Small sample life prediction model training method based on multi-channel source domain transfer learning

Abstract

The invention discloses a small sample life prediction model training method based on multi-channel source domain transfer learning, and relates to the technical field of rotary machinery predictive maintenance. The method comprises the following steps: constructing a five-degree-of-freedom differential equation of the rotating machine; based on the five-degree-of-freedom differential equation, constructing a target nonlinear dynamic model of the rotating machine gradually from a health stage and a degradation stage to a damage stage; performing simulation analysis on the target nonlinear dynamic model to obtain source domain data of a plurality of channels; according to the source domain data of the multiple channels and the target domain data of the rotating machine, a preset life prediction model of the rotating machine is trained, and the preset life prediction model comprises a common feature extractor, a domain specific feature extractor, a domain specific regression device and a domain specific discriminator. According to the method, rich source domain data can be provided for the model when the training samples are scarce, the prediction precision of the model is improved, meanwhile, the difference between the source domain data of different channels can be processed, and the generalization ability of the model is ensured.

Description

Technical Field

[0001] The present application relates to the technical field of predictive maintenance of rotating machinery, and in particular to a small sample life prediction model training method based on multi-channel source domain transfer learning. Background Art

[0002] Industrial equipment is increasingly becoming large-scale, automated, and intelligent. Rotating machinery such as bearings and gears are widely used in aerospace, transportation, and other fields, and their operating conditions are directly related to the life of various types of mechanical equipment. Therefore, life prediction of rotating machinery such as bearings has positive practical significance.

[0003] At present, data measurement is usually performed when rotating machinery is working, and the actual measured data is used as training samples. However, due to the complex working environment of rotating machinery, it is difficult to obtain a large number of training samples, so only a small number of samples or even no samples can be used for training, and the prediction accuracy of the model cannot be guaranteed. Summary of the invention

[0004] In view of this, the present application provides a small sample life prediction model training method based on multi-channel source domain transfer learning, which mainly aims to provide the model with rich source domain data when training samples are scarce, thereby improving the prediction accuracy of the model, and at the same time being able to handle the differences between source domain data of different channels to ensure the generalization ability of the model.

[0005] According to a first aspect of the present application, a small sample life prediction model training method based on multi-channel source domain transfer learning is provided, the method comprising:

[0006] Construct the five-degree-of-freedom differential equations for rotating machinery;

[0007] Based on the five-degree-of-freedom differential equation, a target nonlinear dynamic model of the rotating machinery gradually progressing from a healthy stage, a degraded stage to a damaged stage is constructed;

[0008] Performing simulation analysis on the target nonlinear dynamics model to obtain source domain data of multiple channels;

[0009] Based on the source domain data of the multiple channels and the target domain data of the rotating machinery, a preset life prediction model of the rotating machinery is trained, wherein the preset life prediction model includes a common feature extractor corresponding to the source domain data of the multiple channels, and domain-specific feature extractors, domain-specific regressors and domain-specific discriminators corresponding to the source domain data of the multiple channels respectively.

[0010] According to a second aspect of the present application, a small sample life prediction model training device based on multi-channel source domain transfer learning is provided, the device comprising:

[0011] Construction unit for constructing five-degree-of-freedom differential equations for rotating machinery;

[0012] The construction unit is further used to construct a target nonlinear dynamic model of the rotating machinery gradually from a healthy stage, a degraded stage to a damaged stage based on the five-degree-of-freedom differential equation;

[0013] A simulation unit, used for performing simulation analysis on the target nonlinear dynamics model to obtain source domain data of multiple channels;

[0014] A training unit is used to train a preset life prediction model of the rotating machinery based on the source domain data of the multiple channels and the target domain data of the rotating machinery, wherein the preset life prediction model includes a common feature extractor corresponding to the source domain data of the multiple channels, and domain-specific feature extractors, domain-specific regressors and domain-specific discriminators corresponding to the source domain data of the multiple channels respectively.

[0015] According to a third aspect of the present application, a storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the small sample life prediction model training method based on multi-channel source domain transfer learning is implemented.

[0016] According to the fourth aspect of the present application, an electronic device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein when the processor executes the program, the small sample life prediction model training method based on multi-channel source domain transfer learning is implemented.

[0017] By means of the above technical solution, the present application provides a small sample life prediction model training method based on multi-channel source domain transfer learning. By constructing a target nonlinear dynamic model of a rotating machinery gradually from a healthy stage, a degradation stage to a damaged stage, and performing simulation analysis on the target nonlinear dynamic model, a large amount of simulation data covering the entire cycle and different channel modes can be generated, thereby obtaining a large amount of source domain data of various fault types, and then providing the model with rich source domain data when training samples are scarce, thereby improving the prediction accuracy of the model. In addition, in the process of transfer learning, the present application can align the source domain data and target domain data of different channels to different feature spaces by training a common feature extractor corresponding to the source domain data of multiple channels, and a domain-specific feature extractor, a domain-specific regressor, and a domain-specific discriminator corresponding to the source domain data of multiple channels respectively, thereby being able to handle the differences between the source domain data of different channels and improve the generalization ability of the life prediction model.

