WO2021232655A1 - 一种基于振动特征的高压并联电抗器机械状态评估方法 - Google Patents

一种基于振动特征的高压并联电抗器机械状态评估方法 Download PDF

Info

Publication number
WO2021232655A1
WO2021232655A1 PCT/CN2020/118210 CN2020118210W WO2021232655A1 WO 2021232655 A1 WO2021232655 A1 WO 2021232655A1 CN 2020118210 W CN2020118210 W CN 2020118210W WO 2021232655 A1 WO2021232655 A1 WO 2021232655A1
Authority
WO
WIPO (PCT)
Prior art keywords
value
sequence
voltage shunt
information
shunt reactor
Prior art date
Application number
PCT/CN2020/118210
Other languages
English (en)
French (fr)
Inventor
高树国
孟令明
曾四鸣
范辉
臧谦
贾伯岩
汲胜昌
刘宏亮
孙路
邢超
赵军
Original Assignee
国网河北省电力有限公司电力科学研究院
国家电网有限公司
西安交通大学
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 国网河北省电力有限公司电力科学研究院, 国家电网有限公司, 西安交通大学 filed Critical 国网河北省电力有限公司电力科学研究院
Priority to US17/244,947 priority Critical patent/US20210364481A1/en
Publication of WO2021232655A1 publication Critical patent/WO2021232655A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M7/00—Vibration-testing of structures; Shock-testing of structures
    • G01M7/02—Vibration-testing by means of a shake table
    • G01M7/025—Measuring arrangements
    • 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
    • G01R31/12—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing
    • G01R31/1209—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing using acoustic measurements
    • 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
    • G01R31/12—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing
    • G01R31/1227—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing of components, parts or materials
    • G01R31/1263—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing of components, parts or materials of solid or fluid materials, e.g. insulation films, bulk material; of semiconductors or LV electronic components or parts; of cable, line or wire insulation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs

Definitions

  • the invention relates to the technical field of electrical equipment fault diagnosis, and in particular to a method for evaluating the mechanical state of a high-voltage shunt reactor based on vibration characteristics.
  • the high-voltage shunt reactor is an important reactive power compensation equipment in the power system, which plays an important role in the safe and stable operation of the power system, and its voltage level is above 500kV.
  • high-voltage shunt reactors in an abnormal state or malfunctioning nationwide, so it is of great significance to carry out fault diagnosis for high-voltage shunt reactors.
  • the main form of failure of the high-voltage shunt reactor is the loosening of internal fasteners, which generates a local suspension potential, which in turn generates abnormal partial discharges and increases acetylene, which ultimately leads to equipment shutdown.
  • the current common high-voltage shunt reactor state evaluation methods include oil chromatography, ultra-high frequency method, and ultrasonic method. These methods show high performance in the diagnosis of the late stage of the high-voltage shunt reactor failure (insulation defect). Accuracy, but it is difficult to diagnose its early mechanical failures in time.
  • the vibration method uses the vibration signal on the surface of the high-voltage shunt reactor to detect its mechanical state. It has the advantages of online, non-intrusive and sensitive to early mechanical failures.
  • Related scientific research institutions and enterprises have developed a wide range of high-voltage shunt reactors based on vibro-acoustic signals.
  • the online monitoring instrument of the high-voltage shunt reactor realizes the real-time observation and recording of the vibration signal on the surface of the oil tank of the high-voltage shunt reactor.
  • the surface vibration signals of the high-voltage shunt reactors of different voltage levels and different manufacturers are different.
  • Even the same type of high-voltage shunt reactors produced by the same manufacturer, due to the complexity of the production process control and the mechanical system of the high-voltage shunt reactors Its vibration signal will also be different, and for a particular high-voltage shunt reactor, at different stages of its operation, the vibration signal will also change with the increase of operating years and fluctuations of environmental factors.
  • the current diagnosis and evaluation methods are not sufficient.
  • the selection of the typical healthy vibration signal set and the typical fault vibration signal set and the establishment of the criterion are often based on statistics and experience from a small laboratory model experiment.
  • the results or other statistical results related to the reactor in operation have brought certain difficulties to the evaluation of the state of the high-voltage shunt reactor.
  • the technical problem to be solved by the present invention is to provide a method for evaluating the mechanical state of high-voltage shunt reactors based on vibration characteristics, which uses historical state data of high-voltage shunt reactors and real-time vibration and noise signal data, LSTM neural network time series prediction method, and comparison The deviation between the predicted characteristic value and the actual characteristic value, etc., realizes the evaluation of the mechanical condition of the high-voltage shunt reactor.
  • the technical solution adopted by the present invention is: a method for evaluating the mechanical state of high-voltage shunt reactors based on vibration characteristics, based on the historical state data of the high-voltage shunt reactors and real-time vibration and noise signal data, through LSTM neural network
  • the time series forecasting method compares the deviation between the predicted characteristic value and the actual characteristic value, and measures whether the high-voltage shunt reactor has mechanical defects or failures.
  • the further technical solution is to collect the vibration signal and noise signal on the surface of the high-voltage shunt reactor tank, extract the characteristic value of the signal and form a time series, combine the time series prediction model, calculate the comprehensive deviation measurement factor between the predicted value and the actual value, and pass This factor judgement measures whether mechanical defects or failures with unnatural trends occur inside the fuel tank of the high-voltage shunt reactor.
  • a further technical solution is to use the fundamental frequency amplitude of the vibration signal and the noise signal on the surface of the oil tank of the high-voltage shunt reactor as the characteristic variable.
  • a further technical solution is to use 30 minutes as a sampling period to collect vibration signals and noise signals on the surface of the oil tank of the high-voltage shunt reactor.
  • a further technical solution lies in: predicting the vibration characteristic value of the high-voltage shunt reactor for a period of time in the future through the LSTM neural network.
  • a further technical solution is to calculate a comprehensive deviation measurement factor between the predicted characteristic value and the actual characteristic value based on the prediction result of the vibration characteristic of the LSTM neural network, and use the comprehensive deviation measurement factor as a state evaluation indicator.
  • S1 high-voltage shunt reactor vibration signal and noise signal acquisition and data preprocessing
  • S2 establish LSTM neural network model and set parameters
  • S3 use LSTM neural network to predict signal fundamental frequency amplitude
  • S4 The step of evaluating the operating state of the high-voltage shunt reactor using the comprehensive deviation factor
  • the specific steps of S2 include: S201: determining the number of neurons in the input layer, hidden layer and output layer; S202: the step of constructing a prediction model;
  • the input layer neuron of the neural network is 480, the output layer neuron is 48, and the hidden layer neuron is 24;
  • the neural network model is constructed according to the form of multi-step training on the data. Each step has a time sequence relationship. The notable feature is that in each iteration, the output with current information is used as part of the input for the next time step.
  • a further technical solution is that: the specific steps of S1 include: S101: measuring point arrangement; S102: collecting equipment; S103: collecting period; S104: extracting features from the signal; S105: filtering the feature sequence;
  • the microphone of the fourth channel is fixed in front of the front of the high-voltage shunt reactor box to form four measuring points;
  • Each set of fundamental frequency amplitude forms a characteristic sequence, and the four sets of fundamental frequency amplitude form a total of four characteristic sequences.
  • the time series of the four channels are filtered, and the filtered result is used as the input of the LSTM neural network.
  • the combined filter is defined as follows:
  • Equation 1 f(s) is the fundamental frequency characteristic sequence before filtering for each channel, in the form of a vector. Each characteristic sequence is directly obtained by signal acquisition and fast Fourier transform; y(s) is the original characteristic sequence after the upper Obtained after the formula filtering, the form is a vector; OC is the open-closed morphological filtering method, and CO is the closed-open morphological filtering method.
  • S3 include: S301: input sequence; S302: determine the learning rate, number of iterations and error standard; S303: determine the current moment of forgetting gate; S304: determine the current moment of input gate; S305: confirm Candidate information at the current moment; S306: Determine the information to be retained in the candidate information at the current moment; S307: Determine the unit state at the current moment; S308: Determine the output gate at the current moment; S309: Predict the feature sequence,
  • S3 uses LSTM neural network to predict signal fundamental frequency amplitude
  • the forgetting gate in the time series forecasting model determines the forgotten information and the retained information after receiving the information transmitted from the previous unit state C t-1, and its output
  • f t is the forget gate, which is used to filter the information that needs to be retained and the information that needs to be forgotten in the information at the previous moment, the value;
  • x t is the input at the current moment in the feature sequence, the value;
  • h t-1 is the upper
  • W f is the weight value of the input x t at the current time, the value;
  • U f is the weight value of the output h t-1 at the last time, the value;
  • b f is the offset, the value of the calculation of the forgetting gate;
  • ⁇ is the sigmoid activation function, which is used to map variables between 0 and 1.
  • j is the independent variable of the sigmoid activation function, the value;
  • ⁇ (j) is the result of the independent variable j after the mapping, and its range is between 0 and 1, the value;
  • the input gate in the time series prediction model determines the information to be input after receiving the information of input x t at the current time and output h t-1 at the previous time, namely
  • i t is the input gate for filtering information and data information to be deleted in the current time information need to be retained;
  • b i is the bias value of the calculated input gate;
  • ⁇ is the sigmoid activation function, which is used to map the variable to between 0 and 1;
  • k is the independent variable of the tanh activation function
  • tanh(k) is the result of the independent variable k after being mapped, and its range is between -1 and 1;
  • Input gate i t and candidate information Multiply to obtain the retained information and store it in the unit state Ct of the input gate;
  • the state will be the information of the state of the candidate unit Combine with the information of the state C t-1 at the previous moment; the state of the storage unit at the current moment
  • C t is the value of the current state of the memory cell time; f t value is forgotten door; C t-1 is the state value storing unit on a time; i t is a numerical value input gate; Is the candidate status value at the current moment;
  • the output gate in the time series prediction model determines the information to be output after receiving the information of the input x t at the current time and the output h t-1 at the previous time, that is, the output gate
  • o t is the output gate value, which is used to filter the information that needs to be output from the information at the current moment ;
  • W o is the weight value of the input x t at the current moment;
  • U o is the weight value of the output h t-1 at the previous moment;
  • b o is the offset value of the calculated output gate;
  • ⁇ is the sigmoid activation function, which is used to map the variable between 0 and 1;
  • h t is the current time output value
  • C t is the value of the current state of the memory cell time
  • o t is output of gate value
  • tanh function is active, is used to map a variable between 0 to 1;
  • the network weight value is updated and steps S301-S308 are repeated; before the network reaches the preset number of iterations, if the h t in S308 is The error between t and the true value is less than the error threshold, or the network reaches the preset number of iterations, then the prediction sequence is calculated by the neural network,
  • F is the characteristic sequence calculated by the prediction model, in the form of a one-dimensional sequence; i is the value of the time corresponding to the sequence, and according to the collection period of the collected signal, time i is increased every 30 minutes; f i is the sequence of F The value corresponding to the feature at the i-th moment.
  • the step of evaluating the operating state of the high-voltage shunt reactor using the comprehensive deviation factor in S4 includes S401 calculating the comprehensive deviation factor prediction characteristic sequence, S402 judging whether the comprehensive deviation factor prediction characteristic sequence is greater than the threshold, and S403 warning or repetition.
  • M is the real feature sequence in the form of a one-dimensional sequence
  • i is the value of the time corresponding to the sequence, and according to the collection period of the collected signal, time i is increased every 30 minutes
  • mi is the i-th sequence in M The value corresponding to the feature at the moment; each value in the sequence M is obtained by filtering in step S105;
  • h is greater than the upper limit of the comprehensive deviation measurement factor h max , it indicates that the characteristics of the vibration signal deviate greatly from the ideal value.
  • the internal fastening components of the high-voltage shunt reactor have a certain defect and an alarm is issued.
  • the recommended value range of h max is [0.05, 0.15];
  • a method for evaluating the mechanical state of high-voltage shunt reactors based on vibration characteristics Based on the historical state data of high-voltage shunt reactors and real-time vibration and noise signal data, the LSTM neural network time series prediction method is used to compare the deviation of the predicted eigenvalues from the actual eigenvalues , To measure whether the high-voltage shunt reactor has mechanical defects or failures.
  • the LSTM neural network time series prediction method and the comparison of the deviation between the predicted eigenvalues and the actual eigenvalues, the mechanical state evaluation of the high-voltage shunt reactors is realized.
  • FIG. 1 is a flowchart of the present invention
  • FIG. 2 is a block diagram of the signal acquisition system used in the present invention.
  • Figure 3 is a structural diagram of the LSTM neural network in the present invention.
  • Figure 4 is a screenshot of the vibration signal of the first channel in the present invention.
  • Figure 5 is a screenshot of the second channel vibration signal in the present invention.
  • Figure 6 is a screenshot of the third channel vibration signal in the present invention.
  • Figure 7 is a screenshot of the fourth channel acoustic signal in the present invention.
  • FIG. 8 is a screenshot of the comparison between the predicted feature sequence of the first channel and the real feature sequence in the present invention.
  • FIG. 9 is a screenshot of the comparison between the predicted feature sequence of the second channel and the real feature sequence in the present invention.
  • FIG. 10 is a screenshot of the comparison between the predicted feature sequence of the third channel and the real feature sequence in the present invention.
  • Fig. 11 is a screenshot of the comparison between the predicted feature sequence of the fourth channel and the real feature sequence in the present invention.
  • the present invention discloses a method for evaluating the mechanical state of a high-voltage shunt reactor based on vibration characteristics, including: S1: high-voltage shunt reactor vibration signal and noise signal acquisition and data preprocessing; S2: establishment of LSTM Neural network model and set parameters; S3: use LSTM neural network to predict the amplitude of the fundamental frequency of the signal; S4: use the comprehensive deviation factor to evaluate the operation status of the high-voltage shunt reactor; the details are as follows:
  • the microphone of the fourth channel is fixed in front of the front of the high-voltage shunt reactor box to form four measuring points.
  • Each set of fundamental frequency amplitude forms a characteristic sequence, and the four sets of fundamental frequency amplitude form a total of four characteristic sequences.
  • the time series of the four channels are filtered, and the filtered result is used as the input of the LSTM neural network.
  • the combined filter is defined as follows:
  • Equation 1 f(s) is the fundamental frequency characteristic sequence before filtering for each channel, in the form of a vector. Each characteristic sequence is directly obtained by signal acquisition and fast Fourier transform; y(s) is the original characteristic sequence after the upper Obtained after the formula filtering, the form is a vector; OC is the open-closed morphological filtering method, and CO is the closed-open morphological filtering method.
  • the input layer neurons of the neural network are 480, the output layer neurons are 48, and the hidden layer neurons are 24.
  • the neural network model is constructed according to the form of multi-step training on the data. Each step has a time sequence relationship. The notable feature is that in each iteration, the output with current information is used as part of the input for the next time step.
  • S3 uses LSTM neural network to predict signal fundamental frequency amplitude
  • the learning rate of the neural network is 0.001
  • the number of iterations is 5000
  • the error standard is 0.00001.
  • the forgetting gate in the time series forecasting model determines the forgotten information and the retained information after receiving the information transmitted from the previous unit state C t-1, and its output
  • f t is the forget gate, which is used to filter the information that needs to be retained and the information that needs to be forgotten in the information at the previous moment, the value;
  • x t is the input at the current moment in the feature sequence, the value;
  • h t-1 is the upper
  • W f is the weight value of the input x t at the current time, the value;
  • U f is the weight value of the output h t-1 at the last time, the value;
  • b f is the offset, the value of the calculation of the forgetting gate;
  • ⁇ is the sigmoid activation function, which is used to map variables between 0 and 1.
  • the input gate in the time series prediction model determines the information to be input after receiving the information of input x t at the current time and output h t-1 at the previous time, namely
  • i t is the input gate for filtering information and the information to be deleted information of current time needed reserved value; W i is the current time input weighting values x t, the value, U i is the time output h t-1 weight values, values; b i is the calculation of the input gate bias value; sigmoid activation function [sigma] is used to map a variable between 0-1.
  • Equation 6 k is the independent variable of the tanh activation function; tanh(k) is the result of the independent variable k after being mapped, and its range is a value between -1 and 1.
  • Input gate i t and candidate information Multiply to get the retained information and store it in the cell state Ct of the input gate.
  • the state will be the information of the state of the candidate unit Combined with the information of the state C t-1 at the previous moment.
  • C t is the value of the current state of the memory cell time; f t value is forgotten door; C t-1 is the state value storing unit on a time; i t is a numerical value input gate; Is the candidate status value at the current moment.
  • the output gate in the time series prediction model determines the information to be output after receiving the information of the input x t at the current time and the output h t-1 at the previous time, that is, the output gate
  • o t is the output gate, which is used to filter the information value that needs to be output in the information at the current moment ;
  • W o is the weight value of the input x t at the current moment;
  • U o is the weight value of the output h t-1 at the previous moment;
  • b o is the offset value of the calculated output gate;
