Physical interpretable rolling bearing vibration data intelligent generation method and system

Abstract

The invention relates to a physically interpretable rolling bearing vibration data intelligent generation method and system, and the method comprises the steps: building an equivalent dynamical model corresponding to the motion process of a rolling bearing, obtaining the actual measurement data of small samples in a fault stage and a degradation stage, and taking the actual measurement data as the input of a multi-layer perceptron neural network based on multi-domain joint loss representation, calculating parameters similar to actual parameters required in the fault stage and parameter sequences in the degradation stage are obtained, and future development prediction is carried out on the obtained parameter sequences by using a gated circulation unit neural network and a single-step iteration prediction method in the degradation stage to obtain prediction sequences of the parameters; and inputting the calculation parameters obtained in the fault stage and the prediction sequence parameters obtained in the degradation stage into a dynamic expression to generate vibration data with physical interpretation. The vibration data generated by using the method is similar to actual data in time domain and frequency domain, and the data attributes of the vibration data and the actual data accord with each other, so that the method has clear physical significance.

Description

Technical Field

[0001] The present invention belongs to the field of rolling bearing data generation, and in particular relates to a method and system for intelligently generating physically interpretable rolling bearing vibration data. Background Art

[0002] As the core component of new quality productivity, data is the basis for pattern classification and prediction regression. Its sample diversity, sample completeness and noise interference determine the calculation results and quality of the model. Generative AI learns data structure and statistical features from the latent space of original data and generates new data by establishing a deep generative model, achieving data type balance, data sample expansion and data impurity filtering, and provides key technologies for large language model text generation, image data expansion, embodied data generation, and machine intelligence diagnosis and prediction.

[0003] Vibration data is an important source of mechanical fault diagnosis and degradation prediction. However, the diversity of fault modes, the randomness of degradation failures, and the time-consuming nature of vibration data collection have resulted in a relative lack of data, which is particularly evident in the following aspects: (1) The diversity of fault types leads to the non-uniformity of fault samples, and some fault data sets are small samples, which affects the generalization ability of intelligent diagnosis models. (2) Sudden failures lead to the scarcity of vibration data in the degradation stage. The intelligent prediction model cannot fully extract degradation features from small sample vibration data, which restricts the generalization ability and accuracy of life prediction. (3) High-end equipment such as aircraft engines have a low failure rate due to high reliability requirements, and full life sampling monitoring increases time, labor, and economic costs. Virtual data generated by generative artificial intelligence can balance non-uniform fault samples, expand degradation data, and avoid sampling cost waste, thus solving the above problems to a certain extent. However, the black box characteristics of deep generative models such as autoencoders and adversarial networks lead to the lack of clear physical meaning of the generated signals, images, and other data, and cannot explain the similarity between the attributes of the generated data and the attributes of the measured data. In addition, complex data structures and large, discrete search spaces lead to discrepancies between the latent data features learned by the data generation model and the actual data features. Adding domain knowledge guidance and constraints to the data generation model can make the data interpretable, narrow the search space of the data generation model, and improve the controllability and quality of the generated data.

[0004] In view of this, it is of great significance to design an intelligent generation method for rolling bearing vibration data that is physically interpretable. Summary of the invention

[0005] In order to solve the problems existing in the background technology, the present invention proposes a method for intelligently generating rolling bearing vibration data with physical interpretability, which overcomes the problems of non-uniformity of rolling bearing data in the failure or degradation stage, difficulty in collecting samples, waste of sampling resources, and poor physical interpretability. The data generation method can generate high-quality and highly interpretable data for rolling bearings in the failure or degradation stage.

[0006] In order to achieve the above-mentioned object, in a first aspect, the present invention provides a method for intelligently generating rolling bearing vibration data with physical interpretability, comprising the following steps:

[0007] Establish the rolling bearing dynamics model and its expression;

[0008] Obtain small sample failure stage data or degradation stage data;

[0009] The required data-driven model is designed according to different types of data, and the small sample fault stage data or degradation stage data is input into the data-driven model for training, so as to fit or predict the parameters of the rolling bearing dynamics model, and obtain the rolling bearing dynamics model and its expression after parameter fitting; if the subsequent fault data is generated by using the fault small sample measured data, the data-driven model is two multi-layer perceptron neural networks; if the subsequent degradation data is generated by using the degradation small sample measured data, the data-driven model is three multi-layer perceptron neural networks and three gated recurrent unit neural networks;

[0010] Based on the optimized rolling bearing dynamics model and its expression, fault data or degradation data with physical interpretability is generated.

[0011] Furthermore, when establishing the rolling bearing dynamics model and its expression, the operation process of the rolling bearing is simplified into a two-degree-of-freedom spring-oscillator equivalent model, and a dynamics model with respect to mass, stiffness and damping ratio parameters is constructed.

[0012] Furthermore, the dynamic model with respect to mass, stiffness and damping ratio parameters is:

[0013]

[0014] Where m is the mass of the rolling bearing, Z represents the number of rolling elements, c and K represent the damping ratio and stiffness of the equivalent spring-oscillator, respectively, and μ j represents the contact coefficient of the jth rolling element, F x and F y They represent the forces on the rolling bearing in the x and y directions respectively.

