CN112798888B - Non-invasive fault diagnosis method for vehicle-mounted electrical system of unmanned train - Google Patents
Non-invasive fault diagnosis method for vehicle-mounted electrical system of unmanned train Download PDFInfo
- Publication number
- CN112798888B CN112798888B CN202011616269.4A CN202011616269A CN112798888B CN 112798888 B CN112798888 B CN 112798888B CN 202011616269 A CN202011616269 A CN 202011616269A CN 112798888 B CN112798888 B CN 112798888B
- Authority
- CN
- China
- Prior art keywords
- classifier
- deep learning
- learning model
- data
- time sequence
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 60
- 238000003745 diagnosis Methods 0.000 title claims abstract description 38
- 238000013136 deep learning model Methods 0.000 claims abstract description 40
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 33
- 239000013598 vector Substances 0.000 claims abstract description 23
- 238000010187 selection method Methods 0.000 claims abstract description 18
- 238000012549 training Methods 0.000 claims description 47
- 238000012360 testing method Methods 0.000 claims description 26
- 238000000605 extraction Methods 0.000 claims description 24
- 238000004422 calculation algorithm Methods 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 10
- 230000002787 reinforcement Effects 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 5
- 238000005457 optimization Methods 0.000 claims description 5
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 238000004458 analytical method Methods 0.000 abstract description 14
- 238000004364 calculation method Methods 0.000 description 13
- 239000000306 component Substances 0.000 description 13
- 238000013527 convolutional neural network Methods 0.000 description 10
- 230000008569 process Effects 0.000 description 10
- 239000003795 chemical substances by application Substances 0.000 description 7
- 238000005516 engineering process Methods 0.000 description 7
- 238000012544 monitoring process Methods 0.000 description 7
- 230000006399 behavior Effects 0.000 description 6
- 238000001514 detection method Methods 0.000 description 6
- 210000002569 neuron Anatomy 0.000 description 6
- 239000000284 extract Substances 0.000 description 5
- 239000011159 matrix material Substances 0.000 description 5
- 238000011161 development Methods 0.000 description 4
- 230000018109 developmental process Effects 0.000 description 4
- 230000009471 action Effects 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000005070 sampling Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000004880 explosion Methods 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 239000003990 capacitor Substances 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 239000008358 core component Substances 0.000 description 1
- 238000013075 data extraction Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000009413 insulation Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 230000001502 supplementing effect Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 238000011426 transformation method Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/005—Testing of electric installations on transport means
- G01R31/008—Testing of electric installations on transport means on air- or spacecraft, railway rolling stock or sea-going vessels
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Supply And Distribution Of Alternating Current (AREA)
Abstract
The invention discloses a non-invasive fault diagnosis method for a vehicle-mounted electrical system of an unmanned train, which comprises the following steps: collecting multiple groups of modeling data under different known fault types; obtaining a trained deep learning model based on a deep learning model taking the total voltage and total current time series as input vectors and the voltage and current time series of each electrical device as output vectors; extracting power characteristics; determining a feature selection method and a classifier; collecting multiple groups of measured data, inputting a total voltage time sequence and a total current time sequence corresponding to the measured data into a deep learning model as input vectors, extracting electric power characteristics in each output vector output by the measured data after passing through the deep learning model, selecting the electric power characteristics corresponding to the measured data, inputting the electric power characteristics into a classifier, and outputting the fault type of the electric system to be diagnosed by the classifier. The method has the advantages of high power load decomposition accuracy, high power characteristic analysis performance, high fault diagnosis result accuracy and good timeliness.
Description
Technical Field
The invention belongs to the technical field of fault diagnosis, and particularly relates to a non-invasive fault diagnosis method for a vehicle-mounted electric system of an unmanned train.
Background
With the development of the times, rail transit plays an increasingly important role in the global public transportation field. Urgent government and social demands place ever-increasing demands on the safety, efficiency and operating costs of rail transit. Rail transit intelligence is one of the cores of current and future rail transit industry developments. Unmanned railway vehicles are an important embodiment and core representative of the level of intelligence in the rail transit industry. For an unmanned intelligent train, an electric system is a core component for ensuring the safety and comfort of the whole train.
In order to improve the transportation safety guarantee capability, improve the transportation service quality, improve the transportation efficiency and ensure the safety of passengers, the fault diagnosis and real-time monitoring technology of the electrical system plays an increasingly important role.
Although the traditional invasive electrical system monitoring technology has high detection accuracy, the arrangement cost is high. In addition, with the development of scientific technology, the variety of vehicle-mounted electrical equipment is increasing continuously, and the invasive electrical system monitoring system is difficult to fully cover the electrical system monitoring of the whole train.
With the continuous development of computer science and big data technology, the non-invasive power load identification technology is effectively applied. The technology can effectively reduce the cost and realize the real-time effective monitoring of the electrical system.
