CN111352408B - A fault detection method for multi-condition process industrial process based on evidence K-nearest neighbors - Google Patents

Abstract

The invention discloses a multi-working-condition process industrial process fault detection method based on evidence K nearest neighbor, which belongs to the field of industrial process monitoring. Then calculating the correlation degree of each sample in the training set and each working condition, and establishing an evidence information base; for a sample to be detected, K samples are selected from the training set by the K neighbor model, and evidence information corresponding to the K samples is synthesized by a D-S evidence theory to obtain a final detection result. The fault diagnosis model established based on evidence K nearest neighbor can be well adapted to the multi-working-condition running state of the system; the method can reduce the false alarm rate in the monitoring process and improve the accuracy of fault alarm.

Description

A fault detection method for multi-condition process industrial process based on evidence K-nearest neighbors

technical field

The invention belongs to the field of industrial process monitoring, and in particular relates to a multi-condition process industrial process fault detection method based on evidence K nearest neighbors.

Background technique

With the widespread adoption of DCS, it has become easier to collect data from industrial processes. The development of data mining and database technology also provides strong technical support for the development and application of data-driven modeling methods in industrial processes. Therefore, data-driven process monitoring, control and system optimization are widely used.

As a complex large-scale energy conversion equipment, the boiler is one of the most important parts of the thermal power plant. The stable operation of the boiler is closely related to the power quality. However, the combustion process of the boiler is complex and changeable, and it is difficult to use the traditional mechanism analysis to establish its accurate model; at the same time, the boiler load changes with the external demand. Under different loads, the operation mode of the boiler changes, and the operation process shows obvious Multi-case feature.

For traditional fault diagnosis methods, such as principal components analysis (PCA), partial least squares (partial least squares, PLS), etc., have been used for fault detection and diagnosis in industrial processes for a long time. Such methods are also considered to be very effective. However, it should be noted that methods such as PCA and PLS are formulated under the assumption that the process variables are independent, uniformly distributed and linearly correlated, and assume that the system has only a single normal operating condition. In actual industrial production, it is difficult to meet this assumption, resulting in inaccurate monitoring results and a large number of false alarms.

SUMMARY OF THE INVENTION

In view of the above technical problems, the present invention proposes a multi-condition flow industrial process fault detection method based on evidence K neighbors, which overcomes the deficiencies of the prior art and has good effects.

In order to achieve the above object, the present invention adopts the following technical solutions:

A method for detecting faults in a multi-condition process industrial process based on evidence K-nearest neighbors, comprising the following steps:

Step 1: Obtain the normal historical working condition data of the industrial process, use this data as a training set, standardize the training set, and calculate the standardized value;

Step 2: Divide the standardized training set into multiple normal working conditions through the clustering algorithm;

Step 3: Determine the evidence information according to the correlation criterion. Since each training sample carries the evidence information, the correlation between each training sample and each normal working condition is obtained, and then the evidence information database is established; the evidence information database is all normal conditions. A collection of evidentiary information carried by the sample;

Step 4: Calculate the discount factor γ _i of the training sample according to the local density of each training sample, so that the fault boundary is adapted to the local characteristics of the training sample;

Step 5: Standardize the data to be tested, and find the k samples closest to the data to be tested in the training set;

Step 6: According to the D-S evidence theory Dempster's integration rule, synthesize the evidence information of the k samples in step 5 to obtain the fault detection result.

n为正常工况样本的个数，i代表第i个样本，p为每个样本的维度：Preferably, in step 1, according to formula (1), the normalized value is calculated, the normal working condition data set

n is the number of samples under normal conditions, i represents the ith sample, and p is the dimension of each sample:

是实际值，

是变量x^j的最大值，

是变量x^j的最小值，

是标准化后的值，j表示样本维度。in,

is the actual value,

is the maximum value of the variable x ^j ,

is the minimum value of the variable x ^j ,

is the normalized value, and j represents the sample dimension.

Preferably, the clustering algorithm in the step 2 is any one of K-means clustering, Gaussian mixture model clustering, and density peak clustering.

Preferably, the correlation criterion in step 3 is one of Euclidean distance, Mahalanobis distance, and local density, or a combination of two methods.

