A Fault Classification Method Based on Weighted Probabilistic Neural Network

By using a weighted probabilistic neural network for fault classification, the complexity of fault identification in power plant equipment is solved, and real-time fault classification of power plant equipment under all operating conditions is achieved, thereby improving the safety and economy of the power plant.

CN115017978BActive Publication Date: 2024-11-26SOUTHEAST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210543721.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-05-18
Publication Date
2024-11-26
Estimated Expiration
2042-05-18

AI Technical Summary

Technical Problem

In the context of Industry 4.0, complex and untimely handling of power plant equipment failures can lead to accidents. Traditional methods cannot quickly and accurately identify the type of failure, affecting the safety and economy of the unit.

Method used

A fault classification method based on weighted probabilistic neural networks is adopted. By obtaining fault training sample sets and test sample sets, dimensionality reduction and feature extraction are performed. The reconstructed principal component analysis method and weighted probabilistic neural network are used for training to achieve accurate fault classification.

Benefits of technology

It enables real-time fault classification of power plant equipment under all operating conditions, quickly resolves fault problems, improves the economic efficiency and safety of power plant operation, can handle known and new faults, has fast learning speed and high computational efficiency, and high classification accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115017978B_ABST
    Figure CN115017978B_ABST
Patent Text Reader

Abstract

The present invention discloses a fault classification method based on a weighted probabilistic neural network, which relates to the technical field of machine learning and solves the technical problem that existing power station equipment cannot accurately classify faults. The key point of the technical solution is to measure the class separability of samples by adding weight factors in the pattern layer and summation layer of the weighted probabilistic neural network, so as to further provide information on new faults for fault classification decisions. It can effectively handle known faults and new faults, and realize fault classification under full-operation conditions of power station equipment. At the same time, for new faults, new fault samples can be easily added to the trained network, and only the corresponding hidden layer units need to be added without retraining, thereby improving the economy and safety of power plant operation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of machine learning, and in particular to a fault classification method based on a weighted probabilistic neural network. Background Art

[0002] In the context of "Industry 4.0", power plants are equipped with a large number of instruments to capture the values ​​of hundreds of variables at every moment, forming a huge data set. When a device fails, the abnormal state cannot be quickly and accurately handled only by expert experience due to the complexity and diversity of the information in the data. If the fault is not handled in time, the further development of the fault state will lead to industrial accidents, resulting in casualties and serious economic losses, and even cause unit shutdown. Therefore, a set of intelligent fault classification tools is needed to help determine the fault information and take correct measures to ensure the safety and economy of the unit.

[0003] Fault classification is based on equipment fault detection. According to the characteristics and changes of the operating parameters when the equipment fails, combined with the working principle and thermal performance of the equipment, the fault is analyzed from the data, and fault variable information is provided to quickly and accurately identify the fault category, so as to facilitate the determination of the fault solution. As power plant equipment continues to become larger and more complex, different faults often couple multiple identical fault characteristics. Traditional processing methods cannot accurately identify the fault category. Therefore, accurate fault classification is of great significance for quickly solving fault problems and ensuring safe and stable operation of the unit. Summary of the invention

[0004] The present application provides a fault classification method based on a weighted probabilistic neural network, the technical purpose of which is to accurately classify faults and realize real-time classification of faults under full operating conditions of power station equipment, thereby quickly solving fault problems.

[0005] The above technical objectives of this application are achieved through the following technical solutions:

[0006] A fault classification method based on weighted probabilistic neural network, comprising:

[0007] S1: Based on the historical fault data of power plant equipment, the training sample set and the test sample set of faults are obtained. The training sample set constitutes a sample matrix, and the sample matrix is ​​reduced in dimension to obtain a score matrix and a load matrix;

[0008] S2: The reconstruction-based principal component analysis method is used to extract the fault features in the score matrix and the load matrix, and the extracted fault features are used as input samples of the weighted probability neural network;

[0009] S3: training the weighted probability neural network using the fault characteristics to obtain a weighted probability neural network model;

[0010] S4: Input the test sample set into the trained weighted probability neural network model to obtain the fault classification result.

[0011] The beneficial effects of this application are:

[0012] (1) The system and method provided by the present application have a fast learning speed; the learning rules of the selected method during the training process are simple, the calculation speed is fast, and the training process is completed in one go.

[0013] (2) On the basis of the traditional probabilistic neural network based on Bayesian decision, weight factors are added to the pattern layer and summation layer of the probabilistic neural network to measure the class separability of samples, which is used to further provide information about new faults for fault classification decisions.

[0014] (3) The weighted probabilistic neural network classification model provided by this application can effectively handle known faults and new faults, and can realize fault classification under all operating conditions of power station equipment. At the same time, for new faults, new fault samples can be easily added to the trained network, and only the corresponding hidden layer units need to be added without retraining, which improves the economy and safety of power plant operation.

[0015] (4) The fault classification module in this application classifies samples based on the Bayesian minimum risk criterion, which can make maximum use of prior knowledge. No matter how complex the classification problem is, as long as there are sufficient training samples, the optimal solution under the Bayesian criterion can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of the method described in this application;

[0017] Figure 2 It is a structural diagram of weighted probabilistic neural network;

[0018] Figure 3 This is a schematic diagram of the weight analysis results of known faults;

[0019] Figure 4 Schematic diagram of the weight analysis results of the new fault. DETAILED DESCRIPTION

[0020] The technical solution of the present application will be described in detail below with reference to the accompanying drawings.

[0021] As an important equipment of the steam turbine unit in a power plant, the condenser has an extremely important impact on the vacuum index of the unit. This application takes the condenser system of a power plant as the research object and analyzes its failures under different working conditions. First, the data of the condenser is collected in the SIS system of the power plant, and its operating data under different working conditions and offline historical fault data are obtained. The fault data is marked accordingly as training data for the weighted probability neural network model.

[0022] Table 1 Fault characteristic parameters of condenser system

[0023] serial number name serial number name <![CDATA[F1]]> Condenser temperature <![CDATA[F 14 ]]> Circulating water pump power <![CDATA[F2]]> Condenser pressure <![CDATA[F 15 ]]> Aspirator inlet mass flow <![CDATA[F3]]> Condenser level <![CDATA[F 16 ]]> Mixed pressure at the outlet of the vacuum pump <![CDATA[F4]]> Steam turbine exhaust mass flow rate <![CDATA[F 17 ]]> Mixed temperature at the outlet of the vacuum pump <![CDATA[F5]]> Terminal difference <![CDATA[F 18 ]]> Condenser system mixed pressure drop <![CDATA[F6]]> Subcooling <![CDATA[F 19 ]]> Mixed temperature rise of condenser system <![CDATA[F7]]> Circulating cooling water mass flow <![CDATA[F 20 ]]> Additional heat source <![CDATA[F8]]> Circulating cooling water inlet temperature <![CDATA[F 21 ]]> Air leakage <![CDATA[F9]]> Circulating cooling water outlet temperature <![CDATA[F 22 ]]> Vacuum extractor valve position <![CDATA[F 10 ]]> Circulating cooling water inlet pressure <![CDATA[F 23 ]]> Circulating cooling water control valve position <![CDATA[F 11 ]]> Circulating cooling water outlet pressure <![CDATA[F 24 ]]> Relative roughness of the tube <![CDATA[F 12 ]]> Circulating cooling water temperature rise <![CDATA[F 25 ]]> Condensate pump outlet pressure <![CDATA[F 13 ]]> Circulating cooling water pressure drop <![CDATA[F 26 ]]> Circulating water pump outlet pressure

[0024] When a condenser system fails, all operating parameters of the system will be affected to varying degrees. These parameters include real-time operating data obtained by sensors on site and performance indicators obtained based on thermal calculations. The characteristic parameters used for condenser system fault classification are shown in Table 1 and are used for fault feature extraction and fault information identification.

[0025] like Figure 1 As shown, the fault classification method using weighted probability neural network includes the following steps:

[0026] S1: Based on the historical fault data of power plant equipment, the training sample set and the test sample set of faults are obtained. The training sample set constitutes a sample matrix, and the sample matrix is ​​reduced in dimension to obtain the score matrix and the load matrix.

[0027] Specifically, the fault data of the field equipment is collected by sensors, combined with the fault case library data, extracted operation experience and fault simulation rules, to obtain the fault data of the power station equipment, which is used to construct the training sample set of the classifier and obtain the original observation matrix X. Where N represents the total number of samples, and each sample includes m observation variables, then the original observation matrix X is expressed as:

[0028]

[0029] The original observation matrix X is subjected to dimensionality reduction processing, including:

[0030] (1) Each column of the original observation matrix X is processed with zero mean and unit variance, and the covariance matrix is ​​expressed as:

[0031] Through eigendecomposition, the covariance matrix can be further decomposed into:

[0032]

[0033] (2) Then a sample vector x can be projected into the principal component subspace and the residual subspace respectively, expressed as:

[0034]

[0035] (3) According to the projection of the sample vector x, the original observation matrix X is decomposed into the score matrix T and the load matrix P; in the original observation matrix X, each row represents a sample, and the observation variable x i =[x i (1),xi (2),...,x i (N)]; l represents the number of principal components; i=1,2,...,m.

[0036] S2: The reconstruction-based principal component analysis method is used to extract the fault features in the score matrix and the load matrix, and the extracted fault features are used as input samples of the weighted probabilistic neural network.

[0037] In order to successfully identify the fault characteristics, the indicator SPE reflecting the residual space is used to calculate the observed variable x i The reconstruction contribution value As fault features, then based on the reconstruction contribution value of each sample variable Constitute a fault characteristic data group.

[0038]

[0039]

[0040] This method can indicate the real fault variable, and the fault feature extraction basis is: for the fault variable, the SPE index of the fault variable will have the largest decrease after reconstruction. According to the response degree of each variable after reconstruction, the variable with the largest response is used as the main feature, and this feature and the reconstructed data group are used as the extracted features for the fault category identification here.

[0041] In contribution analysis, the observed variable x i Contribution to the SPE indicator It is expressed as:

[0042]

[0043]

[0044]

[0045] in, represents the i-th column in the identity matrix, with the same direction as x i same; represents the projection matrix; Represents the projection matrix The i-th diagonal element of .

[0046] S3: Training the weighted probability neural network using the fault characteristics to obtain a weighted probability neural network model.

[0047] The weighted probabilistic neural network of the machine learning algorithm is introduced, and the equipment is modeled in detail in combination with prior knowledge. The relationship between the features and the fault data type is described to classify and identify the faults. Class separability represents the ability to distinguish various categories in the feature vector, which is usually used as one of the performance criteria in the classification algorithm, that is, a set of feature vectors can achieve maximum class separability through the classifier. However, the traditional probabilistic neural network assigns the same weight to all pattern layers, which cannot meet the maximization of class separability of different pattern layers. The weighted probabilistic neural network is based on a sound statistical principle, derived from Bayesian decision and non-parametric kernel density estimator, rather than a heuristic method. Due to the existence of weight factors, the summation layer nodes are related to each node of the pattern layer. By selecting appropriate smoothing factor values, the classification results can be guaranteed to be close to the Bayesian optimal decision. The introduction and update of weight factors can improve the classification accuracy of the classifier. The network structure diagram of the weighted probabilistic neural network is shown in the figure below. Figure 2 shown.

[0048] The weighted probabilistic neural network includes a sample input layer, a pattern layer, a summation layer, and a decision output layer which are connected in sequence.

[0049] Sample input layer: input the data to be classified, the number of neurons is the dimension of the sample data after dimensionality reduction; the sample layer is associated with the pattern layer through the Gaussian function.

[0050] Pattern layer: that is, radial base layer, matches the feature vector of input data with each pattern, and calculates the distance between the input sample vector and the center vector of each neuron. The relationship between the sample vector x and the jth center vector of the i-th class is expressed as:

[0051]

[0052] Among them, σ represents the smoothing factor; υ ij Represents the neuron vector x ij The ratio of the “between-class variance” to the “within-class variance”. The weight factor takes into account the class separability of different modes. For high class separability (meaning better class discrimination ability), υ ij The ratio is large; for low class separability, υ ij The ratio is small. x=(x1,x2,...,x m );N i Indicates category C i The total number of samples in .

[0053] Summation layer: The number of summation nodes is equal to the category of the training samples. Neurons calculate the sum of the outputs of pattern nodes of the same category and then take a weighted average.

[0054] Decision output layer: Based on Bayesian classification decision, it determines the category with the maximum posterior probability and outputs the fault category.

[0055] This application derives an analytical formula for the weight factor of the weighted probability neural network based on sensitivity analysis. The weight factor is used to reflect the importance of the qth pattern to the jth category between the pattern layer and the summation layer. The calculation process of the weight factor includes:

[0056] Calculate the sensitivity coefficient S of the neuron in all modes, expressed as:

[0057]

[0058] in, represents the vector parameter of the jth class, then the sensitivity parameter S of the jth class is expressed as:

[0059]

[0060] Among them, S j The elements of the rth column represent: for a specific input pattern q, the calculation of the sensitivity parameters of the kernel density estimation function of the rth neuron corresponding to the jth category in the weighted probability neural network. j The denominator of each term in is a vector, so the gradient needs to be determined as follows:

[0061]

[0062] Aggregate all patterns q and obtain sensitive vectors, q = 1,...,Q j , then the sensitive vector a j It is expressed as:

[0063]

[0064] A j Normalizing the elements in the jth category gives the weight ω of each element in the jth category. j , expressed as:

[0065]

[0066] Among them, when n=∞, Then j Normalized to the interval [0,1], the summation layer of the weighted probability neural network is expressed as:

[0067]

[0068]

[0069] Among them, r represents a j The rth element in the equation, n1, n2 ≥ 1; when calculating the weight factor of the weighted probability neural network, n1 = ∞, n2 = 2

[0070] S4: Input the test sample set into the trained weighted probability neural network model to obtain the fault classification result.

[0071] Specifically, real-time fault data is input into a trained weighted probabilistic neural network model, which then outputs the fault category to determine the root cause of the fault, and the classification result is transmitted to the DCS control system.

[0072] As a specific embodiment, this application uses 70% of the data set to construct training samples, denoted as Tr1, and uses 30% of the data set to construct test samples, denoted as Te1. The principal component analysis method based on reconstruction is applied to extract features, and fault feature samples are obtained as training samples for the classifier, which are input into the weighted probability neural network classifier to output the classification results.

[0073] In order to illustrate the necessity and effectiveness of the weighted probability neural network method, an observation value in the test set Te1 is taken as an example. First, the fault characteristics of the sample are obtained based on the reconstruction principal component analysis method. Then, the fault characteristics of the sample are input into the weighted probability neural network classifier, and the probability values ​​belonging to each category are output. The category with the maximum probability is the predicted fault category.

[0074] The results show that this method can correctly identify the fault state and assign the highest probability value (1.0) to fault 3, as shown in Table 2. This fault corresponds to the abnormal operation of the vacuum pump. Through the fault extraction process, according to the response degree of the variable contribution value, it is known that F 16 The main feature can determine the root cause of the fault. Based on the above fault information, a solution for solving the fault is provided.

[0075] Table 2 Calculation of category probability based on PNN

[0076] Fault category 0 1 2 3 4 5 Probability value 0 0 0 1 0 0

[0077] In order to verify the accuracy of weighted probability neural network classification (WPNN), it is compared with traditional probabilistic neural network (PNN) and five other common classification models, including support vector machine (SVM), random forest (RF), decision tree (DT), extreme random tree (ET), and K nearest neighbor method (KNN). All classification models are trained by dataset Tr1 and tested by dataset Te1. The accuracy of the classification results is shown in Table 3.

[0078] Table 3 Classification accuracy of six classification methods

[0079] WPNN PNN SVM RF DT ET KNN Classification Accuracy 99.8% 98.93% 95.48% 90.38% 89.61% 92.45% 96.98%

[0080] As can be seen from Table 3, the comparison results of the accuracy of 7 different classifiers show that both the traditional probabilistic neural network method and the weighted probabilistic neural network method have good classification performance in the fault classification module and are superior to other types of artificial neural networks and classifiers. The weighted probabilistic neural network method can train the best classifier because the weighted probabilistic neural network method basically has the good performance of the traditional probabilistic neural network and can approximate the probability density function of the multivariate Gaussian function. In addition, the weighted probabilistic neural network also includes a weighting factor that characterizes the separability of the class, and the contribution of each pattern neuron to the network result is accessible.

[0081] New fault classification: In order to verify the effect of weighted probability neural network method on new fault classification, the fifth type of fault samples are removed from the original training set Tr1, and a new training data set Tr2 is constructed, so that fault 5 can be regarded as a new type of fault that does not exist in the training set. In the test set Te1, a sample of the fifth type of fault is used as a test sample.

[0082] Table 4 Calculation of category probability based on WPNN

[0083]

[0084]

[0085] As shown in Table 4, the probability of being assigned to fault 3 is only 0.163, which cannot be directly used as the basis for fault category judgment. Therefore, it is necessary to further determine the fault category based on the weight factor calculation result. The weight factor is calculated by the weighted probability neural network, and the weight factor proportion of the third type of fault sample is obtained. Figure 3 As shown, the proportion of the third type of fault weight factor of the fault test sample is Figure 4 As shown in the figure, the weight factor of the normal third-category fault sample is 89.45%, while the weight of the third-category fault test sample is only 34.04%. The test sample weight calculation results do not show obvious class separability, so the test sample is considered a new type of fault.

[0086] In the face of new faults, it is necessary to combine the information of fault variable isolation in the fault feature extraction process, that is, the condenser water level corresponding to the F3 variable is the main fault feature, F 25 、F 13 It is a secondary fault feature, so the condenser water-filled fault is considered first. However, if it is a simple condenser water-filled fault, the condenser pump outlet pressure variable will not have a large response in the reconstruction analysis, so the condenser water-filled can only be considered as a fault symptom, not the main fault feature. Select F with the largest secondary response amplitude 25 The variable is used as the main feature of the fault. Combined with the condenser water level full sign and the pressure change of the condenser system, it can be determined that the new fault is a condensate pump fault.

[0087] The above are exemplary embodiments of the present application, and the protection scope of the present application is defined by the claims and their equivalents.

Claims

1. A fault classification method based on weighted probabilistic neural network, characterized in that: include: S1: Based on the historical fault data of power plant equipment, the training sample set and the test sample set of faults are obtained. The training sample set constitutes a sample matrix, and the sample matrix is ​​reduced in dimension to obtain a score matrix and a load matrix; S2: The reconstruction-based principal component analysis method is used to extract the fault features in the score matrix and the load matrix, and the extracted fault features are used as input samples of the weighted probability neural network; S3: training the weighted probability neural network using the fault characteristics to obtain a weighted probability neural network model; S4: Input the test sample set into the trained weighted probability neural network model to obtain the fault classification result; In step S1, a training sample set of a classifier is constructed to obtain the original observation matrix X. Where N represents the total number of samples, and each sample includes m observation variables, then the original observation matrix X is expressed as: The original observation matrix X is reduced in dimension and decomposed into the score matrix T and the load matrix P. In the original observation matrix X, each row represents a sample, and the observed variable x i =[x i (1),x i (2),...,x i (N)]; l represents the number of principal components; i = 1, 2, ..., m; In step S2, by calculating the observed variable x i The reconstruction contribution value Obtain fault characteristics and then reconstruct contribution values ​​based on each sample variable Constructing a fault feature data set, including: i Contribution to the SPE indicator Get reconstruction contribution value It is expressed as: in, represents the i-th column in the identity matrix, with the same direction as x i same; represents the projection matrix; Represents the projection matrix The i-th diagonal element of ; In step S2, the weighted probability neural network includes a sample input layer, a pattern layer, a summation layer, and a decision output layer connected in sequence; The pattern layer is used to match the feature vector of the input data with each pattern, and calculate the distance between the input sample and the central vector of each neuron. The relationship between the sample vector x and the jth central vector of the i-th class is expressed as: Among them, σ represents the smoothing factor; υ ij Represents the neuron vector x ij The ratio of the "between-class variance" to the "within-class variance" of x = (x1, x2, ..., x m );N i Indicates category C i The total number of samples in ; The weight factor of the weighted probability neural network is calculated, and the weight factor is used to reflect the importance of the qth pattern to the jth category between the pattern layer and the summation layer. The calculation process of the weight factor includes: Calculate the sensitivity coefficient S of the neuron in all modes, expressed as: in, represents the vector parameter of the jth class, then the sensitivity parameter S of the jth class is expressed as: Among them, S j The elements of the rth column represent: for the input pattern q, the calculation of the sensitivity parameter of the kernel density estimation function of the rth neuron corresponding to the jth category in the weighted probability neural network; Aggregate all patterns q and obtain sensitive vectors, q = 1,...,Q j , then the sensitive vector a j It is expressed as: A j Normalizing the elements in the jth category gives the weight ω of each element in the jth category. j , expressed as: Among them, when n=∞, Then j Normalized to the interval [0,1], the summation layer of the weighted probability neural network is expressed as: Among them, r represents a j For the rth element in , n1,n2≥1; when calculating the weight factor of the weighted probability neural network, n1=∞, n2=2.