Decoupled transfer method driven by degradation knowledge and data fusion based on small sample

Abstract

The application discloses a small sample-based degradation knowledge and data fusion driven transfer decoupling method, and relates to the technical field of predictive maintenance of rotating machinery. The method comprises the following steps: obtaining source domain data and target domain data of the rotating machinery; determining source domain invariant features and source domain private features, target domain invariant features and target domain private features based on an initial life prediction model; constructing a loss function according to the source domain invariant features, the target domain invariant features, the source domain private features, the target domain private features and the predicted life; constructing a degradation mechanism space state model according to the source domain invariant features and the target domain invariant features; gradually estimating the parameters of the degradation mechanism space state model by using a particle filtering algorithm; and when the initial life prediction model continuously updates the weights based on the loss function, using the degradation mechanism space state model after the parameter estimation to guide the weight updating process of the initial life prediction model. The application can improve the generalization ability and interpretability of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of predictive maintenance of rotating machinery, and in particular to a transfer decoupling method driven by fusion of small sample-based degradation knowledge and data. BACKGROUND

[0002] Industrial equipment is increasingly becoming large-scale, automated, and intelligent. Bearings, gears, and other rotating machinery are widely used in aerospace, transportation, and other fields, and their operating conditions are directly related to the service life of various mechanical equipment. Therefore, it is of positive practical significance to predict the service life of bearings and other rotating machinery. In order to solve the small sample problem in the service life prediction of rotating machinery, transfer learning theory is introduced into the prediction model to reduce the distribution difference between the source domain and the target domain.

[0003] At present, most of the research on transfer learning mainly focuses on minimizing the difference in global feature distribution across domains, ignoring the inherent properties of the domain. However, if the global alignment is directly used without proper restriction on the features, the representativeness across domains may be affected, making it difficult to ensure the generalization ability of the model. SUMMARY

[0004] Therefore, the present application provides a transfer decoupling method driven by fusion of small sample-based degradation knowledge and data, which can not only reveal the similarity between domains, but also take into account the uniqueness of the domain, ensuring the training effect of the model and improving the generalization ability of the model.

[0005] According to a first aspect of the present application, a transfer decoupling method driven by fusion of small sample-based degradation knowledge and data is provided, which comprises:

[0006] Obtaining source domain data and target domain data of a rotating machinery, and an initial service life prediction model, the initial service life prediction model comprising a target domain private feature extractor, a source domain private feature extractor, a domain-invariant feature extractor, a domain-invariant feature decoder, a domain-mixed feature decoder, and a service life predictor;

[0007] Based on the source domain data and the target domain data, and the initial service life prediction model, determining the source domain-invariant features and the source domain private features corresponding to the source domain data, the target domain-invariant features and the target domain private features corresponding to the target domain data, and the predicted service life of the rotating machinery;

[0008] According to the source domain-invariant features, the target domain-invariant features, the source domain private features, the target domain private features, and the predicted service life, a loss function is constructed;

[0009] construct a degradation mechanism space state model according to the source domain invariant feature and the target domain invariant feature, and determine initial parameters of the degradation mechanism space state model;

[0010] Based on the initial parameters, the parameters of the degradation mechanism space state model are estimated step by step by using a particle filtering algorithm to obtain a degradation mechanism space state model after parameter estimation;

[0011] When the initial life prediction model continuously updates the weight based on the loss function, the parameter estimation of the degradation mechanism space state model is used to guide the weight update process of the initial life prediction model until a preset iteration number is reached, and a preset life prediction model is output.

[0012] According to a second aspect of the present application, a small sample based degradation knowledge and data fusion driven migration decoupling device is provided, which comprises:

[0013] An acquisition unit is configured to acquire source domain data and target domain data of a rotating machine, and an initial life prediction model, wherein the initial life prediction model comprises a target domain private feature extractor, a source domain private feature extractor, a domain invariant feature extractor, a domain invariant feature decoder, a domain mixed feature decoder, and a life predictor;

[0014] A determination unit is configured to determine source domain invariant features and source domain private features corresponding to the source domain data, target domain invariant features and target domain private features corresponding to the target domain data, and a predicted life of the rotating machine based on the source domain data and the target domain data, and the initial life prediction model;

