A method for industrial data quality analysis

Abstract

The present invention provides an industrial data quality analysis method, comprising the following steps: (1) collecting processing data and performing data verification to match the processing data with the processing tool; (2) grouping the processing data and extracting the characteristic values ​​of the processing data to identify abnormalities in the processing process; (3) fitting the processing data through a prophet model and locating the abnormal data. By constructing a matching degree algorithm from the perspective of the characteristic values ​​of the processing data, a matching index is finally obtained, and it is intuitively obtained from the data whether the production parameters match the processing tool. By extracting the characteristic values ​​of the adopted processing data for dimensionality reduction processing, the judgment of processing abnormalities can be effectively simplified and the analysis efficiency can be improved. The present invention does not need to judge abnormal data after the entire workpiece processing is completed, and can achieve real-time analysis of data in the processing process.

Description

Technical Field

[0001] The present invention relates to the technical field of processing data in industrial production, and in particular to an industrial data quality analysis method. Background Art

[0002] In recent years, with the rapid development of high-speed Internet, big data, cloud computing, artificial intelligence and other technologies, a new industrial revolution is emerging. In the daily industrial production process, it is necessary to collect production data from multiple links and analyze the collected data to determine whether the equipment in the corresponding production process is stable based on the analysis results. However, with the continuous increase in the number of data sensors, the amount of data collected is getting larger and larger. How to judge the accuracy of the data and evaluate the detection data to find abnormal data while ensuring efficiency is crucial.

[0003] The current production data collection mainly has the following technical problems:

[0004] In terms of tool data collection, the collected production parameter signals may not match the processing tools due to on-site operation reasons;

[0005] At present, the method of analyzing whether the machining process is stable is generally to check whether the data trend of the entire machining process is similar to that of the same type of workpiece after the entire workpiece is machined. Therefore, it is difficult to judge whether the workpiece is processed normally for workpieces with large data volume and long machining cycle.

[0006] It is difficult to judge abnormal processes during the processing, and real-time analysis during the processing cannot be achieved. Summary of the invention

[0007] In view of the above technical problems, the purpose of the present invention is to provide an industrial data quality analysis method, which can judge production anomalies during the production process based on the accurate collection of production data and dimensionality reduction processing of the collected data.

[0008] The industrial data quality analysis method provided by the present invention comprises the following steps:

[0009] (1) Collect processing data and perform data verification to match the processing data with the processing tool;

[0010] (2) Group the processing data and extract the feature values ​​of the processing data to identify abnormalities in the processing process;

[0011] (3) Fit the processed data through the prophet model and locate the abnormal data.

[0012] Furthermore, in the step (1), each machining tool is grouped in the form of machining intervals at different tool positions according to the machining sequence, and the tool load corresponding to the machining tool in each machining interval is extracted.

[0013] Furthermore, in the step (1), the processing data at the same continuous tool position are divided into the same processing interval, and the processing intervals at the same tool position but not continuous cannot be merged.

[0014] Furthermore, in step (1), the standard deviation of the tool load in each processing interval of each processing tool is calculated respectively, and all processing intervals are combined to calculate the statistic, and the matching degree between the processing data and the processing tool is determined by the discrete degree of the statistic.

[0015] Furthermore, in step (2), the processing time of each workpiece at different tool positions is grouped, and each group of processing time is divided into divided intervals at the same time interval.

[0016] Furthermore, in the step (2), data characteristic values ​​of the processing parameters in each segmented interval are extracted, and the data characteristic values ​​include median, mode, mean or variance.

[0017] Furthermore, in step (2), different segmentation intervals of each workpiece at the same tool position are integrated, and whether the machining process is abnormal is determined by judging the data feature value.

[0018] Furthermore, in step (3), anomaly detection is performed by calculating the residual between the prophet model prediction value and the integrated production data in step (2), and the residual calculation formula includes:

[0019] in is the predicted value of the prophet model, y is the integrated production data in step (2), and error is the residual.

[0020] Furthermore, in step (3), the abnormal data is located by using the 3σ criterion, and the abnormal value is a value in a group of result values ​​whose deviation from the mean value exceeds three times the standard deviation.

