CN112115643B - Smart train service life non-invasive prediction method - Google Patents

Smart train service life non-invasive prediction method Download PDF

Info

Publication number
CN112115643B
CN112115643B CN202010967935.2A CN202010967935A CN112115643B CN 112115643 B CN112115643 B CN 112115643B CN 202010967935 A CN202010967935 A CN 202010967935A CN 112115643 B CN112115643 B CN 112115643B
Authority
CN
China
Prior art keywords
service life
train
model
key device
key
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010967935.2A
Other languages
Chinese (zh)
Other versions
CN112115643A (en
Inventor
刘辉
尹诗
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Central South University
Original Assignee
Central South University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Central South University filed Critical Central South University
Priority to CN202010967935.2A priority Critical patent/CN112115643B/en
Publication of CN112115643A publication Critical patent/CN112115643A/en
Application granted granted Critical
Publication of CN112115643B publication Critical patent/CN112115643B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G06Q10/06393—Score-carding, benchmarking or key performance indicator [KPI] analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/04—Ageing analysis or optimisation against ageing

Abstract

The invention discloses a non-invasive prediction method for the service life of an intelligent train, which comprises the steps of acquiring a train key equipment signal to obtain sampling data of each key equipment and sampling data of a total measuring point and carrying out data processing; constructing a non-invasive load decomposition model of each key device; processing the historical service life parameters and the key equipment characteristic parameters of each key equipment of the train again; constructing and obtaining a service life prediction model of each key device of the train and a service life integrated prediction model of each key device of the train; constructing a service life prediction model of the whole train; and acquiring real-time signals of the target train to be predicted in real time and predicting the service life of key equipment and the service life of the whole train of the target train to be predicted by adopting the constructed model. The method realizes the non-invasive prediction of the service life of the whole train, and has the advantages of high reliability, good practicability, low cost and good feasibility.

