CN109308522B - GIS fault prediction method based on recurrent neural network - Google Patents

GIS fault prediction method based on recurrent neural network Download PDF

Info

Publication number
CN109308522B
CN109308522B CN201811018590.5A CN201811018590A CN109308522B CN 109308522 B CN109308522 B CN 109308522B CN 201811018590 A CN201811018590 A CN 201811018590A CN 109308522 B CN109308522 B CN 109308522B
Authority
CN
China
Prior art keywords
model
data
recurrent neural
fault
time
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
CN201811018590.5A
Other languages
Chinese (zh)
Other versions
CN109308522A (en
Inventor
苗红霞
常远
朱乾震
张衡
齐本胜
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.)
Changzhou Campus of Hohai University
Original Assignee
Changzhou Campus of Hohai University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Changzhou Campus of Hohai University filed Critical Changzhou Campus of Hohai University
Priority to CN201811018590.5A priority Critical patent/CN109308522B/en
Publication of CN109308522A publication Critical patent/CN109308522A/en
Application granted granted Critical
Publication of CN109308522B publication Critical patent/CN109308522B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/06—Electricity, gas or water supply

Abstract

The invention relates to a GIS fault prediction method based on a recurrent neural network. The device comprises a data processing module and a recurrent neural network identification module. In the data processing module, collected GIS historical parameter data for a long period of time is processed, and then a training and testing sample of the to-be-trained recurrent neural model is constructed by using a mathematical function assignment method. In a recurrent neural network identification module, firstly constructing a recurrent neural model structure, then using the recurrent neural model to detect abnormal points, attaching a label to sample data, then using the sample data with the label determined to train the recurrent neural model, adjusting model parameters, and using the corrected recurrent neural model as a GIS fault prediction model; and finally, inputting test data, and predicting and outputting the probability and the fault type of the GIS which may fail in the future.

