Method, system and equipment for predicting residual life of gearbox bearing of motor train unit under incomplete data condition

Abstract

The invention discloses a motor train unit gearbox bearing residual life prediction method, system and device under the condition of incomplete data. The method comprises the following steps: extracting degradation characteristics of multi-source domain full-life degradation data and target domain non-full-life degradation data; designing a multi-domain local adversarial learning strategy, realizing degradation stage alignment and distribution characteristic alignment of a source domain full-life degradation characteristic and a target domain non-full-life degradation characteristic, accurately mining public degradation information of the two types of degradation characteristics, and further constructing a predictor with high precision and strong generalization ability; and taking the total training loss function as an optimization target to obtain a trained multi-domain local adversarial learning residual life prediction model, thereby realizing accurate prediction of the residual life of the bearing. According to the method, effective feature alignment of the source domain full-life degradation data and the target domain non-full-life degradation data with different degradation stages and different distribution characteristics can be realized, so that the accuracy and generalization of prediction of the residual life of the gearbox bearing of the motor train unit under the condition of incomplete data are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing life prediction, and in particular to a method, system and device for predicting the remaining life of a gearbox bearing of an EMU under the condition of incomplete data. Background Art

[0002] As a key component of the rotating mechanical system, the operating status of the EMU gearbox bearing directly affects the safety, reliability and maintenance economy of the train. Carrying out accurate remaining life prediction is an important prerequisite for predictive maintenance. However, in actual service scenarios, it is difficult to obtain degradation data covering the entire life cycle of the EMU gearbox bearing during operation, and most of them are non-full-life degradation data, which makes the existing remaining life prediction method based on full-life degradation data invalid. Therefore, it is necessary to use full-life degradation data (such as test bench data or line history data) when training the model. However, due to the differences between the accelerated degradation test conditions of the test bench and the actual line operation environment, and the differences in the working conditions of the line history full-life degradation data of different models and operating environments, the test bench full-life degradation data, the line history full-life degradation data and the line real-time degradation data show significant inconsistency in the degradation stage and distribution characteristics, which makes it easy to induce negative transfer when using the existing transfer learning method to directly align features, reduce the prediction accuracy and generalization ability of the model, and affect the reliability of the intelligent operation and maintenance system.

[0003] Therefore, the existing remaining life prediction methods have the following technical limitations: 1) The existing methods generally rely on complete full-life degradation data during model training, and are difficult to be compatible with full-life degradation data and non-full-life degradation data with different degradation stages and distribution characteristics, resulting in limited accuracy and generalization of remaining life prediction, which seriously affects its applicability in real service scenarios; 2) The adopted transfer learning method only focuses on solving the feature alignment problem caused by differences in distribution characteristics, fails to effectively deal with the inconsistency of degradation stages caused by differences in data integrity, and ignores some valuable degradation information.

[0004] In this context, there is an urgent need to develop a method for predicting the remaining life of EMU gearbox bearings under incomplete data conditions, which can fully mine the common degradation information in non-full-life degradation data and full-life degradation data, so as to achieve high-precision and strongly generalized remaining life prediction of railway vehicle gearbox bearings. Summary of the invention

[0005] The present invention provides a method, system and device for predicting the remaining life of a gearbox bearing of an EMU under the condition of incomplete data, so as to solve the above technical problems existing in the existing remaining life prediction method.

[0006] According to the first aspect, an embodiment provides a method for predicting the remaining life of a gearbox bearing of an EMU under the condition of incomplete data, the method comprising:

[0007] The full-life vibration signals and non-full-life vibration signals of the gearbox bearings of the EMU under multiple operating conditions are collected and preprocessed, and divided into multi-source domain data sets and target domain data sets according to the full-life degradation conditions;

[0008] Based on the feature extraction network, the degradation features of the full-life degradation data of each source domain and the non-full-life degradation data of the target domain are extracted to obtain the full-life degradation features of each source domain and the non-full-life degradation features of the target domain;

[0009] A multi-domain local adversarial learning strategy is designed. By introducing a weight domain discriminator and an adversarial domain discriminator, the degradation stage and distribution characteristics of the full-lifetime degradation characteristics of each source domain and the non-full-lifetime degradation characteristics of the target domain are aligned, and the common degradation information of both the source domain and the target domain is mined.

[0010] The predictor is constructed by using the full-life degradation characteristics of each source domain and the corresponding true labels of remaining life that are integrated with public degradation information, using supervised learning method;

[0011] Taking the total training loss function as the optimization target, the network parameters of the feature extraction network, weight domain discriminator, adversarial domain discriminator and predictor are adjusted until a trained multi-domain local adversarial learning remaining life prediction model is obtained;

[0012] The trained multi-domain local adversarial learning remaining life prediction model is used to test the non-full life degradation data of the target domain to obtain the remaining life results.

[0013] Furthermore, the full-life vibration signals and non-full-life vibration signals of the EMU gearbox bearings under multiple operating conditions are collected and preprocessed, and divided into multi-source domain data sets and target domain data sets according to the full-life degradation conditions, specifically including:

[0014] The full-life / part-full-life vibration signals of the gearbox bearings of the EMU under multiple operating conditions are collected at fixed sampling intervals, and the vibration signals obtained each time are taken as a sample;

[0015] Z-Score standardization was used to process each sample to eliminate the differences in statistical characteristics at different time scales;

[0016] The preprocessed samples are used to form a full-life / non-full-life degradation dataset, and are divided into multi-source domain datasets according to the full-life degradation conditions. and the target domain dataset D T , where N S Indicates the number of source domains;

[0017] The multi-source domain dataset contains the full-life degradation data of the test bench and the historical full-life degradation data of the circuit under different operating conditions, covering the complete degradation process from the beginning of early faults to complete failure. Denote the full-life degradation dataset of the j-th source domain as representing the i-th source domain sample, and is the true remaining life label corresponding to it;

[0018] The target domain dataset includes the real-time degradation data of the circuit, only covering the degradation process from the beginning of early faults to the current moment, lacking degradation labels. Denote the non-full-life degradation dataset of the target domain as representing the i-th target domain sample, n T is the target domain sample.

