CN112033463B - Nuclear power equipment state evaluation and prediction integrated method and system - Google Patents
Nuclear power equipment state evaluation and prediction integrated method and system Download PDFInfo
- Publication number
- CN112033463B CN112033463B CN202010908744.9A CN202010908744A CN112033463B CN 112033463 B CN112033463 B CN 112033463B CN 202010908744 A CN202010908744 A CN 202010908744A CN 112033463 B CN112033463 B CN 112033463B
- Authority
- CN
- China
- Prior art keywords
- network model
- convolution network
- time convolution
- nuclear power
- symbol sequence
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01D—MEASURING NOT SPECIALLY ADAPTED FOR A SPECIFIC VARIABLE; ARRANGEMENTS FOR MEASURING TWO OR MORE VARIABLES NOT COVERED IN A SINGLE OTHER SUBCLASS; TARIFF METERING APPARATUS; MEASURING OR TESTING NOT OTHERWISE PROVIDED FOR
- G01D21/00—Measuring or testing not otherwise provided for
- G01D21/02—Measuring two or more variables by means not covered by a single other subclass
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (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)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Monitoring And Testing Of Nuclear Reactors (AREA)
Abstract
The invention relates to a nuclear power equipment state evaluation and prediction integrated method and system. According to the nuclear power equipment state evaluation and prediction integrated method and system, the cycle time convolution network model after structure optimization is adopted, and the prediction result of the service life of the nuclear power equipment can be accurately obtained according to the operation data of the nuclear power equipment. In addition, the nuclear power equipment state evaluation and prediction integration method and system provided by the invention can be used for forming the cycle time convolution network model by adopting small convolution kernel stacking, considering the time attribute of fault characteristics and flexibly adjusting the cycle time convolution network model so as to achieve the purpose of enriching the extracted local characteristics, thereby improving the detection accuracy and the interpretability while avoiding the problems of misjudgment and missed judgment.
Description
Technical Field
The invention relates to the field of nuclear power equipment detection, in particular to a method and a system for integrating state evaluation and prediction of nuclear power equipment.
Background
The nuclear power system has a complex structure, has potential radioactive release danger, and has extremely high requirements on safety. Thus, the reliability requirements for nuclear power systems are very high. Meanwhile, with the requirements of offshore drilling platforms, island power generation and the like, a large number of operators cannot be arranged on related platforms, so that the requirements on the automation and intelligentization level of the operation of the nuclear power plant are very high, and the requirements on unattended operation and unattended operation are strong. The nuclear power system has a severe operating environment, and key equipment of the system can work continuously for a long time, so that faults are very easy to occur, if the faults occur, the faults cannot be found and maintained in time, serious radioactive consequences can be caused, and the life safety of operating personnel and the public is critical. Therefore, the research on the fault diagnosis method of the nuclear power plant system and the equipment has important significance for improving the safety and the reliability of the nuclear power plant.
In the process of state evaluation and service life prediction of the nuclear power equipment, the general technical flow is shown in figure 1, the running state of the equipment is collected through a sensor, data is stored, and on the basis, characteristic engineering analysis, data processing and characteristic extraction are carried out. Then, in order to accurately evaluate the operation state of the nuclear power key equipment and predict the remaining service life of the nuclear power key equipment, the operation state of the nuclear power key equipment needs to be timely and accurately monitored, and once the relevant health indexes are detected to exceed the allowable range of normal operation, the remaining service life of the nuclear power key equipment needs to be predicted by using an intelligent algorithm. Among the key technologies involved in the overall process are mainly the state monitoring technology and the RUL prediction technology.
The state monitoring technology is a comprehensive technology of multidisciplinary cross penetration and is a technical foundation for implementing equipment maintenance and fault diagnosis. Since the early 70 s of the 20 th century, some developed countries have installed on-site condition monitoring systems, such as vibration health monitoring systems of mitsubishi corporation of japan and steam turbine monitoring systems of the central power generation division of the united states. At present, the research aiming at the monitoring of the running state at home and abroad is mainly divided into two types based on data analysis and model. The model-based method has the main advantages that an analysis model is established from the operation mechanism of the system, and the interpretability is strong. The main research results are as follows: william h. et al, who participated in the OECD Halden project, also used a mass and energy conservation model for condition monitoring. But has not been applied to nuclear power plants. The Harbin industry university establishes a dynamic mathematical model suitable for the steam turbine under the load shedding working condition, and simultaneously considers the nonlinearity caused to the unit due to strong disturbance.
The method based on data analysis is characterized by big data, and a data analysis model is constructed through a large amount of training of historical data, but the modeling process is relatively simple, and the universality and the real-time performance are better. Hines J.W. of Tennessee university adopts principal component analysis algorithm and auto-associative nuclear regression method to respectively monitor and correct the sensor in the nuclear power station in real time. Ajamia. an independent component analysis algorithm is applied to detect anomalies in a nuclear turbine. Tan Xiang studied the system level state monitoring method based on principal component analysis, through mutually supporting with sensor level monitoring, can accurately monitor out unusually fast.
And the Remaining Useful Life (RUL) refers to the length from the current time to the end of the Useful Life. And the main task of health state assessment and life prediction is to predict the remaining time of the machine before losing the operation capability based on the state monitoring information. At present, the main context of the residual life prediction research is shown in fig. 2, and the research methods can be roughly divided into 4 types, wherein the 1 st type is a multivariate statistical analysis method, and mainly researches on the RUL prediction technology are carried out by combining related theories such as reliability analysis and probability theory. Class 2 is the use of binding physical mechanisms to build analytical models. The 3 rd category is researched by adopting related algorithms of machine learning and deep learning, and essentially belongs to the category of pattern regression analysis. The 4 th class of mixed model is mainly analyzed by integrating 2 or more algorithms to play a role of mutually making up for deficiencies.
With the rapid development of artificial intelligence and big data technologies, it becomes popular to use machine learning and deep learning related technologies to learn aging and degradation patterns of elements from existing observation history data without establishing complex physical models.
The shallow artificial neural network was proposed and developed rapidly as early as the last 60 s; among them, the feedforward neural network is the most commonly used artificial neural network. Wang et al predict the development trend of the health status indicators using a three-layer feedforward neural network and input the results into a proportional risk model to estimate the hazard rate. The shallow neural network can learn a more complex nonlinear relationship, but cannot accurately describe the time sequence process of element degradation, and the accuracy is poor. A Support Vector machine (SVR) is provided based on a statistical learning theory and a structural risk minimization principle, and can realize minimization of experience risk and a confidence range according to limited data information. Liu et al developed an improved probabilistic SVR model to predict the degradation process of critical components of a nuclear power plant. However, SVR also has some limitations: firstly, it can only provide point prediction, even cannot give accuracy as an artificial neural network; second, the performance of the SVR is highly dependent on the selection and optimization of the hyperparameters.
