0103]

[0104] In the formula, η e is the local change rate of the stress field; is the second-order gradient of the von Mises stress; ‖·‖ e is the Euclidean norm of the variable in the element, indicating the amplitude of the stress gradient in the element; h e is the size of the region element; D eq is the equivalent diameter (unit: m), which is calculated by the projected area A e of the element.

[0105] Specifically, a rigid-flexible coupling dynamics model of the bearing-rotor system is established based on a finite element simulation platform.

[0106] Compared with the traditional simplified model, the bearing contact area is locally refined by adopting a second-order gradient driven adaptive meshing strategy, and a white noise injection mechanism is introduced to simulate the measured environmental disturbance, so that the generated rigid-flexible coupling dynamic model can effectively obtain the vibration response characteristics of the bearing component in the fault state.

[0107] That is, based on the multi-body dynamics theory, the damage morphology of the inner and outer rings and the rolling elements of the bearing is simulated by using the parametric modeling on the finite element simulation platform, the bearing contact area is locally refined by adopting a second-order gradient driven adaptive meshing strategy, and a white noise injection mechanism is introduced to simulate the measured environmental disturbance, and a mathematical model of the bearing-rotor system is constructed as a rigid-flexible coupling dynamic model.

[0108] It should be noted that the mathematical model is discretized by a free tetrahedral meshing method for the bearing-rotor system, and the key contact area of the bearing is meshed.

[0109] In the refinement process, the bearing contact area is locally refined by adopting a second-order gradient driven adaptive meshing strategy. The triggering condition of the adaptive meshing strategy is:

[0110]

[0111] When η e >0.7, it is ensured that the size of the region element meets:

[0112]

[0113] In the formula, η e is the local change rate of the stress field; is the second-order gradient of the von Mises stress; ‖·‖ e is the Euclidean norm of the variable in the element, indicating the amplitude of the stress gradient in the element; h e is the size of the region element; D eq is the equivalent diameter (unit: m), which is calculated by the projected area A e of the element.

[0114] In one specific embodiment, the free tetrahedral mesh is used to locally encrypt the contact area, and the maximum element size is 0.0124m and the minimum element size is 5.29x10 -4 m.

[0115] The total loss function is further described below, and in some embodiments, the total loss function includes:

[0116]

[0117] In the formula, is a total loss function, G A→B is a generator from simulation domain to measured domain; G B→A is a generator from measured domain to simulation domain; D A is a discriminator of simulation domain; D B is a discriminator of measured domain; is a first adversarial loss function; is a second adversarial loss function; is a cycle consistency loss function; λ cyc is a cycle consistency loss weight parameter; λ identity is an identity mapping loss weight parameter; is an identity mapping loss function; λ sp is a spectral constraint loss weight parameter; L sp (G A→B ) is a first spectral constraint loss function, L sp (G B→A ) is a second spectral constraint loss function.

[0118] Specifically, the present application creatively introduces a spectral constraint loss function, jointly optimizes the matching degree of spectral amplitude and phase parameters, and combines an adversarial loss, a cycle consistency loss and an identity mapping loss to form a total loss function, thereby ensuring the dual authenticity of the generated signal in the time-frequency domain.

[0119] The spectral constraint loss forces the generated signal to be aligned with the source signal in the frequency energy distribution through an L1 norm.

[0120]

[0121] In the formula, x is a simulation domain signal, y is a measured domain signal, p data(x) represents the probability distribution of real data in the simulation domain, represents the expectation of the function f(x) when x obeys the p data(x) distribution, ‖‖1 represents an L1 norm, i.e., the sum of the absolute values of each element of a vector, which is used to measure the difference in spectral amplitude between the generated signal and the source signal, represents a spectral transform performed on the input signal, G A→B is a generator from simulation domain to measured domain, is a first spectral constraint loss function, i.e., a spectral constraint loss function in the generation process from simulation domain to measured domain.

[0122] In one specific embodiment, a Fast Fourier Transform (FFT) is used to perform a spectral transform on the input signal, thereby obtaining a frequency domain representation.

[0123] Correspondingly,

[0124] where G B→A is a generator from the measured domain to the simulated domain, L sp (G B→A ) is a second spectral constraint loss function, i.e., a spectral constraint loss function in the generation process from the measured domain to the simulated domain.

[0125] That is, in the training process of step 240, the time domain signal can be reconstructed by inverse fast Fourier transform (IFFT), and compared with the original signal to ensure the physical consistency of cross-domain mapping.