[0015] A first construction unit is configured to construct a loss function according to the source domain invariant features, the target domain invariant features, the source domain private features, the target domain private features, and the predicted life;

[0016] A second construction unit is configured to construct a degradation mechanism space state model according to the source domain invariant features and the target domain invariant features, and determine initial parameters of the degradation mechanism space state model;

[0017] An estimation unit is configured to estimate the parameters of the degradation mechanism space state model step by step by using a particle filtering algorithm based on the initial parameters to obtain a degradation mechanism space state model after parameter estimation;

[0018] A training unit is configured to guide the weight update process of the initial life prediction model by using the parameter estimation of the degradation mechanism space state model when the initial life prediction model continuously updates the weight based on the loss function, and output a preset life prediction model when a preset iteration number is reached.

[0019] According to a third aspect of the present application, a storage medium is provided, and a computer program is stored on the storage medium, and the program is executed by a processor to implement the small sample based degradation knowledge and data fusion driven transfer decoupling method.

[0020] According to a fourth aspect of the present application, an electronic device is provided, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, and the processor executes the program to implement the small sample based degradation knowledge and data fusion driven transfer decoupling method.

[0021] Through the above technical solution, the small sample based degradation knowledge and data fusion driven transfer decoupling method provided by the present application can decouple domain private features and domain invariant features by extracting source domain invariant features, source domain private features, target domain invariant features and target domain private features respectively. Compared with the prior art, the present application not only reveals the similarity between domains, but also takes into account the uniqueness of the domain, greatly improves the generalization ability of the model, and effectively deals with the global domain generalization problem under the small sample condition. In addition, by integrating the degradation mechanism into the life prediction model, the present application can not only enhance the prediction accuracy, but also improve the robustness of the prediction model to abnormal data and noise. Further, in order to improve the perception ability of the prediction model to dynamic time relationship, the present application uses the particle filtering algorithm for forward step-by-step prediction, captures the dependency relationship between long-distance time features, and improves the explainability of the prediction model.

[0022] The above description is only a summary of the technical solutions of the present application. In order to enable a clearer understanding of the technical means of the present application, the content of the specification can be implemented, and in order to enable the above and other purposes, features and advantages of the present application to be more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0023] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The schematic 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:

[0024] Figure 1 A flowchart of a small sample based degradation knowledge and data fusion driven transfer decoupling method provided by an embodiment of the present application is shown;

[0025] Figure 2 A feature alignment diagram provided by an embodiment of the present application is shown;

[0026] Figure 3 A flowchart of a feature extraction method provided by an embodiment of the present application is shown;

[0027] Figure 4 A flowchart of data preprocessing, model training and model testing provided by an embodiment of the application is shown.

[0028] Figure 5 A structural diagram of a degradation knowledge and data fusion driven transfer decoupling device based on small samples provided by an embodiment of the application is shown. DETAILED DESCRIPTION

[0029] Hereinafter, the application will be described in detail with reference to the accompanying drawings and in conjunction with embodiments. It should be noted that the embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0030] The prior art directly adopts global alignment without proper restriction on the features, which affects the representation across domains and makes it difficult to guarantee the generalization ability of the model.

[0031] To solve the above problems, an embodiment of the application provides a degradation knowledge and data fusion driven transfer decoupling method based on small samples, as shown in the figure. Figure 1 The method comprises the following steps.

[0032] Step 10, obtaining source domain data and target domain data of a rotating machine, and an initial life prediction model.

[0033] The rotating machine specifically can include gears, engines, bearings, etc. The initial life prediction model includes a target domain private feature extractor, a source domain private feature extractor, a domain invariant feature extractor, a domain invariant feature decoder, a domain mixed feature decoder and a life predictor, wherein the domain invariant feature decoder and the domain mixed feature decoder are used to disentangle the domain invariant feature and the domain private feature, which can maximize the relative information amount; the domain mixed feature decoder is used to restore the original feature information, ensuring the minimum loss of single-domain relative information amount, while the domain invariant feature decoder is used to reconstruct the original feature, balancing the relative information amount of single-domain and double-domain, and meeting better cross-domain transfer universality. The target domain private feature extractor, the source domain private feature extractor and the domain invariant feature extractor are based on a multi-layer perception framework, the life predictor includes a fully connected layer, and the domain invariant feature decoder and the domain mixed feature decoder include an up-sampling / anti-convolution layer, a convolution layer and a decoding layer.