The deep neural network has stronger pattern recognition capability than the shallow neural network, and the analysis accuracy is obviously higher under the condition of enough data quantity. Currently, the common deep Neural networks include self-encoders, convolutional Neural networks (convolutional Neural networks, RNNs), and their variants. The self-encoder is usually used for feature extraction and manifold learning of data, the convolutional neural network is more used in the fields of image recognition, video tracking and the like, and RNNs can effectively memorize historical information and therefore have the capability of processing explicit symbol sequence data, so that the self-encoder is widely applied to the RUL prediction of time sequence data. Zemouri et al propose a cyclic radial basis function network and use it to predict the RUL of a mechanical device. Malhi et al propose an RNN training method based on competitive learning, aiming at improving the long-term prediction accuracy of RNN. Peng et al propose a new RUL prediction method by using a large sparse matrix instead of a hidden layer, and enhance the RNN performance.
The method provided by the prior art mainly has the following problems: 1. the absolute values of the data are different but the arrangement entropy values are the same, so that misjudgment and missed judgment are easily caused; 2. the accuracy and interpretability of anomaly detection is low.
Therefore, it is an urgent technical problem in the art to provide a new method or system for evaluating and predicting the status of a device to solve the above problems in the prior art.
Disclosure of Invention
The invention aims to provide a nuclear power equipment state evaluation and prediction integrated method and system, which are used for improving the detection accuracy and interpretability while avoiding the problems of misjudgment and missed judgment and giving accurate prediction on the residual life of equipment.
In order to achieve the purpose, the invention provides the following scheme:
a nuclear power plant state assessment and prediction integrated method comprises the following steps:
acquiring operation data of nuclear power equipment to be tested;
acquiring a trained cycle time convolution network model; the cyclic time convolution network model is a convolution network model formed by stacking small convolution kernels, and takes Leaky ReLU as an activation function and cross entropy loss as a loss function;
and determining the service life of the nuclear power equipment to be tested according to the operation data by adopting the trained cycle time convolution network model.
Preferably, before the obtaining the trained cyclic time convolution network model, the method further includes:
acquiring operation data of nuclear power equipment;
performing calibration sampling on the operating data to form a training sample pair;
and training a cycle time convolution network model by adopting the training sample pair to obtain the trained cycle time convolution network model.
Preferably, after the acquiring the operation data of the nuclear power plant, the method further includes:
mapping the same-class data in the operating data to [0,1] by using a dispersion standardization method to obtain characteristic parameters;
performing spatial reconstruction on the characteristic parameters to obtain reconstructed characteristic parameters;
respectively carrying out ascending arrangement on the elements in the reconstructed characteristic parameters by utilizing a multi-scale weighted arrangement entropy to obtain a symbol sequence;
adding a weighting coefficient to each element in the symbol sequence to obtain a weighted symbol sequence, and determining the probability value of each arrangement in the symbol sequence;
determining the multi-scale weighted arrangement entropy corresponding to each symbol sequence under different scales according to the probability value;
determining the influence degree and the change rule of random noise on the operating data according to the multi-scale weighted permutation entropy corresponding to each symbol sequence;
and determining a multi-scale weighted permutation entropy threshold corresponding to each symbol sequence according to the influence degree and the change rule of the operating data, and performing calibration sampling on the operating data according to the multi-scale weighted permutation entropy threshold to form a training sample pair.
Preferably, the construction process of the cyclic time convolutional network model comprises the following steps:
acquiring an initial time convolution network model;
adjusting an activation function in the initial time convolution network model into a Leaky ReLU function, and adjusting a loss function into a cross entropy loss function to obtain an intermediate time convolution network model;
performing parameter optimization on the intermediate time convolution network model by adopting an SGD optimization algorithm to obtain a transition time convolution network model;
and adding a residual convolution structure into the transition time convolution network model to obtain a cycle time convolution network model.
An integrated nuclear power plant state assessment and prediction system comprising:
the first operation data acquisition module is used for acquiring operation data of the nuclear power equipment to be detected;
the cyclic time convolution network model acquisition module is used for acquiring a trained cyclic time convolution network model; the cyclic time convolution network model is a convolution network model which is formed by stacking small convolution kernels, takes Leaky ReLU as an activation function, takes cross entropy loss as a loss function and is formed by stacking small convolution kernels;
and the service life prediction module is used for determining the service life of the nuclear power equipment to be tested according to the operation data by adopting the trained cycle time convolution network model.
Preferably, the method further comprises the following steps:
the second operation data acquisition module is used for acquiring the operation data of the nuclear power equipment;
the training sample pair calibration module is used for carrying out calibration sampling on the operation data to form a training sample pair;
and the cycle time convolution network model training module is used for training the cycle time convolution network model by adopting the training samples to obtain the trained cycle time convolution network model.
Preferably, the method further comprises the following steps:
the characteristic parameter acquisition module is used for mapping the same type data in the operating data to the position between [0 and 1] by using a dispersion standardization method to obtain characteristic parameters;
the characteristic parameter reconstruction module is used for carrying out spatial reconstruction on the characteristic parameters to obtain reconstructed characteristic parameters;
the symbol sequence determining module is used for respectively carrying out ascending arrangement on the elements in the reconstructed characteristic parameters by utilizing the multi-scale weighted arrangement entropy so as to obtain a symbol sequence;
a probability value determining module, configured to add a weighting coefficient to each element in the symbol sequence to obtain a weighted symbol sequence, and determine a probability value appearing in each permutation in the symbol sequence;
the multi-scale weighted permutation entropy determining module is used for determining the multi-scale weighted permutation entropy corresponding to each symbol sequence under different scales according to the probability value;
the influence degree and change rule determining module is used for determining the influence degree and change rule of random noise on the running data according to the multi-scale weighted permutation entropy corresponding to each symbol sequence;
and the threshold determining module is used for determining a multi-scale weighted permutation entropy threshold corresponding to each symbol sequence according to the influence degree and the change rule of the operating data, and performing calibration sampling on the operating data according to the multi-scale weighted permutation entropy threshold to form a training sample pair.
