TWI639908B - Method for detecting and diagnosing an abnormal process - Google Patents

Abstract

A process detection and diagnosis method. This method includes the model building phase as well as the online analysis phase. In the model establishment phase, historical sample data is first provided and grouped to obtain a first data clustering model. Then, based on the historical sample data and the first data grouping model, the normal data threshold corresponding to each data group is determined. Then, at least one outlier sample data of each data group is removed according to a normal data threshold to obtain a second data clustering model. In the online analysis phase, the second data clustering model and the statistical process control method are firstly used to determine whether the target data is abnormal. When the target data is judged to be abnormal, the probability of variation of the data value of the target data is calculated to determine the abnormal process factor.

Description

Process debugging and diagnosis method

The invention relates to a method for detecting and diagnosing a process, and in particular to a method for detecting and diagnosing a process for causing an abnormal process.

The process of debugging and diagnosis is to find out the abnormal operation of the process as early as possible, and to understand the cause and time of the abnormality. In this way, the abnormal process can be improved in time to avoid unnecessary product loss and cost waste, and at the same time improve the quality and yield of the product, and exert the maximum efficiency of the process.

During the process of the process, there are usually some unavoidable variations, such as changes in operating conditions, changes in the external environment, vibration or abnormality of the equipment. In order to monitor the impact of the above variation on the process, most of the current process control is carried out by statistical process control (SPC).

However, univariate SPC basically only allows one or a single type of normal distribution variable, but most of the process operations usually have non-stationary, auto-correlated, and cross-correlation (Cross- Visible) rather than individual process parameters, so using univariate SPC to monitor process variation can be misleading The role, even many errors can not be detected. Therefore, the industry has proposed various multivariate process control methods to improve the shortcomings of single variable SPC. However, although these multivariate process control methods can monitor whether the process is abnormal, they cannot identify the cause of the abnormality in the process.

Therefore, a process debugging and diagnosis method is needed to identify the cause of the abnormality in the process.

The object of the present invention is to provide a method for detecting and diagnosing a process, which can not only monitor whether a process is abnormal, but also timely identify the cause of an abnormality in the process.

According to the above purpose, the process debugging and diagnosis method includes a model establishment phase and an online analysis phase. In the model establishment phase, a plurality of historical sample data are first provided. Next, the historical sample data is grouped to obtain a first data grouping model, wherein the first data grouping model includes a plurality of data groups. Then, based on the historical sample data and the first data grouping model, the normal data threshold corresponding to each data group is determined. Then, at least one outlier sample data of each data group is removed according to a normal data threshold corresponding to each data group to obtain a second data grouping model. In the online analysis phase, first, based on the second data clustering model, it is determined whether the target data is abnormal data. Then, when the target data is abnormal data, the statistical process control method is used to determine whether the target data is within the scope of regulation. Then, when the target data exceeds the scope of control, the probability of variation of the plurality of data values of the target data is calculated. Then, at least one abnormal process factor corresponding to the target data is determined according to the probability of mutation.

According to an embodiment of the invention, the historical sample data described above corresponds to different time points.

According to still another embodiment of the present invention, the step of grouping the historical sample data is performed by a K-means, a K-medoids, or a Gaussian Mixture Model. get on.

According to still another embodiment of the present invention, the step of determining a normal data threshold includes: calculating a plurality of negative log likelihood (NLL) values for each data group of the first data grouping model; The NLL value of the data group is taken as an absolute value and sorted by size. The largest one of the NLL values is at the top, and the smallest one of the NLL values is at the bottom; the NLL value after the absolute value is taken from the smallest one. One of the 95% decided to be the normal data threshold.

According to still another embodiment of the present invention, the step of determining a normal data threshold further includes using an S function (Sigmoid function) to map the NLL value after the absolute value to a range of 0 to 1.

According to an embodiment of the present invention, the step of determining whether the target data is abnormal data according to the second data grouping model comprises: calculating an outlier distance between a center of each data group of the second data grouping model and the target data; According to the outlier distance, it is judged whether the target data is abnormal data.

According to still another embodiment of the present invention, the control range is calculated by: selecting a plurality of normal target data according to a target sample time point corresponding to the target data; and determining a range of the control range according to the normal target data. Upper limit and lower limit of range.

According to still another embodiment of the present invention, the statistical process control method uses an Xbar control map to determine whether the target data is within the scope of regulation.

According to still another embodiment of the present invention, the step of determining the abnormal process factor corresponding to the target data according to the mutation probability comprises: sorting the mutation probability corresponding to the data value according to the size; and selecting at least the highest probability of the mutation probability A chance of mutation. Determining at least one data value according to the selected mutation probability; and determining that at least one process factor corresponding to the at least one data value is an abnormal process factor.