[0015] Furthermore, if the measured data of a small sample of faults is used to generate subsequent fault data, the data-driven model is two multi-layer perceptron neural networks, and the parameters of the rolling bearing dynamics model are fitted or predicted:

[0016] Two multi-layer perceptron neural network models are established, with X 1:n They are used as inputs of two multi-layer perceptron neural network models, and the outputs are defined as c′ and K′;

[0017] Substitute c′ and K′ into the rolling bearing dynamics model expression to generate vibration data, and use fast Fourier transform to perform frequency domain conversion on the generated vibration data to obtain frequency domain characteristics and obtain a frequency domain symmetric sequence. At the same time, use short-time Fourier transform to perform time-frequency domain conversion to obtain time-frequency domain characteristics.

[0018] Two multi-layer perceptron neural network models are trained using a multi-domain joint loss function;

[0019] Repeat the above steps to gradually fit the generated vibration data c′ and K′ to the actual damping ratio and stiffness of the equivalent spring-oscillator.

[0020] Furthermore, the boundary loss function is:

[0021]

[0022]

[0023] where c′ i , K′ i , m′ i c′ 1:n′ , K′ 1:n′ , m′ 1:n′ The i-th element in, ρ is m′ 1:n The magnification factor is m′0, and m′0 is the initial mass of the rolling bearing.

[0024] Furthermore, if the degradation small sample measured data is used to generate subsequent degradation data, the data-driven model is three multi-layer perceptron neural networks and three gated recurrent unit neural networks, and the parameters of the rolling bearing dynamics model are fitted or predicted:

[0025] S311, establish three multi-layer perceptron neural network models, with X′ 1:n′ As the input of three multi-layer perceptron neural network models, the output is defined as the time series c′ 1:n′ , K′ 1:n′ , m′ 1:n′ ;

[0026] S312, c′ 1:n′ , K′1:n′ , m′ 1:n′ Substituting into the rolling bearing dynamics model expression to obtain generated vibration data, performing fast Fourier transform on the generated vibration data to perform frequency domain conversion to obtain frequency domain characteristics, obtaining a frequency domain symmetric sequence, and simultaneously performing time-frequency domain conversion using short-time Fourier transform to obtain time-frequency domain characteristics;

[0027] S313 uses a multi-domain joint loss function to train two multi-layer perceptron neural network models;

[0028] S314, repeat S311-S313, so that the generated vibration data is gradually fitted to the actual equivalent spring-oscillator damping ratio, equivalent spring-oscillator stiffness, and rolling bearing mass time series, until c′ similar to the actual equivalent spring-oscillator damping ratio, equivalent spring-oscillator stiffness, and rolling bearing mass is obtained after the model training is completed. 1:n′ , K′ 1:n′ , m′ 1:n′ ;

[0029] S315, after the three multilayer perceptrons are trained, c′ 1:n′-1 , K′ 1:n′-1 , m′ 1:n′-1 Input three gated recurrent units for the input sequence, and the output is c′ n′ , K′ n′ , m′ n′ ;

[0030] S316, using an L2 loss function to train the gated recurrent unit until the model converges;

[0031] S317, using the single-step iterative prediction method to perform backward prediction to obtain a prediction sequence.

[0032] Furthermore, when generating fault data, the stiffness and damping ratio parameters after the multilayer perceptron neural network training are input into the rolling bearing dynamic model expression after parameter fitting; when generating degradation data, the multilayer perceptron neural network is used to fit the time-varying series of stiffness and damping ratio corresponding to the measured degradation small sample data, and the gated recurrent unit neural network based on the single-step iterative prediction method is used to predict the future mass, stiffness and damping ratio time series, and the prediction results are input into the rolling bearing dynamic model expression after parameter fitting to generate degradation stage vibration data with physical interpretation.

[0033] In a second aspect, the present invention provides a system for intelligently generating rolling bearing vibration data with physical interpretability, comprising a model building module, a data acquisition module, a model optimization module and a data generation module;

[0034] The model building module is used to establish the rolling bearing dynamics model and its expression;

[0035] The data acquisition module is used to acquire the failure stage data or degradation stage data of a small sample;

[0036] The model optimization module designs the required data-driven model according to different types of data, and inputs the small sample fault stage data or degradation stage data into the data-driven model for training, fits or predicts the parameters of the rolling bearing dynamics model, and obtains the rolling bearing dynamics model and its expression after parameter fitting; if the fault small sample measured data is used to generate subsequent fault data, the data-driven model is two multi-layer perceptron neural networks; if the degradation small sample measured data is used to generate subsequent degradation data, the data-driven model is three multi-layer perceptron neural networks and three gated recurrent unit neural networks;

[0037] The data generation module generates fault data or degradation data with physical interpretability based on the rolling bearing dynamics model and its expression after parameter fitting.

[0038] In a third aspect, the present invention may also provide a computer device, including a processor and a memory, the memory being used to store a computer executable program, the processor reading the computer executable program from the memory and executing it, and when the processor executes the computer executable program, it can implement the method for intelligently generating physically interpretable rolling bearing vibration data described in the present invention.

[0039] A computer-readable storage medium is also provided, in which a computer program is stored. When the computer program is executed by a processor, the method for intelligently generating physically interpretable rolling bearing vibration data of the present invention can be implemented.