[0034] The embodiment of the application is mainly applicable to health monitoring of rotating machines, and mainly relates to key equipment such as aircraft engines, railway vehicles and wind driven generators.

[0035] For the embodiment of the application, the original vibration data is first acquired, and then the time domain features, the frequency domain features and the features based on trigonometric functions corresponding to the original vibration data are extracted. The time domain features include entropy, energy, root mean square, kurtosis, root mean square, absolute average value, maximum absolute value, skewness, shape factor, clearance coefficient, pulse factor, wave peak factor, standard deviation, etc. The frequency domain features include the energy of 16 frequency bands generated by four-level wavelet packet decomposition. The features based on trigonometric functions include the standard deviation of inverse hyperbolic cosine, the standard deviation of inverse hyperbolic sine and the standard deviation of inverse tangent, etc. The extracted time domain features, frequency domain features and features based on trigonometric functions are used as the source domain data and the target domain data of the rotating machinery.

[0036] At the same time, the network parameters are initialized, and an initial life prediction model is constructed, which includes six modules, namely a target domain private feature extractor, a source domain private feature extractor, a domain invariant feature extractor, a domain invariant feature decoder, a domain mixed feature decoder and a life predictor.

[0037] Step 20, based on the source domain data and the target domain data, and the initial life prediction model, determining the source domain invariant features and the source domain private features corresponding to the source domain data, the target domain invariant features and the target domain private features corresponding to the target domain data, and the predicted life of the rotating machinery.

[0038] In order to maximize the mining of cross-domain similarity features, the embodiment of the application separates the domain private features and the domain invariant features by feature decoupling, as shown in Figure 2 Compared with the global feature comparison method, the division method of the embodiment of the application not only reveals the similarity between domains, but also takes into account the uniqueness of the domain, greatly improves the generalization ability of the model, and effectively deals with the global domain generalization problem in the case of small sample.

[0039] For the extraction process of the source domain invariant features, the source domain private features, the target domain invariant features and the target domain private features, as shown in Figure 3 The method comprises the following steps:

[0040] Step 21, using the domain invariant feature extractor to extract the source domain invariant original features corresponding to the source domain data and the target domain invariant original features corresponding to the target domain data.

[0041] For the embodiment of the application, the source domain data and the target domain data are respectively input into the domain invariant feature extractor for feature extraction, to obtain the source domain invariant original features and the target domain invariant original features, wherein the source domain invariant original features and the target domain invariant original features are features that have not been decoded.

[0042] Step 22, extracting source domain private original features corresponding to the source domain data and target domain private original features corresponding to the target domain data by using the target domain private feature extractor and the source domain private feature extractor respectively.

[0043] For the embodiment of the application, the source domain data and the source domain invariant original features are input into the source domain private feature extractor for feature extraction to obtain the source domain private original features. Similarly, the target domain data and the target domain invariant original features are input into the target domain private feature extractor for feature extraction to obtain the target domain private original features. The source domain private original features and the target domain private original features are features that have not been decoded.

[0044] Step 23, decoding the source domain invariant original features and the target domain invariant original features by using the domain invariant feature decoder to obtain the source domain invariant features and the target domain invariant features, and decoding the source domain private original features and the target domain private original features by using the domain mixed feature decoder to obtain the source domain private features and the target domain private features.

[0045] For the embodiment of the application, the source domain invariant original features and the target domain invariant original features are input into the domain invariant feature decoder for decoding to obtain the source domain invariant features and the target domain invariant features. At the same time, the source domain private original features and the target domain private original features are input into the domain mixed feature decoder for decoding to obtain the source domain private features and the target domain private features.

[0046] Step 24, inputting the source domain invariant features and the source domain private features into the life predictor for prediction to obtain the predicted life of the rotating machinery.

[0047] For the embodiment of the application, the decoded source domain invariant features and the source domain private features are input into the life predictor (full connection layer) for prediction to obtain the predicted life of the rotating machinery.

[0048] The embodiment of the application can decouple the domain private features and the domain invariant features by extracting the source domain invariant features, the source domain private features, the target domain invariant features and the target domain private features respectively. The embodiment of the application not only reveals the similarity between domains, but also takes into account the uniqueness of the domain, greatly improves the generalization ability of the model, and effectively deals with the global domain generalization problem under the small sample condition.