Description

Smart train service life non-invasive prediction method
Technical Field
The invention belongs to the field of rail transit, and particularly relates to a non-invasive prediction method for service life of a whole intelligent train.
Background
With the development of economic technology and the improvement of living standard of people, the rail transit is widely applied to the production and the life of people, and brings endless convenience to the production and the life of people.
With the continuous development of rail transit technology in China, safe and reliable operation of trains is receiving wide attention. The service life of the whole train has a certain period, and the train with an over-age service easily causes various safety accidents to bring personal injury and property loss, so that the method has important significance for predicting the service life of the whole train.
The current patents for train life prediction mainly relate to the following two aspects:
1. life prediction for individual devices of a train: as disclosed in publication No. CN105973597A, the life of a bearing is predicted by collecting real-time load data using a plurality of sensors disposed on the axle boxes of a train and taking into account the bearing damage over a fixed time.
2. Predicting the service life of the whole train based on multiple sensors: for example, in patent publication No. CN110376003A, the working performance parameters of all relevant components are constantly monitored based on a sensor network disposed inside each component of the train, and the service life of the whole train is predicted by using an intelligent model.
The method can realize the prediction of the service life of single equipment or the whole train, but all the methods adopt an intrusive method, a sensor network needs to be arranged in each part of the train to complete the real-time monitoring of the working performance parameters of the train, the arrangement cost is high, and better economy and feasibility cannot be realized.
Disclosure of Invention
The invention aims to provide a non-invasive prediction method for service life of a whole intelligent train, which has high reliability, good practicability, low cost and better feasibility.
The invention provides a non-intrusive prediction method for the service life of a whole intelligent train, which comprises the following steps:
s1, performing signal acquisition on key equipment of a train in an intrusive mode to obtain sampling data of each key equipment and sampling data of a total measuring point;
s2, preprocessing the sampling data acquired in the step S1 to obtain key equipment characteristic parameters and total measuring point characteristic parameters;
s3, constructing a characteristic relation original model between the total measuring point and each key device, and training the characteristic relation original model by adopting the key device characteristic parameters obtained in the step S2, so as to obtain a non-invasive load decomposition model of each key device;
s4, processing the historical service life parameters of each key device of the train and the key device characteristic parameters obtained in the step S2 again;
s5, constructing a service life prediction original model of each key device and a service life integrated prediction original model of each key device, and training the two constructed models by adopting the data obtained in the step S4, so as to obtain a service life prediction model of each key device of the train and a service life integrated prediction model of each key device of the train;
s6, constructing a service life prediction original model of the whole train, and training the service life prediction original model by adopting the prediction result of the service life prediction model of each key device of the train obtained in the step S5, so as to obtain a service life prediction model of the whole train;
s7, collecting real-time signals of the target train to be predicted in a non-invasive mode in real time, processing data, and predicting the service life of the key equipment of the target train to be predicted and the service life of the whole train through the service life prediction model of each key equipment of the train, the service life integrated prediction model of each key equipment of the train and the service life prediction model of the whole train, which are constructed in the step S5, and the service life prediction model of the whole train, which is constructed in the step S7.
Step S1, performing signal acquisition on the train key equipment in an intrusive manner to obtain sampling data of each key equipment and sampling data of the total measuring point, specifically, obtaining the sampling data of each key equipment and the sampling data of the total measuring point by the following steps:
A. dividing the train into a mechanical part and an electrical part, and selecting corresponding key equipment for each part; meanwhile, selecting a total measuring point of the whole train;
B. selecting trains with different residual service lives;
C. acquiring a vibration acceleration signal and a load active power signal at each key device in the step A aiming at each train selected in the step B, and simultaneously acquiring a vibration acceleration signal and a load active power signal at a total measuring point;
D. and C, performing data processing on the signal data obtained in the step C to obtain final sampling data of each key device and sampling data of the total measuring point.
And D, performing data processing on the signal data obtained in the step C to obtain final sampling data of each key device and the final sampling data of the total measuring point, specifically performing data processing by adopting the following steps:
d-1, if the data are abnormal, deleting the corresponding abnormal data,
d-2, performing data filling by adopting an interpolation method aiming at the missing data and the deleted data;
d-3, using the filled data as a label according to the corresponding train model, equipment model and time to establish an original database D ═ D, l, P, P, a, A](ii) a Wherein d is the model of the train, l is the remaining service life of the whole train, p is the load active power signal of the electrical equipment and p ═ pe,pl,pt,p(t)]，peIs the type of electrical apparatus, plFor the remaining service life of the electrical apparatus, ptIs a time label of the electrical equipment, P (t) is a load active power signal value of the electrical equipment, P is a load active power signal of a total measuring point and P ═ Pl,Pt,P(t)]，PlFor the remaining service life of the electrical part of the train, PtIs the time stamp at the total measured point, p (t) is the value of the load active power signal at the total measured point, a is the load active power signal of the mechanical equipment and a ═ ae,al,at,a(t)]，aeIs the type of mechanical equipment, alFor the remaining service life of the mechanical equipment, atA (t) is a signal value of a vibration acceleration signal of the mechanical equipment, A is a signal value of the vibration acceleration signal of a total measuring point and A is [ A ═ Al,At,A(t)]，AlFor the remaining service life of the mechanical part of the train, AtAnd A (t) is a time label of the total measuring point, and A (t) is a signal value of the vibration acceleration signal of the total measuring point.
The step S2 is to preprocess the sample data obtained in the step S1, so as to obtain key device characteristic parameters and total measurement point characteristic parameters, specifically, the following steps are adopted to obtain the key device characteristic parameters:
a. transforming all the original sampling data acquired in the step S1 into high and low frequency signals by adopting Fourier transform, and denoising by adopting a low-pass filter, thereby obtaining effective characteristic parameters of the total measuring point and effective characteristic parameters of each key device;
b. b, normalizing the effective characteristic parameters obtained in the step a, so as to ensure that the data after normalization meets the standard normal distribution;
c. storing the data information processed in the step b in a uniform format by adopting a sliding window mode, thereby obtaining a final key equipment characteristic parameter p' ═ pe,pl,pt,p'(t)]And a ═ ae,al,at,a'(t)]And a total measured point characteristic parameter P ═ Pl,Pt,P'(t)]And a' ═ al,At,A'(t)](ii) a Wherein P '(t) is the processed load active power signal value of the electrical equipment, a' (t) is the processed signal value of the vibration acceleration signal of the mechanical equipment, P '(t) is the processed load active power signal value at the total measurement point, and a' (t) is the processed signal value of the vibration acceleration signal at the total measurement point.
The step a adopts Fourier transform, specifically adopts the following formula to carry out Fourier transform:
Figure BDA0002683026500000041
wherein X (omega) is a high-low frequency signal after transformation; x (t) is the original sample data before transformation.
The normalization processing in the step b is specifically performed by adopting the following formula:
Figure BDA0002683026500000051
wherein x' is data information after normalization processing;
Figure BDA0002683026500000052
the data information before normalization processing is carried out; mu is the data average value of the data sequence before processing; k is the total number of data of the data sequence before processing.
The sliding window in step c is specifically the width of the sliding window is the length of 100 periodic fundamental frequency signals, and the step length of the sliding window is the length of 100 periodic fundamental frequency signals.
Step S3, constructing a feature relationship original model between the total measurement point and each of the key devices, and training the feature relationship original model by using the key device feature parameters obtained in step S2, so as to obtain a non-intrusive load decomposition model of each of the key devices, specifically obtaining the non-intrusive load decomposition model of each of the key devices by using the following steps:
(1) embedding the convolution layer into a gate control circulation unit neural network, thereby constructing and obtaining a non-invasive load decomposition model;
(2) dividing the data processed in the step S2 into a training set and a verification set, and training the non-invasive load decomposition preliminary models of the key equipment one by one;
(3) and (3) verifying the non-invasive load decomposition preliminary model obtained in the step (2) by using a verification set, and selecting the model with the best error function as the final non-invasive load decomposition model of each key device.
The method comprises the following steps that (1) a convolution layer is built in a gate control cycle unit neural network, so that a non-invasive load decomposition model is built, specifically, the convolution layer is built in the gate control cycle unit neural network, and the obtained neural network comprises an input layer, a convolution layer, a hidden layer and an output layer; the hidden layer comprises a gating circulation unit and a full connection layer; adopting a ReLU function and a linear function as activation functions; adopting a cross entropy function as a loss function; and performing subsequent model training by using an NAG algorithm as a parameter optimization algorithm, and setting a learning rate eta to be 0.01 and an attenuation rate gamma to be 0.9.
In the step (2), the data processed in the step S2 are divided into a training set and a verification set, and the non-intrusive load decomposition preliminary models of the key devices are trained one by one, specifically, the non-intrusive load decomposition preliminary models of the key devices are trained one by adopting the following steps:
(2) 1, taking 80% of the data processed in the step S2 as a training set and 20% of the data as a verification set;
(2) if the key device is an electrical device, taking a total measurement point sequence P '(t) ═ P' (1), P '(2),.. gtoreq.p' (k) of the electrical devices in the training set as a model input, and taking a device sequence P '(t) ═ P' (1), P '(2),. gtoreq.p' (k) ] of each electrical device in the training set as a model output, and training the model;
(2) if the key device is a mechanical device, the model is input by using the total measurement point sequence a '(t) ═ a' (1), a '(2),.. and a' (k)) of the mechanical devices in the training set, and output by using the device sequence a '(t) ═ a' (1), a '(2),. and.. and a' (k)) of each mechanical device in the training set as a model, and the model is trained.
And (3) verifying the non-intrusive load decomposition preliminary model obtained in the step (2) by using a verification set, selecting a model with the best error function as a final non-intrusive load decomposition model of each key device, specifically, verifying the non-intrusive load decomposition preliminary model obtained in the step (2) by using the verification set, taking the root mean square error as an error function, and selecting the model with the minimum root mean square error value as the non-intrusive load decomposition model of each key device.
And step S4, re-processing the historical service life parameters of each key device of the train and the key device characteristic parameters obtained in step S2, specifically, processing by adopting the following steps:
1) using empirical wavelet transform algorithm to obtain data information p from step S1lAnd alAnd the data information p '(t) and a' (t) obtained in step S2 are decomposed into a fixed number of subsequences of high and low frequencies;
2) evaluating the stationarity of the subsequences obtained in the step 1) by using sample entropy;
3) according to the stationarity of the subsequence obtained in the step 2), adopting an ICEEMDAN algorithm to further decompose the unstable subsequence;
4) performing dimensionality reduction by using a principal component analysis method to obtain final processing data pl'＝[pl1',pl2',...,plc']、p”(t)＝[p1”(t),p2”(t),...pc”(t)]、al'＝[al1',al2',...,alc']And a "(t) ═ a1”(t),a2”(t),...,ac”(t)]。
Evaluating the stationarity of the subsequence obtained in the step 1) by using the sample entropy in the step 2), specifically, using m to represent the dimension of the subsequence, using r to represent the similarity of the subsequence, using N to represent the length of the subsequence, thereby obtaining the sample entropy of the subsequence as SampEN (m, r, N), and evaluating: the top 20% of the subsequences ranked from high to low by the value of sample entropy are non-stationary subsequences.
Performing dimensionality reduction by using a principal component analysis method to obtain final processed data, specifically, using principal component analysis as a dimensionality reduction algorithm to obtain subsequences corresponding to a plurality of previous characteristic values with cumulative contribution rates closest to 1 as final processed data, so as to obtain final processed data pl'＝[pl1',pl2',...,plc']、p”(t)＝[p1”(t),p2”(t),...pc”(t)]、al'＝[al1',al2',...,alc']And a "(t) ═ a1”(t),a2”(t),...,ac”(t)]。
Step S5, constructing an original service life prediction model of each key device and an original service life integrated prediction model of each key device, and training the two constructed models by using the data obtained in step S4, so as to obtain a service life prediction model of each key device of the train and an integrated service life prediction model of each key device of the train, specifically obtaining a service life prediction model of each key device of the train by using the following steps:
the method comprises the steps that I, an echo state network model optimized by a dragonfly algorithm is used as an original model, and an original service life prediction model for each key device is constructed; training the constructed original service life prediction model of each key device by adopting the processed data obtained in the step S4, thereby obtaining a service life prediction model of each key device of the train;
taking an echo state network model optimized by a dragonfly algorithm as an original model, and constructing an original model for service life integrated prediction of each key device; and training the constructed original service life integrated prediction model of each key device by adopting the processed data obtained in the step S4, thereby obtaining the service life integrated prediction model of each key device of the train.
Step I, using an echo state network model optimized by a dragonfly algorithm as an original model, and constructing a service life prediction original model aiming at each key device; and training the constructed original service life prediction model of each key device by adopting the processed data obtained in the step S4, thereby obtaining a service life prediction model of each key device of the train, specifically obtaining the service life prediction model of each key device of the train by adopting the following steps:
i-1, when the target key device is an electrical device, the processed subsequence p "(t) ═ p obtained in step S41”(t),p2”(t),...pc”(t)]As model input, the remaining service life subsequence p of each corresponding key devicel'＝[pl1',pl2',...,plc']As an output; when the target key device is a mechanical device, the processed subsequence a "(t) ═ a obtained in step S41”(t),a2”(t),...,ac”(t)]As model input, the residual service life subsequence a of each corresponding key devicel'＝[al1',al2',...,alc']As an output;
i-2, initializing echo network state parameters: randomly generating a reserve pool to enable the spectrum radius to be smaller than R1, randomly generating input connection and output feedback weights, and taking a tanh function as an activation function without changing in the training process;
i-3, generating an initialization population: setting the value range of the dragonfly population quantity as [ N1, N2], and the value range of the maximum iteration number as [ N3, N4 ]; parameters to be optimized of the echo state network are represented by the position of each population individual; N1-N4 are all natural numbers;
i-4, determining an optimization objective function: dividing input data for training an echo state network model into two parts, wherein N5% serves as a training set, and N6% serves as a verification set for verifying the performance of the current parameters; selecting a mean square error as an optimization objective function; n5 and N6 are both positive real numbers, and N5+ N6 is 100;
i-5, outputting the optimal parameters of the echo state network: and optimizing the output connection weight of the echo state network by using the determined optimization objective function, finding a parameter value which enables the mean square error output value to be minimum within the maximum iteration times, using the parameter value as the optimal parameter of the echo state network, and obtaining a service life prediction model of each key device of the train.
Step II, constructing an integrated service life prediction original model of each key device by taking the echo state network model optimized by the dragonfly algorithm as an original model; and training the constructed original service life integrated prediction model of each key device by adopting the processed data obtained in the step S4, thereby obtaining the service life integrated prediction model of each key device of the train, specifically obtaining the service life integrated prediction model of each key device of the train by adopting the following steps:
II-1, when the target key device is an electrical device, the processed subsequence p obtained in step S4l'＝[pl1',pl2',...,plc']As model input, using the residual service life sequence p of each corresponding key device of z trains in the original databasel＝[pl1,pl2,...,plz]As an output; when the target key device is a mechanical device, the processed subsequence a obtained in step S4l'＝[al1',al2',...,alc']As model input, the residual service life sequence a of each corresponding key device of z trainsl＝[al1,al2,...,alz]As an output;
II-2, initializing echo network state parameters: randomly generating a reserve pool to enable the spectrum radius to be smaller than R2, randomly generating input connection and output feedback weights, and taking a tanh function as an activation function without changing in the training process;
II-3, generating an initialization population: setting the value range of the dragonfly population quantity as [ N7, N8], and the value range of the maximum iteration number as [ N9, N10 ]; parameters to be optimized of the echo state network are represented by the position of each population individual; N7-N10 are natural numbers;
II-4, determining an optimization objective function: dividing input data for training an echo state network model into two parts, wherein N11% serves as a training set, and N12% serves as a verification set for verifying the performance of the current parameters; selecting a mean square error as an optimization objective function; n11 and N12 are both positive real numbers, and N11+ N12 is 100;
II-5, outputting the optimal parameters of the echo state network: and optimizing the output connection weight of the echo state network by using the determined optimization objective function, finding a parameter value which enables the mean square error output value to be minimum within the maximum iteration times, and taking the parameter value as the optimal parameter of the echo state network to obtain an integrated service life prediction model of each key device of the train.
Step S6, constructing an original service life prediction model of the whole train, and training the original service life prediction model by adopting the prediction result of the service life prediction model of each key device of the train obtained in step S5, so as to obtain the service life prediction model of the whole train, specifically, obtaining the service life prediction model of the whole train by adopting the following steps:
using the prediction result of the service life prediction model of each key device of the train obtained in the step S5 as model input, and using the train service life sequence as model output;
and ii, training the multi-target ant colony optimization model by adopting the data in the step i, thereby obtaining a final service life prediction model of the whole train.
And step ii, training the multi-target ant colony optimization model by using the data in the step i, so as to obtain a final service life prediction model of the whole train, specifically, obtaining the final service life prediction model of the whole train by using the following steps:
ii-1, initializing parameters: initializing the ant colony number to be M1, the maximum iteration number to be M2 and initializing pheromone related parameters randomly by adopting multi-objective ant colony optimization; m1 and M2 are both positive integers;
ii-2, determining an optimization variable: setting the numbers of electrical equipment and mechanical equipment of the train as m and n respectively, wherein f represents a utilization parameter c1,c2,...,cm+nThe service life of the whole train is obtained by fitting the residual service life of each device, and the optimized variable is the parameter c1,c2,...,cm+n，pli＝[pli(1),pli(2),...,pli(m)]Predicting the residual service life of the m key electrical devices of the ith train obtained in the step S5li＝[ali(1),ali(2),...,ali(n)]If i is 1,2i＝c1pli(1)+c2pli(2)+...cmpli(m)+cm+1ali(1)+cm+2ali(2)+...+cm+ nali(n)；
Ii-3, determining an optimization objective function:
Figure BDA0002683026500000111