[0040] Compared with the prior art, the method of the present invention has the following advantages: a multi-layer perceptron neural network model is used to fit the mass, stiffness and damping ratio, thereby avoiding the limitation of empirical parameter selection; when the multi-layer perceptron neural network model is trained, a multi-domain joint loss function is used, and the contribution of the time domain, frequency domain and time-frequency domain to the model training and the influence of the magnitude difference of stiffness and damping ratio on the model training are considered, thereby ensuring that the generated vibration data has similar properties to the measured data, and the generated data is the result of solving the dynamic equation and has a clear physical meaning. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of a model structure of a method for intelligently generating rolling bearing vibration data with physical interpretability provided by an embodiment of the present invention;

[0042] Figure 2A two-degree-of-freedom spring-oscillator model with a physically interpretable rolling bearing vibration data intelligent generation method provided in an embodiment of the present invention;

[0043] Figure 3 It is a schematic diagram of comparison between the vibration data generated by the invention and the data generated based on the empirical parameters in the CWRU data set in the fault stage for the inner and outer ring faults in the embodiment of the present invention;

[0044] Figure 4 is a predicted range of stiffness c′ obtained by using the present invention in the degradation stage of the bearing 1_1 in an embodiment of the present invention and a time-varying regularity diagram at 0-0.04s and 2399.96-2400s;

[0045] Figure 5 The predicted range of the damping ratio K′ obtained by using the present invention in the degradation stage of the bearing 1_1 in the embodiment of the present invention and the time-varying regularity diagram at 0-0.04s and 2399.96-2400s;

[0046] Figure 6 is a mass m′ prediction result obtained by using the invention in the degradation stage of the bearing 1_1 in an embodiment of the present invention, and a schematic diagram at 0-0.04s and 18.5-18.6min;

[0047] Figure 7 is a schematic diagram of comparing the local frequency spectra of the outer ring measured data and the vibration data generated by the two methods in the fault stage in an embodiment of the present invention;

[0048] Figure 8 is a schematic diagram of comparing the local frequency spectra of the inner ring measured data and the vibration data generated by the two methods in the fault stage in an embodiment of the present invention;

[0049] Fig. 9 is a schematic diagram comparing fault distribution of outer ring measured data in the fault stage and vibration data generated by two methods in an embodiment of the present invention;

[0050] Fig.10 is a schematic diagram comparing fault distribution of inner race measured data in a fault stage and vibration data generated using two methods in an embodiment of the present invention;

[0051] Fig.11 is a schematic diagram comparing the full life measured data of the bearing 1_1 in the embodiment of the present invention and the data generated by using the invention;

[0052] Fig.12 1 is a schematic diagram of the comparison of the local frequency spectrum of the full life measured data and the generated data of the bearing 1_1 data set in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] The structure of the present invention is described in more detail below in conjunction with the accompanying drawings:

[0055] like Figure 1 As shown, the present invention provides a method for intelligently generating rolling bearing vibration data with physical interpretability. First, a dynamic model of the rolling bearing is established. Secondly, measured data at the fault stage and the degradation stage are collected and input into the model. The model at the fault stage is a multi-layer perceptron neural network based on a multi-domain joint loss function, and the model at the degradation stage is a composition of a multi-layer perceptron neural network based on a multi-domain joint loss function and a gated recurrent unit neural network based on a single-step iterative prediction method.

[0056] The working principle of the present invention is described in more detail with reference to the accompanying drawings:

[0057] The present invention proposes a method for intelligently generating rolling bearing vibration data with physical interpretability, which specifically comprises the following steps:

[0058] Step S1, establishing a rolling bearing dynamics model and establishing a dynamics model expression;

[0059] Step S2, obtaining measured fault data or degradation data of a small sample;

[0060] Step S3, designing a required data-driven model according to different types of data, and inputting the data of step S2 into the data-driven model for training, so as to fit or predict the parameters of the rolling bearing dynamics model in step S1, and obtain the rolling bearing dynamics model after parameter fitting;

[0061] Step S4, based on the rolling bearing dynamics model obtained in step S3, generating fault data or degradation data with physical interpretability.

[0062] In step S1, the operation process of the rolling bearing is simplified into a two-degree-of-freedom spring-oscillator model, the structure of which is as follows: Figure 2 As shown. The expression of the dynamic model of the two-dimensional spring-oscillator equivalent rolling bearing with respect to mass, stiffness and damping ratio parameters is constructed. The process of constructing the expression is as follows:

[0063] Step S11, maximum displacement excitation ΔH at the rolling bearing defect: Taking the outer ring fault as an example, the length, width and height of the defect are L, B and H respectively, and the specific calculation method is as follows:

[0064]

[0065] Where D represents the rolling element diameter and ΔH′ represents the theoretical maximum displacement excitation.

[0066] Step S12: displacement excitation function h when the jth rolling element passes through the defect j : Assuming that the micro-topological shape of the defect surface is similar to the change trend of the half-sine excitation function, the specific calculation method is as follows:

[0067]

[0068] where mod(·) represents the remainder operation, and are the angles between the rolling element’s position when it enters the defect and when it exits the defect and the x-axis, θ j Represents the position of the jth rolling element.