[0049] Step 30, constructing a loss function according to the source domain invariant features, the target domain invariant features, the source domain private features, the target domain private features and the predicted life.

[0050] For the embodiment of the application, after extracting the source domain invariant feature, the target domain invariant feature, the source domain private feature and the target domain private feature, a loss function is constructed based on the source domain invariant feature, the target domain invariant feature, the source domain private feature and the target domain private feature. For the construction process of the loss function, step 30 specifically includes: based on the source domain invariant feature, the target domain invariant feature, the source domain private feature and the target domain private feature, respectively calculating the first distribution difference loss between the source domain private feature and the source domain invariant feature, the second distribution difference loss between the target domain private feature and the target domain invariant feature, and the third distribution difference loss between the source domain invariant feature and the target domain invariant feature; based on the predicted life and the actual life of the rotating machinery, calculating the life prediction loss; according to the life prediction loss, the first distribution difference loss, the second distribution difference loss and the third distribution difference loss, constructing the loss function.

[0051] Wherein, the first distribution difference loss, the second distribution difference loss and the third distribution difference loss can be calculated based on the domain adaptive theory, and the specific formula is as follows:

[0052]

[0053] Wherein, X S represents the source domain feature, including the source domain private feature and the source domain invariant feature, X T represents the target domain feature, including the target domain private feature and the target domain invariant feature, n S represents the sample number corresponding to the source domain data, n T represents the sample number corresponding to the target domain data, H is the reproducing Hilbert space, and φ represents the mapping of X S and X T on H.

[0054] After extracting the source domain private feature the source domain invariant feature the target domain private feature and the target domain invariant feature , MMD metric is used as a loss function to measure the distribution, including the first distribution difference loss between and the second distribution difference loss between and the third distribution difference loss L similarity .

[0055] ​Step 40, constructing a degradation mechanism space state model according to the source domain invariant feature and the target domain invariant feature, and determining initial parameters of the degradation mechanism space state model.

[0056] For the embodiment of the present application, after extracting the source domain invariant feature and the target domain invariant feature, a degradation mechanism space state model is established based on the source domain invariant feature and the target domain invariant feature. The degradation mechanism space state model is essentially a double exponential function, which is used to simulate the degradation process of the mechanical equipment. The main use of the model is to process random events in time series and predict the trend of future signals.

[0057] For the construction process of the degradation mechanism space state model, the method comprises: according to the source domain invariant feature and the target domain invariant feature, simulating the degradation process of the rotating mechanical equipment by using a double exponential function to obtain the degradation mechanism space state model.

[0058] ψ(t n ,a,b,c,d)=a*exp(b*t n )+c*exp(d*t n )

[0059] Wherein, a, b, c, d are parameters of the degradation mechanism space state model respectively, t n is a time series.

[0060] Then, the initial parameters of the degradation mechanism space state model are solved by using a maximum likelihood algorithm, that is, the maximum likelihood estimation method is used to make the likelihood function L(Φ / R) reach the maximum value, and at this time, the best initial parameters of the degradation mechanism space state model are obtained.

[0061]

[0062] Step 50, based on the initial parameters, gradually estimating the parameters of the degradation mechanism space state model by using a particle filtering algorithm to obtain the degradation mechanism space state model after parameter estimation.

[0063] For the embodiment of the present application, after determining the initial parameters of the degradation mechanism space state model, a particle filtering algorithm based on resampling is used to gradually estimate the parameters of the degradation mechanism space state model in time sequence and further predict the future state.

[0064] Therefore, step 50 specifically comprises: based on the initial parameters, determining a parameter update formula by using the particle filtering algorithm; according to the parameter update formula, gradually estimating the parameters of the degradation mechanism space state model to obtain the degradation mechanism space state model after parameter estimation.

