Multi-stage self-updating vehicle remaining life prediction and maintenance method

Abstract

The invention relates to a multi-stage self-updating vehicle remaining life prediction and maintenance method. Comprising the following steps: acquiring operation data of each part of an equipment vehicle as characteristic parameters; calculating time domain features and frequency domain features in the feature parameters, and summarizing to construct a feature parameter matrix; constructing an anomaly monitoring model based on a slow feature analysis algorithm, and determining a degradation starting point through the anomaly monitoring model; based on an LSTM model and an LSSVM model as base models, constructing an integrated learning model, and training the integrated learning model by using the feature parameter matrix and the initial degradation point; the trained integrated learning model is used to predict the remaining life of each part of the equipment vehicle, the equipment vehicle is maintained according to the prediction result, and the data is labeled after the maintenance is completed; and updating the tagged data according to the similarity measurement index of the integrated learning model. The life prediction model provided by the invention can give consideration to a trained scene and also has good prediction capability for an untrained scene, and the generalization performance and the general performance of the model are improved.

Description

Technical Field

[0001] The present invention belongs to the field of prediction of remaining life and maintenance of equipment, and specifically relates to a method for predicting and maintaining the remaining life of a vehicle with multi-level self-update. Background Art

[0002] With the maturity of intelligent manufacturing technology and the development of emerging technologies, modern equipment has increasingly rich functions and undertakes more complex tasks. Since it is difficult to conduct full-scenario life degradation experiments on the components of equipment vehicles, it is difficult for data samples to cover all scenarios, which poses a severe challenge to the prediction of remaining life and maintenance of equipment vehicles. The existing methods for predicting and maintaining the remaining life of equipment vehicles use the remaining life prediction model constructed from initial data for full-course prediction, or directly input the predicted results after prediction into the update of the model. On the one hand, due to the difficulty of data covering all scenarios, the model is difficult to obtain high-precision results when facing new situations. On the other hand, due to the rough update method, the model is vulnerable to contamination by abnormal situation data, reducing the reliability of the model. In addition, a large amount of common scenario data will crowd out a small amount of special scenario data, reducing the prediction accuracy of the model for special scenarios.

[0003] Therefore, it is necessary to design a method for predicting and maintaining the remaining life with multi-level and multi-mode updates according to the sample data situation. Summary of the Invention

[0004] In view of the above problems, the object of the present invention is to provide a method for predicting and maintaining the remaining life of equipment with multi-level self-update. This method first decomposes the data by slow feature analysis for time series decomposition and determines the degradation starting point using the SPE statistical limit and S 2 statistical limit. Secondly, the decomposed time series data and the degradation starting point are placed into an integrated learning model combining LSTM and LSSVM for analysis. Then, the influence degree of the base model on the overall model is determined using similarity measurement and the prediction result of the remaining life of the equipment is obtained. Finally, after the equipment maintenance is completed, the sample data is labeled, and three operations of non-update, update without expert knowledge, and update after expert knowledge determination are performed according to the sample situation to complete the multi-level self-update of the integrated model and the slow feature weight ratio.

[0005] The technical solution adopted by the present invention to achieve the above object is:

[0006] A method for predicting and maintaining the remaining life of a vehicle with multi-level self-update, comprising the following steps:

[0007] 1) Collect the operation data of each component of the equipment vehicle as characteristic parameters;

[0008] 2) Calculate the time-domain features and frequency-domain features in the characteristic parameters, and summarize and construct a characteristic parameter matrix;

[0009] 3) Build an anomaly monitoring model based on the slow feature analysis algorithm, and determine the starting point of degradation through the anomaly monitoring model;

[0010] 4) Based on the LSTM model and the LSSVM model as the base models, build an ensemble learning model, and use the characteristic parameter matrix and the initial degradation point to train it;

[0011] 5) Use the trained ensemble learning model to predict the remaining life of each component of the equipment vehicle, perform maintenance on the equipment vehicle according to the prediction results, and label the data after the maintenance is completed;

[0012] 6) Update the labeled data according to the similarity measurement index of the ensemble learning model, and then realize the multi-level self-update of the weight ratio of the ensemble learning model and the slow feature analysis algorithm.

