(348c) Fuzzy Model Based Industrial Batch Dryer Data Analysis

This work describes a pattern-recognition based data analysis of a drying process. The work is motivated by the fact that it is needed to identify the start of the drying process in order to calculate the processing time. The processing time is calculated as the difference between two start-ups (vacuum checks) of subsequent batches, based on the plant measurements. The available plant data consist of basic process measurements (temperature, pressure) without tags which could indicate when a new batch is started. In order to calculate the time elapsed between two vacuum checks a pattern recognition algorithm is developed and applied using the plant data compressed by the PI algorithm. It is considered that based on the absolute pressure value and its gradient it is possible to uniquely identify the vacuum check stage and its pattern. These two parameters will constitute the input space of the pattern recognition problem. The output of the classifier is the degree of membership to the pattern to be recognized, e.g. 0 is the value if the pattern does not belong to the vacuum check pattern, and 1 is the value if it does. In order to perform a supervised training of the data we use the two inputs and the above mentioned output. The next step is the development of a black-box model in the form of a Takagi-Sugeno fuzzy black-box model. After the classifier development the plant data was subjected to classification with the calculated model. Most of the data was classified as belonging or not to the vacuum check profile (output 1 or 0). However, a significant number of data had an output value between 0 and 1. This is due to the fact, that the patterns were not defined with perfect accuracy and because of some data matching errors. In order to assign the points with an intermediate membership values to one of the patterns, a threshold value of 0.7 is chosen. All the data above this value was assigned to the vacuum check pattern and the data below to another process operation pattern. The method was successful in recognizing the vacuum check patterns, and it made a good classification. The start times of the dryer (vacuum checks) were successfully identified and by calculating the time difference between two patterns the drying time was calculated. This data analysis method helped to identify the large drying time variations which are particularly problematic since the dryer is the production bottleneck in the analyzed batch plant. For the tested data set, the pattern recognition method identified 160 batches, 2 process anomalies and it made one classification error; therefore its performance is over 99%. Using this classification method we where able to quantify the variation in drying time which amounts to an average of 43% compared to the minimum drying time (best case batch).