  • ⁇ is the sigmoid activation function, which is used to map the variable between 0 and 1.
  • h t is the current time output value
  • C t is the value of the current state of the memory cell time
  • o t is output gate value
  • tanh function is active, is used to map a variable between 0-1.
  • the network weight value is updated and steps S301-S308 are repeated; before the network reaches the preset number of iterations, if the h t in S308 is The error between t and the true value is less than the error threshold, or the network reaches the preset number of iterations, then the prediction sequence is calculated by the neural network,
  • F is the characteristic sequence calculated by the prediction model, in the form of a one-dimensional sequence; i is the time corresponding to the sequence, the value, according to the collection period of the collected signal, time i is increased every 30 min; f i is the sequence F in The feature and value of the i-th moment.
  • M is the real characteristic sequence in the time period to be studied, in the form of a one-dimensional sequence; i is the time corresponding to the sequence, the value, according to the collection period of the collected signal, time i is increased every 30 minutes; mi is the sequence M The feature and value of the i-th moment in. Each value in the sequence M is obtained by filtering in step S105.
  • h is greater than the upper limit of the comprehensive deviation measurement factor h max , it indicates that the characteristics of the vibration signal deviate greatly from the ideal value.
  • the internal fastening components of the high-voltage shunt reactor have a certain defect and an alarm is issued.
  • the recommended value range of h max is [0.05, 0.15].
  • the first to third vibration sensors are all piezoelectric vibration sensors
  • the capture card is a four-channel capture card
  • the piezoelectric vibration sensor, microphone, four-channel capture card, the computer itself, and the corresponding communication connection technology are existing technologies I won't repeat them here.
  • the method proposed in this scheme is mainly used to realize the operating state evaluation of high-voltage shunt reactors.
  • the current common high-voltage shunt reactor state evaluation methods include oil chromatography, ultra-high frequency method, and ultrasonic method. These methods show high accuracy in the diagnosis of the late stage of the high-voltage shunt reactor failure (insulation defect occurs). However, it is difficult to diagnose its early mechanical failures in time.
  • the vibration method uses the vibration signal on the surface of the high-voltage shunt reactor to detect its mechanical state. It has the advantages of online, non-intrusive and sensitive to early mechanical failures. In order to conduct comprehensive and continuous fault monitoring, it is necessary to use an online monitoring system for the high-voltage shunt reactor. Continuous vibration and noise collection on the surface of the oil tank of the engine.
  • the vibration method realizes the real-time observation and recording of the vibration signal on the surface of the high-voltage shunt reactor tank.
  • the surface vibration signals of the high-voltage shunt reactors of different voltage levels and different manufacturers are different.
  • Even the same type of high-voltage shunt reactors produced by the same manufacturer, due to the complexity of the production process control and the mechanical system of the high-voltage shunt reactors Its vibration signal will also be different, and for a particular high-voltage shunt reactor, at different stages of its operation, the vibration signal will also change with the increase of operating years and fluctuations of environmental factors.
  • the current diagnosis and evaluation methods are not sufficient.
  • the selection of the typical healthy vibration signal set and the typical fault vibration signal set and the establishment of the criterion are often based on statistics and experience from a small laboratory model experiment.
  • the results or other statistical results related to the reactor in operation have brought certain difficulties to the evaluation of the state of the high-voltage shunt reactor.
  • Aiming at the mechanical state assessment of high-voltage shunt reactors based on vibration signals invented a comprehensive utilization of the equipment body to obtain time-varying characteristic data, and the criterion is personalized and adaptive to the current operating environment and operating years of the high-voltage shunt reactor vibration characteristics
  • the evaluation method is of great significance.
  • the technical solution of the present application is suitable for performing sequence prediction fault diagnosis of characteristic values on the premise that a complete high-voltage shunt reactor time series data set is available.
  • the technical solution adopted by the present invention is to continuously collect vibration signals and noise signals from the surface of the high-voltage shunt reactor oil tank according to a fixed collection period, extract the characteristic values in the signals and form a time series, and combine the time series prediction model , Calculate the comprehensive deviation measurement factor between the predicted value and the actual value, and determine whether an unnatural trend mechanical defect or failure occurs inside the reactor tank by the size of the factor, thereby forming an evaluation method for the operating state of the high-voltage shunt reactor.
  • the first vibration sensor of the first channel is fixed at the center of the front of the high-voltage shunt reactor box with a height of 1.9m;
  • the second vibration sensor of the second channel is fixed at the center of the high-voltage shunt reactor box.
  • the center of the side the height is 1.9m;
  • the third vibration sensor of the third channel is arranged at the center of the back of the high-voltage shunt reactor box, and the height is 1.9m;
  • the microphone of the fourth channel is fixed at the center of the front of the high-voltage shunt reactor box In front of the location, it is at a horizontal distance of 1m from the front surface of the high-voltage shunt reactor box, and the height is 1.6m.
  • the first to third vibration sensors are piezoelectric vibration sensors.
  • the capture card is a four-channel capture card. Connect the first vibration sensor to the first channel of the capture card through a cable, and connect the second vibration sensor to the The second channel of the capture card, connect the third vibration sensor to the third channel of the capture card through a cable, connect the microphone to the fourth channel of the capture card through the cable, connect the capture card to the computer through the data cable, and the capture card
  • the collected data is viewed through a computer.
  • the computer is the PC segment, which is a notebook computer or a desktop computer.
  • the vibration signals and noise signals of the four measuring points are continuously collected all-weather synchronously.
  • each channel perform fast Fourier transform on the signals collected by the four channels, extract the fundamental frequency amplitude, and form four characteristic sequences from the four sets of fundamental frequency amplitudes.
  • the time series of the four channels are filtered, and the filtered result is used as the input of the LSTM neural network.
  • the combined filter is defined as follows:
  • f(s) is the fundamental frequency characteristic sequence before filtering for each channel, which has no unit and only represents a vector.
  • Each characteristic sequence is directly obtained through signal acquisition and fast Fourier transform;
  • y(s) is the original characteristic The sequence is obtained after the above-mentioned filtering. It has no unit and only represents a vector;
  • OC is an open-close morphological filtering method, and CO is a closed-open morphological filtering method.
  • the input layer neurons of the neural network are 480, the output layer neurons are 48, and the hidden layer neurons are 24.
  • the neural network model is constructed according to the form of multi-step training on the data.
  • Each step has a time series relationship.
  • the specific feature is that in each iteration, the output with current information is used as the next time step.
  • the hidden layer contains forget gates, input gates and output gates, that is, for the information at the previous moment, a part of the information is selectively deleted; a part of the input at the current moment is selectively left; the output result at the current moment is that of the previous moment The left part of the output and the input information at the current moment.
  • S3 uses LSTM neural network to predict signal fundamental frequency amplitude
  • the learning rate of the neural network is 0.001
  • the number of iterations is 5000
  • the error standard is 0.00001.
  • the forgetting gate in the time series forecasting model determines the forgotten information and the retained information after receiving the information transmitted from the previous unit state C t-1, and its output
  • f t is the forget gate, which is used to filter the information that needs to be retained and the information that needs to be forgotten in the information at the previous moment. It has no unit and the form is a value; x t is the input at the current moment in the feature sequence, without unit, The format is numeric; h t-1 is the output result at the previous moment, without unit, in the form of numeric value; W f is the weight value of the input x t at the current moment, without unit, in the form of numeric value; U f is the output h t at the previous moment The weight value of -1 , without unit, is in the form of a number; b f is the offset for calculating the forgetting gate, and the form is in the form of a number; ⁇ is the sigmoid activation function, which is used to map variables to between 0 and 1, the formula for:
  • Equation 3 j is the independent variable of the sigmoid activation function, without a unit, in the form of a numerical value; ⁇ (j) is the result of the mapping of the independent variable j, and its range is between 0 and 1, without a unit, in the form of a numerical value. Equation 3 specifically explains the calculation method of the sigmoid function in Equation 2.
  • the input gate in the time series prediction model determines the information to be input after receiving the information of input x t at the current time and output h t-1 at the previous time, namely
  • i t is the input gate for filtering information and the information to be deleted information of the current time in the need to retain, unit in the form of numerical value;
  • W i input weighting values x t is the present time, no unit, The form is a numerical value, U i is the weight value of the output h t-1 at the previous time, no unit, and the form is a numerical value;
  • b i is the bias of the calculated input gate, no unit, and the form is a numerical value;
  • ⁇ is the sigmoid activation function, which is used The variable is mapped between 0 and 1, and the explanation is the same as above.
  • Equation 6 specifically explains the calculation method of the tanh function in Equation 5. Both the sigmoid function and the tanh function can be used as activation functions, but the two forms are different, and the range of results obtained is different.
  • Input gate i t and candidate information The selected information is jointly decided and stored in the cell state Ct of the input gate.
  • the input gate i t is equivalent to a threshold as a screening criterion, while the candidate information It is some information that may be selected or eliminated, and the two are multiplied to get the retained information.
  • the state will be the information of the state of the candidate unit Combined with the information of the state C t-1 at the previous moment.
  • C t is the state of the storage unit at the current moment, no unit, in the form of a value
  • f t is the forget gate, no unit, in the form of a value, which is the calculation result in formula 2
  • the form is a numerical value, which is the calculation result in Equation 4; It is the candidate state at the current moment, no unit, in the form of a number.
  • the forget gate f t calculated by formula 2 determines the information that needs to be retained in the state C t-1 at the previous moment
  • the input gate i t calculated by formula 4 determines the candidate state at the current moment Information that needs to be retained.
  • the output gate in the time series prediction model determines the information to be output after receiving the information of the input x t at the current time and the output h t-1 at the previous time, that is, the output gate
  • o t is the output gate, which is used to filter the information that needs to be output in the information at the current moment, without unit, in the form of numerical value
  • W o is the weight value of input x t at the current moment, without unit, in the form of numerical value
  • U o is the weight value of the output h t-1 at the previous time, no unit, in the form of a value
  • b o is the offset of the calculated output gate, no unit, in the form of a value
  • ⁇ is the sigmoid activation function, which is used to map the variable to the Between 0 and 1, the explanation is the same as above.
  • the calculation process of the prediction model is based on the information at the previous time to calculate the information at the current time, and then use the information at the current time to calculate the information at the future time, so as to realize the feature sequence prediction.
  • the network weight value is updated and steps S301-S308 are repeated; before the network reaches the preset number of iterations, if the h t in S308 is The error between t and the true value is less than the error threshold, or the network reaches the preset number of iterations, then the prediction sequence is calculated by the neural network,
  • F is the characteristic sequence calculated by the prediction model, with no unit, in the form of a one-dimensional sequence; i is the time corresponding to the sequence, without unit, in the form of a value. According to the collection period of the collected signal, time i is increased every 30 minutes ; F i is the feature at the i-th moment in the sequence F, no unit, in the form of a number, and each value in the feature sequence is calculated by Equation 9.
  • Each value in the sequence M is filtered by S105.
  • h is the comprehensive deviation factor calculated from the predicted sequence ⁇ f 1 ,f 2 ,...f i ⁇ and the real sequence ⁇ m 1 ,m 2 ,...m i ⁇ , without unit, in the form of a number;
  • i is two Sequence subscript, no unit, in the form of a number, the same as in formula 10 and formula 11;
  • p is the length of the sequence, without unit, in the form of a number;
  • f i is the eigenvalue at a certain moment in the predicted sequence, without unit, in the form Numerical value is the same as f i in formula 10;
  • mi is the characteristic value at a certain moment in the actual sequence, without a unit, in the form of numerical value, which is the same as mi in formula 11.
  • h is greater than the upper limit of the comprehensive deviation measurement factor h max , it indicates that the characteristics of the vibration signal deviate greatly from the ideal value.
  • the internal fastening components of the high-voltage shunt reactor have a certain defect and an alarm is issued.
  • the recommended value range of h max is [0.05, 0.15].
  • the vibration method is used to collect the vibration signal on the surface of the high-voltage shunt reactor oil tank, which has the advantages of online, non-intrusive and sensitive to early mechanical failures. Based on the historical data and real-time measurement data of the four channels of the high-voltage shunt reactor under normal conditions, And by extracting the fundamental frequency amplitude of the vibration signal, it can sensitively reflect the internal mechanical state of the fuel tank.
  • the LSTM neural network time series prediction model can accurately predict the natural change trend of the vibration characteristics of the reactor for a period of time in the future. By comparing the prediction results with the actual data, it is found that the prediction results have high prediction accuracy and can show the future period of time.
  • the internal mechanical state has practical reference significance.
  • This method integrates the vibration sensor and the microphone on the same acquisition system, and synchronously collects the vibration signal and the noise signal. In order to obtain complete data, the system collects vibration and noise on the surface of the fuel tank every 30 minutes.
  • the vibration and noise signal is transformed by fast Fourier transform from the time domain signal to the frequency domain signal.
  • the fundamental frequency characteristics of the internal mechanical state are extracted from the signal and the LSTM neural network prediction model is used.
  • the fluctuation of the vibration signal and noise signal characteristic value on the surface of the fuel tank is closely related to the change of the internal mechanical state.
  • This method builds an LSTM network model to predict the characteristic value of vibration and noise in the normal state in the future. Larger mechanical defects, the actual vibration characteristics on the surface of the fuel tank will inevitably undergo abrupt changes, resulting in huge deviations from the predicted results.
  • the comprehensive deviation measurement factor is calculated, and the operating state of the reactor is evaluated by whether the measurement factor exceeds the threshold.
  • the vibration and noise data of the normal operation of the on-site high-voltage shunt reactor are collected.
  • the collection equipment runs around the clock. The data is collected every 30 minutes and stored in the cloud.
  • Vibration monitoring is performed on the surface of the high-voltage shunt reactor tank.
  • the three vibration acceleration measuring points are located at the center of the front and back of the tank near the windings and the center of the side near the side yoke.
  • the acoustic sensor is located at the center of the front of the tank.
  • the vibration and noise online monitoring system is used to monitor the vibration signal and noise signal of a 1000kV substation reactor in real time.
  • the fundamental frequency amplitude of vibration that can sensitively reflect the internal mechanical state of the high-voltage shunt reactor is selected as the characteristic quantity, and the fundamental frequency amplitude of the signal of each channel is formed into a complete timing sequence to predict the fundamental frequency amplitude for a period of time in the future.
  • the high-voltage shunt reactor vibration characteristic evaluation can clearly judge the mechanical state and characteristic trend by comparing the actual trend with the predicted trend.
  • This application collects the vibration and noise signals of the high-voltage shunt reactor through an online monitoring system, and selects the vibration fundamental frequency amplitude that can sensitively reflect the internal mechanical state of the high-voltage shunt reactor as the characteristic quantity.
  • This application forms a complete time series of data to predict the amplitude of the fundamental frequency of vibration, and proposes a high-voltage shunt reactor vibration characteristic evaluation based on the LSTM neural network prediction model. By comparing the actual trend and the predicted trend, it can be clearly judged Mechanical state and characteristic trend.