[0013] The step 3) includes the following steps:

[0014] 3.1) Use slow feature analysis to decompose the normal operation samples in the same operation cycle in the characteristic parameter matrix and screen out k anomaly-related slow features according to the weight ratio;

[0015] 3.2) Calculate the SPE statistic and the S 2 statistic of the slow feature respectively;

[0016] 3.3) Use the kernel density estimation method to calculate the S 2 and the SPE control limits. When the continuous 5 acquisition points of the two statistics exceed the control limits, it is considered that the equipment has an anomaly, and the starting over-limit point is recorded as the initial degradation point.

[0017] The kernel density estimation method is specifically:

[0018]

[0019] Among them, is the kernel density estimation of y, y is the original data to be estimated, y i is the observed value; n is the number of samples; b is the bandwidth; G(·) is the kernel function.

[0020] The step 4) is specifically:

[0021] Taking the LSTM model and the LSSVM model as the base models of the ensemble learning model, and using spatio-temporal similarity measurement to determine the proportion of the two base models. The spatio-temporal similarity measurement includes spatial similarity measurement and temporal similarity measurement. The values obtained after normalizing the two similarity measurement indicators are used as the proportion of the two base models. The closer the value is to 1, the larger the proportion of the LSSVM model. Among them:

[0022] The spatial similarity measurement is a hybrid similarity measurement index that fuses the cosine of the angle and the Euclidean distance, and is used to characterize the coupling relationship between feature variables;

[0023] The temporal similarity measurement uses the Mahalanobis distance in multi-scale time series as the similarity measurement index, and is used to characterize the temporal characteristics of feature variables after excluding the correlation between feature variables.

[0024] The spatial similarity measurement S1(x h ,x q ) is specifically:

[0025]

[0026] Among them, x h is the historical sample, x q is the query sample, d1(x h ,x q ) is the Euclidean distance between x h and x q , θ1(x h ,x q ) is the cosine of the angle between the two, λ1 is a hybrid weight coefficient between 0 and 1, is an adjustable parameter, d1(x h ,x q ) is specifically:

[0027] d1(x h ,x q ) = ‖x h -x q ‖2

[0028] θ1(x h ,x q ) is specifically:

[0029]

[0030] The temporal similarity measurement S2(x h ,x q ) is specifically:

[0031]

[0032] Among them, K is the number of selected time series, ui is the weight ratio, d2(x h , x q ) is the Mahalanobis distance between x h and x q . t h and t q are respectively the time series widths corresponding to x h and x q . d t (t h , t q ) is the Mahalanobis distance between t h and t q . and are adjustable parameters, λ2 is the mixing weight coefficient between 0 and 1, u i Specifically:

[0033]

[0034] Among them, U i is the time series weight ratio of the i-th feature;

[0035] d2(x h , x q ) Specifically:

[0036]

[0037] d t (t h , t q ) Specifically:

[0038]

[0039] Step 6) includes the following steps:

[0040] 6.1) Calculate the mixed similarity index S X of the labeled samples:

[0041]

[0042] Among them, N is the number of historical samples;

[0043] 6.2) Based on the mixed similarity index, determine whether it is necessary to update the time series weight ratio, the LSSVM base model, or the LSTM base model;

[0044] 6.3) Update the mixed similarity index of the historical data. If any update is performed, add the sample data to the historical samples.

[0045] Step 6.2) Specifically:

[0046] When S X ≥ 95%, it is not updated;

[0047] When 70% ≤ S X < 95%,

[0048] e X = result real - result LSSVM

[0049] where e X is the prediction difference, result real is the true value, result lSSVM is the prediction value of the LSSVM base model;

[0050] Calculate the constraint calculation formula L(w, b, e X , α) of the LSSVM base model and update the LSSVM model:

[0051]

[0052] where H new is the updated time series weight ratio, H old is the original time series weight ratio, H x is the time series weight ratio to be updated, and N is the number of historical samples;

[0053] When 40% ≤ S X < 70%,

[0054]

[0055] When S X < 40%,

[0056] If it passes the expert knowledge test, then

[0057] e X = result real - result LSTM

[0058] where result LSTM is the prediction value of the LSTM base model;

[0059] Update the LSTM model

[0060]

[0061] If it does not pass, delete the sample data and jump out of the algorithm.