[0065] Specifically, first, the historical observation data R 1:n-1 ∈Rn-1 The prior probability density function at time n is obtained, and the specific formula is as follows:

[0066] p(s n / R 1:n-1 )=∫p(s n / s n-1 )p(s n-1 / R 1:n-1 )ds n-1

[0067] Wherein, p(s n / s n-1 ) is the probability density function of the state transition function, and p(s n-1 / R 1:n-1 ) is the probability density function at time n-1. After obtaining the observation value at time n, the particle is updated and the posterior probability density function at time n is obtained, and the specific formula is as follows:

[0068]

[0069] Wherein, Indicates the particle weight at time n, Indicates the particle obtained by importance sampling from the distribution p(s n / R 1:n ). After obtaining the observation value at time n, the updating method (parameter updating formula) of the particle weight is as follows,

[0070]

[0071] After updating the parameters at each time by the particle filter, the degradation mechanism space state model can be correctly solved respectively, and the next life prediction can be performed according to the established two-stage double exponential model.

[0072] Step 60, when the initial life prediction model continuously updates the weight based on the loss function, the degradation mechanism space state model after the parameter estimation is used to guide the weight updating process of the initial life prediction model, until a preset iteration number is reached, and a preset life prediction model is output.

[0073] Wherein, the preset iteration number can be set according to actual business requirements, and the embodiment of the present application does not make specific limitation.

[0074] For the embodiment of the present application, after training the preset life prediction model, the target domain prediction data of the rotating machinery can be obtained, and the target domain prediction data 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 application is as follows, Figure 4As shown, specifically includes data processing, model training and model testing.

[0075] The embodiment of the application provides a small sample-based degradation knowledge and data fusion driven transfer decoupling method, compared with the prior art, by respectively extracting source domain invariant features, source domain private features, target domain invariant features and target domain private features, the domain private features and the domain invariant features can be decoupled, the method discloses the similarity between domains, and the uniqueness of the domain is also considered, so that the generalization ability of the model is greatly improved, and the global domain generalization problem in the small sample situation is effectively solved. In addition, by integrating the degradation mechanism into the life prediction model, the prediction accuracy can be improved, and the robustness of the prediction model to abnormal data and noise can be improved. Further, in order to improve the perception ability of the prediction model to dynamic time relationship, the particle filtering algorithm is used for forward step-by-step prediction, the dependence relationship between long-distance time features is captured, and the explainability of the prediction model is improved.

[0076] Further, as Figure 1 and Figure 3 The embodiment provides a small sample-based degradation knowledge and data fusion driven transfer decoupling device, as shown in the specific implementation of the method, Figure 5 The device comprises an acquisition unit 101, a determination unit 102, a first construction unit 103, a second construction unit 104, an estimation unit 105 and a training unit 106.

[0077] The acquisition unit 101 can be used for acquiring source domain data and target domain data of a rotating machine, and an initial life prediction model, wherein the initial life prediction model comprises a target domain private feature extractor, a source domain private feature extractor, a domain invariant feature extractor, a domain invariant feature decoder, a domain mixed feature decoder and a life predictor.

[0078] The determination unit 102 can be used for determining source domain invariant features and source domain private features corresponding to the source domain data, target domain invariant features and target domain private features corresponding to the target domain data, and a predicted life of the rotating machine based on the source domain data and the target domain data and the initial life prediction model.

[0079] The first construction unit 103 can be used for constructing a loss function according to the source domain invariant features, the target domain invariant features, the source domain private features, the target domain private features and the predicted life.

[0080] The second construction unit 104 can be used for constructing a degradation mechanism space state model according to the source domain invariant features and the target domain invariant features, and determining initial parameters of the degradation mechanism space state model.

[0081] The estimation unit 105 can be configured to estimate parameters of the degradation mechanism space state model based on the initial parameters using a particle filter algorithm to obtain a parameter-estimated degradation mechanism space state model.

[0082] The training unit 106 can be configured to guide a weight updating process of the initial life prediction model using the parameter-estimated degradation mechanism space state model while the initial life prediction model continuously updates weights based on the loss function, and output a preset life prediction model when a preset number of iterations is reached.

[0083] In some embodiments, the determination unit 102 can be specifically configured to extract source domain invariant original features corresponding to the source domain data and target domain invariant original features corresponding to the target domain data using the domain invariant feature extractor; extract source domain private original features corresponding to the source domain data and target domain private original features corresponding to the target domain data using the target domain private feature extractor and the source domain private feature extractor; decode the source domain invariant original features and the target domain invariant original features using the domain invariant feature decoder to obtain the source domain invariant features and the target domain invariant features, and decode the source domain private original features and the target domain private original features using the domain mixed feature decoder to obtain the source domain private features and the target domain private features; and input the source domain invariant features and the source domain private features into the life predictor for prediction to obtain the predicted life of the rotating machinery.

