WO2022213600A1 - 基于lstm-cnn的核电站蒸发器涡流信号类型识别方法 - Google Patents

基于lstm-cnn的核电站蒸发器涡流信号类型识别方法 Download PDF

Info

Publication number
WO2022213600A1
WO2022213600A1 PCT/CN2021/129548 CN2021129548W WO2022213600A1 WO 2022213600 A1 WO2022213600 A1 WO 2022213600A1 CN 2021129548 W CN2021129548 W CN 2021129548W WO 2022213600 A1 WO2022213600 A1 WO 2022213600A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
lstm
eddy current
network
cnn
Prior art date
Application number
PCT/CN2021/129548
Other languages
English (en)
French (fr)
Inventor
张洋
孔玉莹
张军
唐博
林戈
丁伯愿
万象
杨乾飞
Original Assignee
中广核检测技术有限公司
苏州热工研究院有限公司
中国广核集团有限公司
中国广核电力股份有限公司
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 中广核检测技术有限公司, 苏州热工研究院有限公司, 中国广核集团有限公司, 中国广核电力股份有限公司 filed Critical 中广核检测技术有限公司
Priority to GB2308984.0A priority Critical patent/GB2616996A/en
Publication of WO2022213600A1 publication Critical patent/WO2022213600A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N27/00—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means
    • G01N27/72—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables
    • G01N27/82—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables for investigating the presence of flaws
    • G01N27/90—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables for investigating the presence of flaws using eddy currents
    • 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
    • G06N3/044—Recurrent networks, e.g. Hopfield 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
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • 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
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N27/00—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means
    • G01N27/72—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables
    • G01N27/82—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables for investigating the presence of flaws
    • G01N27/90—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables for investigating the presence of flaws using eddy currents
    • G01N27/9046—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables for investigating the presence of flaws using eddy currents by analysing electrical signals
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • G06F2218/04—Denoising
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
    • Y02E30/00—Energy generation of nuclear origin
    • Y02E30/30—Nuclear fission reactors