Figure BDA0002683026500000112
in the formula IiReal service life data of the ith train in the z trains obtained in the step S1; f. ofiFitting results of the service life of the entire vehicle of the ith train;
Figure BDA0002683026500000113
the average value of the real data of the service life of the whole z trains is obtained;
Figure BDA0002683026500000114
the average value of the fitting results of the service life of the whole z trains is obtained;
ii-4. performing multi-objective optimization: enabling all ants to traverse all paths, calculating an optimization objective function value and recording a non-dominant solution corresponding to the optimal condition;
ii-5, continuously updating the search path and recording the optimal path, stopping searching when the maximum iteration times is reached, and outputting a non-dominated solution set;
ii-6, evaluating the performance of the non-dominated solution set by using the verification set to obtain the average absolute percentage error of the model output value
Figure BDA0002683026500000115
And (5) as an evaluation index, obtaining a non-dominated solution with the minimum E value as an optimal parameter to be output, and obtaining a final service life prediction model of the whole train.
Step S7, collecting real-time signals of a target train to be predicted in a non-intrusive mode in real time, after data processing, predicting the service life of the key equipment of the target train to be predicted and the service life integrated prediction model of the key equipment of the train through the service life prediction model of each key equipment of the train and constructed in step S5 and the service life prediction model of the whole train constructed in step S7, specifically, collecting vibration acceleration data information and load active power information on a total measuring point of the target train to be predicted in a non-intrusive mode in real time on the target train to be predicted which normally runs, performing data processing in the mode S2, decomposing the model obtained in step S3, performing data reprocessing in the mode S4, and finally performing service life prediction on each equipment through the model obtained in step S5, and predicting the service life of the whole vehicle by using the model obtained in the step S6.
The intelligent non-intrusive prediction method for the service life of the whole intelligent train provided by the invention has the advantages that basic data are obtained in an intrusive mode, a prediction model is established, then the working performance parameters of the train are monitored by using a non-intrusive load decomposition technology aiming at a target train, and the collected working performance parameters of the train are subjected to model prediction by using the prediction model, so that the non-intrusive prediction of the service life of the whole intelligent train is realized; moreover, the method has the advantages of high reliability, good practicability, low cost and good feasibility.
Drawings
FIG. 1 is a schematic process flow diagram of the process of the present invention.
Detailed Description
FIG. 1 is a schematic flow chart of the method of the present invention: the invention provides a non-invasive prediction method for service life of an intelligent train, which comprises the following steps:
s1, carrying out signal acquisition on key equipment of a train in an intrusive mode to obtain sampling data of each key equipment and sampling data of a total measuring point; specifically, the method comprises the following steps of obtaining sampling data of each key device and sampling data of a total measuring point:
A. dividing the train into a mechanical part and an electrical part, and selecting corresponding key equipment for each part; meanwhile, selecting a total measuring point of the whole train;
B. selecting trains with different residual service lives;
C. for each train selected in the step B, acquiring a vibration acceleration signal a (t) and a load active power signal p (t) at each key device in the step A, and simultaneously acquiring a vibration acceleration signal A (t) and a load active power signal P (t) at a total measuring point;
D. c, performing data processing on the signal data obtained in the step C to obtain final sampling data of each key device and sampling data of a total measuring point; specifically, the following steps are adopted for data processing:
d-1, if the data are abnormal, deleting the corresponding abnormal data,
d-2, performing data filling by adopting an interpolation method aiming at the missing data and the deleted data;
d-3, using the filled data as a label according to the corresponding train model, equipment model and time to establish an original database D ═ D, l, P, P, a, A](ii) a Wherein d is the model of the train, l is the remaining service life of the whole train, p is the load active power signal of the electrical equipment and p ═ pe,pl,pt,p(t)]，peIs the type of electrical equipment, plFor the remaining service life of the electrical apparatus, ptIs a time label of the electrical equipment, P (t) is a load active power signal value of the electrical equipment, P is a load active power signal of a total measuring point and P ═ Pl,Pt,P(t)]，PlFor the remaining service life of the electrical part of the train, PtIs the time stamp at the total measured point, p (t) is the value of the load active power signal at the total measured point, a is the load active power signal of the mechanical equipment and a ═ ae,al,at,a(t)]，aeIs the type of mechanical equipment, alFor the remaining service life of the mechanical equipment, atA (t) is a signal value of a vibration acceleration signal of the mechanical equipment, A is a signal value of the vibration acceleration signal of a total measuring point and A is [ A ═ Al,At,A(t)]，AlFor the remaining service life of the mechanical part of the train, AtAt the total measuring pointA (t) is a signal value of the vibration acceleration signal at the total measurement point;
s2, preprocessing the sampling data acquired in the step S1 to obtain key equipment characteristic parameters and total measuring point characteristic parameters; specifically, the key equipment characteristic parameters are obtained by adopting the following steps:
a. transforming all the original sampling data acquired in the step S1 into high and low frequency signals by adopting Fourier transform, and denoising by adopting a low-pass filter so as to obtain effective characteristic parameters of a total measuring point and effective characteristic parameters of each key device; specifically, the following formula is adopted to carry out Fourier transform:
Figure BDA0002683026500000141
wherein X (omega) is a transformed high-low frequency signal; x (t) is original sampling data before transformation;
b. b, normalizing the effective characteristic parameters obtained in the step a, so as to ensure that the data after normalization meets the standard normal distribution; specifically, the normalization processing is performed by adopting the following formula:
Figure BDA0002683026500000142
wherein x' is data information after normalization processing;
Figure BDA0002683026500000143
the data information before normalization processing is carried out; mu is the data average value of the data sequence before processing; k is the total data number of the data sequence before processing;
c. storing the data information processed in the step b in a uniform format by adopting a sliding window mode, thereby obtaining a final key equipment characteristic parameter p' ═ pe,pl,pt,p'(t)]And a' ═ ae,al,at,a'(t)]And a total measured point characteristic parameter P ═ Pl,Pt,P'(t)]And a ═ al,At,A'(t)](ii) a Wherein P '(t) is a processed load active power signal value of the electrical equipment, a' (t) is a processed signal value of a vibration acceleration signal of the mechanical equipment, P '(t) is a processed load active power signal value at a total measurement point, and a' (t) is a processed signal value of the vibration acceleration signal at the total measurement point;
in specific implementation, the width of the sliding window is the length of 100 periodic fundamental frequency signals, and the step length of the sliding window is the length of 100 periodic fundamental frequency signals;
s3, constructing a characteristic relation original model between the total measuring point and each key device, and training the characteristic relation original model by adopting the key device characteristic parameters obtained in the step S2, so as to obtain a non-invasive load decomposition model of each key device; specifically, the non-invasive load decomposition model of each key device is obtained by the following steps:
(1) embedding the convolutional layer into a gate control cycle unit neural network, thereby constructing and obtaining a non-invasive load decomposition model; the method specifically comprises the steps that a convolution layer is built in a gate control circulation unit neural network, and the obtained neural network comprises an input layer, the convolution layer, a hidden layer and an output layer; the hidden layer comprises a gating circulation unit and a full connection layer; adopting a ReLU function and a linear function as activation functions; adopting a cross entropy function as a loss function; performing subsequent model training by using an NAG algorithm as a parameter optimization algorithm, and setting a learning rate eta to be 0.01 and an attenuation rate gamma to be 0.9;
(2) dividing the data processed in the step S2 into a training set and a verification set, and training the non-invasive load decomposition preliminary models of the key equipment one by one; specifically, the non-invasive load decomposition initial model of each key device is trained one by adopting the following steps:
(2) 1, taking 80% of the data processed in the step S2 as a training set and 20% of the data as a verification set;
(2) if the key device is an electrical device, taking a total measurement point sequence P '(t) ═ P' (1), P '(2),.. gtoreq.p' (k) of the electrical devices in the training set as a model input, and taking a device sequence P '(t) ═ P' (1), P '(2),. gtoreq.p' (k) ] of each electrical device in the training set as a model output, and training the model;
(2) if the key device is a mechanical device, using the total measured point sequence a '(t) ═ a' (1), a '(2),. and.a' (k) of the mechanical devices in the training set as model input, and using the device sequence a '(t) ═ a' (1), a '(2),. and.a' (k) ] of each mechanical device in the training set as model output, training the model;
(3) verifying the non-invasive load decomposition preliminary model obtained in the step (2) by using a verification set, and selecting a model with the best error function as a final non-invasive load decomposition model of each key device; verifying the non-intrusive load decomposition preliminary model obtained in the step (2) by using a verification set, taking the root mean square error as an error function, and selecting the model with the minimum root mean square error value as the non-intrusive load decomposition model of each key device;
s4, processing the historical service life parameters of each key device of the train and the characteristic parameters of the key devices obtained in the step S2 again; the method specifically comprises the following steps:
1) using empirical wavelet transform algorithm to obtain data information p from step S1lAnd alAnd the data information p '(t) and a' (t) obtained in step S2 are decomposed into a fixed number of subsequences of high and low frequencies;
2) evaluating the stationarity of the subsequences obtained in the step 1) by using sample entropy; specifically, m represents the dimension of the subsequence, r represents the similarity of the subsequence, and N represents the length of the subsequence, so that the sample entropy of the subsequence is SampEN (m, r, N), and the evaluation is as follows: ranking the top 20% of subsequences from high to low according to the value of sample entropy as unstable subsequences;