[0021] The industrial data quality analysis method of the present invention constructs a matching algorithm from the perspective of the characteristic values ​​of the processing data, and finally obtains a matching index, which can intuitively determine from the data whether the production parameters match the processing tools.

[0022] By extracting the characteristic values ​​of the adopted processing data and performing dimensionality reduction processing, the judgment of processing anomalies can be effectively simplified and the analysis efficiency can be improved.

[0023] The processing data is fitted through the model and the abnormal data is located, which solves the problem that it is difficult to judge whether the processing is normal during the processing of the workpiece due to the long processing cycle and the need to collect a large amount of data. There is no need to judge after the entire workpiece is processed, and real-time analysis of the data during the processing can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0025] Figure 1 It is a flow chart of the industrial data quality analysis method in the present invention;

[0026] Figure 2 A signal diagram for processing data in the prior art;

[0027] Figure 3 This is the data signal diagram after feature extraction;

[0028] Figure 4 b is the data signal diagram after feature extraction;

[0029] Figure 5 for Figure 3 and Figure 4 Figure 2. The two types of data merged into one graph;

[0030] Figure 6 This is the data signal diagram after prophet algorithm fitting;

[0031] Figure 7 This is the b data signal diagram after prophet algorithm fitting;

[0032] Figure 8 is the 3σ distribution diagram of the residual error of a data;

[0033] Fig. 9 is the b data residual 3σ distribution diagram;

[0034] Fig.10 This is the adaptive 3σ distribution diagram of the data residual. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0036] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] Some embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0038] See also Figure 1 The industrial data quality analysis method provided by the present invention mainly includes the following steps:

[0039] (1) Collect processing data and perform data verification to match the processing data with the processing tool;

[0040] (2) Group the processing data and extract the feature values ​​of the processing data to identify abnormalities in the processing process;

[0041] (3) Fit the processed data through the prophet model and locate the abnormal data.

[0042] The industrial data quality analysis method of the present invention is mainly used to determine whether a machining tool operates normally.

[0043] In the past, tool data collection, such as Figure 2 When multiple machining tools correspond to multiple tool loads, the signals may not match due to on-site operation reasons. For example, a tool load collected does not match its corresponding machining tool, but actually matches other tools.

[0044] For example, the collected sl1 data does not match tn1, but actually matches tn2. tn1 and tn2 represent machining tools at different tool positions, while sl1 and sl2 are the corresponding tool loads.

[0045] In the existing solution, by drawing parameter diagrams such as sl1, sl2, tn1, tn2, etc., the matching is judged according to the tool (tn) tool change diagram and the load (sl) trend. However, this solution requires the collection of a large number of parameters, and the accuracy cannot be guaranteed in the judgment.

[0046] In the data verification process of the present invention, a new signal matching method is proposed. If sl1 corresponds to tn1, then when the tool position of tn1 is not switched (i.e., the process of processing with the same processing tool), the fluctuation of the sl1 value change in the corresponding process will not be very large; then for sl1 in this interval, its variance, standard deviation and other statistics reflecting the degree of data dispersion should also be small. Therefore, the corresponding sl value can be divided into intervals according to the tool position of tn1, and the statistics in each interval are calculated respectively, and finally a comprehensive comparison is made.

[0047] Specifically, each machining tool is grouped in the form of machining sections at different tool positions according to the machining sequence, and the tool load corresponding to the machining tool in each machining section is extracted.

[0048] Take the machining tool of tn1 as an example, extract the tool position number in tn1 according to the machining order, if it is the same as the previous tool position number, but the order is discontinuous, it will not be merged with the previous machining interval. For example, in tn1, the tool position is from aba, then the first three a's are one machining interval, the last two a's are another machining interval, and so on, and the machining intervals with the same but discontinuous tool position numbers will not be merged.

