A process monitoring method based on mutual k-nearest neighbors

By introducing mutual k nearest neighbor method and elbow rule in the k nearest neighbor process monitoring method, the pseudo-nearest neighbor problem under the influence of outliers is solved, effective outlier point elimination and fault detection are achieved, and monitoring performance is improved.

CN114609993BActive Publication Date: 2024-11-05HUZHOU UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210154037.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-02-20
Publication Date
2024-11-05
Estimated Expiration
2042-02-20

AI Technical Summary

Technical Problem

The existing k-nearest neighbor-based process monitoring method may find pseudo-nearest neighbors when processing data with outliers, resulting in a degradation of monitoring performance and inability to effectively eliminate outliers and perform fault detection.

Method used

Using the mutual k nearest neighbor (MkNN) method, through the selection of different k values ​​and the elbow rule, outliers in the training sample are eliminated, and better k values ​​are used in the fault detection stage to calculate the average cumulative distance statistics of the sample. Identify and remove outliers, determine the fault detection threshold, and improve the reliability of detection.

Benefits of technology

It realizes simultaneous outlier point removal and fault detection, improves fault detection performance, eliminates the influence of pseudo-nearest neighbors, and improves the accuracy and stability of process monitoring.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114609993B_ABST
    Figure CN114609993B_ABST
Patent Text Reader

Abstract

The present invention discloses a process monitoring method based on mutual k-nearest neighbors, which belongs to the field of industrial process monitoring and fault detection, and is mainly used to solve the fault detection problem when outliers exist in industrial processes. The present invention simultaneously realizes the removal of outliers and fault detection, thereby improving the fault detection performance. Specifically, before the training stage, some outliers in the training samples are eliminated, so that the outliers no longer affect the determination of the threshold when establishing the model; in the fault detection stage, the present invention uses more valuable and reliable information than the k-nearest neighbor method, and eliminates the influence of pseudo neighbors. In addition, the present invention solves the mutual protection problem of outliers, thereby further improving the process monitoring performance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of industrial process monitoring and fault detection, and in particular relates to a process monitoring method based on mutual k-nearest neighbors. Background Art

[0002] The k-nearest neighbor (kNN)-based process monitoring method uses the distance between samples to perform fault detection, which can overcome the problems of nonlinearity, multimodality, non-Gaussianity, etc. in process data. However, the data collected in actual industrial processes usually contain a certain amount of noise or even outliers, and the quality of the data cannot be guaranteed. When processing the above process data with outliers, the neighbors found by the kNN rule may not be real and credible neighbors, but a pseudo neighbor, which will degrade the performance of process monitoring. Therefore, in order to solve the problems existing in the prior art, a process monitoring method that can simultaneously achieve outlier removal and fault detection should be developed. Summary of the invention

[0003] The purpose of the present invention is to provide a process monitoring method based on mutual k-nearest neighbors in view of the deficiencies in the prior art.

[0004] The objective of the present invention is achieved through the following technical solutions:

[0005] The detailed steps of a process monitoring method based on mutual k-nearest neighbors are as follows:

[0006] (1) Collect the data set of normal operating conditions of the process, that is, the training set, denoted as m1 is the number of samples and n1 is the number of variables.

[0007] (2) Determine the value of parameter k. Different k values ​​are used in the outlier removal stage and the fault detection stage, denoted as k1 and k2 respectively. Since the larger the k value, the higher the probability that the sample has mutual k-nearest neighbors, after multiple attempts with different k values, k1 and k2 with the best performance are selected. Then, the mutual k-nearest neighbor (MkNN) rule is used to remove outliers in X, and then the result is returned.

[0008] (3) Calculate the MkNN average cumulative distance statistic for each sample in Y, denoted as The calculation formula is as follows:

[0009]

[0010] in, represents the average cumulative distance from sample i to its mutual k nearest neighbors, Represents the square of the Euclidean distance from sample i to its jth mutual nearest neighbor.