3) according to the stationarity of the subsequence obtained in the step 2), adopting an ICEEMDAN algorithm to further decompose the unstable subsequence;
4) performing dimensionality reduction by using a principal component analysis method to obtain final processing data pl'＝[pl1',pl2',...,plc']、p”(t)＝[p1”(t),p2”(t),...pc”(t)]、al'＝[al1',al2',...,alc']And a "(t) ═ a1”(t),a2”(t),...,ac”(t)](ii) a Specifically, principal component analysis is adopted as a dimension reduction algorithm, subsequences corresponding to a plurality of characteristic values with the accumulated contribution rate closest to 1 are obtained and used as final processing data, and therefore final processing data p are obtainedl'＝[pl1',pl2',...,plc']、p”(t)＝[p1”(t),p2”(t),...pc”(t)]、al'＝[al1',al2',...,alc']And a "(t) ═ a1”(t),a2”(t),...,ac”(t)]；
S5, constructing a service life prediction original model of each key device and a service life integrated prediction original model of each key device, and training the two constructed models by adopting the data obtained in the step S4, so as to obtain a service life prediction model of each key device of the train and a service life integrated prediction model of each key device of the train; specifically, the service life prediction model of each key device of the train is obtained by the following steps:
the method comprises the steps that I, an echo state network model optimized by a dragonfly algorithm is used as an original model, and an original service life prediction model for each key device is constructed; training the constructed original service life prediction model of each key device by adopting the processed data obtained in the step S4, thereby obtaining a service life prediction model of each key device of the train; specifically, the service life prediction model of each key device of the train is obtained by the following steps:
i-1. when the target key device is an electrical device, the processed subsequence p "(t) ═ p obtained in step S41”(t),p2”(t),...pc”(t)]As model input, the remaining service life subsequence p of each corresponding key devicel'＝[pl1',pl2',...,plc']As an output; when the target key device is a mechanical device, the processed subsequence a "(t) ═ a obtained in step S41”(t),a2”(t),...,ac”(t)]As model input, with each corresponding keyRemaining service life subsequence a of equipmentl'＝[al1',al2',...,alc']As an output;
i-2, initializing echo network state parameters: randomly generating a reserve pool to enable the spectrum radius to be smaller than R1 (preferably 1), randomly generating input connection and output feedback weights, enabling the input connection and the output feedback weights not to change any more in the training process, and adopting a tanh function as an activation function;
i-3, generating an initialization population: setting the value range of the dragonfly population quantity as [ N1, N2] (preferably [20,100]), and the value range of the maximum iteration number as [ N3, N4] (preferably [100,200 ]); parameters to be optimized of the echo state network are represented by the position of each population individual; N1-N4 are all natural numbers;
i-4, determining an optimization objective function: dividing input data for training the echo state network model into two parts, wherein N5% (preferably 80%) is used as a training set, and N6% is used as a verification set (preferably 20%) for verifying the performance of the current parameters; selecting a mean square error as an optimization objective function; n5 and N6 are both positive real numbers, and N5+ N6 is 100;
i-5, outputting the optimal parameters of the echo state network: optimizing the output connection weight of the echo state network by using the determined optimization objective function, finding a parameter value which enables the mean square error output value to be minimum within the maximum iteration times, using the parameter value as the optimal parameter of the echo state network, and obtaining a service life prediction model of each key device of the train;
taking an echo state network model optimized by a dragonfly algorithm as an original model, and constructing an original model for service life integrated prediction of each key device; training the constructed original service life integrated prediction model of each key device by adopting the processed data obtained in the step S4, thereby obtaining the service life integrated prediction model of each key device of the train; specifically, the service life integrated prediction model of each key device of the train is obtained by the following steps:
II-1, when the target key device is an electrical device, the processed subsequence p obtained in step S4l'＝[pl1',pl2',...,plc']As model input, using the residual service life sequence p of each corresponding key device of z trains in the original databasel＝[pl1,pl2,...,plz]As an output; when the target key device is a mechanical device, the processed subsequence a obtained in step S4l'＝[al1',al2',...,alc']As model input, the residual service life sequence a of each corresponding key device of z trainsl＝[al1,al2,...,alz]As an output;
II-2, initializing echo network state parameters: randomly generating a reserve pool to enable the spectrum radius to be smaller than R2 (preferably 1), randomly generating input connection and output feedback weights, enabling the input connection and the output feedback weights not to change any more in the training process, and adopting a tanh function as an activation function;
II-3, generating an initialization population: setting the value range of the dragonfly population quantity as [ N7, N8] (preferably [20,100]), and the value range of the maximum iteration number as [ N9, N10] (preferably [100,200 ]); parameters to be optimized of the echo state network are represented by the position of each population individual; N7-N10 are all natural numbers;
II-4, determining an optimization objective function: dividing input data for training the echo state network model into two parts, wherein N11% is used as a training set (preferably 80%), and N12% is used as a verification set (preferably 20%) for verifying the performance of the current parameters; selecting a mean square error as an optimization objective function; n11 and N12 are both positive real numbers, and N11+ N12 is 100;
II-5, outputting the optimal parameters of the echo state network: optimizing the output connection weight of the echo state network by using the determined optimization objective function, finding a parameter value which enables the mean square error output value to be minimum within the maximum iteration times, using the parameter value as the optimal parameter of the echo state network, and obtaining an integrated service life prediction model of each key device of the train;
s6, constructing a service life prediction original model of the whole train, and training the service life prediction original model by adopting the prediction result of the service life prediction model of each key device of the train obtained in the step S5, so as to obtain a service life prediction model of the whole train; specifically, a service life prediction model of the whole train is obtained by adopting the following steps:
using the prediction result of the service life prediction model of each key device of the train obtained in the step S5 as model input, and using the train service life sequence as model output;
training the multi-target ant colony optimization model by adopting the data in the step i, so as to obtain a final service life prediction model of the whole train; and (3) taking 80% of the data in the step i as a test set and 20% as a training set, and obtaining a final service life prediction model of the whole train by adopting the following steps:
ii-1, initializing parameters: by adopting multi-objective ant colony optimization, initializing the ant colony number to be M1 (preferably 100), the maximum iteration number to be M2 (preferably 200), and randomly initializing pheromone related parameters; m1 and M2 are both positive integers;
ii-2, determining an optimization variable: setting the numbers of electrical equipment and mechanical equipment of the train as m and n respectively, wherein f represents a utilization parameter c1,c2,...,cm+nThe service life of the whole train is obtained by fitting the residual service life of each device, and the optimized variable is the parameter c1,c2,...,cm+n，pli＝[pli(1),pli(2),...,pli(m)]Predicting the residual service life of the m key electrical devices of the ith train obtained in the step S5li＝[ali(1),ali(2),...,ali(n)]Predicting the remaining service life of the n key mechanical devices of the ith train obtained in step S5, wherein i is 1,2i＝c1pli(1)+c2pli(2)+...cmpli(m)+cm+1ali(1)+cm+2ali(2)+...+cm+nali(n)；
And ii-3, determining an optimization objective function:
Figure BDA0002683026500000201
Figure BDA0002683026500000202
in the formula IiReal service life data of the ith train in the z trains obtained in the step S1; f. ofiFitting results of the service life of the entire vehicle of the ith train;
Figure BDA0002683026500000203
the average value of the real data of the service life of the whole z trains is obtained;
Figure BDA0002683026500000204
the average value of the fitting results of the service life of the whole z trains is obtained;
ii-4. performing multi-objective optimization: enabling all ants to traverse all paths, calculating an optimization objective function value and recording a non-dominant solution corresponding to the optimal condition;
ii-5, continuously updating the search path and recording the optimal path, stopping searching when the maximum iteration times is reached, and outputting a non-dominated solution set;
ii-6, evaluating the performance of the non-dominated solution set by using the verification set to obtain the average absolute percentage error of the model output value
Figure BDA0002683026500000205
As an evaluation index, obtaining a non-dominated solution with the minimum E value as an optimal parameter output, and obtaining a final service life prediction model of the whole train;
s7, collecting real-time signals of the target train to be predicted in a non-invasive mode in real time, processing data, and predicting the service life of the key equipment of the target train to be predicted and the service life of the whole train through the service life prediction model of each key equipment of the train, the service life integrated prediction model of each key equipment of the train and the service life prediction model of the whole train, which are constructed in the step S5, and the service life prediction model of the whole train, which is constructed in the step S7; the method specifically includes the steps that on a target train to be predicted which normally runs, vibration acceleration data information and load active power information on a total measuring point of the target train to be predicted are collected in a non-invasive mode in real time, data processing is conducted in the mode of S2, then the model obtained in the step S3 is used for decomposition, data reprocessing is conducted in the mode of S4, then service life prediction of all devices is conducted through the model obtained in the step S5, service life prediction results of all devices are analyzed, for the devices which are about to reach the service life limit, if the devices can be maintained or replaced, device model information is recorded and reported, maintenance or replacement work of related technicians is waited, service life data of the corresponding devices are updated in real time, and finally service life prediction of the whole train is conducted through the model obtained in the step S6.

