CN109615003B - Power failure prediction method based on ELM-CHMM - Google Patents

Power failure prediction method based on ELM-CHMM Download PDF

Info

Publication number
CN109615003B
CN109615003B CN201811488243.9A CN201811488243A CN109615003B CN 109615003 B CN109615003 B CN 109615003B CN 201811488243 A CN201811488243 A CN 201811488243A CN 109615003 B CN109615003 B CN 109615003B
Authority
CN
China
Prior art keywords
model
degradation
elm
voltage signal
state
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
CN201811488243.9A
Other languages
Chinese (zh)
Other versions
CN109615003A (en
Inventor
杨京礼
刘晓东
张天瀛
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.)
Harbin Institute of Technology
Original Assignee
Harbin Institute of Technology
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 Harbin Institute of Technology filed Critical Harbin Institute of Technology
Priority to CN201811488243.9A priority Critical patent/CN109615003B/en
Publication of CN109615003A publication Critical patent/CN109615003A/en
Application granted granted Critical
Publication of CN109615003B publication Critical patent/CN109615003B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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
    • G01R31/40—Testing power supplies
    • 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
    • G06N3/048—Activation functions
    • 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

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)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Supply And Distribution Of Alternating Current (AREA)

Abstract

The invention discloses a power failure prediction method based on ELM-CHMM, and relates to a power failure prediction method. The invention aims to solve the problem of low accuracy of fault prediction in the existing method. The process is as follows: dividing the voltage signal data into a training data set and a testing data set, and preprocessing the training data set to obtain a reconstructed voltage signal matrix; establishing an ELM model; inputting the test data set into an ELM model, and outputting a voltage signal predicted by the ELM model; extracting characteristic parameters of a training data set; establishing a CHMM state prediction model; extracting characteristic parameters of voltage signal data predicted by an ELM model, and respectively inputting the characteristic parameters into a CHMM model; and obtaining an ELM-CHMM model and obtaining the state of the power supply fault to be detected. The method is used for the field of power failure prediction.