Definitions

  • the invention belongs to the field of nuclear power detection equipment, in particular to an LSTM-CNN-based identification method for eddy current signal types of evaporators in nuclear power plants.
  • In-service inspection of nuclear power plants usually uses eddy current testing technology (ET) to track and detect the degradation of heat transfer tubes, and the accurate and rapid analysis of eddy current data of a large number of steam generator heat transfer tubes has become one of the main tasks and difficulties.
  • ETD eddy current testing technology
  • personnel with relevant qualifications compare the multi-channel data of the signal, and finally give a definite conclusion.
  • the analysis work is repetitive and complex, and the analyst is easily fatigued and increases the possibility of human error.
  • the purpose of the present invention is to provide a method for identifying the eddy current signal type of the evaporator of the nuclear power plant based on LSTM-CNN, to intelligently identify the signal type represented by the time series data of the eddy current signal of the evaporator of the nuclear power plant, and to realize the application of the LSTM-CNN technology in the non-destructive testing. .
  • the present invention adopts the following technical scheme: a method for identifying the eddy current signal type of the evaporator of a nuclear power plant based on LSTM-CNN, which is characterized in that it comprises the following steps:
  • A. Collect the eddy current data of the heat transfer tube of the evaporator of the nuclear power plant
  • the eddy current data of the evaporator heat transfer tube includes N channels, and the data of each channel includes a horizontal component and a vertical component.
  • the calibration process refers to the time t.
  • the eddy current data of t1 is expressed as (ch 1x ,ch 1y ,ch 2x ,ch 2y ,...ch Nx ,ch Ny ), that is, the eddy current data when recording t1 is expressed as (ch 1x ,ch 1y ,ch 2x ,ch 2y ,...ch Nx ,ch Ny ), the eddy current data when recording t2 is expressed as (ch 1x' ,ch 1y' ,ch 2x' ,ch 2y' ,...ch Nx' ,ch Ny' ), and so on;
  • the calibrated data to construct a time series based on a time window.
  • the preferred parameter for the length of the time window is M sampling points, and the data of this time window can be expressed as [t 1 , t 2 ,..., t M ];
  • step D Perform differential processing on the time series data in step C, that is, the difference between the current moment data of each channel and the data of each channel at the previous moment, expressed as [t 2 -t 1 , t 3 -t 2 , ..., t M -T M-1 ], namely ⁇ t 1 , ⁇ t 2 , ..., ⁇ t M-1 ;
  • the input time series of LSTM network ⁇ t 1 , ⁇ t 2 ,..., ⁇ t M-1 the time series feature information is extracted by LSTM network;
  • the input time series of CNN network ⁇ t 1 , ⁇ t 2 ,..., ⁇ t M-1 The CNN network is composed of L convolutional layers, L maximum pooling layers, and fully connected layers; the output of the LSTM network and the CNN network are integrated, and the network parameters are optimized by using the principle of triple loss function, and the input signal is represented in the form of a vector. ;
  • step E Use step E to construct a defect signal feature database, and characterize it in the form of the vector, calculate the Euclidean distance between the vector feature of the input signal and all defect vector features in the database, and compare the Euclidean distances of all defects one by one to determine the category of the signal , and finally realize the classification of eddy current signals;
  • the LSTM network adopts a single-layer one-way architecture and is composed of multiple LSTM units.
  • Each of the LSTM units includes a forget gate, an input gate and an output gate.
  • the LSTM unit continuously updates and memorizes the information of the previous time series, and retains For valid information and forgetting useless information, the number of normal distribution Gaussian noise expansion data samples is added to the data, and the Gaussian noise expansion data is 5%-30% of the number of historical samples.
  • the present invention also provides an LSTM-CNN-based method for identifying the eddy current signal type of the evaporator of a nuclear power plant, which includes the following steps:
  • A. Collect the eddy current data of the heat transfer tube of the evaporator of the nuclear power plant
  • the eddy current data of the evaporator heat transfer tube includes N channels, and the data of each channel includes a horizontal component and a vertical component.
  • the calibration process refers to the time t.
  • the eddy current data of is represented as (ch1x,ch1y,ch2x,ch2y,...chNx,chNy), that is, the eddy current data at the time of recording t1 is represented as (ch 1x ,ch 1y ,ch 2x ,ch 2y ,...ch Nx ,ch Ny ),
  • the eddy current data when recording t2 is expressed as (ch 1x' ,ch 1y' ,ch 2x' ,ch 2y' ,...ch Nx' ,ch Ny' ), and so on;
  • the calibrated data to construct a time series based on a time window.
  • the preferred parameter for the length of the time window is M sampling points, and the data of this time window can be expressed as [t1, t2, ..., tM];
  • step C Perform differential processing on the time series data in step C, that is, the difference between the current moment data of each channel and the data of each channel at the previous moment, expressed as [t2-t1, t3-t2,..., tM- TM-1], namely ⁇ t1, ⁇ t2,..., ⁇ tM-1;
  • the input time series ⁇ t1, ⁇ t2,..., ⁇ tM-1 of the LSTM network the time series feature information is extracted by the LSTM network
  • the input time series ⁇ t1, ⁇ t2,..., ⁇ tM-1 of the CNN network the CNN network consists of L volumes It is composed of accumulation layer, L maximum pooling layer, and fully connected layer
  • the output of LSTM network and CNN network is integrated, and the network parameters are optimized by using the principle of triple loss function, so that the input signal can be represented in the form of a vector;
  • step E Use step E to construct a defect signal feature database, and characterize it in the form of the vector, calculate the Euclidean distance between the vector feature of the input signal and all defect vector features in the database, and compare the Euclidean distances of all defects one by one to determine the category of the signal , and finally achieve eddy current signal classification.
  • the LSTM network adopts a single-layer one-way architecture, which is composed of multiple LSTM units, and each of the LSTM units includes a forget gate, an input gate, and an output gate.
  • the LSTM unit continuously updates and memorizes the information of the previous time series, retains the effective information and forgets the useless information.
  • the Gaussian noise expansion data is 5%-30% of the number of historical samples.
  • a method for identifying the eddy current signal type of a nuclear power plant evaporator based on LSTM-CNN (long short-term memory network combined with convolutional neural network) provided by the present invention adopts a time series processing method, which can effectively filter out the differential processing of time series signals. Noise in the signal, extract signal feature information from the change trend of the signal;
  • An LSTM-CNN-based method for identifying the eddy current signal type of an evaporator of a nuclear power plant takes the eddy current signal of the evaporator heat transfer tube containing multiple channels as input, and the constructed LSTM-CNN network can extract multiple channels at the same time.
  • the characteristic information of the eddy current signal is realized to realize the purpose of identifying and classifying the eddy current signal;
  • a LSTM-CNN-based nuclear power plant evaporator eddy current signal type identification method provided by the present invention, by constructing an LSTM-CNN network, the LSTM network can be used to solve the signal classification problem from a time series perspective, and then extract the time series feature information, The local feature information in the time series feature information can be extracted by using the CNN network, the feature information of the LSTM and CNN networks can be combined, and the network parameters can be optimized by using the principle of triple loss function.
  • the input signal can represent its features in the form of vectors
  • a defect signal database By constructing a defect signal database and characterizing it in the form of the vector, comparing the vector feature of the input signal with the Euclidean distance of the vector feature in the defect database, and determining the category of the signal according to the Euclidean distance.
  • Fig. 1 is the network flow chart of the present invention
  • FIG. 2 is a diagram of a variation situation of training loss in the present invention.
  • LSTM Long Short-term Memory
  • CNN Convolutional Neural Networks
  • the identification method of eddy current signal type of evaporator of nuclear power plant based on LSTM-CNN includes the following steps:
  • A. Collect the eddy current data of the heat transfer tube of the evaporator of the nuclear power plant
  • the eddy current data of the evaporator heat transfer tube includes N channels, and the data of each channel includes a horizontal component and a vertical component.
  • the calibration process is to convert the time t
  • the eddy current data of t1 is expressed as (ch 1x ,ch 1y ,ch 2x ,ch 2y ,...ch Nx ,ch Ny ), that is, the eddy current data when recording t1 is expressed as (ch 1x ,ch 1y ,ch 2x ,ch 2y ,...ch Nx ,ch Ny ), the eddy current data when recording t2 is expressed as (ch 1x' ,ch 1y' ,ch 2x' ,ch 2y' ,...ch Nx' ,ch Ny' ), and so on, N can be extracted at the same time
  • the characteristic information of the channel realizes the purpose of identifying and classifying the eddy current signal;
  • the calibrated data to construct a time series based on a time window.
  • the preferred parameter for the length of the time window is M sampling points, and the data of this time window can be expressed as [t 1 , t 2 ,..., t M ];
  • step D Perform differential processing on the time series data in step C, that is, the difference between the current moment data of each channel and the data of each channel at the previous moment, expressed as [t 2 -t 1 , t 3 -t 2 , ..., t M -T M-1 ], namely ⁇ t 1 , ⁇ t 2 , ..., ⁇ t M-1 , can effectively filter out the noise in the signal, and extract the signal characteristic information from the change trend of the signal;