[0019] Furthermore, based on the feature extraction network, the degradation features of the full-life degradation data of each source domain and the non-full-life degradation data of the target domain are extracted to obtain the full-life degradation features of each source domain and the non-full-life degradation features of the target domain, specifically including:

[0020] Select N S Transformer encoders with the same structure but independent parameters as feature extractors; each Transformer encoder includes an input embedding layer, positional encoding, multi-head self-attention mechanism layer, feed-forward neural network, and layer normalization, and has the ability to extract multi-scale temporal features;

[0021] For each source domain j ∈ {1, 2,..., N S}, N S represents the number of source domains. The corresponding full-life degradation data of the source domain and the non-full-life degradation data of the target domain are input into the j-th Transformer encoder in parallel to automatically learn the degradation laws of the source domain and target domain data, thereby obtaining the degradation features of the two types of data respectively. The specific feature extraction process is expressed as:

[0022]

[0023] where Transformer j (·) represents the j-th Transformer encoder, and are the degradation features extracted from the source domain and the target domain under the j-th Transformer encoder respectively.

[0024] Furthermore, the degradation stage alignment specifically includes:

[0025] Evaluate the matching degree of the full - life degradation characteristics of each source domain and the non - full - life degradation characteristics of the target domain in the degradation stage through the weight domain discriminator, and generate transferable weights to specifically screen out the source domain full - life degradation characteristics that are most relevant to the current degradation stage of the target domain non - full - life degradation characteristics. The specific steps are as follows:

[0026] For the non - full - life degradation characteristics of the target domain and N S full - life degradation characteristics of the source domain, construct N S weight domain discriminators;

[0027] During the training process, label each source domain full - life degradation characteristic as the positive class and the target domain non - full - life degradation characteristic as the negative class, and use the binary cross - entropy loss function to supervise the training of the corresponding weight domain discriminator. The loss function of the j - th weight domain discriminator is defined as follows:

[0028]

[0029] Among them, D j (·) represents the j - th weight domain discriminator; E(·) represents the expected value of the distribution function; represents the j - th source domain full - life degradation characteristic obeys its corresponding distribution p(S); represents the target domain non - full - life degradation characteristic obeys its corresponding distribution p(T);

[0030] In each training batch, use the constructed weight domain discriminator to calculate the confidence degrees of each source domain full - life degradation characteristic and the target domain non - full - life degradation characteristic belonging to their respective domains, and convert the confidence degrees of the two types of degradation characteristics into transferable weights to provide a quantitative basis for the alignment of the degradation stages, which is specifically expressed as:

[0031]

[0032]

[0033] Among them, is the confidence degree that the j - th source domain full - life degradation characteristic is judged to be the source domain; is the confidence degree that the target domain non - full - life degradation characteristic is judged to be the target domain; n Y is the target domain sample; ε is a constant; the transferable weight is always between (0, 1). The closer it is to 1, the closer the degradation stages of the source domain full - life degradation characteristic and the target domain non - full - life degradation characteristic are. At this time, the source domain full - life degradation characteristic has a higher transfer value. On the contrary, it indicates that there are significant differences in the degradation stages of the two;

[0034] For each source domain j∈{1,2,…,N S},N S represents the number of source domains, performs L2 regularization on the transferable weights, and the regularization term is defined as follows:

[0035]

[0036] Among them, ⊙ represents element-by-element multiplication, is the total number of samples in the source domain;

[0037] Based on the transferable weights after L2 regularization, the full-life degradation features of each source domain are targeted for screening, the full-life degradation features of the source domain that are in the same degradation stage as the non-full-life degradation features of the target domain are strengthened, and the full-life degradation features of the source domain that are in different degradation stages from the non-full-life degradation features of the target domain are suppressed, thereby achieving alignment of the degradation stages. The specific calculation process is as follows:

[0038]

[0039] in, is the transferable weight after L2 regularization of the jth.

[0040] Furthermore, the distribution characteristic alignment specifically includes:

[0041] The adversarial domain discriminator is introduced to map the full-life degradation features of each source domain and the non-full-life degradation features of the target domain after the degradation stage is aligned to the same feature space, so as to effectively transfer and merge the two types of degradation features under different distribution characteristics, thereby mining the common degradation information of both the source domain and the target domain. The specific steps include:

[0042] Build N S The adversarial domain discriminators discriminate the source domain degraded features and the target domain degraded features after alignment in the degradation stage, and calculate the discrimination loss of each adversarial domain discriminator based on the binary cross entropy loss function:

[0043]

[0044] in, represents the jth adversarial domain discriminator; E(·) represents the expected value of the distribution function; represents the degradation characteristics of the source domain over the entire life after the jth degradation stage is aligned Obey its corresponding distribution p(S); Indicates the non-full life degradation characteristics of the target domain Obey its corresponding distribution p(T);

[0045] Meanwhile, a gradient reversal layer is introduced between the feature extractor and the adversarial domain discriminator to achieve adversarial learning, prompting the feature extractor to learn the common degradation information of each source domain and the target domain, so as to further achieve the feature alignment of each source domain and the target domain.

[0046] Furthermore, using the full-life degradation features of each source domain that fuse the common degradation information and the corresponding true remaining life labels, a predictor is constructed in a supervised learning manner, specifically including:

[0047] Construct N S predictors; for each source domain j ∈ {1, 2, …, N S}, where N S represents the number of source domains, input the degradation features after fusing the common degradation information into the corresponding predictor to obtain the predicted remaining life value; then use the mean square error loss function to calculate the deviation between the predicted remaining life value and the true remaining life label as the prediction loss of each predictor. The specific formula is as follows:

[0048]

[0049] where and respectively represent the predicted remaining life value and the true label value of the j-th predictor, is the total number of source domain samples.