Preferably, the method further comprises a cyclic time convolution network model building module, wherein the cyclic time convolution network model building module comprises:
the time convolution network model obtaining unit is used for obtaining an initial time convolution network model;
the function substitution unit is used for adjusting the activation function in the initial time convolution network model into a Leaky ReLU function and adjusting the loss function into a cross entropy loss function to obtain an intermediate time convolution network model;
the transition time convolution network model determining unit is used for performing parameter optimization on the intermediate time convolution network model by adopting an SGD optimization algorithm to obtain a transition time convolution network model;
and the cyclic time convolution network model building unit is used for adding the residual convolution structure into the transition time convolution network model to obtain the cyclic time convolution network model.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects:
according to the nuclear power equipment state evaluation and prediction integrated method and system, the cycle time convolution network model after structure optimization is adopted, and the prediction result of the service life of the nuclear power equipment can be accurately obtained according to the operation data of the nuclear power equipment.
In addition, the nuclear power equipment state evaluation and prediction integration method and system provided by the invention can be used for forming the cycle time convolution network model by adopting small convolution kernel stacking, considering the time attribute of fault characteristics and flexibly adjusting the cycle time convolution network model so as to achieve the purpose of enriching the extracted local characteristics, thereby improving the detection accuracy and the interpretability while avoiding the problems of misjudgment and missed judgment.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.
FIG. 1 is a general technical flow diagram of state estimation and life prediction for electrically operated valves according to the prior art;
FIG. 2 is a classification diagram of a remaining useful life prediction technique in the prior art;
FIG. 3 is a flow chart of an integrated method for evaluating and predicting the state of a nuclear power plant in accordance with the present invention;
FIG. 4 is a flow chart illustrating a technique for monitoring the status and predicting the life of an electric gate valve in a nuclear power plant according to an embodiment of the present invention;
FIG. 5 is a schematic diagram of a modeling process of a cyclic time convolution network model according to an embodiment of the present invention;
fig. 6 is a schematic structural diagram of a nuclear power plant state evaluation and prediction integrated system provided by the invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The invention aims to provide a nuclear power equipment state evaluation and prediction integrated method and system, which are used for improving the detection accuracy and interpretability while avoiding the problems of misjudgment and missed judgment and giving accurate prediction on the residual life of equipment.
In order to make the aforementioned objects, features and advantages of the present invention more comprehensible, the present invention is described in detail with reference to the accompanying drawings and the detailed description thereof.
The overall design idea and the achieved technical effect of the technical scheme provided by the invention are as follows:
according to the method, firstly, the characteristics of nonlinearity and time-varying property of characteristic parameters are combined, phase space reconstruction is respectively carried out on the characteristic parameters collected in the nuclear power equipment, sequence combination of each characteristic parameter is obtained, and anomaly detection based on multi-scale weighted arrangement entropy is convenient to carry out subsequently. On the basis, the phase space reconstruction results of each feature are spliced in rows according to feature dimensions to form three-dimensional data, wherein the first dimension represents time, the second dimension represents embedding dimensions, and the third row represents dimensions of feature parameters. In this way, data support can be provided for subsequent life prediction based on the cycle time convolutional network. Since the premise for performing the life prediction is to know the abnormal occurrence time of the nuclear power plant, and then perform subsequent prediction analysis based on this time. Therefore, the invention firstly solves the problem of abnormality detection, some students carry out abnormality detection by using methods such as permutation entropy, principal component analysis and independent component analysis at present, and few students carry out abnormality detection by using multi-scale weighted permutation entropy. The invention can reflect the important information of the characteristic parameters on a plurality of scales by adopting the multi-scale weighted permutation entropy, and reserve the useful amplitude information carried in the characteristic parameters by weighting, thereby avoiding the uncertainty and instability caused by the fact that the permutation entropy only reflects the change characteristics on one scale, avoiding the misjudgment and the missed judgment caused by the same permutation entropy value due to different absolute values of the data, and improving the accuracy and the interpretability of the anomaly detection.
After finding the abnormal occurrence point, the invention provides a life prediction method based on a cycle time convolution network. Firstly, the time attribute of fault characteristics can be considered by utilizing causal convolution in a time convolution network, and the receptive field can be increased or reduced according to requirements through expanding convolution, so that the adjustment is flexible; and finally, abundant local features can be extracted. On the basis, in order to enhance the fitting and regression capability of time series data, the invention adopts the GRU gated recurrent neural network to fit the evolution of the global sequence, and can provide a more accurate life prediction result.
Based on the above design concept, the present invention provides a flow chart of a nuclear power plant state assessment and prediction integration method, as shown in fig. 3, the nuclear power plant state assessment and prediction integration method includes:
step 100: and acquiring the operation data of the nuclear power equipment to be tested.
Step 200: and acquiring a trained cycle time convolution network model. The cyclic time convolution network model is a convolution network model formed by stacking small convolution kernels, and the convolution network model takes Leaky ReLU as an activation function and cross entropy loss as a loss function.
Step 300: and determining the service life of the nuclear power equipment to be tested according to the operation data by adopting a trained cycle time convolution network model.
Preferably, before step 200, the method further comprises:
operational data of the nuclear power plant is acquired.
And performing calibration sampling on the operation data to form a training sample pair.
And training the cycle time convolution network model by adopting the training sample pair to obtain the trained cycle time convolution network model.
Preferably, after step 100, the method further comprises:
and mapping the homogeneous data in the operation data to the range between [0,1] by using a dispersion standardization method to obtain characteristic parameters.
And carrying out spatial reconstruction on the characteristic parameters to obtain the reconstructed characteristic parameters.
And respectively carrying out ascending arrangement on the elements in the reconstructed characteristic parameters by utilizing the multi-scale weighted arrangement entropy to obtain a symbol sequence.
And adding a weighting coefficient into each element in the symbol sequence to obtain a weighted symbol sequence, and determining the probability value of each permutation in the symbol sequence.
And determining the multi-scale weighted arrangement entropy corresponding to each symbol sequence under different scales according to the probability value.
And determining the influence degree and the change rule of the random noise on the operation data according to the multi-scale weighted permutation entropy corresponding to each symbol sequence.
And determining a multi-scale weighted permutation entropy threshold corresponding to each symbol sequence according to the influence degree and the change rule of the operation data, and performing calibration sampling on the operation data according to the multi-scale weighted permutation entropy threshold to form a training sample pair.
Preferably, the process of constructing the cyclic time convolutional network model includes:
an initial time convolutional network model is obtained.
And adjusting an activation function in the initial time convolution network model into a Leaky ReLU function, and adjusting a loss function into a cross entropy loss function to obtain an intermediate time convolution network model.
And optimizing parameters of the intermediate time convolution network model by adopting an SGD optimization algorithm to obtain a transition time convolution network model.
And adding the residual convolution structure into the transition time convolution network model to obtain a cycle time convolution network model.
The following provides a specific embodiment to further illustrate the scheme of the present invention, and the specific embodiment of the present invention is described by taking monitoring of an electric gate valve in a nuclear power plant as an example, and in a specific application, the scheme of the present invention is also applicable to monitoring of the service life of other nuclear power plant components.