  • the LSTM network adopts a single-layer one-way architecture and consists of multiple LSTM units. Each LSTM unit contains a forget gate, an input gate and an output gate.
  • the input time series of the LSTM network ⁇ t 1 , ⁇ t 2 ,..., ⁇ t M- 1. Use the LSTM network to extract the time series feature information.
  • the LSTM unit continuously updates and memorizes the information of the previous time series, retains the effective information and forgets the useless information;
  • the input time series of the CNN network ⁇ t 1 , ⁇ t 2 ,..., ⁇ t M-1 the CNN network consists of L convolutional layers, L maximum pooling layers, and fully connected layers;
  • the output of the LSTM network and the CNN network is integrated, and the network parameters are optimized using the principle of triple loss function, so that the input signal is in the form of a vector After a large amount of data training and learning, the input signal can represent its characteristic information in the form of a vector;
  • step E Use step E to construct a defect signal feature database, and represent it in the form of a vector, calculate the Euclidean distance between the vector feature of the input signal and all defect vector features in the database, compare the Euclidean distances of all defects one by one to determine the category of the signal, and finally Implement eddy current signal classification.
  • the eddy current signal of the heat transfer tube of the evaporator of the nuclear power plant can be integrated into 10 channels, and the data of each channel includes the horizontal component and the vertical component; the preferred parameter of the time window length is 150 sampling points; Pooling layer 1, convolution layer 2, max pooling layer 2, convolution layer 3, max pooling layer 3, convolution layer 4, max pooling layer 4, fully connected layer.
  • the flow chart of the method for identifying the eddy current signal type of the nuclear power plant evaporator based on LSTM-CNN specifically includes the following steps:
  • A. Collect the eddy current data of the heat transfer tube of the evaporator of the nuclear power plant for a period of time
  • the eddy current data of the evaporator heat transfer tube includes N channels, and the data of each channel includes a horizontal component and a vertical component.
  • the calibration process is to convert the time t
  • the eddy current data of is represented as (ch 1x , ch 1y , ch 2x , ch 2y , ..., ch 10x , ch 10y ), that is, the eddy current data when recording t1 is represented as (ch 1x , ch 1y , ch 2x , ch 2y ,... ch Nx , ch Ny ), the eddy current data when recording t2 is expressed as (ch 1x' ,ch 1y' ,ch 2x' ,ch 2y' ,...ch Nx' ,ch Ny' ), and so on;
  • step C Perform differential processing on the time series data in step C, that is, the difference between the current moment data of each channel and the data of each channel at the previous moment can be expressed as [t 2 -t 1 , t 3 -t 2 , ..., t 150 -t 149 ], namely [ ⁇ t 1 , ⁇ t 2 , ..., ⁇ t 149 ], judging the signal type from the change trend of time series data can not only effectively filter out the noise in the signal, but also meet the requirements of manual analysis.
  • the signal trend identification method adopted, the data after time series difference processing can be regarded as a 149*20 matrix
  • steps A-D complete the calibration and differential processing of the historical data of the eddy current signal of the evaporator heat transfer tube of the nuclear power plant, and complete the manual labeling according to the category of the signal characteristics in the time window, and classify the signals in the data set according to requirements, including but not limited to Heat transfer tube NDD (free segment signal), TSP (support signal) and grinding traces (MBM) in the manufacturing stage, a 3-dimensional one-hot vector is used to indicate the category of each time window signal, and NDD is expressed as [1,0 ,0], TSP is expressed as [0,1,0], MBM is expressed as [0,0,1]; historical data is divided into training data and test data, the training data is used to train the model, and the test data is used to verify the reliability of the model
  • NDD free segment signal
  • TSP support signal
  • MBM grinding traces
  • the input of the LSTM network is the time series of [ ⁇ t 1 , ⁇ t 2 , ..., ⁇ t 149 ], and the time series feature information can be learned by using the LSTM network;
  • the input of the CNN network is [ ⁇ t 1 , ⁇ t 2 , ..., ⁇ t 149 ], the CNN network consists of convolution layer 1, max pooling layer 1, convolution layer 2, max pooling layer 2, convolution layer 3, max pooling layer 3, convolution layer 4, max pooling layer 4.
  • the LSTM network and the CNN network are integrated, and the network parameters are optimized by using the principle of triple loss function, which can generate a network that uses vector features to represent the characteristics of the input signal;
  • the LSTM network adopts a single-layer one-way architecture and consists of multiple LSTMs.
  • Each LSTM unit contains a forget gate, an input gate and an output gate.
  • the number of LSTMs is 149, and the output of the LSTM network is a 1*1*512 vector.
  • the LSTM unit can continuously update and memorize the information of the previous time series.
  • the forget gate determines how much of the unit state at the previous moment is retained to the current moment
  • the input gate determines how much of the network's input at the current moment is saved to the unit state
  • the output gate controls how much of the unit state is output to the current output value of the LSTM:
  • W f , W i , W c are weight matrices; b f , b t , b c are bias vectors; ⁇ represents the sigmoid function, whose value ranges from 0 to 1; tanh represents the hyperbolic tangent function, whose value range is -1 ⁇ 1; C t represents memory update, and its value is related to input information, forget gate information and memory C t-1 at the previous moment; h t is the output result of the hidden layer;
  • the input of the CNN network model is the output result of LSTM, and the preferred output matrix size is 149*149.
  • the maximum pooling layer 3, the convolution layer 4, the maximum pooling layer 4, the fully connected layer, and the triple loss function is used to train the network;
  • CNN network convolution layer 1 using linear rectification (Relu) activation function, the preferred convolution kernel size is 7*1, the preferred number of convolution kernels is 64, and the preferred step size is 1;
  • the maximum pooling layer 1 of the CNN network adopts the max-pooling method.
  • the preferred size of the pooling layer is 2*1, and the preferred step size is 2;
  • CNN network convolution layer 2 using Relu activation function, the preferred convolution kernel size is 7*1, the preferred number of convolution kernels is 64, and the preferred step size is 1;
  • the maximum pooling layer 2 of the CNN network adopts the common max-pooling method, the preferred size of the pooling layer is 2*1, and the preferred step size is 2;
  • CNN network convolution layer 3 using the common Relu activation function, the preferred convolution kernel size is 5*1, the preferred number of convolution kernels is 128, and the preferred step size is 1;
  • the maximum pooling layer 3 of the CNN network adopts the common max-pooling method.
  • the preferred size of the pooling layer is 2*1, and the preferred step size is 2;
  • the preferred convolution kernel size is 3*1, the preferred number of convolution kernels is 256, and the preferred step size is 1;
  • the maximum pooling layer 2 of the CNN network adopts the common max-pooling method, the preferred size of the pooling layer is 2*1, and the preferred step size is 2;
  • the fully connected layer of the CNN network the number of neurons is preferred, and the output of the CNN network is a vector of 1*1*512;
  • the LSTM network output and the CNN network use the full connection layer to fuse the output results, and the number of neurons is optimized, which can form an output vector of 1*1*128;
  • the triplet loss function is:
  • the manually labeled training data is fed into the network model in step E, one forward transfer and one reverse transfer process can complete the update of network parameters, that is, one training is completed.
  • the parameters of the network model can be obtained, and the network structure, weight and other information of the model can be saved as a file in a specified format.
  • the input signal can be represented as a 1*1*128 vector form using the network model;
  • defect signal that has been determined as the defect type, and characterize it as a 1*1*128 vector form to form a defect signal feature library; represent the new input signal as a 1*1*128 vector, and combine it with the defect feature library.
  • the Euclidean distances between the defect signals are compared one by one, and the signal category corresponding to the smallest Euclidean distance is regarded as the category to which the new input signal belongs;
  • the present invention introduces an index evaluation model of accuracy rate, call rate and equilibrium average (F1-Score) in the model testing process.
  • TP True Positive
  • TN True Negative
  • FP Fale Positive, false Positive
  • FN Fale Negative, false negative
  • the actual positive class sample is predicted as a negative class sample by the model.
  • the historical data of the eddy currents of the evaporator heat transfer tubes of different units of a nuclear power plant are used and divided into training data and test data.
  • the training data set has a total of 132,320 pieces, including 76,875 NDD data, 36,134 TSP, and 19,311 MBM data;
  • the test data set has a total of 144,676 pieces, of which NDD data is 82,319, TSP is 36981, MBM is 25376.
  • steps 5, 6, and 7, the network is trained for 100 times, and the loss changes during the training process are shown in Figure 2.
  • the signal recognition and classification effects of the trained model in the test data set are shown in Table 1.
  • the prediction accuracy of the model is 0.9889, and the classification and recognition accuracy of the model is high. It can be seen from Table 2 that the model has high precision, recall and F1-Score in each classification, and it has a good classification effect in NDD, TSP, and MBM signal recognition.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Chemical & Material Sciences (AREA)
  • Chemical Kinetics & Catalysis (AREA)
  • Electrochemistry (AREA)
  • Analytical Chemistry (AREA)
  • Biochemistry (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Monitoring And Testing Of Nuclear Reactors (AREA)
  • Testing And Monitoring For Control Systems (AREA)