Description

GIS fault prediction method based on recurrent neural network
Technical Field
The invention relates to a GIS (geographic information System) fault prediction method of a recurrent neural network, belonging to the technical field of pre-installed substations.
Background
In recent years, with the increase in voltage class and the increase in power consumption, people have been increasingly demanding more reliability in power supply. Gas Insulated switchgear GIS (gas Insulated switchgear) occupies an important link in the power distribution link due to the characteristics of small occupied area, high reliability, strong shock resistance and the like. Once the GIS fails, large-scale power grid paralysis can be caused, and great influence is brought to the daily life of the people and necessary industrial production. The maintenance of the faulty GIS not only needs to be carried out by power failure maintenance and dismantling the external shell for maintenance, but also costs longer maintenance time, and further causes great loss to national economy. And the GIS fault can be effectively checked by timely predicting the fault, and the danger is relieved before the GIS fault endangers the power grid. In this case, how to predict the failure before the failure of the GIS device occurs becomes a demand for power distribution safety. The method has the advantages that the fault type and the fault occurrence probability are judged quickly and accurately before the GIS fault occurs, and the method has important significance for reducing power failure time and improving maintenance efficiency and equipment utilization rate.
Unlike traditional modeling or training of small amounts of data, deep learning can mine and learn complex structures, i.e., feature extraction, inside data through training of large amounts of data. The concrete application can be abstracted into a process of sending original data and signals into a neural network as low-level features and representing information which people want to obtain as high-level. The main structure of deep learning includes Convolutional Neural Network (CNN), Deep Belief Network (DBN), and Recurrent Neural Network (RNN). Unlike CNNs that retain some of the historical inputs through sliding windows, the Recurrent Neural Network (RNN) employed herein can retain all of the historical inputs, with high classification accuracy by abstracting the correlation between the historical inputs. Furthermore, the RNN has good universality and good identification effect on various input signals, and is widely applied to solving the problem related to time series. At present, the RNN has been successfully applied to the fields of natural language processing, computer vision, disease prediction and the like, and good effects are achieved. The fault prediction is similar to language processing and disease prediction, most sensor signals of various devices are time series, and signals at the current moment are inseparably connected with historical data, so that the fault prediction requirement can be met by using the RNN.
Disclosure of Invention
The invention solves the problems: aiming at the defects of the existing GIS fault prediction technology, a GIS fault prediction method based on a recurrent neural network is provided, and the defects of the prior art are filled.
The technical scheme of the invention is as follows: a GIS fault prediction method based on a recurrent neural network comprises a data processing module and a recurrent neural network identification module. The method comprises the steps of establishing a training sample in a data processing module by using a mathematical function assignment method, firstly establishing a cyclic neural model structure in a cyclic neural network recognition module, then using the cyclic neural model to detect abnormal points, attaching labels to training sample data, then using the training sample established by the data processing module to train the cyclic neural model, adjusting model parameters, and determining a GIS fault prediction model. And finally, inputting the data into a cyclic neural model to carry out fault prediction on the GIS. The method specifically comprises the following steps:
1) collecting data of three signals of vibration, air pressure and current on a GIS sensor for a long time as historical data;
2) in a data processing module, establishing a training and testing sample matrix by using a mathematical function assignment method;
3) in the recurrent neural network identification module, firstly, a recurrent neural model structure is determined, then, the recurrent neural model is used for abnormal point detection, a label is attached to sample data, then, a training sample is input to adjust model parameters, and finally, test data is input to the recurrent neural model to obtain a prediction result.
In the step 1), collecting GIS vibration, air pressure and current data of the GIS sensor for a long period of time, namely 183 days or more.
The data processing module in the step 2) comprises the following steps:
21) sampling historical data of the extracted vibration, air pressure and current signals by taking a day as a unit;
22) constructing training sample data of the to-be-trained recurrent neural model by a mathematical function assignment method, and constructing test data by using the same method;
in the step 22), the method for constructing the sample matrix by assigning the mathematical function comprises the following steps: the fault waveforms of two different sensors of the same signal are collected, and a plurality of sampling points are respectively extracted from two fault waveform functions a (1, t) and a (2, t) to form a five-dimensional matrix b. The matrix has 180 columns representing 180 samples each. The matrix is as follows:
Figure GDA0003344609150000021
where b is the input matrix, a (1, t) and a (2, t) are two fault waveform functions, respectively, and t is the abscissa of the history data, i.e., the number of days.
The step 3) specifically includes the following steps:
31) determining the specific structure of the recurrent neural model, including the number of hidden layers and the number of neurons in each layer, and building the recurrent neural model; the constructed recurrent neural model adopts a model with 1 input layer, 4 hidden layers and 1 output layer, and the number of neurons contained in each hidden layer is 10;
32) carrying out abnormal point detection by using the built cyclic nerve model, and attaching a label to training sample data;
33) inputting the labeled training sample data into a recurrent neural model for training, so that the recurrent neural model automatically adjusts model parameters, and taking the trained recurrent neural model as a GIS fault prediction model;
34) and inputting the current GIS signal data serving as input data into a GIS fault prediction model, and learning the input data through the GIS fault prediction model to finally obtain the probability and fault signal type of the GIS possible hidden fault at the current moment.