[0018] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0020] Figure 1 A schematic flow chart of a method for training a small sample life prediction model based on multi-channel source domain transfer learning provided in an embodiment of the present application is shown;

[0021] Figure 2 A schematic diagram of a five-degree-of-freedom dynamic model provided in an embodiment of the present application is shown;

[0022] Figure 3 A schematic diagram of a process for constructing a nonlinear dynamics model provided in an embodiment of the present application is shown;

[0023] Figure 4 A schematic diagram of the process of the preset life prediction model training method provided in an embodiment of the present application is shown;

[0024] Figure 5 A schematic diagram of the process of source domain data generation, model training, and model testing provided by an embodiment of the present application is shown;

[0025] Figure 6 A structural schematic diagram of a small sample life prediction model training device based on multi-channel source domain transfer learning provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0026] The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0027] Since the working environment of rotating machinery is relatively complex, it is difficult to obtain a large number of training samples, resulting in only a small number of samples or even no samples available for training, and the prediction accuracy of the model cannot be guaranteed.

[0028] In order to solve the above problems, the embodiment of the present invention provides a small sample life prediction model training method based on multi-channel source domain transfer learning, such as Figure 1 As shown, the method includes:

[0029] Step 10: Construct the five-degree-of-freedom differential equation of the rotating machinery.

[0030] The rotating machinery may specifically include gears, engines, bearings, etc. The bearings include a plurality of rolling elements, a bearing inner ring, and a bearing outer ring.

[0031] The embodiments of the present invention are mainly applicable to health monitoring of rotating machinery, mainly involving key equipment such as aircraft engines, railway vehicles and wind turbines.

[0032] The embodiment of the present invention now takes a bearing as an example to explain in detail the process of constructing a five-degree-of-freedom differential equation and a target nonlinear dynamic model, but is not limited to bearings.

[0033] Specifically, the five-degree-of-freedom differential equation of the bearing is constructed as follows,

[0034]

[0035] Where m, k and c represent mass, stiffness and damping respectively, x and y represent horizontal vibration and vertical vibration respectively, subscripts s, p and r represent the bearing inner ring, bearing outer ring and unit resonator respectively, F s represents the external radial load, f x and f y are the horizontal nonlinear contact force and vertical nonlinear contact force applied to the inner ring and outer ring of the bearing, and g is the gravitational acceleration.

[0036] Step 20: Based on the five-degree-of-freedom differential equation, a target nonlinear dynamic model of the rotating machinery gradually progressing from a healthy stage, a degraded stage to a damaged stage is constructed.

[0037] The target nonlinear dynamic model refers to the nonlinear dynamic model of the final damage stage.

[0038] In order to simulate the vibration behavior of the bearing during its entire service life, the embodiment of the present invention constructs a five-degree-of-freedom dynamic model for different degradation modes, including the inner ring of the bearing, the outer ring of the bearing and the rolling element, such as Figure 2 As shown, the five-degree-of-freedom dynamic model includes a nonlinear dynamic model in the healthy stage, a nonlinear dynamic model in the degradation stage, and a nonlinear dynamic model in the injury stage.

[0039] For the specific process of building nonlinear dynamic models, such as Figure 3 As shown, including:

[0040] Step 21: construct a first contact deformation formula between any one of the plurality of rolling elements and the inner ring of the bearing and the outer ring of the bearing in the healthy stage.

[0041] For the embodiment of the present invention, when the bearing is running in a healthy stage, there is no damage on the bearing inner ring, the bearing outer ring and the rolling element. Taking any rolling element and the bearing inner ring and the bearing outer ring in contact therewith as the research objects, the first contact deformation formula for expressing the corresponding deformation is as follows:

[0042] δ j =(x s -x p )cosθ j +(y s -y p )sinθ j +c

[0043] Among them, θ j is the angular position of the jth rolling element, δ j is the deformation in the healthy stage, and c is the radial clearance of the bearing.

[0044] Step 22: Substitute the first contact deformation formula into the five-degree-of-freedom differential equation to construct a nonlinear dynamic model of the bearing in the healthy stage.

[0045] For the embodiment of the present invention, by substituting the first contact deformation formula into the above five-degree-of-freedom differential equation, a nonlinear dynamic model in the healthy stage (normal working condition) can be obtained.

[0046] Step 23: Perform deformation adjustment on the nonlinear dynamic model in the healthy stage to obtain the nonlinear dynamic model of the bearing in the degraded stage.

[0047] For the embodiment of the present invention, for the degradation stage, based on the first contact deformation formula, a second contact deformation formula between any one of the rolling elements and the inner ring of the bearing and the outer ring of the bearing in the case of a bearing outer ring defect, a third contact deformation formula between any one of the rolling elements and the inner ring of the bearing and the outer ring of the bearing in the case of a bearing inner ring defect, and a fourth contact deformation formula between any one of the rolling elements and the inner ring of the bearing and the outer ring of the bearing in the case of a rolling element defect are respectively constructed; based on the second contact deformation formula, the third contact deformation formula and the fourth contact deformation formula, the nonlinear dynamic model of the healthy stage is deformed and adjusted to obtain the nonlinear dynamic model of the bearing in the degradation stage.

[0048] Specifically, in the degradation stage, in order to simulate the impact of damage to different parts on the full-cycle degradation of the bearing, the embodiment of the present invention simulates three fault environments and divides the bearing degradation into the degradation processes of the inner ring of the bearing, the outer ring of the bearing and the rolling element. The defects in the degradation stage are caused by surface dents. When the rolling element passes through the defect, the change in surface morphology will generate excitation, which is reflected in the corresponding vibration signal.

[0049] For inner ring defects, the second contact deformation formula used to express the contact deformation between the rolling element and the raceway is as follows:

[0050] δ j =(x s -x p )cosθ j +(y s -y p )sinθ j +c+δ 0

[0051]

[0052] Among them, δ j is the deformation amount of the inner ring defect in the degradation stage, δ 0 is the deformation adjustment amount for the inner ring defect of the bearing in the degradation stage, d is the rolling element diameter, L in is the width of the inner ring defect.

