Control rod driving mechanism fault diagnosis method based on reeliefF feature selection
Through the combination of releaseF feature selection and support vector machine model, the problem of insufficient feature selection in CRDM fault diagnosis is solved, more accurate fault diagnosis is achieved, and fault identification and positioning capabilities are improved.
Patent Information
- Application Number
- CN202411152499.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-21
- Publication Date
- 2024-12-06
AI Technical Summary
In the fault diagnosis of control rod driving mechanism (CRDM), the prior art relies on complex electromechanical models and data driving methods, resulting in insufficient feature selection, affecting the accuracy of fault diagnosis, and there are limitations in determining the number of features.
Using a method based on releaseF feature selection, through data acquisition, preprocessing, feature extraction and support vector machine model training, the number of features is gradually increased to determine the optimal number of features, and the effectiveness of feature selection and the accuracy of fault diagnosis are improved.
It improves the effectiveness of feature selection, improves the shortcomings of unsupervised learning methods, solves the problem of determining the number of features, realizes more accurate fault diagnosis, and enhances the ability to quickly identify CRDM faults and locate fault causes.
Smart Images

Figure CN119089286A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a control rod drive mechanism fault diagnosis method based on reliefF feature selection. Background Art
[0002] The control rod drive mechanism (CRDM) is an important part of a nuclear reactor. It consists of a series of mechanical-motor links such as a reluctance motor, bearings, and screws. The reluctance motor provides operating power and adopts VF open-loop control. The given excitation signal is a three-phase voltage, and the feedback signals that can be collected are three-phase current and vibration signals. The specific function of CRDM is to lift, lower, hold or quickly insert the control rods to complete the reactor startup, adjust the reactor power, maintain power, stop the reactor, and quickly shut down the reactor in the event of an accident (also known as accident quick insertion or accident rod drop).
[0003] In order to ensure the safe operation of nuclear power plants, it is necessary to detect the CRDM status. The current common practice is to perform regular detection of the CRDM status. However, regular detection of CRDM has the following disadvantages: First, regular detection is time-consuming and labor-intensive, and relies heavily on the experience and technical level of engineers to determine the status and diagnose faults, which poses the risk of missed detection and excessive maintenance; second, it is impossible to grasp the current operating status and health status of the equipment in a timely manner, and it is impossible to effectively warn in the early stages of fault accumulation and when a fault occurs. Therefore, it is necessary to carry out research on data-driven intelligent monitoring technology, realize the rapid identification of CRDM faults through three-phase current and vibration data, assist staff to quickly locate the cause of the fault, and improve the safety and efficiency of reactor operation.
[0004] In order to solve the above problems, the following technical solutions are adopted in the prior art:
[0005] At present, the research on CRDM is mainly based on simulation research and real-time monitoring and signal processing. Simulation research requires the construction of complex physical models and control systems, and requires a lot of computing resources, so real-time monitoring of CRDM and processing of monitoring data are more widely used in CRDM fault diagnosis. The common steps of real-time detection and signal processing are to receive data from sensors in real time, extract features from the data, and finally use the extracted features as the input of the classifier to realize fault identification, thereby helping inspectors to take necessary maintenance measures in time. It is worth noting that if all the extracted feature values are used to form a feature vector as the input of the classifier, its dimension will be too high, which will not only increase the complexity of the algorithm, but also bring the risk of overfitting. Therefore, after feature extraction, feature dimensionality reduction will be used to select the most discriminative features, and the dimension of the data set will be reduced by reducing the number of features in the data. This process helps to simplify the complexity of the model, reduce computational costs, and eliminate noise and redundant information in the data, thereby improving the generalization ability of the model.
[0006] The feature selection method based on principal component analysis (PCA) is often used. It converts a high-dimensional data set into a low-dimensional data set through linear transformation while retaining the main structural information of the data set. First, the data is standardized, and then the covariance matrix of the data is calculated and the eigenvalue decomposition of the covariance matrix is performed to obtain the eigenvalues and corresponding eigenvectors. The principal component is selected according to the size of the eigenvalue, and a projection matrix is constructed. Finally, the data is projected onto the selected principal component to achieve data dimensionality reduction.
[0007] Feature selection methods based on Linear Discriminant Analysis (LDA) are also common, which achieves data dimensionality reduction by finding a new feature space that maximizes the distance between classes and minimizes the distance within classes. LDA first requires the calculation of the mean vector of the category, then the calculation of the intra-class scatter matrix and the inter-class scatter matrix, and then finds the projection direction by solving the generalized eigenvalue problem, and finally maps the data to a new low-dimensional space. There are many other feature selection methods that can also be used for feature selection of CRDM datasets.