Description

Power failure prediction method based on ELM-CHMM
Technical Field
The invention relates to a power failure prediction method based on ELM-CHMM.
Background
The marine power supply system is responsible for the electric energy supply of the electric equipment of the ship, ensures the normal and stable work of various complicated electronic equipment on the ship and the living requirements of workers on the ship, and is the 'blood' of the ship. Therefore, the stability and the reliability of the marine power supply system are very important, and the fault of the power supply needs to be accurately predicted in real time in order to provide stable power supply guarantee for the electrical equipment on the ship.
Because the ship works on the sea for a long time, the ship is influenced by severe environmental factors such as dampness, corrosion, vibration, impact and the like, and is also subjected to radiation interference of other equipment, the possibility of power failure is greatly improved, and the performance and the service life of the power supply are seriously influenced. Once the power supply for the ship fails, partial electric equipment on the ship can be damaged, even the ship cannot work normally, and personnel safety on the ship is threatened. Therefore, a proper method is adopted to timely and effectively find out potential faults, so that workers can timely take preventive maintenance conveniently, and loss is reduced.
In the traditional various fitting and predicting problems, a widely applied method is that a single hidden layer feedforward neural network utilizes sample data to train and fit a mapping function, and the method is suitable for the problem of nonlinear data modeling with large data volume. Meanwhile, the method has many defects, such as large calculation amount, long training time, easy falling into local optimum of the traditional feedforward neural network, difficulty in obtaining a global optimum solution, sensitivity to parameter selection and the like, and is not suitable for a fault prediction scene with high accuracy requirement.
In recent years, an Extreme Learning Machine (ELM), which is a simple and effective single hidden layer feedforward neural network SLFNs learning algorithm, has been proposed and widely used. The analogy to the biological brain is that the connection of the hidden layer to the output layer is adjusted according to different applications to realize compression, feature learning, sparse coding, clustering, regression fitting and classification. On the basis of exerting the advantages of the traditional single hidden layer feedforward neural network, the ELM algorithm overcomes the bottleneck of the traditional neural network, avoids repeated iterative solution of a gradient descent method, effectively improves the calculation speed, avoids local minimum problems to obtain a global optimal solution, realizes as little human intervention as possible, and obtains higher accuracy.
Disclosure of Invention
The invention aims to solve the problem of low failure prediction accuracy of the existing method, and provides a power failure prediction method based on ELM-CHMM.
A power failure prediction method based on ELM-CHMM comprises the following specific processes:
the method comprises the steps of firstly, collecting voltage signal data of four states of light degradation, moderate degradation, severe degradation and complete failure of a marine power supply, dividing the voltage signal data of each state into a training data set and a testing data set, preprocessing the voltage signal data of the training data set, and obtaining a reconstructed voltage signal matrix
Figure BDA0001895070150000021
In the formula, X, Y is a reconstructed output voltage signal matrix of the marine power supply,
Figure BDA0001895070150000022
is a one-dimensional matrix [ x ] 1 ,x 2 ...x m ]，
Figure BDA0001895070150000023
Is a one-dimensional matrix [ x ] 2 ,x 3 ...x m+1 ]，
Figure BDA0001895070150000024
Is a one-dimensional matrix [ x ] N-m ,x N-m+1 ...x N-1 ]，x m+1 Supply voltage of 1 time point after the mth time point, x m+2 Supply voltage of 2 time points backward from the m time point, x N The power supply voltage is output for the Nth time point;
step two, taking the voltage signal matrix X reconstructed in the step one as the input of the ELM model, and taking the voltage signal matrix Y reconstructed in the step one as the output of the ELM model;
training an ELM model by using an input-output voltage signal matrix to obtain ELM model parameters, and establishing the ELM model;
inputting the voltage signal corresponding to the test data set in the step one into the built ELM model, and outputting the voltage signal predicted by the ELM model;
the ELM model is an extreme learning machine model;
thirdly, decomposing and reconstructing voltage signal data corresponding to the training data set in each state collected in the first step by utilizing wavelet packet analysis according to the frequency domain characteristics of the power output voltage signal, and extracting characteristic parameters of voltages in four states of slight degradation, moderate degradation, severe degradation and complete failure;
step four, training a CHMM state prediction model according to the characteristic parameters of the four state voltages obtained in the step three to obtain model parameters of the power supply in four states of slight degradation, moderate degradation, severe degradation and complete failure, and completing establishment of the CHMM state prediction model of the four power supply states;
the CHMM is a continuous hidden Markov model;
step five, decomposing and reconstructing the voltage signal data predicted by the ELM model in the step two by utilizing wavelet packet analysis, extracting characteristic parameters of voltages in four states of slight degradation, moderate degradation, severe degradation and complete failure, and respectively inputting the characteristic parameters into CHMM state prediction models in four power states;
if the probability that the prediction result is consistent with the power state of the test data set is greater than 85%, power failure prediction is accurate, an ELM-CHMM power failure prediction model is obtained, a power failure voltage signal to be tested is input into the ELM-CHMM power failure prediction model, the states of the power failure to be tested are obtained, the states include four states of slight degradation, moderate degradation, severe degradation and complete failure, and the evaluation of the power failure degree is achieved;
and if the probability that the prediction result is consistent with the power state of the test data set is less than or equal to 85%, the power failure prediction is inaccurate.
The invention has the beneficial effects that:
the invention provides a power failure prediction method based on ELM-CHMM. For power supply data, data trend prediction is carried out by using ELM; and (3) carrying out state evaluation on the predicted data by utilizing the established Continuous Hidden Markov Model (CHMM) to judge the state of the predicted data. The method combines the advantages of high ELM operation speed, strong generalization capability, capability of obtaining global optimal solution and the like, and the advantages of better classification capability of CHMM, capability of establishing the corresponding relation between observation data and hidden states, capability of carrying out accurate log-likelihood probability calculation and the like, improves the power failure prediction precision and speed, shortens the prediction time, and solves the problems of local optimization and sensitivity to parameter selection. The CHMM model is divided into a mild fault state, a moderate fault state, a severe fault state and a complete fault state according to the fault degree. Therefore, the prediction of the power failure state can be completed, and the approximate time of the power failure with different degrees can be given by conversion according to the sampling interval.
In the power failure prediction algorithm provided by the invention, because the type of the hidden layer activation function has certain influence on the prediction result of the ELM model, the effect is relatively better by comparing the Sigmoid function, and the speed and the precision can be faster, so that the Sigmoid function is selected as the activation function of the ELM model. Training the ELM and BP neural networks by using training data, testing the trained model by using test data, calculating Root Mean Square Error (RMSE) and prediction time, and comparing to obtain that the prediction precision of the power failure is improved by about 10.8%, the calculation speed is improved by about 78.8%, and meanwhile, comparing the original data and the prediction data, the corresponding degradation states are consistent, namely the prediction result is consistent with the real result. Therefore, the fault prediction algorithm provided by the invention can obviously improve the real-time performance and accuracy of power supply fault prediction.