100‧‧‧Process detection and diagnosis methods

110‧‧‧Model establishment phase

111-114‧‧‧Steps

113a-113c‧‧‧Steps

120‧‧‧Online analysis stage

121-125‧‧‧Steps

121a-121b‧‧‧Steps

122a-122b‧‧‧Steps

124a-124d‧‧‧Steps

200‧‧‧four-stage air compressor

210‧‧‧First stage air compressor

220‧‧‧Second stage air compressor

230‧‧‧third-stage air compressor

240‧‧‧Fourth-stage air compressor

P1_in, P1_out, T1_in, T1_out‧‧‧ signals

P2_in, P2_out, T2_in, T2_out‧‧‧ signals

P3_in, P3_out, T3_in, T3_out‧‧‧ signals

P4_in, P4_out, T4_in, T4_out‧‧‧ signals

For a more complete understanding of the embodiments and the advantages thereof, the following description is made with reference to the accompanying drawings, wherein: FIG. 1 is a flow chart showing a process debugging and diagnosis method according to an embodiment of the present invention; 2] is a schematic diagram showing the structure of a four-stage air compressor using a process debugging and diagnosis method; [Fig. 3] is a flow chart showing the steps of determining a normal data threshold according to an embodiment of the present invention; [Fig. 4] A flow chart showing the steps of determining whether the target data is abnormal data according to an embodiment of the present invention; [FIG. 5] is a schematic flowchart showing an SPC monitoring method according to an embodiment of the present invention; FIG. 6 is a schematic flow chart showing an abnormal process factor determining step according to an embodiment of the present invention.

Embodiments of the invention are discussed in detail below. However, it will be appreciated that the embodiments provide many applicable concepts that can be implemented in a wide variety of specific content. The examples discussed and disclosed are illustrative only and are not intended to limit the scope of the invention.

Please refer to FIG. 1 and FIG. 2 simultaneously. FIG. 1 is a schematic flow chart of a process debugging and diagnosis method 100 according to an embodiment of the present invention, and FIG. 2 is a schematic diagram of a four-stage air compressor applying the process debugging and diagnosis method 100. Schematic diagram of 200. The process debugging and diagnosis method 100 of the embodiment of the invention is used to monitor the process of a system to monitor whether the process is abnormal, and to identify the cause of the abnormality in the process. In the present embodiment, the process error detection and diagnosis method 100 is applied to the four-stage air compressor 200 to monitor its process. However, embodiments of the invention are not limited to the application of a four-stage air compressor. The process debug and diagnostic method 100 can be applied to other systems that require monitoring and error diagnosis.

As shown in FIG. 2, the four-stage air compressor 200 includes a first-stage air compressor 210, a second-stage air compressor 220, a third-stage air compressor 230, and a fourth-stage air compressor 240, wherein signals P1_in and P1_out Corresponding to the input gas flow rate and the output gas flow rate of the first stage air compressor 210, respectively, the signals T1_in and T1_out respectively correspond to the input gas temperature and the output gas temperature of the first stage air compressor 210; the signals P2_in and P2_out respectively correspond to The input gas flow and the output gas flow of the second stage air compressor 220, and the signal T2_in and T2_out corresponds to the input gas temperature and the output gas temperature of the second stage air compressor 220; the signals P2_in and P2_out respectively correspond to the input gas flow and the output gas flow of the third stage air compressor 230, and the signals T3_in and T3_out correspond respectively. The input gas temperature and the output gas temperature to the third stage air compressor 230; the signals P4_in and P4_out correspond to the input gas flow and the output gas flow of the fourth stage air compressor 240, respectively, and the signals T4_in and T4_out correspond to the fourth The input gas temperature and the output gas temperature of the stage air compressor 240.

Returning to FIG. 1, the process debug and diagnostic method 100 includes a model establishment phase 110 and an online analysis phase 120. The model establishment phase 110 is used to establish a data clustering model, while the online analysis phase 120 is used to determine whether the target data is abnormal by using a data clustering model and a statistical process control (SPC) method, and to find out the cause of the abnormality.

In the model establishment phase 110, step 111 is first performed. To provide a plurality of historical sample data as a training model. These historical sample data correspond to different time points, and each historical sample data contains multiple data values. In this embodiment, the data values include, but are not limited to, the foregoing 16 signals P1_in, P1_out, T1_in, T1_out, P2_in, P2_out, T2_in, T2_out, P3_in, P3_out, T3_in, T3_out, P4_in, P4_out, T4_in, and T4_out.

Next, step 112 is performed to group the historical sample data to divide the historical sample data into a plurality of data groups, and obtain the first data grouping model. In this embodiment, step 112 is to utilize a mixture model. (Gaussian Mixture Model) is used to perform grouping of historical sample data, but embodiments of the present invention are not limited thereto. In other embodiments of the invention, K-means or K-medoids may be utilized to perform grouping of historical sample data. In addition, the number of groups can be determined by methods such as Gap Statistics and Silhouette indicators.