[0053] For outer ring defects, the third contact deformation formula used to express the contact deformation between the rolling element and the raceway is as follows:

[0054] δ j =(x s -x p )cosθ j +(y s -y p )sinθ j +c-δ 0

[0055]

[0056] Among them, δ j is the deformation of the outer ring defect in the degradation stage, δ 0 is the deformation adjustment amount for the bearing outer ring defects in the degradation stage, d is the rolling element diameter, L out is the width of the outer ring defect.

[0057] For rolling element defects, the fourth contact deformation formula used to express the contact deformation between the rolling element and the raceway is as follows:

[0058] δ j =(x s -x p )cosθ j +(y s -y p )sinθ j +c+δ 0

[0059] δ 0 =h ball

[0060] Among them, δ j is the deformation of the rolling element in the degradation stage, δ 0 is the deformation adjustment amount for rolling element defects in the degradation stage, d is the rolling element diameter, h is ball is the width of the rolling element defect.

[0061] After the second contact deformation formula, the third contact deformation formula and the fourth contact deformation formula are obtained, the deformation variables in the nonlinear dynamic model of the healthy stage are replaced or adjusted by them to obtain the nonlinear dynamic model of the degraded stage.

[0062] Step 24: Perform deformation adjustment on the nonlinear dynamic model in the degradation stage to obtain the nonlinear dynamic model of the bearing in the damage stage.

[0063] For the embodiment of the present invention, for the damage stage, based on the second contact deformation formula, the third contact deformation formula and the fourth contact deformation formula, the fifth contact deformation formula between any one of the rolling elements and the inner ring of the bearing and the outer ring of the bearing in the case of a bearing outer ring defect, the sixth contact deformation formula between any one of the rolling elements and the inner ring of the bearing and the outer ring of the bearing in the case of a bearing inner ring defect, and the seventh contact deformation formula between any one of the rolling elements and the inner ring of the bearing and the outer ring of the bearing in the case of a rolling element defect are respectively constructed; based on the fifth contact deformation formula, the sixth contact deformation formula and the seventh contact deformation formula, the nonlinear dynamic model of the degradation stage is deformed and adjusted to obtain the nonlinear dynamic model of the bearing in the damage stage.

[0064] Specifically, in the damage stage, the defect size increases, and the vibration excitation generated is more intense. Specifically, it can be simplified as a rectangular defect. As the defect size increases, the rolling element can contact the bottom of the defect.

[0065] For inner ring damage, the fifth contact deformation formula used to express the contact deformation between the rolling element and the raceway is as follows:

[0066] δ j =(x s -x p )cosθ j +(y s -y p )sinθ j +c-δ 0

[0067]

[0068] Among them, δ j is the deformation of the inner ring at the damage stage, δ 0 is the deformation adjustment amount for bearing inner ring damage in the damage stage, ΔΦ 4 h is the edge angle from when the rolling element contacts the edge of the defect to when it completely falls into the bottom of the defect. in It is the width of the rectangular defect produced on the inner track.

[0069] For outer ring damage, the sixth contact deformation formula used to express the contact deformation between the rolling element and the raceway is as follows:

[0070] δ j =(x s -x p )cosθ j +(y s -y p )sinθ j +c+δ 0

[0071]

[0072] Among them, δ j is the deformation of the outer ring in the damage stage, δ 0 h is the deformation adjustment amount for bearing outer ring damage in the damage stage, out It is the width of the rectangular defect produced on the outer ring track.

[0073] For rolling element damage, the seventh contact deformation formula used to express the contact deformation between the rolling element and the raceway is as follows:

[0074] δ j =(x s -x p )cosθ j +(y s -y p )sinθ j +c+δ 0

[0075] δ 0 =h ball

[0076] Among them, δ j is the deformation of the rolling element in the damage stage, δ 0 h is the deformation adjustment for rolling element damage in the damage stage, ball is the width of the rolling element defect.

[0077] After the fifth contact deformation formula, the sixth contact deformation formula and the seventh contact deformation formula are obtained, the deformation variables in the nonlinear dynamic model of the degradation stage are replaced or adjusted by them to obtain the nonlinear dynamic model of the damage stage.

[0078] Step 25: Determine the nonlinear dynamic model of the damage stage as the target nonlinear dynamic model.

[0079] The embodiment of the present invention simulates various complex bearing degradation processes by establishing nonlinear dynamic models under different degradation modes (such as inner race fault, outer race fault and rolling element fault). Compared with the simplified assumption of only considering a single degradation mode, the embodiment of the present invention can generate a large amount of simulation data covering the entire cycle and different channel modes, thereby providing rich source domain data for the model when training samples are scarce.

[0080] Step 30: Perform simulation analysis on the target nonlinear dynamics model to obtain source domain data of multiple channels.

[0081] The source domain data of the multiple channels include source domain data of bearing inner race faults, source domain data of bearing outer race faults and source domain data of rolling element faults.

[0082] For the embodiment of the present invention, in order to obtain source domain data of multiple channels, step 103 specifically includes: solving the equation corresponding to the target nonlinear dynamic model to obtain the time domain waveforms and frequency domain waveforms of the multiple channels; based on the time domain waveforms and frequency domain waveforms, determining the source domain data of the multiple channels.