[0050] Furthermore, taking the total training loss function as the optimization objective, adjust the network parameters of the feature extraction network, the weight domain discriminator, the adversarial domain discriminator and the predictor until a trained multi-domain local adversarial learning remaining life prediction model is obtained, specifically including:

[0051] By fusing the losses of N S weight domain discriminators, the L2 regularization loss, the adversarial domain discriminator loss and the predictor loss in the model, obtain the total training loss function:

[0052]

[0053] where λ, β and γ are the hyperparameters of the weight domain discriminator loss the L2 regularization loss the adversarial domain discriminator loss respectively, is the predictor loss; for each source domain j ∈ {1, 2, …, N S}, where N S represents the number of source domains;

[0054] Taking the minimization of the total training loss function as the optimization objective, use the stochastic gradient descent algorithm to iteratively adjust the parameters of the remaining life prediction model until the training error converges to obtain a trained prediction model.

[0055] Furthermore, the trained multi-domain local adversarial learning remaining life prediction model is used to test the non-full life degradation data of the target domain to obtain the remaining life results, including:

[0056] The non-full-life degradation data of the target domain is input into the trained multi-domain local adversarial learning remaining life prediction model to obtain N S The output of the predictor Take the mean as the final remaining life result:

[0057]

[0058] in, is the output of the jth predictor, j∈{1,2,…,N s}.

[0059] According to the second aspect, an embodiment provides a system for predicting the remaining life of a gearbox bearing of an EMU under the condition of incomplete data, the system comprising:

[0060] The data acquisition and preprocessing module is used to collect the full-life vibration signals and non-full-life vibration signals of the EMU gearbox bearings under multiple operating conditions and perform preprocessing, and divide them into multi-source domain data sets and target domain data sets according to the full-life degradation conditions;

[0061] A feature extraction module is used to extract degradation features of the full-life degradation data of each source domain and the non-full-life degradation data of the target domain based on the feature extraction network, so as to obtain the full-life degradation features of each source domain and the non-full-life degradation features of the target domain;

[0062] The multi-domain local adversarial learning module is used to design a multi-domain local adversarial learning strategy. By introducing a weight domain discriminator and an adversarial domain discriminator, the degradation stage alignment and distribution characteristics alignment of the full-lifetime degradation characteristics of each source domain and the non-full-lifetime degradation characteristics of the target domain are achieved, and the common degradation information of both the source domain and the target domain is mined;

[0063] A predictor construction module is used to construct a predictor using supervised learning method by using the full-life degradation characteristics of each source domain and the corresponding true labels of remaining life that integrate the common degradation information;

[0064] A model training module is used to adjust the network parameters of the feature extraction network, the weight domain discriminator, the adversarial domain discriminator and the predictor with the total training loss function as the optimization target until a trained multi-domain local adversarial learning remaining life prediction model is obtained;

[0065] The remaining life prediction module is used to use the trained multi-domain local adversarial learning remaining life prediction model to test the non-full life degradation data of the target domain to obtain the remaining life result.

[0066] According to a third aspect, an embodiment provides an electronic device, the device comprising: a processor and a memory;

[0067] The memory is used to store one or more program instructions;

[0068] The processor is used to run one or more program instructions to execute the steps of a method for predicting the remaining life of a gearbox bearing of an EMU under incomplete data conditions as described in any of the above items.

[0069] According to the fourth aspect, an embodiment provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of a method for predicting the remaining life of a gearbox bearing of a EMU under incomplete data conditions as described in any of the above items are implemented.

[0070] The present invention provides a method, system and device for predicting the remaining life of a gearbox bearing of a train under the condition of incomplete data, which has the following beneficial effects:

[0071] (1) The present invention proposes a method for predicting the remaining life of a gearbox bearing of a train under incomplete data conditions. Existing remaining life prediction methods rely on full life degradation data. The present invention deeply mines and fully activates the potential value of non-full life degradation data, and effectively combines the complementary advantages of non-full life degradation data and full life degradation data to achieve high-precision and strong generalization ability remaining life prediction.

[0072] (2) The present invention proposes a multi-domain local adversarial learning method. Existing feature alignment methods mainly deal with the situation where there are differences in distribution characteristics, and cannot effectively solve the problem of inconsistent degradation stages. The present invention combines degradation stage alignment with distribution characteristic alignment to achieve accurate matching of full-life degradation features and non-full-life degradation features in the local degradation stage, as well as effective transmission and fusion of source domain degradation features and target domain degradation features, thereby accurately mining the common degradation information of source domain full-life degradation data and target domain non-full-life degradation data, and capturing common degradation laws. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 A flowchart of a method for predicting the remaining life of a gearbox bearing of an EMU under incomplete data conditions provided by an embodiment of the present invention;

[0074] Figure 2Flowchart of the multi - domain local adversarial learning algorithm in a method for predicting the remaining life of a train gearbox bearing under incomplete data conditions provided by an embodiment of the present invention;

[0075] Figure 3 Schematic diagram of the alignment of the full - life degradation data of the j - th source domain and the non - full - life degradation data of the target domain in terms of degradation stage and distribution characteristics in a method for predicting the remaining life of a train gearbox bearing under incomplete data conditions provided by an embodiment of the present invention;

[0076] Figure 4 Schematic diagram of the structure principle of the Transformer encoder network in a method for predicting the remaining life of a train gearbox bearing under incomplete data conditions provided by an embodiment of the present invention;

[0077] Figure 5 Schematic diagram of the structure of a system for predicting the remaining life of a train gearbox bearing under incomplete data conditions provided by an embodiment of the present invention;

[0078] Figure 6 Schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. Detailed implementation manners

[0079] The present invention will be further described in detail below in conjunction with the accompanying drawings through specific implementation manners. Similar elements in different implementation manners are labeled with related similar element numbers. In the following implementation manners, many detailed descriptions are provided to enable a better understanding of the present invention. However, those skilled in the art can easily recognize that some of the features can be omitted in different situations, or can be replaced by other elements, materials, or methods. In some cases, some operations related to the present invention are not shown or described in the specification to avoid the core part of the present invention being overwhelmed by excessive descriptions. For those skilled in the art, it is not necessary to describe these related operations in detail, and they can fully understand the related operations based on the description in the specification and general technical knowledge in the field.