[0084] In some embodiments, the first construction unit 103 can be specifically configured to calculate a first distribution difference loss between the source domain private features and the source domain invariant features, a second distribution difference loss between the target domain private features and the target domain invariant features, and a third distribution difference loss between the source domain invariant features and the target domain invariant features based on the source domain invariant features, the target domain invariant features, the source domain private features, and the target domain private features; calculate a life prediction loss based on the predicted life and an actual life of the rotating machinery; and construct a loss function according to the life prediction loss, the first distribution difference loss, the second distribution difference loss, and the third distribution difference loss.

[0085] In some embodiments, the first construction unit 103 can be specifically configured to simulate a degradation process of the rotating machinery using a double exponential function according to the source domain invariant features and the target domain invariant features to obtain the degradation mechanism space state model.

[0086] ​In some embodiments, the first construction unit 103 can be further specifically configured to solve initial parameters of the degradation mechanism space state model by using a maximum likelihood algorithm.

[0087] In some embodiments, the estimation unit 105 can be specifically configured to determine a parameter update formula by using the particle filtering algorithm based on the initial parameters, and gradually estimate parameters of the degradation mechanism space state model according to the parameter update formula to obtain a parameter-estimated degradation mechanism space state model.

[0088] In some embodiments, the device further comprises a prediction unit.

[0089] The acquisition unit 101 can be further configured to acquire target domain prediction data of the rotating machinery.

[0090] The prediction unit can be configured to input the target domain prediction data into the preset life prediction model for life prediction to obtain a life prediction result of the rotating machinery.

[0091] It should be noted that other corresponding descriptions of each functional unit involved in the device provided by the embodiments of the present application based on small sample degradation knowledge and data fusion driven migration decoupling can be referred to the corresponding descriptions in Figure 1 and Figure 2 , which will not be repeated here.

[0092] Based on the above methods as shown in Figure 1 and Figure 3 , correspondingly, the present embodiment further provides a storage medium having a computer program stored thereon, which is executed by a processor to implement the above-mentioned small sample degradation knowledge and data fusion driven migration decoupling method as shown in Figure 1 and Figure 3 .

[0093] Based on such 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, a U disk, a mobile hard disk, etc.), and includes a plurality of instructions for making an electronic device (which can be a personal computer, a server, or a network device, etc.) execute the method of each implementation scenario of the present application.

[0094] Based on the above methods as shown in Figure 1 and Figure 3 , and Figure 5In order to achieve the above-mentioned purposes, the electronic device provided by the embodiment of the present application can be a personal computer, a tablet computer, a server, or other network devices, etc., which comprises a storage medium and a processor. Figure 1 and Figure 3 The small sample based degradation knowledge and data fusion driven migration decoupling method shown.

[0095] Optionally, the entity device can further 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 can include a display, an input unit such as a keyboard, etc. The optional user interface can further include a USB interface, a card reader interface, etc. The network interface can optionally include a standard wired interface, a wireless interface (such as a WI-FI interface), etc.

[0096] Those skilled in the art can understand that the above-mentioned entity device structure provided by the embodiment does not constitute a limitation on the entity device, and can include more or fewer components, or combine certain components, or different component arrangements.

[0097] The storage medium can further include an operating system and a network communication module. The operating system is a program for managing hardware and software resources of the entity device, and supports the running of information processing programs and other software and / or programs. The network communication module is used to realize communication between the components inside the storage medium, and communication with other hardware and software in the information processing entity device.

[0098] Through the above description of the embodiments, those skilled in the art can clearly understand that the present application can be realized by means of software and necessary general hardware platforms, or by hardware.

[0099] The embodiment of the present application can decouple the domain private feature and the domain invariant feature by extracting the source domain invariant feature, the source domain private feature, the target domain invariant feature and the target domain private feature respectively. This way not only reveals the similarity between domains, but also takes into account the uniqueness of the domain, greatly improves the generalization ability of the model, and effectively deals with the global domain generalization problem under the small sample situation. In addition, by incorporating the degradation mechanism into the life prediction model, the present application can not only enhance the prediction accuracy, but also improve the robustness of the prediction model to abnormal data and noise. Further, in order to improve the perception ability of the prediction model to dynamic time relationship, the present application uses the particle filtering algorithm for forward step-by-step prediction, captures the dependency relationship between long-distance time features, and improves the explainability of the prediction model.