[0126] wherein, in some embodiments, the adversarial loss function is defined as:

[0127]

[0128] wherein x is a simulated domain signal, y is a measured domain signal, denotes the expectation of the function f(y) when y is subject to p data(x) distribution.

[0129] wherein the cycle consistency loss function is defined as:

[0130]

[0131] wherein pdata(y) denotes the probability distribution of the real data in the measured domain, denotes the expectation of the function f(y) when y is subject to pdata(y) distribution.

[0132] It should be explained that the cycle consistency loss (Cycle-Consistency Loss) is used to ensure that the data can maintain its original features after being converted by the bidirectional generator. That is, after the cycle conversion of G A→B and G B→A , the input A domain data should be as consistent as possible with the initial data, and vice versa. This loss function punishes the L1 norm of the generated data and the original data, ensuring that the generator does not just learn a simple mapping, but also performs reasonable conversion based on preserving key features of the data. This loss ensures that the conversion process of CycleGAN is reversible, so that the generated B domain fault signal can still recover to the original A domain simulation signal after reverse mapping.

[0133] wherein the identity mapping loss function is defined as:

[0134]

[0135] It needs to be explained that the identity-mapping loss is used to ensure that the generator can retain the characteristics of the original signal when processing the same domain data. This means that when the real B domain data is input, the generator G A→B The output should be as close as possible to the input B domain data, so as to avoid unnecessary conversion. This loss further enhances the processing capability of CycleGAN for complex signals in actual applications while ensuring that the same domain input signal does not change significantly. The total loss function is the sum of the adversarial loss, the cycle consistency loss, the identity-mapping loss and the spectral constraint loss.

[0136] The following describes the training of the fault detection model. In some embodiments, the fault detection model is trained based on a neural network using a pre-constructed training data set, including:

[0137] Step 310: An initial classification model is constructed based on a neural network.

[0138] In some embodiments, the initial classification model includes an input layer, a convolutional layer, a fully connected layer, and an output layer.

[0139] Further, in one specific embodiment, the initial classification model includes a four-layer convolutional neural network (CNN), which has the following structure: an input layer that receives a 4096-point vibration signal; a first convolutional layer that uses a large-size convolutional kernel (64x1 kernel) to extract global time-domain features and enhances non-linear expression capability through a ReLU activation function; a second convolutional layer that uses a small-size convolutional kernel (24x1 kernel) to refine local features and reduces the dimension through a pooling operation; a fully connected layer that maps the features to a high-dimensional space and prevents overfitting through Dropout regularization; and an output layer that uses a Softmax function to output fault class probability, significantly improving the classification accuracy of the health state, inner circle fault, and outer circle fault.

[0140] On this basis, the data received by the input layer is sequentially passed through the convolutional layer (64x1 and 24x1 kernel), the ReLU activation function, and the fully connected layer, and finally the Softmax outputs the fault class probability.

[0141] First, in some embodiments, the number of output channels of the convolutional layer is 20, the second layer is increased to 30, and the dimension of the fully connected layer is 500, significantly improving the classification robustness in small sample scenarios.

[0142] Step 320: The initial classification model is trained according to the real-simulated domain training sample pairs in the training data set to learn the time-domain features of the real domain and simulated domain training samples until the confusion matrix and / or classification accuracy meet the preset conditions, and the training result is output as the fault detection model.

[0143] training a deep learning classification model based on the generated training data set, evaluating the diagnostic model performance through a confusion matrix and / or a classification accuracy, and outputting a training result as a fault detection model when the confusion matrix and / or the classification accuracy meet a preset condition, so as to obtain the fault detection model supporting adaptive diagnosis under variable rotating speed and variable load conditions.

[0144] Further, in one embodiment, in order to verify the generalization ability of the bearing fault detection method provided by the application, the Case Western Reserve University (CWRU) public bearing data set is used for experimental analysis.

[0145] The test platform of the embodiment includes an SKF-6205-2RS type deep groove ball bearing, the data set covers three types of healthy state (HEA), inner ring fault (IRF) and outer ring fault (ORF), the fault size is 0.53 mm, the experimental condition is 3HP / 1730r / min, the sampling frequency is 12 kHz, the single sample length is 4096 points, and 500 samples of each type are selected.

[0146] The experimental results of the comparative model are shown in Table 1.

[0147] Table 1 shows the experimental results of the comparative model.

[0148]

[0149] Table 1 shows that the comprehensive accuracy of FB-CycleGAN (the fault detection model of the application) on the CWRU data set is 97.07%, close to the training result of the measured data (97.20%), and significantly better than the traditional CycleGAN (89.40%) and WGAN (90.27%).