[0008] Although the above method can solve certain problems, it has the following defects:
[0009] Existing technologies rely on electromechanical models for fault behavior analysis, which requires complex modeling and a lot of calculations. In addition, if data-driven methods are used for CRDM fault diagnosis, there will also be limitations in feature selection, and the most relevant features cannot be effectively extracted, affecting the accuracy of fault diagnosis.
[0010] The feature selection method based on PCA is an unsupervised learning method that only focuses on the variance of the data and does not consider the category label of the data. In addition, PCA only focuses on the direction with the largest variance, and the features it selects are not necessarily the features most relevant to the target variable. In short, the features obtained by using PCA dimensionality reduction may not be able to distinguish various fault categories well.
[0011] The feature selection method based on LDA is different from PCA. It is a supervised learning method that uses category label information to make the projected data more clearly distinguished in the new space. In the case of obvious category distinction, LDA can effectively reduce the dimension and retain the difference between categories, but one disadvantage of LDA is that it does not directly determine the number of features. If the number of features is to be determined, it can only be judged by experience or by the cumulative contribution of the discriminant power, and the number of features is selected when the cumulative contribution reaches a certain threshold, such as 85%.
[0012] Based on this, the present invention is proposed. Summary of the invention
[0013] The purpose of the present invention is to provide a control rod drive mechanism fault diagnosis method based on reliefF feature selection, so as to improve the effectiveness of feature selection and enhance the accuracy of fault diagnosis.
[0014] To achieve the above object, the present invention provides the following technical solutions:
[0015] The control rod drive mechanism fault diagnosis method based on reliefF feature selection includes the following steps:
[0016] S100 data acquisition, collecting feedback signal sample data of the control rod drive mechanism (CRDM) operation;
[0017] S200 data preprocessing, grouping, data cleaning and normalization of the collected feedback signal sample data;
[0018] S300 feature extraction, extracting features in the time domain and frequency domain of the preprocessed feedback signal to obtain a feature data set;
[0019] S400 feature selection, using reliefF to sort the features in the feature dataset and select the most relevant features;
[0020] The S500 fault diagnosis model training adopts the support vector machine model (SVM) and trains the model, and determines the parameters of the model during the training process;
[0021] The number of S600 features is determined. The classification accuracy of the support vector machine model is observed by gradually increasing the number of features until the classification performance stops increasing.
[0022] S700 fault diagnosis, inputs subsequent data into the trained fault diagnosis model for fault diagnosis.
[0023] The present invention provides a preferred solution, in the data acquisition step S100, the feedback signal includes a three-phase current and a vibration signal.
[0024] The present invention provides a preferred solution, in which the data preprocessing step S200 includes removing abnormal values and noise in the collected sample data.
[0025] The present invention provides a preferred solution. In the S300 feature extraction step, the time domain features are extracted and divided into two categories: dimensional and dimensionless. The dimensional time domain features are at least related to the motor speed and the working state of the load when the control rod drive mechanism (CRDM) is in operation; the dimensionless time domain features are irrelevant to the working state of the control rod drive mechanism when it is in operation.
[0026] The present invention provides a preferred solution, in the feature extraction step S300, the time domain features reflect time and signal amplitude, including mean, peak, root mean square, skewness and kurtosis.
[0027] The present invention provides a preferred solution, in the feature extraction step S300, fast Fourier transform (FFT) is used in the frequency domain to extract features, and the frequency domain features at least reflect the amplitude and phase information of each frequency band.
[0028] The present invention provides a preferred solution, in which the feature selection step S400 specifically includes the following sub-steps:
[0029] S410 selects an instance from a random category, selects a point of the same category that is closest to the instance, and then selects a point of a different category that is closest to the instance;
[0030] Calculate the distance between points within a class and between classes, assign different weights based on the distance within a class and the distance between points between classes, and update the weight for each feature;
[0031] The feature with the highest score is selected as the most important feature, and finally a feature vector is obtained from large to small based on the weight.
[0032] The present invention provides a preferred solution, in which the fault diagnosis model training step S500 specifically includes the following sub-steps:
[0033] The S510 support vector machine model uses the radial basis function kernel (RBF kernel);
[0034] S520 sets the hyper parameters of the SVM classifier;
[0035] S530 divides the CRDM data set into a training set and a test set, and performs model training and final result prediction respectively.