Landscapes

  • Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Acoustics & Sound (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种基于振动特征的高压并联电抗器机械状态评估方法，涉及电气设备故障诊断技术领域，基于高压并联电抗器的历史状态数据和实时振动噪声信号数据，通过LSTM神经网络时间序列预测方法，对比预测特征值与实际特征值的偏差，衡量高压并联电抗器是否出现机械缺陷或故障，包括：S1，高压并联电抗器振动信号和噪声信号采集及数据预处理；S2，建立LSTM神经网络模型并设置参数；S3，利用LSTM神经网络预测信号基频幅值；S4，对比预测特征值与实际特征值的偏差等，利用综合偏差因子在对高压并联电抗器机械状态运行进行评估。

Description

一种基于振动特征的高压并联电抗器机械状态评估方法 技术领域
本发明涉及电气设备故障诊断技术领域，尤其涉及一种基于振动特征的高压并联电抗器机械状态评估方法。
背景技术
高压并联电抗器是电力系统中重要的无功补偿设备，对电力系统的安全稳定运行起着重要的作用，其电压等级在500kV以上。近年来全国范围内有多台高压并联电抗器处于异常状态或发生故障，因此对高压并联电抗器开展故障诊断具有十分重要的意义。根据发生故障的高压并联电抗器解体报告，高压并联电抗器主要故障形式为内部紧固件松动，产生局部悬浮电位，进而产生异常的局部放电、乙炔增高，最终导致设备停运。
据相关资料显示，目前常见的高压并联电抗器状态评估方法有油色谱法、特高频法、超声法，这些方法在高压并联电抗器故障后期(出现绝缘缺陷)陷诊断方面表现出较高的准确性，但对于其早期机械故障难以及时诊断。振动法利用高压并联电抗器油箱表面的振动信号检测其机械状态，具有在线、非侵入和对早期机械故障反应灵敏的优点，相关科研机构和企业研制了种类繁多的基于振动声学信号的高压并联电抗器在线监测仪器，实现了高压并联电抗器油箱表面振动信号的实时观测和记录。但是不同电压等级、不同生产厂家的高压并联电抗器表面振动信号有所不同，即使是同一生产厂家生产的同一种类高压并联电抗器，由于生产工艺把控和高压并联电抗器本身机械系统的复杂性，其振动信号也会不同，并且对于某台特定的高压并联电抗器，在其运行的不同阶段，振动信号也会随运行年限增长和环境因素的波动而发生改变，目前的诊断评估方法未充分考虑被观测高压并联电抗器在不同时间维度下本身的运行特征，其典型健康振动信号集、典型故障振动信号集的选择和判据的确立往往基于自于一次实验室小模型实验的统计、经验结果或其它在运电抗器相关统计结果，给高压并联电抗器状态评估带来了一定困难。
因此针对基于振动信号的高压并联电抗器机械状态评估，发明一种综合利用设备本体获取随时间变化的特征数据，判据具有个性化和能自适应当前运行环境和运行年限的高压并联电抗器振动特征评估方法具有重要意义，该方法具有坚实的工程背景和广阔的应用前景。
现有技术问题及思考：
如何解决高压并联电抗器机械状态评估的技术问题。
发明内容
本发明所要解决的技术问题是提供一种基于振动特征的高压并联电抗器机械状态评估方法，其通过高压并联电抗器的历史状态数据和实时振动噪声信号数据、LSTM神经网络时间序列预测方法、对比预测特征值与实际特征值的偏差等，实现了高压并联电抗器机械状态评估。
为解决上述技术问题，本发明所采取的技术方案是：一种基于振动特征的高压并联电抗器机械状态评估方法，基于高压并联电抗器的历史状态数据和实时振动噪声信号数据，通过LSTM神经网络时间序列预测方法，对比预测特征值与实际特征值的偏差，衡量高压并联电抗器是或否出现机械缺陷或故障。
进一步的技术方案在于：采集高压并联电抗器油箱表面的振动信号和噪声信号，提取信号中的特征值并组成时间序列，结合时间序列预测模型，计算预测值与实际值的综合偏差衡量因子，通过该因子判断衡量高压并联电抗器的油箱内部是或否发生非自然趋势的机械缺陷或故障。
进一步的技术方案在于：将高压并联电抗器油箱表面的振动信号和噪声信号的基频幅值作为特征变量。
进一步的技术方案在于：将30分钟作为一个采样周期，采集高压并联电抗器油箱表面的振动信号和噪声信号。
进一步的技术方案在于：通过LSTM神经网络预测高压并联电抗器未来一段时间内的振动特征值。
进一步的技术方案在于：基于LSTM神经网络对振动特征的预测结果，计算预测特征值与实际特征值的综合偏差衡量因子，将综合偏差衡量因子作为状态评估指标。
进一步的技术方案在于：包括S1：高压并联电抗器振动信号和噪声信号采集及数据预处理；S2：建立LSTM神经网络 模型并设置参数；S3：利用LSTM神经网络预测信号基频幅值；S4：利用综合偏差因子对高压并联电抗器运行状态进行评估的步骤；所述S2的具体步骤包括：S201：确定输入层、隐层和输出层神经元个数；S202：构建预测模型的步骤；
S2建立LSTM神经网络模型并设置参数
S201确定输入层、隐层和输出层神经元个数
确定神经网络的输入层神经元为480，输出层神经元为48，隐层神经元为24；
S202构建预测模型
根据对数据进行多步长训练的形式来构建神经网络模型，每个步长具有时序关系，显著的特征在于，在每一次迭代中，具有当前信息的输出作为下一时间步的部分输入。
进一步的技术方案在于：所述S1的具体步骤包括：S101：测点布置；S102：采集设备；S103：采集周期；S104：从信号中提取特征；S105：对特征序列滤波的步骤；
S1高压并联电抗器振动信号和噪声信号采集及数据预处理
S101测点布置：
将第一通道的第一振动传感器固定在高压并联电抗器箱体的正面，第二通道的第二振动传感器固定在高压并联电抗器箱体的侧面，第三通道的第三振动传感器固定在高压并联电抗器箱体的背面，第四通道的传声器固定在高压并联电抗器箱体正面的前方，形成四个测点；
S102采集设备：
将第一振动传感器通过线缆连接至采集卡的第一通道，将第二振动传感器通过线缆连接至采集卡的第二通道，将第三振动传感器通过线缆连接至采集卡的第三通道，将传声器通过线缆连接至采集卡的第四通道，将采集卡通过数据线连接至电脑；
S103采集周期
以30min为采集周期，对三个测点的振动信号和一个测点的噪声信号进行采集；
S104从信号中提取特征
对四通道采集到的信号进行快速傅里叶变换，提取基频幅值，每一组基频幅值形成一个特征序列，四组基频幅值共形成四个特征序列，其中每一通道的序列形式为f(s)＝{f 1,f 2,…f s}；
S105对特征序列滤波
对四个通道的时间序列进行滤波处理，将滤波后的结果作为LSTM神经网络的输入，组合滤波器定义如下：
Figure PCTCN2020118210-appb-000001
式1中，f(s)为每个通道过滤前的基频特征序列，形式为向量，每个特征序列经过信号采集、快速傅里叶变换直接得到；y(s)为原始特征序列经过上式滤波后获得，形式为向量；OC为开-闭形态学滤波方式，CO为闭-开形态学滤波方式。
进一步的技术方案在于：所述S3的具体步骤包括：S301：输入序列；S302：确定学习率、迭代次数和误差标准；S303：确定当前时刻遗忘门；S304：确定当前时刻输入门；S305：确定当前时刻候选信息；S306：确定当前时刻候选信息中要保留下的信息；S307：确定当前时刻单元状态；S308：确定当前时刻输出门；S309：预测特征序列，
S3利用LSTM神经网络预测信号基频幅值
S301输入序列
将S103步骤中过滤后的序列作为输入；
S302确定学习率、迭代次数和误差标准
确定神经网络学习率为0.001，迭代次数为5000，误差标准为0.00001；
S303确定当前时刻遗忘门
进行预测的过程中，时间序列预测模型中的遗忘门在接收到上一个单元状态C t-1传送过来的信息后确定从中遗忘的信息和保留的信息，其输出
f t＝σ(W f·x t+U fh t-1+b f)           (2)
式2中，f t为遗忘门，用来筛选上一时刻的信息中需要保留的信息和需要遗忘的信息，数值；x t为特征序列中当前时刻的输入，数值；h t-1为上一时刻的输出结果，数值；W f为当前时刻输入x t的权重值，数值；U f为上时刻输出h t-1的权重值，数值；b f为计算遗忘门的偏置，数值；σ为sigmoid激活函数，用来将变量映射到在0到1之间，公式为：
Figure PCTCN2020118210-appb-000002
式3中，j为sigmoid激活函数的自变量，数值；σ(j)为自变量j经过映射后的结果，其范围在0到1之间，数值；
S304确定当前时刻输入门
时间序列预测模型中的输入门在接收到当前时刻输入x t和上一时刻输出h t-1的信息后确定要输入的信息，即
i t＝σ(W ix t+U ih t-1+b i)         (4)
式4中，i t为输入门，用来筛选当前时刻的信息中需要保留的信息和需要删除的信息数值；W i为当前时刻输入x t的权重值，U i为上时刻输出h t-1的权重值；b i为计算输入门的偏置数值；σ为sigmoid激活函数，用来将变量映射到在0到1之间；
S305确定当前时刻候选信息
Figure PCTCN2020118210-appb-000003
式5中，
Figure PCTCN2020118210-appb-000004
表示当前时刻的候选信息数值；
Figure PCTCN2020118210-appb-000005
包含着当前时刻输入x t和上一时刻输出h t-1的信息；W c为当前时刻输入x t的权重值，数值；U c为上时刻输出h t-1的权重值；b c为计算当前候选信息的偏置数值；tanh为激活函数，计算结果介值于-1到1之间，公式为：
Figure PCTCN2020118210-appb-000006
式6中，k为tanh激活函数的自变量；tanh(k)为自变量k经过映射后的结果，其范围在-1到1之间的数值；
S306确定当前时刻候选信息中要保留下的信息
输入门i t和候选信息
Figure PCTCN2020118210-appb-000007
相乘得到保留下来的信息并存储到输入门的单元状态Ct中；
S307确定当前时刻单元状态
更新当前时刻的存储单元状态，该状态将候选单元状态的信息
Figure PCTCN2020118210-appb-000008
与上一时刻状态C t-1的信息结合起来；当前时刻的存储单元状态
Figure PCTCN2020118210-appb-000009
式7中，C t为当前时刻的存储单元状态数值；f t为遗忘门数值；C t-1为上一时刻的存储单元状态数值；i t为输入门数值；
Figure PCTCN2020118210-appb-000010
为当前时刻的候选状态数值；
S308确定当前时刻输出门
时间序列预测模型中的输出门在接收到当前时刻输入x t和上一时刻输出h t-1的信息后确定要输出的信息，即输出门
o t＝σ(W ox t+U oh t-1+b o)           (8)
式8中，o t为输出门数值，用来筛选当前时刻的信息中需要输出的信息；W o为当前时刻输入x t的权重值；U o为上时刻输出h t-1的权重值；b o为计算输出门的偏置数值；σ为sigmoid激活函数，用来将变量映射到在0到1之间；
则当前时刻的输出为
h t＝o t*tanh(C t)             (9)
式9中，h t为当前时刻的输出数值；C t为当前时刻的存储单元状态数值；o t为输出门数值；tanh为激活函数，用来将变量映射到在0到1之间；
S309预测特征序列
网络在达到预设迭代次数之前，若S308中的h t与真实值误差不小于误差阈值，则更新网络权重值且重复步骤S301-S308；网络在达到预设迭代次数之前，若S308中的h t与真实值误差小于误差阈值，或者网络达到预设迭代次数，则通过神经网络计算得到预测序列，
F＝{f 1,f 2,…f i}            (10)
式10中，F为预测模型计算得到的特征序列，形式为一维序列；i为序列对应的时刻的数值，根据采集信号的采集周期，时刻i每30min增加一次；f i为序列F中的第i个时刻的特征对应的数值。
进一步的技术方案在于：所述S4利用综合偏差因子对高压并联电抗器运行状态进行评估的步骤包括S401计算综合偏差因子预测特征序列、S402判断综合偏差因子预测特征序列是否大于阈值和S403告警或重复以上步骤的步骤，
S4利用综合偏差因子对高压并联电抗器运行状态进行评估
S401计算综合偏差因子预测特征序列
M＝{m 1,m 2,…m i}             (11)
式11中，M为真实的特征序列，形式为一维序列；i为序列对应的时刻的数值，根据采集信号的采集周期，时刻i每30min增加一次；m i为序列M中的第i个时刻的特征对应的数值；序列M中的每个值都由S105步骤过滤获得；
将预测序列F i＝{f 1,f 2,…f i}与实际的待研究序列M i＝{m 1,m 2,…m i}进行对比，并计算综合偏差因子h，综合偏差衡量因子定义如下：
Figure PCTCN2020118210-appb-000011
式12中，h为预测序列{f 1,f 2,…f i}和真实序列{m 1,m 2,…m i}计算得到的综合偏差因子，数值；i为两序列下标，数值，与公式10和公式11中的一样；p为序列长度，数值；
S402判断综合偏差因子预测特征序列是否大于阈值