Abstract

一种基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法，所述方法包括：标定涡流信号各通道数据；采用时间窗口处理标定后数据；采用差分方式处理时间序列数据；通过LSTM网络提取时间序列的时间特征信息；CNN网络可以提取时间序列局部特征信息；融合LSTM与CNN网络的特征信息，利用三元组损失原理，经大量数据训练学习后，输入信号可以向量形式表示其特征信息；构建缺陷信号特征数据库，并以所述向量形式表征，对比输入信号的向量特征与缺陷库内向量特征间的欧氏距离，根据欧氏距离大小确定信号所属类别，最终实现涡流信号分类的目的。

Description

基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法 技术领域
本发明属于核电检测设备领域，特别涉及一种基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法。
背景技术
核电站在役检查通常使用涡流检测技术(ET)对传热管降质现象进行跟踪检测，而对大量蒸汽发生器传热管涡流数据的准确快速分析成为主要工作及难点之一。传统的数据分析方法由具备相关资质的人员对信号进行多通道数据对比，最终给出确定结论。分析工作重复繁杂，分析人员极易疲劳增加人因失误的可能性。
发明内容
本发明的目的是提供一种基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法，对核电站蒸发器涡流信号时间序列数据表征的信号类型进行智能识别，实现LSTM-CNN技术在无损检测中的应用。
为解决上述技术问题，本发明采用如下技术方案：一种基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法，其特征在于，其包括以下步骤：
A.采集核电站蒸发器传热管涡流数据；
B.对采集到的核电站蒸发器传热管涡流数据进行标定处理，蒸发器传热管涡流数据包含N个通道，每个通道的数据包含水平分量以及垂直分量，所述标定处理指将t时刻的涡流数据表示为(ch 1x,ch 1y,ch 2x,ch 2y,…ch Nx,ch Ny)，即记录t1时的涡流数据表示为(ch 1x,ch 1y,ch 2x,ch 2y,…ch Nx,ch Ny)，记录t2时的涡流数据表示为(ch 1x’,ch 1y’,ch 2x’,ch 2y’,…ch Nx’,ch Ny’),以此类推；
C.利用标定后的数据构建基于时间窗口的时间序列，时间窗口长度优选的参数为M个采样点，该时间窗口的数据可以表示为[t 1,t 2,…，t M]；
D.对步骤C中的时间序列数据做差分处理，即每个通道的当前时刻数据与上一时刻的每个通道数据各自做差，表示为[t 2-t 1,t 3-t 2,…，t M-T M-1]，即δt 1,δt 2,…，δt M-1；
E.LSTM网络的输入时间序列δt 1,δt 2,…，δt M-1，利用LSTM网络提取 到时间序列特征信息；CNN网络的输入时间序列δt 1,δt 2,…，δt M-1，CNN网络由L个卷积层、和L个最大池化层、以及全连接层构成；融合LSTM网络与CNN网络的输出，利用三元组损失函数原理优化网络参数，实现输入信号以向量形式表征；
F.利用步骤E构建缺陷信号特征数据库，并以所述向量形式表征，计算输入信号的向量特征与数据库内所有缺陷向量特征间的欧氏距离，逐一比较所有缺陷欧氏距离大小确定信号所属类别，最终实现涡流信号分类；
所述LSTM网络采用单层单向架构，由多个LSTM单元组成，每个所述LSTM单元内包含遗忘门、输入门以及输出门,所述LSTM单元持续更新和记忆前期时间序列的信息，保留有效信息以及遗忘无用信息,在所述数据中加入正态分布高斯噪声扩展数据样本的数量,所述高斯噪声扩展数据为历史样本数量的5％-30％。
本发明还提供了一种基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法，其包括以下步骤：
A.采集核电站蒸发器传热管涡流数据；
B.对采集到的核电站蒸发器传热管涡流数据进行标定处理，蒸发器传热管涡流数据包含N个通道，每个通道的数据包含水平分量以及垂直分量，所述标定处理指将t时刻的涡流数据表示为(ch1x,ch1y,ch2x,ch2y,…chNx,chNy)，即记录t1时的涡流数据表示为(ch 1x,ch 1y,ch 2x,ch 2y,…ch Nx,ch Ny)，记录t2时的涡流数据表示为(ch 1x’,ch 1y’,ch 2x’,ch 2y’,…ch Nx’,ch Ny’),以此类推；
C.利用标定后的数据构建基于时间窗口的时间序列，时间窗口长度优选的参数为M个采样点，该时间窗口的数据可以表示为[t1,t2,…，tM]；
D.对步骤C中的时间序列数据做差分处理，即每个通道的当前时刻数据与上一时刻的每个通道数据各自做差，表示为[t2-t1,t3-t2,…，tM-TM-1]，即δt1,δt2,…，δtM-1；
E.LSTM网络的输入时间序列δt1,δt2,…，δtM-1，利用LSTM网络提取到时间序列特征信息；CNN网络的输入时间序列δt1,δt2,…，δtM-1，CNN网络由L个卷积层、和L个最大池化层、以及全连接层构成；融合LSTM网络与CNN网络的输出，利用三元组损失函数原理优化网络参数，实现输入信号以向量形式表征；
F.利用步骤E构建缺陷信号特征数据库，并以所述向量形式表征，计算输入信号的向量特征与数据库内所有缺陷向量特征间的欧氏距离，逐一比较所有 缺陷欧氏距离大小确定信号所属类别，最终实现涡流信号分类。
优化的，所述LSTM网络采用单层单向架构，由多个LSTM单元组成，每个所述LSTM单元内包含遗忘门、输入门以及输出门。
进一步的，所述LSTM单元持续更新和记忆前期时间序列的信息，保留有效信息以及遗忘无用信息。
优化的，在所述数据中加入正态分布高斯噪声扩展数据样本的数量。
进一步的，所述高斯噪声扩展数据为历史样本数量的5％-30％。
本发明的有益效果在于：
1.本发明提供的一种基于LSTM-CNN(长短期记忆网络结合卷积神经网络)的核电站蒸发器涡流信号类型识别方法，采用时间序列处理方式，对时间序列信号差分处理可以有效的滤除信号中的噪声，从信号的变化趋势提取信号特征信息；