At present, the difficulty of real-time monitoring and analysis of a train electrical system is gradually increased along with the change of the power load of the train electrical system and the increase of different power load types.
The monitoring of indexes such as current, voltage, power and the like is very important for realizing the stability of an electric system and fault analysis and early warning. To realize non-invasive fault diagnosis of an electrical system, it is most important to realize real-time decomposition of power loads such as total power and the like, extract power characteristics of different electrical devices, and realize real-time diagnosis and analysis of the different electrical devices according to the power characteristics. At present, the mainstream power load decomposition method is a hidden markov method based on unsupervised learning, which is effectively used, but as the types of power loads increase, the recognition capability thereof is difficult to be further improved.
On the other hand, the information redundancy can be effectively reduced for the power characteristic analysis of different electrical equipment data, and the detection accuracy is improved. At present, the mainstream feature selection method is a heuristic method based on population, and although the method is widely researched, the ultimate feature analysis performance cannot be effectively broken through due to limited evolutionary capability.
Finally, real-time and high-precision fault type diagnosis is realized, and effective classifier selection is very important. The traditional RNN method cannot realize parallel computation, so that the computation time of a model is long, and the requirement of the diagnosis effectiveness of an electrical system cannot be met.
Disclosure of Invention
The present invention is directed to provide a method for non-invasive fault diagnosis of an electrical system mounted on an unmanned train, which overcomes the above-mentioned drawbacks of the prior art.
In order to solve the technical problem of low accuracy of power load decomposition in the existing non-invasive electric system fault diagnosis method, the technical scheme adopted by the invention is as follows:
a non-invasive fault diagnosis method for a vehicle-mounted electric system of an unmanned train is characterized by comprising the following steps:
a, collecting multiple groups of modeling data under M different known fault types, wherein each group of modeling data comprises total voltage data of an electrical system, total current data of the electrical system, voltage data of each electrical device in the electrical system and current data of each electrical device in the electrical system under each fault type; after converting each group of modeling data into a time sequence with the same rated working frequency, dividing a plurality of groups of converted modeling data into a training set and a test set according to a set proportion;
step B, training the deep learning model by using a training set based on a deep learning model taking the total voltage time sequence and the total current time sequence as input vectors and the voltage time sequence of each electrical device and the current time sequence of each electrical device as output vectors, testing the deep learning model by using a testing set, updating the deep learning model and continuing training when the accuracy of the testing result does not meet the set requirement, and otherwise terminating the training and obtaining the trained deep learning model;
step C, extracting the electric power characteristics in each output vector output by the training set after passing through the deep learning model by using a characteristic extraction method;
step D, selecting the extracted power features by using a feature selection method, training the classifier by using a training set based on the classifier which takes the selected power features as input quantity and takes the power system fault diagnosis type as output quantity, testing the classifier by using a test set, continuing training based on the updated feature selection method and the classifier when the fault type diagnosis accuracy of the test result does not meet the set requirement, and otherwise terminating the training and determining the feature selection method and the trained classifier;
step E, collecting a plurality of groups of measured data of the electrical system to be diagnosed, wherein each group of measured data comprises total voltage data and total current data of the electrical system to be diagnosed, converting each group of measured data into a total voltage time sequence and a total current time sequence with the same rated working frequency, inputting the total voltage time sequence and the total current time sequence corresponding to the measured data as input vectors into the deep learning model determined in the step B, extracting the electric power characteristics in each output vector output after the measured data passes through the deep learning model by using the characteristic extraction method in the step C, and selecting the electric power characteristics corresponding to the measured data by using the characteristic selection method determined in the step D, and inputting the power characteristics corresponding to the selected obtained measured data into the classifier determined in the step D, and outputting the fault type of the electrical system to be diagnosed by the classifier.
By means of the method, the deep learning method is used as a new non-invasive power load decomposition method to decompose the total voltage and total current data, and compared with a traditional unsupervised learning method, the method can effectively realize accurate electric load classification by analyzing the characteristics and actual conditions of the existing big data, reduces analysis errors and improves the accuracy of power load decomposition.
Preferably, the deep learning model in step B includes a CNN framework and a QRNN framework. The CNN framework is used for extracting the characteristics of the acquired original data, and the QRNN framework is used for realizing power load decomposition, so that the dimensionality reduction of the original total voltage data and the original total current data is effectively realized, and better data characteristics are obtained subsequently. The CNN method can effectively extract deep features of the electrical equipment load time sequence. The raw load data is said to have long and short term dependencies. Whereas the QRNN structure is able to efficiently analyze time series with long and short term dependencies. In addition, compared with the traditional RNN, the QRNN has the capability of parallel operation calculation, so that the QRNN can effectively relieve the problems of low training efficiency, gradient explosion, shaving loss and the like. Therefore, the present invention adopts CNN-QRNN as the algorithm of power load decomposition.
Preferably, the feature extraction method in step C includes a physical feature extraction method and a time series feature extraction method.
