Anomaly detection for yield improvement in glass production

Predictive maintenance using manufacturing sensor data has attracted attention for reducing defects and selecting appropriate actions. This paper proposes an anomaly detection method using lasso regression and group-wise variable selection based on FTA (Fault Tree Analysis) domain knowledge. We evaluated our approach using real factory data and found that its precision and false positive rate are 66% and 30%, respectively. Moreover, we validate that the visualization of the contribution rate for anomaly detection is helpful for factory maintenance.


INTRODUCTION
Flat glass manufacturing facilities consist of melting furnace that melts raw materials at high temperature, float bath that form flat glass on molten tin, and annealing furnace that cool glass gradually.These facilities are hot and sealed, and it is difficult to monitor their internal conditions directly, so they are managed using various types of sensing information.At the startup of the production line, the production equipment is set to an appropriate state to prevent the production of defects.However, as glass production continues, the equipment condition gradually deviates from the appropriate production conditions, and defects may occur.Therefore, we have decided to develop an online detection method for detecting signs of abnormalities in manufacturing equipment based on sensing information, so that appropriate equipment maintenance actions can be taken based on early detection of abnormal conditions of the equipment.
Figure 1 Glass production facility (from AGC HP)

Flow of Proposal Method
Figure 2 shows the overall structure of the abnormality sign detection method.Process data is periodically transferred from the manufacturing line to the database.The new abnormality detection method obtains process data from the database and outputs a health index (abnormal degree) to the database.When the visualized health index obtained from the database indicates an abnormality, appropriate manufacturing actions are taken at the factory process site.4th Asia Pacific Conference of the Prognostics and Health Management, Tokyo, Japan, September 11 -14, 2023 OS05-04 Figure 3 shows the health index calculation flow for the abnormality level.Using process data under normal quality conditions as training data, a prediction equation is created using the LASSO model, and the health index is calculated from the residual difference between the predicted value and the actual measured value.Details are described in Chapter 2.4.In the predictive creation model study, in addition to LASSO, Auto-associative kernel regression model and Autoencoder were tested.Among them, LASSO, which obtained the most consistent results, was adopted as the current model.
Figure 3: Health Index Calculation Method

Training Data and Estimation Goals
The goal was to create a method that can manage current production conditions not to deviate from the training data.

Figure Training data and anomaly data
The sensed process data used were stored in a database at a sampling rate of once every 10 minutes.The health index was calculated at the same time intervals and monitored online during production to allow for proactive response when abnormal signs were detected.The target detection time for abnormal signs was set to 2 hours before an abnormality occurred.
Figure 5: Health Index Schematic Graph

Variable Selection Based on Causal Analysis
The defects that were the subject of this anomaly detection method had many factors and had been a problem for a long time.First, FTA was used to analyze the causal relationships that generate the defects, and to estimate the "physical quantity" that causes the defect and the "location" where the defect is generated.From the FTA, the important sensor groups and the objective and explanatory variables of the regression analysis were determined for each failure cause classification (Figure 7).Failure causes were categorized by location (Part 1-4) and by the four variables (a measured control value, an external measurement, a control setpoint, an internal measurement), finally creating 10 variable groups.In the glass process flow, Part 1 ~ 4 of the four equipment sometimes make defects.Glass is being transported from Part 1 upstream of the process to Part 4 downstream of the process continuously.Also, each facility has a control device, which controls variables to be managed during operation and measures whether they can be controlled correctly.Furthermore, many variables inside the equipment and outside the equipment have also been measured.And the causal relationship to the current anomaly can be explained with 4 variables among them using LASSO regression and domain knowledge.The health index was calculated for each of these groups, and the largest health index (means highest abnormality) from the group was the OverAll health index.In the production site, when the OverAll health index exceeded the threshold, it was determined to abnormal

Analysis Anomaly Detection Algorithm
The original program of the method was commissioned to Predictronics Corp. and customized in-house at AGC.The HI (health index) was calculated by determining the residuals from the predicted values and the measured values that were obtained for each variable group.Furthermore, from each health index, the contribution of the variables used for each health index was calculated.The equations ( 1 Thresholds were set using ROC curves (Figure 9).Specifically, the threshold value was the value at which the difference between the precise rate and false positive rate of the Overall health index was the largest value.The intent was to maximize method performance in this way.The predictive performance during the approximately oneyear trial period was 66% precision rate and 30% false positive rate.Our goal was to achieve a precision rate of 80% or higher and a false positive rate of 30% or lower.Although the precision rate fell slightly short of our goal, we were able to raise the values quite close to our target.We would like to further improve the precision rate and false positive rate in the future.

Validation of Variable Selection
For each health index variable group, we confirmed whether the appropriate explanatory variables were selected and whether deviations from the training data were consistent with the occurrence of product defects.Figure 10 Health index configuration for variable groups 2 and 9 Figure 11 shows the health index actual performance graphs for variable groups 2 and 9 as an example.There is a good agreement between the time of product defect occurrence and the actual results of health index.Thus, it was found that not only an increase in the health index of a single part, as in variable group 2, but also a change in the relationship between variables of multiple parts, as in variable group 9, can cause the health index to increase, and also, product defects occur actually.Figure 10 shows the health index structure of variable groups 2 and 9.
Figure 11: Health index results for variable groups 2 and 9

Evaluation of the contribution of each variable
When the Overall Health Index exceeds the threshold, it is determined that an abnormality has occurred in the manufacturing process, and furthermore, manufacturing actions are taken to check which variable group is abnormal and to ensure that the part and physical quantity of that variable group is normal.In addition, within the variable group that shows abnormality, we referred to which variable worsened the health index the most by looking at the contribution ratio.
Figure 12 shows the percentage contribution of each variable in the Health Index of Variable Group 9.This health index was calculated for Part1-4 variables.The contribution rate of Part 4 in the abnormality at this time was relatively low compared to the other Parts 1-3, indicating that it was the manufacturing actions of Parts 1-3 that improved the health index.By examining the contribution rate in this way, we can learn the direction of manufacturing actions for improvement.

CONCLUSION
As an abnormality detection method for improving the yield rate of glass manufacturing, we have implemented the abnormality sign detection method that foreshadows abnormality from a variety of variables by setting up a group of variables.We have confirmed that this method is effective as a method for detecting signs of abnormality in large facilities such as glass manufacturing where there are many sensing values.We would like to utilize this system for other product defects and equipment maintenance in the future.

Figure 2 :First
Figure 2: Overall structure of the detection method for abnormal signs

Figure 6 :
Figure 6: Causal analysis resulting in glass defects (FTA) control measurement values, control set values, internal measurements, and external measurements.

Figure 7 :
Figure 7: Contents of each variable group

Figure 8
Figure 8 Definition of precise rate and false positive rate

Figure 9
Figure 9 Determination of Overall H.I thresholds using ROC curves

Figure
Figure shows the health index structure of variable groups 2 and 9. Variable Group 2 consists of the measured control values of Part 2, and Variable Group 9 consists of the measured control values of Parts 1 through 4. Variable group 9 expresses the balance of the measured control values of Part1 to 4.
Figure 12 Variable Group 9 Health Index Contribution of each variable