[0083] Specifically, when solving the equations, key characteristic frequencies such as BPFO, BPFI, and BSF can be calculated. By removing the DC component from the frequency domain signal, the signals of the bearing inner ring, bearing outer ring, and rolling element damage stage are further analyzed in detail in the time domain and frequency domain. In the time domain diagram, obvious pulse signals can be observed, especially when the rolling element passes through the defective part, the vibration signal presents regular pulses. In the frequency domain analysis, especially the envelope spectrum observation, the simulation signal of the bearing inner ring damage shows a peak that is highly consistent with the theoretical calculation frequency, and the error is maintained within an acceptable range.

[0084] Step 40: training a preset life prediction model of the rotating machinery based on the source domain data of the plurality of channels and the target domain data of the rotating machinery.

[0085] The preset life prediction model includes a common feature extractor corresponding to the source domain data of the multiple channels, and domain-specific feature extractors, domain-specific regressors and domain-specific discriminators corresponding to the source domain data of the multiple channels. The common feature extractor is composed of three convolutional layers, which are used to extract common representations from the source domain data and the target domain data to map the original data to a common feature space; the domain-specific feature extractor is composed of three convolutional layers, each of which is followed by an average pooling layer and an activation layer; the domain-specific regressor and the domain-specific discriminator are composed of fully connected layers.

[0086] It should be noted that the number of model layers in the embodiment of the present invention is not limited to the above, and may also be other numbers of layers.

[0087] The training process for the preset life prediction model, such as Figure 4 As shown, specifically including:

[0088] Step 41: construct an initial common feature extractor corresponding to the source domain data of the multiple channels, and an initial domain-specific feature extractor, an initial domain-specific regressor, and an initial domain-specific discriminator corresponding to the source domain data of the multiple channels respectively.

[0089] Among them, the number of initial common feature extractors is one, and the number of initial domain-specific feature extractors, initial domain-specific regressors and initial domain-specific discriminators is the same as the number of channels.

[0090] For the embodiment of the present invention, by initializing network parameters, an initial common feature extractor, an initial domain-specific feature extractor, an initial domain-specific regressor and an initial domain-specific discriminator are respectively constructed.

[0091] Step 42: inputting the source domain data of the plurality of channels and the target domain data into the initial common feature extractor for feature extraction to obtain common features between the source domain data of the plurality of channels and the target domain data;

[0092] For example, the source domain data and target domain data of channel A are input into the initial common feature extractor to extract the common features between the source domain data and target domain data of channel A. At the same time, the source domain data and target domain data of channel B are input into the initial common feature extractor to extract the common features between the source domain data and target domain data of channel B.

[0093] Step 43: input the source domain data of the multiple channels and their corresponding common features into the corresponding initial domain-specific feature extractor for feature extraction to obtain domain-specific features of the source domain data of the multiple channels.

[0094] For the embodiments of the present invention, due to the unique differences between different source domain data, it is necessary to design an independent feature extractor for each specific domain to capture the uniqueness of each domain. The structure of each initial domain-specific feature extractor is unified, including three convolutional layers, each of which is followed by an average pooling layer and an activation layer. It should be noted that the number of convolutional layers is not limited to the above, and can also be other numbers of layers.

[0095] For example, the source domain data of channel A and its corresponding common features are input to the initial domain-specific feature extractor corresponding to channel A for feature extraction. Similarly, the source domain data of channel B and its corresponding common features are input to the initial domain-specific feature extractor corresponding to channel B for feature extraction.

[0096] The embodiment of the present invention takes into account that the source domain data under different channel degradation modes have certain differences in the feature level. If it is difficult to directly align with the target domain data into a feature space, the embodiment of the present invention adopts a multi-channel source domain migration model domain nonlinear dynamics knowledge fusion method. The migration model aligns the source domain data and target domain data of each channel into different feature spaces and optimizes the differences.

[0097] Step 44: input the domain-specific features of the source domain data of the multiple channels into the corresponding initial domain-specific regressors and initial domain-specific discriminators for life prediction, and obtain the predicted lifespans output by the multiple initial domain-specific regressors.

[0098] Among them, the number of initial domain-specific regressors and initial domain-specific discriminators is the same as the number of initial domain-specific feature extractors.

[0099] The embodiment of the present invention adopts multiple initial domain-specific feature extractors and multiple initial domain-specific regressors, so as to ensure that source domain data and target domain data from different channels pass through their respective feature extractors and regressors without interfering with each other.

[0100] Step 45: construct a loss function based on the common features, the domain-specific features, and the predicted lifespans output by the multiple initial domain-specific regressors.

[0101] For the embodiment of the present invention, when constructing the loss function, step 45 specifically includes: calculating the regression error loss based on the predicted lifespan output by the multiple initial domain-specific regressors and the actual lifespan of the rotating machinery; calculating the migration domain adaptation loss based on the source domain data of the multiple channels and the target domain data; calculating the output difference loss between different initial domain-specific regressors based on the predicted lifespan output by the multiple initial domain-specific regressors; calculating the total adversarial loss of multiple initial domain-specific discriminators based on the common features and the domain-specific features; constructing the loss function based on the regression error loss, the migration domain adaptation loss, the output difference loss and the total adversarial loss.

[0102] Specifically, the specific formula of the final constructed loss function is as follows:

[0103]

[0104] in, are the training parameters of the common feature extractor, is the training parameter of the i-th domain-specific feature extractor, are the training parameters of the i-th domain-specific discriminator, is the training parameter of the i-th domain-specific regressor, L reg is the regression error loss, L MMD is the transfer domain adaptation loss, L alig is the output difference loss between different domain-specific regressors, L adv is the total adversarial loss of multiple domain-specific discriminators, and α, β, and γ are the weight factors corresponding to different loss functions in the balanced training process.