[0150] The ablation experiment shows that removing the FFT (Fast Fourier Transform) module causes the outer ring fault misjudgment rate to rise to 8.6%, and removing the Bi-LSTM (Bidirectional Long Short-Term Memory Network) module causes the inner ring fault misjudgment rate to increase to 11.6%, further verifying the necessity of frequency domain constraint and time series modeling.

[0151] The diagnostic precision confusion matrix of the embodiment is shown in Table 2. Figure 3 Table 2 shows the diagnostic precision confusion matrix. Figure 3 In Table 2, the horizontal axis is the predicted label, and the vertical axis is the true label. Figure 3 (a) is the diagnostic precision confusion matrix of the simulation signal, Figure 3 (b) is the diagnostic precision confusion matrix of the CycleGAN generated signal, Figure 3 (c) is the diagnostic precision confusion matrix of the FB-CycleGAN generated signal, Figure 3 (d) is the diagnostic precision confusion matrix of the real experimental signal.

[0152] It can be seen from the confusion matrix that the FB-CycleGAN (the fault detection model of the application) has an accuracy of 96.8%, 97% and 97.4% in the health, inner ring and outer ring fault classification, respectively, and no significant cross misjudgment occurs. Feature visualization analysis shows that the envelope spectrum of the generated signal is highly consistent with the measured signal in the fault characteristic frequency (such as BPFO = 114.13 Hz, BPFI = 179.3 Hz) and harmonic component, verifying the physical interpretability of the method.

[0153] The index conditions of the comparative model in the time domain root mean square error (RMSE) and the frequency domain maximum mean difference (MMD) are shown in Table 2.

[0154] Table 2 Comparison of diagnostic accuracy of different data training models for CWRU dataset

[0155]

[0156] As can be seen from Table 2, the enhanced data generated by the frequency domain bidirectional cycle generative adversarial network provided by the application reduces the time domain root mean square error (RMSE) and the frequency domain maximum mean difference (MMD) by 61.6% and 98.5%, respectively, compared with the original simulation data, and the fault classification accuracy is improved to 87.83% (self-built test bench) and 97.07% (CWRU dataset).

[0157] In summary, the bearing fault detection method provided by the application effectively solves the problem of bearing fault sample scarcity by fusing finite element simulation and frequency domain bidirectional CycleGAN. Two experiments show that the time-frequency characteristics of the generated signal are highly consistent with the measured data, which is significantly improved compared with traditional methods.

[0158] The bearing fault detection device provided by the application is described below. The bearing fault detection device described below can be correspondingly referred to the bearing fault detection method described above. Figure 4 is a structural schematic diagram of the bearing fault detection device provided by the application, as Figure 4 shown, the device comprises the following modules:

[0159] The data acquisition unit 410 is configured to acquire a bearing signal to be detected.

[0160] The fault detection unit 420 is configured to input the bearing signal to be detected into a pre-trained fault detection model to obtain a fault detection result.

[0161] The fault detection model is based on a neural network and is trained using a pre-constructed training data set. The training data set is input into a pre-trained frequency domain bidirectional cycle generative adversarial network to generate a measured signal sample and a simulation signal sample.

[0162] The bearing fault detection device provided by the application inputs the measured signal sample and the simulation signal sample into the pre-trained frequency domain bidirectional recurrent generative adversarial network to generate the training data set, and the training data set comprises:

[0163] A plurality of groups of measured signal samples and a plurality of groups of simulation signal samples are obtained.

[0164] The measured signal sample is converted into a measured domain frequency domain tensor, and the simulation signal sample is converted into a simulation domain frequency domain tensor.

[0165] The generator is constructed based on the bidirectional long short-term memory network, the discriminator is constructed based on the PatchGAN, and the initial recurrent generative adversarial network is obtained.

[0166] The initial recurrent generative adversarial network is trained by using the pre-constructed total loss function based on the measured domain frequency domain tensor and the simulation domain frequency domain tensor, and the frequency domain bidirectional recurrent generative adversarial network is obtained.

[0167] The measured domain frequency domain tensor and the simulation domain frequency domain tensor are input into the frequency domain bidirectional recurrent generative adversarial network to obtain a measured domain-simulation domain enhanced data tensor pair.

[0168] The measured domain-simulation domain enhanced data tensor pair is subjected to inverse fast Fourier transform to obtain a measured domain-simulation domain training sample pair, and then a training data set is obtained.

[0169] The simulation signal sample is obtained by inputting simulation parameters into a pre-constructed rigid-flexible coupling dynamics model.