As shown in fig. 4, a basic flow for monitoring the service life of the electric gate valve by using the integrated method provided by the present invention includes:
step 1: the method comprises the steps of firstly collecting operation data obtained by an acoustic emission sensor, an acceleration sensor and process parameter sensors such as pressure difference, temperature and flow rate arranged on the electric gate valve, and then storing the operation data in calculation through a data collection board card.
Step 2: and processing the signals of the acceleration sensor and the non-stationary signals acquired by the acoustic emission to obtain the operation data after denoising. The time domain characteristics processed by the sensor can be directly obtained by adopting mature commercial software (data processing software conventionally used in the prior art). For the frequency domain and the time-frequency domain characteristics, the wavelet transform, the variational modal decomposition, the hilbert transform and other methods need to be adopted for processing, which is considered to be a mature technology in the field of the invention, and therefore, the related contents are not described.
And step 3: performing data characteristic engineering on the parameters acquired in the step 1 and the step 2, respectively mapping all data values of the same parameter between [0,1] by using a dispersion standardization method, and converting a function: and x is (x-min)/(max-min), wherein x is the single characteristic obtained in the steps 1 and 2, max is the maximum value of the sample, and min is the minimum value of the sample, so that the influence of inconsistent dimensions and overlarge and undersize data on the data training and subsequent analysis processes is avoided.
And 4, step 4: spatial reconstruction of the input data. For each feature parameter obtained in step 3, assuming that each feature parameter can be represented as a segment of symbol sequence { X (i), i ═ 1, 2.., N }, the present invention performs phase space reconstruction on the feature parameter to obtain X (1), X (2., X (N- (m-1) τ). Where x (i) { x (i), x (i + τ),. -, x (i + (m-1) τ) }, i ═ 1, 2. -, N- (m-1) τ, m is the embedding dimension, τ is the time delay factor. In the implementation process of the invention, m is 30, and the time delay factor is 1.
And 5: after the phase space reconstruction in step 4, the reconstructed characteristic parameters are spliced in columns, and the two-dimensional input data (dimension N × D) in step 3 can be converted into a three-dimensional stacked data block of (N-m +1) × (dimension m × D), where N is the total timing length and D is the dimension of the characteristic parameter. Since the time delay factor takes 1, the total data input length is (N-num _ steps +1) for the algorithm of the present invention. In this way, the input data at each moment is not an isolated characteristic parameter at a certain moment, but a combination of data in a period of time can support a subsequent life prediction model for calculation.
Step 6: and calculating the multi-scale weighted permutation entropy. Rearranging m elements in each characteristic parameter X (i) obtained through phase space reconstruction in the step 4 according to an ascending order:
X(i)＝{x(i+(j 1 -1)τ)≤x(i+(j 2 -1)τ)≤...≤x(i+(j m -1)τ)}
if x (i + (j) i1 -1)τ)＝x(i+(j i2 -1) τ) are ordered by the magnitude of the value of j, i.e., when j is present i1 ≤j i2 Having x (i + (j) i1 -1)τ)≤x(i+(j i2 -1) τ). Thus, any vector x (i) can obtain a symbol sequence z (i) ═ j 1 ,j 2 ,...,j m ]Wherein i is 1,2, k is less than or equal to m! . m different symbols [ j 1 ,j 2 ,...,j m ]Total m! Different arrangements correspond to a total of m! A different sequence of symbols, Z (i) ═ j 1 ,j 2 ,...,j m ]Is m! One of the symbol sequences.
Although the method of permutation entropy and the like can also be used for anomaly detection and identification, the permutation entropy is an average entropy parameter used for measuring the complexity of a one-dimensional time sequence, and the dynamic sudden change of a time signal is detected by comparing the sizes of adjacent data, so that the method has the characteristics of simple calculation, suitability for on-line monitoring and strong noise resistance, the problem of non-optimization is often caused by the fact that the permutation entropy only can be concerned about the entropy value of one scale in a phase space, and meanwhile, the error is large when the data with large change amplitude are connected together because specific amplitude information is not concerned.
In order to solve the above problems, the invention adopts the multi-scale weighted permutation entropy to detect the abnormality, and compared with the method without the weighted permutation entropy, the method retains the useful amplitude information carried in the intrinsic mode function through weighting, thereby improving the expressive force of the data characteristics. For the life prediction technology, machine learning technologies such as a common artificial neural network and a support vector machine are sleeved in the process to also realize life prediction, but the accuracy of life prediction is low because the time sequence characteristic of the characteristic parameter is not considered;
and 7: in the operation step according to permutation entropy obtained in step 6, a weighting coefficient is added to each reconstructed component after the amplitude information of the symbol sequence is considered:
wherein the content of the first and second substances,
is a weighted average.
And step 8: calculating the probability p of each permutation in the symbol sequence after coarse graining under different scales according to the weighted symbol sequence obtained in the step 7 w (π ji ) Comprises the following steps:
wherein p is w (π ji ) The variable is expressed as a whole for the probability of each arrangement mode in the symbol sequence after coarse-grained under different scales, ji is the serial number of the arrangement mode, and T is the total time step.
And step 9: for each permutation occurrence probability under different scales obtained in the step 8, the multi-scale weighted permutation entropy under the symbol sequence
Is defined as:
wherein, the first and the second end of the pipe are connected with each other,
for real-time weighted permutation entropy at different scales, H pw The value range of (m, tau) is H which is more than or equal to 0 pw (m, tau) is less than or equal to 1, m is embedding dimension, tau is time delay factor. H pw The size of the (m, tau) value represents the symbol sequenceDegree of complexity and randomness of H pw The larger (m, τ) the more random the symbol sequence is. Conversely, the more regular the symbol sequence is illustrated.
Step 10: based on the multi-scale weighted permutation entropy calculation result of step 9, the influence degree and change rule of random noise on the operation parameters are obtained by adopting a threshold value method for short, so that the statistical threshold values of the weighted permutation entropies under different scales can be set, and the interference of the random noise on the abnormal monitoring can be avoided.