[0105] In calculating the regression error loss L reg When , the regression error loss can be determined based on the difference between the predicted lifespan output by the initial domain-specific regressor and the true lifespan.

[0106] In calculating the transfer domain adaptation loss L MMD When , the difference between the source domain data and the target domain data can be evaluated by calculating the MMD distance, and the difference is used to represent the migration domain adaptation loss L MMD , the specific calculation formula is as follows:

[0107]

[0108] Among them, X S is the source domain data, n S is the number of samples corresponding to the source domain data, X T is the target domain data, n T is the number of samples corresponding to the target domain data, H is the regenerated Hilbert space, and φ represents X Sand X T Mapping on H.

[0109] When calculating the output difference loss between different domain-specific regressors, the specific formula is as follows:

[0110]

[0111] in, and represents the output between different domain-specific regressors, i.e., predicted lifespan, and N represents the number of domain-specific regressors.

[0112] When calculating the total adversarial performance of multiple domain-specific discriminators, the specific formula is as follows:

[0113]

[0114] Among them, E S,i and E T,i The i-th group of source domain data and target domain data are sequentially extracted by the common feature extractor and the domain-specific feature extractor, respectively. adv Represents the total adversarial loss of multiple domain-specific discriminators.

[0115] Therefore, according to the above formula, the regression error loss, migration domain adaptation loss, output difference loss and total adversarial loss can be calculated. By adding the regression error loss, migration domain adaptation loss, output difference loss and total adversarial loss, the total loss function can be obtained.

[0116] Step 46: Based on the loss function, iteratively train the initial common feature extractor, the initial domain-specific feature extractor, the initial domain-specific regressor and the initial domain-specific discriminator until a preset number of iterations is reached, then stop the training and output the common feature extractor, the domain-specific feature extractor, the domain-specific regressor and the domain-specific discriminator.

[0117] For the embodiment of the present invention, after constructing the total loss function, based on the total loss function, the parameters in the initial common feature extractor, the initial domain-specific feature extractor, the initial domain-specific regressor and the initial domain-specific discriminator are continuously updated and iterated until a preset number of iterations is reached, and the iteration is stopped, and finally the trained common feature extractor, domain-specific feature extractor, domain-specific regressor and domain-specific discriminator are output.

[0118] The transfer learning model for integrating multi-channel source domains provided by the embodiment of the present invention aligns the source domain data of each channel with the target domain data into different feature spaces, and designs different adapters for feature extraction and modeling. The feature optimization strategy can effectively reduce the feature differences between multi-source domains, and improve the generalization ability and prediction accuracy of the model. In addition, the embodiment of the present invention combines the simulation data generated by the kinetic model with the multi-source domain transfer learning model to achieve an effective fusion of data-driven and knowledge-driven, thereby achieving more accurate and reliable remaining life prediction in small sample scenarios, thereby solving the dual challenges of data scarcity and distribution differences.

[0119] Step 47: Determine the preset life prediction model based on the common feature extractor, the domain-specific feature extractor, the domain-specific regressor and the domain-specific discriminator.

[0120] For the embodiment of the present invention, after the preset life prediction model is constructed, the target domain test data of the rotating machinery is input into the preset life prediction model for life prediction, and the life prediction result of the rotating machinery is obtained. The overall process of the embodiment of the present invention is as follows: Figure 5 As shown, it includes: simulation original domain generation, model training and model testing.

[0121] The embodiment of the present invention provides a small sample life prediction model training method based on multi-channel source domain transfer learning. By constructing a target nonlinear dynamic model of a rotating machinery gradually from a healthy stage, a degradation stage to a damaged stage, and performing simulation analysis on the target nonlinear dynamic model, a large amount of simulation data covering the entire cycle and different channel modes can be generated, thereby obtaining a large amount of source domain data of various fault types, and then providing the model with rich source domain data when training samples are scarce, thereby improving the prediction accuracy of the model. In addition, in the process of transfer learning, the present application can align the source domain data and target domain data of different channels to different feature spaces by training a common feature extractor corresponding to the source domain data of multiple channels, and a domain-specific feature extractor, a domain-specific regressor, and a domain-specific discriminator corresponding to the source domain data of multiple channels. Therefore, the differences between the source domain data of different channels can be processed, thereby improving the generalization ability of the life prediction model.

[0122] Further, as Figure 1 , Figure 3 and Figure 4 The specific implementation of the method shown in the embodiment provides a small sample life prediction model training device based on multi-channel source domain transfer learning, such as Figure 6 As shown, the device includes: a construction unit 101, a simulation unit 102 and a training unit 103.

[0123] The construction unit 101 can be used to construct a five-degree-of-freedom differential equation of a rotating machine.

[0124] The construction unit 101 can also be used to construct a target nonlinear dynamic model of the rotating machinery gradually from a healthy stage, a degraded stage to a damaged stage based on the five-degree-of-freedom differential equation.

[0125] The simulation unit 102 may be used to perform simulation analysis on the target nonlinear dynamics model to obtain source domain data of multiple channels.

[0126] The training unit 103 can be used to train a preset life prediction model of the rotating machinery based on the source domain data of the multiple channels and the target domain data of the rotating machinery, wherein the preset life prediction model includes a common feature extractor corresponding to the source domain data of the multiple channels, and domain-specific feature extractors, domain-specific regressors and domain-specific discriminators corresponding to the source domain data of the multiple channels respectively.