[0080] In addition, the features, operations, or characteristics described in the specification can be combined in any appropriate manner to form various implementation manners. At the same time, the steps or actions in the method description can also be reordered or adjusted in an obvious manner by those skilled in the art. Therefore, the various sequences in the specification and drawings are only for clearly describing a certain embodiment and do not mean that they are necessary sequences, unless it is stated that a certain sequence must be followed.

[0081] As a core component in the rotating machinery of EMUs, the gearbox bearings' operating status is directly related to the vehicle's safety and maintenance costs. Therefore, conducting effective remaining life prediction for EMU gearbox bearings is the key to achieving predictive maintenance. However, for the actually-serviced EMU gearbox bearings, their degradation data are non-full-life degradation data, that is, only covering part of the degradation stage, which is difficult to support the prediction of the remaining life. Therefore, it is necessary to introduce full-life degradation data (such as test bench data or line historical data) to participate in model training. However, due to the differences in vehicle types and operating conditions, the degradation stages of non-full-life degradation data and full-life degradation data are inconsistent and there are significant differences in distribution characteristics. If the existing transfer learning methods are still used for feature alignment, it is easy to cause negative transfer, which instead reduces the prediction effect and affects the reliability of intelligent operation and maintenance.

[0082] To address the above problems, a method for predicting the remaining life of EMU gearbox bearings under incomplete data conditions provided by the first embodiment of the present invention aims to solve the problem that there are significant differences in both the degradation stage and distribution characteristics between the source domain full-life degradation data and the target domain non-full-life degradation data. It mainly includes six steps, mainly including data collection and preprocessing, feature extraction, multi-domain local adversarial learning, predictor construction, model training, and remaining life prediction. The following will be combined with Figure 1 to be described in detail.

[0083] S1. In view of the characteristics of time-varying operating conditions of EMU gearbox bearings during operation, collect the full-life vibration signals and non-full-life vibration signals under multiple operating conditions, and after preprocessing, form a full-life / non-full-life degradation data set. Among them, the full-life degradation data of the gearbox bearings include the test bench full-life degradation data and line historical full-life degradation data under different operating conditions, as a multi-source domain data set; the non-full-life degradation data of the gearbox bearings is the line real-time degradation data, as the target domain data set.

[0084] Specifically, step S1 specifically includes the following steps:

[0085] S1.1: Collect the full-life / non-full-life vibration signals of EMU gearbox bearings under multiple operating conditions at a fixed sampling interval. Each vibration signal obtained by sampling is used as a sample, and is stored as a CSV format file according to the sampling time. The file name is represented by a continuous digital serial number indicating the sampling order, that is, 1.csv, 2.csv,..., n.csv, where n is the total number of samples.

[0086] S1.2: Since the EMU gearbox bearings are in a non-stationary state for a long time during actual service, use Z-Score standardization to process each sample to eliminate the statistical characteristic differences on different time scales.

[0087] S1.3: Construct a full-life / non-full-life degradation dataset from several segments of samples processed in S1.2, and divide it into multi-source domain datasets and a target domain dataset D according to the full-life degradation situation. and the target domain dataset D t , N s represents the number of source domains;

[0088] Among them, the multi-source domain datasets include the full-life degradation data of test benches under different operating conditions and the historical full-life degradation data of lines, covering the complete degradation process from the start of early faults to complete failure. Denote the full-life degradation dataset of the j-th source domain as represents the i-th source domain sample, and

[0089] is its corresponding remaining life true label. The target domain dataset is the real-time degradation data of the line, only covering the degradation process from the start of early faults to the current moment, lacking degradation labels. Denote the non-full-life degradation dataset of the target domain as

[0090] S2. Extract degradation features from the full-life degradation data of each source domain and the non-full-life degradation data of the target domain based on the Transformer encoder network to obtain the full-life degradation features of each source domain and the non-full-life degradation features of the target domain.

[0091] Specifically, step S2 specifically includes the following steps:

[0092] S2.1: In order to mine sufficient feature information from the degradation data of the EMU gearbox bearings, select N s Transformer encoders with the same structure but independent parameters as feature extractors. For each source domain j ∈ {1, 2,..., N S}}, input the corresponding full-life degradation data of the source domain and the non-full-life degradation data of the target domain into the j-th Transformer encoder in parallel, automatically learn the degradation laws of the source domain and target domain data, and thus obtain the degradation features of the two types of data respectively. The specific feature extraction process can be expressed as:

[0093]

[0094] where, Transformer J (·) represents the j-th Transformer encoder, and They are the degradation features extracted from the source domain and the target domain under the jth Transformer encoder. Each Transformer encoder includes an input embedding layer, a position encoding, a multi-head self-attention mechanism layer, a feedforward neural network, and layer normalization. It has the ability to extract multi-scale time series features, providing rich feature information for subsequent feature alignment and remaining life prediction.

[0095] S3, designs a multi-domain local adversarial learning strategy to achieve accurate matching and effective migration of the full-lifetime degradation features of each source domain and the non-full-lifetime degradation features of the target domain through degradation stage alignment and distribution characteristic alignment.

[0096] The degradation stage alignment module uses a weighted domain discriminator to evaluate the matching degree between the full-life degradation features of each source domain and the non-full-life degradation features of the target domain in terms of degradation stages, and generates transferable weights to specifically select the full-life degradation features of the source domain that are most relevant to the current degradation stage of the non-full-life degradation features of the target domain.

[0097] The distribution characteristic alignment module introduces an adversarial domain discriminator to map the full-lifetime degradation features of each source domain and the non-full-lifetime degradation features of the target domain after the degradation stage is aligned to the same feature space, promoting the effective transmission and fusion of the two types of degradation features under different distribution characteristics, thereby mining the common degradation information of both the source domains and the target domains.

[0098] Specifically, step S3 specifically includes the following steps:

[0099] S3.1: For the non-full-life degradation characteristics of the target domain and from N s The source domain lifetime degradation characteristics are constructed s A weight domain discriminator consisting of two fully connected layers. During the training process, each source domain full-life degradation feature is labeled as a positive class, and the target domain non-full-life degradation feature is labeled as a negative class, and the corresponding weight domain discriminator is supervised and trained using a binary cross entropy loss function. Specifically, the loss function of the jth weight domain discriminator is defined as follows:

[0100]

[0101] Among them, D j (·) represents the jth weight domain discriminator; E(·) represents the expected value of the distribution function; Represents the lifetime degradation characteristics of the jth source domain Obey its corresponding distribution p(S); Indicates the non-full life degradation characteristics of the target domain Obey its corresponding distribution p(T).

[0102] S3.2: To further improve the transfer effectiveness of the full-life degradation features in the source domain to the non-full-life degradation features in the target domain, in each training batch, the weight domain discriminator constructed in S3.1 is used to calculate the confidence levels of each source domain's full-life degradation features and the target domain's non-full-life degradation features belonging to their respective domains, and the confidence levels of the two types of degradation features are converted into transferable weights, providing a quantitative basis for the alignment of degradation stages. Specifically, it can be expressed as follows:

[0103]

[0104] Among them, is the confidence level that the j-th source domain's full-life degradation feature is judged to be in the source domain; is the confidence level that the target domain's non-full-life degradation feature is judged to be in the target domain (relative to the j-th weight domain discriminator). ε = 10 -6 , and the transferable weight is always between (0, 1). The closer it is to 1, the closer the degradation stages of the source domain's full-life degradation feature and the target domain's non-full-life degradation feature are. At this time, the source domain's full-life degradation feature has a higher transfer value; on the contrary, there are significant differences in their degradation stages.

[0105] S3.3: For each source domain j ∈ {1, 2,..., N S}, L2 regularization is performed on the transferable weights. The regularization term is defined as follows:

[0106]

[0107] Among them, ⊙ represents element-wise multiplication.

[0108] S3.4: Based on the L2-regularized transferable weights, targeted screening is performed on the full-life degradation features of each source domain, strengthening the full-life degradation features of the source domain that are in the same degradation stage as the non-full-life degradation features of the target domain, and suppressing the full-life degradation features of the source domain that are in different degradation stages from the non-full-life degradation features of the target domain, thereby achieving the alignment of degradation stages. The specific calculation process is as follows:

[0109]

[0110] Among them, is the j-th L2-regularized transferable weight.

[0111] S3.5: To reduce the differences in distribution characteristics between the full-life degradation features of each source domain after the alignment of degradation stages and the non-full-life degradation features of the target domain, N S adversarial domain discriminators composed of two fully connected layers are constructed to discriminate the degradation features from the source domain and the target domain respectively, and the discrimination losses of each adversarial domain discriminator are calculated based on the binary cross-entropy loss function:

[0112]

[0113] Among them, represents the j-th adversarial domain discriminator; E(·) represents the expected value of the distribution function; represents the full-life degradation characteristics of the source domain after alignment in the j-th degradation stage obeys its corresponding distribution p(S); represents the non-full-life degradation characteristics of the target domain obeys its corresponding distribution p(T).

[0114] Meanwhile, a gradient reversal layer is introduced between the feature extractor and the adversarial domain discriminator to achieve adversarial learning, prompting the feature extractor to learn the common degradation information of each source domain and the target domain, so as to further achieve the feature alignment of each source domain and the target domain.

[0115] S4. Use the full-life degradation characteristics of each source domain that fuse the common degradation information, and combine the corresponding true remaining life labels, and construct a predictor in a supervised learning manner.

[0116] Specifically, step S4 specifically includes the following steps:

[0117] S4.1: Construct N S predictors composed of two fully connected layers. For each source domain, input the degradation characteristics after fusing the common degradation information into the corresponding predictor to obtain the predicted remaining life value; then use the mean square error loss function to calculate the deviation between the predicted remaining life value and the true remaining life label as the prediction loss of each predictor:

[0118]

[0119] Among them, and respectively represent the predicted remaining life value and the true label value of the j-th predictor.

[0120] S5. After executing steps S2 to S4, obtain a multi-domain local adversarial learning remaining life prediction model. During the training process, calculate the total training loss function, and use the optimizer to iteratively adjust the parameters of the multi-domain local adversarial learning remaining life prediction model to obtain a trained prediction model.

[0121] Specifically, step S5 specifically includes the following steps:

[0122] S5.1: After executing the above steps S2 to S4, accordingly obtain a multi-domain local adversarial learning remaining life prediction model. By fusing N in the model SThe weighted domain discriminator loss, L2 regularization loss, adversarial domain discriminator loss and predictor loss are used to obtain the total training loss function:

[0123]

[0124] Among them, λ, β, and γ are the hyperparameters of the weighted domain discriminator loss, L2 regularization loss, and adversarial domain discriminator loss, respectively.

[0125] S5.2: Taking the minimization of the total training loss function as the optimization goal, the parameters of the remaining life prediction model are iteratively adjusted using the stochastic gradient descent algorithm until the training error converges to obtain a trained prediction model.

[0126] S6, use the trained multi-domain local adversarial learning remaining life prediction model to test the target domain non-full life degradation data to obtain the remaining life result.