[0100] Those skilled in the art can understand that the modules or flows in the drawings are not necessarily required for implementing the present application. Those skilled in the art can understand that the modules in the devices in the implementation scenarios can be distributed in the devices in the implementation scenarios according to the description of the implementation scenarios, or can be changed to be located in one or more devices different from the implementation scenarios. The modules in the above implementation scenarios can be combined into one module, or can be further split into multiple sub-modules.

[0101] The above serial numbers of the present application are only for description, and do not represent the advantages and disadvantages of the implementation scenarios. The above disclosure is only some specific implementation scenarios of the present application, however, the present application is not limited thereto, and any variations that can be thought of by those skilled in the art shall fall within the protection scope of the present application.

Claims

1. A small sample based degradation knowledge and data fusion driven transfer decoupling method, characterized in that, The method comprises the following steps: acquiring source domain data and target domain data of a rotating machine, and an initial life prediction model, wherein the initial life prediction model comprises a target domain private feature extractor, a source domain private feature extractor, a domain invariant feature extractor, a domain invariant feature decoder, a domain mixed feature decoder, and a life predictor; based on the source domain data and the target domain data, and the initial life prediction model, determining source domain invariant features and source domain private features corresponding to the source domain data, target domain invariant features and target domain private features corresponding to the target domain data, and a predicted life of the rotating machine; constructing a loss function according to the source domain invariant features, the target domain invariant features, the source domain private features, the target domain private features, and the predicted life; constructing a degradation mechanism space state model according to the source domain invariant features and the target domain invariant features, and determining initial parameters of the degradation mechanism space state model; based on the initial parameters, gradually estimating parameters of the degradation mechanism space state model by using a particle filtering algorithm to obtain a degradation mechanism space state model after parameter estimation; when the initial life prediction model continuously updates weights based on the loss function, using the degradation mechanism space state model after parameter estimation to guide the weight updating process of the initial life prediction model until a preset iteration number is reached, and outputting a preset life prediction model; wherein, based on the source domain data and the target domain data, and the initial life prediction model, determining the source domain invariant features and the source domain private features corresponding to the source domain data, the target domain invariant features and the target domain private features corresponding to the target domain data, and the predicted life of the rotating machine, comprises: extracting source domain invariant original features corresponding to the source domain data and target domain invariant original features corresponding to the target domain data by using the domain invariant feature extractor; extracting source domain private original features corresponding to the source domain data and target domain private original features corresponding to the target domain data by using the target domain private feature extractor and the source domain private feature extractor; decoding the source domain invariant original features and the target domain invariant original features by using the domain invariant feature decoder to obtain the source domain invariant features and the target domain invariant features, and decoding the source domain private original features and the target domain private original features by using the domain mixed feature decoder to obtain the source domain private features and the target domain private features; inputting the source domain invariant features and the source domain private features into the life predictor for prediction to obtain the predicted life of the rotating machine; the loss function is constructed according to the source domain invariant features, the target domain invariant features, the source domain private features, the target domain private features, and the predicted life, comprising: calculating a first distribution difference loss between the source domain private feature and the source domain invariant feature, a second distribution difference loss between the target domain private feature and the target domain invariant feature, and a third distribution difference loss between the source domain invariant feature and the target domain invariant feature based on the source domain invariant feature, the target domain invariant feature, the source domain private feature and the target domain private feature; calculating a life prediction loss based on the predicted life and the actual life of the rotating machinery; constructing a loss function according to the life prediction loss, the first distribution difference loss, the second distribution difference loss and the third distribution difference loss; the calculation formulas of the first distribution difference loss, the second distribution difference loss and the third distribution difference loss are as follows: wherein, represent source domain features, including the source domain private features and the source domain invariant features, represent target domain features, including the target domain private features and the target domain invariant features, represent the number of samples corresponding to the source domain data, represent the number of samples corresponding to the target domain data, H is a reproducing Hilbert space, while represent and a mapping on H .