[0049] Group sl1 and sl2 corresponding to tn1 respectively, as shown in the following table:

[0050]

[0051] sl1 is divided into multiple processing intervals such as [2,3,2], [8,9], [1,2]…, denoted as sl1 i , i = 1, 2, L, the number of parameters contained in each interval is recorded as n1 i ,i=1,2,L; similarly, sl2sl1 is divided into multiple intervals such as [3,4,7], [1,5], [5,9]…, denoted as sl2 i , i = 1, 2, L, the number of parameters contained in each interval is recorded as n2 i ,i=1,2,L.

[0052] Calculate sl1 separately i The standard deviation of , and all intervals are combined to form a statistic:

[0053]

[0054] where std(sl1i ) indicates sl1 i The standard deviation of

[0055] For sl2 i Similarly, calculate:

[0056] where std(sl1 i ) indicates sl1 i The standard deviation of .

[0057] Compare t1s1 and t1s2. If t1s1<t1s2, it means that after grouping by tn1 tool number, the discrete degree of sl1 data interval is lower than that of sl2, indicating that in the processing without tool change, sl1 load is more balanced. Therefore, it is judged that tn1 and sl1 are more matched.

[0058] The standard deviation of sl1 grouped by tn1 is 10.60719508125356

[0059] The standard deviation of sl1 grouped by tn2 is 11.045757828988163

[0060] The standard deviation of sl2 grouped by tn1 is 11.971683582018622

[0061] The standard deviation of sl2 grouped by tn2 is 11.60572722209523

[0062] The standard deviation of sl1 grouped by tn1 is 7.972101509411196

[0063] The standard deviation of sl1 grouped by tn2 is 9.367583262459625

[0064] The standard deviation of sl1 grouped by tn1 is 9.737578730713613

[0065] The standard deviation of sl2 grouped by tn2 is 9.759835847631278

[0066] The above different groups are divided into 4 groups of data, as shown in the following table. The matching degree analysis is carried out on each group of data respectively, which is actually a comparison within each group, and a total of four groups of data are compared. Combining the above 4 groups and the difference between the standard deviations in each group, the smaller the difference between the standard deviations in each group, the smaller the discrete degree between the tn and sl data, and the higher the corresponding matching degree.

[0067] From the results of the above two cases, it can be seen that the difference between the standard deviation of sl2 grouped according to tn1 and the standard deviation of sl2 grouped according to tn2 is the smallest. Therefore, sl1 corresponds to tn1, and sl2 corresponds to tn2.

[0068] Group 1 tn1 sl1 tn2 sl1 Group 2 tn1 sl2 tn2 sl2 Group 3 tn1 sl1 tn2 sl1 Group 4 tn1 sl1 tn2 sl2

[0069] The amount of data collected during the production process is huge. Whether the machining tool parameters are stable and whether the machining process is normal is generally determined after the entire workpiece is processed. This is done by checking whether the data trend during the entire machining process is similar to that of workpieces of the same type. Alternatively, a machining cycle is extracted from the data of one day, and the data of the entire cycle is analyzed to determine whether the machining is normal. However, the above cannot achieve analysis and judgment during the machining process.

[0070] On the basis of ensuring the accuracy of signal matching, the present invention starts with data segmentation, extracts segmented data, and then performs comparative analysis according to the one-to-one correspondence of data positions.

[0071] Firstly, the processing time of each workpiece at different tool positions is grouped, and each group of processing time is divided into segments at the same time interval to obtain segmentation intervals.

[0072] Specifically, the data is first classified according to cc, where cc represents the processing workpiece number;

[0073] Secondly, classify the data according to the tn1 number. For example, if the tn1 tool position is switched from 1-3-1..., the processing time of the first occurrence of 1 is the first interval, recorded as t1; the processing time of the next occurrence of 3 is the second interval, recorded as t2; the processing time of the second occurrence of 1 is the third interval, recorded as t3, and so on, to the interval t of the last processing tool number. n ;

[0074] Then, t i ,i=1,2,L divides the data according to the time interval of 2s as a small interval. If the length of the last small interval after division is less than 2s, it will be discarded. After the operation, the data is divided into t ij ,i=1,2,L n,j=1,2,L m; represents multiple 2s interval segmentation intervals.