The abnormal point detection method comprises the following steps: when the fault signal has a slight change, a point a at which the fault signal starts to change is detected, called a fault signal starting point, and then a point B at which the fault actually occurs is detected, and a label of [0.1-1] is attached to a sample corresponding to a time period of T-B-a.
The abnormal point detection step comprises:
321) firstly, selecting an abnormal point moment v according to experience, wherein the value of v is late in time;
322) inputting data into a cyclic neural model for first training, and reserving data with a time margin M between a v value and a fault trend starting point A without putting the data into the cyclic neural model for training. Assume that the selected value is v 1;
323) assigning the output value of the fault from the abnormal moment v1 by using P, and enabling the change of the P value to be the same as the change trend of the actual fault;
324) and assigning a function P (T) to data from an abnormal point to a fault occurrence time B, wherein a fault model time domain T is B-v 1.
Figure GDA0003344609150000031
Where the argument T represents the time from the occurrence of the failure trend, and changes from 0 to T, and when T ═ T, the failure occurs, so T is the time from the discovery of the failure trend to the occurrence of the failure.
325) After the first training is finished, the same data is put into the test and the output value is observed;
326) the recurrent nerve detects the failure trend in a time interval delta T before the first selected v1, let n be the data number of the failure trend detected in the time interval delta T, and when the time interval delta T is within the time interval delta T, the data number of the failure trend in unit time
Figure GDA0003344609150000032
Satisfy the requirement of
Figure GDA0003344609150000033
(where u is a scaling factor and u ∈ (0,1)), let v2 be v1- Δ T when v2 is already closer to the failure trend starting point a than v 1;
327) training the network by re-labeling the data, and repeating the process until the network is trained
Figure GDA0003344609150000041
Or finishing training all data in the M time period.
328) Number of data with failure trend per unit time
Figure GDA0003344609150000042
Or when the data are completely trained, stopping updating the v value, wherein the v value is infinitely close to the moment of the starting point A of the fault trend.
In the above step 324), the final data is assigned using formula (1), and since the previously suspected abnormal point v is late, the value P is larger, formula (1) is modified according to formula (2):
Figure GDA0003344609150000043
where the argument t represents the time since the failure trend occurred, B represents the time at which the failure occurred, and v represents the finally updated v value.
In step 328), the method for determining the starting point with the failure trend includes: comparing the output value Pd of the undetermined time point with the average output value in the health state of the equipment, and judging according to the formula (3):
Figure GDA0003344609150000044
pd is an output value of the undetermined point, Pk is an output value of the health state at the moment k, s is a proportionality coefficient, s is set to be a value slightly larger than 1, and n is the number of data.
In the step 33), the training process of the recurrent neural model includes the following specific steps:
331) initialization: randomly initializing parameters of a recurrent neural model, wherein the parameters comprise three weight matrixes U, W and V and two bias matrixes b and c;
332) forward propagation: inputting training sample data into a recurrent neural model, obtaining a predicted value of the recurrent neural model under an initial model parameter through a forward propagation algorithm, and adjusting the model parameter by making a difference with a label of the training sample. The method comprises the following specific steps:
3321) calculating the hidden state h of the model at time t(t)，h(t)Can be inputted by x(t)And hidden state h at the previous moment(t-1)The formula is obtained as follows:
h(t)＝σ(Ux(t)+Wh(t-1)+b) (5)
the activation function sigma is generally tan h, the bias matrix b is the bias of linear relation, and the weight matrix U and W are linear relation parameters of the recurrent neural model. x is the number of(t)Representing the input of a training sample at time t, h(t-1)Representing the hidden state of the model at time t-1.
3322) Using the above formulaCalculated hidden state h(t)To calculate the output o of the model at time t(t)The formula is as follows:
o(t)＝Vh(t)+c (4)
wherein the weight matrix V and the bias matrix c are both recurrent neural model parameters.
333) And (3) back propagation: and (3) performing back propagation calculation on the recurrent neural model, calculating an error by comparing the previous output with the sample label, further correcting the model parameter by using a gradient descent method iteration according to the error, and adjusting the recurrent neural model parameter, wherein the recurrent neural model parameter comprises three weight matrixes U, W and V and two bias matrixes b and c. The method comprises the following specific steps:
3331) define the final loss as L, an
Figure GDA0003344609150000051
Wherein L is(t)For the loss function, t represents the time instant and τ represents the final time instant.
3332) Calculating a weight matrix V and a bias matrix c, wherein the specific formula is as follows:
Figure GDA0003344609150000052
Figure GDA0003344609150000053
3333) and calculating the weight matrix W, U and the bias matrix b, wherein the formulas are respectively as follows:
Figure GDA0003344609150000054
Figure GDA0003344609150000055
Figure GDA0003344609150000056
wherein delta(t)Representing the gradient of the hidden state at the t position, the function diag represents taking the diagonal elements of the matrix,
Figure GDA0003344609150000057
representing the predicted output at time t, y(t)Representing the actual output of the sample at time t.
334) And (3) repeatedly determining final parameters: after the adjusted model parameters are obtained, the final parameters are determined by the following specific steps:
3341) re-inputting the cyclic neural model after the parameters are adjusted by using the same training sample;
3342) and comparing the error between the output result and the sample label.
3343) If the error meets the requirement, determining the model parameters;
3344) if the difference is still large and does not meet the requirement, the steps 332) and 333) are repeated to adjust the parameters until the error meets the requirement;
3345) determining the final model parameters includes: and weighting the matrix U, W and V and the bias matrixes b and c, and taking the recurrent neural model at the moment as a GIS fault prediction model.
The final prediction output form in step 34) above is as follows:
Figure GDA0003344609150000061