2. The method of claim 1, wherein, the constructing of the degradation mechanism space state model according to the source domain invariant feature and the target domain invariant feature comprises: simulating the degradation process of the rotating machinery by using a double exponential function according to the source domain invariant feature and the target domain invariant feature to obtain the degradation mechanism space state model.

3. The method of claim 1, wherein, the determining of the initial parameters of the degradation mechanism space state model comprises: solving the initial parameters of the degradation mechanism space state model by using a maximum likelihood algorithm.

4. The method of claim 1, wherein, the step-by-step estimation of the parameters of the degradation mechanism space state model based on the initial parameters by using a particle filtering algorithm to obtain the degradation mechanism space state model after parameter estimation comprises: determining a parameter update formula by using the particle filtering algorithm based on the initial parameters; estimating the parameters of the degradation mechanism space state model step by step according to the parameter update formula to obtain the degradation mechanism space state model after parameter estimation.

5. The method of claim 1, wherein, the method further comprises: obtaining target domain prediction data of the rotating machinery; inputting the target domain prediction data into the preset life prediction model for life prediction to obtain a life prediction result of the rotating machinery.

6. A small sample based degradation knowledge and data fusion driven transfer decoupling apparatus, characterized in that, comprise: an acquisition unit, configured to acquire source domain data and target domain data of rotating machinery, and an initial life prediction model, wherein the initial life prediction model comprises a target domain private feature extractor, a source domain private feature extractor, a domain invariant feature extractor, a domain invariant feature decoder, a domain mixed feature decoder and a life predictor; a determination unit, configured to determine source domain invariant features and source domain private features corresponding to the source domain data, target domain invariant features and target domain private features corresponding to the target domain data, and a predicted life of the rotating machinery based on the source domain data and the target domain data and the initial life prediction model; a first construction unit, configured to construct a loss function according to the source domain invariant features, the target domain invariant features, the source domain private features, the target domain private features and the predicted life; a second construction unit, configured to construct a degradation mechanism space state model according to the source domain invariant features and the target domain invariant features, and determine initial parameters of the degradation mechanism space state model; An estimation unit is configured to estimate parameters of the degradation mechanism space state model based on the initial parameters by using a particle filter algorithm to obtain a parameter-estimated degradation mechanism space state model; A training unit is configured to guide a weight updating process of the initial life prediction model by using the parameter-estimated degradation mechanism space state model when the initial life prediction model continuously updates the weights based on the loss function, and output a preset life prediction model when a preset iteration number is reached. The determination unit is specifically configured to extract source domain invariant original features corresponding to the source domain data and target domain invariant original features corresponding to the target domain data by using the domain invariant feature extractor; The target domain private feature extractor and the source domain private feature extractor are used to extract source domain private original features corresponding to the source domain data and target domain private original features corresponding to the target domain data, respectively; the domain invariant feature decoder is used to decode the source domain invariant original features and the target domain invariant original features, respectively, to obtain the source domain invariant features and the target domain invariant features, and the domain mixed feature decoder is used to decode the source domain private original features and the target domain private original features, respectively, to obtain the source domain private features and the target domain private features; and the source domain invariant features and the source domain private features are input into the life predictor for prediction to obtain the predicted life of the rotating machinery. The first construction unit is specifically configured to calculate a first distribution difference loss between the source domain private features and the source domain invariant features, a second distribution difference loss between the target domain private features and the target domain invariant features, and a third distribution difference loss between the source domain invariant features and the target domain invariant features based on the source domain invariant features, the target domain invariant features, the source domain private features and the target domain private features; calculate a life prediction loss based on the predicted life and the actual life of the rotating machinery; and construct a loss function according to the life prediction loss, the first distribution difference loss, the second distribution difference loss and the third distribution difference loss. The calculation formulas of the first distribution difference loss, the second distribution difference loss and the third distribution difference loss are as follows: wherein, represents source domain features, including the source domain private features and the source domain invariant features, represents target domain features, including the target domain private features and the target domain invariant features, represents the number of samples corresponding to the source domain data, represents the number of samples corresponding to the target domain data, H is a reproducing Hilbert space, while represents and a mapping on H .

7. A storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to implement the method in any one of claims 1 to 5.

8. An electronic device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, The processor executes the computer program to implement the method in any one of claims 1 to 5.

Images