Step 11: after an anomaly detection threshold value based on the multi-scale weighted arrangement entropy is set, a cycle time convolution network model is further established. Compared with a large convolution kernel, the time convolution network model structure formed by stacking small convolution kernels can more effectively extract the nonlinear characteristics of the measured data, as shown in fig. 5. For dilation convolution and causal convolution, convolution is only a common convolution when the dilation coefficient d is 1. The larger the expansion coefficient, the longer the input range. Therefore, the invention can freely adjust the structure of TCN (time convolution network model) by changing the size of the expansion factor and the convolution kernel, can increase the feeling and establish the long-range dependence relationship between sequence elements. The time convolution network of the invention adopts a formula (3) to extract the characteristics, wherein l is the first convolution layer, k is a convolution kernel, b is a bias parameter,
is the output of the l-th layer,
is the input of the l-1 layer, and the characteristic diagram is M j 。
Step 12: the activation functions involved in the time convolution network are all adjusted to be Leaky ReLU functions, dead nodes can be avoided on the basis of the ReLU activation functions, and nonlinear characteristics in data can be reflected better.
Step 13: and (5) on the basis of the step 6, dropout operation is applied to the formed time convolution network model, so that the time convolution network can be more stable to prevent the over-fitting phenomenon.
Step 14: in order to solve the problems that gradient disappearance and the like can be caused by a depth TCN model, the invention refers to a residual convolution structure in a residual network. As shown in fig. 5, after the operations of step 5, step 6 and step 7 are performed for 2 times, a residual convolution structure is added, the input data and the output result form a serial-parallel structure, the data feature dimensions are enriched, and long-time and fine-grained local correlation features are formed.
Step 15: a gru (gated current unit) unit performs serialized modeling on the features extracted in step 14, and captures an evolution mode of the sequence. At each point in time, the GRU receives an input feature vector S of sequence elements t And combined with the hidden state h at the previous moment t-1 Weighted update of the hidden state h at the present time t . The iterative update formula is as follows:
where, is the dot product operation, reset the threshold r t And updating the threshold z t And controlling the information update of each hidden layer. W * And U * Representing a matrix of coefficients, b * A bias vector is represented for adaptively selecting and discarding historical information that constructs elements of the current sequence. Given sequence feature matrix S ═ S 1 ,s 2 ,...,s n-k Inputting GRU in sequence for iterative computation to obtain hidden layer vectors { h } 1 ,h 2 ,...,h n-k }. Each hidden vector is considered to contain the logical information of the next sequence element. On the basis, the final life prediction result is obtained by inputting the life prediction result into the full-connection network.
The long-time and short-time memory network and other deep learning technologies have high accuracy, but large calculation amount and low calculation speed. Aiming at the problem, the invention adopts a cycle time convolution network model, can fully extract the depth characteristics of the original data, and meanwhile, the time convolution network has the regression analysis capability of the time sequence data; on the basis of a time convolution network, a GRU unit is introduced, so that the prediction of the time sequence correlation characteristics can be further enhanced, and the effect of enhancing the service life prediction accuracy is finally achieved.
Step 16: defining a loss function and optimizing parameters. The present invention uses cross entropy loss as a loss function. In order to optimize the weight and the bias in the time convolution network, an SGD optimization algorithm is adopted to solve the network in the training process, so that the value of a loss function is as small as possible, and finally, network structure parameters which best meet the classification characteristics of the fault mode of the nuclear power device are obtained. In the calculation process of each back propagation, the learning rate of the first 5 iterations is set to be 0.001, the learning rate is not attenuated, and the attenuation rate of the learning rate of each subsequent iteration is set to be 0.99. With the increase of the number of training rounds and the reduction of training errors, the time convolution network model can continuously approximate to the actual fault characteristic.
And step 17: in the invention, 5-fold verification is adopted for training and testing, 80% of the data processed in the step 5 is selected as training data, and the other 20% of the data is used as test data to output the life prediction result predicted by the model. On the basis, the initial time convolution network and the GRU unit parameters are adopted to train the network. In the training process of the model, all data are divided into a plurality of batches of training samples in order to improve the training speed and efficiency, and the processed data are input into a time convolution network and a GRU unit for training after being randomly disordered to reduce uncertainty.
Step 18: after the training process of the whole state evaluation and life prediction is completed in steps 3-17, the model is applied to the actual analysis process. First, data acquisition and normalization are performed in a manner completely consistent with steps 1,2, and 3.
Step 19: and (5) calculating actual operation data by utilizing the calculation processes of the step 4 and the step 6-10 in the modeling stage to obtain the weighted arrangement entropy values of all the characteristic parameters in the same step 10 under multiple scales.
Step 20: and (3) comparing the real-time weighted permutation entropy obtained in the step (19) with the statistical threshold obtained in the step (10), and if the real-time weighted permutation entropy under different scales of all the characteristic parameters does not exceed the statistical threshold obtained in the step (10), continuously monitoring. If the real-time weighted arrangement entropy under different scales of a certain characteristic parameter exceeds the statistical threshold, the operation parameter is abnormal, and the service life needs to be further predicted.
Step 21: after the abnormal state is detected in the step 20 and the life prediction process needs to be carried out, the method can utilize the cycle time convolution network model optimized in the steps 11-17 to predict the residual service life of the nuclear power equipment in the actual operation process, and finally obtain the residual service life value. The obtained related result of the residual service life value can be referred by maintenance and decision-making personnel so as to take related measures in time, and the economy can be improved while the safety is ensured.
Based on the specific embodiments provided above, the advantages of the present invention are explained:
1. the reason for the high accuracy is the overall implementation of step 4, step 5, step 6, step 7, step 8, step 9, step 10 and steps 11-15 in part 2.
2. And 3, phase space reconstruction is respectively carried out on all characteristic parameters, so that each instantaneous single data point can be converted into a combination of a section of time sequence data in the phase space, and the multi-scale weighted arrangement entropy can be conveniently calculated subsequently.
3. In step 4, symbol sequence data after feature combination can be obtained by combining each feature parameter of phase space reconstruction in step 3 according to columns, namely, original 2-dimensional data is converted into a three-dimensional data group with symbol sequence attributes, so that the subsequent residual service life prediction process can not focus on a certain single instant any more, but focus on a section of symbol sequence.
4. And 6, processing the symbols by a phase space reconstruction technology according to the principle of the multi-scale arrangement entropy, representing the dynamic mutation of the symbol sequence of the intrinsic mode function obtained after the variation mode decomposition processing, performing coarse graining processing on the symbol sequence, facilitating the subsequent use of the multi-scale weighting arrangement entropy, and then calculating the arrangement entropy under the set scale. In order to ensure the real-time property of feature extraction, the invention adopts the maximum overlapping moving window method for selection, and can carry out coarse graining treatment on the symbol sequence in the moving window.
5. And 7-9, the weighted permutation entropy is considered on the basis of the multi-scale permutation entropy, amplitude information is not considered in the multi-scale permutation entropy, and only the permutation sequence of the amplitudes is considered, so that the objective and inaccurate calculation result is easily caused. The multi-scale weighted permutation entropy can keep useful amplitude information carried by the signal, so that the method has better robustness and stability, and has unique capability of extracting complexity information from data with spike characteristics or suddenly changed amplitudes.