[0127] In some embodiments, the rotating machine includes a bearing, and the bearing includes a plurality of rolling elements, a bearing inner ring, and a bearing outer ring. The construction unit 101 includes: a first construction module, an adjustment module, and a first determination module.

[0128] The first construction module may be used to construct a first contact deformation formula between any one of the plurality of rolling elements and the inner ring of the bearing and the outer ring of the bearing in the healthy stage.

[0129] The first construction module can also be used to substitute the first contact deformation formula into the five-degree-of-freedom differential equation to construct a nonlinear dynamic model of the bearing in the healthy stage.

[0130] The adjustment module can be used to perform deformation adjustment on the nonlinear dynamic model in the healthy stage to obtain the nonlinear dynamic model of the bearing in the degraded stage.

[0131] The adjustment module may also be used to perform deformation adjustment on the nonlinear dynamic model in the degradation stage to obtain the nonlinear dynamic model of the bearing in the damage stage.

[0132] The first determination module may be configured to determine the nonlinear dynamic model at the damage stage as the target nonlinear dynamic model.

[0133] In some embodiments, the adjustment module can be specifically used to construct, for the degradation stage, based on the first contact deformation formula, a second contact deformation formula between any one of the rolling elements and the inner ring and the outer ring of the bearing in the case of a bearing outer ring defect, a third contact deformation formula between any one of the rolling elements and the inner ring and the outer ring of the bearing in the case of a bearing inner ring defect, and a fourth contact deformation formula between any one of the rolling elements and the inner ring and the outer ring of the bearing in the case of a rolling element defect; based on the second contact deformation formula, the third contact deformation formula and the fourth contact deformation formula, the nonlinear dynamic model of the healthy stage is deformed and adjusted to obtain the nonlinear dynamic model of the bearing in the degradation stage.

[0134] In some embodiments, the adjustment module can also be specifically used to construct, for the damage stage, based on the second contact deformation formula, the third contact deformation formula and the fourth contact deformation formula, the fifth contact deformation formula between any one of the rolling elements and the inner ring and the outer ring of the bearing in the case of a bearing outer ring defect, the sixth contact deformation formula between any one of the rolling elements and the inner ring and the outer ring of the bearing in the case of a bearing inner ring defect, and the seventh contact deformation formula between any one of the rolling elements and the inner ring and the outer ring of the bearing in the case of a rolling element defect; based on the fifth contact deformation formula, the sixth contact deformation formula and the seventh contact deformation formula, the nonlinear dynamic model of the degradation stage is deformed and adjusted to obtain the nonlinear dynamic model of the bearing in the damage stage.

[0135] In some embodiments, the simulation unit 102 can be specifically used to solve the equations corresponding to the target nonlinear dynamic model to obtain the time domain waveforms and frequency domain waveforms of the multiple channels; based on the time domain waveforms and frequency domain waveforms, determine the source domain data of the multiple channels.

[0136] In some embodiments, the training unit 103 includes: a second construction module, an extraction module and a prediction module, a training module and a second determination module.

[0137] The second construction module can be used to construct an initial common feature extractor corresponding to the source domain data of the multiple channels, as well as an initial domain-specific feature extractor, an initial domain-specific regressor and an initial domain-specific discriminator corresponding to the source domain data of the multiple channels respectively.

[0138] The extraction module can be used to input the source domain data of the multiple channels and the target domain data into the initial common feature extractor for feature extraction, so as to obtain the common features between the source domain data of the multiple channels and the target domain data.

[0139] The extraction module can also be used to input the source domain data of the multiple channels and their corresponding common features into the corresponding initial domain-specific feature extractor for feature extraction to obtain domain-specific features of the source domain data of the multiple channels.

[0140] The prediction module can be used to input the domain-specific features of the source domain data of the multiple channels into the corresponding initial domain-specific regressors and initial domain-specific discriminators for life prediction, and obtain the predicted lifespans output by the multiple initial domain-specific regressors.

[0141] The second construction module can also be used to construct a loss function based on the common features, the domain-specific features and the predicted lifespans output by the multiple initial domain-specific regressors.

[0142] The training module can be used to iteratively train the initial common feature extractor, the initial domain-specific feature extractor, the initial domain-specific regressor and the initial domain-specific discriminator based on the loss function until a preset number of iterations is reached, then stop the training and output the common feature extractor, the domain-specific feature extractor, the domain-specific regressor and the domain-specific discriminator.

[0143] The second determination module can be used to determine the preset life prediction model based on the common feature extractor, the domain-specific feature extractor, the domain-specific regressor and the domain-specific discriminator.

[0144] In some embodiments, the second construction module can be specifically used to calculate the regression error loss based on the predicted lifespans output by the multiple initial domain-specific regressors and the actual lifespan of the rotating machinery; calculate the migration domain adaptation loss based on the source domain data of the multiple channels and the target domain data; calculate the output difference loss between different domain-specific regressors based on the predicted lifespans output by the multiple initial domain-specific regressors; calculate the total adversarial loss of multiple domain-specific discriminators based on the common features and the domain-specific features; and construct the loss function based on the regression error loss, the migration domain adaptation loss, the output difference loss and the total adversarial loss.

[0145] In some embodiments, the device further includes: an acquisition unit and a prediction unit.

[0146] The acquisition unit may be used to acquire target domain test data of the rotating machinery.