[0127] Specifically, step S6 specifically includes the following steps:

[0128] S6.1: Input the target domain non-full life degradation data into the trained multi-domain local adversarial learning remaining life prediction model to obtain N S The output of the predictor Take the mean as the final remaining life result:

[0129]

[0130] Based on the above content, Figure 2 As shown, the multi-domain local adversarial learning strategy in the embodiment of the present invention realizes the accurate matching and effective migration of the full-life degradation characteristics of each source domain and the non-full-life degradation characteristics of the target domain through the alignment of degradation stages and distribution characteristics. The degradation stage alignment module evaluates the matching degree of the full-life degradation characteristics of each source domain and the non-full-life degradation characteristics of the target domain in terms of degradation stage through the weight domain discriminator, and generates transferable weights, and specifically selects the full-life degradation characteristics of the source domain that are most relevant to the current degradation stage of the non-full-life degradation characteristics of the target domain; the distribution characteristic alignment module introduces the adversarial domain discriminator, and maps the full-life degradation characteristics of each source domain and the non-full-life degradation characteristics of the target domain after the degradation stage alignment to the same feature space, so as to promote the effective transmission and fusion of the two types of degradation characteristics under different distribution characteristics, thereby mining the common degradation information of both the source domain and the target domain.

[0131] like Figure 3As shown, in the embodiment of the present invention, after the degradation stage is aligned, a large weight is assigned to the source domain full-life degradation feature that is in a similar degradation stage to the target domain non-full-life degradation feature, while a small weight is assigned to the source domain full-life degradation feature that is not in a similar degradation stage, thereby effectively alleviating the global feature mismatch problem and improving the application value of the source domain full-life degradation feature in life prediction. After the distribution characteristics are aligned, the peaks of the probability density curves of the source domain full-life degradation feature and the target domain non-full-life degradation feature tend to be consistent and the overlapping area is large, indicating that the distribution characteristics of the two tend to be consistent.

[0132] like Figure 4 As shown, the Transformer encoder network in the embodiment of the present invention includes an input embedding layer, a position encoding, a multi-head self-attention mechanism layer, a feedforward neural network, and layer normalization. The full-life degradation data of each source domain and the non-full-life degradation data of the target domain are input into the network in parallel, and the extracted degradation features can provide rich feature information for subsequent feature alignment and remaining life prediction.

[0133] Corresponding to the above-disclosed method for predicting the remaining life of a gearbox bearing of an EMU under incomplete data conditions, an embodiment of the present invention further discloses a system for predicting the remaining life of a gearbox bearing of an EMU under incomplete data conditions, such as Figure 5 As shown, it specifically includes:

[0134] The data acquisition and preprocessing module is used to collect the full-life vibration signals and non-full-life vibration signals of the EMU gearbox bearings under multiple operating conditions and perform preprocessing, and divide them into multi-source domain data sets and target domain data sets according to the full-life degradation conditions;

[0135] A feature extraction module is used to extract degradation features of the full-life degradation data of each source domain and the non-full-life degradation data of the target domain based on the feature extraction network, so as to obtain the full-life degradation features of each source domain and the non-full-life degradation features of the target domain;

[0136] The multi-domain local adversarial learning module is used to design a multi-domain local adversarial learning strategy. By introducing a weight domain discriminator and an adversarial domain discriminator, the degradation stage alignment and distribution characteristics alignment of the full-lifetime degradation characteristics of each source domain and the non-full-lifetime degradation characteristics of the target domain are achieved, and the common degradation information of both the source domain and the target domain is mined;

[0137] A predictor construction module is used to construct a predictor using supervised learning method by using the full-life degradation characteristics of each source domain and the corresponding true labels of remaining life that integrate the common degradation information;

[0138] A model training module, which is used to adjust the network parameters of the feature extraction network, the weight domain discriminator, the adversarial domain discriminator, and the predictor with the total training loss function as the optimization objective until a trained multi-domain local adversarial learning remaining life prediction model is obtained;

[0139] A remaining life prediction module, which is used to test the non-full-life degradation data of the target domain by using the trained multi-domain local adversarial learning remaining life prediction model to obtain the remaining life result.

[0140] It should be noted that the detailed description of a remaining life prediction system for the EMU gearbox bearing under the condition of incomplete data provided in the embodiment of the present invention can refer to the relevant description of a remaining life prediction method for the EMU gearbox bearing under the condition of incomplete data provided in the embodiment of the present invention, which will not be elaborated here.

[0141] In addition, the embodiment of the present invention also provides an electronic device, as Figure 6 shown, the device includes: a processor and a memory; the memory is used to store one or more program instructions; the processor is used to run one or more program instructions to execute the steps of a remaining life prediction method for the EMU gearbox bearing under the condition of incomplete data as described in any one of the above.

[0142] It should be noted that the detailed description of an electronic device provided in the embodiment of the present invention can refer to the relevant description of a remaining life prediction method for the EMU gearbox bearing under the condition of incomplete data provided in the embodiment of the present application, which will not be elaborated here.

[0143] In addition, the embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of a remaining life prediction method for the EMU gearbox bearing under the condition of incomplete data as described in any one of the above.

[0144] It should be noted that the detailed description of a computer-readable storage medium provided in the embodiment of the present invention can refer to the relevant description of a remaining life prediction method for the EMU gearbox bearing under the condition of incomplete data provided in the embodiment of the present application, which will not be elaborated here.

[0145] Those skilled in the art can understand that all or part of the functions of the various methods in the above embodiments can be implemented in a hardware manner or in a computer program manner. When all or part of the functions in the above embodiments are implemented in a computer program manner, the program can be stored in a computer-readable storage medium, and the storage medium can include: read-only memory, random access memory, magnetic disk, optical disk, hard disk, etc. The above functions can be realized by a computer executing the program. For example, the program is stored in the memory of the device, and when the processor executes the program in the memory, the above all or part of the functions can be realized. In addition, when all or part of the functions in the above embodiments are implemented in a computer program manner, the program can also be stored in a storage medium such as a server, another computer, magnetic disk, optical disk, flash drive or mobile hard disk, and saved to the memory of the local device by downloading or copying, or the system of the local device is updated. When the processor executes the program in the memory, all or part of the functions in the above embodiments can be realized.

[0146] The above uses specific examples to illustrate the present invention, which is only for helping to understand the present invention and is not intended to limit the present invention. For those skilled in the art of the present invention, according to the idea of the present invention, several simple deductions, deformations or substitutions can also be made.