Drawings
FIG. 1 is a block diagram of ELM-CHMM prediction according to the present invention.
Detailed Description
The first specific implementation way is as follows: the present embodiment is described with reference to fig. 1, and a specific process of the power failure prediction method based on ELM-CHMM in the present embodiment is as follows:
the method comprises the steps of firstly, collecting voltage signal data of four states of light degradation, moderate degradation, severe degradation and complete failure of a marine power supply, dividing the voltage signal data of each state into a training data set and a testing data set, preprocessing the voltage signal data of the training data set, and obtaining a reconstructed voltage signal matrix
Figure BDA0001895070150000031
In the formula, X, Y is a reconstructed output voltage signal matrix of the marine power supply,
Figure BDA0001895070150000041
is a one-dimensional matrix [ x ] 1 ,x 2 ...x m ]，
Figure BDA0001895070150000042
Is a one-dimensional matrix [ x ] 2 ,x 3 ...x m+1 ]，
Figure BDA0001895070150000043
Is a one-dimensional matrix [ x ] N-m ,x N-m+1 ...x N-1 ]，x m+1 Supply voltage of 1 time point after the mth time point, x m+2 Supply voltage of 2 time points backward from the m time point, x N The power supply voltage is output for the Nth time point;
the power supply is slightly degraded to the state that the power supply loss of the power supply is less than or equal to 20%;
when the power supply is moderately degraded, the power supply loss of the power supply is more than 20 percent and less than or equal to 30 percent;
when the power supply is severely degraded to the state that the power supply loss of the power supply is more than 30 percent and less than or equal to 50 percent;
when the power supply completely fails, the power supply loss of the power supply is more than 50%;
step two, taking the voltage signal matrix X reconstructed in the step one as the input of the ELM model, and taking the voltage signal matrix Y reconstructed in the step one as the output of the ELM model;
training an ELM model by using an input-output voltage signal matrix to obtain ELM model parameters, and establishing the ELM model;
inputting the voltage signal corresponding to the test data set in the step one into the built ELM model, and outputting the voltage signal predicted by the ELM model;
the ELM model is an extreme learning machine model;
thirdly, decomposing and reconstructing voltage signal data corresponding to the training data set in each state collected in the first step by utilizing wavelet packet analysis according to the frequency domain characteristics of the power output voltage signal, and extracting characteristic parameters of voltages in four states of slight degradation, moderate degradation, severe degradation and complete failure;
step four, training a CHMM state prediction model according to the characteristic parameters of the four state voltages obtained in the step three to obtain model parameters of the power supply in four states of slight degradation, moderate degradation, severe degradation and complete failure, and completing establishment of the CHMM state prediction model of the four power supply states;
the CHMM is a continuous hidden Markov model;
step five, decomposing and reconstructing the voltage signal data predicted by the ELM model in the step two by utilizing wavelet packet analysis, extracting characteristic parameters of voltages in four states of slight degradation, moderate degradation, severe degradation and complete failure, and respectively inputting the characteristic parameters into CHMM state prediction models in four power states;
if the probability that the prediction result is consistent with the power state of the test data set is greater than 85%, power failure prediction is accurate, an ELM-CHMM power failure prediction model is obtained, a power failure voltage signal to be tested is input into the ELM-CHMM power failure prediction model, the states of the power failure to be tested are obtained, the states include four states of slight degradation, moderate degradation, severe degradation and complete failure, and the evaluation of the power failure degree is achieved;
and if the probability that the prediction result is consistent with the power state of the test data set is less than or equal to 85%, the power failure prediction is inaccurate.
The second embodiment is as follows: the first embodiment is different from the first embodiment in that output voltage signal data in four states of light degradation, moderate degradation, severe degradation and complete failure of the marine power supply are collected in the first step, and the output voltage signal data are preprocessed to obtain a reconstructed output voltage signal matrix X, Y; the specific process is as follows:
when the power supply is degraded or fails, the output voltage signal changes accordingly. The method realizes the prediction of the progressive faults of the marine power supply by monitoring the output voltage of the marine power supply. The output voltage trends of five power supplies are predicted by taking the output voltages of the power supplies in the mild fault state, the moderate fault state, the severe fault state and the complete fault state as original data and training and modeling an Extreme Learning Machine (ELM) model.
In order to better support ELM modeling prediction, phase space reconstruction is carried out on an original data sequence according to a prediction model structure of the ELM to obtain high-dimensional input data, so that an incidence relation among the data is obtained and the information quantity as large as possible is mined; the specific process is as follows:
the output voltage signal sequence of four states of light degradation, moderate degradation, severe degradation and complete failure of the marine power supply is X N ＝{x 1 ,x 2 ,…,x N }；
Given a sequence of values x for a one-dimensional time point n Assuming that a non-linear mapping exists between the first m sequence values:
x n ＝f(x n-m ,...,x n-2 ,x n-1 ),n∈(1,...,N),m∈(0,...,N-1) (1)
in the formula, x n The power supply voltage (any voltage value in the interval), x, outputted for the nth time point n-m Supply voltage, x, m time points forward of the nth time point n-2 Supply voltage, x, 2 points forward of the nth point in time n-1 The power supply voltage is one power supply voltage ahead at the nth time point, N is a serial value label of the time point, and N = N-m is satisfied; n is the number of original data of the output voltage, m is the embedding dimension, and f is a nonlinear mapping model;
a sequence value x of one-dimensional time points n The reconstruction is in the form of a matrix,
Figure BDA0001895070150000051
in the formula, X, Y is a reconstructed output voltage signal matrix of the marine power supply,
Figure BDA0001895070150000052
is a one-dimensional matrix [ x ] 1 ,x 2 ...x m ]，
Figure BDA0001895070150000053
Is a one-dimensional matrix [ x ] 2 ,x 3 ...x m+1 ]，
Figure BDA0001895070150000054
Is a one-dimensional matrix [ x ] N-m ,x N-m+1 ...x N-1 ]，x m+1 Supply voltage of 1 time point after the mth time point, x m+2 Supply voltage of 2 time points backward from the m time point, x N The power supply voltage outputted for the Nth time point.
Other steps and parameters are the same as those in the first embodiment.
The third concrete implementation mode: the second step is different from the first or second specific embodiment in that the voltage signal matrix X reconstructed in the first step is used as the input of the ELM model, and the voltage signal matrix Y reconstructed in the first step is used as the output of the ELM model;
training an ELM model by using an input-output voltage signal matrix to obtain ELM model parameters, and establishing the ELM model;
inputting the voltage signal corresponding to the test data set in the step one into the built ELM model, and outputting the voltage signal predicted by the ELM model;
the specific process is as follows:
for M sets of training samples (X) j ,Y j ) If the actual output of the ELM model network is equal to the desired output, then the mathematical model expression form g (x) of the activation function is as follows:
Figure BDA0001895070150000061
wherein, X j And Y j For any set of input and output matrices of the ELM model,
Figure BDA0001895070150000062
the number of nodes of the hidden layer of the ELM model network, M is the number of training sample groups, beta i To connect the weights of the ith hidden layer node, b i Implicit layer node bias, ω, for ELM model networks i The connection weight of the input layer and the hidden layer of the ELM model network is obtained;
equation (3) can be abbreviated as: h β = Y;
Figure BDA0001895070150000063
wherein H is an output matrix of an ELM model network hidden layer;
randomly generating hidden node parameters (omega) i ,b i ) Calculating the hidden layer output matrix H, and then calculating the weight beta, beta = H -1 Y; taking beta as an ELM model parameter, and establishing an ELM model;
and (4) inputting the voltage signal corresponding to the test data set in the step one into the built ELM model, and outputting the voltage signal predicted by the ELM model.
Other steps and parameters are the same as those in the first or second embodiment.
The fourth concrete implementation mode: this embodiment is different from the first to third embodiments in that
As the power supply degrades or fails, a corresponding change in the output voltage signal is reflected in the energy distribution. The invention realizes the characteristic extraction of the power output voltage by utilizing wavelet packet analysis. Firstly, decomposing and reconstructing voltage signal data corresponding to a training data set in each state collected in the step one by utilizing wavelet packet analysis according to the frequency domain characteristics of a power supply output voltage signal, and extracting characteristic parameters of voltages in four states of slight degradation, moderate degradation, severe degradation and complete failure;
the specific process is as follows:
setting voltage signal data corresponding to the training data set in each state acquired in the step one as S, and reconstructing at the qth node of the p layer after wavelet packet decomposition to obtain a discrete point signal S pq (k) K =1,2, L is the number of points after signal reconstruction;
the signal energy E in each equidistant frequency band after wavelet packet decomposition and reconstruction pq And total energy E is calculated as:
Figure BDA0001895070150000071
Figure BDA0001895070150000072
obtaining a frequency domain characteristic vector T extracted by the wavelet packet,
Figure BDA0001895070150000073
and the frequency domain characteristic vector T extracted from the wavelet packet is the ratio of the energy of each decomposed frequency band to the total energy, and voltage characteristic parameters of a training data set of a light degeneration power supply, a medium degeneration power supply, a heavy degeneration power supply and a complete fault power supply are obtained.
Other steps and parameters are the same as those in one of the first to third embodiments.
The fifth concrete implementation mode: the difference between the present embodiment and one of the first to fourth embodiments is that, in the fourth step, according to the characteristic parameters of the four state voltages obtained in the third step, the CHMM state evaluation model is trained to obtain model parameters of the power supply in four states of light degradation, moderate degradation, severe degradation and complete failure, so as to complete the establishment of the CHMM state evaluation model in the four power supply states; the specific process is as follows:
hidden Markov Models (HMMs) are a special markov process that describes the statistical properties of a stochastic process by means of probabilistic models. It is a double stochastic process with hidden markov chains of finite state numbers to describe the state transitions and a set of display stochastic functions to describe the correspondence between observations and each state.
The hidden markov model has the following five elements:
u is the total number of states in the hidden Markov model, namely four states of slight degradation, moderate degradation, severe degradation and complete failure;
v, V is the total number of observed values observed in each state in the hidden Markov model, and the total number of the observed values is the number of characteristic parameters of the voltages of the four states obtained in the third step;
π＝{π ξ the initial state probability distribution represents the probability of being in the xi state at the beginning, and xi is slight degradation, moderate degradation, severe degradation or complete failure;
A＝{a ξζ a is a transition probability distribution matrix of the state, and represents the probability that the state is transitioned from xi to zeta, and zeta is light degradation, moderate degradation, heavy degradation or complete failure;
B＝{b ζ (r) }, B is an observation probability distribution matrix, which represents the probability that an observation vector is the r-th one when in state ζ (a number of observation vectors input below are said to be the r-th one, and one observation vector represents an observation vector in one state);
generally, a hidden markov model is represented by λ = (U, V, pi, a, B), which can be simplified to λ = (pi, a, B);
taking the characteristic parameters of the four state voltages obtained in the step three as an observation vector O Original ；
Will observe vector O Original As an input of the hidden markov model HMM, λ = (pi, a, B) is used as a parameter of the hidden markov model HMM, and the hidden markov model HMM is trained to obtain model parameters λ of the light degradation fault HMM1, the moderate degradation fault HMM2, the heavy degradation fault HMM3, and the complete fault HMM4, respectively 1 、λ 2 、λ 3 、λ 4 And finishing the establishment of the CHMM state evaluation model of the four power states.
Other steps and parameters are the same as in one of the first to fourth embodiments.
The sixth specific implementation mode: the difference between this embodiment mode and one of the first to fifth embodiment modes is that in the fifth step
Decomposing and reconstructing the voltage signal data predicted by the ELM model in the second step by utilizing wavelet packet analysis, extracting characteristic parameters of voltages in four states of slight degradation, moderate degradation, severe degradation and complete failure, and respectively inputting the characteristic parameters into CHMM state prediction models in four power states;
if the probability that the prediction result is consistent with the power state of the test data set is larger than 85%, power failure prediction is accurate, an ELM-CHMM power failure prediction model is obtained, a power failure voltage signal to be tested is input into the ELM-CHMM power failure prediction model, the state of the power failure to be tested is obtained, the state comprises four states of slight degradation, moderate degradation, severe degradation and complete failure, and the evaluation on the power failure degree is realized;
and if the probability that the prediction result is consistent with the power state of the test data set is less than or equal to 85%, the power failure prediction is inaccurate.
The specific process is as follows:
decomposing and reconstructing the voltage signal data predicted by the ELM model in the second step by utilizing wavelet packet analysis, extracting characteristic parameters of voltages in four states of light degradation, moderate degradation, severe degradation and complete failure, taking the characteristic parameters of each state as a group, and respectively passing each group through a system with four model parameters of lambda 1 、λ 2 、λ 3 、λ 4 The HMM1, HMM2, HMM3 and HMM4 models utilize the forward-backward algorithm of the CHMM to make each group obtain four log-likelihood probabilities, which are respectively P (O) Prediction |λ 1 )、P(O Prediction |λ 2 )、P(O Prediction |λ 3 )、P(O Prediction |λ 4 )；
The state type corresponding to each group of the maximum log-likelihood probability values is a power failure prediction result, if the probability that the prediction result is consistent with the power state of the test data set is greater than 85%, the power failure prediction is accurate, a power failure prediction model of the ELM-CHMM is obtained, a power failure voltage signal to be tested is input into the power failure prediction model of the ELM-CHMM, the state of the power failure to be tested is obtained, the states comprise four states of slight degradation, moderate degradation, severe degradation and complete failure, and the evaluation of the power failure degree is realized;
and if the probability that the prediction result is consistent with the power state of the test data set is less than or equal to 85%, the power failure prediction is inaccurate.
Other steps and parameters are the same as those in one of the first to fifth embodiments.
The following examples were used to demonstrate the beneficial effects of the present invention:
the first embodiment is as follows:
the preparation method comprises the following steps:
ELM model training
Figure BDA0001895070150000101
Figure BDA0001895070150000111
CHMM model training
Figure BDA0001895070150000121
Figure BDA0001895070150000131
Figure BDA0001895070150000141
The present invention is capable of other embodiments and its several details are capable of modifications in various obvious respects, all without departing from the spirit and scope of the present invention.