若h大于综合偏差衡量因子上限h max,说明振动信号的特征与理想值偏差较大，高压并联电抗器内部紧固部件产生了一定的缺陷，发出报警，建议h max取值范围为[0.05,0.15]；
S403告警或重复以上步骤
若综合偏差衡量因子h小于误差因子上限h max，继续采集数据，并基于当前数据执行S1的步骤。
采用上述技术方案所产生的有益效果在于：
一种基于振动特征的高压并联电抗器机械状态评估方法，基于高压并联电抗器的历史状态数据和实时振动噪声信号数据，通过LSTM神经网络时间序列预测方法，对比预测特征值与实际特征值的偏差，衡量高压并联电抗器是或否出现机械缺陷或故障。其通过高压并联电抗器的历史状态数据和实时振动噪声信号数据、LSTM神经网络时间序列预测方法、对比预测特征值与实际特征值的偏差等，实现了高压并联电抗器机械状态评估。
详见具体实施方式部分描述。
附图说明
图1是本发明的流程图；
图2是本发明使用的信号采集系统的构架图；
图3是本发明中LSTM神经网络的结构图；
图4是本发明中第一通道振动信号的屏幕截图；
图5是本发明中第二通道振动信号的屏幕截图；
图6是本发明中第三通道振动信号的屏幕截图；
图7是本发明中第四通道声学信号的屏幕截图；
图8是本发明中第一通道预测特征序列与真实特征序列对比的屏幕截图；
图9是本发明中第二通道预测特征序列与真实特征序列对比的屏幕截图；
图10是本发明中第三通道预测特征序列与真实特征序列对比的屏幕截图；
图11是本发明中第四通道预测特征序列与真实特征序列对比的屏幕截图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本申请一部分实施例，而不是全部的实施例。以下对至少一个示例性实施例的描述实际上仅仅是说明性的，决不作为对本申请及其应用或使用的任何限制。基于本申请中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
在下面的描述中阐述了很多具体细节以便于充分理解本申请，但是本申请还可以采用其他不同于在此描述的其它方式来实施，本领域技术人员可以在不违背本申请内涵的情况下做类似推广，因此本申请不受下面公开的具体实施例的限制。
如图1～图3所示，本发明公开了一种基于振动特征的高压并联电抗器机械状态评估方法包括：S1：高压并联电抗器振动信号和噪声信号采集及数据预处理；S2：建立LSTM神经网络模型并设置参数；S3：利用LSTM神经网络预测信号基频幅值；S4：利用综合偏差因子对高压并联电抗器运行状态进行评估的步骤；具体如下：
S1高压并联电抗器振动信号和噪声信号采集及数据预处理
S101测点布置
将第一通道的第一振动传感器固定在高压并联电抗器箱体的正面，第二通道的第二振动传感器固定在高压并联电抗器箱体的侧面，第三通道的第三振动传感器固定在高压并联电抗器箱体的背面，第四通道的传声器固定在高压并联电抗器箱体正面的前方，形成四个测点。
S102采集设备
将第一振动传感器通过线缆连接至采集卡的第一通道，将第二振动传感器通过线缆连接至采集卡的第二通道，将第三振动传感器通过线缆连接至采集卡的第三通道，将传声器通过线缆连接至采集卡的第四通道，将采集卡通过数据线连接至电脑。
S103采集周期
以30min为采集周期，对三个测点的振动信号和一个测点的噪声信号进行采集。
S104从信号中提取特征
对四通道采集到的信号进行快速傅里叶变换，提取基频幅值，每一组基频幅值形成一个特征序列，四组基频幅值共形成四个特征序列，其中每一通道的序列形式为f(s)＝{f 1,f 2,…f s}。
S105对特征序列滤波
对四个通道的时间序列进行滤波处理，将滤波后的结果作为LSTM神经网络的输入，组合滤波器定义如下：
Figure PCTCN2020118210-appb-000012
式1中，f(s)为每个通道过滤前的基频特征序列，形式为向量，每个特征序列经过信号采集、快速傅里叶变换直接得到；y(s)为原始特征序列经过上式滤波后获得，形式为向量；OC为开-闭形态学滤波方式，CO为闭-开形态学滤波方式。
S2建立LSTM神经网络模型并设置参数
S201确定输入层、隐层和输出层神经元个数
确定神经网络的输入层神经元为480，输出层神经元为48，隐层神经元为24。
S202构建预测模型
根据对数据进行多步长训练的形式来构建神经网络模型，每个步长具有时序关系，显著的特征在于，在每一次迭代中，具有当前信息的输出作为下一时间步的部分输入。
S3利用LSTM神经网络预测信号基频幅值
S301输入序列
将S103步骤中过滤后的序列作为输入。
S302确定学习率、迭代次数和误差标准
确定神经网络学习率为0.001，迭代次数为5000，误差标准为0.00001。
S303确定当前时刻遗忘门
进行预测的过程中，时间序列预测模型中的遗忘门在接收到上一个单元状态C t-1传送过来的信息后确定从中遗忘的信息和保留的信息，其输出
f t＝σ(W f·x t+U fh t-1+b f)          (2)
式2中，f t为遗忘门，用来筛选上一时刻的信息中需要保留的信息和需要遗忘的信息，数值；x t为特征序列中当前时刻的输入，数值；h t-1为上一时刻的输出结果，数值；W f为当前时刻输入x t的权重值，数值；U f为上时刻输出h t-1的权重值，数值；b f为计算遗忘门的偏置，数值；σ为sigmoid激活函数，用来将变量映射到在0到1之间，公式为：
Figure PCTCN2020118210-appb-000013
式3中，j为sigmoid激活函数的自变量，数值；σ(j)为自变量j经过映射后的结果，其范围在0到1之间，数值。
S304确定当前时刻输入门
时间序列预测模型中的输入门在接收到当前时刻输入x t和上一时刻输出h t-1的信息后确定要输入的信息，即
i t＝σ(W ix t+U ih t-1+b i)             (4)
式4中，i t为输入门，用来筛选当前时刻的信息中需要保留的信息和需要删除的信息，数值；W i为当前时刻输入x t的权重值，数值，U i为上时刻输出h t-1的权重值，数值；b i为计算输入门的偏置，数值；σ为sigmoid激活函数，用来将变量映射到在0到1之间。
S305确定当前时刻候选信息
Figure PCTCN2020118210-appb-000014
式5中，
Figure PCTCN2020118210-appb-000015
表示当前时刻的候选信息数值；
Figure PCTCN2020118210-appb-000016
包含着当前时刻输入x t和上一时刻输出h t-1的信息；W c为当前时刻输入x t的权重值；U c为上时刻输出h t-1的权重值；b c为计算当前候选信息的偏置数值；tanh为激活函数，计算结果介值于-1到1之间，公式为：
Figure PCTCN2020118210-appb-000017
式6中，k为tanh激活函数的自变量；tanh(k)为自变量k经过映射后的结果，其范围在-1到1之间的数值。
S306确定当前时刻候选信息中要保留下的信息
输入门i t和候选信息
Figure PCTCN2020118210-appb-000018
相乘得到保留下来的信息并存储到输入门的单元状态Ct中。
S307确定当前时刻单元状态
更新当前时刻的存储单元状态，该状态将候选单元状态的信息
Figure PCTCN2020118210-appb-000019
与上一时刻状态C t-1的信息结合起来。当前时刻的存储单元状态
Figure PCTCN2020118210-appb-000020
式7中，C t为当前时刻的存储单元状态数值；f t为遗忘门数值；C t-1为上一时刻的存储单元状态数值；i t为输入门数值；
Figure PCTCN2020118210-appb-000021
为当前时刻的候选状态数值。
S308确定当前时刻输出门
时间序列预测模型中的输出门在接收到当前时刻输入x t和上一时刻输出h t-1的信息后确定要输出的信息，即输出门
o t＝σ(W ox t+U oh t-1+b o)        (8)
式8中，o t为输出门，用来筛选当前时刻的信息中需要输出的信息数值；W o为当前时刻输入x t的权重值；U o为上时刻输出h t-1的权重值；b o为计算输出门的偏置数值；σ为sigmoid激活函数，用来将变量映射到在0到1之间。
则当前时刻的输出为
h t＝o t*tanh(C t)           (9)
式9中，h t为当前时刻的输出数值；C t为当前时刻的存储单元状态数值；o t为输出门，数值；tanh为激活函数，用来将变量映射到在0到1之间。
S309预测特征序列
网络在达到预设迭代次数之前，若S308中的h t与真实值误差不小于误差阈值，则更新网络权重值且重复步骤S301-S308；网络在达到预设迭代次数之前，若S308中的h t与真实值误差小于误差阈值，或者网络达到预设迭代次数，则通过神经网络计算得到预测序列，
F＝{f 1,f 2,…f i}               (10)
式10中，F为预测模型计算得到的特征序列，形式为一维序列；i为序列对应的时刻，数值，根据采集信号的采集周期，时刻i每30min增加一次；f i为序列F中的第i个时刻的特征，数值。
S4利用综合偏差因子对高压并联电抗器运行状态进行评估
S401计算综合偏差因子预测特征序列
M＝{m 1,m 2,…m i}             (11)
式11中，M为待研究时间段内真实的特征序列，形式为一维序列；i为序列对应的时刻，数值，根据采集信号的采集周期，时刻i每30min增加一次；m i为序列M中的第i个时刻的特征，数值。序列M中的每个值都由S105步骤过滤获得。
将预测序列F i＝{f 1,f 2,…f i}与实际的待研究序列M i＝{m 1,m 2,…m i}进行对比，并计算综合偏差因子h，综合偏差衡量因子定义如下：
Figure PCTCN2020118210-appb-000022
式12中，h为预测序列{f 1,f 2,…f i}和真实序列{m 1,m 2,…m i}计算得到的综合偏差因子，数值；i为两序列下标，数值，与公式10和公式11中的一样；p为序列长度，数值。
S402判断综合偏差因子预测特征序列是否大于阈值
若h大于综合偏差衡量因子上限h max,说明振动信号的特征与理想值偏差较大，高压并联电抗器内部紧固部件产生了一定的缺陷，发出报警，建议h max取值范围为[0.05,0.15]。
S403告警或重复以上步骤
若综合偏差衡量因子h小于误差因子上限h max，继续采集数据，并基于当前数据执行S1的步骤。
其中，第一至第三振动传感器均为压电式振动传感器，采集卡为四通道采集卡，压电式振动传感器、传声器、四通道采集卡和电脑本身以及相应的通信连接技术为现有技术在此不再赘述。
本申请的发明构思：
本方案提出的方法主要用于实现高压并联电抗器的运行状态评估。
1、目前常见的高压并联电抗器状态评估方法有油色谱法、特高频法、超声法，这些方法在高压并联电抗器故障后期(出现绝缘缺陷)陷诊断方面表现出较高的准确性，但对于其早期机械故障难以及时诊断。振动法利用高压并联电抗器油箱表面的振动信号检测其机械状态，具有在线、非侵入和对早期机械故障反应灵敏的优点，为了进行全面持续的故障监测，有必要采用在线监测系统对高压并联电抗器油箱表面进行持续的振动噪声采集。
2、振动法实现了高压并联电抗器油箱表面振动信号的实时观测和记录。但是不同电压等级、不同生产厂家的高压并联电抗器表面振动信号有所不同，即使是同一生产厂家生产的同一种类高压并联电抗器，由于生产工艺把控和高压并联电抗器本身机械系统的复杂性，其振动信号也会不同，并且对于某台特定的高压并联电抗器，在其运行的不同阶段，振动信号也会随运行年限增长和环境因素的波动而发生改变，目前的诊断评估方法未充分考虑被观测高压并联电抗器在不同时间维度下本身的运行特征，其典型健康振动信号集、典型故障振动信号集的选择和判据的确立往往基于自于一次实验室小模型实验的统计、经验结果或其它在运电抗器相关统计结果，给高压并联电抗器状态评估带来了一定困难。针对基于振动信号的高压并联电抗器机械状态评估，发明一种综合利用设备本体获取随时间变化的特征数据，判据具有个性化和能自适应当前运行环境和运行年限的高压并联电抗器振动特征评估方法具有重要意义。
因为本申请选用的在线监测系统可以全面地监测内部机械状态，并能将数据形成完整的时序序列，同时为了预测振动噪声信号特征值，提出基于LSTM神经网络的高压并联电抗器振动特征评估。目前未发现与本方案接近的技术方案。
本申请的目的：
本申请的技术方案适用于在具备完备高压并联电抗器时间序列数据集的前提下，对其进行特征值的序列预测故障诊断。
本申请的技术贡献：
高压并联电抗器在实际运行过程中，铁芯和绕组由于受到应力作用产生振动，电抗器内部机械状态与振动噪声信号特征联系紧密，因此，提前预测振动噪声信号的特征值可有效提前关注电抗器运行状态并及时采取措施。
为解决上述技术问题，本发明所采取的技术方案是：按照固定采集周期持续对高压并联电抗器油箱表面采集振动信号和噪声信号，提取信号中的特征值并组成时间序列，结合时间序列预测模型，计算预测值与实际值的综合偏差衡量因子，由该因子的大小判断电抗器油箱内部是否发生非自然趋势的机械缺陷或故障，从而形成高压并联电抗器运行状态的评估方法。
S1高压并联电抗器振动信号和噪声信号采集及数据预处理
S101测点布置
如图2所示，将第一通道的第一振动传感器固定在高压并联电抗器箱体的正面中心位置，高度为1.9m；第二通道的第二振动传感器固定在高压并联电抗器箱体的侧面中心位置，高度为1.9m；第三通道的第三振动传感器布置在高压并联电抗器箱体的背面中心位置，高度为1.9m；第四通道的传声器固定在高压并联电抗器箱体正面中心位置的前方，与高压并联电抗器箱体的正表面水平距离1m的位置，高度为1.6m。
S102采集设备
第一至第三振动传感器均为压电式振动传感器，采集卡为四通道采集卡，将第一振动传感器通过线缆连接至采集卡的第一通道，将第二振动传感器通过线缆连接至采集卡的第二通道，将第三振动传感器通过线缆连接至采集卡的第三通道，将传声器通过线缆连接至采集卡的第四通道，将采集卡通过数据线连接至电脑，采集卡采集到的数据通过电脑查看。电脑即PC段，为笔记本电脑或者台式电脑。
S103采集周期
以30min为采集周期，对四个测点的振动信号和噪声信号进行同步地全天候持续采集。
S104从信号中提取特征
分别对四通道采集到的信号进行快速傅里叶变换，提取基频幅值，并由四组基频幅值形成四个特征序列，其中各通道的序列形式为f(s)＝{f 1,f 2,…f s}。
S105对特征序列滤波
对四个通道的时间序列进行滤波处理，将滤波后的结果作为LSTM神经网络的输入，其中组合滤波器定义如下：
Figure PCTCN2020118210-appb-000023
式1中，f(s)为每个通道过滤前的基频特征序列，无单位，仅表示向量，每个特征序列经过信号采集、快速傅里叶变换直接得到；y(s)为原始特征序列经过上式滤波后获得，无单位，仅表示向量；OC为开-闭形态学滤波方式，CO为闭-开形态学滤波方式。
S2建立LSTM神经网络模型并设置参数
S201确定输入层、隐层和输出层神经元个数
确定神经网络的输入层神经元为480，输出层神经元为48，隐层神经元为24。
S202构建预测模型
如图3所示，根据对数据进行多步长训练的形式来构建神经网络模型，每个步长具有时序关系，具体的特征在于在每一次迭代中，具有当前信息的输出作为下一时间步的部分输入。隐藏层中包含遗忘门、输入门和输出门，即对于上一时刻的信息，选择性地删除一部分；对当前时刻的输入，选择性地留下一部分；当前时刻的输出结果为上一时刻的输出和当前 时刻的输入信息中被留下的部分。
S3利用LSTM神经网络预测信号基频幅值
S301输入序列
以其中一个通道为例，将S103步骤中过滤后的序列作为输入。
S302确定学习率、迭代次数和误差标准
确定神经网络学习率为0.001，迭代次数为5000，误差标准为0.00001。
S303确定当前时刻遗忘门
进行预测的过程中，时间序列预测模型中的遗忘门在接收到上一个单元状态C t-1传送过来的信息后确定从中遗忘的信息和保留的信息，其输出
f t＝σ(W f·x t+U fh t-1+b f)         (2)
式2中，f t为遗忘门，用来筛选上一时刻的信息中需要保留的信息和需要遗忘的信息，无单位，形式为数值；x t为特征序列中当前时刻的输入，无单位，形式为数值；h t-1为上一时刻的输出结果，无单位，形式为数值；W f为当前时刻输入x t的权重值，无单位，形式为数值；U f为上时刻输出h t-1的权重值，无单位，形式为数值；b f为计算遗忘门的偏置，无单位，形式为数值；σ为sigmoid激活函数，用来将变量映射到在0到1之间，公式为：
Figure PCTCN2020118210-appb-000024
式3中，j为sigmoid激活函数的自变量，无单位，形式为数值；σ(j)为自变量j经过映射后的结果，其范围在0到1之间，无单位，形式为数值。式3具体解释了式2中sigmoid函数的计算方式。
S304确定当前时刻输入门
时间序列预测模型中的输入门在接收到当前时刻输入x t和上一时刻输出h t-1的信息后确定要输入的信息，即
i t＝σ(W ix t+U ih t-1+b i)           (4)
式4中，i t为输入门，用来筛选当前时刻的信息中需要保留的信息和需要删除的信息，无单位，形式为数值；W i为当前时刻输入x t的权重值，无单位，形式为数值，U i为上时刻输出h t-1的权重值，无单位，形式为数值；b i为计算输入门的偏置，无单位，形式为数值；σ为sigmoid激活函数，用来将变量映射到在0到1之间，解释同上。
S305确定当前时刻候选信息
Figure PCTCN2020118210-appb-000025
式5中，
Figure PCTCN2020118210-appb-000026
表示当前时刻的候选信息，无单位，形式为数值；
Figure PCTCN2020118210-appb-000027
包含着当前时刻输入x t和上一时刻输出h t-1的 信息；W c为当前时刻输入x t的权重值，无单位，形式为数值；U c为上时刻输出h t-1的权重值，无单位，形式为数值；b c为计算当前候选信息的偏置，无单位，形式为数值；tanh为激活函数，计算结果介值于-1到1之间，公式为：
Figure PCTCN2020118210-appb-000028
式6中，k为tanh激活函数的自变量，无单位，形式为数值；tanh(k)为自变量k经过映射后的结果，其范围在-1到1之间，无单位，形式为数值。式6具体解释了式5中tanh函数的计算方式。sigmoid函数和tanh函数都可以做为激活函数，但两者形式不同，得到的结果范围不同。
S306确定当前时刻候选信息中要保留下的信息
输入门i t和候选信息
Figure PCTCN2020118210-appb-000029
共同决定选择的信息并存储到输入门的单元状态Ct中。输入门i t相当于一个门槛，作为筛选标准，而候选信息
Figure PCTCN2020118210-appb-000030
就是一些可能被选上也可能被淘汰的信息，两者相乘就得到了保留下来的信息。
S307确定当前时刻单元状态
更新当前时刻的存储单元状态，该状态将候选单元状态的信息
Figure PCTCN2020118210-appb-000031
与上一时刻状态C t-1的信息结合起来。当前时刻的存储单元状态
Figure PCTCN2020118210-appb-000032
式7中，C t为当前时刻的存储单元状态，无单位，形式为数值；f t为遗忘门，无单位，形式为数值，即为式2中的计算结果；C t-1为上一时刻的存储单元状态，无单位，形式为数值，最初时刻的存储单元状态只与最初时刻的输入x 1有关，即C 1＝tanh(W ix 1)；i t为输入门，无单位，形式为数值，即为式4中的计算结果；
Figure PCTCN2020118210-appb-000033
为当前时刻的候选状态，无单位，形式为数值。由式2计算得到的遗忘门f t决定了上一时刻状态C t-1需要保留的信息，由式4计算得到的输入门i t决定了当前时刻候选状态
Figure PCTCN2020118210-appb-000034
需要保留的信息。
S308确定当前时刻输出门
时间序列预测模型中的输出门在接收到当前时刻输入x t和上一时刻输出h t-1的信息后确定要输出的信息，即输出门
o t＝σ(W ox t+U oh t-1+b o)            (8)
式8中，o t为输出门，用来筛选当前时刻的信息中需要输出的信息，无单位，形式为数值；W o为当前时刻输入x t的权重值，无单位，形式为数值；U o为上时刻输出h t-1的权重值，无单位，形式为数值；b o为计算输出门的偏置，无 单位，形式为数值；σ为sigmoid激活函数，用来将变量映射到在0到1之间，解释同上。
则当前时刻的输出为
h t＝o t*tanh(C t)             (9)
式9中，h t为当前时刻的输出，无单位，形式为数值；C t为当前时刻的存储单元状态，无单位，形式为数值；o t为输出门，即为式8中的计算结果，无单位，形式为数值；tanh为激活函数，用来将变量映射到在0到1之间，解释同上；输出门o t决定了当前信息可以输出的部分。
针对上述过程的描述需解释的方面：预测模型的计算过程整体围绕上一时刻的信息来计算当前时刻的信息，再利用当前时刻信息，计算未来时刻的信息，从而实现特征序列预测。
S309预测特征序列
网络在达到预设迭代次数之前，若S308中的h t与真实值误差不小于误差阈值，则更新网络权重值且重复步骤S301-S308；网络在达到预设迭代次数之前，若S308中的h t与真实值误差小于误差阈值，或者网络达到预设迭代次数，则通过神经网络计算得到预测序列，
F＝{f 1,f 2,…f i}            (10)
式10中，F为预测模型计算得到的特征序列，无单位，形式为一维序列；i为序列对应的时刻，无单位，形式为数值，根据采集信号的采集周期，时刻i每30min增加一次；f i为序列F中的第i个时刻的特征，无单位，形式为数值，特征序列中的每个值都由式9计算获得。
S4利用综合偏差因子对高压并联电抗器运行状态进行评估
S401计算综合偏差因子预测特征序列
M＝{m 1,m 2,…m i}             (11)
式11中，M为待研究时间段内真实的特征序列，无单位，形式为一维序列；i为序列对应的时刻，无单位，形式为数值，根据采集信号的采集周期，时刻i每30min增加一次；m i为序列M中的第i个时刻的特征，无单位，形式为数值；式11中的i与式10中的i意义相同，代表时刻同步。序列M中的每个值都由S105过滤获得。
将预测序列F i＝{f 1,f 2,…f i}与实际的待研究序列M i＝{m 1,m 2,…m i}进行对比，并计算综合偏差因子h，综合偏差衡量因子定义如下：
Figure PCTCN2020118210-appb-000035
式12中，h为预测序列{f 1,f 2,…f i}和真实序列{m 1,m 2,…m i}计算得到的综合偏差因子，无单位，形式为数值；i为两序列下标，无单位，形式为数值，与公式10和公式11中的一样；p为序列长度，无单位，形式为数值；f i为预测序列中某时刻的特征值，无单位，形式为数值，与公式10中f i相同；m i为实际序列中某时刻的特征值，无单位， 形式为数值，与公式11中的m i相同。
S402判断综合偏差因子预测特征序列是否大于阈值
若h大于综合偏差衡量因子上限h max,说明振动信号的特征与理想值偏差较大，高压并联电抗器内部紧固部件产生了一定的缺陷，发出报警，建议h max取值范围为[0.05,0.15]。
S403告警或重复以上步骤
若综合偏差衡量因子h小于误差因子上限h max，继续采集数据，并基于当前数据重复以上步骤。
采用上述技术方案所产生的有益效果在于：
利用振动法对高压并联电抗器油箱表面的振动信号进行采集，具有在线、非侵入和对早期机械故障反应灵敏的优点，基于高压并联电抗器正常状态下四个通道的历史数据以及实时测量数据，并通过提取振动信号基频幅值，能够灵敏地反应油箱内部机械状态。
结合LSTM神经网络时间序列预测模型，可以准确预测出电抗器未来一段时间的振动特征的自然变化趋势，通过对比预测结果与实际数据，发现预测结果预测精确度较高，能够表现出未来一段时间的内部机械状态，具有实际的参考意义。
通过对未来一段时间的预测结果与实际特征值比较，并计算综合偏差衡量因子，可以通过设定阈值来判断电抗器内部是否出现机械方面的缺陷，从而形成了高压并联电抗器运行状态的评估方法。
技术方案分项说明：
该方法将振动传感器与传声器集于同一采集系统上，同步采集振动信号和噪声信号。为了获取完备的数据，该系统每30分钟对油箱表面的振动和噪声进行一次采集。
为了准确评估电抗器的运行状态，对振动噪声信号进行快速傅里叶变换，由时域信号转化为频域信号，提取信号中能体现内部机械状态的基频特征并采用LSTM神经网络预测模型。
油箱表面的振动信号和噪声信号特征值的波动，与内部机械状态的变化有紧密的联系，该方法通过构建LSTM网络模型，从而预测未来一天内正常状态下的振动噪声特征值，如果油箱内部发生较大的机械缺陷，则油箱表面的实际振动特征必然会发生突变，与预测结果产生巨大偏差，通过计算出综合偏差衡量因子，由该衡量因子是否超过阈值来完成电抗器运行状态的评估。
本申请的优点：
采集到现场高压并联电抗器正常运行的振动噪声数据，分别有三个通道的振动信号和一个通道的声学信号，采集设备全天候运行，每30分钟采集一次数据，并存储在云端。对高压并联电抗器油箱表面进行振动监测，三个振动加速度测点分别位于靠近绕组的油箱正面中心和背面中心以及靠近旁轭的侧面中心，声学传感器位于油箱正面中心。利用振动噪声在线监测系统对某1000kV变电站电抗器振动信号和噪声信号进行实时监测。选取能够灵敏地反应高压并联电抗器内部机械状态的振动基频幅值作为特征量，将各通道的信号基频幅值形成完整的时序序列，对未来一段时间的基频幅值进行预测，提出基于LSTM神经网络预测模型的高压并联电抗器振动特征评估，通过对比实际趋势和预测出来的趋势，能够明显判断机械状态和特征走向。
本申请保密运行一段时间后，现场技术人员反馈的有益之处在于：
1、本申请通过在线监测系统对高压并联电抗器振动噪声信号进行采集，选取能够灵敏地反应高压并联电抗器内部机械状态的振动基频幅值作为特征量。
2、本申请将数据形成完整的时序序列，对振动基频幅值进行预测，提出基于LSTM神经网络预测模型的高压并联电抗器振动特征评估，通过对比实际趋势和预测出来的趋势，能够明显判断机械状态和特征走向。