Preferably, in step 3, according to formula (2), the evidence information is calculated for each training sample:

表示第i个样本与所有工况相关度的总和，m_i({ω_q})表示第i个样本属于第q个正常工况的概率。Among them, q represents the qth normal operating condition, d _iq represents the correlation between the ith sample and the qth normal operating condition, and c represents the total number of normal operating conditions;

Represents the sum of the correlation between the ith sample and all working conditions, and m _i ({ω _q }) represents the probability that the ith sample belongs to the qth normal working condition.

Preferably, in step 4, according to formula (3), the discount factor γ _i of the training sample is calculated:

Among them, D _i is the set of distances between the sample _xi and the k training samples closest to the sample _xi , and ε is the parameter in the quantile function.

Preferably, in step 6, the following steps are included:

Step 6.1: According to formula (4), calculate the adaptive discount factor γ ^s :

Among them, λ is used to adjust the size of γ ^s , and γ _i is the discount factor of training samples;

Step 6.2: According to formula (5), calculate the weight value of the training sample:

Among them, d _i represents the Euclidean distance between the test sample and the training sample, and γ ^s is the adaptive discount factor;

Step 6.3: According to formula (6), calculate the mass function of the sample to be tested belonging to each working condition:

表示距离待测样本最近的K个训练样本的索引集合，d_si表示待测样本与训练样本之间的距离；Among them, α is the confidence level of the training set, and m _i (A) is the evidence information of the sample to be tested;

Represents the index set of K training samples closest to the sample to be tested, and _dsi represents the distance between the sample to be tested and the training sample;

Step 6.4: According to formulas (7), (8), (9), synthesize the mass function to obtain the diagnosis result;

Among them, K ₁ is the normalization factor, m ^s (Ω) represents the failure probability of the test sample, and m ^s ({ω _q }) represents the probability that the test sample belongs to each working condition.

Beneficial technical effects brought by the present invention:

(1) The fault diagnosis model based on K-nearest neighbors can be well adapted to the operating state of the system under multiple operating conditions; (2) This method can reduce the false alarm rate in the monitoring process and improve the accuracy of fault alarms.

Description of drawings

Figure 1 is a flow chart of the method of the present invention.

FIG. 2 is a schematic diagram of a projection of a training set clustering result on a two-dimensional space according to an embodiment of the present invention.

FIG. 3 is a schematic diagram of a fault detection result of an embodiment of the present invention using the method for detecting faults in a multi-condition process industrial process based on the evidence K nearest neighbors of the present invention.

Detailed ways

The present invention is described in further detail below in conjunction with the accompanying drawings and specific embodiments:

The data of this example is generated by a coal mill in a boiler system of a power plant, and a total of 2880 samples and 28 variables are collected. The fault interval is 2133-2500. Take the first 1000 normal samples as the training set and the rest as the test set. As shown in Figure 1, the method of the present invention includes the following two major stages: offline modeling and online monitoring, respectively, and the offline modeling stage specifically includes:

Step 1: Normalize the training set;

是实际值，

是变量x^j的最大值，

是变量x^j的最小值，

是标准化后的值，j表示样本维度。In the above formula,

is the actual value,

is the maximum value of the variable x ^j ,

is the minimum value of the variable x ^j ,

is the normalized value, and j represents the sample dimension.

Step 2: Use PCA to project the high-dimensional data space composed of the process data into the low-dimensional feature subspace, as shown in Figure 2; observe the optimal number of clusters, and cluster the training set through the K-means clustering algorithm. The parameter K =5, divide the normal working condition into two working conditions, w1 and w2;

Step 3: Calculate the center points of w1 and w2 respectively, calculate the correlation between each training set and w1, w2 to establish an evidence information base, and calculate the discount factor γ _i of each training sample;

In the above formula, D _i is the distance set between the training sample _xi and the K nearest neighbor model to select the training sample, and the parameter ε = 0.85.

The online monitoring stage mainly includes:

Step 1: For a sample _xi to be tested, first standardize. Find the 5 nearest neighbor samples in the training set to form a set Θ.

Step 2: Calculate the adaptive discount factor γ ^s of the sample _xi to be tested according to the discount factor corresponding to each training set in Θ.