Claims (6)

1. A power failure prediction method based on ELM-CHMM is characterized in that: the method comprises the following specific processes:
the method comprises the steps of firstly, collecting voltage signal data of four states of light degradation, moderate degradation, severe degradation and complete failure of a marine power supply, dividing the voltage signal data of each state into a training data set and a testing data set, preprocessing the voltage signal data of the training data set, and obtaining a reconstructed voltage signal matrix
Figure FDA0003828667740000011
In the formula, X, Y is a reconstructed output voltage signal matrix of the marine power supply,
Figure FDA0003828667740000012
is a one-dimensional matrix
Figure FDA0003828667740000013
Is a one-dimensional matrix [ x ] 2 ,x 3 …x m+1 ]，
Figure FDA0003828667740000014
Is a one-dimensional matrix [ x ] N-m ,x N-m+1 …x N-1 ]，x m+1 Supply voltage of 1 time point backward from the mth time point, x m+2 Supply voltage of 2 time points backward from the m time point, x N The power supply voltage is output for the Nth time point;
step two, taking the voltage signal matrix X reconstructed in the step one as the input of the ELM model, and taking the voltage signal matrix Y reconstructed in the step one as the output of the ELM model;
training an ELM model by using an input and output voltage signal matrix to obtain ELM model parameters, and establishing the ELM model;
the ELM model is an extreme learning machine model;
inputting the voltage signal corresponding to the test data set in the step one into the built ELM model, and outputting the voltage signal predicted by the ELM model;
thirdly, decomposing and reconstructing voltage signal data corresponding to the training data set in each state collected in the first step by utilizing wavelet packet analysis according to the frequency domain characteristics of the power output voltage signal, and extracting characteristic parameters of voltages in four states of slight degradation, moderate degradation, severe degradation and complete failure;
step four, training a CHMM state prediction model according to the characteristic parameters of the four state voltages obtained in the step three to obtain model parameters of the power supply in four states of slight degradation, moderate degradation, severe degradation and complete failure, and completing establishment of the CHMM state prediction model of the four power supply states;
the CHMM is a continuous hidden Markov model;
step five, decomposing and reconstructing the voltage signal data predicted by the ELM model in the step two by utilizing wavelet packet analysis, extracting characteristic parameters of voltages in four states of slight degradation, moderate degradation, severe degradation and complete failure, and respectively inputting the characteristic parameters into CHMM state prediction models in four power states;
if the probability that the prediction result is consistent with the power state of the test data set is greater than 85%, power failure prediction is accurate, an ELM-CHMM power failure prediction model is obtained, a power failure voltage signal to be tested is input into the ELM-CHMM power failure prediction model, the states of the power failure to be tested are obtained, the states include four states of slight degradation, moderate degradation, severe degradation and complete failure, and the evaluation of the power failure degree is achieved;
and if the probability that the prediction result is consistent with the power state of the test data set is less than or equal to 85%, the power failure prediction is inaccurate.
2. The ELM-CHMM-based power failure prediction method of claim 1, wherein: acquiring output voltage signal data of four states of light degradation, moderate degradation, severe degradation and complete failure of the marine power supply in the first step, and preprocessing the output voltage signal data to obtain a reconstructed output voltage signal matrix X, Y; the specific process is as follows:
the output voltage signal sequence of four states of light degradation, moderate degradation, severe degradation and complete failure of the marine power supply is X N ＝{x 1 ,x 2 ,…,x N }；
Given a one-dimensional sequence of time points x n Assuming a non-linear mapping relationship with the first m sequence values:
x n ＝f(x n-m ,...,x n-2 ,x n-1 ),n∈(1,...,N),m∈(0,...,N-1) (1)
in the formula, x n Supply voltage, x, output for the nth point in time n-m Supply voltage, x, m time points forward of the nth time point n-2 Supply voltage, x, 2 points forward of the nth point in time n-1 The power supply voltage is one power supply voltage ahead at the nth time point, N is a serial value label of the time point, and N = N-m is satisfied; n is the number of original data of the output voltage, m is the embedding dimension, and f (·) is a nonlinear mapping model;
a sequence value x of one-dimensional time point n The reconstruction is in the form of a matrix,
Figure FDA0003828667740000021
in the formula, X, Y is a reconstructed output voltage signal matrix of the marine power supply,
Figure FDA0003828667740000022
is a one-dimensional matrix [ x ] 1 ,x 2 ...x m ]，
Figure FDA0003828667740000023
Is aDimension matrix x 2 ,x 3 ...x m+1 ]，
Figure FDA0003828667740000024
Is a one-dimensional matrix [ x ] N-m ,x N-m+1 ...x N-1 ]，x m+1 Supply voltage of 1 time point after the mth time point, x m+2 Supply voltage of 2 time points backward from the m time point, x N The power supply voltage outputted for the Nth time point.
3. The ELM-CHMM-based power failure prediction method of claim 1 or 2, wherein: in the second step, the voltage signal matrix X reconstructed in the first step is used as the input of the ELM model, and the voltage signal matrix Y reconstructed in the first step is used as the output of the ELM model;
training an ELM model by using an input-output voltage signal matrix to obtain ELM model parameters, and establishing the ELM model;
inputting the voltage signal corresponding to the test data set in the step one into the built ELM model, and outputting the voltage signal predicted by the ELM model;
the specific process is as follows:
for M sets of training samples (X) j ,Y j ) If the actual output of the ELM model network is equal to the desired output, then the mathematical model of the activation function is expressed in the form g (x) as follows:
Figure FDA0003828667740000031
wherein, X j And Y j For any set of input and output matrices of the ELM model,
Figure FDA0003828667740000032
the number of hidden layer nodes of the ELM model network, M is the number of training sample groups, beta i To connect the weights of the ith hidden layer node, b i Implicit layer node bias, ω, for ELM model networks i For network input of ELM modelThe connection weight of the layer and the hidden layer;
equation (3) is abbreviated as: h β = Y;
Figure FDA0003828667740000033
wherein H is an output matrix of an ELM model network hidden layer;
randomly generating hidden node parameters (omega) i ,b i ) Calculating the hidden layer output matrix H, and then calculating the weight beta, beta = H -1 Y; taking beta as an ELM model parameter, and establishing an ELM model;
and (4) inputting the voltage signal corresponding to the test data set in the step one into the built ELM model, and outputting the voltage signal predicted by the ELM model.
4. The ELM-CHMM-based power failure prediction method of claim 3, wherein: in the third step, according to the frequency domain characteristics of the voltage signals output by the power supply, decomposing and reconstructing the voltage signal data corresponding to the training data set in each state acquired in the first step by utilizing wavelet packet analysis, and extracting characteristic parameters of voltages in four states of slight degradation, moderate degradation, severe degradation and complete failure;
the specific process is as follows:
and C, enabling voltage signal data corresponding to the training data set in each state acquired in the step I to be S, and reconstructing at the qth node of the p layer after wavelet packet decomposition to obtain a discrete point signal S pq (k) K =1,2, L is the number of points after signal reconstruction;
the signal energy E in each equidistant frequency band after wavelet packet decomposition and reconstruction pq And total energy E is calculated as:
Figure FDA0003828667740000041
Figure FDA0003828667740000042
obtaining a frequency domain characteristic vector T' extracted by the wavelet packet,
Figure FDA0003828667740000043
and the frequency domain characteristic vector T' extracted by the wavelet packet is the ratio of the energy of each decomposed frequency band to the total energy, and voltage characteristic parameters of a training data set of the power supply with light degradation, medium degradation, heavy degradation and complete failure are obtained.
5. The ELM-CHMM-based power failure prediction method of claim 4, wherein: in the fourth step, according to the characteristic parameters of the four state voltages obtained in the third step, training a CHMM state evaluation model to obtain model parameters of the power supply in four states of slight degradation, moderate degradation, severe degradation and complete failure, and completing establishment of the CHMM state evaluation model of the four power supply states; the specific process is as follows:
the hidden markov model has the following five elements:
u is the total number of states in the hidden Markov model, namely four states of light degradation, moderate degradation, heavy degradation and complete failure;
v, V is the total number of observed values observed in each state in the hidden Markov model, and the total number of the observed values is the number of characteristic parameters of the voltages of the four states obtained in the third step;
π＝{π ξ the initial state probability distribution represents the probability of being in the xi state at the beginning, and xi is slight degradation, moderate degradation, severe degradation or complete failure;
A＝{a ξζ a is a transition probability distribution matrix of the state, and represents the probability that the state is transferred from xi to zeta, and zeta is light degradation, moderate degradation, severe degradation or complete failure;
B＝{b ζ (r), B is an observed probability distribution matrix representing when in state ζThe probability that the observation vector is the r-th bar;
a hidden Markov model is represented by lambda = (U, V, pi, A, B), and simplified into lambda = (pi, A, B);
taking the characteristic parameters of the four state voltages obtained in the step three as an observation vector O Original ；
Will observe vector O Original As an input of the hidden markov model HMM, λ = (pi, a, B) is used as a parameter of the hidden markov model HMM, and the hidden markov model HMM is trained to obtain model parameters λ of the light degradation fault HMM1, the moderate degradation fault HMM2, the heavy degradation fault HMM3, and the complete fault HMM4, respectively 1 、λ 2 、λ 3 、λ 4 And finishing the establishment of the CHMM state evaluation model of the four power states.
6. The ELM-CHMM-based power failure prediction method of claim 5, wherein: the concrete process of the step five is as follows:
decomposing and reconstructing the voltage signal data predicted by the ELM model in the second step by utilizing wavelet packet analysis, extracting characteristic parameters of voltages in four states of light degradation, moderate degradation, heavy degradation and complete failure, taking the characteristic parameters of each state as a group, and respectively passing each group through a system with four model parameters of lambda 1 、λ 2 、λ 3 、λ 4 The HMM1, HMM2, HMM3 and HMM4 models utilize the forward-backward algorithm of the CHMM to make each group obtain four log-likelihood probabilities, which are respectively P (O) Prediction |λ 1 )、P(O Prediction |λ 2 )、P(O Prediction |λ 3 )、P(O Prediction |λ 4 )；
The state type corresponding to each group of the maximum log-likelihood probability values is a power failure prediction result, if the probability that the prediction result is consistent with the power state of the test data set is greater than 85%, the power failure prediction is accurate, a power failure prediction model of the ELM-CHMM is obtained, a power failure voltage signal to be tested is input into the power failure prediction model of the ELM-CHMM, the state of the power failure to be tested is obtained, the states comprise four states of slight degradation, moderate degradation, severe degradation and complete failure, and the evaluation of the power failure degree is realized;
and if the probability that the prediction result is consistent with the power state of the test data set is less than or equal to 85%, the power failure prediction is inaccurate.
CN201811488243.9A 2018-12-06 2018-12-06 Power failure prediction method based on ELM-CHMM Active CN109615003B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811488243.9A CN109615003B (en) 2018-12-06 2018-12-06 Power failure prediction method based on ELM-CHMM

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811488243.9A CN109615003B (en) 2018-12-06 2018-12-06 Power failure prediction method based on ELM-CHMM

Publications (2)

Publication Number Publication Date
CN109615003A CN109615003A (en) 2019-04-12
CN109615003B true CN109615003B (en) 2022-11-08

Family

ID=66006652

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811488243.9A Active CN109615003B (en) 2018-12-06 2018-12-06 Power failure prediction method based on ELM-CHMM

Country Status (1)

Country Link
CN (1) CN109615003B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112816898B (en) * 2021-01-26 2022-03-01 三一重工股份有限公司 Battery failure prediction method and device, electronic equipment and storage medium
CN117214766B (en) * 2023-11-09 2024-02-09 深圳市蓝之洋科技有限公司 Mobile power supply fault detection method, device and equipment

Citations (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102606557A (en) * 2012-01-16 2012-07-25 北京航空航天大学 Health evaluation method for hydraulic system based on fault observer and SOM (self-organized mapping)
CN102830341A (en) * 2012-08-28 2012-12-19 南京航空航天大学 Online intelligent fault prediction method for power electronic circuit based on RS-CMAC (rough sets and cerebellar model articulation controller)
CN102867132A (en) * 2012-10-16 2013-01-09 南京航空航天大学 Aviation direct-current converter online fault combined prediction method based on fractional order wavelet transformation
CN103176128A (en) * 2013-03-28 2013-06-26 华南理工大学 Method and system for forcasting state of wind generating set and diagnosing intelligent fault
CN103995237A (en) * 2014-05-09 2014-08-20 南京航空航天大学 Satellite power supply system online fault diagnosis method
CN104463347A (en) * 2014-11-06 2015-03-25 电子科技大学 Method for electronic product degenerate state trend prediction with singular signals
CN104807639A (en) * 2015-04-23 2015-07-29 广西大学 Fault diagnosis method and device for rolling bearing of running gear of locomotive
CN105095918A (en) * 2015-09-07 2015-11-25 上海交通大学 Multi-robot system fault diagnosis method
CN105334472A (en) * 2015-10-26 2016-02-17 安徽理工大学 Online remaining life prediction method for mining intrinsic safety power supply
CN105975749A (en) * 2016-04-28 2016-09-28 上海交通大学 Bearing health assessment and prediction method and system
CN107169243A (en) * 2017-06-27 2017-09-15 电子科技大学 A kind of fuel cell Forecasting Methodology based on HSMM and empirical model
CN107202027A (en) * 2017-05-24 2017-09-26 重庆大学 A kind of large fan operation trend analysis and failure prediction method
CN108090427A (en) * 2017-12-07 2018-05-29 上海电机学院 Fault Diagnosis of Gear Case method based on flock of birds algorithm and Hidden Markov Model
CN108168924A (en) * 2017-12-20 2018-06-15 东北石油大学 A kind of reciprocating compressor life-span prediction method based on VMD and MFSS models
CN108256173A (en) * 2017-12-27 2018-07-06 南京航空航天大学 A kind of Gas path fault diagnosis method and system of aero-engine dynamic process
CN108763654A (en) * 2018-05-03 2018-11-06 国网江西省电力有限公司信息通信分公司 A kind of electrical equipment fault prediction technique based on Weibull distribution and hidden Semi-Markov Process
CN108892014A (en) * 2018-09-19 2018-11-27 歌拉瑞电梯股份有限公司 A kind of elevator internal contracting brake fault early warning method based on Elman neural network

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030184307A1 (en) * 2002-02-19 2003-10-02 Kozlowski James D. Model-based predictive diagnostic tool for primary and secondary batteries
US10949762B2 (en) * 2015-09-30 2021-03-16 Tata Consultancy Services Limited Methods and systems for optimizing hidden Markov Model based land change prediction