Claims (10)

  1. 一种基于振动特征的高压并联电抗器机械状态评估方法，其特征在于：基于高压并联电抗器的历史状态数据和实时振动噪声信号数据，通过LSTM神经网络时间序列预测方法，对比预测特征值与实际特征值的偏差，衡量高压并联电抗器是或否出现机械缺陷或故障。
  2. 根据权利要求1所述的一种基于振动特征的高压并联电抗器机械状态评估方法，其特征在于：采集高压并联电抗器油箱表面的振动信号和噪声信号，提取信号中的特征值并组成时间序列，结合时间序列预测模型，计算预测值与实际值的综合偏差衡量因子，通过该因子判断衡量高压并联电抗器的油箱内部是或否发生非自然趋势的机械缺陷或故障。
  3. 根据权利要求2所述的一种基于振动特征的高压并联电抗器机械状态评估方法，其特征在于：将高压并联电抗器油箱表面的振动信号和噪声信号的基频幅值作为特征变量。
  4. 根据权利要求2所述的一种基于振动特征的高压并联电抗器机械状态评估方法，其特征在于：将30分钟作为一个采样周期，采集高压并联电抗器油箱表面的振动信号和噪声信号。
  5. 根据权利要求1所述的一种基于振动特征的高压并联电抗器机械状态评估方法，其特征在于：通过LSTM神经网络预测高压并联电抗器未来一段时间内的振动特征值。
  6. 根据权利要求1所述的一种基于振动特征的高压并联电抗器机械状态评估方法，其特征在于：基于LSTM神经网络对振动特征的预测结果，计算预测特征值与实际特征值的综合偏差衡量因子，将综合偏差衡量因子作为状态评估指标。
  7. 根据权利要求1所述的一种基于振动特征的高压并联电抗器机械状态评估方法，其特征在于：包括以下步骤：步骤S1：高压并联电抗器振动信号和噪声信号采集及数据预处理；步骤S2：建立LSTM神经网络模型并设置参数；步骤S3：利用LSTM神经网络预测信号基频幅值；步骤S4：利用综合偏差因子对高压并联电抗器运行状态进行评估的步骤；所述步骤S2的具体实施步骤包括：步骤S201：确定输入层、隐层和输出层神经元个数；步骤S202：构建预测模型-；
    在步骤S201中，确定神经网络的输入层神经元为480，输出层神经元为48，隐层神经元为24；
    在步骤S202中，根据对数据进行多步长训练的形式来构建神经网络模型，每个步长具有时序关系，在每一次迭代中，具有当前信息的输出作为下一时间步的部分输入。
  8. 根据权利要求7所述的一种基于振动特征的高压并联电抗器机械状态评估方法，其特征在于：所述步骤S1的具体实施步骤包括：步骤S101：测点布置；步骤S102：采集设备；步骤S103采集周期；步骤S104：从信号中提取特征；步骤S105：对特征序列滤波；所述步骤S3的具体实施步骤包括：步骤S301：输入序列；步骤S302：确定学习率、迭代次数和误差标准；步骤S303：确定当前时刻遗忘门；步骤S304：确定当前时刻输入门；步骤S305：确定当前时刻候选信息；步骤S306：确定当前时刻候选信息中要保留下的信息；步骤S307：确定当前时刻单元状态；步骤S308：确定当前时刻输出门和步骤S309：预测特征序列；
    在步骤S301中，将S103步骤中过滤后的序列作为输入；
    在步骤S302中，确定神经网络学习率为0.001，迭代次数为5000，误差标准为0.00001；
    在步骤S303中，进行预测的过程中，时间序列预测模型中的遗忘门在接收到上一个单元状态C t-1传送过来的信息后确定从中遗忘的信息和保留的信息，其输出
    f t＝σ(W f·x t+U fh t-1+b f)  (2)
    式(2)中，f t为遗忘门，用来筛选上一时刻的信息中需要保留的信息和需要遗忘的信息数值；x t为特征序列中当前时刻的输入数值；h t-1为上一时刻的输出结果数值；W f为当前时刻输入x t的权重值；U f为上时刻输出h t-1的权重值；b f为计算遗忘门的偏置数值；σ为sigmoid激活函数，用来将变量映射到在0到1之间，公式为：
    Figure PCTCN2020118210-appb-100001
    式(3)中，j为sigmoid激活函数的自变量；σ(j)为自变量j经过映射后的结果，其范围在0到1之间数值；
    在步骤S304中，时间序列预测模型中的输入门在接收到当前时刻输入x t和上一时刻输出h t-1的信息后确定要输入的信息，即
    i t＝σ(W ix t+U ih t-1+b i)  (4)
    式(4)中，i t为输入门，用来筛选当前时刻的信息中需要保留的信息和需要删除的信息；W i为当前时刻输入x t的权重值，U i为上时刻输出h t-1的权重值；b i为计算输入门的偏置数值；σ为sigmoid激活函数，用来将变量映射到在0到1之间；
    在步骤S305中，当前时刻候选信息满足下列公式：
    Figure PCTCN2020118210-appb-100002
    式(5)中，
    Figure PCTCN2020118210-appb-100003
    表示当前时刻的候选信息数值；
    Figure PCTCN2020118210-appb-100004
    包含着当前时刻输入x t和上一时刻输出h t-1的信息；W c为当前时刻输入x t的权重值；U c为上时刻输出h t-1的权重值；b c为计算当前候选信息的偏置数值；tanh为激活函数，计算结果介值于-1到1之间，公式为：
    Figure PCTCN2020118210-appb-100005
    式(6)中，k为tanh激活函数的自变量；tanh(k)为自变量k经过映射后的结果，其范围在-1到1之间数值；
    在步骤S306中，输入门i t和候选信息
    Figure PCTCN2020118210-appb-100006
    相乘得到保留下来的信息并存储到输入门的单元状态Ct中；
    在步骤S307中，更新当前时刻的存储单元状态，该状态将候选单元状态的信息
    Figure PCTCN2020118210-appb-100007
    与上一时刻状态C t-1的信息结合起来；当前时刻的存储单元状态满足下列公式：
    Figure PCTCN2020118210-appb-100008
    式(7)中，C t为当前时刻的存储单元状态数值；f t为遗忘门数值；C t-1为上一时刻的存储单元状态数值；i t为输入门数值；
    Figure PCTCN2020118210-appb-100009
    为当前时刻的候选状态数值；
    在步骤S308中，时间序列预测模型中的输出门在接收到当前时刻输入x t和上一时刻输出h t-1的信息后确定要输出的信息，即输出门满足下列公式：
    o t＝σ(W ox t+U oh t-1+b o)   (8)
    式(8)中，o t为输出门数值，用来筛选当前时刻的信息中需要输出的信息；W o为当前时刻输入x t的权重值；U o为上时刻输出h t-1的权重值；b o为计算输出门的偏置数值；σ为sigmoid激活函数，用来将变量映射到在0到1之间；
    则当前时刻的输出为
    h t＝o t*tanh(C t)  (9)
    式(9)中，h t为当前时刻的输出，数值；C t为当前时刻的存储单元状态，数值；o t为输出门，数值；tanh为激活函数，用来将变量映射到在0到1之间；
    在步骤S309中，网络在达到预设迭代次数之前，若S308中的h t与真实值误差不小于误差阈值，则更新网络权重值且重复步骤S301-S308；网络在达到预设迭代次数之前，若S308中的h t与真实值误差小于误差阈值，或者网络达到预设迭代次数，则通过神经网络计算得到预测序列，计算公式如下：
    F＝{f 1,f 2,…f i}  (10)
    式(10)中，F为预测模型计算得到的特征序列，形式为一维序列；i为序列对应的时刻的数值，根据采集信号的采集周期，时刻i每30min增加一次；f i为序列F中的第i个时刻的特征数值。
  9. 根据权利要求7所述的一种基于振动特征的高压并联电抗器机械状态评估方法，其特征在于：所述步骤S1的具体实施步骤包括：步骤S101：测点布置；步骤S102：采集设备；步骤S103采集周期；步骤S104：从信号中提取特征；步骤S105：对特征序列滤波；所述步骤S4的具体实施步骤包括：步骤S401：计算综合偏差因子预测特征序列；步骤S402：判断综合偏差因子预测特征序列是否大于阈值；步骤S403：告警或重复以上步骤-；
    在步骤S401中，真实的特征序列定义如下：
    M＝{m 1,m 2,…m i}  (11)
    式(11)中，M为真实的特征序列，形式为一维序列；i为序列对应的时刻的数值，根据采集信号的采集周期，时刻i每30min增加一次；mi为序列M中的第i个时刻的特征，数值；序列M中的每个值都由S105步骤过滤获得；
    将预测序列F i＝{f 1,f 2,…f i}与实际的待研究序列M i＝{m 1,m 2,…m i}进行对比，并计算综合偏差因子h，综合偏差衡量因子定义如下：
    Figure PCTCN2020118210-appb-100010
    式(12)中，h为预测序列{f 1,f 2,…f i}和真实序列{m 1,m 2,…m i}计算得到的综合偏差因子，数值；i为 两序列下标，数值，与公式(10)和公式(11)中的一样；p为序列长度，数值；
    在步骤S402中，若h大于综合偏差衡量因子上限h max,说明振动信号的特征与理想值偏差较大，高压并联电抗器内部紧固部件产生了一定的缺陷，发出报警，建议h max取值范围为[0.05,0.15]；
    在步骤S403中，若综合偏差衡量因子h小于误差因子上限h max，继续采集数据，并基于当前数据执行S1的步骤。
  10. 根据权利要求7所述的一种基于振动特征的高压并联电抗器机械状态评估方法，其特征在于：所述步骤S1的具体实施步骤包括：步骤S101：测点布置；步骤S102：采集设备；步骤S103采集周期；步骤S104：从信号中提取特征；步骤S105：对特征序列滤波的步骤；
    在步骤S101中，将第一通道的第一振动传感器固定在高压并联电抗器箱体的正面，第二通道的第二振动传感器固定在高压并联电抗器箱体的侧面，第三通道的第三振动传感器固定在高压并联电抗器箱体的背面，第四通道的传声器固定在高压并联电抗器箱体正面的前方，形成四个测点；
    在步骤S102中，将第一振动传感器通过线缆连接至采集卡的第一通道，将第二振动传感器通过线缆连接至采集卡的第二通道，将第三振动传感器通过线缆连接至采集卡的第三通道，将传声器通过线缆连接至采集卡的第四通道，将采集卡通过数据线连接至电脑；
    在步骤S103中，以30min为采集周期，对三个测点的振动信号和一个测点的噪声信号进行采集；
    在步骤S104中，对四通道采集到的信号进行快速傅里叶变换，提取基频幅值，每一组基频幅值形成一个特征序列，四组基频幅值共形成四个特征序列，其中每一通道的序列形式为f(s)＝{f 1,f 2,…f s}；
    在步骤S105中，对四个通道的时间序列进行滤波处理，将滤波后的结果作为LSTM神经网络的输入，组合滤波器定义如下：
    Figure PCTCN2020118210-appb-100011
    式(1)中，f(s)为每个通道过滤前的基频特征序列，形式为向量，每个特征序列经过信号采集、快速傅里叶变换直接得到；y(s)为原始特征序列经过上式滤波后获得，形式为向量；OC为开-闭形态学滤波方式，CO为闭-开形态学滤波方式。
PCT/CN2020/118210 2020-05-20 2020-09-27 一种基于振动特征的高压并联电抗器机械状态评估方法 WO2021232655A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/244,947 US20210364481A1 (en) 2020-05-20 2021-04-29 Method for evaluating mechanical state of high-voltage shunt reactor based on vibration characteristics

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010432433.XA CN111638028B (zh) 2020-05-20 2020-05-20 一种基于振动特征的高压并联电抗器机械状态评估方法
CN202010432433.X 2020-05-20

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/244,947 Continuation-In-Part US20210364481A1 (en) 2020-05-20 2021-04-29 Method for evaluating mechanical state of high-voltage shunt reactor based on vibration characteristics

Publications (1)

Publication Number Publication Date
WO2021232655A1 true WO2021232655A1 (zh) 2021-11-25

Family

ID=72328939

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/118210 WO2021232655A1 (zh) 2020-05-20 2020-09-27 一种基于振动特征的高压并联电抗器机械状态评估方法

Country Status (2)

Country Link
CN (1) CN111638028B (zh)
WO (1) WO2021232655A1 (zh)

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114091984A (zh) * 2022-01-21 2022-02-25 北京智盟信通科技有限公司 一种电力变压器运行状态评估方法和设备
CN114564877A (zh) * 2021-12-29 2022-05-31 西安交通大学 一种滚动轴承寿命预测方法、系统、设备及可读存储介质
CN114819054A (zh) * 2022-03-11 2022-07-29 西北工业大学 一种基于物理信息神经网络的电力电子系统状态监测方法
CN114810513A (zh) * 2022-06-24 2022-07-29 江苏奥派电气科技有限公司 基于5g通信的风力发电机轴承振动故障智能监测系统
CN115289606A (zh) * 2022-07-13 2022-11-04 杭州安脉盛智能技术有限公司 除湿机在线故障诊断方法、系统、服务器及存储介质
CN115329639A (zh) * 2022-08-26 2022-11-11 中国长江三峡集团有限公司 一种风电站维护方法、平台和电子设备
CN115563475A (zh) * 2022-10-25 2023-01-03 南京工业大学 一种挖掘机液压系统的压力软传感器
CN116151169A (zh) * 2022-11-30 2023-05-23 西安西电变压器有限责任公司 预测电抗器噪声强度的方法和装置
CN116205375A (zh) * 2023-04-18 2023-06-02 北京市南水北调团城湖管理处 一种泵站机组运行状态预测方法和系统
CN116256054A (zh) * 2023-05-15 2023-06-13 广东电网有限责任公司阳江供电局 一种桥臂电抗器的故障监测方法、系统、设备和介质
CN116307190A (zh) * 2023-03-21 2023-06-23 信通院(江西)科技创新研究院有限公司 一种基于蓝牙mesh网络的果园环境对产量的预测方法
CN117349797A (zh) * 2023-12-04 2024-01-05 四川航空股份有限公司 基于人工智能的飞机故障检测方法及系统
CN117889943A (zh) * 2024-03-13 2024-04-16 浙江维度仪表有限公司 基于机器学习的气体超声波流量计检验方法及系统

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111638028B (zh) * 2020-05-20 2022-05-10 国网河北省电力有限公司电力科学研究院 一种基于振动特征的高压并联电抗器机械状态评估方法
CN112114215A (zh) * 2020-09-17 2020-12-22 陕西师范大学 基于误差反向传播算法的变压器老化评估的方法及系统
CN112098066A (zh) * 2020-09-21 2020-12-18 陕西师范大学 基于门控循环单元的高压并联电抗器故障诊断方法及系统
CN113033695B (zh) * 2021-04-12 2023-07-25 北京信息科技大学 一种电子器件故障的预测方法
CN113111564B (zh) * 2021-05-24 2023-05-23 扬州大学 基于自适应预测区间内置式芦苇收割机健康状态评估方法
CN113671361A (zh) * 2021-06-25 2021-11-19 海南电网有限责任公司海南输变电检修分公司 基于多源信号融合的高压断路器特性参数预测方法及系统
CN114924157B (zh) * 2022-06-16 2022-12-20 中国矿业大学(北京) 一种基于5g传输的并联电抗器状态监测方法及系统
CN115859835B (zh) * 2023-02-20 2023-06-20 西安华创马科智能控制系统有限公司 基于顶板振动波的来压预测方法及装置
CN117670260B (zh) * 2024-01-31 2024-04-30 成都三一能源环保技术有限公司 一种基于数据分析的井下钻井仪器使用周期管理系统

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101409986B1 (ko) * 2013-07-15 2014-06-20 시그널링크 주식회사 진동모니터링 결함진단장치
WO2015076435A1 (ko) * 2013-11-22 2015-05-28 이선휘 진동 특성을 이용한 설비의 결함 진단 방법
CN108303253A (zh) * 2017-12-06 2018-07-20 华南理工大学 基于长短时记忆循环神经网络的轴承早期故障识别方法
CN110175425A (zh) * 2019-05-31 2019-08-27 重庆大学 一种基于mmalstm的齿轮剩余寿命的预测方法
CN110210516A (zh) * 2019-04-25 2019-09-06 合刃科技（武汉）有限公司 基于深度学习的机床异常识别方法、系统及终端设备
CN110361207A (zh) * 2019-07-25 2019-10-22 中南大学 一种智能列车走行部在线状态预测系统及其方法
CN110534118A (zh) * 2019-07-29 2019-12-03 安徽继远软件有限公司 基于声纹识别和神经网络的变压器/电抗器故障诊断方法
CN110702418A (zh) * 2019-10-10 2020-01-17 山东超越数控电子股份有限公司 航空发动机故障预测方法
CN110956154A (zh) * 2019-12-11 2020-04-03 哈尔滨高斯触控科技有限公司 一种基于cnn-lstm的振动信息地形分类识别方法
CN111638028A (zh) * 2020-05-20 2020-09-08 国网河北省电力有限公司电力科学研究院 一种基于振动特征的高压并联电抗器机械状态评估方法

Family Cites Families (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5675087A (en) * 1994-02-15 1997-10-07 The Babcock & Wilcox Company Fastener characterization with an electromagnetic acoustic transducer
US5521840A (en) * 1994-04-07 1996-05-28 Westinghouse Electric Corporation Diagnostic system responsive to learned audio signatures
DE19528426C2 (de) * 1995-08-02 2003-04-30 Siemens Ag Verfahren zur Ermittlung von Zustandsparametern eines chemischen Reaktors unter Verwendung künstlicher neuronaler Netze
JP2002122571A (ja) * 2000-10-12 2002-04-26 Kenzo Miya 欠陥検査方法と欠陥検査装置
EP3343717A1 (en) * 2016-12-27 2018-07-04 Vito NV Hierarchical implicit controller for shielded system in a grid
CN108761228B (zh) * 2018-03-30 2020-05-19 大连理工大学 量测大数据下的电力变压器运行状态评估方法
CN108584592B (zh) * 2018-05-11 2019-10-11 浙江工业大学 一种基于时间序列预测模型的电梯轿厢振动异常预警方法
CN109033612B (zh) * 2018-07-20 2023-05-05 广西电网有限责任公司电力科学研究院 一种基于振动噪声及bp神经网络的变压器故障诊断方法
CN109813420A (zh) * 2019-01-18 2019-05-28 国网江苏省电力有限公司检修分公司 一种基于Fuzzy-ART的并联电抗器故障诊断方法
CN109932053B (zh) * 2019-03-19 2021-10-08 国网江苏省电力有限公司检修分公司 一种用于高压并联电抗器的状态监测装置及方法
CN110672323B (zh) * 2019-09-02 2021-03-26 佛山科学技术学院 一种基于神经网络的轴承健康状态评估方法及装置
CN110619182A (zh) * 2019-09-24 2019-12-27 长沙理工大学 一种基于wams大数据的输电线路参数辨识及输电网建模方法
CN110929769B (zh) * 2019-11-14 2023-02-10 国网吉林省电力有限公司超高压公司 一种基于振动和声音的电抗器机械类故障联合检测模型、方法及装置
CN111178612B (zh) * 2019-12-19 2023-06-09 绍兴大明电力设计院有限公司 一种网格用户基于大数据odps引擎的lstm负荷预测方法

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101409986B1 (ko) * 2013-07-15 2014-06-20 시그널링크 주식회사 진동모니터링 결함진단장치
WO2015076435A1 (ko) * 2013-11-22 2015-05-28 이선휘 진동 특성을 이용한 설비의 결함 진단 방법
CN108303253A (zh) * 2017-12-06 2018-07-20 华南理工大学 基于长短时记忆循环神经网络的轴承早期故障识别方法
CN110210516A (zh) * 2019-04-25 2019-09-06 合刃科技（武汉）有限公司 基于深度学习的机床异常识别方法、系统及终端设备
CN110175425A (zh) * 2019-05-31 2019-08-27 重庆大学 一种基于mmalstm的齿轮剩余寿命的预测方法
CN110361207A (zh) * 2019-07-25 2019-10-22 中南大学 一种智能列车走行部在线状态预测系统及其方法
CN110534118A (zh) * 2019-07-29 2019-12-03 安徽继远软件有限公司 基于声纹识别和神经网络的变压器/电抗器故障诊断方法
CN110702418A (zh) * 2019-10-10 2020-01-17 山东超越数控电子股份有限公司 航空发动机故障预测方法
CN110956154A (zh) * 2019-12-11 2020-04-03 哈尔滨高斯触控科技有限公司 一种基于cnn-lstm的振动信息地形分类识别方法
CN111638028A (zh) * 2020-05-20 2020-09-08 国网河北省电力有限公司电力科学研究院 一种基于振动特征的高压并联电抗器机械状态评估方法

Cited By (22)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114564877A (zh) * 2021-12-29 2022-05-31 西安交通大学 一种滚动轴承寿命预测方法、系统、设备及可读存储介质
CN114564877B (zh) * 2021-12-29 2024-05-10 西安交通大学 一种滚动轴承寿命预测方法、系统、设备及可读存储介质
CN114091984A (zh) * 2022-01-21 2022-02-25 北京智盟信通科技有限公司 一种电力变压器运行状态评估方法和设备
CN114819054A (zh) * 2022-03-11 2022-07-29 西北工业大学 一种基于物理信息神经网络的电力电子系统状态监测方法
CN114819054B (zh) * 2022-03-11 2024-02-27 西北工业大学 一种基于物理信息神经网络的电力电子系统状态监测方法
CN114810513A (zh) * 2022-06-24 2022-07-29 江苏奥派电气科技有限公司 基于5g通信的风力发电机轴承振动故障智能监测系统
CN115289606A (zh) * 2022-07-13 2022-11-04 杭州安脉盛智能技术有限公司 除湿机在线故障诊断方法、系统、服务器及存储介质
CN115329639B (zh) * 2022-08-26 2023-10-20 中国长江三峡集团有限公司 一种风电站维护方法、平台和电子设备
CN115329639A (zh) * 2022-08-26 2022-11-11 中国长江三峡集团有限公司 一种风电站维护方法、平台和电子设备
CN115563475A (zh) * 2022-10-25 2023-01-03 南京工业大学 一种挖掘机液压系统的压力软传感器
CN116151169A (zh) * 2022-11-30 2023-05-23 西安西电变压器有限责任公司 预测电抗器噪声强度的方法和装置
CN116151169B (zh) * 2022-11-30 2024-03-12 西安西电变压器有限责任公司 预测电抗器噪声强度的方法和装置
CN116307190A (zh) * 2023-03-21 2023-06-23 信通院(江西)科技创新研究院有限公司 一种基于蓝牙mesh网络的果园环境对产量的预测方法
CN116307190B (zh) * 2023-03-21 2024-03-26 信通院(江西)科技创新研究院有限公司 一种基于蓝牙mesh网络的果园环境对产量的预测方法
CN116205375B (zh) * 2023-04-18 2023-09-01 北京市南水北调团城湖管理处 一种泵站机组运行状态预测方法和系统
CN116205375A (zh) * 2023-04-18 2023-06-02 北京市南水北调团城湖管理处 一种泵站机组运行状态预测方法和系统
CN116256054B (zh) * 2023-05-15 2023-08-04 广东电网有限责任公司阳江供电局 一种桥臂电抗器的故障监测方法、系统、设备和介质
CN116256054A (zh) * 2023-05-15 2023-06-13 广东电网有限责任公司阳江供电局 一种桥臂电抗器的故障监测方法、系统、设备和介质
CN117349797A (zh) * 2023-12-04 2024-01-05 四川航空股份有限公司 基于人工智能的飞机故障检测方法及系统
CN117349797B (zh) * 2023-12-04 2024-02-06 四川航空股份有限公司 基于人工智能的飞机故障检测方法及系统
CN117889943A (zh) * 2024-03-13 2024-04-16 浙江维度仪表有限公司 基于机器学习的气体超声波流量计检验方法及系统
CN117889943B (zh) * 2024-03-13 2024-05-14 浙江维度仪表有限公司 基于机器学习的气体超声波流量计检验方法及系统

Also Published As

Publication number Publication date
CN111638028B (zh) 2022-05-10
CN111638028A (zh) 2020-09-08

Similar Documents

Publication Publication Date Title
WO2021232655A1 (zh) 一种基于振动特征的高压并联电抗器机械状态评估方法
US20210364481A1 (en) Method for evaluating mechanical state of high-voltage shunt reactor based on vibration characteristics
CN112763929B (zh) 一种储能电站系统电池单体健康预测方法及装置
Lindemann et al. Anomaly detection and prediction in discrete manufacturing based on cooperative LSTM networks
CN110533294B (zh) 一种基于人工智能技术的核电厂运行故障报警方法
Liu et al. Combined forecasting method of dissolved gases concentration and its application in condition-based maintenance
Qin et al. Remaining useful life prediction for rotating machinery based on optimal degradation indicator
Hu et al. A hybrid model for predicting the degradation trend of hydropower units based on deep learning
Kim et al. Digital twin approach for on-load tap changers using data-driven dynamic model updating and optimization-based operating condition estimation
CN115983084A (zh) 一种燃料电池剩余使用寿命预测方法、系统、设备及终端
CN112906764A (zh) 基于改进bp神经网络的通信安全设备智能诊断方法及系统
Barati Faster than real-time prediction of disruptions in power grids using PMU: Gated recurrent unit approach
CN116029699A (zh) 一种基于数据孪生的电力信息系统及运维方法
US20240044695A1 (en) Train compartment vibration monitoring method and vibration signal feature library establishment and application methods
CN117554821A (zh) 基于深度学习网络的储能电池在线监测管理方法及系统
CN117192369A (zh) 基于数字孪生技术的曳引电机监测与诊断方法
CN117371207A (zh) 一种特高压换流阀状态评价方法、介质及系统
Sun et al. Jointed Task of Multi-scale CNN Based Power Transformer Fault Diagnosis with Vibration and Sound Signals
Dang et al. seq2graph: Discovering dynamic non-linear dependencies from multivariate time series
CN115081304A (zh) 抽水蓄能机组轴系健康性能趋势区间预测的方法及装置
KR20230082887A (ko) 유지보수 이력에 기반한 학습데이터를 이용하는 장비의 고장 예측 장치 및 방법
CN112307918A (zh) 一种基于模糊神经网络的变压器直流偏磁的诊断方法
CN112098066A (zh) 基于门控循环单元的高压并联电抗器故障诊断方法及系统
Zheng et al. Research on Predicting Remaining Useful Life of Equipment Based on Health Index
CN117056814B (zh) 一种变压器声纹振动故障诊断方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20936774

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20936774

Country of ref document: EP

Kind code of ref document: A1