[0069] Step S13, contact deformation δ between the jth rolling element and the raceway j Considering the influence of contact deformation between rolling element and raceway on displacement excitation during the operation of rolling bearing, the specific calculation method is as follows:

[0070] δ j =δ x cosθ j +δ y sinθ j -ε-h j (4)

[0071] Step S14, constructing a two-dimensional spring-oscillator equivalent rolling bearing dynamics model, the specific calculation method is as follows:

[0072]

[0073]

[0074] Where m is the mass of the rolling bearing, Z represents the number of rolling elements, c and K represent the damping ratio and stiffness of the equivalent spring-oscillator, respectively, and μ j represents the contact coefficient of the jth rolling element, F x and F y They represent the forces on the rolling bearing in the x and y directions respectively.

[0075] The small sample measured vibration data of the rolling bearing failure stage defined in step S2 is X 1:n ={X1,X2,…,X n} T , the small sample measured vibration data in the degradation stage is X′ 1:n′={X′1,X′2,…,X′ n′} T , where n and n′ are the lengths of small sample measured vibration data in the fault stage and degradation stage, respectively.

[0076] In step S3, if the input data is a small sample measured data at the fault stage, the data-driven model will generate the c and K values ​​of the equivalent dynamic model at the fault stage. The specific generation process is as follows:

[0077] Step S301, establish two multi-layer perceptron neural network models, with X 1:n As the input of two multilayer perceptron neural network models, the output is defined as c' and K'. The calculation process of each multilayer perceptron is as follows:

[0078] c′=AF(w Fc3 AF(w Fc2 AF(w Fc1 X 1:n +b Fc1 )+b Fc2 )+b Fc3 ) (7)

[0079] K′=AF(w FK3 AF(w FK2 AF(w FK1 X 1:n +b FK1 )+b FK2 )+b FK3 ) (8)

[0080] in, are trainable weights, is the learnable bias. AF(·) represents the activation function operation.

[0081] Step S302: Substitute c′ and K′ into formula (5) to generate vibration data, denoted as Y 1:n ={Y1,Y2,…,Y n}, use fast Fourier transform to perform frequency domain conversion, obtain frequency domain characteristics, and obtain frequency domain symmetric sequence, which are recorded as and At the same time, short-time Fourier transform is used to convert the time-frequency domain to obtain the time-frequency domain features, which are recorded as and

[0082] Step S303, construct a multi-domain joint loss function based on the time domain, frequency domain and time-frequency domain features of the measured signal and the generated signal to train the two multi-layer perceptron neural network models, and the expression is as follows:

[0083] Loss=Loss b +αLoss t +βLoss f +γLoss tf (9)

[0084] Loss b =Loss c +ηLoss K (10)

[0085]

[0086] Among them, Loss b is the boundary loss function, Loss t is the time domain loss function, Loss f is the frequency domain loss function, Loss tf is the time-frequency domain loss function, α, β, γ are weight coefficients, η is the balance coefficient, Y i and X i Y 1:n and X 1:n The i-th element in Y i f and They are and The i-th element in and They are and The i-th element in .

[0087] Step S303, repeating steps S301-S303 to generate vibration data Y 1:n Gradually fit to X 1:n , that is, c′ and K′ are gradually fitted to the actual damping ratio of the equivalent spring-oscillator and the stiffness of the equivalent spring-oscillator, until c′ and K′ similar to the actual damping ratio c and the stiffness K of the equivalent spring-oscillator are obtained after the model training is completed.

[0088] In step S3, if the input data is the measured data of a small sample in the degradation stage, the data-driven model will generate the c, K, and m time series of the equivalent dynamic model in the degradation stage. The specific generation process is as follows:

[0089] Step S311, establish three multi-layer perceptron neural network models, with X′ 1:n′ As the input of three multi-layer perceptron neural network models, the output is defined as the time series c′ 1:n′ , K′ 1:n′ , m′ 1:n′ , the calculation process of each multilayer perceptron is as follows:

[0090] c′ 1:n′ =AF(w Dc3 AF(w Dc2 AF(w Dc1 X 1:n +b Dc1 )+b Dc2 )+b Dc3 ) (16)

[0091] K′ 1:n′ =AF(w DK3 AF(w DK2 AF(w DK1 X 1:n +b DK1 )+b DK2 )+b DK3 ) (17)

[0092] m′ 1:n′ =AF(w Dm3 AF(w Dm2 AF(w Dm1 X 1:n +b Dm1 )+b Dm2 )+b Dm3 ) (18)

[0093] in are trainable weights,

[0094] is the learnable bias, R e (·) represents the ReLU activation function.

[0095] Step S312: c′ 1:n′ , K′ 1:n′ , m′ 1:n′ Substituting into formula (5) to obtain the generated vibration data, denoted as Y′ 1:n′ ={Y′1,Y′2,…,Y′ n′}Use fast Fourier transform to perform frequency domain conversion, obtain frequency domain characteristics, and obtain frequency domain symmetric sequence, which are recorded as and At the same time, short-time Fourier transform is used to convert the time-frequency domain to obtain the time-frequency domain features, which are recorded as and

[0096] S313, two multi-layer perceptron neural network models are trained using a multi-domain joint loss function, wherein the multi-domain joint loss function is modified to:

[0097]

[0098] where c′ i , K′ i , m′ i c′ 1:n′ , K′ 1:n′ , m′ 1:n′ The i-th element in, ρ is m′ 1:n The magnification factor, m′0 is the initial mass of the rolling bearing, Y′ i and X′ i Y′ 1:n′ and X′ 1:n′ The i-th element in and They are and The i-th element in and They are and The i-th element in .