[0062] The present invention has the following beneficial effects and advantages:

[0063] 1. The present invention integrates two base models in an ensemble learning model, namely the LSTM model and the LSSVM model, enabling the proposed remaining life prediction model to have good prediction ability for both trained scenarios and untrained scenarios while taking into account the trained scenarios, thereby improving the generalization performance and general performance of the model.

[0064] 2. The present invention uses a similarity metric index that fuses spatio-temporal information. Among the similarity metric indices representing time information, the characteristics of slow feature analysis are combined to construct a similarity metric index using multi-scale time series, enabling it to more accurately evaluate the similarity between test samples and training samples, thus better distinguishing trained scenarios and untrained scenarios, providing a basis for judging the influence degree of the results of the base models in the ensemble learning model on the overall result and subsequent multi-level self-update methods, and improving the accuracy of the remaining life prediction model based on ensemble learning.

[0065] 3. The present invention adopts a model update method of multi-level self-update. In the way of not updating with high coincidence, it prevents the weight ratio of special scenario data from being occupied; in the way of self-updating of duplicate data, it reduces the update pressure brought by a large amount of duplicate data; in the way of updating after expert appraisal of new type data, it reduces the possible model contamination caused by new type data. In the way of multi-level self-update, the generalization of the model is enhanced, both trained scenarios and untrained scenarios are taken into account, and the model accuracy is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a flow diagram of the present invention.

[0067] Figure 2 It is a flow chart of a multi-level self-update method for predicting and maintaining the remaining life of a vehicle.

[0068] Figure 3 It is a flow chart of multi-level self-update.

[0069] Figure 4 It is a structure diagram of LSTM. DETAILED DESCRIPTION OF THE INVENTION

[0070] The following further describes the present invention in detail with reference to the drawings and embodiments.

[0071] As Figure 1 shown, a method for predicting and maintaining the remaining life of equipment with multi-level self-update includes the following steps:

[0072] Step 1: Collect the operation data of each component of the equipment vehicle as characteristic parameters and store them in a database;

[0073] Step 2: Preprocess the characteristic parameters obtained in Step 1 and construct a sample feature matrix;

[0074] Step 3. Use slow features to perform time series decomposition on the sample feature matrix obtained in Step 2, and establish a monitoring model according to the weight ratio and calculate the monitoring statistics SPE and S 2 , and train an early anomaly detection model to obtain the initial degradation point;

[0075] Step 4. Use the preprocessed time series and the initial degradation point to train two base models of the ensemble learning model, namely the LSTM model and the LSSVM model, and use spatio-temporal fusion similarity measurement to determine the influence of the two base models on the prediction results to obtain the remaining useful life prediction model;

[0076] Step 5. Perform remaining useful life prediction on each component of the equipment vehicle, that is, after preprocessing the data, use the early anomaly detection model to locate the initial degradation point, and then use the trained ensemble learning model to perform remaining useful life prediction;

[0077] Step 6. Repair the equipment vehicle according to the remaining useful life prediction results in Step 5, and perform labeling processing on the data after the repair is completed;

[0078] Step 7. Record the new label data in Step 6, record the data and store it in the database, and perform three operations of non-update, update without expert knowledge, and update after expert knowledge determination according to the similarity measurement results to complete the multi-level self-update of the ensemble model and the slow feature weight ratio.

[0079] The operation data of each component of the equipment vehicle collected includes:

[0080] Establish a sample library and a historical database, collect the operation data of each component of the equipment vehicle under normal driving in each working condition as samples, perform preliminary screening on the samples, analyze and identify the abnormal data in the samples, and perform correction, complementation or deletion processing on the data.

[0081] The feature parameters are preprocessed and a feature parameter matrix is constructed, including:

[0082] Calculate the time domain features and frequency domain features of the feature parameters, and summarize the time domain features and frequency domain features of each feature parameter to construct a feature parameter matrix. Use the zero-mean algorithm to preprocess the feature parameter matrix under multiple working conditions to obtain the sample feature matrix under a single working condition and store it in the historical database as a modeling sample.

[0083] The initial degradation point includes:

[0084] Use slow feature analysis to perform time series decomposition on the normal operation samples in the same operation cycle in the modeling samples and establish an anomaly monitoring model according to the weight ratio. The anomaly monitoring model evaluates the current state through the SPE control limit and S 2 control limit. When the SPE statistic of the new sample and S2 The statistics are all stably above the SPE control limit and S 2 When the control limit is reached, it is recognized as an abnormal state point, and the earliest abnormal state point is used as the initial degradation point.

[0085] The training of the ensemble learning model includes:

[0086] Using multiple groups of dimension-reduced data to train two base models respectively, where the base models include an LSTM model and an LSSVM model. Adjust the model parameters of the base models according to multiple trainings, and use the converged model parameters as the initial parameters of the ensemble learning model.

[0087] The spatio-temporal similarity metric includes:

[0088] Using a hybrid similarity metric method that fuses space and time. Among them, the method of spatial similarity metric is a hybrid similarity metric index that fuses the cosine of the included angle and the Euclidean distance, which is used to characterize the coupling relationship between each feature variable. The method of temporal similarity metric is to use the Mahalanobis distance under multi-scale time series as the similarity metric index, which is used to characterize the temporal characteristics of feature variables after excluding the correlation between feature variables. The two similarity metric indexes are normalized to obtain the hybrid similarity metric index.

[0089] The prediction of the remaining useful life of each component of the equipment vehicle includes:

[0090] Calculate the time-domain features and frequency-domain features of a sample at a certain moment in the sample library, and summarize the time-domain features and frequency-domain features of each sample to construct a feature parameter matrix. Inject the feature parameter matrix into the anomaly monitoring model to obtain the initial degradation point. Inject the feature parameter matrix and the initial degradation point into the remaining useful life prediction model to obtain the remaining useful life prediction values of each component of the equipment vehicle.

[0091] The data tagging process includes:

[0092] After the repair is completed, the machine stores the initial degradation point, spatio-temporal similarity metric index, and remaining useful life prediction value of this section of sample data in the data segment in the form of tags.

[0093] The multi-level self-update includes:

[0094] Analyze the labeled data after the repair is completed. When the new data has a too high degree of coincidence with the historical data, no update is performed. When the new data has a high degree of coincidence with the historical data, the model automatically performs an update, which includes updating the LSSVM base model and the time series weight ratio. When the new data has a medium degree of coincidence with the historical data, only the time series weight ratio is updated. When the new data has a low degree of coincidence with the historical data, it needs to be handed over to an expert for knowledge introduction for judgment. If the data conforms to the operation law of the equipment, the model is manually updated, which includes retraining the LSTM base model and updating the time series weight ratio.

[0095] In step 4, two base models, LSTM and LSSVM, are used for training. Among them, the LSTM model, as a pure data-driven method, is used to enhance the prediction effect of the model on data with a low degree of coincidence with historical data. The lower the degree of coincidence, the greater its proportion. LSSVM, as a traditional machine learning method, has strong interpretability and is used to enhance the prediction effect of the model on data with a high degree of coincidence with historical data. The higher the degree of coincidence, the greater its proportion.

[0096] The spatio-temporal similarity metric described in step 4. This similarity metric method combines spatial similarity and temporal similarity. Among them, spatial similarity is used to characterize the similarity of the relationship between feature variables. Temporal similarity is used to characterize the similarity of the feature variables themselves after removing the mutuality between variables. Different from the traditional temporal similarity metric method, it constructs a temporal similarity metric index in the way of multi-scale time series according to the characteristics of the SFA decomposition time series, which is more in line with the structural characteristics of SFA, thus improving the adaptability of the similarity metric index.

[0097] The multi-level self-update described in step 7 updates the labeled data in different ways according to the similarity metric index. On the one hand, this update method reduces the manual pressure by quickly processing the frequently occurring labeled data; on the other hand, it reduces the pollution of the fault data to the model update through the expert identification of new type samples, improving the reliability of the model; finally, through the segmented update of the anomaly monitoring model and the remaining life prediction model, it prevents a large amount of data of the same type from occupying the time series weight ratio of other types of data, and can enhance the prediction model while improving the monitoring comprehensiveness of the monitoring model.

[0098] Embodiment

[0099] As Figure 2 shown, a method for predicting the remaining life of equipment based on ensemble learning includes the following steps:

[0100] Step1: Collect the operation data of each component of the equipment vehicle as feature parameters and store them in the database;

[0101] The characteristic variables that can characterize the state of the equipped vehicle include engine speed, load, torque, engine water temperature, coolant temperature, oil pressure, and oil temperature. These parameters can be collected through sensors on the equipped vehicle.