three rows a, b and c of the matrix are used for representing the result of pattern recognition and representing the probability of the current GIS fault, the closer the values of a, b and c are to 1, the more easily the fault occurs, and i represents the state of the GIS at different moments and the result of GIS fault prediction at that moment. When a fault occurs, each signal containing an anomaly is encoded as follows:
signals containing anomalies abc
Abnormality of current 001
Abnormal air pressure 010
Abnormality of vibration 100
The invention achieves the following beneficial effects:
the GIS fault prediction method based on the recurrent neural network comprises faults caused by the abnormality of three different signals, namely vibration, air pressure and current. The method for detecting the abnormal points fully utilizes a large amount of label-free samples which are easy to obtain, labels are attached to the samples, and the problems that the conventional method for obtaining the labeled samples is high in difficulty and consumes manpower and material resources are solved. And for noise signals accompanying in the actual application process of the sensor, the circulating neural network is used to effectively identify the noise signals of the sensor in normal operation and filter the noise signals so that the noise signals do not influence the fault prediction result. By utilizing the fitting capacity of the recurrent neural network to the time series data and combining the current signal of the sensor and the historical data to predict the fault of the GIS equipment, the full utilization of the measured data can be realized, and the prediction accuracy is improved. The whole method has strong adaptability and generalization capability and has certain social value and practical significance.
Description of the drawings:
FIG. 1 is a flow chart of an implementation of the present invention;
fig. 2 is a diagram of a recurrent neural model training process.
The specific implementation mode is as follows:
the invention is further described below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.
As shown in fig. 1, a GIS fault prediction method based on a recurrent neural network includes the following steps:
step 1: collecting data of three signals of vibration, air pressure and current on a GIS sensor for a long time as historical data;
step 2: in a data processing module, establishing a training and testing sample matrix by using a mathematical function assignment method, wherein the data processing module comprises the following steps;
1) sampling historical data of the extracted vibration, air pressure and current signals by taking a day as a unit;
2) and constructing training sample data of the to-be-trained recurrent neural model by a mathematical function assignment method, and constructing test data by using the same method. The method for assigning the mathematical function is as follows:
the fault waveforms of two different sensors of the same signal are collected, and a plurality of sampling points are respectively extracted from two fault waveform functions a (1, t) and a (2, t) to form a five-dimensional matrix b. The matrix has 180 columns representing 180 samples each. The matrix is as follows:
Figure GDA0003344609150000071
where b is the input matrix, a (1, t) and a (2, t) are two fault waveform functions, respectively, and t is the abscissa of the history data, i.e., the number of days.
And 3, in the recurrent neural network identification module, firstly determining the structure of the recurrent neural model, then using the recurrent neural model to detect abnormal points, attaching labels to sample data, and then inputting training samples to adjust model parameters. The method specifically comprises the following steps:
1) and determining the specific structure of the recurrent neural model, including the number of hidden layers and the number of neurons in each layer, and constructing the recurrent neural model. The constructed recurrent neural model adopts a model with 1 input layer, 4 hidden layers and 1 output layer, and the number of neurons contained in each hidden layer is 10.
2) And (3) carrying out abnormal point detection by using the built cyclic nerve model, and attaching labels to training sample data, wherein the labels respectively correspond to vibration signal faults, air pressure signal faults and current signal faults and are normal. The abnormal point detection method comprises the following steps: when the fault signal has slight change, a point A where the fault signal starts to change is detected and called as a fault trend starting point, then a point B where the fault really occurs is detected, and a label of [0.1-1] is attached to a sample corresponding to a time period of T-B-A. The abnormal point detection method comprises the following steps:
21) firstly, selecting an abnormal point moment v according to experience, wherein the value of v is late in time;
22) inputting data into a cyclic neural model for first training, and reserving data with a time margin M between a v value and a fault trend starting point A without putting the data into the cyclic neural model for training. Assume that the selected value is v 1;
23) assigning the output value of the fault from the abnormal moment v1 by using P, and enabling the change of the P value to be the same as the change trend of the actual fault;
24) and assigning a function P (T) to data from an abnormal point to a fault occurrence time B, wherein a fault model time domain T is B-v 1.
Figure GDA0003344609150000081
Where the argument T represents the time from the occurrence of the failure trend, and changes from 0 to T, and when T ═ T, the failure occurs, so T is the time from the discovery of the failure trend to the occurrence of the failure. Then, formula (1) is corrected according to formula (2):
Figure GDA0003344609150000082
where the argument t represents the time since the failure trend occurred, B represents the time at which the failure occurred, and v represents the finally updated v value.
25) After the first training is finished, the same data is put into the test and the output value is observed;
26) the recurrent nerve detects the failure trend in a time interval delta T before the first selected v1, let n be the data number of the failure trend detected in the time interval delta T, and when the time interval delta T is within the time interval delta T, the data number of the failure trend in unit time
Figure GDA0003344609150000083
Satisfy the requirement of
Figure GDA0003344609150000084
(where u is a scaling factor and u ∈ (0,1)), let v2 be v1- Δ T when v2 is already closer to the failure trend starting point a than v 1;
27) training the network by re-labeling the data, and repeating the process until the network is trained
Figure GDA0003344609150000085
Or finishing training all data in the M time period.
28) Number of data with failure trend per unit time
Figure GDA0003344609150000086