[0036] The present invention provides a preferred solution, in the step of determining the number of features in S600, the evaluation indexes of the classification accuracy of the support vector machine model adopt F-measure and G-mean.
[0037] The present invention provides a preferred solution, in the step S600 of determining the number of features, the number of features is selected in the range of 3 to 22.
[0038] Compared with the prior art, the above technical solution has the following advantages:
[0039] (1) The present invention improves the effectiveness of feature selection, improves the shortcomings of the PCA unsupervised learning method, and solves the problem of determining the number of LDA features. The method of the present invention can solve the problem that feature selection cannot effectively select the most suitable number of features in the current CRDM fault diagnosis process.
[0040] (2) The present invention adopts the reliefF algorithm to sort features, which can more effectively screen out the most discriminative features, and can gradually increase the number of features as the input of the classifier, and determine the optimal number of features through the classification results fed back by the classifier. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0042] Figure 1 A flowchart of a control rod drive mechanism fault diagnosis method based on reliefF feature selection provided by a specific embodiment of the present invention;
[0043] Figure 2 A flowchart of the reliefF algorithm used in a specific embodiment of the present invention is provided;
[0044] Figure 3 A weight graph based on the total characteristics of the reliefF algorithm provided by a specific embodiment of the present invention;
[0045] Figure 4 A scatter point distribution diagram of the optimal feature dimension provided by a specific embodiment of the present invention;
[0046] Figure 5A comparison diagram of random feature dimension scatter distribution diagrams provided by a specific embodiment of the present invention;
[0047] Figure 6 A diagram showing the predicted classification results of a CRDM test set provided in a specific embodiment of the present invention;
[0048] Figure 7 An evaluation index score diagram of a CRDM test set provided in a specific embodiment of the present invention. DETAILED DESCRIPTION
[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0050] Please refer to Figure 1 The control rod drive mechanism fault diagnosis method based on reliefF feature selection in this embodiment includes the following steps:
[0051] S100 data acquisition collects feedback signal sample data of the control rod drive mechanism (CRDM) operation, including three-phase current, vibration signals and other feedback signals.
[0052] S200 data preprocessing, grouping, data cleaning, normalization and other operations are performed on the collected feedback signal sample data; noisy data is prone to appear in the process of data collection, and it is necessary to remove outliers and noise in the collected samples to avoid affecting the subsequent process.
[0053] S300 feature extraction performs feature extraction in the time domain and frequency domain on the preprocessed feedback signal (three-phase current and vibration signal) to obtain a feature data set; the features in the time domain reflect the time and signal amplitude; the fast Fourier transform (FFT) is used in the frequency domain to extract features, which reflects information such as the amplitude and phase of each frequency band.
[0054] Common time domain features include mean, peak, root mean square, skewness, kurtosis, etc. When extracting time domain features, they can be divided into two categories: dimensional and dimensionless. The dimensional time domain features are related to the working conditions of the CRDM system motor, such as the speed and load. For example, the mean reflects the concentration of the signal, which reflects the stable component in the signal. The peak value refers to the maximum value of the signal over a period of time. The peak value of the current signal can reflect the fluctuation of the load. The dimensionless time domain features are independent of the state of the CRDM system during operation, and they are more sensitive to early stage faults in the equipment. For example, skewness and kurtosis are used to describe the degree to which the signal deviates from the normal distribution. When local defects occur in mechanical links such as bearings, the skewness and kurtosis will change significantly.
[0055] Some faults may be limited when only observed from the time domain characteristics. In this case, they can be analyzed from the perspective of the frequency domain. For example, a stuck rod fault will inevitably cause the frequency of the current and vibration signals to change. Compared with the signal in the time domain, the change trend of the vibration signal is relatively obvious. In the spectrum of the current and vibration signals, the proportion and energy of each frequency component will be revealed. The spectrum under normal conditions is manifested as some stable characteristic frequencies without obvious high-frequency noise, but the amplitude of the noise frequency in the spectrum under a stuck rod fault will increase significantly. Therefore, the change in frequency characteristics can also reflect to a certain extent whether a fault has occurred in the CRDM system. The spectrum under normal operation and various fault conditions can be seen. Figure 1 There must be differences. Therefore, for some faults, frequency domain analysis can be used. The time domain information of the current signal and vibration signal collected by the sensor is converted into frequency domain information through fast Fourier transform (FFT):
[0056] S=[s(1),s(2),s(3),...,s(k)]
[0057] Where k represents the length of the spectrum signal, and each coefficient contains the frequency, phase and amplitude information of each frequency band.
[0058] After feature extraction is completed, the extracted features are normalized first, and then feature selection is performed.
[0059] S400 feature selection is a process of evaluating and selecting the most relevant features in a data set to improve model performance and reduce computational complexity. This embodiment uses reliefF to sort the features in a feature data set and select the most relevant features. In order to ensure that the number of selected features can optimize classification performance, this embodiment uses the reliefF algorithm. As a distance-based feature sorting algorithm, the idea is to evaluate the importance of each feature by weighted calculation of the distance between adjacent sample points between different categories. The importance of the feature is calculated by calculating the contribution of each feature to the distance between different categories. The features that contribute more to the distance are relatively more important. By using the ReliefF algorithm for feature sorting, the most discriminative features can be screened out more effectively, and the number of features can be gradually increased as the input of the classifier, and the optimal number of features can be determined by the classification results fed back by the classifier.
[0060] In summary:
[0061] (1) Select an instance from a random category, then select a point from the same category that is closest to the instance, and then select a point from another category that is closest to the instance.
[0062] (2) Calculate the distance between intra-class points and inter-class points, assign different weights based on the intra-class distance and inter-class distance, and update the weight of each feature. If the feature makes the intra-class distance smaller and the inter-class distance larger, it proves that the feature is conducive to classification, and the weight of the feature increases; otherwise, it is not conducive to classification, and its weight decreases.
[0063] (3) The feature with the highest score is selected as the most important feature, and finally a feature vector from large to small based on weights is obtained.
[0064] Specifically, they are as follows:
[0065] Please refer to Figure 2 , Figure 2 The reliefF algorithm flow adopted by the present invention is demonstrated.
[0066] S410 inputs a feature data set, where the number of samples is n and the number of features is k;
[0067] S420 initializes the weight vector W to 0;
[0068] S430 sets parameter t (number of iterations);
[0069] S440 randomly selects an instance, namely, sample Xi;
[0070] S450 finds the nearest neighboring samples of Xi, namely the same type samples Ti and the different type samples Yi;
[0071] S460 updates the weight vector W;
[0072] S470 determines whether the loop is finished, if so, executes the next step, if not, returns to step S440;
[0073] S480 outputs the feature weight W.
[0074] In this embodiment, 60 features of three-phase current signals and vibration signals are extracted in the feature extraction stage. After normalization, they are sorted using the reliefF algorithm. The results are as follows: Figure 3 As shown. Figure 3 The information in the figure shows the weight of each eigenvalue. It can be seen that the features with relatively high discrimination are the mean value of the vibration signal, the maximum value of the vibration signal, the root mean square value of the vibration signal, and the absolute mean value of the vibration signal. The mean value of the vibration signal represents the stable component and average amplitude of the vibration signal, and the absolute mean value of the vibration signal represents the average amplitude of the overall deviation of the vibration signal from zero. They are directly related to the vibration amplitude of the CRDM system. Different faults cause different vibration effects, so they can be used for differentiation. The maximum value of the vibration signal reflects whether the system has suffered serious impact and failure, and the root mean square value of the vibration signal reflects the energy size and vibration intensity of the vibration signal. These signals can reflect the current operating status of the CRDM system and provide information for distinguishing faults.
[0075] The data in the data set are divided into normal data and fault data. The fault data are further divided into stuck data, stalled data and idling data. For visualization, this embodiment also draws and displays a scatter plot in the reselected feature dimension. This embodiment selects the first three most important features as axes. It can be seen from the distribution of the scatter plots of each category that the selected optimal features can clearly distinguish the four categories of data, which provides verification for the feasibility of subsequent fault classification. In order to verify that the features selected by the reliefF algorithm are helpful for classification, this embodiment makes a comparison group, randomly selects 3 features as axes, and compares the two to observe the distribution of scatter points of different categories. Figure 4 and Figure 5 It can be seen that it is proved that the features selected after feature sorting are more helpful for fault classification.
[0076] S500 fault diagnosis model training specifically includes the following sub-steps:
[0077] The S510 support vector machine model (SVM) uses a radial basis function kernel (RBF kernel) to determine the parameters of the model during the model training process;
[0078] S520 sets the parameters of the SVM classifier, mainly including the penalty coefficient c and the kernel function parameter g;
[0079] S530 divides the CRDM data set obtained after feature selection into a training set and a test set, and performs model training and final result prediction respectively.
[0080] The reason for using SVM is that the data set after reliefF feature selection effectively removes redundant features and screens out the most distinguishable features in the feature space. The purpose of SVM is to find the classification surface that maximizes the category interval in the feature space, so that the classifier can find a simpler and clearer classification surface.
[0081] The number of S600 features is determined. By gradually increasing the number of features, the classification accuracy of the support vector machine model is observed until the classification performance no longer increases. Due to the particularity of the nuclear power industry, normal data often outnumbers fault data, so there is a slight imbalance in the data set. The number of normal data samples is 1027, while the number of data samples for sticking, stalling, and idling are 201, 204, and 279, respectively. This will result in a high overall classification accuracy, but in fact the fault classification accuracy is low, resulting in the classification accuracy not being able to fully reflect the classification performance of the classifier (the trained SVM support vector machine model). Therefore, the evaluation indicators of the classifier performance use the F-measure and the geometric mean (G-mean), where the F-measure value can well reflect the recognition rate of a certain type of sample, and the G-mean reflects the recognition of positive and negative samples at the same time, which is very suitable as an evaluation indicator for unbalanced data sets. In order to reduce the impact of chance on the experiment and enhance the reliability of the evaluation results, this embodiment conducts 20 experiments under each different input, and takes the arithmetic average of the F-measure and G-mean of each category to evaluate the overall classification performance.
[0082] As shown in Table 1 below, when the number of features increases gradually from 1, the classification performance gradually improves until the number of features reaches 3, the classification effect no longer changes and the classification effect reaches the best, because the F-measure and G-mean of all categories are 1, proving that no samples are misclassified during classification. Therefore, in the end, this embodiment selects the first three features in the feature ranking as the input of the SVM.
[0083] Table 1 Classification performance under different feature numbers
[0084]
[0085] From the above results, we can see that when the number of features is selected as 3, the final classification result has reached the best effect, and there are no samples that are misclassified. First, from the feature dimension visualization diagram, it can be observed that there is no cross-over and aliasing between all categories in the three-dimensional space, and they are all linearly separable, so the categories are relatively easy to distinguish; second, the imbalance between the categories is not too obvious. In the collected CRDM data set, the number of normal data samples is 1027, while the number of data samples for sticking, stalling, and idling are 201, 204, and 279, respectively. Therefore, the maximum imbalance between the categories is only about 5:1, and there is almost no data imbalance problem between the three faults, so it has little impact on the results in the final classification.
[0086] It is worth noting that the more features are selected, the more accurate the results are. In the process of gradually increasing the number of features, it can be found that the classification performance has not been increasing or maintained at the optimal state. As shown in Table 2 below, when the number of features is increased to 23, the classification performance begins to decline until the number of features increases to 60. The classification performance has been fluctuating, but it has never reached the optimal state, as shown in the following table. Analyzing the reasons, one is because of data overfitting. When the amount of data is too large, it causes the distribution of the training set data to be overfitted, resulting in poor fitting effect on the test set; the second is because of the interference of related features. In the feature sorting stage, a weight value is obtained from large to small, so the features ranked later can be understood as having a negative impact on the classification, resulting in a decrease in classification performance after the increase in feature input.
[0087] Table 2 Changes in classification performance after gradually increasing the number of features
[0088]
[0089] After confirming the optimal number of input features, the classification effect is verified on the test set. The classification results are as follows: Figure 6 and Figure 7 As shown. Among them, Figure 6 It is a single classification result prediction graph. Figure 7 It is a box plot of the evaluation index of the 20 classification results. From the results, we can see that the CRDM fault diagnosis effect is very good after the reliefF feature dimension reduction.
[0090] S700 fault diagnosis, inputs subsequent data into the trained fault diagnosis model for fault diagnosis.
[0091] Based on the above embodiments, the present invention proposes a CRDM fault diagnosis method based on ReliefF feature selection. Through data collection, data preprocessing, feature extraction, feature selection, fault diagnosis model training, feature quantity determination and fault diagnosis, the ReliefF algorithm is used to sort the features, combined with time domain and frequency domain feature extraction, and a support vector machine model (SVM) is used to train and optimize the fault diagnosis model, and finally efficient and accurate CRDM fault diagnosis is achieved, the effectiveness of feature selection is improved, the shortcomings of the PCA unsupervised learning method are improved, and the problem of LDA feature quantity determination is solved. Each embodiment in this specification focuses on the differences from other embodiments, and the same and similar parts between the embodiments can be referred to each other.
[0092] The above is a detailed introduction to a control rod drive mechanism fault diagnosis method based on reliefF feature selection provided by the present application. This article uses specific examples to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present application.
Claims
1. A control rod drive mechanism fault diagnosis method based on reliefF feature selection, characterized in that: The steps include: S100 data acquisition, collecting feedback signal sample data of the control rod drive mechanism operation; S200 data preprocessing, grouping, data cleaning and normalization of the collected feedback signal sample data; S300 feature extraction, extracting features in the time domain and frequency domain of the preprocessed feedback signal to obtain a feature data set; S400 feature selection, using reliefF to sort the features in the feature data set and select the most relevant features to obtain a data set containing only the most relevant features; S500 fault diagnosis model training, using support vector machine model and training the model based on the data set after feature selection, and determining the model parameters in the process of training the model; The number of S600 features is determined. The classification accuracy of the support vector machine model is observed by gradually increasing the number of features until the classification performance stops increasing. S700 fault diagnosis, inputs subsequent data into the trained fault diagnosis model for fault diagnosis.
2. The control rod drive mechanism fault diagnosis method based on reliefF feature selection according to claim 1, characterized in that: In the data collection step S100, the feedback signal includes a three-phase current and a vibration signal.
3. The control rod drive mechanism fault diagnosis method based on reliefF feature selection according to claim 1, characterized in that: The data preprocessing step S200 includes removing outliers and noise from the collected sample data.
4. The control rod drive mechanism fault diagnosis method based on reliefF feature selection according to claim 1, characterized in that: In the S300 feature extraction step, the time domain features are extracted and divided into two categories: dimensional and dimensionless. The dimensional time domain features are at least related to the motor speed and the working state of the load when the control rod drive mechanism is in operation; the dimensionless time domain features are irrelevant to the working state of the control rod drive mechanism when it is in operation.
5. The control rod drive mechanism fault diagnosis method based on reliefF feature selection according to claim 1, characterized in that: In the feature extraction step S300, the time domain features reflect the time and signal amplitude, including mean, peak, root mean square, skewness and kurtosis.
6. The control rod drive mechanism fault diagnosis method based on reliefF feature selection according to claim 1, characterized in that: In the feature extraction step S300, fast Fourier transform is used in the frequency domain to extract features, and the frequency domain features at least reflect the amplitude and phase information of each frequency band.
7. The control rod drive mechanism fault diagnosis method based on reliefF feature selection according to claim 1, characterized in that: The feature selection step S400 specifically includes the following sub-steps: S410 selects an instance from a random category, selects a point of the same category that is closest to the instance, and then selects a point of a different category that is closest to the instance; The distance between points of the same type is calculated, which is called the intra-class point distance; the distance between points of different types is calculated, which is called the inter-class point distance. The weight of each feature is updated according to the different weights of the intra-class point distance and the inter-class point distance. The feature with the highest score is selected as the most important feature, and finally a feature vector is obtained from large to small based on the weight.
8. The control rod drive mechanism fault diagnosis method based on reliefF feature selection according to claim 1, characterized in that: The S500 fault diagnosis model training step specifically includes the following sub-steps: The S510 support vector machine model uses a radial basis function kernel; S520 sets the hyper parameters of the SVM classifier; S530 divides the CRDM data set after feature selection into a training set and a test set, and performs model training and final result prediction respectively.
9. The control rod drive mechanism fault diagnosis method based on reliefF feature selection according to claim 1, characterized in that: In the step of determining the number of features in S600, the evaluation indicators of the classification accuracy of the support vector machine model are F-measure and G-mean.
10. The control rod drive mechanism fault diagnosis method based on reliefF feature selection according to claim 1, characterized in that: In the step of determining the number of features in S600 , the number of features is selected in a range of 3 to 22.
Citation Information
Patent Citations
Pressurized water reactor nuclear power station control bar driving mechanism on-line monitoring and fault diagnosing method
CN101067976A
High-dimensional unbalanced data classification method based on mixed sampling and feature selection
CN111695626A
Photovoltaic fault arc feature selection method combining filtering type evaluation strategy and packaging type evaluation strategy
CN113076986A