[0099] Step S314, repeating steps S311-S313 to generate vibration data Y' 1:n′ Gradually fit to X′ 1:n′ , that is, c′ 1:n′ , K′ 1:n′ , m′ 1:n′ Gradually fit to the actual equivalent spring-oscillator damping ratio, equivalent spring-oscillator stiffness, and rolling bearing mass time series until c′ similar to the actual equivalent spring-oscillator damping ratio, equivalent spring-oscillator stiffness, and rolling bearing mass is obtained after the model training is completed. 1:n′ , K′ 1:n′ , m′ 1:n′ .

[0100] Step S315: After the training of the three multilayer perceptrons is completed, c′ 1:n′-1 , K′ 1:n′-1 , m′ 1:n′-1 Input three gated recurrent units for the input sequence, and the output is c′ n′ , K′ n′ , m′ n′ , the calculation process of each gated recurrent unit is as follows:

[0101]

[0102] in, are trainable weights, is the trainable bias, h c,n′ 、h K,n′ 、h m,n′ Calculate c′ n′ , K′ n′ , m′ n′ , tanh(·) represents the Tanh function, σ(·) represents the Sigmoid function, and ⊙ represents the dot product operation.

[0103] Step S316, using L2 loss function to train the gated recurrent unit to accelerate model convergence, the expression is as follows:

[0104] Loss = (c′ n′ -c n′ ) 2 +η 2 (K′ n′ -K n′ ) 2 +ρ 2 (m′ n′ -m n′ ) 2 (29)

[0105] Step S317, after the gated recurrent unit training is completed, it fits c′ 1:n′-1 , K′ 1:n′-1 , m′ 1:n′-1 The development trend mapping function is then used to predict backward using the single-step iterative prediction method. The specific calculation process is as follows:

[0106] Step S3171, with c' 2:n′ , K′ 2:n′ , m′ 2:n′ Input the input sequence into the three trained gated recurrent units and predict c′ n′+1 , K′ n′+1 , m′ n′+1 .

[0107] Step S3172, construct a new data set (c′ 2:n′ ,c′ n′+1 ),(K′ 2:n′ ,K′ n′+1 )、(m′ 2:n′ ,m′ n′+1 ) and substitute it into the trained gated recurrent unit to predict c′ n′+2 , K′ n′+2 , m′ n′+2 .

[0108] Step S3173, construct a new data set (c 3:n′ ,c′ n′+1 ,c′ n′+2)、(K 3:n′ ,K′ n′+1 ,K′ n′+2 )、(m 3:n′ ,m′ n′+1 ,m′ n′+2 ), repeat the iterative thinking of S3171-S3172, and finally get the predicted sequence c′ n′+1:n′+t , K′ n′+1:n′+t , m′ n′+1:n′+t . Where t is the length of vibration data corresponding to the complete failure of the rolling bearing. Then c′ 1:n′+t , K′ 1:n′+t , m′ 1:n′+t Generate c, K, and m time series of the equivalent dynamic model in the degradation stage for the data-driven model.

[0109] In step S4, c′ and K′ of the fitted fault stage or c′ predicted in the degradation stage are n′+1:n′+t , K′ n′+1:n′+t , m′ n′+1:n′+t Substituting into formula (5) can generate data of rolling bearing failure stage or degradation stage. Different from the existing methods, the present invention uses multi-layer perceptron and gated recurrent unit to fit and predict c, K, and m in the dynamic equation respectively, avoiding the limitation of empirical parameter selection. At the same time, a multi-domain joint loss function is proposed to iteratively update the parameters of the multi-layer perceptron, so that the differences between the generated data and the measured vibration data in the time domain, frequency domain and time-frequency domain are considered in the model training process, ensuring that the generated data is more consistent with the properties of the measured vibration data.

[0110] The present invention is further described below by taking the CWRU rolling bearing fault data set and the XJTU-SY rolling bearing full life data set as examples in conjunction with the accompanying drawings.

[0111] The present invention uses the CWRU rolling bearing fault data set to verify the vibration data generation of the rolling bearing fault stage, and uses the XJTU-SY rolling bearing full life data set to verify the vibration data generation of the rolling bearing degradation stage. When verifying the vibration data generation of the fault stage, the inner ring fault and outer ring fault of the rolling bearing under the 1750rpm working condition are selected as examples, and the rotation frequency is determined to be 29Hz, and the fault characteristic frequencies of the inner ring fault and the outer ring fault are 104.53Hz and 157.97Hz respectively. When verifying the vibration data generation of the degradation stage, the full life data set of bearing 1_1 is selected, and the failure frequency of the rolling bearing is determined to be 107.9Hz, and the rotation frequency is 35Hz.