Then, step 113 is performed to determine a normal data threshold corresponding to each data group according to the historical sample data and the first data grouping model. In this embodiment, it is assumed that the historical sample data is mostly normal, but may contain a certain proportion of abnormal events, so the normal data threshold is set here to exclude the historical sample data of the abnormal events.

Please refer to FIG. 3. FIG. 3 is a schematic flowchart diagram of a determining step 113 of a normal data threshold value according to an embodiment of the present invention. In step 113, step 113a is first performed to calculate a negative log likelihood (NLL) value for each data group of the first data clustering model. Next, step 113b is performed to take the NLL value of each data group as an absolute value and sort by size. In this embodiment, the largest of the NLL values is located at the top of the sequence, and the smallest of the NLL values is located at the bottom of the sequence, but embodiments of the present invention are not limited thereto. Then, step 113c is performed to determine the NLL value after taking the absolute value from the smallest one to the 95% of the upper one to the above-mentioned normal data threshold. In other words, in the present embodiment, 5% of the historical sample data is determined as the abnormal data.

Considering that the numerical range of NLL values may be between positive and negative infinity, there may be a range of performance far beyond the training data for the target data to be tested. Therefore, in other embodiments of the present invention, the NLL value output by the first data grouping model is corresponding to between 0 and 1. Here, the N-value obtained by taking the absolute value is used to correspond to the range of 0 to 1 by the Sigmoid function. The S function is expressed as follows: Among them, the constants a and c are used to control the shape of the S function IF(x, a, c), and x is the NLL value after taking the absolute value.

Please return to FIG. 1. After step 113, proceed to step 114 to remove the outlier sample data of each data group according to the normal data threshold corresponding to each data group to obtain the second data. Group model. Since this embodiment uses 95% as the threshold for normal data, 5% of the historical sample data is determined to be removed from the outlier sample data.

After the model establishment phase 110 is completed, the second data clustering model can be used to perform the online analysis phase 120 to debug and diagnose the target data on the line.

In the online analysis phase 120, step 121 is first performed to determine whether the target data is abnormal data according to the second data grouping model. Please refer to FIG. 4 , which is a flow chart of the step 121 of determining whether the target data is abnormal data according to an embodiment of the present invention. In step 121, step 121a is first performed to calculate the outlier distance between the center of each data group of the second data grouping model and the target data. For example, the distance from the center of the i-th signal value of the target data dl (one of the above 16 signals) can be expressed as follows: Where u _q,i is the i-th dimension of the qth group of the model; Σ _{q, i} is the i-th variation on the diagonal of the co-variation matrix of the qth Mixture; the distance d _l,i is the signal The value t _l,i to the i-th dimension of each Mixture center u _q, the Mahalanobis distance of _i multiplied by the weight of this Mixture (Weight) The sum of them.

Next, step 121b is performed to determine whether the target data is abnormal data based on the outlier distance. In the present embodiment, the threshold of the distance can be set to determine whether the target data deviates too much, but the embodiment of the present invention is not limited thereto. When the outlier distance is greater than the distance threshold, the target data is initially determined to be abnormal data.

Returning to FIG. 1, after step 121, step 122 is performed to determine whether the target data is within the scope of regulation by using a statistical process control (SPC). In the embodiment of the present invention, in order to avoid the occurrence of a false alarm, the embodiment of the present invention separately monitors each instrument signal value d _{l, i} by the SPC method. In the present embodiment, step 121 is monitored by the SPC Xbar control map. However, embodiments of the invention are not limited thereto.

Please refer to FIG. 5. FIG. 5 is a schematic flowchart diagram of an SPC monitoring method according to an embodiment of the present invention. In step 122, step 122a is first performed to select a plurality of normal target data in a rolling manner according to a time point corresponding to the target data.

Next, step 122b is performed to determine the range upper limit value and the range lower limit value of the control range based on the normal target data. The upper and lower limits of the scope of control are calculated using the following equations: Among them, UCL _{l, i} is the upper limit of regulation; LCL _{l, i} is the lower limit of regulation; for average value; for Standard deviation; b is the length of Xbar; d _l,i,j is the sum of the signal value t _l,i,j to the i-th dimension of the center of each Mixture and the Mahalanobis distance of the j-th dimension multiplied by the weight of Mixture .

Referring back to FIG. 1, when it is determined in step 122 that the target data exceeds the SPC control range, step 123 is performed to calculate the probability of mutation corresponding to the data value of the target data. In this embodiment, the probability of mutation is calculated as follows: Where P _l (i) is the probability of mutation; d _l,j is the sum of the signal value t _l,j to the jth dimension of the jth dimension of each Mixture center multiplied by the weight of Mixture.