The invention adopts a mixed characteristic data extraction method based on physical characteristics and time sequence characteristics, thereby not only adopting the traditional physical characteristic extraction method to analyze the basic physical characteristics of each electrical device, but also adopting the time sequence characteristic extraction method to extract the deep-layer voltage and current fluctuation characteristics of each electrical device. The mixed feature extraction method provides more detailed information data and provides effective help for fault classification of a subsequent classifier.
As a preferable mode, in the step C, the time series feature extraction method includes:
decomposing the voltage time sequence and the current time sequence output by the deep learning model to obtain an IMF component with the minimum sum of K bandwidths;
and calculating the multi-scale permutation entropy of each IMF component, and taking the obtained permutation entropy characteristic data as time sequence characteristics.
As a preferable mode, in order to solve the technical problem of low feature analysis performance in the existing non-invasive electrical system fault diagnosis method, in the step D, the feature selection method is a reinforcement learning method.
The invention adopts the reinforcement learning method to replace the traditional heuristic method for feature selection, and the reinforcement learning feature selection method can be continuously evolved in the learning process, has good optimization performance and high intelligent degree, so that the feature analysis capability is far higher than that of the heuristic method, thereby ensuring that the subsequent classifier has stronger feature sensitivity and improving the feature analysis performance.
As a preferable mode, in order to solve the technical problems of long calculation time and poor timeliness existing in the existing non-invasive electrical system fault diagnosis method, the classifier in the step D is an SRU classifier.
On the basis of meeting the precision requirement, it is very important to select a classifier with good timeliness. In order to improve timeliness on the basis of ensuring performance, the fault identification model adopted by the invention uses the latest SRU classifier as the electrical equipment state classifier. The SRU classifier effectively combines the advantages of the CNN, the RNN and the GPU, greatly improves the calculation efficiency of the model, and enables the model to realize state detection and fault diagnosis of each electrical device more timely.
As a preferred mode, the activation function in the CNN framework is a ReLU function, and the optimization algorithm in the CNN framework is an Adam algorithm.
Preferably, the voltage time series and the current time series output by the deep learning model are decomposed by a variation modal decomposition method, so that the time series data decomposition has better robustness.
Compared with the prior art, the method has the advantages of high power load decomposition accuracy, high power characteristic analysis performance, high fault diagnosis result accuracy and good timeliness.
Drawings
FIG. 1 is a schematic flow chart of the method of the present invention.
Detailed Description
Taking the fault diagnosis of the vehicle-mounted electrical system on the unmanned train as an example, as shown in fig. 1, the invention firstly adopts the deep learning technology to replace the original hidden markov method to realize the real-time decomposition of the power load; then, extracting basic features and deep time sequence fluctuation features of the data by using a physical feature extraction method and a mixed VMD-MPE time sequence feature extraction method; moreover, the reinforcement learning method is used for replacing the traditional heuristic method to realize the selection of the electric power characteristics of different electrical equipment; and finally, the real-time state analysis and fault classification diagnosis of the electrical equipment are realized by adopting an SRU classifier meeting the requirements of high precision and effectiveness.
Specific implementation procedures are given below.
Step A: unmanned train electrical system signal acquisition and preprocessing
Because different electrical equipment has different indexes of fault conditions, sudden changes of voltage, current, power and the like can be the basis for judging the faults of different electrical equipment. The diversity and the sufficiency of the data are ensured, and the method is particularly important for the training of a subsequent neural network and a classifier. Therefore, according to the invention, firstly, a plurality of groups of modeling data under M different known fault types on the unmanned train are collected at the same time by a sampling frequency of 10kHz, wherein each group of modeling data comprises total voltage data of the electrical system, total current data of the electrical system, voltage data of each electrical device in the electrical system and current data of each electrical device in the electrical system under each fault type. And then, converting the building module data into load time sequence sequences U and I with the same rated working frequency (such as the rated working frequency of 50Hz) by a mode of solving an effective value, and correspondingly obtaining a total voltage time sequence, a total current time sequence, a voltage time sequence of each electrical device and a current time sequence of each electrical device.
In the process of manually acquiring data, the data is interfered by factors such as human factors, bad detection tools and the like, so that error information and missing values may exist, and the original sampling data needs to be preprocessed in the sampling process. On one hand, the usability and diversity of the sample are ensured by preprocessing data by means of eliminating error information, supplementing missing values by methods such as interpolation fitting and the like. On the other hand, label information needs to be added to the data, and the originally acquired data corresponds to each fault state of the electrical system one by one, so that the required label information is provided for the training of the subsequent classifier.
In the present embodiment, the electric apparatus mainly includes a power transformer, an engine, a pantograph, an inverter, a converter, an ac motor, and the like. For different electrical devices, fault signatures are set according to different fault types (short circuit, open circuit, joint discharge, and open phase).