[0170] The rigid-flexible coupling dynamics model is a finite element model constructed based on the multi-body dynamics theory, simulating the vibration response characteristics of the inner and outer ring faults of the bearing by using parameterized modeling, and introducing a noise injection mechanism to simulate the measured environmental interference.

[0171] The finite element model is a model obtained by discretizing the bearing by using a free tetrahedron mesh division algorithm and locally refining the contact area of the bearing by using a second-order gradient driven adaptive mesh division strategy.

[0172] The trigger condition of the adaptive mesh division strategy is:

[0173]

[0174] When η e >0.7, the size of the region unit is ensured to satisfy:

[0175]

[0176] wherein η e is the local rate of change of the stress field; is the second order gradient of von Mises stress; ‖·‖ e is the Euclidean norm of the variables within the element, representing the magnitude of the stress gradient within the element; h e is the size of the zone element; D eq is the equivalent diameter (unit: m), calculated by the projected area A e of the element.

[0177] The bearing fault detection device provided by the application comprises:

[0178]

[0179] wherein, is a total loss function, G A→B is a generator from a simulation domain to a measured domain; G B→A is a generator from a measured domain to a simulation domain; D A is a discriminator of the simulation domain; D B is a discriminator of the measured domain; is a first adversarial loss function; is a second adversarial loss function; is a cycle consistency loss function; λ cyc is a cycle consistency loss weight parameter; λ identity is an identity mapping loss weight parameter; is an identity mapping loss function; λ sp is a spectral constraint loss weight parameter; L sp (G A→B ) is a first spectral constraint loss function, L sp (G B→A ) is a second spectral constraint loss function.

[0180] The bearing fault detection device provided by the application is based on a neural network, and a training data set is used to train the fault detection model, and the bearing fault detection device comprises the following steps:

[0181] An initial classification model is constructed based on a neural network;

[0182] The initial classification model is trained according to the measured domain-simulation domain training sample pair in the training data set, so as to learn the time domain features of the training samples in the measured domain and the simulation domain, until the confusion matrix and / or the classification accuracy rate meet a preset condition, and the training result is output as the fault detection model.

[0183] The bearing fault detection device provided by the application generates a fault sample with fault characteristics and measured noise characteristics through a pre-trained frequency domain bidirectional cycle generative adversarial network, so as to obtain a training data set, and trains a fault detection model by using the training data set, thereby solving the problems of a lack of fault samples, a simulation-measured data difference and insufficient feature fidelity in the prior art, and improving the precision and generalization ability of bearing fault detection.

[0184] Figure 5 An example of a schematic diagram of a physical structure of an electronic device is shown in FIG. 1. Figure 5 As shown in FIG. 1, the electronic device can include a processor 510, a communications interface 520, a memory 530 and a communications bus 540, wherein the processor 510, the communications interface 520 and the memory 530 complete mutual communication through the communications bus 540. The processor 510 can invoke a logical instruction in the memory 530 to execute a bearing fault detection method, which includes: acquiring a bearing signal to be detected; inputting the bearing signal to be detected into a pre-trained fault detection model to obtain a fault detection result; wherein the fault detection model is based on a neural network and is trained by using a pre-constructed training data set, and the training data set is generated by inputting a measured signal sample and a simulation signal sample into a pre-trained frequency domain bidirectional cycle generative adversarial network.

[0185] In addition, the logical instruction in the memory 530 described above can be implemented in the form of a software function unit and sold or used as an independent product, and can be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk and various program codes that can be stored in the medium.

[0186] In another aspect, the present application also provides a computer program product comprising a computer program, which can be stored on a non-transitory computer-readable storage medium, and the computer program can be executed by a processor to enable a computer to perform the bearing fault detection method provided by the above methods, which comprises: obtaining a bearing signal to be detected; inputting the bearing signal to be detected into a pre-trained fault detection model to obtain a fault detection result; wherein the fault detection model is based on a neural network and is trained using a pre-constructed training data set, and the training data set is generated by inputting measured signal samples and simulated signal samples into a pre-trained frequency domain bidirectional cyclic generative adversarial network.

[0187] In another aspect, the present application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a bearing fault detection method provided by the above methods, which comprises: obtaining a bearing signal to be detected; inputting the bearing signal to be detected into a pre-trained fault detection model to obtain a fault detection result; wherein the fault detection model is based on a neural network and is trained using a pre-constructed training data set, and the training data set is generated by inputting measured signal samples and simulated signal samples into a pre-trained frequency domain bidirectional cyclic generative adversarial network.

[0188] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present embodiment scheme according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0189] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary universal hardware platforms, and of course, can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in terms of the contribution to the prior art, can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform the methods described in the various embodiments or some parts of the embodiments.