In the above formula, the parameter λ=0.2, and γ _i is the discount factor of the training sample;

Step 3: Calculate the weight between the sample _xi to be tested and each training set in Θ:

In the above formula, d _i represents the Euclidean distance between the test sample and the training sample, and γ ^s is the adaptive discount factor;

Step 4: Calculate the mass function of the sample to be tested belonging to each working condition:

Among them, the parameter α=0.95 is the confidence level of the training set, and m _i (A) is the evidence information of the training sample;

Step 5: Use Dempster's integration rule to synthesize the evidence carried by each training set in Θ to obtain the final detection result. It is stipulated that when m ^s (Ω)=0.2 is the fault early warning boundary, and m ^s (Ω)=0.9 is the fault boundary. The monitoring results are shown in Figure 3;

Among them, K ₁ is the normalization factor, m ^s (Ω) represents the failure probability of the test sample, and m ^s ({ω _q }) represents the probability that the test sample belongs to each working condition.

The above description is not intended to limit the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also belong to the present invention. protected range.

Claims (6)

1. A multi-working-condition process industrial process fault detection method based on evidence K nearest neighbor is characterized by comprising the following steps: the method comprises the following steps:

step 1: acquiring normal historical working condition data of an industrial process, taking the normal historical working condition data as a training set, standardizing the training set, and calculating a standardized value;

step 2: dividing the standardized training set into a plurality of normal working conditions through a clustering algorithm;

and step 3: determining evidence information according to a relevancy criterion, and obtaining the relevancy of each training sample and each normal working condition as each training sample carries the evidence information, thereby establishing an evidence information base; the evidence information base is a set of evidence information carried by all normal samples;

and 4, step 4: calculating discount factor gamma of training samples according to local density of each training sample_iAdapting the fault boundary to the local characteristics of the training sample;

and 5: standardizing the data to be detected, and finding out k samples closest to the data to be detected in a training set;

step 6: synthesizing the evidence information of the k samples in the step 5 according to a Dempster' S integration rule of a D-S evidence theory to obtain a fault detection result; the method comprises the following steps:

step 6.1: calculating an adaptive discount factor gamma according to equation (4)^s：

Wherein λ is used to adjust γ^sSize of (a), γ_iA discount factor for a training sample;

step 6.2: according to equation (5), the weight values of the training samples are calculated:

wherein d is_iRepresenting Euclidean distance, gamma, of the sample to be measured and the training sample^sIs an adaptive discount factor;

step 6.3: and (3) calculating a mass function of the sample to be tested belonging to each working condition according to the formula (6):

where α is the confidence of the training set, m_i(A) The evidence information of the sample to be detected is obtained;

index set representing K training samples closest to the sample to be measured, d_siRepresenting the distance between the sample to be tested and the training sample;

step 6.4: synthesizing a mass function according to the formulas (7), (8) and (9) to obtain a diagnosis result;

wherein, K₁As a normalization factor, m^s(Ω) represents the failure probability of the test sample, m^s({ω_q} represents the probability that the test sample is subject to each condition.

2. The evidence K nearest neighbor based multi-condition process industrial process fault detection method according to claim 1, wherein: in step 1, the normalized value, normal condition data set, is calculated according to equation (1)

n is the number of normal working condition samples, i represents the ith sample, and p is the dimension of each sample:

wherein,

is the actual value of the,

is a variable x^jThe maximum value of (a) is,

is a variable x^jThe minimum value of (a) is determined,

is a normalized value and j represents the sample dimension.

3. The evidence K nearest neighbor based multi-condition process industrial process fault detection method according to claim 1, wherein: the clustering algorithm in the step 2 is any one of a K-means clustering method, a Gaussian mixture model clustering method and a density peak value clustering method.

4. The evidence K nearest neighbor based multi-condition process industrial process fault detection method according to claim 1, wherein: and the relevancy criterion in the step 3 is one or a combination of two methods of Euclidean distance, Mahalanobis distance and local density.

5. The evidence K nearest neighbor based multi-condition process industrial process fault detection method according to claim 1, wherein: in step 3, evidence information is calculated for each training sample according to equation (2):

wherein q represents the q-th normal operating condition, d_iqThe correlation degree of the ith sample and the qth normal working condition is shown, and c represents the total number of the normal working conditions;

represents the sum of the correlation of the ith sample with all the working conditions, m_i({ω_qAnd) represents the probability that the ith sample belongs to the qth normal condition.

6. The evidence K nearest neighbor based multi-condition process industrial process fault detection method according to claim 1, wherein: in step 4, the fold of the training sample is calculated according to equation (3)Deduction factor gamma_i：

Wherein D is_iIs a sample x_iAnd distance sample x_iThe distance between the nearest k training samples is combined into a set, which is a parameter within the quantile function.

Images