And finally, dividing the preprocessed and converted multiple groups of modeling data into a training set and a testing set so as to ensure that the neural network and the classifier are fully trained subsequently and the performance of the model is accurately tested.
And B: CNN-QRNN-based raw data decomposition
The power load resolution task is very similar to the voice recognition task. When the electrical equipment operates, the voltage, the current and other sequences of the electrical equipment have obvious time sequence characteristics. Therefore, the neural network has strong adaptability to the modeling of the decomposition model. In the step, based on a deep learning model taking a total voltage time sequence and a total current time sequence as input vectors and taking a voltage time sequence of each electrical device and a current time sequence of each electrical device as output vectors, the deep learning model is trained by using a training set, meanwhile, the deep learning model is tested by using a test set, when the accuracy of a test result does not meet a set requirement, the deep learning model is updated and training is continued, otherwise, the training is terminated and the trained deep learning model is obtained. And step B, modeling each electrical device one by one to achieve the purpose of non-invasive electrical system power load decomposition.
Step B1: CNN-QRNN initial parameter frame structure setting
In order to realize the power load decomposition of the non-invasive electric system, the input of the deep learning model is the total voltage time series of the unmanned train electric system and the total current time series of the electric system, and the output result is the voltage time series of each electric device and the current time series of each electric device after decomposition.
The non-invasive electric system power load decomposition is realized by adopting a CNN-QRNN deep learning model, and the deep learning model comprises two frames, namely a CNN frame and a QRNN frame. CNN is used to extract features of the original data. The QRNN is used for realizing power load decomposition, and the framework can effectively realize the dimensionality reduction of original voltage and current data, so that a subsequent QRNN structure obtains better data characteristics. For the CNN structure, an activation function and an optimization algorithm are important parameters for optimizing the feature extraction capability of the model. The invention adopts a ReLU function and an Adam algorithm as an activation function and an optimization algorithm. The CNN method can effectively extract deep features of the electrical equipment load time sequence. The raw load data is said to have long and short term dependencies. Whereas the QRNN structure is able to efficiently analyze time series with long and short term dependencies. In addition, compared with the traditional RNN, the QRNN has the capability of parallel operation calculation, so that the QRNN can effectively relieve the problems of low training efficiency, gradient explosion, shaving loss and the like. Therefore, the present invention adopts CNN-QRNN as the algorithm of power load decomposition.
Step B2: training and verification of CNN-QRNN deep learning model
The method realizes the power load decomposition of the electrical system through the mixed CNN-QRNN deep learning model. And dividing the collected multiple groups of modeling data into a training set and a testing set according to the ratio of 3: 1. The training set is used to train the parameters of the CNN-QRNN deep learning model. The test set is used for verifying the decomposition precision and robustness of the deep learning model.
The input vector I is: i ═ I(1),I(2),…,I(i),…,I(n),U(1),U(2)…,U(i),…,U(n)]，I(i)Representing the total current time series, U(i)Representing the total voltage time series and n representing the number of sample points. The number of neurons in the output layer is determined according to the total number of electrical devices in the electrical system.
In the invention, the total number of the electrical equipment in the unmanned train is set to be Z, and the output vector O is as follows: o ═ Io(1),Io(1),…,Io(k),…,Io(z),Uo(z),Uo(1),…,Uo(k),…,Uo(z)]，Io(j)Representing the current time series, U, of the k-th electrical device after decompositiono(j)Representing the voltage time series of the kth electrical device after decomposition.
In the training process, the deep learning model is tested by using the test set, and when the precision of the deep learning model meets the requirement that the decomposition accuracy is higher than 99%, the CNN-QRNN model is used for decomposing the power load condition of the unmanned train in operation in real time. The decomposed voltage time series and current time series of each electrical device are subjected to data processing in the next step and are used for realizing state detection and fault diagnosis of different electrical devices.
And C: extracting power characteristics of different decomposed electrical devices
The invention is based on a mixed physical characteristic extraction method and a VMD-MPE time sequence characteristic extraction method, extracts the characteristics containing the physical information of original electrical equipment and deep time sequence stability information, and trains a subsequent SRU classifier through the extracted data characteristics.
Step C1: physical feature extraction
Generally, the core of the electrical equipment feature extraction is analysis and extraction of various load characteristics. The physical characteristics of the load are the most common load characteristics including active power, reactive power, impedance, admittance, etc. The physical features are typically superposable and the model is well able to analyze and distinguish the physical features. The focus of the invention is the analysis of classifier performance, so only the identification part of the classification experiment is loaded. Before the experiment, data mining needs to be performed on the voltage time series and the current time series of each electrical device obtained after the decomposition in step B, and the power feature vector is obtained through the following calculation.
In the formula of UkRepresenting the effective value of the voltage, IkRepresenting the effective value of the current, uk(t) represents the actual measured value of the current, ik(T) represents the actual measured value of the current, TkIs the total time of the measurement.