Claims (9)

1. A non-invasive prediction method for service life of a whole intelligent train comprises the following steps:
s1, carrying out signal acquisition on key equipment of a train in an intrusive mode to obtain sampling data of each key equipment and sampling data of a total measuring point;
s2, preprocessing the sampling data acquired in the step S1 to obtain key equipment characteristic parameters and total measuring point characteristic parameters;
s3, constructing a characteristic relation original model between the total measuring point and each key device, and training the characteristic relation original model by adopting the key device characteristic parameters obtained in the step S2, so as to obtain a non-invasive load decomposition model of each key device;
s4, processing the historical service life parameters of each key device of the train and the characteristic parameters of the key devices obtained in the step S2 again; the method specifically comprises the following steps:
1) using empirical wavelet transform algorithm to obtain data information p from step S1lAnd alAnd the data information p '(t) and a' (t) obtained in step S2 are decomposed into a fixed number of subsequences of high and low frequencies;
2) evaluating the stationarity of the subsequences obtained in the step 1) by using sample entropy;
3) according to the stationarity of the subsequence obtained in the step 2), adopting an ICEEMDAN algorithm to further decompose the unstable subsequence;
4) performing dimensionality reduction by using a principal component analysis method to obtain final processing data pl'＝[pl1',pl2',...,plc']、p”(t)＝[p1”(t),p2”(t),...pc”(t)]、al'＝[al1',al2',...,alc']And a "(t) ═ a1”(t),a2”(t),...,ac”(t)]；
S5, constructing a service life prediction original model of each key device and a service life integrated prediction original model of each key device, and training the two constructed models by adopting the data obtained in the step S4, so as to obtain a service life prediction model of each key device of the train and a service life integrated prediction model of each key device of the train;
s6, constructing a service life prediction original model of the whole train, and training the service life prediction original model by adopting the prediction result of the service life prediction model of each key device of the train obtained in the step S5, so as to obtain a service life prediction model of the whole train;
s7, collecting real-time signals of the target train to be predicted in a non-invasive mode in real time, processing data, and predicting the service life of the key equipment of the target train to be predicted and the service life of the whole train through the service life prediction model of each key equipment of the train, the service life integrated prediction model of each key equipment of the train and the service life prediction model of the whole train, which are constructed in the step S5, and the service life prediction model of the whole train, which is constructed in the step S7.
2. The intelligent non-intrusive prediction method for the service life of the whole train as claimed in claim 1, wherein in step S1, signal acquisition is performed on key devices of the train in an intrusive manner to obtain sampling data of each key device and sampling data of a total measuring point, specifically, the following steps are performed to obtain the sampling data of each key device and the sampling data of the total measuring point:
A. dividing the train into a mechanical part and an electrical part, and selecting corresponding key equipment for each part; meanwhile, selecting a total measuring point of the whole train;
B. selecting trains with different residual service lives;
C. acquiring a vibration acceleration signal and a load active power signal at each key device in the step A aiming at each train selected in the step B, and simultaneously acquiring a vibration acceleration signal and a load active power signal at a total measuring point;
D. and C, performing data processing on the signal data obtained in the step C to obtain final sampling data of each key device and sampling data of the total measuring point.
3. The intelligent non-intrusive prediction method for service life of the whole train as claimed in claim 2, wherein in step S2, the sampled data obtained in step S1 is preprocessed to obtain key device characteristic parameters and total measurement point characteristic parameters, specifically, the key device characteristic parameters are obtained by the following steps:
a. transforming all the original sampling data acquired in the step S1 into high and low frequency signals by adopting Fourier transform, and denoising by adopting a low-pass filter, thereby obtaining effective characteristic parameters of the total measuring point and effective characteristic parameters of each key device;
b. b, normalizing the effective characteristic parameters obtained in the step a, so as to ensure that the data after normalization meets the standard normal distribution;
c. storing the data information processed in the step b in a uniform format by adopting a sliding window mode, thereby obtaining a final key equipment characteristic parameter p' ═ pe,pl,pt,p'(t)]And a ═ ae,al,at,a'(t)]And a total measured point characteristic parameter P ═ Pl,Pt,P'(t)]And a ═ al,At,A'(t)](ii) a Wherein P '(t) is a processed load active power signal value of the electrical equipment, a' (t) is a processed signal value of a vibration acceleration signal of the mechanical equipment, P '(t) is a processed load active power signal value at a total measurement point, and a' (t) is a processed signal value of the vibration acceleration signal at the total measurement point;pethe model of the electrical equipment; p is a radical oflThe remaining service life of the electrical equipment; p is a radical oftIs a time tag of the electrical device; a iseThe model of mechanical equipment; a islThe remaining service life of the mechanical equipment; a istA time tag for the mechanical device; p islThe remaining service life of the electrical part of the train; ptTime labels at the total survey points; a. thelThe remaining service life of the mechanical part of the train is represented; a. thetTime stamp at the total endpoint.
4. The intelligent non-intrusive prediction method for service life of the whole train as claimed in claim 3, wherein the step S3 is to construct a feature relationship original model between the total measurement point and each of the key devices, and train the feature relationship original model by using the feature parameters of the key devices obtained in the step S2, so as to obtain a non-intrusive load decomposition model of each of the key devices, specifically, the following steps are adopted to obtain the non-intrusive load decomposition model of each of the key devices:
(1) embedding the convolutional layer into a gate control cycle unit neural network, thereby constructing and obtaining a non-invasive load decomposition model;
(2) dividing the data processed in the step S2 into a training set and a verification set, and training the non-invasive load decomposition preliminary models of the key equipment one by one;
(3) and (3) verifying the non-invasive load decomposition preliminary model obtained in the step (2) by using a verification set, and selecting the model with the best error function as the final non-invasive load decomposition model of each key device.
5. The intelligent non-intrusive prediction method for service life of the whole intelligent train is characterized in that in the step (1), the convolutional layer is built in a gated cyclic unit neural network, so that a non-intrusive load decomposition model is constructed, specifically, the convolutional layer is built in the gated cyclic unit neural network, and the obtained neural network comprises an input layer, a convolutional layer, a hidden layer and an output layer; the hidden layer comprises a gating circulation unit and a full connection layer; adopting a ReLU function and a linear function as activation functions; adopting a cross entropy function as a loss function; and performing subsequent model training by using an NAG algorithm as a parameter optimization algorithm, and setting a learning rate eta to be 0.01 and an attenuation rate gamma to be 0.9.
6. The intelligent non-intrusive forecasting method for the service life of the whole train as recited in claim 5, wherein the step S5 is implemented by constructing an original service life forecasting model of each key device and an original service life integrated forecasting model of each key device, and training the two constructed models by using the data obtained in the step S4, so as to obtain the service life forecasting models of each key device of the train and the service life integrated forecasting models of each key device of the train, specifically, the service life forecasting models of each key device of the train are obtained by using the following steps:
the method comprises the steps that I, an echo state network model optimized by a dragonfly algorithm is used as an original model, and an original service life prediction model for each key device is constructed; training the constructed original service life prediction model of each key device by adopting the processed data obtained in the step S4, thereby obtaining a service life prediction model of each key device of the train;