[0075] For each 2s segmentation interval, the processing parameters such as variables ss1 and sl1 are collected, ss1 is the tool speed, and the data feature values ​​in each segmentation interval are extracted, such as median, mode, mean, variance, etc., that is, the large amount of data originally within 2s is reduced to one-dimensional data;

[0076] For each workpiece, select the same interval data of the same tool position, such as the t 11Indicates that the processing tool position is No. 1, the first 2s segmentation interval data, integrates the data on the same tool position of each workpiece, that is, integrates the data of each workpiece on the same tool position in different segmentation intervals together, forming data in which each row represents a workpiece and each column is the characteristic value of the processing process data. By analyzing the data, it can be determined whether the processing is smooth and normal.

[0077] The processed data is shown in the following table:

[0078]

[0079] The first row in the table shows the first 2s processing time when the workpiece number (cc) is 1 and the tn1 tool number is 9 (indeed 9), and the ss1 and sl1 values ​​after feature extraction; the second row shows the second 2s processing time when the workpiece number is 1 and the tn1 tool number is 9, and the ss1 and sl1 values ​​after feature extraction. Then select the first data of tn1=9 of each workpiece (that is, the feature value extracted from the first 2s processing data) as a data, and plot sl1, as shown below: Figure 3 shown.

[0080] exist Figure 3 The data first fluctuated around 11, then had a big drop, and then stabilized around 7. After considering the actual data, it was found that the time when the drop occurred was the process of tool damage and tool replacement. According to past experience, when the tool is severely worn, the tool load of the workpiece will be higher. After the tool is replaced, the tool load of the new tool for processing the workpiece will be reduced, which can also be intuitively reflected in the figure.

[0081] Similarly, select the second data of tn1=9 of each workpiece (i.e., the feature value extracted from the second 2s of processing data) as b data. Plot sl1, as shown in Figure 4 shown.

[0082] Under the same conditions, tool load and Figure 3 The only difference is that there is a large fluctuation at 11100. Through the detection of raw data, it can be seen that the tool load in the current machining process is high, which can be regarded as an abnormal situation in the machining process.

[0083] Combine the two images, see Figure 5 .from Figure 5 It can be seen more intuitively that the processing data trends at different time points in the same processing process are similar. Therefore, the original data can be reduced in dimension by grouping the data and collecting eigenvalues, and the collected data can be guaranteed to be scientific and effective.

[0084] The time interval between cells is an adjustable parameter. In addition to the 2s given above, it can also be adjusted according to the time used in different processing processes. In actual practice, it can also be set to 3s, 5s or other times to meet the actual needs under different processing conditions.

[0085] After determining the abnormal situation of the machining process, the machining data is fitted through the prophet model and the abnormal data is located.

[0086] In the field of time series analysis, there is a common analysis method called Decomposition of Time Series, which divides the time series y into t It is divided into several parts, namely the seasonal term S t , trend term T t , the remaining term R t , that is, for all t ≥ 0, we have

[0087] y t =S t +T t +R t

[0088] In addition to the addition form, there is also the multiplication form, which is:

[0089] y t =S t ×T t ×R t

[0090] The above formula is equivalent to lny t =lnS t +lnT t +lnR t Therefore, sometimes when predicting a model, we first take the logarithm and then decompose the time series to get the multiplication form. The prophet algorithm makes necessary improvements and optimizations based on this method.

[0091] Generally speaking, in real life and production, in addition to seasonal terms, trend terms, and residual terms, there are usually holiday effects. Therefore, in the prophet algorithm, the author considers the above four terms at the same time, namely:

[0092] y(t)=g(t)+s(t)+h(t)+ε t

[0093] g(t) is the trend function and represents non-periodic variation in values, s(t) represents periodic variation (such as weekly and yearly seasonality), and h(t) represents the effect of holidays that occur on a potentially irregular schedule. The error term εt Represents any special variation that the model cannot accommodate and assumes a normal distribution.

[0094] Trending Items

[0095] Prophet implements two trend models, namely the saturated growth model based on logistic regression and the piecewise linear model. The first is the trend term based on logistic regression:

[0096]

[0097] C is the bearing capacity, k is the growth rate, and m is the displacement parameter.