(1) In a typical ac circuit, the output power has a real component and a reactive component. The instantaneous active power is different and constantly changing. The active power and the reactive power of the electrical equipment are calculated by the following formulas:
in the formula, Sk＝UkIk，PkIs active power, QkIs reactive power, SkIs the total power.
(2) Power factor: which is the ratio of the active power to the apparent power of the ac circuit. From the active and reactive power, the power factor that can be calculated is:
in the formula, λkRepresenting the power factor.
(3) Resistance force: the calculation formula is as follows:
in the formula, ZkRepresenting the resistance.
(4) Impedance: in a circuit with a resistor, an inductor and a capacitor, the impedance of the alternating current is called impedance. The calculation formula of the impedance is as follows:
in the formula, RkWhich represents the impedance of the circuit board,
representing the phase difference.
(5) The calculation formula of the reactance is as follows:
in the formula, XkRepresenting the reactance.
(6) The formula for the admittance is as follows:
in the formula, YkRepresenting the admittance.
(7) The calculation formula of the conductance is as follows:
in the formula, YkRepresenting the conductance.
(8) The calculation formula of susceptance is as follows:
in the formula, BkRepresenting the susceptance.
(9) The calculation formulas for the real and imaginary parts of the currents 1-9 th harmonic are as follows:
in the formula, HikRepresenting the real and imaginary parts of the ith harmonic.
Through the method, the model of the invention calculates the physical characteristic information of each electrical device.
Step C2: permutation entropy feature extraction based on VMD decomposition
In the field of speech recognition and the like, Empirical Mode Decomposition (EMD) is widely used to decompose raw data and extract features. However, EMD lacks theoretical definition and recursive structure, and in addition, it has problems of modal aliasing, end-point effects, decomposition stopping criteria, and the like. And the VMD adopts a variation model, so that the VMD can adaptively decompose relevant frequency bands, and therefore, the VMD has better stick property for time series data decomposition. First, the VMD adaptively decomposes the voltage and current time series into K closely surrounding center frequency ωkAnd the nearby IMF components, and each decomposed IMF component is redefined into an amplitude modulation-frequency modulation signal. Then, the VMD method realizes the time domain variable by utilizing a Fourier equidistant transformation method
And λ1(t) efficient conversion into the frequency domain
And λ1(x) In that respect Wherein,
which is representative of the original signal, is,
representing the central frequency, λ, of the signal1(t) represents the Lagrangian operator,
and λ1(x) Respectively representing the original signal after fourier transformation, the signal center frequency and the lagrange operator. And finally, obtaining an IMF component with the minimum sum of the K bandwidths in the time domain by applying an inverse Fourier transform method.
The multi-scale arrangement entropy (MPE) refers to the arrangement entropy under different scales, the IMF components are subjected to coarse graining processing on multiple scales, and then the arrangement entropy value of the IMF components after coarse graining at each scale is calculated. And carrying out coarse graining treatment on the obtained IMF components, namely dividing the IMF components in sequence by using a window with the length of s, and then calculating the average value in each time window to obtain the IMF components after coarse graining. The permutation entropy characteristic data is obtained through time sequence reconstruction and normalization processing, the size of MPE values represents the random degree of the time sequence, the smaller the value of the MPE values is, the more regular the time sequence is, otherwise, the closer the time sequence is to the random. After the voltage time series and the current time series are subjected to variation modal decomposition, multi-scale permutation entropy information needs to be extracted from each IMF component and used as input features of a subsequent classifier.
Step D: feature selection, SRU classifier training and validation
Before inputting feature data into the SRU classifier, feature selection of the input feature vectors is necessary to reduce redundancy of the model, which can greatly reduce the computational cost of the classifier. Meanwhile, the accuracy of the classifier can be further effectively improved by extracting the effective features.
The invention selects the extracted power characteristics by using a reinforcement learning algorithm, and specifically comprises the following steps:
step D101: the state matrix S and the action matrix a are initialized. The state matrix S represents the state of each feature. The action matrix a is an action that adds or deletes a feature.
S＝[s1,s2,...,sm] (14)
a＝[Δs1,Δs2,Δs3,...,Δsm] (15)
Wherein s ismIs the state of the m-th feature, smIs 0 or 1(0 means that the feature is not used, and 1 means that the feature is used). Δ smAn act of adding or deleting an mth feature.
Behavior a is a strategy selected according to ε -greedy:
ε∈(0,1) (17)
wherein epsilon is the exploration probability.
Step D102: a reward R is established which will affect the behavior of the agent. In the invention, the accuracy of the SRU classification result is used as a return.
Step D103: the agent performs operation a according to the current state and environment.