Or when the data are completely trained, stopping updating the v value, wherein the v value is infinitely close to the moment of the starting point A of the fault trend. The method for judging the starting point with the fault trend comprises the following steps: comparing the output value Pd of the undetermined time point with the average output value in the health state of the equipment, and judging according to the formula (3):
Figure GDA0003344609150000087
pd is an output value of the undetermined point, Pk is an output value of the health state at the moment k, s is a proportionality coefficient, s is set to be a value slightly larger than 1, and n is the number of data.
3) The method comprises the following steps of inputting labeled training sample data into a cyclic neural model for training, automatically adjusting model parameters, and taking the trained cyclic neural model as a GIS fault prediction model, wherein the steps are as follows:
31) initialization: randomly initializing parameters of a recurrent neural model, wherein the parameters comprise three weight matrixes U, W and V and two bias matrixes b and c;
32) forward propagation: inputting training sample data into a recurrent neural model, obtaining a predicted value of the recurrent neural model under an initial model parameter through a forward propagation algorithm, and adjusting the model parameter by making a difference with a label of the training sample. The method comprises the following specific steps:
321) calculating the hidden state h of the model at time t(t)，h(t)Can be inputted by x(t)And hidden state h at the previous moment(t-1)The formula is obtained as follows:
h(t)＝σ(Ux(t)+Wh(t-1)+b) (5)
the activation function sigma is generally tan h, the bias matrix b is the bias of linear relation, and the weight matrix U and W are linear relation parameters of the recurrent neural model. x is the number of(t)Representing the input of a training sample at time t, h(t-1)Representing the hidden state of the model at time t-1.
322) Hidden state h calculated using the above formula(t)To calculate the output o of the model at time t(t)The formula is as follows:
o(t)＝Vh(t)+c (4)
wherein the weight matrix V and the bias matrix c are both recurrent neural model parameters.
33) And (3) back propagation: and (3) performing back propagation calculation on the recurrent neural model, calculating an error by comparing the previous output with the sample label, further correcting the model parameter by using a gradient descent method iteration according to the error, and adjusting the recurrent neural model parameter, wherein the recurrent neural model parameter comprises three weight matrixes U, W and V and two bias matrixes b and c. The method comprises the following specific steps:
331) define the final loss as L, an
Figure GDA0003344609150000091
Wherein L is(t)For the loss function, t represents the time instant and τ represents the final time instant.
332) Calculating a weight matrix V and a bias matrix c, wherein the specific formula is as follows:
Figure GDA0003344609150000092
Figure GDA0003344609150000093
333) and calculating the weight matrix W, U and the bias matrix b, wherein the formulas are respectively as follows:
Figure GDA0003344609150000101
Figure GDA0003344609150000102
Figure GDA0003344609150000103
wherein delta(t)Representing the gradient of the hidden state at the t position, the function diag represents taking the diagonal elements of the matrix,
Figure GDA0003344609150000104
representing the predicted output at time t, y(t)Representing the actual output of the sample at time t.
34) And (3) repeatedly determining final parameters: and (3) repeatedly determining final parameters: after the adjusted model parameters are obtained, the final parameters are determined by the following specific steps:
341) re-inputting the cyclic neural model after the parameters are adjusted by using the same training sample;
342) and comparing the error between the output result and the sample label.
343) If the error meets the requirement, determining the model parameters;
344) if the difference is still large and does not meet the requirement, the steps 332) and 333) are repeated to adjust the parameters until the error meets the requirement;
345) determining the final model parameters includes: and weighting the matrix U, W and V and the bias matrixes b and c, and taking the recurrent neural model at the moment as a GIS fault prediction model.
4) And inputting the current GIS signal data serving as input data into a GIS fault prediction model, and learning the input data through the GIS fault prediction model to finally obtain the probability and fault signal type of the GIS possible hidden fault at the current moment. The final prediction output form is as follows:
Figure GDA0003344609150000105
three rows a, b and c of the matrix are used for representing the result of pattern recognition and representing the probability of the current GIS fault, the closer the values of a, b and c are to 1, the more easily the fault occurs, and i represents the state of the GIS at different moments and the result of GIS fault prediction at that moment. When a fault occurs, each signal containing an anomaly is encoded as follows:
signals containing anomalies abc
Abnormality of current 001
Abnormal air pressure 010
Vibration deviceOften times 100
And 4, inputting data into the recurrent neural model to obtain a prediction result.
Example 1
The input signals are taken from a power plant and include vibration signals, current signals and air pressure signals.
Assigning values to the data of 5 time units before the fault occurs, putting the assigned data into a network for training, and putting all the data into a network for testing again to obtain the following results:
Time 131 132 133 134 135 136 137 138 139 140
a 0.100 0.035 0.032 0.043 0.272 0.401 0.544 0.999 0.924 0.910
c 0 0 0 0 0 0 0 0 0 0
b 0 0 0 0 0 0 0 0 0 0
the output average value of the health state is 0.0294, the proportionality coefficient s is 1.3, it can be seen that the time is 131-135, three of five data of the GIS fault prediction value a satisfy the formula (3), so that the data with the time of 131-140 are re-labeled and put into network training, all data are put into network testing again, and the result is as follows:
Time 126 127 128 129 130 131 132 133 134 135
a 0.070 0.369 0.200 0.062 0.704 0.354 0.297 0.378 0.451 0.559
b 0 0 0 0 0 0 0 0 0 0
c 0 0 0 0 0 0 0 0 0 0
the data was entered into MATLAB and the image was drawn. When t is 120, a failure does not occur, but a slight signal change starts, and when t is 140, a failure occurs, and a signal containing abnormal information is a vibration signal.
The above examples are provided for the purpose of describing the present invention only, and are not intended to limit the scope of the present invention. The scope of the invention is defined by the appended claims. Various equivalent substitutions and modifications can be made without departing from the spirit and principles of the invention, and are intended to be within the scope of the invention.