[0102] Step2: Preprocess the obtained characteristic parameters and construct a sample feature matrix;

[0103] This step calculates the time-domain features and frequency-domain features of the state characteristics of the equipped vehicle collected in the database, and summarizes and constructs a feature parameter matrix. The calculation formulas for the selected time-domain features and frequency-domain features are shown in Table 1:

[0104] Table 1 Selection of Time-Domain Features and Frequency-Domain Features

[0105]

[0106] Step3: Obtain training sample data to construct and train an anomaly monitoring model, and determine the starting point of degradation through the anomaly monitoring model;

[0107] Through the slow feature analysis algorithm, use the historical normal operation samples of the equipped vehicle to complete the construction of the anomaly monitoring model.

[0108] The calculation in the slow feature analysis algorithm is as follows:

[0109] For the m-dimensional input signal vector X(t) = [X1(t), x2(t), …, X m (t)], assuming the change function g j (x) (j = 1, 2, …, m), can make the output signal S j (t) = [S1(t), S2(t), …, S m (t)] change as slowly as possible and be sorted according to the degree of slowness. Then the purpose of SFA is to obtain g j (x), and its objective function is:

[0110]

[0111] Among them: is the difference value between the slow feature and the original signal.

[0112] The constraint conditions of the above formula are as follows:

[0113]

[0114] Among them <S j > t is the mean value of the output signal; <S j 2 > t is the variance of the output signal; <S i Sj >[[]] t t is the covariance of the output signal.

[0115] In the basic SFA monitoring model, a linear mapping is commonly used to extract the slow features of the signal. For the linear mapping, the output signal where w is the linear transformation vector, also known as the weight vector. W is the linear transformation matrix, also known as the weight matrix. Substituting the output signal into the constraint function and the objective function, we can obtain:

[0116]

[0117] where is the difference of the output signal, and the SFA objective function can be optimized as:

[0118]

[0119] It can be transformed into a generalized eigenvalue problem:

[0120]

[0121] where Ω is a diagonal matrix containing the singular values. Arranging the eigenvalues of Ω from smallest to largest gives the sorting of the slow features from slow to fast.

[0122] According to the weight ratio, k slow features related to anomalies are selected and denoted as s k , then the calculation formula of the S 2 statistic is as follows:

[0123] S 2 = s k T s k

[0124] The calculation formula of the SPE statistic is as follows:

[0125]

[0126] where H is the feature matrix; P k is the orthogonal matrix of k slow features related to anomalies; W k is the load matrix of k slow features related to anomalies;

[0127] Since the slow features do not follow a Gaussian distribution, the kernel density estimation method is used to calculate the control limits of S 2 and SPE. The probability density estimate of y calculated by the kernel density estimation method is:

[0128]

[0129] where y iis the observed value; n is the number of samples; b is the bandwidth; G(·) is the kernel function. When the statistic stably exceeds the control limit, it can be considered that the equipment is abnormal, and the starting point of exceeding the limit is recorded as the initial degradation point.

[0130] Step4: Add the sample data and the initial degradation point into the ensemble learning model, and determine the influence proportion of the base model according to the hybrid similarity metric, and finally obtain the prediction result.

[0131] Ensemble learning includes two base models: LSSVM and LSTM. The LSSVM base model is an SVM model based on statistical theory, which can transform the quadratic optimization problem into solving a system of linear equations. For a given training set T, the function f obtained by training LSSVM transforms the input into the output, thus realizing regression. Its regression model can be expressed as:

[0132]

[0133] where is the weight vector of f, b is the bias term, is the non-linear mapping function that maps the input vector to a high-dimensional feature space. To find the weight vector and bias term that minimize the model prediction error, the following constraints need to be satisfied:

[0134]

[0135] where e i is the error term, representing the deviation between the model and the true value. Introduce the Lagrange multiplier α i , and define the Lagrangian function:

[0136]

[0137] where γ is the regularization parameter. Take the partial derivatives of the Lagrangian function and set them equal to 0 to get:

[0138]

[0139] Solving the above equations can obtain the model parameters α and b, thus determining the model.