2.本发明提供的一种基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法，将包含多个通道的蒸发器传热管涡流信号作为输入，构建的LSTM-CNN网络可以同时提取多个通道的特征信息，实现对涡流信号识别分类的目的；
3.本发明提供的一种基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法，通过构建LSTM-CNN网络，利用LSTM网络可以从时间序列角度解决信号分类问题，进而提取到时间序列特征信息，利用CNN网络可以提取时间序列特征信息中的局部特征信息，融合LSTM与CNN网络的特征信息，利用三元组损失函数原理优化网络参数，经大量数据训练学习后，输入信号可以向量形式表示其特征信息；通过构建缺陷信号数据库，并以所述向量形式表征，对比输入信号的向量特征与缺陷库内向量特征欧氏距离，根据欧氏距离大小确定信号所属类别。
附图说明
图1为本发明的网络流程图；
图2为本发明的一个训练损失变化情况图。
具体实施方式
下面结合附图所示的实施例对本发明作以下详细描述：
LSTM：LSTM(Long Short-term Memory)是指长短期记忆网络，其独特的设 计结构，处理时间序列数据时有明显的优势，CNN:(Convolutional Neural Networks)是指卷积神经网络，其仿造生物的视觉机制构建，具有表征学习能力。
基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法，其包括以下步骤：
A.采集核电站蒸发器传热管涡流数据；
B.对采集到的核电站蒸发器传热管涡流数据进行标定处理，蒸发器传热管涡流数据包含N个通道，每个通道的数据包含水平分量以及垂直分量，所述标定处理为将t时刻的涡流数据表示为(ch 1x,ch 1y,ch 2x,ch 2y,…ch Nx,ch Ny)，即记录t1时的涡流数据表示为(ch 1x,ch 1y,ch 2x,ch 2y,…ch Nx,ch Ny)，记录t2时的涡流数据表示为(ch 1x’,ch 1y’,ch 2x’,ch 2y’,…ch Nx’,ch Ny’),以此类推，可以同时提取N个通道的特征信息，实现对涡流信号识别分类的目的；
C.利用标定后的数据构建基于时间窗口的时间序列，时间窗口长度优选的参数为M个采样点，该时间窗口的数据可以表示为[t 1,t 2,…，t M]；
D.对步骤C中的时间序列数据做差分处理，即每个通道的当前时刻数据与上一时刻的每个通道数据各自做差，表示为[t 2-t 1,t 3-t 2,…，t M-T M-1]，即δt 1,δt 2,…，δt M-1，可以有效的滤除信号中的噪声，从信号的变化趋势提取信号特征信息；
E.LSTM网络采用单层单向架构，由多个LSTM单元组成，每个LSTM单元内包含遗忘门、输入门以及输出门，LSTM网络的输入时间序列δt 1,δt 2,…，δt M-1，利用LSTM网络提取到时间序列特征信息，LSTM单元持续更新和记忆前期时间序列的信息，保留有效信息以及遗忘无用信息；CNN网络的输入时间序列δt 1,δt 2,…，δt M-1，CNN网络由L个卷积层、和L个最大池化层、以及全连接层构成；融合LSTM网络与CNN网络的输出，利用三元组损失函数原理优化网络参数，实现输入信号以向量形式表征，经大量数据训练学习后，输入信号可以向量形式表示其特征信息；
F.利用步骤E构建缺陷信号特征数据库，并以向量形式表征，计算输入信号的向量特征与数据库内所有缺陷向量特征间的欧氏距离，逐一比较所有缺陷欧氏距离大小确定信号所属类别，最终实现涡流信号分类。
实施例一
设定核电站蒸发器传热管涡流信号可整合成10个通道，每个通道的数据包含水平分量以及垂直分量；时间窗口长度优选的参数为150个采样点；CNN网络由卷积层1、最大池化层1、卷积层2、最大池化层2、卷积层3、最大池化层3、 卷积层4、最大池化层4、全连接层构成。
则在本实施例中，基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法的流程图，具体包括以下步骤:
A.采集一段时间内的核电站蒸发器传热管涡流数据；
B.对采集到的核电站蒸发器传热管涡流数据进行标定处理，蒸发器传热管涡流数据包含N个通道，每个通道的数据包含水平分量以及垂直分量，所述标定处理为将t时刻的涡流数据表示为(ch 1x，ch 1y，ch 2x，ch 2y，…，ch 10x，ch 10y)，即记录t1时的涡流数据表示为(ch 1x,ch 1y,ch 2x,ch 2y,…ch Nx,ch Ny)，记录t2时的涡流数据表示为(ch 1x’,ch 1y’,ch 2x’,ch 2y’,…ch Nx’,ch Ny’),以此类推；
C.利用标定处理后的数据构建基于时间窗口的时间序列，该时间窗口的数据可以表示为[t 1，t 2，…t 150]；
D.对步骤C中的时间序列数据做差分处理，即每个通道的当前时刻数据与上一时刻的每个通道数据各自做差，可表示为[t 2-t 1，t 3-t 2，…，t 150-t 149]，即[Δt 1，Δt 2，…，Δt 149]，从时间序列数据变化趋势判断信号类型不仅可以有效的滤除信号中的噪声，同时也符合人工分析时采用的信号趋势识别方法，经时间序列差分处理后的数据可视为149*20的矩阵，
E.按照步骤A-D，将核电站蒸发器传热管涡流信号历史数据完成标定以及差分处理，并且根据时间窗口内信号特征所属类别完成人工标注，根据需求将数据集中的信号进行分类，包括但不限于传热管NDD(自由段信号)，TSP(支撑信号)以及制造阶段打磨痕迹(MBM)，采用一个3维的one-hot向量表示每一时间窗口信号所属类别，将NDD表示为[1,0,0]，TSP表示为[0,1,0]，MBM表示为[0,0,1]；历史数据分为训练数据以及测试数据，训练数据用来训练模型，测试数据用来验证模型可靠性，对于历史数据中样本数量少的类别，加入N(0,0.001)的正态分布高斯噪声扩展样本数量，其数量为历史样本数量的15％，以增加模型的鲁棒性；