[0112] Step 1, establishing a rolling bearing dynamics model and establishing a dynamics model expression;

[0113] In step 1, the operation process of the rolling bearing is simplified into a two-degree-of-freedom spring-oscillator model, and the expression of the dynamic model of the two-dimensional spring-oscillator equivalent rolling bearing with respect to mass, stiffness and damping ratio parameters is constructed, as shown in formula (5), which is the key to training the model and generating vibration data.

[0114] Step 2, obtain the measured fault data and degradation data of a small sample; the small sample measured data are taken from two stages, namely the fault stage and the degradation stage. In the fault stage, the CWRU rolling bearing fault data set is selected as the measured data. In the degradation stage, since the data set used is a full life data set, the 3-δ criterion is used to separate the healthy stage and the degradation stage in the full life data set of bearing 1_1, and the measured data of the healthy stage and the degradation stage are used as input data respectively, where in the degradation stage, the measured data length is taken as 3000.

[0115] Step 3, designing the required data-driven model according to different types of data, and substituting the data in step 2 into the data-driven model for training to fit or predict the parameters of the kinetic model in step 1;

[0116] In step 3, during the fault phase, two multilayer perceptron neural network models are established and trained using a multi-domain joint loss function to obtain c′=85, K′=8.51×10 8 , c′=50, K′=3.51×10 8 In the degradation stage, three multi-layer perceptron neural network models are established, and the models are trained using the multi-domain joint loss function to obtain c1′ :n , K1′ :n , m′ 1:n Time series.

[0117] Step 4, substitute the parameters fitted or predicted in step 3 into the dynamic model of step 1 to generate fault data or degradation data with physical interpretability.

[0118] In step 4, during the fault stage, there is no need to predict the parameters output after the model training is completed. The obtained fitting parameters are directly substituted into formula (5) to generate the vibration data of the fault stage, such as Figure 3 In the degradation stage, it is necessary to predict the future trend of the parameters. Therefore, it is necessary to introduce a gated recurrent unit neural network based on a single-step iterative prediction method into the model to predict the subsequent development trend of c′, K′, and m′. The prediction result is input into the expression in step 1 to generate the vibration data of the degradation stage.

[0119] like Figure 4Shown are the predicted range of c′ in bearing 1_1 and the time-varying regularity diagram at 0-0.04s and 2399.96-2400s.

[0120] like Figure 5 Shown are the predicted range of K′ in bearing 1_1 and its time-varying regularity at 0-0.04s and 2399.96-2400s.

[0121] like Figure 6 Shown are the predicted results of m′ in bearing 1_1 and the time-varying law diagram at 18.5-18.6min.

[0122] In order to verify the effectiveness of the method described in the present invention, a comparative description is given based on the vibration data generated by the present invention as described above.

[0123] To verify the effectiveness of this method at the fault stage, the following methods are adopted:

[0124] Select c = 300, K = 1.3 × 10 10 As an empirical parameter, it is substituted into formula (5) to generate vibration data based on the empirical parameter, and compared with the vibration data generated by using the parameters obtained in step 3 at the fault stage.

[0125] To verify the effectiveness of this method in the degradation stage, the following methods are adopted:

[0126] Since the data set is a full-life data set, the full-life generated data is compared with the measured full-life vibration data. At this time, the full-life generated data only includes the vibration data of the degradation stage obtained in step 4, so it is necessary to use the measured data of the healthy stage obtained in step 2, and use the method of generating vibration data in the fault stage of this invention to obtain the generated data in the healthy stage, and then splice it with the data generated in the degradation stage to obtain the full-life generated data for comparison.

[0127] Depend on Figure 3 It can be seen that the vibration data generated in the fault stage using this invention is highly consistent with the data generated based on empirical parameters in the time domain.

[0128] Depend on Figure 7 and Figure 8 It can be seen that after fast Fourier transform is performed on the vibration data generated by this invention at the fault stage, the generated data based on empirical parameters, and the measured data, it can be obviously seen that there is a large difference between the generated data and the measured data in the frequency domain, and the vibration data generated by this invention at the fault stage is closer to the frequency spectrum of the measured data.

[0129] Depend on Fig. 9 and Fig.10It can be seen that after performing Hilbert transform and fast Fourier transform on the vibration data generated by the invention at the fault stage, the generated data based on empirical parameters and the measured data, from the envelope diagrams of the three about the inner and outer circles and the fault frequency distribution, it can be seen that the three can clearly see the outer ring fault frequency of 104Hz and its multiples and the inner ring fault frequency of 158Hz and its multiples. However, the vibration data generated by the invention at the fault stage, whether it is an inner ring fault or an outer ring fault, has a larger amplitude at the fault frequency and its multiples, a higher energy, and a more obvious change in the multiples.

[0130] Fig.11 The following is a comparison between the measured data and the generated data of the full life of bearing 1_1. It can be seen that the development trends of the two are the same; Fig.12 It can be seen that the spectrum of the data set of bearing 1_1 has a fault frequency and rotation frequency close to the measured data. This phenomenon also proves that it is consistent with the spectrum evolution results of the rolling bearing degradation stage.