Claims

1. A method for predicting the remaining life of the gearbox bearing of the multiple unit under the condition of incomplete data, characterized in that The method includes: Collecting the full-life vibration signals and non-full-life vibration signals of the EMU gearbox bearings under multiple operating conditions and performing preprocessing, and dividing them into a multi-source domain dataset and a target domain dataset according to the full-life degradation situation; Extracting degradation features from the full-life degradation data of each source domain and the non-full-life degradation data of the target domain based on a feature extraction network to obtain the full-life degradation features of each source domain and the non-full-life degradation features of the target domain; Designing a multi-domain local adversarial learning strategy, and by introducing a weight domain discriminator and an adversarial domain discriminator, achieving the alignment of the degradation stages and the distribution characteristics of the full-life degradation features of each source domain and the non-full-life degradation features of the target domain, and mining the common degradation information that combines each source domain and the target domain; Using the full-life degradation features of each source domain that fuse the common degradation information and the corresponding remaining life true labels, and constructing a predictor in a supervised learning manner; Taking the total training loss function as the optimization objective, adjusting the network parameters of the feature extraction network, the weight domain discriminator, the adversarial domain discriminator and the predictor until a trained multi-domain local adversarial learning remaining life prediction model is obtained; Using the trained multi-domain local adversarial learning remaining life prediction model to test the non-full-life degradation data of the target domain to obtain the remaining life results.

2. The method for predicting the remaining life of the EMU gearbox bearing under the condition of incomplete data as described in claim 1, wherein Collecting the full-life vibration signals and non-full-life vibration signals of the EMU gearbox bearings under multiple operating conditions and performing preprocessing, and dividing them into a multi-source domain dataset and a target domain dataset according to the full-life degradation situation, specifically including: Collecting the full-life / non-full-life vibration signals of the EMU gearbox bearings under multiple operating conditions at a fixed sampling interval, and taking the vibration signal obtained each time as a sample; Processing each sample using Z-Score standardization to eliminate the statistical characteristic differences on different time scales; Construct a full-life / non-full-life degradation dataset from several preprocessed segments of samples, and divide it into multi-source domain datasets according to the full-life degradation situation and the target domain dataset D T , where N S represents the number of source domains; The multi-source domain dataset contains the full-life degradation data of the test bench and the historical full-life degradation data of the line under different operating conditions, covering the complete degradation process from the beginning of early faults to complete failure. Denote the full-life degradation dataset of the j-th source domain as which represents the i-th source domain sample, and is the corresponding true remaining life label, and is the total number of source domain samples; The target domain dataset includes real-time line degradation data, which only covers the degradation process from the beginning of early faults to the current moment and lacks degradation labels. Denote the target domain non-full-life degradation dataset as denote the i-th target domain sample, n T is a target domain sample.

3. The method for predicting the remaining life of the gearbox bearing of the multiple unit under the condition of incomplete data according to claim 1, wherein, Extracting degradation features from the full-life degradation data of each source domain and the non-full-life degradation data of the target domain based on a feature extraction network to obtain the full-life degradation features of each source domain and the non-full-life degradation features of the target domain, specifically including: Select N S Transformer encoders with the same structure but independent parameters are used as feature extractors; each Transformer encoder includes an input embedding layer, positional encoding, multi-head self-attention mechanism layer, feed-forward neural network, and layer normalization, and has the ability to extract multi-scale temporal features; For each source domain \(j\in\{1,2,\ldots,N\}\) S}, where \(N\) S represents the number of source domains, the corresponding full-life degradation data of the source domain and the non-full-life degradation data of the target domain are input into the \(j\)-th Transformer encoder in parallel to automatically learn the degradation laws of the source domain and target domain data, so as to obtain the degradation characteristics of the two types of data respectively. The specific feature extraction process is expressed as: Among them, Transformer j (·) represents the j-th Transformer encoder, and are the degraded features extracted from the source domain and the target domain under the j-th Transformer encoder, respectively.

4. The method for predicting the remaining life of the gearbox bearing of the EMU under the condition of incomplete data according to claim 1, characterized in that The specific alignment of the degradation stages includes: Evaluating the matching degree of the full-life degradation features of each source domain and the non-full-life degradation features of the target domain in the degradation stage through a weight domain discriminator, and generating transferable weights, and specifically screening out the source domain full-life degradation features that are most relevant to the current degradation stage of the non-full-life degradation features of the target domain. The specific steps include: For the non-full-life degradation characteristics of the target domain and the full-life degradation characteristics of N S source domains, construct N S weight domain discriminators; During the training process, labeling each source domain full-life degradation feature as a positive class and the non-full-life degradation feature of the target domain as a negative class, and using the binary cross-entropy loss function to supervise the training of the corresponding weight domain discriminator. The loss function of the jth weight domain discriminator is defined as follows: Among them, D j (·) represents the j-th weight domain discriminator; E(·) represents the expected value of the distribution function; represents the j-th source domain full-life degradation feature obeys its corresponding distribution p(S); represents the target domain non-full-life degradation feature obeys its corresponding distribution p(T); In each training batch, using the constructed weight domain discriminator to calculate the confidence levels of the full-life degradation features of each source domain and the non-full-life degradation features of the target domain belonging to their respective domains, and converting the confidence levels of the two types of degradation features into transferable weights to provide a quantitative basis for the alignment of the degradation stages, specifically expressed as: Among them, is the full-life degradation feature of the j-th source domain is the confidence level determined as the source domain; is the non-full-life degradation feature of the target domain is the confidence level determined as the target domain; n T is the target domain sample; ε is a constant; the transferable weight always lies between (0, 1). The closer it is to 1, the closer the degradation stages of the full-life degradation feature of the source domain and the non-full-life degradation feature of the target domain are. At this time, the full-life degradation feature of the source domain has a higher transfer value. On the contrary, it indicates that there are significant differences in the degradation stages of the two. For each source domain \(j\in\{1,2,\ldots,N\}\) S}, where \(N\) S represents the number of source domains, L2 regularization is performed on the transferable weights, and the regularization term is defined as follows: where ⊙ represents element-wise multiplication, is the total number of source domain samples; Based on the transferable weights after L2 regularization, the full-life degradation features of each source domain are targeted for screening, the full-life degradation features of the source domain that are in the same degradation stage as the non-full-life degradation features of the target domain are strengthened, and the full-life degradation features of the source domain that are in different degradation stages from the non-full-life degradation features of the target domain are suppressed, thereby achieving alignment of the degradation stages. The specific calculation process is as follows: Among them, is the transferable weight after the j-th L2 regularization.