Claims (6)

1. A gas insulated switchgear fault prediction method based on a recurrent neural network is characterized in that a data processing module and a recurrent neural network identification module are utilized; the method comprises the following specific steps:
1) collecting data of three signals of vibration, air pressure and current on a sensor of the gas insulated switchgear for a long time as historical data;
2) in a data processing module, establishing a training and testing sample matrix by using a mathematical function assignment method;
in the step 2), a training and testing sample matrix is established by using a mathematical function assignment method, and the method comprises the following steps:
21) sampling historical data of the extracted vibration, air pressure and current signals by taking a day as a unit;
22) constructing training sample data of the to-be-trained recurrent neural model by a mathematical function assignment method, and constructing test data by using the same method; in step 22), the method for constructing the sample matrix by using the mathematical function assignment comprises the following steps: collecting fault waveforms of two different sensors of the same signal, and respectively extracting a plurality of sampling points in two fault waveform functions a (1, t) and a (2, t) to combine into a five-dimensional matrix b; 180 columns of the matrix represent 180 samples respectively; the matrix is as follows:
Figure FDA0003354258980000011
wherein b is an input matrix, a (1, t) and a (2, t) are two fault waveform functions respectively, and t is the abscissa of the historical data, namely the number of days;
3) in a recurrent neural network identification module, firstly determining a recurrent neural model structure, then using the recurrent neural model to detect abnormal points, attaching labels to sample data, then inputting training samples to adjust model parameters, and finally inputting test data to the recurrent neural model to obtain a prediction result;
31) determining the specific structure of the recurrent neural model, including the number of hidden layers and the number of neurons in each layer, and building the recurrent neural model; the constructed recurrent neural model adopts a model with 1 input layer, 4 hidden layers and 1 output layer, and the number of neurons contained in each hidden layer is 10;
32) carrying out abnormal point detection by using the built cyclic nerve model, and attaching a label to training sample data;
33) inputting the labeled training sample data into a cyclic neural model for training, so that the cyclic neural model automatically adjusts model parameters, and taking the trained cyclic neural model as a fault prediction model of the gas insulated switchgear;
34) and inputting the current signal data of the gas insulated switchgear as input data into a gas insulated switchgear fault prediction model, and learning the input data through the gas insulated switchgear fault prediction model to finally obtain the probability and fault signal type of the possible hidden fault of the gas insulated switchgear at the current moment.
In step 32), the method for detecting the abnormal point comprises the following steps: when the fault signal has slight change, detecting a point A where the fault signal begins to change, called as a fault trend starting point, then detecting a point B where the fault really occurs, and attaching a label of [0.1-1] to a sample corresponding to a time period T-B-A;
the abnormal point detection step comprises:
321) firstly, selecting an abnormal point moment v according to experience, wherein the value of v is late in time;
322) inputting data into a cyclic neural model for first training, reserving data of a time margin M between a v value and a fault trend starting point A, and not putting the data into the cyclic neural model for training; assume that the selected value is v 1;
323) assigning the output value of the fault from the abnormal moment v1 by using P, and enabling the change of the P value to be the same as the change trend of the actual fault;
324) assigning a function P (T) to data from an abnormal point to a fault occurrence time B, wherein a fault model time domain T is B-v 1;
Figure FDA0003354258980000021
wherein, the independent variable T represents the time from the occurrence of the failure trend, and changes from 0 to T, and when T equals to T, the failure occurs, so T is the time from the discovery of the failure trend to the occurrence of the failure;
325) after the first training is finished, the same data is put into the test and the output value is observed;
326) the recurrent nerve detects the failure trend in a time interval delta T before the first selected v1, let n be the data number of the failure trend detected in the time interval delta T, and when the time interval delta T is within the time interval delta T, the data number of the failure trend in unit time
Figure FDA0003354258980000022
Satisfy the requirement of
Figure FDA0003354258980000023
When u is a proportionality coefficient and u is belonged to (0,1), making v2 equal to v1- Δ T, and when v2 is closer to the failure trend starting point a than v 1;
327) training the network by re-labeling the data, and repeating the process until the network is trained
Figure FDA0003354258980000024
Or finishing training all data in the M time period.
328) Number of data with failure trend per unit time
Figure FDA0003354258980000025
Or when the data are completely trained, stopping updating the v value, wherein the v value is infinitely close to the moment of the starting point A of the fault trend.
2. The fault prediction method of the gas insulated switchgear based on the recurrent neural network as claimed in claim 1, wherein: in the step 1), collecting vibration, air pressure and current data of the gas insulated switchgear for a long period of time, namely 183 days or more, of the gas insulated switchgear.