[0098] With the development of hardware and the upgrade of information volume, neither the carrying capacity nor the growth rate will be constant, so the model modifies the fixed carrying capacity and growth rate into a carrying capacity and growth rate function that changes with time, C=C(t), k=k(t).

[0099] In reality, time series usually have change points, that is, where the trend changes. When the change point is set, the model is a piecewise logistic regression model:

[0100]

[0101] where a(t)=(a1(t),L,a S (t)) T ,δ=(δ1,L,δ S ) T ,γ=(γ1,L,γ S ) T .

[0102] δ i Represents the rate of change over time, and the growth rate over time is k+a T δ,

[0103] in

[0104] Seasonality

[0105] prophet uses Fourier series to build a periodic model:

[0106]

[0107] P represents the period, and the parameter can be expressed as β = [a1, b1, L, a N ,b N ] T .

[0108] The adjustment of N acts as a low-pass filter. Therefore, the seasonal term of the time series is: s(t) = X(t)β and the initialization of β is β:Normal(0,σ 2 ), indicating that β has a mean of 0 and a variance of σ 2 The larger the value of σ, the more obvious the seasonal effect; the smaller the value, the less obvious the seasonal effect.

[0109] Events (holidays) usually have a significant impact on time series, such as National Day. Prophet has built-in holidays for many countries, and users can also customize holidays.

[0110] in κ=(κ1,L,κ L ) T .

[0111] Where κ: Normal(0,v 2 ) means that κ has a mean of 0 and a variance of v 2 The normal distribution of is affected by the index v. The default value is 10. When the value is larger, the impact of holidays on the model is greater; when the value is smaller, the effect of holidays on the model is smaller. Users can adjust it according to their own situation.

[0112] Model Fitting

[0113] According to the above explanation, our time series can be constructed by growth term, seasonal term and holiday term, ie

[0114] y(t)=g(t)+s(t)+h(t)+ε t .

[0115] The prophet model is used to fit the data in the following two tables respectively, and the results are as follows Figure 6 , Figure 7 As shown, the distribution of outliers can be clearly seen.

[0116]

[0117]

[0118] In the above two tables, a row of data represents a processing data of a workpiece, where raw represents the original processing data ( Figure 6 The black dots in the figure), yhat represents the data predicted by the prophet model ( Figure 6 middle curve), t represents the current processing time.

[0119] After fitting the data, the key is how to determine if there are abnormal values ​​in the processing process. Here, it is necessary to remove the trend and periodicity of the time series. The residual of the sequence is calculated by subtracting the predicted value given by the prediction model from the true value. The residual does not contain periodicity and trend, and can be used for subsequent anomaly detection. The formula is as follows:

[0120] in is the predicted value, y is the actual data, and error is the residual.

[0121] In the present invention, the abnormal data is located by the 3σ criterion, and the abnormal value is a value in a group of result values ​​whose deviation from the average value exceeds three times the standard deviation.

[0122] The 3σ principle is commonly used in engineering science. This principle has been used in fault warning. The data is collected by sensors. These data are assumed to conform to the normal distribution. Then, after model identification, the 3σ principle is used to accurately locate abnormal data. In practical applications, the k value in k-sigma can be determined according to the business scenario.

[0123] The 3σ principle can be simply described as follows: if the data follows a normal distribution, an outlier is defined as a value in a set of result values ​​that deviates from the mean by more than three times the standard deviation. That is, under the assumption of a normal distribution, the probability of a value that is three times σ (standard deviation) away from the mean value is very small (as shown in the following formula), so it can be considered an outlier.

[0124] P(|error-μ(error)|>3σ)≤0.003.