F.LSTM网络的输入为[Δt 1，Δt 2，…，Δt 149]的时间序列，利用LSTM网络可以学习到时间序列特征信息；CNN网络的输入为[Δt 1，Δt 2，…，Δt 149]的时间序列，CNN网络由卷积层1、最大池化层1、卷积层2、最大池化层2、卷积层3、最大池化层3、卷积层4、最大池化层4、全连接层构成；融合LSTM网络与CNN网络，利 用三元组损失函数原理优化网络参数，可生成用向量特征表示输入信号特征的网络；LSTM网络采用单层单向架构，由多个LSTM单元组成，其中每个LSTM单元内包含遗忘门、输入门以及输出门，LSTM数量为149，LSTM网络的输出为1*1*512的向量，LSTM单元可以持续更新和记忆前期时间序列的信息。遗忘门决定了上一时刻的单元状态有多少保留到当前时刻，输入门它决定了当前时刻网络的输入有多少保存到单元状态，输出门控制单元状态有多少输出到LSTM的当前输出值：
遗忘门：f t＝σ(W f·[h t-1，x t]+b f)
输入门：i t＝σ(W i·[h t-1，x t]+b i)
记忆更新：
Figure PCTCN2021129548-appb-000001
Figure PCTCN2021129548-appb-000002
输出门：o t＝σ(W f·[h t-1，x t]+b o)
h t＝o t*tanh(C t)
W f，W i，W c为权重矩阵；b f，b t，b c为偏置向量；σ表示sigmoid函数，其取值范围0～1；tanh表示双曲正切函数，其取值范围为-1～1；C t表示记忆更新，其取值与输入信息、遗忘门信息以及上一时刻的记忆C t-1有关；h t为隐含层输出结果；
CNN网络模型的输入为LSTM的输出结果，优选的输出矩阵大小为149*149，由CNN网络卷积层1、最大池化层1、卷积层2、最大池化层2、卷积层3、最大池化层3、卷积层4、最大池化层4、全连接层构成，使用三元组损失函数训练网络；
CNN网络卷积层1，采用线性整流(Relu)激活函数，优选的卷积核大小为7*1，优选的卷积核的数量为64，优选的步长为1；
CNN网络最大池化层1，采用最大池化(max-pooling)方式，优选的池化层的尺寸大小为2*1，优选的步长为2；
CNN网络卷积层2，采用Relu激活函数，优选的卷积核大小为7*1，优选的卷积核的数量为64，优选的步长为1；
CNN网络最大池化层2，采用常见的max-pooling方式，优选的池化层的尺寸大小为2*1，优选的步长为2；
CNN网络卷积层3，采用常见的Relu激活函数，优选的卷积核大小为5*1，优选的卷积核的数量为128，优选的步长为1；
CNN网络最大池化层3，采用常见的max-pooling方式，优选的池化层的尺寸大小为2*1，优选的步长为2；
CNN网络卷积层4，采用常见的Relu激活函数，优选的卷积核大小为3*1，优选的卷积核的数量为256，优选的步长为1；
CNN网络最大池化层2，采用常见的max-pooling方式，优选的池化层的尺寸大小为2*1，优选的步长为2；
CNN网络全连接层，优选神经元数量，CNN网络的输出为1*1*512的向量；
LSTM网络输出与CNN网络采用全连接层融合输出结果，优选神经元数量，可构成1*1*128的输出向量；
三元组损失函数为：
Figure PCTCN2021129548-appb-000003
式中
Figure PCTCN2021129548-appb-000004
表示相同信号类别之间的欧氏距离、
Figure PCTCN2021129548-appb-000005
表示不同信号类别之间的欧氏距离、α表示相同信号类别之间的欧氏距离与不同信号类别之间的欧氏距离最小间隔；
将人工标注的训练数据喂入步骤E的网络模型，一次正向传递以及一次反向的传递过程，可完成网络参数的更新，即完成一次训练。若干次训练后，可得到网络模型的参数，并将模型的网络结构、权重等信息保存为指定格式文件。输入信号可利用网络模型表征为1*1*128的向量形式；
选取已被确定为缺陷类型的缺陷信号，并将其表征为1*1*128的向量形式，形成缺陷信号特征库；将新的输入信号以1*1*128向量表示，并与缺陷特征库之间的各缺陷信号间的欧氏距离逐一比对，欧氏距离最小的所对应的信号类别即认为新的输入信号所属类别；
为了评价步骤F所训练的模型效果，本发明在模型测试过程中引入精确率、召唤率以及均衡平均数(F1-Score)指标评价模型。TP(True Positive，真阳性)：实际为正类样本被模型预测为正类样本；TN(True Negative，真阴性)：实际为负类样本被模型预测为负类样本；FP(False Positive，假阳性)：实际为负类样本被模型预测为正类样本；FN(False Negative，假阴性)：实际为正类样本被模型预测为负类样本。
精确率：
Figure PCTCN2021129548-appb-000006
召唤率：
Figure PCTCN2021129548-appb-000007
F1-Score：
Figure PCTCN2021129548-appb-000008
在本实例中，采用了某核电站的不同机组的蒸发器传热管涡流历史数据，并将其分为训练数据以及测试数据。以其中NDD、TSP、MBM分类测试为例：训练数据集共计132320条，其中NDD数据为76875条、TSP为36134条、MBM为19311条；测试数据集共计144676条，其中NDD数据为82319条、TSP为36981条、MBM为25376条。按步骤5、步骤6、步骤7网络进行100次训练，训练过程损失变化情况如图2所示。训练所得模型在测试数据集中信号识别分类效果如表1所示，模型的预测精度为0.9889，模型的分类识别精度较高。从表2可以得到模型的各个分类上均有很高的精确率、召回率以及F1-Score，其在NDD、TSP、MBM信号识别上有很好的分类效果。
Figure PCTCN2021129548-appb-000009
表1本发明提供的涡流信号识别分类方法测试集识别结果
类别 Precision(％) Recall(％) F1-Score
NDD 99.96 99.88 0.9992
TSP 99.96 95.93 0.9790
MBM 94.20 100 0.9701
表2本发明提供的涡流信号识别分类方法性能对比
上述实施例只为说明本发明的技术构思及特点，其目的在于让熟悉此项技术的人士能够了解本发明的内容并据以实施，并不能以此限制本发明的保护范围。凡根据本发明精神所作的等效变化或修饰，都应涵盖在本发明的保护范围之内。