3. The fault prediction method of the gas insulated switchgear based on the recurrent neural network as claimed in claim 1, wherein: in step 324), the final data is assigned using formula (1), since the previously suspected abnormal point v is late, the value P is larger, and formula (1) is modified according to formula (2):
Figure FDA0003354258980000026
where the argument t represents the time since the failure trend occurred, B represents the time at which the failure occurred, and v represents the finally updated v value.
4. The fault prediction method of the gas insulated switchgear based on the recurrent neural network as claimed in claim 1, wherein: in step 326), the method for determining the starting point with the failure trend includes: the output value P of the undetermined time pointdComparing the output average value with the output average value in the health state of the equipment, and judging according to the formula (3);
Figure FDA0003354258980000027
wherein P isdOutputting a value, P, for a pending time pointkAnd s is a proportionality coefficient and is set to a value slightly larger than 1, and n is the data number.
5. The fault prediction method of the gas insulated switchgear based on the recurrent neural network as claimed in claim 1, wherein: step 33), a training process of the recurrent neural model, which comprises the following specific steps:
331) initialization: randomly initializing parameters of a recurrent neural model, wherein the parameters comprise three weight matrixes U, W and V and two bias matrixes b and c;
332) forward propagation: inputting training sample data into a cyclic neural model, obtaining a predicted value of the cyclic neural model under an initial model parameter through a forward propagation algorithm, and adjusting the model parameter by making a difference with a label of the training sample; the method comprises the following specific steps:
3321) calculating the hidden state h of the model at time t(t)，h(t)By inputting x(t)And hidden state h at the previous moment(t-1)The formula is obtained as follows:
h(t)＝σ(Ux(t)+Wh(t-1)+b) (5)
wherein the activation function sigma is tanh, the bias matrix b is the bias of linear relation, and the weight matrix U and W are linear relation parameters of the recurrent neural model; x is the number of(t)Representing the input of a training sample at time t, h(t-1)Representing the hidden state of the model at the moment t-1;
3322) hidden state h calculated using the above formula(t)To calculate the output o of the model at time t(t)The formula is as follows:
o(t)＝Vh(t)+c (4)
wherein the weight matrix V and the bias matrix c are both recurrent neural model parameters;
333) and (3) back propagation: performing back propagation calculation on the recurrent neural model, and calculating an error by comparing the previous output with a sample label, thereby further correcting the model parameters by using a gradient descent method iteration according to the error, and adjusting the recurrent neural model parameters, wherein the parameters comprise three weight matrixes U, W and V and two bias matrixes b and c; the method comprises the following specific steps:
3331) define the final loss as L, an
Figure FDA0003354258980000031
Wherein L is(t)For the loss function, t represents the time, and τ represents the final time;
3332) calculating a weight matrix V and a bias matrix c, wherein the specific formula is as follows:
Figure FDA0003354258980000032
Figure FDA0003354258980000033
3333) and calculating the weight matrix W, U and the bias matrix b, wherein the formulas are respectively as follows:
Figure FDA0003354258980000034
Figure FDA0003354258980000035
Figure FDA0003354258980000036
wherein delta(t)Representing the gradient of the hidden state at the t position, the function diag represents taking the diagonal elements of the matrix,
Figure FDA0003354258980000037
representing the predicted output at time t, y(t)Representing the actual output of the sample at time t;
334) and (3) repeatedly determining final parameters: after the adjusted model parameters are obtained, the final parameters are determined by the following specific steps:
3341) re-inputting the cyclic neural model after the parameters are adjusted by using the same training sample;
3342) and comparing the error between the output result and the sample label.
3343) If the error meets the requirement, determining the model parameters;
3344) if the difference is still large and does not meet the requirement, the steps 332) and 333) are repeated to adjust the parameters until the error meets the requirement;
3345) determining the final model parameters includes: and weighting the matrix U, the matrix W, the matrix V and the bias matrixes b and c, and taking the recurrent neural model at the moment as a gas insulated switchgear fault prediction model.
6. The fault prediction method of the gas insulated switchgear based on the recurrent neural network as claimed in claim 1, wherein: the final prediction output form in step 34) is as follows:
Figure FDA0003354258980000041
three rows a, b and c of the matrix are used for representing the result of pattern recognition and representing the probability of the fault occurrence of the current gas insulated switchgear, the closer the values of a, b and c are to 1, the more easily the fault occurs, and i represents the state of the gas insulated switchgear at different moments and the result of the fault prediction of the gas insulated switchgear at that moment; when a fault occurs, each signal containing an anomaly is encoded as follows:
signal containing anomalies: abc;
current anomaly: 001;
abnormal air pressure: 010;
abnormal vibration: 100.
CN201811018590.5A 2018-09-03 2018-09-03 GIS fault prediction method based on recurrent neural network Active CN109308522B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811018590.5A CN109308522B (en) 2018-09-03 2018-09-03 GIS fault prediction method based on recurrent neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811018590.5A CN109308522B (en) 2018-09-03 2018-09-03 GIS fault prediction method based on recurrent neural network