5. A method for predicting the remaining life of the gearbox bearing of an EMU under the condition of incomplete data, as claimed in claim 4, wherein The distribution characteristic alignment specifically includes: The adversarial domain discriminator is introduced to map the full-life degradation features of each source domain and the non-full-life degradation features of the target domain after the degradation stage is aligned to the same feature space, so as to effectively transfer and merge the two types of degradation features under different distribution characteristics, thereby mining the common degradation information of both the source domain and the target domain. The specific steps include: Construct N S adversarial domain discriminators, which respectively discriminate the source domain degraded features and the target domain degraded features after alignment in the degradation stage, and calculate the discrimination loss of each adversarial domain discriminator based on the binary cross-entropy loss function: Among them, represents the j-th adversarial domain discriminator; E(·) represents the expected value of the distribution function; represents the source domain full-life degradation feature after alignment in the j-th degradation stage obeys its corresponding distribution p(S); represents the target domain non-full-life degradation feature obeys its corresponding distribution p(T); At the same time, a gradient reversal layer is introduced between the feature extractor and the adversarial domain discriminator to realize adversarial learning, which enables the feature extractor to learn the common degradation information of each source domain and target domain, so as to further achieve feature alignment of each source domain and target domain.

6. A method for predicting the remaining life of the gearbox bearing of the multiple unit under the condition of incomplete data as described in claim 1, characterized in that By using the full-life degradation characteristics of each source domain that integrates public degradation information and the corresponding true labels of remaining life, a predictor is constructed using supervised learning, including: Construct N S predictors; for each source domain j ∈ {1, 2, …, N S}, where N S represents the number of source domains, input the degradation features after fusing the common degradation information into the corresponding predictors to obtain the remaining useful life prediction values; then use the mean square error loss function to calculate the deviation between the remaining useful life prediction values and the true remaining useful life labels as the prediction loss of each predictor. The specific formula is as follows: Among them, and respectively represent the predicted remaining life value and the true label value of the j-th predictor, is the total number of source domain samples.

7. The remaining life prediction method of the EMU gearbox bearing under the condition of incomplete data according to claim 1, characterized in that Taking the total training loss function as the optimization target, the network parameters of the feature extraction network, weight domain discriminator, adversarial domain discriminator and predictor are adjusted until a trained multi-domain local adversarial learning remaining life prediction model is obtained, including: By fusing the N weight domain discriminator losses, L2 regularization losses, adversarial domain discriminator losses, and predictor losses in the fusion model, the total training loss function is obtained: S ​ where λ, β, and γ are the hyperparameters of the weight domain discriminator loss L2 regularization loss adversarial domain discriminator loss respectively, and is the predictor loss; for each source domain j ∈ {1, 2, …, N S}, N S represents the number of source domains; Taking minimization of the total training loss function as the optimization objective, the parameters of the remaining life prediction model are iteratively adjusted using the stochastic gradient descent algorithm until the training error converges to obtain a trained prediction model.

8. A method for predicting the remaining life of the gearbox bearing of an EMU under the condition of incomplete data as described in claim 1, characterized in that The trained multi-domain local adversarial learning remaining life prediction model is used to test the non-full life degradation data of the target domain to obtain the remaining life results, including: Input the non-full-life degradation data of the target domain into the trained multi-domain local adversarial learning remaining useful life prediction model to obtain the output results of N S predictors Calculate the mean as the final remaining useful life result: Among them, is the output result of the j-th predictor, where j ∈ {1, 2, …, N S}.

9. A remaining life prediction system for the gearbox bearings of EMUs under the condition of incomplete data, characterized in that, The system comprises: The data acquisition and preprocessing module is used to collect the full-life vibration signals and non-full-life vibration signals of the EMU gearbox bearings under multiple operating conditions and perform preprocessing, and divide them into multi-source domain data sets and target domain data sets according to the full-life degradation conditions; A feature extraction module is used to extract degradation features of the full-life degradation data of each source domain and the non-full-life degradation data of the target domain based on the feature extraction network, so as to obtain the full-life degradation features of each source domain and the non-full-life degradation features of the target domain; The multi-domain local adversarial learning module is used to design a multi-domain local adversarial learning strategy. By introducing a weight domain discriminator and an adversarial domain discriminator, the degradation stage alignment and distribution characteristics alignment of the full-lifetime degradation characteristics of each source domain and the non-full-lifetime degradation characteristics of the target domain are achieved, and the common degradation information of both the source domain and the target domain is mined; A predictor construction module is used to construct a predictor using supervised learning method by using the full-life degradation characteristics of each source domain and the corresponding true labels of remaining life that integrate the common degradation information; A model training module is used to adjust the network parameters of the feature extraction network, the weight domain discriminator, the adversarial domain discriminator and the predictor with the total training loss function as the optimization target until a trained multi-domain local adversarial learning remaining life prediction model is obtained; The remaining useful life prediction module is used to test the non-full-life degradation data of the target domain by using the trained multi-domain local adversarial learning remaining useful life prediction model to obtain the remaining useful life result.

10. An electronic device, characterized in that, The device includes: a processor and a memory; The memory is used to store one or more program instructions; The processor is used to run one or more program instructions to execute the steps of a method for predicting the remaining useful life of the EMU gearbox bearing under the condition of incomplete data as described in any one of claims 1 to 8.

Images