Patent Citations (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102606557A (en) * 2012-01-16 2012-07-25 北京航空航天大学 Health evaluation method for hydraulic system based on fault observer and SOM (self-organized mapping)
CN102830341A (en) * 2012-08-28 2012-12-19 南京航空航天大学 Online intelligent fault prediction method for power electronic circuit based on RS-CMAC (rough sets and cerebellar model articulation controller)
CN102867132A (en) * 2012-10-16 2013-01-09 南京航空航天大学 Aviation direct-current converter online fault combined prediction method based on fractional order wavelet transformation
CN103176128A (en) * 2013-03-28 2013-06-26 华南理工大学 Method and system for forcasting state of wind generating set and diagnosing intelligent fault
CN103995237A (en) * 2014-05-09 2014-08-20 南京航空航天大学 Satellite power supply system online fault diagnosis method
CN104463347A (en) * 2014-11-06 2015-03-25 电子科技大学 Method for electronic product degenerate state trend prediction with singular signals
CN104807639A (en) * 2015-04-23 2015-07-29 广西大学 Fault diagnosis method and device for rolling bearing of running gear of locomotive
CN105095918A (en) * 2015-09-07 2015-11-25 上海交通大学 Multi-robot system fault diagnosis method
CN105334472A (en) * 2015-10-26 2016-02-17 安徽理工大学 Online remaining life prediction method for mining intrinsic safety power supply
CN105975749A (en) * 2016-04-28 2016-09-28 上海交通大学 Bearing health assessment and prediction method and system
CN107202027A (en) * 2017-05-24 2017-09-26 重庆大学 A kind of large fan operation trend analysis and failure prediction method
CN107169243A (en) * 2017-06-27 2017-09-15 电子科技大学 A kind of fuel cell Forecasting Methodology based on HSMM and empirical model
CN108090427A (en) * 2017-12-07 2018-05-29 上海电机学院 Fault Diagnosis of Gear Case method based on flock of birds algorithm and Hidden Markov Model
CN108168924A (en) * 2017-12-20 2018-06-15 东北石油大学 A kind of reciprocating compressor life-span prediction method based on VMD and MFSS models
CN108256173A (en) * 2017-12-27 2018-07-06 南京航空航天大学 A kind of Gas path fault diagnosis method and system of aero-engine dynamic process
CN108763654A (en) * 2018-05-03 2018-11-06 国网江西省电力有限公司信息通信分公司 A kind of electrical equipment fault prediction technique based on Weibull distribution and hidden Semi-Markov Process
CN108892014A (en) * 2018-09-19 2018-11-27 歌拉瑞电梯股份有限公司 A kind of elevator internal contracting brake fault early warning method based on Elman neural network

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
A data-driven prognostic approach based on wavelet transform and extreme learning machine;S. Laddada等;《The 5th International Conference on Electrical Engineering – Boumerdes (ICEE-B)》;20171031;第1-4页 *
SW-ELM:A summation wavelet extreme learning machine algorithm with a priori parameter initialization;KamranJaved等;《Neurocomputing》;20140110;第123卷;第299-307页 *
基于AP-HMM混合模型的充电桩故障诊断;林越等;《广西师范大学学报（自然科学版）》;20180131;第36卷(第1期);第25-33页 *
基于ELM与HMM的序列飞机目标识别算法研究;成杰;《中国优秀硕士学位论文全文数据库 信息科技辑》;20180415(第4期);第I138-2774页 *
基于马尔科夫链的船舶输电***短路故障预测方法研究;许奇歆;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》;20180115(第1期);第C036-116页 *

Also Published As

Publication number Publication date
CN109615003A (en) 2019-04-12

Similar Documents

Publication Publication Date Title
Zheng et al. A data-driven approach for remaining useful life prediction of aircraft engines
CN109814527B (en) Industrial equipment fault prediction method and device based on LSTM recurrent neural network
CN108038300B (en) Optical fiber state evaluation method based on improved membership function combined with neural network
Liu et al. A novel method for online health prognosis of equipment based on hidden semi-Markov model using sequential Monte Carlo methods
CN111626345A (en) Multi-stage deep convolution transfer learning fault diagnosis method between different bearing devices
CN109615003B (en) Power failure prediction method based on ELM-CHMM
CN114429153A (en) Lifetime learning-based gearbox increment fault diagnosis method and system
CN114722952B (en) Bearing fault detection and classification integrated method based on representation learning
CN113537649B (en) Power grid load prediction method and system based on neural network and dynamic mode decomposition
CN112185104A (en) Traffic big data restoration method based on countermeasure autoencoder
CN111815806A (en) Method for preprocessing flight parameter data based on wild value elimination and feature extraction
CN111680875A (en) Unmanned aerial vehicle state risk fuzzy comprehensive evaluation method based on probability baseline model
Xie et al. Neural-network based structural health monitoring with wireless sensor networks
Sun et al. Lightweight bidirectional long short-term memory based on automated model pruning with application to bearing remaining useful life prediction
CN115510950A (en) Aircraft telemetry data anomaly detection method and system based on time convolution network
CN115902642A (en) Battery state of charge estimation method and device, electronic equipment and storage medium
Ren et al. DLformer: A dynamic length transformer-based network for efficient feature representation in remaining useful life prediction
CN113609766B (en) Soft measurement method based on depth probability hidden model
CN113780160A (en) Electric energy quality disturbance signal classification method and system
Chandra et al. Cooperative neuro-evolution of Elman recurrent networks for tropical cyclone wind-intensity prediction in the south pacific region
Dang et al. seq2graph: Discovering dynamic non-linear dependencies from multivariate time series
CN113740671A (en) Fault arc identification method based on VMD and ELM
Joslyn et al. A supervised learning approach to water quality parameter prediction and fault detection
Dotche et al. Evaluating Solar Energy Harvesting using Artificial Neural Networks: A Case study in Togo
Xiaoqian A sensor fault diagnosis algorithm for uav based on neural network

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