[0140] As Figure 3 shown, the LSTM neural network is composed of basic components of LSTM cells, including components such as input gates, output gates, forget gates, and cell states. Its basic structure is as Figure 4 shown

[0141] At each moment, when the input of the LSTM neural network is X(t) = (X1, X2,..., X n ), the output of the hidden layer is H(t) = (H1, H2,..., H n ), and the output of the output layer is Y(t) = (Y1, Y2,..., Y n), the cell state is C(t) = (C1, C2, …, C n ), and there is

[0142]

[0143] where σ is the Sigmod activation function; w is the weight corresponding to different gates in different states; b i , b f , b o are the biases corresponding to different gates respectively.

[0144] The similarity metric index combines the spatial similarity metric index and the temporal similarity metric index. The spatial similarity metric index combines the Euclidean distance and the cosine of the included angle, and its definition method is as follows:

[0145]

[0146] where x h is the historical sample; x q is the query sample; d1(x h , x q ) is the Euclidean distance between x h and x q ; θ1(x h , x q ) is the cosine of the included angle between the two, and λ1 is the mixing weight coefficient between 0 and 1, is an adjustable parameter. The calculation formula of d1(x h , x q ) is as follows:

[0147] d1(x h , x q ) = ||x h - x q ‖2

[0148] The calculation formula of θ1(x h , x q ) is as follows:

[0149]

[0150] The temporal similarity metric index combines the Mahalanobis distance and the feature matrix in slow feature analysis, and its representation is as follows:

[0151]

[0152] where K is the number of selected time series; u i is the weight ratio; d2(x h , x q ) is the Mahalanobis distance between and, t h and t qAre x respectively h And x q The corresponding time - series width; d t (t h , t q ) is the Mahalanobis distance between t h And t q ; And Are adjustable parameters; λ2 is the mixing weight coefficient between 0 and 1. u i The calculation formula of is as follows:

[0153]

[0154] The calculation formula of d2(x h , x q ) is as follows:

[0155]

[0156] d t (t h , t q ) is as follows:

[0157]

[0158] Step.5

[0159] As Figure 3 Shown, update the labeled data according to the similarity - measurement index. The update steps are as follows:

[0160]

[0161] e X = result real - result LSSVM

[0162] Calculate L(w, b, e X , α) (update the LSSVM model)

[0163] (Update the time - series weight ratio, N is the number of historical samples)

[0165] When 40% ≤ S X < 70%

[0166] (Update the time - series weight ratio, N is the number of historical samples)

[0168] When S p < 40% If it passes the expert - knowledge test, then

[0170] eX = result real - result LSTM

[0171] Update the LSTM model

[0172] (Update the time series weight ratio, N is the number of historical samples)

[0174] If not passed, delete the sample data and jump out of the algorithm

[0176] 3. Update the hybrid similarity index of historical data

[0177] The implementation methods of each step described in this specification are closely linked and cannot be implemented independently without any link. The similarities and differences between each implementation method can be detailed in the specification part. It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not expressly listed, or elements inherent to such process, method, article or device.

[0179] The above is only the implementation mode of the present invention and is not used to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, extension, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A multi-level self-updating method for predicting the remaining life of a vehicle and maintenance, characterized in that, It includes the following steps: 1) Collect the operation data of each component of the equipment vehicle as characteristic parameters; 2) Calculate the time-domain characteristics and frequency-domain characteristics in the characteristic parameters, and summarize and construct a characteristic parameter matrix; 3) Construct an anomaly monitoring model based on the slow feature analysis algorithm, and determine the starting point of degradation through the anomaly monitoring model; 4) Based on the LSTM model and the LSSVM model as the base models, construct an ensemble learning model, and use the characteristic parameter matrix and the initial degradation point to train it; 5) Use the trained ensemble learning model to predict the remaining life of each component of the equipment vehicle, perform maintenance on the equipment vehicle according to the prediction results, and perform labeling processing on the data after the maintenance is completed; 6) Update the labeled data according to the similarity metric index of the ensemble learning model, and then realize the multi-level self-update of the weight ratio of the ensemble learning model and the slow feature analysis algorithm.