Step D104: based on the evaluation function F, the agent calculates and updates the Q-table according to the behavior a and the state S. When using the Q-learning algorithm as the feature selection method, agent receives rewards from the environment, updating the state and Q tables by adjusting the behavior of the input feature changes. When using the Q-learning algorithm as an optimizer, agent receives rewards from the environment and updates the state and Q tables by adjusting the behavior of parameter changes. The calculation formula for the Q table update is as follows:
Qn+1(Sn,an)＝Qn(Sn,an)+βn(R(Sn,an)+γmaxQn(Sn+1,an+1)-Qn(Sn,an)) (18)
in the formula, anRepresenting the behavior of agent at the nth time node; snRepresenting the node state of agent at the nth time; rnIs the return the model gets at this time; gamma is a discount parameter; β represents the learning speed.
And training the classifier by using the training set based on the classifier which takes the selected power characteristics as input quantity and takes the power system fault diagnosis type as output quantity, testing the classifier by using the testing set, continuing training based on the updated characteristic selection method and the classifier when the fault type diagnosis accuracy of the test result does not meet the set requirement, and otherwise terminating the training and determining the characteristic selection method and the trained classifier, wherein the state matrix S is the final selection result of the model input characteristics.
The selected power characteristic vector M is equal to [ M [ ]1,m2,...,mn]And as the input of a training sample, the fault diagnosis type of the power system is used as the output, and a simple recurrent neural network model is trained, so that the classifier for real-time electric equipment state detection and fault diagnosis is obtained. In the invention, the main fault types of the electrical equipment comprise single-phase grounding short circuit, two-phase short circuit, three-phase short circuit, joint discharge, phase failure, insulation system fault and other fault types. The SRU classifier training and verifying method comprises the following steps:
step D201: initializing parameters of an SRU classifier
The number n, n of neurons in the input layer of the SRU classifier is determined after feature selection is carried out by a reinforced learning model. The input vector M is M ═ M1,m2,...,mn]Wherein m isiRepresenting the input features. The number of neurons in the hidden layer is 100, and the vector H output by the hidden layer is: h ═ H(1),H(2),H(3)...,H(100)]In which H is(i)Representing the basic parameters of the ith neuron. The number of neurons in the output layer is M determined according to the type of failure of each electrical device, and the output vector O is: o ═ O(1),O(2),O(3),...,O(M)]In which O is(i)Representing the type of fault diagnosis of the electrical equipment.
Step D202: training of SRU classifier
The main purpose of a simple recursive network is to solve the output O of the RNN at time TtDependent on the output O of the previous momentt-1This results in a problem that parallel computation cannot be performed. The SRU method parallelizes the computation process by simplifying the computation of the loop unit. Most of its calculations at each step are recursive, which makes parallelization easy.
In the training process of the SRU classifier, the input gate, the forgetting gate and the reset gate can realize parallel computation. Thus, the output of the SRU classifier and the SRU classifier internal parameters can be calculated very quickly and compactly. The simple recursive network can effectively realize parallel operation in the training process, and the training speed is greatly improved. In the present invention, a widely used Sigmoid function is selected as an activation function of each neuron.
Step D203: validation of SRU classifier
The maximum iteration number of the SRU classifier is set to 10000, and the training learning rate is set to 0.05.
And dividing the collected data after decomposition and feature selection into a training set and a test set according to the ratio of 3:1 after the data are processed according to the process. The training set is used to train the SRU classifier. The test set is used to calculate the error of the SRU classifier and evaluate the accuracy and robustness of the SRU classifier. And when the fault type diagnosis accuracy of the test result does not meet the set requirement, continuing training based on the updated feature selection method and the classifier, otherwise, terminating the training and determining the feature selection method and the trained classifier.
By integrating the processes, a final non-invasive fault diagnosis model of the unmanned train vehicle-mounted electrical system is obtained, and the final non-invasive fault diagnosis model can be used for executing the step E to diagnose the fault of the unmanned train vehicle-mounted electrical system in the process of waiting to run in real time and judging the specific running state.
Step E: real-time diagnosis
Collecting a plurality of groups of measured data of the electrical system to be diagnosed, wherein each group of measured data comprises total voltage data and total current data of the electrical system to be diagnosed, converting each group of measured data into a total voltage time sequence and a total current time sequence with the same rated working frequency, inputting the total voltage time sequence and the total current time sequence corresponding to the measured data into the deep learning model determined in the step B as input vectors, extracting the electric power characteristics in each output vector output by the measured data after passing through the deep learning model by using the characteristic extraction method in the step C, selecting the electric power characteristics corresponding to the measured data by using the state matrix S determined in the step D, inputting the electric power characteristics corresponding to the selected measured data into the SRU classifier determined in the step D, and outputting the fault type of the electrical system to be diagnosed by the SRU classifier, and finally obtaining the state and fault diagnosis result of each electrical device.
While the present invention has been described with reference to the embodiments shown in the drawings, the present invention is not limited to the embodiments, which are illustrative and not restrictive, and it will be apparent to those skilled in the art that various changes and modifications can be made therein without departing from the spirit and scope of the invention as defined in the appended claims.