[0131] Based on the concept of the method, the present invention can provide a physically interpretable rolling bearing vibration data intelligent generation system, including a model building module, a data acquisition module, a model optimization module and a data generation module;

[0132] The model building module is used to establish the rolling bearing dynamics model and its expression;

[0133] The data acquisition module is used to acquire the failure stage data or degradation stage data of a small sample;

[0134] The model optimization module designs the required data-driven model according to different types of data, and inputs the small sample fault stage data or degradation stage data into the data-driven model for training, fits or predicts the parameters of the rolling bearing dynamics model, and obtains the rolling bearing dynamics model and its expression after parameter fitting; if the fault small sample measured data is used to generate subsequent fault data, the data-driven model is two multi-layer perceptron neural networks; if the degradation small sample measured data is used to generate subsequent degradation data, the data-driven model is three multi-layer perceptron neural networks and three gated recurrent unit neural networks;

[0135] The data generation module generates fault data or degradation data with physical interpretability based on the rolling bearing dynamics model and its expression after parameter fitting.

[0136] On the other hand, the present invention provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the method for intelligently generating physically interpretable rolling bearing vibration data described in the present invention can be implemented.

[0137] The computer device may be a laptop computer, a desktop computer or a workstation.

[0138] The present invention can also provide a computer device, including a processor and a memory, the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and when the processor executes the computer executable program, it can implement the method for intelligently generating physically interpretable rolling bearing vibration data described in the present invention.

[0139] The processor may be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC), or an off-the-shelf field programmable gate array (FPGA).

[0140] The memory of the present invention may be an internal storage unit of a laptop computer, a desktop computer or a workstation, such as a memory or a hard disk; or an external storage unit, such as a mobile hard disk or a flash memory card.

[0141] Computer-readable storage media may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules or other data. Computer-readable storage media may include: read-only memory (ROM), random access memory (RAM), solid-state drive (SSD) or optical disk, etc. Among them, random access memory may include resistance random access memory (ReRAM) and dynamic random access memory (DRAM).

[0142] In summary, the present invention relates to an intelligent method for generating rolling bearing vibration data based on multi-domain joint loss representation and physical interpretability. First, a dynamic model is constructed for the motion process of the rolling bearing to obtain the relationship between mass, stiffness and damping ratio. Second, the measured data of small samples in the fault stage and degradation stage are obtained as the input of the multi-layer perceptron neural network based on multi-domain joint loss representation to obtain the calculation parameters similar to the actual parameters required in the fault stage and the parameter sequences in the degradation stage. At the same time, in the degradation stage, the gated recurrent unit neural network and the single-step iterative prediction method are used to predict the future development of the obtained parameter sequences to obtain the predicted sequences of the parameters. Finally, the calculation parameters obtained in the fault stage and the predicted sequence parameters obtained in the degradation stage are input into the dynamic expression to generate vibration data with more physical interpretability. When training the multi-layer perceptron neural network model, a multi-domain joint loss function is used, and the contribution of the time domain, frequency domain and time-frequency domain to the model and the influence of the order of magnitude difference of stiffness and damping ratio on the model are considered, so as to ensure that the generated vibration data has similar properties to the measured data and is physically interpretable; a multi-layer perceptron neural network model and a gated recurrent unit neural network model based on a single-step iterative prediction method are used to fit and predict the parameters, avoiding the limitations of empirical parameter selection, and ensuring that the generated vibration data has clear physical meaning and clearer interpretability; based on the above results and analysis, it is shown that the present invention can indeed ensure the generation of more physically interpretable vibration data in the fault stage and the degradation stage.

[0143] The above contents are only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A method for intelligently generating physically interpretable rolling bearing vibration data, characterized in that: The following steps are involved: Establish the rolling bearing dynamics model and its expression; Obtain failure stage data or degradation stage data of a small sample; The required data-driven model is designed according to different types of data, and the small sample fault stage data or degradation stage data is input into the data-driven model for training, so as to fit or predict the parameters of the rolling bearing dynamics model, and obtain the rolling bearing dynamics model and its expression after parameter fitting; if the subsequent fault data is generated by using the fault small sample measured data, the data-driven model is two multi-layer perceptron neural networks; if the subsequent degradation data is generated by using the degradation small sample measured data, the data-driven model is three multi-layer perceptron neural networks and three gated recurrent unit neural networks; Based on the rolling bearing dynamics model and its expression after parameter fitting, fault data or degradation data with physical interpretability are generated.

2. The method for intelligently generating rolling bearing vibration data with physical interpretability according to claim 1, characterized in that: When establishing the rolling bearing dynamics model and its expression, the operation process of the rolling bearing is simplified into a two-degree-of-freedom spring-oscillator equivalent model, and a dynamics model with respect to mass, stiffness and damping ratio parameters is constructed.

3. The method for intelligently generating physically interpretable rolling bearing vibration data according to claim 2, characterized in that: The dynamic model with respect to mass, stiffness and damping ratio parameters is: Where m is the mass of the rolling bearing, Z represents the number of rolling elements, c and K represent the damping ratio and stiffness of the equivalent spring-oscillator, respectively, and μ j represents the contact coefficient of the jth rolling element, F x and F y They represent the forces on the rolling bearing in the x and y directions respectively.