[0011] (4) Based on the distance statistics calculated in (3), use the elbow rule to convert Identify and remove outliers with mutual protection in the

[0012] (5) The MkNN average cumulative distance statistic of the outliers with mutual protection determined in (4) is converted from Remove from the list and return

[0013] (6) Determine the fault detection threshold of the MkNN average cumulative distance statistic, denoted as The calculation formula is as follows:

[0014]

[0015] in, is calculated in step (5) According to the result of descending sorting, [m3(1-α)] represents the integer part of m3(1-α).

[0016] (7) For a data sample of the process running online, it is recorded as Determine whether it contains mutual k-nearest neighbors.

[0017] (8) If z has no mutual k-nearest neighbors, z is determined to be a faulty sample; otherwise, proceed to the next step.

[0018] (9) Calculate the average cumulative distance statistic between z and its k nearest neighbors, denoted as Will With detection threshold Compare, if z is a fault, otherwise z is detected as a normal sample.

[0019] The beneficial effects of the present invention are as follows: the present invention realizes the elimination of outliers and fault detection at the same time, thereby improving the fault detection performance. Specifically, before the training stage, some outliers in the training samples are eliminated, so that the outliers no longer affect the determination of the threshold when establishing the model; in the fault detection stage, the present invention uses more valuable and reliable information than kNN, and eliminates the influence of pseudo neighbors. In addition, the present invention solves the mutual protection problem of outliers, thereby further improving the process monitoring performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart of the present invention.

[0021] Figure 2 This is a graph showing the decrease in the average cumulative distance statistic of the mutual k nearest neighbors of the Tennessee-Eastman process for different proportions of outliers.

[0022] Figure 3 This is the result of detecting fault 7 in the Tennessee-Eastman process based on the k-nearest neighbor method.

[0023] Figure 4 This is the result of detecting fault 7 of the Tennessee-Eastman process based on the mutual k-nearest neighbor method. DETAILED DESCRIPTION

[0024] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0025] The present invention provides a process monitoring method based on mutual k-nearest neighbors, such as Figure 1 As shown, the steps are as follows:

[0026] The first step is to collect a data set of normal operating conditions of the process, namely the training set, denoted as m1 is the number of samples and n1 is the number of variables.

[0027] The second step is to determine the value of parameter k. The k values ​​used in the outlier removal stage and the fault detection stage are different, denoted as k1 and k2 respectively. Since the larger the k, the higher the probability that the sample has mutual k-nearest neighbors, after multiple attempts with different k values, k1 and k2 with the best performance are selected, and then the mutual k-nearest neighbor (MkNN) rule is used to remove outliers in X, and then the result is returned. That is, given a sample y, if y i Belongs to one of the k nearest neighbors of y, and y also belongs to y i One of the k nearest neighbors of , then y i is the mutual neighbor of y, using M k (y) = {y i} indicates that if this condition is not met, y does not have mutual k-nearest neighbors, and M k (y) = Φ, the formula is as follows:

[0028] M k (y) = {y i ∈X|y i ∈N k (y)∧y∈N k (y i )}

[0029] Among them, N k (y) represents the k nearest neighbor set of sample y, N k (y i ) represents sample y i The k-nearest neighbor set of X. Apply the above formula to all samples in X, remove the samples whose mutual k-nearest neighbor set is an empty set, that is, the samples that do not have mutual k-nearest neighbors, and then return

[0030] The third step is to calculate the MkNN average cumulative distance statistic for each sample in Y, denoted as The calculation formula is as follows:

[0031]

[0032] in, represents the average cumulative distance from sample i to its mutual k nearest neighbors, represents the square of the Euclidean distance from sample i to its jth nearest neighbor, d i,j =||x i -x j ||,i=1,...,m2,i≠j.