Claims (6)

  1. 一种基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法，其特征在于，其包括以下步骤：
    A.采集核电站蒸发器传热管涡流数据；
    B.对采集到的核电站蒸发器传热管涡流数据进行标定处理，蒸发器传热管涡流数据包含N个通道，每个通道的数据包含水平分量以及垂直分量，所述标定处理指将t时刻的涡流数据表示为(ch 1x,ch 1y,ch 2x,ch 2y,…ch Nx,ch Ny)，即记录t1时的涡流数据表示为(ch 1x,ch 1y,ch 2x,ch 2y,…ch Nx,ch Ny)，记录t2时的涡流数据表示为(ch 1x’,ch 1y’,ch 2x’,ch 2y’,…ch Nx’,ch Ny’),以此类推；
    C.利用标定后的数据构建基于时间窗口的时间序列，时间窗口长度优选的参数为M个采样点，该时间窗口的数据可以表示为[t 1,t 2,…，t M]；
    D.对步骤C中的时间序列数据做差分处理，即每个通道的当前时刻数据与上一时刻的每个通道数据各自做差，表示为[t 2-t 1,t 3-t 2,…，t M-T M-1]，即δt 1,δt 2,…，δt M-1；
    E.LSTM网络的输入时间序列δt 1,δt 2,…，δt M-1，利用LSTM网络提取到时间序列特征信息；CNN网络的输入时间序列δt 1,δt 2,…，δt M-1，CNN网络由L个卷积层、和L个最大池化层、以及全连接层构成；融合LSTM网络与CNN网络的输出，利用三元组损失函数原理优化网络参数，实现输入信号以向量形式表征；
    F.利用步骤E构建缺陷信号特征数据库，并以所述向量形式表征，计算输入信号的向量特征与数据库内所有缺陷向量特征间的欧氏距离，逐一比较所有缺陷欧氏距离大小确定信号所属类别，最终实现涡流信号分类；
    所述LSTM网络采用单层单向架构，由多个LSTM单元组成，每个所述LSTM单元内包含遗忘门、输入门以及输出门,所述LSTM单元持续更新和记忆前期时间序列的信息，保留有效信息以及遗忘无用信息,在所述数据中加入正态分布高斯噪声扩展数据样本的数量,所述高斯噪声扩展数据为历史样本数量的5％-30％。
  2. 一种基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法，其特征在于，其包括以下步骤：
    A.采集核电站蒸发器传热管涡流数据；
    B.对采集到的核电站蒸发器传热管涡流数据进行标定处理，蒸发器传热管涡流数据包含N个通道，每个通道的数据包含水平分量以及垂直分量，所述标定处理指将t时刻的涡流数据表示为(ch 1x,ch 1y,ch 2x,ch 2y,…ch Nx,ch Ny)，即记录 t1时的涡流数据表示为(ch 1x,ch 1y,ch 2x,ch 2y,…ch Nx,ch Ny)，记录t2时的涡流数据表示为(ch 1x’,ch 1y’,ch 2x’,ch 2y’,…ch Nx’,ch Ny’),以此类推；
    C.利用标定后的数据构建基于时间窗口的时间序列，时间窗口长度优选的参数为M个采样点，该时间窗口的数据可以表示为[t 1,t 2,…，t M]；
    D.对步骤C中的时间序列数据做差分处理，即每个通道的当前时刻数据与上一时刻的每个通道数据各自做差，表示为[t 2-t 1,t 3-t 2,…，t M-T M-1]，即δt 1,δt 2,…，δt M-1；
    E.LSTM网络的输入时间序列δt 1,δt 2,…，δt M-1，利用LSTM网络提取到时间序列特征信息；CNN网络的输入时间序列δt 1,δt 2,…，δt M-1，CNN网络由L个卷积层、和L个最大池化层、以及全连接层构成；融合LSTM网络与CNN网络的输出，利用三元组损失函数原理优化网络参数，实现输入信号以向量形式表征；
    F.利用步骤E构建缺陷信号特征数据库，并以所述向量形式表征，计算输入信号的向量特征与数据库内所有缺陷向量特征间的欧氏距离，逐一比较所有缺陷欧氏距离大小确定信号所属类别，最终实现涡流信号分类。
  3. 根据权利要求2所述的基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法，其特征在于：所述LSTM网络采用单层单向架构，由多个LSTM单元组成，每个所述LSTM单元内包含遗忘门、输入门以及输出门。
  4. 根据权利要求3所述的基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法，其特征在于：所述LSTM单元持续更新和记忆前期时间序列的信息，保留有效信息以及遗忘无用信息。
  5. 根据权利要求2所述的基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法，其特征在于：在所述数据中加入正态分布高斯噪声扩展数据样本的数量。
  6. 根据权利要求5所述的基于LSTM-CNN的核电站蒸发器涡流信号类型识别方法，其特征在于：所述高斯噪声扩展数据为历史样本数量的5％-30％。
PCT/CN2021/129548 2021-04-08 2021-11-09 基于lstm-cnn的核电站蒸发器涡流信号类型识别方法 WO2022213600A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
GB2308984.0A GB2616996A (en) 2021-04-08 2021-11-09 Method for recognizing type of vortex signal of evaporator of nuclear power plant on basis of LSTM-CNN

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110377204.7A CN113283288B (zh) 2021-04-08 2021-04-08 基于lstm-cnn的核电站蒸发器涡流信号类型识别方法
CN202110377204.7 2021-04-08