[0125] After 3σ processing of the data in the above table, the results are as follows: Figure 8 , Fig. 9 For example Figure 8 The curve in the figure corresponds to the difference between raw (original data) and yhat (prophet predicted data) in the first table; the three marked points are detected outliers; the two straight lines correspond to the upper and lower limits of the interval (μ-3σ,μ+3σ) constructed by the 3σ principle. The situation of the outliers is as follows:

[0126]

[0127]

[0128] Similarly, Fig. 9 Corresponding to the data in the second table, the outliers are as follows:

[0129]

[0130] The above scheme constructs a 3σ interval for the overall data. At the same time, historical data and current real data can be used for adaptive 3σ analysis. In this way, the constructed interval will change with the change of data. Fig.10 .

[0131] In the actual production process, due to factors such as tool wear, the load signal sl will gradually increase over time. Then the 3σ interval constructed by the overall data will gradually deviate over time. Therefore, it is necessary to use the current data for adaptive 3σ analysis.

[0132] The principles of adaptive 3σ analysis include:

[0133] Where i>k, k is a custom constant. The current processing data is x i , select the first k processing data x i-k ,x i-k+1 ,L x i , calculate the adaptive 3σ interval. When i≤k, the 3σ interval is calculated using the overall data.

[0134] By adopting adaptive 3σ analysis, abnormal points in the processing process can be located more accurately, improving the accuracy of judgment.

[0135] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for analyzing industrial data quality, characterized in that: The following steps are involved: (1) Collect processing data and perform data verification to match the processing data with the processing tool; (2) Group the processing data and extract the feature values ​​of the processing data to identify abnormalities in the processing process; (3) Fit the processed data through the prophet model and locate the abnormal data; In the step (1), each machining tool is grouped in the form of machining intervals at different tool positions according to the machining order, and the tool load corresponding to the machining tool in each machining interval is extracted; The processing data on the same continuous tool position will be divided into the same processing interval. The processing intervals with the same tool position but not continuous cannot be merged. Calculate the standard deviation of tool load in each machining interval of each machining tool, and combine all machining intervals to calculate statistics. The matching degree between machining data and machining tools is determined by the discrete degree of statistics, including: Group sl1 and sl2 respectively according to tn1, and divide sl1 into multiple processing intervals, recorded as sl1 i ,i=1,2,...l1, the number of parameters contained in each interval is recorded as nl i ,i=1,2,...,l1; Similarly, sl2 is divided into multiple intervals, denoted as sl2 j ,j=1,2,...l2, the number of parameters contained in each interval is recorded as n2 j ,j=1,2,...,l2; where tn1 represents the machining tool at different tool positions, sl1 and sl2 are tool loads; Calculate sl1 separately i The standard deviation of , and all intervals are combined to form a statistic: where std(sl1 i ) indicates sl1 i The standard deviation of For sl2 j Similarly, calculate: where std(sl2 j ) means sl2 j The standard deviation of Compare t1s1 and t1s2. If t1s1<t1s2, it means that after grouping by tn1 tool number, the discrete degree of sl1 data interval is lower than that of sl2, indicating that in the processing without tool change, sl1 load is more balanced. Therefore, it is judged that tn1 and sl1 are more matched.

2. The industrial data quality analysis method according to claim 1, characterized in that: In the step (2), the processing time of each workpiece at different tool positions is grouped, and each group of processing time is divided into divided intervals at the same time interval.

3. The industrial data quality analysis method according to claim 2, characterized in that: In the step (2), data characteristic values ​​of the processing parameters in each segmented interval are extracted, and the data characteristic values ​​include median, mode, mean or variance.

4. The industrial data quality analysis method according to claim 3, characterized in that: In the step (2), the different segmentation intervals of each workpiece on the same tool position are integrated, and whether the machining process is abnormal is determined by judging the data feature value.

5. The industrial data quality analysis method according to claim 4, characterized in that: In step (3), anomaly detection is performed by calculating the residual between the prophet model prediction value and the integrated production data in step (2), and the residual calculation formula includes: in is the predicted value of the prophet model, y is the integrated production data in step (2), and error is the residual.

6. The industrial data quality analysis method according to claim 5, characterized in that: In the step (3), the abnormal data is located by using the 3σ criterion, and the abnormal value is a value in a group of result values ​​whose deviation from the mean value exceeds three times the standard deviation.

Images