[0033] The fourth step is to use the elbow rule to convert Identify and remove outliers with mutual protection in the First, The statistics of each sample in are arranged in descending order, and then the elbow curve is drawn. All samples before the elbow point are outliers with mutual protection phenomenon. Finally, these identified outliers are removed from Y and returned

[0034] The fifth step is to convert the MkNN average cumulative distance statistics corresponding to the outliers with mutual protection phenomenon determined in the fourth step from Remove from the list and return

[0035] The sixth step is to determine the fault detection threshold of the MkNN average cumulative distance statistic, denoted as The calculation formula is as follows:

[0036]

[0037] in, is calculated in step (5) According to the results arranged in descending order, [m3(1-α)] represents the integer part of m3(1-α), and the confidence level is set to 99%, then α=0.01.

[0038] Step 7: For a data sample of the online process, record it as Use the formula in the second step to determine whether it contains mutual k-nearest neighbors in Y, that is, determine whether the neighbors of z's k-nearest neighbors in Y contain z.

[0039] Step 8: If z has no mutual k-nearest neighbors, that is, M k (z) = Φ, then z is directly determined to be a faulty sample and the detection ends. In this case, the samples do not have mutual k-nearest neighbors, so there is no need to calculate their MkNN average cumulative distance statistics; if M k(z)≠Φ, proceed to the next step.

[0040] In the ninth step, the distance calculation formula in the third step is used to calculate the average cumulative distance statistic between z and its k nearest neighbors, which is recorded as Will The detection threshold determined in step 6 Compare, if z is a fault, otherwise z is detected as a normal sample.

[0041] Application Examples

[0042] The effectiveness of the method of the present invention is illustrated below in conjunction with the specific process of Tennessee Eastman (TE). There are 52 process variables (22 continuous measurement variables, 19 component measurement variables and 11 control variables), the sampling interval is 3 minutes, the process data includes a training set, a validation set, and 21 fault sets, the training set X is 960 normal samples, the validation set is 480 normal samples, each fault set is 960 samples (of which the first 160 are normal samples), the fault 7 data in the TE process is used as the test set, and the outlier introduction ratio is 1%, 2%, 3%, 4%, and 5%.

[0043] Next, the implementation steps of the present invention are described in detail in conjunction with the specific process:

[0044] 1. First, determine the value of parameter k. Different k values ​​are used in the outlier removal stage and the fault detection stage, denoted as k1 and k2 respectively. Since the larger the k value, the higher the probability that the sample has mutual k-nearest neighbors, after multiple attempts with different k values, k1 and k2 with the best performance are selected, which are 42 and 45 respectively. Then, some outliers in X are removed using the mutual k-nearest neighbor rule, and then Y is returned.

[0045] 2. Calculate the MkNN average cumulative distance statistic for each sample in Y, denoted as Then, the elbow rule is used to identify and remove outliers in Y that have mutual protection phenomenon, and return to T.

[0046] 3. The MkNN average cumulative distance statistics of the outliers with mutual protection determined in 2 are Remove from the list and return The fault detection threshold is then determined.

[0047] Find the mutual k nearest neighbors of each sample in the test set in T, calculate the MkNN average cumulative distance statistic of each sample in the test, and then compare it with the threshold. If it exceeds the control limit, it is a faulty sample, otherwise it is a normal sample.

[0048] 4. Online monitoring of fault 7 of the Tennessee-Eastman process, Figure 2 It is the MkNN average cumulative distance statistic decline curve with different outlier ratios, which is used to exclude outliers with mutual protection phenomenon. The fault detection rate and false alarm rate of the k-nearest neighbor-based method (FD-kNN) and the method of the present invention (FD-MkNN) are shown in Table 1 and Table 2. Figure 3 and Figure 4 This is the monitoring result of fault 7. It can be clearly seen that the monitoring method based on mutual k-nearest neighbors is significantly better than the k-nearest neighbor method. In comparison, with the increase in the proportion of outliers, the k-nearest neighbor method degrades seriously, while the method of the present invention always maintains a high fault detection rate.