Claims (6)
1. A non-invasive fault diagnosis method for a vehicle-mounted electrical system of an unmanned train is characterized by comprising the following steps:
step A, collecting multiple groups of modeling data of the unmanned train under M different known fault types, wherein each group of modeling data comprises total voltage data of an electrical system, total current data of the electrical system, voltage data of each electrical device in the electrical system and current data of each electrical device in the electrical system under each fault type; after converting each group of modeling data into a time sequence with the same rated working frequency, dividing a plurality of groups of converted modeling data into a training set and a test set according to a set proportion;
step B, training the deep learning model by using a training set based on a deep learning model taking the total voltage time sequence and the total current time sequence as input vectors and the voltage time sequence of each electrical device and the current time sequence of each electrical device as output vectors, testing the deep learning model by using a testing set, updating the deep learning model and continuing training when the accuracy of the testing result does not meet the set requirement, and otherwise terminating the training and obtaining the trained deep learning model;
step C, extracting the electric power characteristics in each output vector output by the training set after passing through the deep learning model by using a characteristic extraction method;
step D, selecting the extracted power features by using a feature selection method, training the classifier by using a training set based on the classifier which takes the selected power features as input quantity and takes the power system fault diagnosis type as output quantity, testing the classifier by using a test set, continuing training based on the updated feature selection method and the classifier when the fault type diagnosis accuracy of the test result does not meet the set requirement, and otherwise terminating the training and determining the feature selection method and the trained classifier; the feature selection method is a reinforcement learning method; the classifier is an SRU classifier;
step E, collecting a plurality of groups of measured data of the electrical system to be diagnosed, wherein each group of measured data comprises total voltage data and total current data of the electrical system to be diagnosed, converting each group of measured data into a total voltage time sequence and a total current time sequence with the same rated working frequency, inputting the total voltage time sequence and the total current time sequence corresponding to the measured data as input vectors into the deep learning model determined in the step B, extracting the electric power characteristics in each output vector output after the measured data passes through the deep learning model by using the characteristic extraction method in the step C, and selecting the electric power characteristics corresponding to the measured data by using the characteristic selection method determined in the step D, and inputting the power characteristics corresponding to the selected obtained measured data into the classifier determined in the step D, and outputting the fault type of the electrical system to be diagnosed by the classifier.
2. The method according to claim 1, wherein the deep learning model in step B comprises a CNN frame and a QRNN frame.
3. The method for non-invasive diagnosis of a fault in an on-board electrical system of an unmanned train as set forth in claim 1, wherein the feature extraction method in the step C includes a physical feature extraction method and a time series feature extraction method.
4. The method for non-invasive diagnosis of a fault in an on-board electrical system of an unmanned train according to claim 3, wherein in the step C, the time series feature extraction method includes:
decomposing the voltage time sequence and the current time sequence output by the deep learning model to obtain an IMF component with the minimum sum of K bandwidths;
and calculating the multi-scale permutation entropy of each IMF component, and taking the obtained permutation entropy characteristic data as time sequence characteristics.
5. The method for non-invasive diagnosis of faults in an on-board electrical system of an unmanned train as claimed in claim 2, wherein the activation function in the CNN framework is a ReLU function and the optimization algorithm in the CNN framework is an Adam algorithm.
6. The method for non-invasive diagnosis of a fault in an on-board electrical system of an unmanned train as claimed in claim 4, wherein the time series of voltage and the time series of current output from the deep learning model are decomposed by a variational modal decomposition method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011616269.4A CN112798888B (en) | 2020-12-30 | 2020-12-30 | Non-invasive fault diagnosis method for vehicle-mounted electrical system of unmanned train |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011616269.4A CN112798888B (en) | 2020-12-30 | 2020-12-30 | Non-invasive fault diagnosis method for vehicle-mounted electrical system of unmanned train |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112798888A CN112798888A (en) | 2021-05-14 |
| CN112798888B true CN112798888B (en) | 2021-12-17 |
Family
ID=75804832