Then, step 124 is performed to determine at least one abnormal process factor corresponding to the target data according to the probability of mutation. Please refer to FIG. 6 , which illustrates a flow of an abnormal process factor determining step according to an embodiment of the present invention. schematic diagram. In step 124, step 124a is first performed to sort the probability of mutation corresponding to the data value by size. Next, step 124b is performed to sequentially select at least one mutation probability from the largest of the mutation probabilities. For example, choose the largest of the mutation probabilities. Then, step 124c is performed to select the data value according to the selected probability of mutation. For example, select the data value corresponding to the maximum mutation probability. Next, step 124d is performed to determine the process factor corresponding to the data value selected in step 124c as the abnormal process factor. For example, if the value of the data selected in step 124c is the value of the signal T1_out, it means that the temperature of the output gas of the first-stage air compressor is abnormal and needs to be detected.

In other embodiments of the invention, a probability threshold may also be preset to determine the probability of the selected covariance. For example, when the probability of the total variability exceeds the probability threshold, the corresponding data value is selected, and the process is determined as an abnormal process factor. However, embodiments of the invention are not limited thereto.

After the end of step 124, step 125 is followed to receive the next target data to be analyzed, and then returns to step 121.

According to the above description, the embodiment of the present invention applies the similarity principle to establish an early warning model of the process/device to perform on-the-spot monitoring of the process state of the process/device, and simultaneously calculate the operating state of the process/device, so that the process/device can be When an exception occurs, find out the cause of the exception. In addition, the embodiment of the present invention also applies the SPC method to consider the operating state of the process/device to avoid unnecessary false alarms.

Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention, and any one of ordinary skill in the art can make some changes and refinements without departing from the spirit and scope of the present invention. The scope of the invention is defined by the scope of the appended claims.

Claims (8)

A method for detecting and diagnosing a process includes: a model establishing stage, comprising: providing a plurality of historical sample data; grouping the historical sample data to obtain a first data grouping model, wherein the first data grouping model comprises a plurality of data groups; determining, according to the historical sample data and the first data grouping model, a normal data threshold corresponding to each of the data groups; and corresponding to each of the data groups a normal data threshold to remove at least one outlier sample data of each of the data groups to obtain a second data clustering model; and an online analysis phase for analyzing a target data, wherein the online analysis phase The method comprises: determining whether the target data is abnormal data according to the second data grouping model; and when the target data is abnormal data, using a statistical process control method to determine whether the target data is located within a control range; when the target data exceeds In the scope of the control, the probability of complex mutations of the plurality of data values of the target data is calculated; Variation of the plurality of probability to determine at least a factor of exception process information corresponding to the target; wherein some of the history of the sample data for the step of clustering based algorithm to average K (K-means), K center point algorithm (- K- Medoids) or Gaussian Mixture Model. The process debugging and diagnosis method of claim 1, wherein the historical sample data corresponds to different time points. The process of detecting and diagnosing the process according to claim 1, wherein the step of determining the normal data threshold corresponding to each of the data groups comprises: calculating each of the data groups of the first data grouping model a plurality of negative log likelihood (NLL) values of the group; the NLL values of each of the data groups are taken as absolute values and sorted by size, wherein one of the NLL values is the largest Above, the smallest of the NLL values is located at the bottom; and the NLL values determined by taking the absolute value are one of the 95% of the lowest ones being the normal data threshold. The process of detecting and diagnosing the process according to claim 3, wherein the step of determining the normal data threshold corresponding to each of the data groups further comprises: using an S function (Sigmoid function) to take an absolute value The NLL values correspond to a range of 0~1. The process of detecting and diagnosing the method according to claim 1, wherein the step of determining whether the target data is abnormal data according to the second data grouping model comprises: calculating each of the data groups of the second data grouping model The distance between the center of the group and one of the target materials; and determining whether the target data is abnormal according to the distance of the outlier. The process debugging and diagnosis method according to claim 1, wherein the control range is calculated by: selecting a plurality of normal target data according to a target sample time point corresponding to the target data; and according to the These normal target data determine the upper limit of one of the scope of the control and the lower limit of a range. The process debugging and diagnosis method according to claim 1, wherein the statistical process control method uses an Xbar control chart to determine whether the target data is within the control range. The process of detecting and diagnosing the process according to claim 1, wherein the step of determining the at least one abnormal process factor corresponding to the target data according to the mutation probability comprises: the mutations corresponding to the data values The probability is sorted according to the size; at least one mutation probability is selected in order from the largest one of the mutation chances; Selecting at least one data value according to the selected at least one mutation probability; and determining that the at least one process factor corresponding to the at least one data value is the at least one abnormal process factor.