6. Step 10, the noise interference level in the historical normal operation data can be calculated through a multi-scale weighted arrangement entropy formula, and the monitoring threshold can be reasonably determined through calculation, so that the problems of misdiagnosis and missed diagnosis caused by improper threshold setting are avoided.
7. Step 11, a convolution neural network model structure formed by stacking small convolution kernels is established, and compared with a large convolution kernel, the nonlinear features of the measured data can be extracted more effectively. Meanwhile, the size of the receptive field of the time convolution network can be adjusted arbitrarily by adjusting the size of the expansion factor and the convolution kernel.
8. In the step 12, a Leaky ReLU function is used as an activation function, so that dead nodes can be avoided on the basis of the ReLU activation function, the sparse model can better mine relevant characteristics, training data are fitted, and nonlinear characteristics in the data can be reflected better.
9. Step 13 adopts dropout operation in the neural network, which can prevent overfitting of the neural network result of the invention, so that the obtained life prediction result is more stable and does not generate excessive fluctuation.
10. Step 14, by setting the input and output serial-parallel structure in each layer of residual convolution based on the residual convolution, the causal time sequence relation among data can be deeply memorized, and a better fault diagnosis effect is achieved.
11. And step 15, after the characteristics are extracted through the time convolution network, modeling the long-distance dependence relationship of characteristic parameters on the symbol sequence through the GRU threshold recurrent neural network, capturing the evolution mode of sequence elements and fitting a global sequence, so that more accurate residual service life prediction can be realized.
12. Step 16 sets the learning rate of the previous 5 iterations to 0.001 in the calculation process of each back propagation, does not attenuate the learning rate, and sets the attenuation rate of the learning rate of each subsequent iteration to 0.99. Through the change of the learning rate, the most appropriate weight and bias can be found more accurately in the back propagation calculation process, and finally the accuracy of the model is improved.
13. And step 20, comparing the weighted permutation entropy under different scales calculated in real time with a corresponding statistical threshold, and if the weighted permutation entropy exceeds the statistical threshold, indicating that the operation parameters are abnormal and sending an alarm. The timeliness of the abnormity monitoring structure can be guaranteed, meanwhile, through the comparison of the statistical threshold value and the real value, an abnormity detection result is visual and vivid, and the abnormity judgment of operators can be well supported.
In addition, aiming at the provided nuclear power equipment state evaluation and prediction integration method, the invention also correspondingly provides a nuclear power equipment state evaluation and prediction integration system. The integrated system includes:
the first operation data acquisition module 1 is used for acquiring operation data of the nuclear power equipment to be detected.
And the cyclic time convolution network model acquisition module 2 is used for acquiring the trained cyclic time convolution network model. The cyclic time convolution network model is a convolution network model formed by stacking small convolution kernels, and the convolution network model takes Leaky ReLU as an activation function and cross entropy loss as a loss function.
And the service life prediction module 3 is used for determining the service life of the nuclear power equipment to be tested according to the operation data by adopting the trained cycle time convolution network model.
As a preferred embodiment of the present invention, the integrated system may further include:
and the second operation data acquisition module is used for acquiring the operation data of the nuclear power equipment.
And the training sample pair calibration module is used for performing calibration sampling on the operation data to form a training sample pair.
And the cycle time convolution network model training module is used for training the cycle time convolution network model by adopting the training samples to obtain the trained cycle time convolution network model.
As another preferred embodiment of the present invention, the integrated system further includes:
and the characteristic parameter acquisition module is used for mapping the same-class data in the operating data to the range between [0 and 1] by using a dispersion standardization method so as to obtain characteristic parameters.
And the characteristic parameter reconstruction module is used for carrying out spatial reconstruction on the characteristic parameters to obtain the reconstructed characteristic parameters.
And the symbol sequence determining module is used for respectively carrying out ascending arrangement on the elements in the reconstructed characteristic parameters by utilizing the multi-scale weighted arrangement entropy so as to obtain a symbol sequence.
And the probability value determining module is used for adding a weighting coefficient into each element in the symbol sequence to obtain a weighted symbol sequence and determining the probability value of each arrangement in the symbol sequence.
And the multi-scale weighted arrangement entropy determining module is used for determining the multi-scale weighted arrangement entropy corresponding to each symbol sequence under different scales according to the probability value.
And the influence degree and change rule determining module is used for determining the influence degree and change rule of the random noise on the operation data according to the multi-scale weighted arrangement entropy corresponding to each symbol sequence.
And the threshold determining module is used for determining the multi-scale weighted permutation entropy threshold corresponding to each symbol sequence according to the influence degree and the change rule of the operating data and calibrating and sampling the operating data according to the multi-scale weighted permutation entropy threshold to form a training sample pair.
As another preferred embodiment of the present invention, the integrated system further includes a cyclic time convolution network model building module, where the cyclic time convolution network model building module includes:
and the time convolution network model obtaining unit is used for obtaining an initial time convolution network model.
And the function substitution unit is used for adjusting the activation function in the initial time convolution network model into a Leaky ReLU function and adjusting the loss function into a cross entropy loss function to obtain an intermediate time convolution network model.
And the transition time convolution network model determining unit is used for performing parameter optimization on the intermediate time convolution network model by adopting an SGD optimization algorithm to obtain the transition time convolution network model.
And the cycle time convolution network model building unit is used for adding the residual convolution structure into the transition time convolution network model to obtain the cycle time convolution network model.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. For the system disclosed by the embodiment, the description is relatively simple because the system corresponds to the method disclosed by the embodiment, and the relevant points can be referred to the method part for description.
The principle and the implementation mode of the invention are explained by applying a specific example, and the description of the embodiment is only used for helping to understand the method and the core idea of the invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the foregoing, the description is not to be taken in a limiting sense.