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011616269.4A Active CN112798888B (en) | 2020-12-30 | 2020-12-30 | Non-invasive fault diagnosis method for vehicle-mounted electrical system of unmanned train |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112798888B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN114266383A (en) * | 2021-11-26 | 2022-04-01 | 华能碳资产经营有限公司 | Power equipment state identification method based on multi-scale permutation entropy algorithm |
| CN114815629B (en) * | 2022-07-01 | 2022-09-16 | 天津市职业大学 | Method for evaluating working state of intelligent networking automobile inductor |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102662142A (en) * | 2012-03-15 | 2012-09-12 | 南京航空航天大学 | Prediction method for multi-parameter identification fault of power electronic circuit based on RVM-QNN |
| CN110109015A (en) * | 2019-05-31 | 2019-08-09 | 中南大学 | A kind of asynchronous motor Fault monitoring and diagnosis method based on deep learning |
| CN110347547A (en) * | 2019-05-27 | 2019-10-18 | 中国平安人寿保险股份有限公司 | Log method for detecting abnormality, device, terminal and medium based on deep learning |
| DE102019203631A1 (en) * | 2019-03-18 | 2020-09-24 | Robert Bosch Gmbh | Method and device for modeling a power semiconductor module |
| CN111967512A (en) * | 2020-08-07 | 2020-11-20 | 国网江苏省电力有限公司电力科学研究院 | Abnormal electricity utilization detection method, system and storage medium |
| CN112067916A (en) * | 2019-09-20 | 2020-12-11 | 武汉理工大学 | Time series data intelligent fault diagnosis method based on deep learning |
| CN112101116A (en) * | 2020-08-17 | 2020-12-18 | 北京无线电计量测试研究所 | Analog circuit fault diagnosis method based on deep learning |
-
2020
- 2020-12-30 CN CN202011616269.4A patent/CN112798888B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102662142A (en) * | 2012-03-15 | 2012-09-12 | 南京航空航天大学 | Prediction method for multi-parameter identification fault of power electronic circuit based on RVM-QNN |
| DE102019203631A1 (en) * | 2019-03-18 | 2020-09-24 | Robert Bosch Gmbh | Method and device for modeling a power semiconductor module |
| CN110347547A (en) * | 2019-05-27 | 2019-10-18 | 中国平安人寿保险股份有限公司 | Log method for detecting abnormality, device, terminal and medium based on deep learning |
| CN110109015A (en) * | 2019-05-31 | 2019-08-09 | 中南大学 | A kind of asynchronous motor Fault monitoring and diagnosis method based on deep learning |
| CN112067916A (en) * | 2019-09-20 | 2020-12-11 | 武汉理工大学 | Time series data intelligent fault diagnosis method based on deep learning |
| CN111967512A (en) * | 2020-08-07 | 2020-11-20 | 国网江苏省电力有限公司电力科学研究院 | Abnormal electricity utilization detection method, system and storage medium |
| CN112101116A (en) * | 2020-08-17 | 2020-12-18 | 北京无线电计量测试研究所 | Analog circuit fault diagnosis method based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| 基于深度学习的时间序列数据异常检测方法;胡姣姣 等;《信息与控制》;20190131;第48卷(第1期);1-8 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112798888A (en) | 2021-05-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112798888B (en) | Non-invasive fault diagnosis method for vehicle-mounted electrical system of unmanned train | |
| CN110161343B (en) | Non-invasive real-time dynamic monitoring method for external powered device of intelligent train | |
| Lu et al. | Fault severity recognition of aviation piston pump based on feature extraction of EEMD paving and optimized support vector regression model | |
| CN109164343B (en) | Transformer fault diagnosis method based on characteristic information quantization and weighted KNN | |
| CN107219457B (en) | Frame-type circuit breaker fault diagnosis and degree assessment method based on operation attachment electric current | |
| CN109947086A (en) | Mechanical breakdown migration diagnostic method and system based on confrontation study | |
| CN110705456A (en) | Micro motor abnormity detection method based on transfer learning | |
| CN110070229A (en) | The short term prediction method of home electrical load | |
| CN111239549A (en) | Power distribution fault rapid positioning method based on discrete wavelet transform | |
| CN107101813A (en) | A kind of frame-type circuit breaker mechanical breakdown degree assessment method based on vibration signal | |
| CN108535648A (en) | Method of Motor Fault Diagnosis and system | |
| CN113295413B (en) | Traction motor bearing fault diagnosis method based on indirect signals | |
| CN115049058B (en) | Compression method and device of topology recognition model, electronic equipment and medium | |
| CN112307963A (en) | Converter transformer running state identification method based on vibration signals | |
| CN116552306B (en) | Monitoring system and method for direct current pile | |
| CN114462508A (en) | Power transformer health state assessment method based on multi-mode neural network | |
| CN113627375A (en) | Planetary gear fault diagnosis method and system, storage medium and computing device | |
| CN113533952A (en) | CEEMD and phase space reconstruction-based mechanical fault diagnosis method for tap changer of on-load tap-changing transformer | |
| CN113705396A (en) | Motor fault diagnosis method, system and equipment | |
| Liang et al. | Generalized composite multiscale diversity entropy and its application for fault diagnosis of rolling bearing in automotive production line | |
| CN115166514A (en) | Motor fault identification method and system based on self-adaptive spectrum segmentation and denoising | |
| CN117219124B (en) | Switch cabinet voiceprint fault detection method based on deep neural network | |
| CN113740671A (en) | Fault arc identification method based on VMD and ELM | |
| CN115564075B (en) | Main and auxiliary integrated fault collaborative diagnosis method and system for urban power grid | |
| CN116340812A (en) | Transformer partial discharge fault mode identification method and system |
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 |