4. The method for intelligently generating rolling bearing vibration data with physical interpretability according to claim 3 is characterized in that: If the measured data of a small sample of faults is used to generate subsequent fault data, the data-driven model is two multi-layer perceptron neural networks, and the parameters of the rolling bearing dynamics model that are fitted or predicted include: Two multi-layer perceptron neural network models are established, with X 1:n They are used as inputs of two multi-layer perceptron neural network models, and the outputs are defined as c′ and K′; Substitute c′ and K′ into the rolling bearing dynamics model expression to generate vibration data, and use fast Fourier transform to perform frequency domain conversion on the generated vibration data to obtain frequency domain characteristics and obtain a frequency domain symmetric sequence. At the same time, use short-time Fourier transform to perform time-frequency domain conversion to obtain time-frequency domain characteristics. Based on the time domain, frequency domain and time-frequency domain features of the measured and generated signals, a multi-domain joint loss function is constructed to train two multi-layer perceptron neural network models; Repeat the above steps to gradually fit the generated vibration data c′ and K′ to the actual damping ratio and stiffness of the equivalent spring-oscillator.

5. The method for intelligently generating rolling bearing vibration data with physical interpretability according to claim 4, characterized in that: The boundary loss function is: where c′ i , K′ i , m′ i c′ 1:n′ , K′ 1:n′ , m′ 1:n′ The i-th element in, ρ is m′ 1:n The magnification factor is m′0, and m′0 is the initial mass of the rolling bearing.

6. The method for intelligently generating physically interpretable rolling bearing vibration data according to claim 3, characterized in that: If the degradation small sample measured data is used to generate subsequent degradation data, the data-driven model is three multi-layer perceptron neural networks and three gated recurrent unit neural networks. The parameters of the rolling bearing dynamics model that are fitted or predicted include: S311, establish three multi-layer perceptron neural network models, with X1′ :n′ As the input of three multi-layer perceptron neural network models, the output is defined as the time series c1′ :n′ , K1′ :n′ , m′ 1:n′ ; S312, c1′ :n′ , K1′ :n′ , m′ 1:n′ Substituting into the rolling bearing dynamics model expression to obtain generated vibration data, performing fast Fourier transform on the generated vibration data to perform frequency domain conversion to obtain frequency domain characteristics, obtaining a frequency domain symmetric sequence, and simultaneously performing time-frequency domain conversion using short-time Fourier transform to obtain time-frequency domain characteristics; S313 uses a multi-domain joint loss function to train two multi-layer perceptron neural network models; S314, repeat S311-S313, so that the generated vibration data is gradually fitted to the actual equivalent spring-oscillator damping ratio, equivalent spring-oscillator stiffness, and rolling bearing mass time series, until c1′ similar to the actual equivalent spring-oscillator damping ratio, equivalent spring-oscillator stiffness, and rolling bearing mass is obtained after the model training is completed. :n′ , K1′ :n′ , m′ 1:n′ ; S315, after the three multilayer perceptrons are trained, c1′ :n′-1 , K1′ :n′-1 , m′ 1:n′-1 Input three gated recurrent units for the input sequence, and the output is c′ n′ , K″, m′ n′ ; S316, using an L2 loss function to train the gated recurrent unit until the model converges; S317, using the single-step iterative prediction method to perform backward prediction to obtain a prediction sequence.

7. The method for intelligently generating physically interpretable rolling bearing vibration data according to claim 1, characterized in that: When generating fault data, the stiffness and damping ratio parameters after the multilayer perceptron neural network training are input into the rolling bearing dynamic model expression after parameter fitting; when generating degradation data, the multilayer perceptron neural network is used to fit the time-varying series of stiffness and damping ratio corresponding to the measured degradation small sample data, and the gated recurrent unit neural network based on the single-step iterative prediction method is used to predict the future mass, stiffness and damping ratio time series, and the prediction results are input into the rolling bearing dynamic model expression after parameter fitting to generate vibration data in the degradation stage with physical interpretation.

8. A system for intelligently generating rolling bearing vibration data with physical interpretability, characterized in that: It includes a model building module, a data acquisition module, a model optimization module and a data generation module; The model building module is used to establish the rolling bearing dynamics model and its expression; The data acquisition module is used to acquire the failure stage data or degradation stage data of a small sample; The model optimization module designs the required data-driven model according to different types of data, and inputs the small sample fault stage data or degradation stage data into the data-driven model for training, fits or predicts the parameters of the rolling bearing dynamics model, and obtains the rolling bearing dynamics model and its expression after parameter fitting; if the fault small sample measured data is used to generate subsequent fault data, the data-driven model is two multi-layer perceptron neural networks; if the degradation small sample measured data is used to generate subsequent degradation data, the data-driven model is three multi-layer perceptron neural networks and three gated recurrent unit neural networks; The data generation module generates fault data or degradation data with physical interpretability based on the rolling bearing dynamics model and its expression after parameter fitting.

9. A computer device, characterized in that: It includes a processor and a memory, the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and when the processor executes the computer executable program, it can implement the intelligent generation method of rolling bearing vibration data with physical interpretability as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A computer program is stored in the computer-readable storage medium. When the computer program is executed by a processor, the method for intelligently generating rolling bearing vibration data with physical interpretability as described in any one of claims 1 to 7 can be implemented.

Images