[0049] Table 1: Fault detection rates (%) of FD-kNN and FD-MkNN methods for TE process fault 7

[0050] method No outliers 1% Outliers 2% Outliers 3% outliers 4% outliers 5% outliers FD-kNN 100.00 79.75 25.25 25.25 20.38 18.75 FD-MkNN 100.00 99.63 97.63 94.75 93.63 92.75

[0051] Table 2: Fault false alarm rate (%) for TE process fault 7 by FD-kNN and FD-MkNN methods

[0052] method No outliers 1% Outliers 2% Outliers 3% outliers 4% outliers 5% outliers FD-kNN 0.00 0.00 0.00 0.00 0.00 0.00 FD-MkNN 0.00 0.00 0.00 0.00 0.00 0.00

[0053] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modification and change made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A process monitoring method based on mutual k-nearest neighbors, characterized in that: The method comprises the following steps: (1) Collect the data set of normal operating conditions of the process, that is, the training set, denoted as m1 is the number of samples, n1 is the number of variables; (2) Determine the value of parameter k. Different k values ​​are used in the outlier removal stage and the fault detection stage, denoted as k1 and k2 respectively. The larger the k value, the higher the probability that the sample has mutual k-nearest neighbors. After multiple attempts with different k values, k1 and k2 with the best performance are selected. Then, the mutual k-nearest neighbor (MkNN) rule is used to remove outliers in the training set, and the number of samples m2 is obtained. The training set is denoted as (3) Calculation The MkNN average cumulative distance statistic of each sample in is recorded as The calculation formula is as follows: in, represents the average cumulative distance from sample i to its mutual k nearest neighbors, Represents the square of the Euclidean distance from sample i to its jth nearest neighbor; (4) Based on the distance statistics calculated in step (3), use the elbow rule to convert The outliers with mutual protection phenomenon are identified and removed, and the number of samples m3 is obtained. The training set is recorded as (5) The MkNN average cumulative distance statistic of the outliers with mutual protection determined in step (4) is converted from Remove from the list and return (6) Determine the fault detection threshold of the MkNN average cumulative distance statistic, denoted as The calculation formula is as follows: in, i=1,…,m3 is the U calculated in step (5) i 2 According to the results of descending order, [m3(1-α)] represents the integer part of m3(1-α); (7) For a data sample of the process running online, it is recorded as Determine whether it contains mutual k-nearest neighbors; (8) If z has no mutual k-nearest neighbors, z is determined to be a faulty sample; otherwise, proceed to the next step; (9) Calculate the average cumulative distance statistic between z and its k nearest neighbors, denoted as Will With detection threshold Compare, if z is a fault, otherwise z is detected as a normal sample.

2. The process monitoring method based on mutual k-nearest neighbors according to claim 1 is characterized in that: The step (2) is specifically as follows: using the mutual k-nearest neighbor (MkNN) rule to remove outliers in X, that is, given a sample y, if y i is one of the k nearest neighbors of y, and y is also one of the k nearest neighbors of yi, then yi is a mutual neighbor of y, and M k (y) = {y i } indicates that if this condition is not met, y does not have mutual k-nearest neighbors, and M k (y) = Φ, the formula is as follows: M k (and)={and i ∈X|y i ∈N k (y)∧y∈N k (and i )} Among them, N k (y) represents the k nearest neighbor set of sample y, N k (y i ) represents sample y i k nearest neighbor set, apply the above formula to all samples in X, remove the samples whose mutual k nearest neighbor set is an empty set, that is, the samples without mutual k nearest neighbors, and then return 3. The process monitoring method based on mutual k-nearest neighbors according to claim 1 is characterized in that: The step (4) is specifically as follows: using the elbow rule to identify and remove outliers in Y that have mutual protection phenomenon; first, The statistics of each sample in are arranged in descending order, and then the elbow curve is drawn. All samples before the elbow position are outliers with mutual protection phenomenon. Finally, these identified outliers are removed from Y and returned