Claims (2)
1. A nuclear power plant state assessment and prediction integration method is characterized by comprising the following steps:
acquiring operation data of nuclear power equipment to be tested;
acquiring a trained cycle time convolution network model; the cyclic time convolution network model is a convolution network model formed by stacking small convolution kernels, and takes Leaky ReLU as an activation function and cross entropy loss as a loss function;
determining the service life of the nuclear power equipment to be tested according to the operation data by adopting the trained cycle time convolution network model;
the construction process of the cycle time convolution network model comprises the following steps:
acquiring an initial time convolution network model;
adjusting an activation function in the initial time convolution network model into a Leaky ReLU function, and adjusting a loss function into a cross entropy loss function to obtain an intermediate time convolution network model;
performing parameter optimization on the intermediate time convolution network model by adopting an SGD optimization algorithm to obtain a transition time convolution network model;
adding a residual convolution structure into the transition time convolution network model to obtain a cycle time convolution network model;
before the obtaining of the trained cyclic time convolutional network model, the method further includes:
acquiring operation data of nuclear power equipment;
mapping the same-class data in the operating data to [0,1] by using a dispersion standardization method to obtain characteristic parameters;
performing spatial reconstruction on the characteristic parameters to obtain reconstructed characteristic parameters;
respectively carrying out ascending arrangement on elements in the reconstructed characteristic parameters by utilizing a multi-scale weighted arrangement entropy to obtain a symbol sequence;
adding a weighting coefficient to each element in the symbol sequence to obtain a weighted symbol sequence, and determining the probability value of each arrangement in the symbol sequence;
determining a multi-scale weighted arrangement entropy corresponding to each symbol sequence under different scales according to the probability value;
determining the influence degree and the change rule of random noise on the operating data according to the multi-scale weighted permutation entropy corresponding to each symbol sequence;
determining a multi-scale weighted permutation entropy threshold corresponding to each symbol sequence according to the influence degree and the change rule of the operating data, and performing calibration sampling on the operating data according to the multi-scale weighted permutation entropy threshold to form a training sample pair;
and training a cycle time convolution network model by adopting the training sample pair to obtain the trained cycle time convolution network model.
2. An integrated nuclear power plant state assessment and prediction system, comprising:
the first operation data acquisition module is used for acquiring operation data of the nuclear power equipment to be detected;
the cyclic time convolution network model acquisition module is used for acquiring a trained cyclic time convolution network model; the cyclic time convolution network model is a convolution network model formed by stacking small convolution kernels, and takes Leaky ReLU as an activation function and cross entropy loss as a loss function;
the service life prediction module is used for determining the service life of the nuclear power equipment to be tested according to the operation data by adopting the trained cycle time convolution network model;
the system also comprises a module for constructing the cyclic time convolution network model, wherein the module for constructing the cyclic time convolution network model comprises:
the time convolution network model obtaining unit is used for obtaining an initial time convolution network model;
the function substitution unit is used for adjusting an activation function in the initial time convolution network model into a Leaky ReLU function and adjusting a loss function into a cross entropy loss function to obtain an intermediate time convolution network model;
the transition time convolution network model determining unit is used for performing parameter optimization on the intermediate time convolution network model by adopting an SGD optimization algorithm to obtain a transition time convolution network model;
the cyclic time convolution network model building unit is used for adding a residual convolution structure into the transition time convolution network model to obtain a cyclic time convolution network model;
the second operation data acquisition module is used for acquiring the operation data of the nuclear power equipment;
the characteristic parameter acquisition module is used for mapping the same type data in the operating data to the position between [0 and 1] by using a dispersion standardization method to obtain characteristic parameters;
the characteristic parameter reconstruction module is used for carrying out spatial reconstruction on the characteristic parameters to obtain reconstructed characteristic parameters;
the symbol sequence determining module is used for respectively carrying out ascending arrangement on the elements in the reconstructed characteristic parameters by utilizing the multi-scale weighted arrangement entropy so as to obtain a symbol sequence;
a probability value determining module, configured to add a weighting coefficient to each element in the symbol sequence to obtain a weighted symbol sequence, and determine a probability value appearing in each permutation in the symbol sequence;
the multi-scale weighted permutation entropy determining module is used for determining the multi-scale weighted permutation entropy corresponding to each symbol sequence under different scales according to the probability value;
the influence degree and change rule determining module is used for determining the influence degree and change rule of random noise on the running data according to the multi-scale weighted permutation entropy corresponding to each symbol sequence;
the threshold value determining module is used for determining a multi-scale weighted permutation entropy threshold value corresponding to each symbol sequence according to the influence degree and the change rule of the operating data, and performing calibration sampling on the operating data according to the multi-scale weighted permutation entropy threshold value to form a training sample pair;
and the cycle time convolution network model training module is used for training the cycle time convolution network model by adopting the training samples to obtain the trained cycle time convolution network model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010908744.9A CN112033463B (en) | 2020-09-02 | 2020-09-02 | Nuclear power equipment state evaluation and prediction integrated method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010908744.9A CN112033463B (en) | 2020-09-02 | 2020-09-02 | Nuclear power equipment state evaluation and prediction integrated method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112033463A CN112033463A (en) | 2020-12-04 |
| CN112033463B true CN112033463B (en) | 2022-09-06 |
Family
ID=73591114
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010908744.9A Active CN112033463B (en) | 2020-09-02 | 2020-09-02 | Nuclear power equipment state evaluation and prediction integrated method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112033463B (en) |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112613646A (en) * | 2020-12-08 | 2021-04-06 | 上海交通大学烟台信息技术研究院 | Equipment state prediction method and system based on multi-dimensional data fusion |
| CN113551765A (en) * | 2021-08-17 | 2021-10-26 | 中冶北方（大连）工程技术有限公司 | Sound spectrum analysis and diagnosis method for equipment fault |