Publications (1)

Publication Number Publication Date
WO2022213600A1 true WO2022213600A1 (zh) 2022-10-13

Family

ID=77276404

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/129548 WO2022213600A1 (zh) 2021-04-08 2021-11-09 基于lstm-cnn的核电站蒸发器涡流信号类型识别方法

Country Status (3)

Country Link
CN (1) CN113283288B (zh)
GB (1) GB2616996A (zh)
WO (1) WO2022213600A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116187203A (zh) * 2023-04-20 2023-05-30 武汉大学 一种流域水质预测方法、系统、电子设备及存储介质
CN117522371A (zh) * 2023-09-27 2024-02-06 华能核能技术研究院有限公司 一种核电厂缺陷自动判定执行专业方法及系统

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113283288B (zh) * 2021-04-08 2023-08-18 中广核检测技术有限公司 基于lstm-cnn的核电站蒸发器涡流信号类型识别方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020048119A1 (en) * 2018-09-04 2020-03-12 Boe Technology Group Co., Ltd. Method and apparatus for training a convolutional neural network to detect defects
CN111274395A (zh) * 2020-01-19 2020-06-12 河海大学 基于卷积和长短期记忆网络的电网监控告警事件识别方法
US20200201313A1 (en) * 2018-12-21 2020-06-25 Canvass Analytics Predictive analysis of industrial processes
CN112232483A (zh) * 2020-10-15 2021-01-15 昆明理工大学 一种结合cnn和lstm的航班平均票价预测方法
CN112561165A (zh) * 2020-12-16 2021-03-26 南京航空航天大学 一种基于组合模型的多维时序数据预测方法
CN113283288A (zh) * 2021-04-08 2021-08-20 中广核检测技术有限公司 基于lstm-cnn的核电站蒸发器涡流信号类型识别方法

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109820525A (zh) * 2019-01-23 2019-05-31 五邑大学 一种基于cnn-lstm深度学习模型的驾驶疲劳识别方法
CN110209715B (zh) * 2019-05-31 2022-09-13 中广核检测技术有限公司 用于涡流检测数据分析的数据预分析方法
CN110806692A (zh) * 2019-10-21 2020-02-18 上海海事大学 一种基于cnn-latm组合模型的波浪补偿预测方法
CN111337234A (zh) * 2020-03-09 2020-06-26 西南交通大学 一种基于实时监测的tbm刮刀寿命预测系统及方法
CN111832417B (zh) * 2020-06-16 2023-09-15 杭州电子科技大学 基于cnn-lstm模型和迁移学习的信号调制样式识别方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020048119A1 (en) * 2018-09-04 2020-03-12 Boe Technology Group Co., Ltd. Method and apparatus for training a convolutional neural network to detect defects
US20200201313A1 (en) * 2018-12-21 2020-06-25 Canvass Analytics Predictive analysis of industrial processes
CN111274395A (zh) * 2020-01-19 2020-06-12 河海大学 基于卷积和长短期记忆网络的电网监控告警事件识别方法
CN112232483A (zh) * 2020-10-15 2021-01-15 昆明理工大学 一种结合cnn和lstm的航班平均票价预测方法
CN112561165A (zh) * 2020-12-16 2021-03-26 南京航空航天大学 一种基于组合模型的多维时序数据预测方法
CN113283288A (zh) * 2021-04-08 2021-08-20 中广核检测技术有限公司 基于lstm-cnn的核电站蒸发器涡流信号类型识别方法

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116187203A (zh) * 2023-04-20 2023-05-30 武汉大学 一种流域水质预测方法、系统、电子设备及存储介质
CN117522371A (zh) * 2023-09-27 2024-02-06 华能核能技术研究院有限公司 一种核电厂缺陷自动判定执行专业方法及系统

Also Published As

Publication number Publication date
CN113283288B (zh) 2023-08-18
CN113283288A (zh) 2021-08-20
GB2616996A (en) 2023-09-27

Similar Documents

Publication Publication Date Title
WO2022213600A1 (zh) 基于lstm-cnn的核电站蒸发器涡流信号类型识别方法
CN110232395B (zh) 一种基于故障中文文本的电力系统故障诊断方法
CN110135459B (zh) 一种基于双三元组深度度量学习网络的零样本分类方法
CN112756759B (zh) 点焊机器人工作站故障判定方法
CN111191726B (zh) 一种基于弱监督学习多层感知器的故障分类方法
CN113484738A (zh) 基于多特征信息融合的电路故障诊断方法
CN112508105A (zh) 一种采油机故障检测与检索方法
CN111079836A (zh) 基于伪标签方法和弱监督学习的过程数据故障分类方法
CN108875772A (zh) 一种基于堆叠稀疏高斯伯努利受限玻尔兹曼机和强化学习的故障分类模型及方法
CN109034483A (zh) 一种基于质量功能配置的检测规划方法
CN113642255A (zh) 基于多尺度卷积循环神经网络的光伏发电功率预测方法
CN114578011A (zh) 一种基于多传感器多源数据融合的水质监测方法
CN114330486A (zh) 基于改进Wasserstein GAN的电力系统不良数据辨识方法
CN115757103A (zh) 基于树结构的神经网络测试用例生成方法
CN116502163A (zh) 基于多特征融合和深度学习的振动监测数据异常检测方法
CN115096357A (zh) 一种基于ceemdan-pca-lstm的室内环境质量预测方法
Liu et al. Fault diagnosis of subway indoor air quality based on local fisher discriminant analysis
CN116521872B (zh) 一种认知和情感的联合识别方法、系统及电子设备
CN108920428B (zh) 一种基于联合模糊扩张原理的模糊距离判别方法
CN117078960A (zh) 一种基于图像特征提取的近红外光谱分析方法及系统
CN115452957B (zh) 基于注意力原型网络的小样本金属损伤识别方法
CN116738332A (zh) 一种结合注意力机制的飞行器多尺度信号分类识别与故障检测方法
CN111178383A (zh) 一种基于多核支持向量机的电池单体品质分类方法
CN115907178A (zh) 一种净生态系统co2交换量的预测方法
CN112906746B (zh) 一种基于结构方程模型的多源航迹融合评估方法

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: 21935817

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 202308984

Country of ref document: GB

Kind code of ref document: A

Free format text: PCT FILING DATE = 20211109

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 22.02.2024)