Publications (2)

Publication Number Publication Date
CN109308522A CN109308522A (en) 2019-02-05
CN109308522B true CN109308522B (en) 2022-02-22

Family

ID=65224525

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811018590.5A Active CN109308522B (en) 2018-09-03 2018-09-03 GIS fault prediction method based on recurrent neural network

Country Status (1)

Country Link
CN (1) CN109308522B (en)

Families Citing this family (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109872014A (en) * 2019-03-29 2019-06-11 永安行科技股份有限公司 Hydrogen energy source moped day usage amount and fault condition prediction technique
CN110210516A (en) * 2019-04-25 2019-09-06 合刃科技（武汉）有限公司 Lathe abnormality recognition method, system and terminal device based on deep learning
CN110647924B (en) * 2019-09-05 2022-04-22 国网河北省电力有限公司电力科学研究院 GIS equipment state evaluation method based on support vector description and K-nearest neighbor algorithm
CN110737732A (en) * 2019-10-25 2020-01-31 广西交通科学研究院有限公司 electromechanical equipment fault early warning method
CN111160454B (en) * 2019-12-27 2023-12-22 北京旋极信息技术股份有限公司 Quick change signal detection method and device
CN111210231A (en) * 2019-12-31 2020-05-29 联想(北京)有限公司 Information processing method and information processing apparatus
CN111865863B (en) * 2020-07-20 2021-07-20 山东大学 RNN neural network-based OFDM signal detection method
CN112633550B (en) * 2020-11-23 2023-07-18 成都唐源电气股份有限公司 RNN-based contact network fault trend prediction method, equipment and storage medium
CN112670996B (en) * 2020-12-03 2022-05-17 浙江大学 CNN-based GPU parallel power flow out-of-limit discrimination method
CN112650204B (en) * 2020-12-30 2021-12-24 中南大学 Intelligent track unmanned vehicle fault gene identification method and system
CN112836577B (en) * 2020-12-30 2024-02-20 中南大学 Intelligent traffic unmanned vehicle fault gene diagnosis method and system
CN112926752B (en) * 2021-01-26 2022-07-01 西华大学 Intelligent power transmission network fault line detection and fault recovery method considering energy storage system
CN112883639A (en) * 2021-02-03 2021-06-01 国网浙江省电力有限公司宁波供电公司 GIS equipment service life prediction device and method based on machine learning
CN113866572A (en) * 2021-09-06 2021-12-31 西安理工大学 Direct-current fault arc detection and positioning method under condition of access of multiple power electronic devices
CN117372880B (en) * 2023-12-07 2024-02-09 天津市祥途测绘科技有限公司 Road engineering supervision system and method based on remote sensing image

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107769972A (en) * 2017-10-25 2018-03-06 武汉大学 A kind of power telecom network equipment fault Forecasting Methodology based on improved LSTM
CN108241298A (en) * 2018-01-09 2018-07-03 南京航空航天大学 A kind of aerogenerator method for diagnosing faults based on FWA-RNN models
CN108268935A (en) * 2018-01-11 2018-07-10 浙江工业大学 A kind of PM2.5 concentration values Forecasting Methodology and system based on sequential Recognition with Recurrent Neural Network

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107769972A (en) * 2017-10-25 2018-03-06 武汉大学 A kind of power telecom network equipment fault Forecasting Methodology based on improved LSTM
CN108241298A (en) * 2018-01-09 2018-07-03 南京航空航天大学 A kind of aerogenerator method for diagnosing faults based on FWA-RNN models
CN108268935A (en) * 2018-01-11 2018-07-10 浙江工业大学 A kind of PM2.5 concentration values Forecasting Methodology and system based on sequential Recognition with Recurrent Neural Network

Also Published As

Publication number Publication date
CN109308522A (en) 2019-02-05

Similar Documents

Publication Publication Date Title
CN109308522B (en) GIS fault prediction method based on recurrent neural network
CN107909206B (en) PM2.5 prediction method based on deep structure recurrent neural network
CN111460728B (en) Method and device for predicting residual life of industrial equipment, storage medium and equipment
CN107786369A (en) Based on the perception of IRT step analyses and LSTM powerline network security postures and Forecasting Methodology
CN108958217A (en) A kind of CAN bus message method for detecting abnormality based on deep learning
CN110516735B (en) Natural gas pipeline event classification method based on LSTM network and Adam algorithm
CN110119853B (en) Water supply network leakage alarm threshold value selection method based on time series monitoring data analysis
CN112765908B (en) Gas turbine engine combustion chamber stability analysis method based on image sequence analysis
CN111814699B (en) Deep learning earthquake prediction method for SWARM electromagnetic satellite data
CN112000015A (en) Intelligent BIT design method for heavy-duty gas turbine control system controller module based on LSTM and bio-excitation neural network
CN115951014A (en) CNN-LSTM-BP multi-mode air pollutant prediction method combining meteorological features
CN112215279B (en) Power grid fault diagnosis method based on immune RBF neural network
CN110737948A (en) method for predicting residual life of aero-engine based on deep FNN-LSTM hybrid network
CN111122811A (en) Sewage treatment process fault monitoring method of OICA and RNN fusion model
CN113516304A (en) Space-time joint prediction method and device for regional pollutants based on space-time graph network
CN116307152A (en) Traffic prediction method for space-time interactive dynamic graph attention network
CN113470316A (en) Debris flow monitoring and early warning method based on self-coding single classification model
CN105183659A (en) Software system behavior anomaly detection method based on multi-level mode predication
CN105823801B (en) A kind of electronic nose drift compensation method based on deepness belief network feature extraction
CN114818579A (en) Analog circuit fault diagnosis method based on one-dimensional convolution long-short term memory network
CN111738348B (en) Power data anomaly detection method and device
CN111079348B (en) Method and device for detecting slowly-varying signal
CN116400168A (en) Power grid fault diagnosis method and system based on depth feature clustering
CN116432359A (en) Variable topology network tide calculation method based on meta transfer learning
CN110262447A (en) A kind of ACS closed-loop system Fault Locating Method based on ANN

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