[0147] The prediction unit may be used to input the target domain test data into the preset life prediction model to perform life prediction, thereby obtaining a life prediction result of the rotating machinery.

[0148] It should be noted that for other corresponding descriptions of the functional units involved in the small sample life prediction model training device based on multi-channel source domain transfer learning provided by the embodiment of the present invention, please refer to Figure 1 , Figure 3 and Figure 4 The corresponding description in will not be repeated here.

[0149] Based on the above Figure 1 , Figure 3 and Figure 4 The method shown in the embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the above-mentioned Figure 1 , Figure 3 and Figure 4 The shown method is a small sample lifetime prediction model training method based on multi-channel source domain transfer learning.

[0150] Based on this understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, USB flash drive, mobile hard disk, etc.), including a number of instructions for enabling an electronic device (which can be a personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of the present application.

[0151] Based on the above Figure 1 , Figure 3 and Figure 4 The method shown, and Figure 6 In order to achieve the above-mentioned purpose, the embodiment of the present application also provides an electronic device, which can be a personal computer, a tablet computer, a server, or other network equipment, etc. The device includes a storage medium and a processor; the storage medium is used to store a computer program; the processor is used to execute the computer program to achieve the above-mentioned Figure 1 , Figure 3 and Figure 4 The shown method is a small sample lifetime prediction model training method based on multi-channel source domain transfer learning.

[0152] Optionally, the above-mentioned physical device may also include a user interface, a network interface, a camera, a radio frequency (RF) circuit, a sensor, an audio circuit, a WI-FI module, etc. The user interface may include a display, an input unit such as a keyboard, etc., and the optional user interface may also include a USB interface, a card reader interface, etc. The network interface may optionally include a standard wired interface, a wireless interface (such as a WI-FI interface), etc.

[0153] Those skilled in the art will appreciate that the above-mentioned physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or a combination of certain components, or different arrangements of components.

[0154] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the above-mentioned physical device, and supports the operation of the information processing program and other software and / or programs. The network communication module is used to realize the communication between the components inside the storage medium, and the communication with other hardware and software in the information processing physical device.

[0155] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform, or by hardware.

[0156] The embodiment of the present invention constructs a target nonlinear dynamic model of the rotating machinery gradually from the healthy stage, the degradation stage to the damage stage, and performs simulation analysis on the target nonlinear dynamic model, so as to generate a large amount of simulation data covering the whole cycle and different channel modes, thereby being able to obtain a large amount of source domain data of various fault types, and then being able to provide the model with rich source domain data when training samples are scarce, thereby improving the prediction accuracy of the model. In addition, in the process of transfer learning, the present application can align the source domain data and target domain data of different channels to different feature spaces by training a common feature extractor corresponding to the source domain data of multiple channels, and a domain-specific feature extractor, a domain-specific regressor, and a domain-specific discriminator corresponding to the source domain data of multiple channels respectively, so as to be able to handle the differences between the source domain data of different channels and improve the generalization ability of the life prediction model.

[0157] Those skilled in the art will appreciate that the accompanying drawings are only schematic diagrams of a preferred implementation scenario, and the modules or processes in the accompanying drawings are not necessarily necessary for implementing the present application. Those skilled in the art will appreciate that the modules in the devices in the implementation scenario can be distributed in the devices of the implementation scenario according to the description of the implementation scenario, or can be changed accordingly and located in one or more devices different from the present implementation scenario. The modules of the above-mentioned implementation scenario can be combined into one module, or can be further split into multiple submodules.

[0158] The above serial numbers of this application are only for description and do not represent the advantages and disadvantages of the implementation scenarios. The above disclosure is only a few specific implementation scenarios of this application, but this application is not limited to them, and any changes that can be thought of by technicians in this field should fall within the scope of protection of this application.

Claims

1. A small sample life prediction model training method based on multi-channel source domain transfer learning, characterized in that: include: Construct the five-degree-of-freedom differential equations for rotating machinery; Based on the five-degree-of-freedom differential equation, a target nonlinear dynamic model of the rotating machinery gradually progressing from a healthy stage, a degraded stage to a damaged stage is constructed; Performing simulation analysis on the target nonlinear dynamics model to obtain source domain data of multiple channels; Based on the source domain data of the multiple channels and the target domain data of the rotating machinery, a preset life prediction model of the rotating machinery is trained, wherein the preset life prediction model includes a common feature extractor corresponding to the source domain data of the multiple channels, and domain-specific feature extractors, domain-specific regressors and domain-specific discriminators corresponding to the source domain data of the multiple channels respectively.

2. The method according to claim 1, characterized in that The rotating machine includes a bearing, and the bearing includes a plurality of rolling elements, a bearing inner ring, and a bearing outer ring. Based on the five-degree-of-freedom differential equation, a target nonlinear dynamic model of the rotating machine gradually changing from a healthy stage, a degraded stage to a damaged stage is constructed, including: Constructing a first contact deformation formula between any one of the plurality of rolling elements and the inner ring of the bearing and the outer ring of the bearing in the healthy stage; Substituting the first contact deformation formula into the five-degree-of-freedom differential equation to construct a nonlinear dynamic model of the bearing in the healthy stage; Performing deformation adjustment on the nonlinear dynamic model in the healthy stage to obtain the nonlinear dynamic model of the bearing in the degraded stage; Performing deformation adjustment on the nonlinear dynamic model of the degradation stage to obtain the nonlinear dynamic model of the bearing in the damage stage; The nonlinear dynamic model of the damage stage is determined as the target nonlinear dynamic model.

3. The method according to claim 2, characterized in that The deformation adjustment of the nonlinear dynamic model in the healthy stage to obtain the nonlinear dynamic model of the bearing in the degraded stage includes: For the degradation stage, based on the first contact deformation formula, a second contact deformation formula between any one of the rolling elements and the inner ring and the outer ring of the bearing in the case of a bearing outer ring defect, a third contact deformation formula between any one of the rolling elements and the inner ring and the outer ring of the bearing in the case of a bearing inner ring defect, and a fourth contact deformation formula between any one of the rolling elements and the inner ring and the outer ring of the bearing in the case of a rolling element defect are respectively constructed; Based on the second contact deformation formula, the third contact deformation formula and the fourth contact deformation formula, deformation adjustment is performed on the nonlinear dynamic model in the healthy stage to obtain the nonlinear dynamic model of the bearing in the degraded stage; The deformation adjustment of the nonlinear dynamic model in the degradation stage to obtain the nonlinear dynamic model of the bearing in the damage stage includes: For the damage stage, based on the second contact deformation formula, the third contact deformation formula and the fourth contact deformation formula, a fifth contact deformation formula between any one of the rolling elements and the inner ring and the outer ring of the bearing in the case of a bearing outer ring defect, a sixth contact deformation formula between any one of the rolling elements and the inner ring and the outer ring of the bearing in the case of a bearing inner ring defect, and a seventh contact deformation formula between any one of the rolling elements and the inner ring and the outer ring of the bearing in the case of a rolling element defect are respectively constructed; Based on the fifth contact deformation formula, the sixth contact deformation formula and the seventh contact deformation formula, the nonlinear dynamic model of the degradation stage is deformed and adjusted to obtain the nonlinear dynamic model of the bearing in the damage stage.

4. The method according to claim 1, characterized in that The simulation analysis of the target nonlinear dynamics model is performed to obtain source domain data of multiple channels, including: Solving equations corresponding to the target nonlinear dynamics model to obtain time domain waveforms and frequency domain waveforms of the multiple channels; Based on the time domain waveform and the frequency domain waveform, source domain data of multiple channels are determined.

5. The method according to any one of claims 1 to 4, characterized in that: The step of training a preset life prediction model of the rotating machinery according to the source domain data of the plurality of channels and the target domain data of the rotating machinery comprises: Constructing an initial common feature extractor commonly corresponding to the source domain data of the multiple channels, and an initial domain-specific feature extractor, an initial domain-specific regressor, and an initial domain-specific discriminator respectively corresponding to the source domain data of the multiple channels; Inputting the source domain data of the multiple channels and the target domain data into the initial common feature extractor for feature extraction respectively, and obtaining common features between the source domain data of the multiple channels and the target domain data respectively; Inputting the source domain data of the plurality of channels and the corresponding common features thereof into the corresponding initial domain-specific feature extractor for feature extraction to obtain domain-specific features of the source domain data of the plurality of channels; Inputting domain-specific features of the source domain data of the multiple channels into corresponding initial domain-specific regressors and initial domain-specific discriminators for life prediction, and obtaining predicted lifespans output by multiple initial domain-specific regressors; constructing a loss function according to the common features, the domain-specific features, and the predicted lifespans output by the multiple initial domain-specific regressors; Based on the loss function, iteratively train the initial common feature extractor, the initial domain-specific feature extractor, the initial domain-specific regressor and the initial domain-specific discriminator until a preset number of iterations is reached, stop the training, and output the common feature extractor, the domain-specific feature extractor, the domain-specific regressor and the domain-specific discriminator; The preset life prediction model is determined according to the common feature extractor, the domain-specific feature extractor, the domain-specific regressor and the domain-specific discriminator.

6. The method according to claim 5, characterized in that The step of constructing a loss function according to the common features, the domain-specific features and the predicted lifespans output by the multiple initial domain-specific regressors comprises: Calculating regression error loss based on the predicted lifespans output by the multiple initial domain-specific regressors and the actual lifespan of the rotating machinery; Calculating a migration domain adaptation loss according to the source domain data of the plurality of channels and the target domain data; Calculating output difference losses between different domain-specific regressors based on the predicted lifespans output by the multiple initial domain-specific regressors; Calculating a total adversarial loss of a plurality of domain-specific discriminators based on the common features and the domain-specific features; The loss function is constructed according to the regression error loss, the transfer domain adaptation loss, the output difference loss and the total adversarial loss.

7. The method according to claim 1, characterized in that The method further comprises: Acquiring target domain test data of the rotating machinery; The target domain test data is input into the preset life prediction model to perform life prediction, thereby obtaining a life prediction result of the rotating machinery.

8. A small sample life prediction model training device based on multi-channel source domain transfer learning, characterized in that: include: Construction unit for constructing five-degree-of-freedom differential equations for rotating machinery; The construction unit is further used to construct a target nonlinear dynamic model of the rotating machinery gradually from a healthy stage, a degraded stage to a damaged stage based on the five-degree-of-freedom differential equation; A simulation unit, used for performing simulation analysis on the target nonlinear dynamics model to obtain source domain data of multiple channels; A training unit is used to train a preset life prediction model of the rotating machinery based on the source domain data of the multiple channels and the target domain data of the rotating machinery, wherein the preset life prediction model includes a common feature extractor corresponding to the source domain data of the multiple channels, and domain-specific feature extractors, domain-specific regressors and domain-specific discriminators corresponding to the source domain data of the multiple channels respectively.

9. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. An electronic device comprising a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

Images