[0190] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit the same; and although the present application has been described in detail with reference to the foregoing embodiments, it should be appreciated by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features thereof can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A bearing fault detection method characterized by, The method comprises the following steps: acquiring a bearing signal to be detected; inputting the bearing signal to be detected into a pre-trained fault detection model to obtain a fault detection result; wherein the fault detection model is based on a neural network and is trained by using a pre-constructed training data set, and the training data set is generated by inputting measured signal samples and simulated signal samples into a pre-trained frequency domain bidirectional recurrent generative adversarial network.

2. The bearing fault detection method of claim 1, wherein, The training data set is generated by inputting the measured signal samples and the simulated signal samples into the pre-trained frequency domain bidirectional recurrent generative adversarial network, which comprises the following steps: acquiring a plurality of groups of measured signal samples and a plurality of groups of simulated signal samples; converting the measured signal samples into measured domain frequency domain tensors and converting the simulated signal samples into simulated domain frequency domain tensors; constructing a generator based on a bidirectional long short-term memory network and constructing a discriminator based on PatchGAN to obtain an initial recurrent generative adversarial network; training the initial recurrent generative adversarial network by using a pre-constructed total loss function according to the measured domain frequency domain tensors and the simulated domain frequency domain tensors to obtain the frequency domain bidirectional recurrent generative adversarial network; inputting the measured domain frequency domain tensors and the simulated domain frequency domain tensors into the frequency domain bidirectional recurrent generative adversarial network to obtain a measured domain-simulated domain enhanced data tensor pair; performing inverse fast Fourier transform on the measured domain-simulated domain enhanced data tensor pair to obtain a measured domain-simulated domain training sample pair, and further obtaining a training data set.

3. The bearing failure detection method according to claim 1 or 2, characterized by, The simulated signal samples are obtained by inputting simulation parameters into a pre-constructed rigid-flexible coupling dynamics model; wherein the rigid-flexible coupling dynamics model is a finite element model constructed based on the multi-body dynamics theory, simulating the vibration response characteristics of bearing inner and outer ring faults by parameterized modeling, and introducing a noise injection mechanism to simulate the measured environmental interference.

4. The bearing fault detection method of claim 3, wherein, The finite element model is a model obtained by discretizing the bearing through a free tetrahedron mesh division algorithm and locally refining the contact area of the bearing through a second-order gradient driven adaptive mesh division strategy; wherein the trigger condition of the adaptive mesh division strategy is: when η e > 0.7 ensures that the area cell size satisfies: where η e is the local rate of change of the stress field; is the second order gradient of the von Mises stress; ||·|| e is the Euclidean norm of the variables within the element, representing the magnitude of the stress gradient within the element; h e is the size of the zone element; D eq is the equivalent diameter (unit: m) calculated from the element projected area A e and the element volume V 5. The bearing fault detection method of claim 2, wherein, The total loss function comprises: wherein, G is the total loss function, A→B G is a generator from simulation domain to real domain; B→A D is a discriminator from real domain; A D is a discriminator from simulation domain; B D is a discriminator from real domain; L is a first adversarial loss function; L is a second adversarial loss function; L is a cycle consistency loss function; cyc L is a cycle consistency loss weight parameter; identity L is an identity mapping loss weight parameter; L is an identity mapping loss function; sp L is a spectral constraint loss weight parameter; sp (G A→B ) is a first spectral constraint loss function, sp (G B→A ) is a second spectral constraint loss function.

6. The bearing fault detection method of claim 2, wherein, training the fault detection model based on the neural network by using the pre-constructed training data set, which comprises the following steps: constructing an initial classification model based on the neural network; training the initial classification model according to the measured domain-simulated domain training sample pairs in the training data set to learn the time domain features of the measured domain and the simulated domain training samples until the confusion matrix and / or the classification accuracy rate meet the preset condition, and outputting the training result as the fault detection model.

7. A bearing fault detection apparatus characterized by, The method comprises the following steps: a data acquisition unit for acquiring a bearing signal to be detected; a fault detection unit for inputting the bearing signal to be detected into a pre-trained fault detection model to obtain a fault detection result; wherein the fault detection model is based on a neural network and is trained by using a pre-constructed training data set, and the training data set is generated by inputting measured signal samples and simulated signal samples into a pre-trained frequency domain bidirectional recurrent generative adversarial network.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The computer program is executed by the processor to implement the bearing fault detection method according to any one of claims 1 to 6. 9.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the bearing fault detection method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the bearing fault detection method according to any one of claims 1 to 6.

Images