| CN113673165B (en) * | 2021-08-26 | 2023-07-25 | 中国核动力研究设计院 | Reliability evaluation method and system for nuclear grade pipeline |
| CN114098681A (en) * | 2021-11-12 | 2022-03-01 | 南京海量物联科技有限公司 | Intelligent blood pressure prediction method based on TCN model and PPG signal |
| CN114386451B (en) * | 2021-12-03 | 2024-05-03 | 中铁第一勘察设计院集团有限公司 | Contact net hanger fault diagnosis alarm method based on sensor information perception |
| CN114510992A (en) * | 2021-12-27 | 2022-05-17 | 上海梦象智能科技有限公司 | Equipment switch state detection method based on deep learning |
| CN114897292B (en) * | 2022-03-28 | 2023-04-07 | 南京林业大学 | Fault diagnosis method and system based on multi-scale residual convolution variational network |
| CN115452358B (en) * | 2022-09-22 | 2024-01-26 | 中国核动力研究设计院 | Accelerated test method for action life of electric valve |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109977996A (en) * | 2019-02-12 | 2019-07-05 | 华北水利水电大学 | Hydro-structure operating state monitoring system based on Time Series Clustering fusion |
| CN110059377A (en) * | 2019-04-02 | 2019-07-26 | 西南交通大学 | A kind of fuel battery service life prediction technique based on depth convolutional neural networks |
| CN110108456A (en) * | 2019-04-16 | 2019-08-09 | 东南大学 | A kind of rotating machinery health evaluating method of depth convolutional neural networks |
| CN110555479A (en) * | 2019-09-05 | 2019-12-10 | 北京工业大学 | fault feature learning and classifying method based on fusion of 1DCNN and GRU |
| CN110633792A (en) * | 2019-10-22 | 2019-12-31 | 西安交通大学 | End-to-end bearing health index construction method based on convolution cyclic neural network |
| CN110738274A (en) * | 2019-10-26 | 2020-01-31 | 哈尔滨工程大学 | nuclear power device fault diagnosis method based on data driving |
| CN111191740A (en) * | 2020-01-10 | 2020-05-22 | 福州大学 | Fault diagnosis method for rolling bearing |
| CN111325112A (en) * | 2020-01-31 | 2020-06-23 | 贵州大学 | Cutter wear state monitoring method based on depth gate control circulation unit neural network |
| CN111340282A (en) * | 2020-02-21 | 2020-06-26 | 山东大学 | DA-TCN-based method and system for estimating residual service life of equipment |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105719002A (en) * | 2016-01-18 | 2016-06-29 | 重庆大学 | Wind turbine generator state parameter abnormity identification method based on combination prediction |
| US10387768B2 (en) * | 2016-08-09 | 2019-08-20 | Palo Alto Research Center Incorporated | Enhanced restricted boltzmann machine with prognosibility regularization for prognostics and health assessment |
| US20200063665A1 (en) * | 2018-01-25 | 2020-02-27 | Dalian University Of Technology | Aero-engine full flight envelope model adaptive modification method based on deep learning algorithm |
-
2020
- 2020-09-02 CN CN202010908744.9A patent/CN112033463B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109977996A (en) * | 2019-02-12 | 2019-07-05 | 华北水利水电大学 | Hydro-structure operating state monitoring system based on Time Series Clustering fusion |
| CN110059377A (en) * | 2019-04-02 | 2019-07-26 | 西南交通大学 | A kind of fuel battery service life prediction technique based on depth convolutional neural networks |
| CN110108456A (en) * | 2019-04-16 | 2019-08-09 | 东南大学 | A kind of rotating machinery health evaluating method of depth convolutional neural networks |
| CN110555479A (en) * | 2019-09-05 | 2019-12-10 | 北京工业大学 | fault feature learning and classifying method based on fusion of 1DCNN and GRU |
| CN110633792A (en) * | 2019-10-22 | 2019-12-31 | 西安交通大学 | End-to-end bearing health index construction method based on convolution cyclic neural network |
| CN110738274A (en) * | 2019-10-26 | 2020-01-31 | 哈尔滨工程大学 | nuclear power device fault diagnosis method based on data driving |
| CN111191740A (en) * | 2020-01-10 | 2020-05-22 | 福州大学 | Fault diagnosis method for rolling bearing |
| CN111325112A (en) * | 2020-01-31 | 2020-06-23 | 贵州大学 | Cutter wear state monitoring method based on depth gate control circulation unit neural network |
| CN111340282A (en) * | 2020-02-21 | 2020-06-26 | 山东大学 | DA-TCN-based method and system for estimating residual service life of equipment |
Non-Patent Citations (4)
| Title |
|---|
| Convolution and Long Short-Term Memory Hybrid Deep Neural Networks for Remaining Useful Life Prognostics;Kong ZM, et al;《APPLIED SCIENCES-BASEL》;20191031;第9卷(第19期);第1-18页 * |
| Remaining Useful Life Prediction Techniques of Electric Valves for Nuclear Power Plants with Convolution Kernel and LSTM;Wang H, et al;《SCIENCE AND TECHNOLOGY OF NUCLEAR INSTALLATIONS》;20200828;第2020卷;第1-13页 * |
| 基于BP-SDG的核动力装置典型故障诊断研究;武茂浦 等;《中国核科学技术进展报告》;20171031;第5卷;第469页倒数第5行-第473页第5行 * |
| 基于深度多尺度卷积神经网络的剩余使用寿命预测模型;原建洋等;《科技创新导报》;20200521;第17卷(第15期);第243-244页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112033463A (en) | 2020-12-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112033463B (en) | Nuclear power equipment state evaluation and prediction integrated method and system | |
| Xia et al. | Multi-stage fault diagnosis framework for rolling bearing based on OHF Elman AdaBoost-Bagging algorithm | |
| CN110334764A (en) | Rotating machinery intelligent failure diagnosis method based on integrated depth self-encoding encoder | |
| CN115758212B (en) | Mechanical equipment fault diagnosis method based on parallel network and transfer learning | |
| CN111412977A (en) | Preprocessing method for vibration sensing data of mechanical equipment | |
| CN103868692A (en) | Rotary machine fault diagnosis method based on kernel density estimation and K-L divergence | |
| CN104880217B (en) | A kind of fault sensor signal reconstruct method based on the measured value degree of association | |
| Zhang et al. | Early fault detection method of rolling bearing based on MCNN and GRU network with an attention mechanism | |
| CN112036042B (en) | Power equipment abnormality detection method and system based on variational modal decomposition | |
| CN111122162B (en) | Industrial system fault detection method based on Euclidean distance multi-scale fuzzy sample entropy | |
| CN105424366A (en) | Bearing fault diagnosis method based on EEMD adaptive denoising | |
| CN111881594B (en) | Non-stationary signal state monitoring method and system for nuclear power equipment | |
| CN113743016B (en) | Engine residual life prediction method based on self-encoder and echo state network | |
| Zhang et al. | Rotating Machinery Remaining Useful Life Prediction Scheme Using Deep‐Learning‐Based Health Indicator and a New RVM | |
| CN111595541A (en) | Multi-dimensional structure damage identification method based on mass vibration transmissibility data convolutional neural network processing | |
| Qin et al. | Remaining useful life prediction for rotating machinery based on optimal degradation indicator | |
| CN110060368A (en) | Mechanical method for detecting abnormality based on potential feature coding | |
| CN109472097A (en) | A kind of transmission line of electricity on-line monitoring equipment method for diagnosing faults | |
| Kang et al. | A dual-experience pool deep reinforcement learning method and its application in fault diagnosis of rolling bearing with unbalanced data | |
| Zhang et al. | Graph neural network-based bearing fault diagnosis using Granger causality test | |
| CN116793666A (en) | Wind turbine generator system gearbox fault diagnosis method based on LSTM-MLP-LSGAN model | |
| CN114046816B (en) | Sensor signal fault diagnosis method based on lightweight gradient lifting decision tree | |
| Zhang et al. | Fault Diagnosis with Bidirectional Guided Convolutional Neural Networks Under Noisy Labels | |
| CN114266013B (en) | Transmission system vibration decoupling method based on deep learning virtual perception network | |
| Guo | Wind turbine generator bearing Condition Monitoring with NEST method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |