CN115017978A - A Fault Classification Method Based on Weighted Probabilistic Neural Network - Google Patents

Abstract

The invention discloses a fault classification method based on a weighted probability neural network, relates to the technical field of machine learning, and solves the technical problem that existing power station equipment cannot accurately classify faults. A weight factor is added to the pattern layer and summation layer to measure the class separability of the samples, which is used to further provide information of new types of faults for fault classification decisions. It can effectively deal with known faults and new faults, and 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, only need to add corresponding hidden layer units, no need to retrain, improve the economy and safety of power plant operation.

Description

A Fault Classification Method Based on Weighted Probabilistic Neural Network

technical field

The present application relates to the technical field of machine learning, and in particular, to a fault classification method based on a weighted probability neural network.

Background technique

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 huge data sets. When the equipment fails, due to the complex and diverse information in the data, it is impossible to quickly and accurately handle the abnormal state only through expert experience. If the fault treatment is not timely, the further development of the fault state will lead to industrial accidents, resulting in casualties and serious economic losses, and even lead to the shutdown of the unit. Therefore, a set of intelligent fault classification tools is needed to help determine fault information and take corrective measures to ensure the safety and economy of the unit.

Fault classification is based on equipment fault detection. According to the characteristics and changes of operating parameters when the equipment fails, combined with the working principle and thermal performance of the equipment, the fault is analyzed from the data, providing fault variable information, and quickly and accurately identifying the fault category. Easy to identify solutions to failures. With the increasing size and complexity of power station equipment, different faults are often coupled with multiple identical fault characteristics. Traditional processing methods cannot accurately identify the type of fault. Therefore, accurate fault classification is important for quickly solving fault problems and ensuring the safe and stable operation of units. have important significance.

SUMMARY OF THE INVENTION

The present application provides a fault classification method based on a weighted probability neural network, the technical purpose of which is to accurately classify faults, realize real-time fault classification under all operating conditions of power station equipment, and quickly solve fault problems.

The above-mentioned technical purpose of the present application is achieved through the following technical solutions:

A fault classification method based on weighted probability neural network, comprising:

S1: Obtain a training sample set and a test sample set of the fault based on the historical fault data of the power station equipment, the training sample set constitutes a sample matrix, and the sample matrix is dimensionally reduced to obtain a score matrix and a load matrix;

S2: Use the reconstruction-based principal component analysis method to extract the fault features in the score matrix and the load matrix, and use the extracted fault features as the input samples of the weighted probability neural network;

S3: train a weighted probability neural network through the fault feature to obtain a weighted probability neural network model;

S4: Input the test sample set into the trained weighted probability neural network model to obtain fault classification results.

The beneficial effects of this application are:

(1) The system and method provided by this application have a fast learning speed; the selected method has simple learning rules in the training process, fast calculation speed, and the training process is completed at one time.

(2) On the basis of the traditional probabilistic neural network based on Bayesian decision-making, a weight factor is added to the pattern layer and summation layer of the probabilistic neural network to measure the class separability of the sample, which is used to further provide information on new faults for fault classification decisions.

(3) The weighted probability neural network classification model provided by the present application can effectively deal with known faults and new faults, and both 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, only need to add corresponding hidden layer units, no need to retrain, improve the economy and safety of power plant operation.

(4) The fault classification module in this application classifies the samples based on the Bayesian minimum risk criterion, which can utilize prior knowledge to the greatest extent. No matter how complex the classification problem is, as long as there are sufficient training samples, it is guaranteed to obtain The optimal solution under the Bayesian criterion.

Description of drawings

Fig. 1 is a flow chart of the method described in this application;

Figure 2 is a structural diagram of a weighted probability neural network;

FIG. 3 is a schematic diagram of a weight analysis result of a known fault;

Figure 4 is a schematic diagram of the weight analysis result of the new fault.

Detailed ways

The technical solutions of the present application will be described in detail below with reference to the accompanying drawings.

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

Table 1 Condenser system fault characteristic parameters

Numbering name Numbering name F₁ condenser temperature F₁₄ Circulating water pump power F₂ condenser pressure F₁₅ Aspirator inlet mass flow F₃ Condenser level F₁₆ Aspirator outlet mixing pressure F₄ Steam turbine exhaust mass flow F₁₇ Aspirator outlet mixing temperature F₅ end difference F₁₈ Condenser system mixing pressure drop F₆ supercooling F₁₉ Condenser system mixing temperature rise F₇ Circulating cooling water mass flow F₂₀ Additional heat source F₈ Circulating cooling water inlet temperature F₂₁ air leakage F₉ Circulating cooling water outlet temperature F₂₂ Aspirator valve position F₁₀ Circulating cooling water inlet pressure F₂₃ Circulating cooling water control valve position F₁₁ Circulating cooling water outlet pressure F₂₄ Relative roughness inside the tube F₁₂ Circulating cooling water temperature rise F₂₅ Condensate pump outlet pressure F₁₃ Circulating cooling water pressure drop F₂₆ Circulating water pump outlet pressure

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

As shown in Figure 1, the fault classification method using weighted probability neural network, the implementation steps include:

S1: Obtain a training sample set and a test sample set of faults based on the historical fault data of the power station equipment. The training sample set constitutes a sample matrix, and the sample matrix is dimensionally reduced to obtain a score matrix and a load matrix.

其中，N表示样本总数，且每个样本包括m个观测变量，则原始观测矩阵X表示为：Specifically, the fault data of field equipment is collected by sensors, combined with fault case database data, operation experience extraction and fault simulation rules, to obtain fault data of power station equipment, which is used to construct the training sample set of the classifier, and the original observation matrix X is obtained,

Among them, N represents the total number of samples, and each sample includes m observation variables, then the original observation matrix X is expressed as:

Perform dimensionality reduction on the original observation matrix X, including:

(1) Perform zero mean and unit variance processing on each column of the original observation matrix X, and obtain the covariance matrix expressed as:

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

(2) Then a sample vector x can be projected to the principal subspace and the residual subspace, respectively, expressed as:

l表示主元个数；i＝1,2,...,m。(3) Decompose the original observation matrix X into a score matrix T and a load matrix P according to the projection of the above-mentioned sample vector x; wherein, in the original observation matrix X, each row represents a sample, and the observation variable x _i =[x _i (1) , _xi (2),..., _xi (N)];

l represents the number of pivots; i=1,2,...,m.

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 the input samples of the weighted probability neural network.

作为故障特征，然后基于各个样本变量的重构贡献值

构成故障特征数据组。In order to successfully identify the fault features, the index SPE, which reflects the residual space, is used to calculate the reconstruction contribution value of the observed variable x _i .

as fault features, and then based on the reconstructed contribution value of each sample variable

Constitute the fault characteristic data set.

This method can indicate the real fault variable, and the fault feature extraction is based on the following: 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 set are used as the extracted features for the classification of faults here.

表示为：In the contribution analysis, the contribution value of the observed variable x _i to the indicator SPE

Expressed as:

表示单位矩阵中第i列，方向与x_i相同；

表示投影矩阵；

表示投影矩阵

的第i个对角元素。in,

Represents the i-th column in the identity matrix, the direction is the same as x _i ;

represents the projection matrix;

represents the projection matrix

The ith diagonal element of .

S3: Train a weighted probabilistic neural network through the fault feature to obtain a weighted probabilistic neural network model.

The weighted probabilistic neural network of the machine learning algorithm is introduced, and the equipment is modeled in detail by combining the prior knowledge, and the relationship between the characteristics and the fault data type is described, and the fault is classified and identified. Class separability represents the ability to distinguish various classes in a feature vector, and is usually used as one of the performance criteria in classification algorithms, that is, a set of feature vectors can achieve maximum class separability through the classifier. However, traditional probabilistic neural networks assign the same weights to all model layers, which cannot maximize the class separability of different model layers. Weighted probabilistic neural networks are based on well-established statistical principles, derived from Bayesian decision-making and nonparametric kernel density estimators, rather than heuristics. Due to the existence of the weight factor, the summation layer node is related to each node of the pattern layer. By selecting the appropriate smoothing factor value, the classification result can be guaranteed to be close to the Bayesian optimal decision. The introduction and update of the weight factor can improve the classification of the classifier. precision. The network structure diagram of the weighted probability neural network is shown in Figure 2.

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

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

Pattern layer: the radial base layer, which matches the feature vector of the 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 j-th center vector of the i-th class is expressed as:

Among them, σ represents the smoothing factor; υ _ij represents the ratio of the "inter-class variance" and "intra-class variance" of the neuron vector x _ij , the weight factor takes into account the class separation of different modes, and for high class separability (meaning more Good class discrimination ability), the υ _ij ratio is larger; for low class separability, the υ _ij ratio is smaller. x=(x ₁ , x ₂ , . . . , x _m ); N _i represents the total number of samples in category C _i .

Summation layer: The number of summation nodes is equal to the category of training samples, neurons calculate the sum of the output of mode nodes of the same category, and then do a weighted average.

Decision output layer: Based on Bayesian classification decision, the category with the largest posterior probability is determined, and the category of fault is output.

This application derives the analytical formula of the weight factor of the weighted probability neural network based on the 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 of the weight factor The process includes:

Calculate the sensitivity coefficient S of neurons in all modes, expressed as:

表示第j类的向量参数，则第j类的敏感度参数S表示为：in,

represents the vector parameter of the jth class, then the sensitivity parameter S of the jth class is expressed as:

Among them, the element in the rth column of S _j represents: for a specific input mode 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. Since the denominator of each term in S _j is a vector, the gradient needs to be determined as follows:

Aggregate all patterns q and obtain the sensitive vector, q=1,...,Q _j , then the sensitive vector a _j is expressed as:

Normalize the elements in a _j to get the weight ω _j of each element in the jth category, which is expressed as:

则w_j被归一化为[0,1]区间，此时加权概率神经网络的求和层表示为：Among them, when n=∞,

Then w _j is normalized to the [0,1] interval, and the summation layer of the weighted probability neural network is expressed as:

Among them, r represents the rth element in a _j , n ₁ , n ₂ ≥1; when calculating the weighted probability neural network weight factor, n ₁ =∞, n ₂ =2

S4: Input the test sample set into the trained weighted probability neural network model to obtain fault classification results.

Specifically, the real-time fault data is input into the trained weighted probability neural network model, and then the fault category is output to determine the root cause of the fault, and the classification result is transmitted to the DCS control system.

As a specific example, in the present application, 70% of the data set used to construct training samples is denoted as Tr ₁ , and 30% of the data set used to construct test samples is denoted as Te ₁ . The reconstruction-based principal component analysis method is used for feature extraction, and the fault feature samples are obtained as the training samples of the classifier, which are input to the weighted probability neural network classifier, and the classification results are output.

For the necessity and effectiveness of the weighted probabilistic neural network approach, take an observation in the test set Te ₁ as an example. Firstly, the fault features of the sample are obtained based on the reconstructed principal component analysis method. Then, the fault features of the sample are input into the weighted probability neural network classifier, and the probability values belonging to each category are output, and the maximum probability category is the predicted fault category.

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

Table 2 Calculation of category probability based on PNN

fault category 0 1 2 3 4 5 probability value 0 0 0 1 0 0

To verify the accuracy of weighted probabilistic neural network (WPNN)-based classification, it was 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), K-Nearest Neighbor (KNN). All classification models are trained by dataset Tr ₁ and tested using dataset Te _1. The accuracy of the classification results is shown in Table 3.

Table 3 Classification accuracy of six classification methods

WPNN PNN SVM RF DT ET KNN Classification accuracy 99.8% 98.93% 95.48% 90.38% 89.61% 92.45% 96.98%

It can be seen from Table 3 that 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 better than other types of artificial neural networks and classifiers. The weighted probability neural network method can train the optimal classifier, because the weighted probability neural network method basically has the good performance of the traditional probability neural network, and can approximate the probability density function of the multivariate Gaussian function. In addition to this, weighted probabilistic neural networks include weighting factors that characterize class separability, accessible to the contribution of each pattern neuron to the network outcome.

New fault classification: In order to verify the effect of the weighted probability neural network method on the new fault classification, the fifth type of fault samples are removed from the original training set Tr ₁ , and a new training data set Tr ₂ is constructed, so fault 5 can be used as the training set. A new type of failure that does not exist. In the test set Te ₁ , a sample of the fifth type of failure is used as the test sample.

Table 4 Calculation of class probability based on WPNN

It can be seen from Table 4 that the probability of being assigned to fault 3 is only 0.163, which cannot be directly used as the basis for judging the fault type. Therefore, it is necessary to further determine the fault type according to the calculation result of the weight factor. Through the calculation of the weight factors through the weighted probability neural network, the proportion of the weight factors of the third type of fault samples is shown in Figure 3, and the proportion of the third type of fault weight factors of the fault test samples is shown in Figure 4. The weight factor of the normal third type of fault samples is 89.45%, while the weight of the third type of fault test samples is only 34.04%. The test sample weight calculation results do not show obvious class separability, so the test sample is considered as a novel fault.

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, and _F25 and _F13 are the secondary fault features, so the condenser is the first consideration. Full water failure. However, if the condenser is simply full of water failure, the pressure variable at the outlet of the condensate pump will not have a relatively large response in the reconstruction analysis, so the condenser full of water can only be considered as a symptom of the failure, not the main feature of the failure. The F ₂₅ variable with the largest secondary response amplitude is selected as the main fault feature. Combined with the sign of the condenser water level full of water and the pressure change of the condenser system, it can be determined that the new type of fault is the fault of the condensate pump.

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 (5)

1. a fault classification method based on weighted probability neural network, is characterized in that, comprises: S1: Obtain a training sample set and a test sample set of the fault based on the historical fault data of the power station equipment, the training sample set constitutes a sample matrix, and the sample matrix is dimensionally reduced to obtain a score matrix and a load matrix; S2: Use the reconstruction-based principal component analysis method to extract the fault features in the score matrix and the load matrix, and use the extracted fault features as the input samples of the weighted probability neural network; S3: train a weighted probability neural network through the fault feature to obtain a weighted probability neural network model; S4: Input the test sample set into the trained weighted probability neural network model to obtain fault classification results. 2. The fault classification method according to claim 1, wherein in the step S1, a training sample set of the classifier is constructed to obtain the original observation matrix X,

Among them, N represents the total number of samples, and each sample includes m observation variables, then the original observation matrix X is expressed as:

Perform dimensionality reduction processing on the original observation matrix X, and decompose the original observation matrix X into a score matrix T and a load matrix P; among them, in the original observation matrix X, each row represents a sample, and the observation variable x _i =[x _i (1) , _xi (2),..., _xi (N)];

l represents the number of pivots; i=1,2,...,m. 3. The fault classification method according to claim 2, wherein in the step S2, by calculating the reconstruction contribution value of the observed variable x _i

Obtain fault features, and then reconstruct contribution values based on each sample variable

Constitute the fault feature data set, including: the contribution value of the observed variable x _i to the index SPE

get the reconstruction contribution value

Expressed as:

in,

Represents the i-th column in the identity matrix, the direction is the same as x _i ;

represents the projection matrix;

represents the projection matrix

The ith diagonal element of . 4. The fault classification method according to claim 3, wherein in the step S2, the weighted probability neural network comprises 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 to calculate the distance between the input sample and the center vector of each neuron, then the relationship between the sample vector x and the jth center vector of the i-th class represents the for:

Among them, σ represents the smoothing factor; υ _ij represents the ratio of "between-class variance" and "within-class variance" of neuron vector x _ij ; x=(x ₁ , x ₂ ,...,x _m ); N _i represents The total number of samples in category C _i . 5. The fault classification method according to claim 4, wherein the weighting factor of the weighted probability neural network is calculated, and the weighting factor is used to reflect the qth pattern between the pattern layer and the summation layer. The importance of j categories, the calculation process of the weight factor includes: Calculate the sensitivity coefficient S of neurons 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, the element of the rth column of S _j represents: for the input mode 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 the sensitive vector, q=1,...,Q _j , then the sensitive vector a _j is expressed as:

Normalize the elements in a _j to get the weight ω _j of each element in the jth category, which is expressed as:

Among them, when n=∞,

Then w _j is normalized to the [0,1] interval, and the summation layer of the weighted probability neural network is expressed as:

Among them, r represents the rth element in a _j , n ₁ , n ₂ ≥1; when calculating the weighted probability neural network weight factor, n ₁ =∞, n ₂ =2.

Images