2. A method for predicting the remaining life and maintenance of a vehicle with multi-level self-update according to claim 1, characterized in that, The step 3) includes the following steps: 3.1) Use slow feature analysis to decompose the normal operation samples in the same operation cycle in the characteristic parameter matrix and screen out k anomaly-related slow features according to the weight ratio; 3.2) Calculate the SPE statistic and the S 2 statistic of the slow feature respectively; 3.3) Calculate S using the kernel density estimation method 2 and the SPE control limits. When the continuous 5 acquisition points of both statistics exceed the control limits, it is considered that the equipment has an anomaly, and the starting point of exceeding the limit is recorded as the initial degradation point.

3. A multi-level self-updating method for predicting the remaining life of a vehicle and maintenance, according to claim 1, characterized in that, The kernel density estimation method is specifically: Among them, is the kernel density estimate of y, where y is the original data to be estimated, and y i is the observed value; n is the number of samples; b is the bandwidth; G(·) is the kernel function.

4. A multi-level self-updating method for predicting the remaining life of a vehicle and maintenance, according to claim 1, characterized in that, The step 4) is specifically: Take the LSTM model and the LSSVM model as the base models of the ensemble learning model, and use the spatio-temporal similarity metric to determine the proportion of the two base models. The spatio-temporal similarity metric includes the spatial similarity metric and the temporal similarity metric. The value obtained by normalizing the two similarity metric indexes is used as the proportion of the two base models. The closer the value is to 1, the larger the proportion of the LSSVM model. Among them: The spatial similarity metric is a hybrid similarity metric index that fuses the cosine of the included angle and the Euclidean distance, and is used to characterize the coupling relationship between each feature variable; The temporal similarity metric is the Mahalanobis distance under the multi-scale time series as the similarity metric index, and is used to characterize the temporal characteristics of the feature variables after excluding the correlation between the feature variables.

5. A multi-level self-updating method for predicting the remaining life of a vehicle and maintenance, according to claim 4, characterized in that, The spatial similarity metric S1(x h , x q ) is specifically as follows: Among them, x h is a historical sample, and x q is a query sample. d1(x h , x q ) is the Euclidean distance between x h and x q . θ1(x h , x w ) is the cosine of the included angle between the two. λ1 is a mixing weight coefficient between 0 and 1, is an adjustable parameter. d1(x h , x q ) is specifically: θ1(x h ,x q ) Specifically:

6. A multi-level self-updating method for predicting the remaining life of a vehicle and maintenance according to claim 4, characterized in that The time similarity metric S2(x h , x q ) is specifically as follows: where K is the number of selected time series, and u i is the weight ratio, and d2(x h , x q ) is the Mahalanobis distance between x h and x q . t h and t q are the widths of the time series corresponding to x h and x q respectively. d t (t h , t q ) is the Mahalanobis distance between t h and t q . and are adjustable parameters, λ2 is the mixing weight coefficient between 0 and 1, and u i is specifically: Among them, U i is the time series weight ratio of the i-th feature; d2(x h ,x q )Specifically: d t (t h ,t q Specifically:

7. A method for predicting the remaining life and maintenance of a vehicle with multi-level self-update according to claim 1, characterized in that, The step 6) includes the following steps: 6.1) Calculate the similarity index S of the labeled sample mixture X : Among them, N is the number of historical samples; 6.2) Judge whether it is necessary to update the time series weight ratio, the LSSVM base model or the LSTM base model based on the hybrid similarity index; 6.3) Update the hybrid similarity index of the historical data. If any update is performed, add the sample data to the historical samples.

8. A method for predicting the remaining life and maintenance of a vehicle with multi-level self-update according to claim 7, characterized in that, The step 6.2) is specifically: When S X ≥ 95%, do not update; When 70% ≤ S X < 95%, e X = result real - result LSSVM Among them, e X is the predicted difference, and result real is the true value, and result LSSVM is the predicted value of the LSSVM base model; Calculate the constraint calculation formula \(L(w,b,\xi,\alpha)\) of the LSSVM base model and update the LSSVM model: X ,\(\alpha\), and update the LSSVM model: Among them, H new is the updated time series weight ratio, H old is the original time series weight ratio, H x is the time series weight ratio to be updated, and N is the number of historical samples; When 40% ≤ S X < 70%, When S X <40%, If it passes the expert knowledge test, then e X = result real -result LSTM Among them, result LSTM is the predicted value of the LSTM-based model; Update the LSTM model If it does not pass, delete the sample data and jump out of the algorithm.

Images