taking an echo state network model optimized by a dragonfly algorithm as an original model, and constructing an original model for service life integrated prediction of each key device; and training the constructed original service life integrated prediction model of each key device by adopting the processed data obtained in the step S4, thereby obtaining the service life integrated prediction model of each key device of the train.
7. The intelligent non-intrusive forecasting method for the service life of the whole intelligent train as claimed in claim 6, is characterized in that the echo state network model optimized by the dragonfly algorithm in the step I is used as an original model, and an original service life forecasting model for each key device is constructed; and training the constructed original service life prediction model of each key device by using the processed data obtained in the step S4, so as to obtain a service life prediction model of each key device of the train, specifically obtaining the service life prediction model of each key device of the train by using the following steps:
i-1. when the target key device is an electrical device, the processed subsequence p "(t) ═ p obtained in step S41”(t),p2”(t),...pc”(t)]As model input, the remaining service life subsequence p of each corresponding key devicel'＝[pl1',pl2',...,plc']As an output; when the target key device is a mechanical device, the processed subsequence a "(t) ═ a obtained in step S41”(t),a2”(t),...,ac”(t)]As model input, the residual service life subsequence a of each corresponding key devicel'＝[al1',al2',...,alc']As an output;
i-2, initializing echo network state parameters: randomly generating a reserve pool to enable the spectrum radius to be smaller than R1, randomly generating input connection and output feedback weights, and taking a tanh function as an activation function without changing in the training process;
i-3, generating an initialization population: setting the value range of the dragonfly population quantity as [ N1, N2], and the value range of the maximum iteration number as [ N3, N4 ]; parameters to be optimized of the echo state network are represented by the position of each population individual; N1-N4 are natural numbers;
i-4, determining an optimization objective function: dividing input data for training an echo state network model into two parts, wherein N5% serves as a training set, and N6% serves as a verification set for verifying the performance of the current parameters; selecting a mean square error as an optimization objective function; n5 and N6 are both positive real numbers, and N5+ N6 is 100;
i-5, outputting the optimal parameters of the echo state network: optimizing the output connection weight of the echo state network by using the determined optimization objective function, finding a parameter value which enables the mean square error output value to be minimum within the maximum iteration times, using the parameter value as the optimal parameter of the echo state network, and obtaining a service life prediction model of each key device of the train;
step II, constructing an integrated service life prediction original model of each key device by taking the echo state network model optimized by the dragonfly algorithm as an original model; and training the constructed original service life integrated prediction model of each key device by adopting the processed data obtained in the step S4, thereby obtaining the service life integrated prediction model of each key device of the train, specifically obtaining the service life integrated prediction model of each key device of the train by adopting the following steps:
II-1, when the target key device is an electrical device, the processed subsequence p obtained in step S4l'＝[pl1',pl2',...,plc']As model input, using the residual service life sequence p of each corresponding key device of z trains in the original databasel＝[pl1,pl2,...,plz]As an output; when the target key device is a mechanical device, the processed subsequence a obtained in step S4l'＝[al1',al2',...,alc']As model input, the residual service life sequence a of each corresponding key device of z trainsl＝[al1,al2,...,alz]As an output;
II-2, initializing echo network state parameters: randomly generating a reserve pool to enable the spectrum radius to be smaller than R2, randomly generating input connection and output feedback weights, and taking a tanh function as an activation function without changing in the training process;
II-3, generating an initialization population: setting the value range of the dragonfly population quantity as [ N7, N8], and the value range of the maximum iteration number as [ N9, N10 ]; parameters to be optimized of the echo state network are represented by the position of each population individual; N7-N10 are all natural numbers;
II-4, determining an optimization objective function: dividing input data for training an echo state network model into two parts, wherein N11% serves as a training set, and N12% serves as a verification set for verifying the performance of the current parameters; selecting a mean square error as an optimization objective function; n11 and N12 are both positive real numbers, and N11+ N12 is 100;
II-5, outputting the optimal parameters of the echo state network: and optimizing the output connection weight of the echo state network by using the determined optimization objective function, finding a parameter value which enables the mean square error output value to be minimum within the maximum iteration times, and taking the parameter value as the optimal parameter of the echo state network to obtain an integrated service life prediction model of each key device of the train.
8. The intelligent full-train service life non-intrusive prediction method of claim 7, is characterized in that the service life prediction original model of the full-train is constructed in step S6, the service life prediction original model is trained by using the prediction results of the service life prediction models of the key devices of the train obtained in step S5, so as to obtain the service life prediction model of the full-train, and specifically, the service life prediction model of the full-train is obtained by using the following steps:
using the prediction result of the service life prediction model of each key device of the train obtained in the step S5 as model input, and using the train service life sequence as model output;
and ii, training the multi-target ant colony optimization model by adopting the data in the step i, thereby obtaining a final service life prediction model of the whole train.
9. The intelligent full-train service life non-intrusive prediction method as recited in claim 8, wherein the data in step ii is used to train the multi-objective ant colony optimization model, so as to obtain a final service life prediction model of the full-train, specifically, the following steps are used to obtain the final service life prediction model of the full-train:
ii-1, initializing parameters: initializing the ant colony number to be M1, the maximum iteration number to be M2 and initializing pheromone related parameters randomly by adopting multi-objective ant colony optimization; m1 and M2 are both positive integers;
ii-2, determining an optimization variable: setting the numbers of electrical equipment and mechanical equipment of the train as m and n respectively, wherein f represents a utilization parameter c1,c2,...,cm+nIs fitted with the residual service life of each deviceThe service life of the whole train is optimized by using the parameter c1,c2,...,cm+n，pli＝[pli(1),pli(2),...,pli(m)]Predicting the residual service life of the m key electrical devices of the ith train obtained in the step S5li＝[ali(1),ali(2),...,ali(n)]If i is 1, 2.. times, z, the fitting result of the service life of the entire vehicle of the ith train is fi＝c1pli(1)+c2pli(2)+...cmpli(m)+cm+1ali(1)+cm+2ali(2)+...+cm+nali(n)；
And ii-3, determining an optimization objective function:
Figure FDA0003621058180000081
Figure FDA0003621058180000082
in the formula IiReal service life data of the ith train in the z trains obtained in the step S1; f. ofiFitting results of the service life of the entire vehicle of the ith train;
Figure FDA0003621058180000083
the average value of the real data of the service life of the whole z trains is obtained;
Figure FDA0003621058180000084
the average value of the fitting results of the service life of the whole z trains is obtained;
ii-4. performing multi-objective optimization: enabling all ants to traverse all paths, calculating an optimization objective function value and recording a non-dominant solution corresponding to the optimal condition;
ii-5, continuously updating the search path and recording the optimal path, stopping searching when the maximum iteration times is reached, and outputting a non-dominated solution set;
ii-6, evaluating the performance of the non-dominated solution set by using the verification set to obtain the average absolute percentage error of the model output value
Figure FDA0003621058180000085
And (5) as an evaluation index, obtaining a non-dominated solution with the minimum E value as an optimal parameter to be output, and obtaining a final service life prediction model of the whole train.
CN202010967935.2A 2020-09-15 2020-09-15 Smart train service life non-invasive prediction method Active CN112115643B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010967935.2A CN112115643B (en) 2020-09-15 2020-09-15 Smart train service life non-invasive prediction method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010967935.2A CN112115643B (en) 2020-09-15 2020-09-15 Smart train service life non-invasive prediction method

Publications (2)

Publication Number Publication Date
CN112115643A CN112115643A (en) 2020-12-22
CN112115643B true CN112115643B (en) 2022-06-10

Family

ID=73802731

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010967935.2A Active CN112115643B (en) 2020-09-15 2020-09-15 Smart train service life non-invasive prediction method

Country Status (1)

Country Link
CN (1) CN112115643B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112800855B (en) * 2021-01-04 2022-07-12 中南大学 Non-invasive real-time fault monitoring method for train bogie
CN112816052B (en) * 2021-01-04 2022-02-08 中南大学 Train compartment vibration monitoring method, vibration signal characteristic library establishment and application method
CN113285448B (en) * 2021-05-25 2022-10-04 国网四川省电力公司电力科学研究院 Non-invasive traction load decomposition method for electrified railway

Citations (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8781982B1 (en) * 2011-09-23 2014-07-15 Lockheed Martin Corporation System and method for estimating remaining useful life
CN105973597A (en) * 2016-05-27 2016-09-28 北京交通大学 Test and prediction method for service life of bearing of axle box of train
CN109145319A (en) * 2017-06-16 2019-01-04 哈尔滨理工大学 Key equipment cutting tool method for predicting residual useful life based on deep neural network
CN110161343A (en) * 2019-06-12 2019-08-23 中南大学 A kind of non-intrusion type real-time dynamic monitoring method of intelligence train exterior power receiving device
CN110361180A (en) * 2019-07-25 2019-10-22 中南大学 Intelligent train pantograph military service performance dynamic monitoring and appraisal procedure and its system
CN110376003A (en) * 2019-07-25 2019-10-25 中南大学 Intelligent whole train service life prediction technique and its system based on BIM
CN110428066A (en) * 2019-07-25 2019-11-08 中南大学 It is a kind of intelligence Electric device presence assessment with operation platform and its method
CN110555230A (en) * 2019-07-12 2019-12-10 北京交通大学 rotary machine residual life prediction method based on integrated GMDH framework
CN110803192A (en) * 2018-08-06 2020-02-18 中车永济电机有限公司 Train-mounted PHM equipment and high-speed rail train
CN110807257A (en) * 2019-11-04 2020-02-18 中国人民解放军国防科技大学 Method for predicting residual life of aircraft engine
CN111191400A (en) * 2019-12-31 2020-05-22 上海钧正网络科技有限公司 Vehicle part service life prediction method and system based on user fault reporting data
CN111242357A (en) * 2020-01-06 2020-06-05 北京锦鸿希电信息技术股份有限公司 Neural network learning-based train-mounted equipment fault prediction method and device

Patent Citations (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8781982B1 (en) * 2011-09-23 2014-07-15 Lockheed Martin Corporation System and method for estimating remaining useful life
CN105973597A (en) * 2016-05-27 2016-09-28 北京交通大学 Test and prediction method for service life of bearing of axle box of train
CN109145319A (en) * 2017-06-16 2019-01-04 哈尔滨理工大学 Key equipment cutting tool method for predicting residual useful life based on deep neural network
CN110803192A (en) * 2018-08-06 2020-02-18 中车永济电机有限公司 Train-mounted PHM equipment and high-speed rail train
CN110161343A (en) * 2019-06-12 2019-08-23 中南大学 A kind of non-intrusion type real-time dynamic monitoring method of intelligence train exterior power receiving device
CN110555230A (en) * 2019-07-12 2019-12-10 北京交通大学 rotary machine residual life prediction method based on integrated GMDH framework
CN110361180A (en) * 2019-07-25 2019-10-22 中南大学 Intelligent train pantograph military service performance dynamic monitoring and appraisal procedure and its system
CN110376003A (en) * 2019-07-25 2019-10-25 中南大学 Intelligent whole train service life prediction technique and its system based on BIM
CN110428066A (en) * 2019-07-25 2019-11-08 中南大学 It is a kind of intelligence Electric device presence assessment with operation platform and its method
CN110807257A (en) * 2019-11-04 2020-02-18 中国人民解放军国防科技大学 Method for predicting residual life of aircraft engine
CN111191400A (en) * 2019-12-31 2020-05-22 上海钧正网络科技有限公司 Vehicle part service life prediction method and system based on user fault reporting data
CN111242357A (en) * 2020-01-06 2020-06-05 北京锦鸿希电信息技术股份有限公司 Neural network learning-based train-mounted equipment fault prediction method and device

Non-Patent Citations (6)

* Cited by examiner, † Cited by third party
Title
A Model-Based Method for Remaining Useful Life Prediction of Machinery;Yaguo Lei等;《IEEE Transactions on Reliability 》;20160727;第1314-1326页 *
Fatigue life reliability evaluation in a high-speed train bogie frame using accelerated life and numerical test;Y Lu等;《Reliability Engineering & System Safety》;20190831;第221-232页 *
Non-intrusive Load Identification Method Based on Integrated Intelligence Strategy;Ren Zhiren等;《2019 25th International Conference on Automation and Computing 》;20191111;第1-6页 *
动车组转向架轴箱剩余寿命预测方法研究;赵珂等;《软件》;20200315(第3期);第219-224页 *
城轨车辆关键部件状态修技术与系统解决方案;刘可安等;《机车电传动》;20200710(第04期);第1-7页 *
基于大数据和云计算的车辆智能运维模式;曹勇等;《城市轨道交通研究》;20200410(第04期);第69-73页 *

Also Published As

Publication number Publication date
CN112115643A (en) 2020-12-22

Similar Documents

Publication Publication Date Title
CN112115643B (en) Smart train service life non-invasive prediction method
CN112395959B (en) Power transformer fault prediction and diagnosis method and system based on audio frequency characteristics
CN109948117A (en) A kind of satellite method for detecting abnormality fighting network self-encoding encoder
CN111442827B (en) Optical fiber passive online monitoring system for transformer winding vibration
CN110702418A (en) Aircraft engine fault prediction method
CN110610035A (en) Rolling bearing residual life prediction method based on GRU neural network
CN108168924B (en) Reciprocating compressor service life prediction method based on VMD (model-based decomposition) and MFSS (model-based decomposition) models
Cacciola et al. Aspects about air pollution prediction on urban environment
KR100795227B1 (en) Method and apparatus for analyzing signal pattern of sensor array
Zhao et al. A novel cap-LSTM model for remaining useful life prediction
CN112070322A (en) High-voltage cable line running state prediction method based on long-short term memory network
CN113138005A (en) Vehicle random load pattern recognition system based on feature extraction and neural network
Chen et al. Real-time bearing remaining useful life estimation based on the frozen convolutional and activated memory neural network
CN115659583A (en) Point switch fault diagnosis method
CN115375026A (en) Method for predicting service life of aircraft engine in multiple fault modes
CN116415723A (en) Wind speed prediction method and device for railway disaster prevention system
CN114563150B (en) Bridge health online detection module generation method, detection method, tool box and device
CN117074961A (en) Variation self-encoder power battery abnormality detection method, system, device and storage medium
CN115563880A (en) Enterprise power consumption abnormity detection method based on Isolated forest-variable point enhancement
Qiu et al. A piecewise method for bearing remaining useful life estimation using temporal convolutional networks
CN115809716A (en) Flight path prediction method and device based on multi-resolution analysis and wavelet packet reconstruction
Zhang et al. Remaining useful life prediction for rolling bearings with a novel entropy-based health indicator and improved particle filter algorithm
CN114595721A (en) Vehicle part strain load prediction method based on road condition recognition
CN113673774A (en) Aero-engine remaining life prediction method based on self-encoder and time sequence convolution network
Melendez et al. Self-supervised Multi-stage Estimation of Remaining Useful Life for Electric Drive Units

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant