CC BY
18II. ИНФОРМАЦИОННЫЕ СИСТЕМЫ. АВТОМАТИЗАЦИЯ И СИСТЕМЫ УПРАВЛЕНИЯ
УДК 004.622
M. Kohlert1, T.B. Chistyakova2
ADVANCED PROCESS DATA ANALYSIS AND ON-LINE EVALUATION FOR COMPUTER-AIDED MONITORING IN POLYMER FILM INDUSTRY
St. Petersburg State Institute of Technology (Technical University), Moskovsky Pr., 26, St Petersburg, 190013, Russia Mondi Gronau GmbH, Joebksweg 11, 48599 Gronay, Germany e-mail:chistb@mail.ru
Modern polymer manufactures are confronted high costs of raw materials or energy, high technological complexity, monthly changing production range, high quality requirements, lack of machine workers' knowledge on real-time management. Polymeric films production control and monitoring systems use great arrays of production data and decentralized heterogeneous data storage devices and lack qualified machine workers and quality engineers. Within all the production data only 5% of manufacturing data is used for error cause analysis due to missing of data source linkage. Therefore, the resulting conclusions about production, quality, and engineering surveillance are often biased. Increasing losses of money in the range of billions of dollars/euros are caused by the previously mentioned problems leading to material waste, customer claims for damages, safety issues, and high energy consumption. This work selects the above-mentioned topics with high influence on material waste to apply them to polymer film industry case study: material waste and process monitoring.
Keywords: novelty detection, multi-sensor data analysis, monitoring, recommendation, real-time evaluation
М. Колерт, Т.Б. Чистякова
ИНТЕЛЛЕКТУАЛЬНЫЙ АНАЛИЗ И МОНИТОРИНГ В РЕЖИМЕ РЕАЛЬНОГО ВРЕМЕНИ ХАРАКТЕРИСТИК ПРОИЗВОДСТВА ПОЛИМЕРНЫХ ПЛЕНОК
Санкт-Петербургский государственный технологический институт (технический университет), Московский пр., 26, Санкт-Петербург, 190013, Россия ООО «Монди Гронау» е-таН:сЬ^Ь@таИ.ш
Современные производства полимерных пленок характеризуются высокой стоимостью сырьевых материалов и энергии, сложностью технологических процессов, ежемесячным изменением ассортимента продукции, высокими требованиями к качеству продукции, недостатком у оперативного персонала знаний, для управления в режиме реального времени. Системы управления и мониторинга в производстве полимерных пленок характеризуются использованием больших массивов данных и децентрализованных, гетерогенных устройств хранения информации, недостатком квалифицированных операторов и инженеров по качеству. Из всего массива производственных данных только до 5% параметров производства используется для анализа ошибок и брака из-за сложных связей между источниками данных. Поэтому результирующие выводы о качестве продукции и данные инженерного контроля зачастую оказываются искаженными. Указанные выше проблемы, ведущие к потерям сырьевых материалов, жалобам клиентов на качество продукции, возникновению проблем безопасности и высокому потреблению энергии, являются причиной возрастающих потерь денежных средств в размерах миллиардов долларов/евро. Предложенные методы интеллектуального анализа данных производства больших объемов в режиме реального времени для мониторинга и принятия решений по управлению производством позволяют снизить потери сырьевых, энергетических ресурсов и материалов.
Ключевые слова: прогнозирование, мультисенсорный анализ данных, мониторинг, рекомендации, реальное время
1 Kohlert Michael, Head of Information and Automatization Department, Mondi Gronau GmbH e-mail: michael.kohlert@mondigroup.com
Колерт Михаель, руководитель отдела управления информацией и автоматизации ООО «Монди Гронау», e-mail: michael.kohlert@mondigroup.com
2 Chistyakova Tamara B., Dr Sci. (Eng.), vice rector for innovation, Professor, Heat of the Department of Computer-Alded Design and Control Systems, e-mail: chistb@mail.ru
Чистякова Тамара Балабековна, д-р техн. наук, профессор, проректор по инновациям, зав. каф. систем автоматизированного проектирования и управления СПбГТИ(ТУ), e-mail: chistb@mail.ru
ReceivedMay 29, 2015
Дата поступления - 29 мая 2015 года
DOI: 10.15217/issn998984-9.2015.29.83
Introduction
Modern industry systems' development strategy (Industry 4.0) is aimed on application of artificial intelligence and expert knowledge for industry system's control, especially power-consuming with high production quality requirements. Blown extrusion is one of the most difficult to control, has the biggest number of parameters and power-consuming from all the polymeric films production methods.
The blown film extrusion usually has a lower dosing stage feeding the extruders with, e.g., polyethylene. The molten granulate is injected into a ring that forms the blow (warm temperature inside, cooled from outside), which moves up to a traversal. There, the blow is pressed to a film by traversal rotation and then guided to the winder below (Figure 1).
The blown film extrusion process has low material waste and stable process conditions, in comparison to cast extrusion. A regular machine generates about 600-1000 kg output per hour (10-20 t per day), with an energy consumption of 250 kW/h (efficiency 0.6 kWh/kg).
About five programmable logic controllers (e.g. S5, S7) regulate the machine behavior, such as temperature and speed, monitored and adjusted by the machine staff. Additional quality measurement systems are integrated for statistical process control(Figure 2) [1].
Figure 1: Blown Film Extrusion Line Overview
Figure 2: Optical Control System [OCS15]
For this purpose, sensor applications are installed for specific attribute monitoring, such as optical control systems (defects: gels, holes, contaminations) and color (L*a*b) or thickness measurements (Mm).
Datasets and Acquisition
Polymer manufacturing machines, equipped with hundreds of sensors from the dosing stage to the winder stage, collect heterogeneous sensor data (e.g. on temperature, pressure, speed, or energy consumption) from the machine controllers (e.g., S7) via TCP-IP (Figure 3). To increase efficiency of applied control systems modern methods and solutions for data processing are used [2-4].
Parameters
Cast Extrusion Line
Sensory Informationen 20 - 300 Parametersat each stage
Acquisition SPS and Data Collector
Winderl Winder! _Y
Sensory Sample Points: The earlier a possible prediction the lower the costs Figure 3: Sensory locations for datasets in the extrusion process of rigid film production: 1. dosing level, 2. extruders, 3. Winder
Within each production minute, sensor data is stored on database servers in specific format types (e.g. numerical or char values) validated by technical experts for further downstream process analysis.
The applied preselected dataset of a total of 21.900 minutes includes normal and conspicuous data taken from three month of typical elastic extrusion production by supervised selection with about 160 dimensions (sensory locations) for this work describes the extrusion process for 120 pm flat-film. The data acquisition was achieved by data collectors from the machine control with each dimension corresponding to one sensory point during the extrusion process. The available (experimental) data consists of three datasets,
one dataset used for the missing data problem, and two datasets used for the faulty data problem.
Experimental Methods and Results
Two approaches were investigated, first the one-class approach to recognize abnormal condition states, and second building up on the one-class results a trajectory approach for earlier prediction of abnormal condition states with a two-class classification by analyzing condition states in sequence. In the following, the results for both types of approach are presented.
The complete approach cycle focusing on the settings for the knowledge discovery processes is presented in Figure 4.
Figure 4: Complete Processing Cycle for novelty and trajectory prediction
The experimental datasets for the normal and abnormal conditions were executed by hold-out for off-line training (70% of the data), validation (15%), and testing (15%). The tested methods are listed in the following result tables.
As shown onFigure5, the best settings for the AutoEncoder Neural Networks were achieved with 8 hidden neurons and a rejection rate of 0.00001, leading to the deterministic empirical, lowest repeatedly received specificity rate of 97.6% after the execution of 10 runs, with an experimental standard deviation (STD) rate smaller 2.5% [5, 6].
Methods OCC Dataset Condition Mean in STD In
Class [%] [%]
1 Neural Networks Dataset 1 Normal 97.6 0.4
(Auto-Encoder) Dataset 2 Normal 99.1 0.1
Dataset 3 Normal 99.5 0.4
2 Support Vector Dataset 1 Normal 99.9 0.1
Machines (RBF) Dataset 2 Normal 99.6 0.2
Dataset 3 Normal 99.S 0.1
3 /¿-Nearest-Neighbor Dataset 1 Normal 98.8 0.2
Classifier Dataset 2 Normal 98.7 1.1
Dataset 3 Normal 99.9 0.1
4 NOVClass Dataset 1 Normal 99.8 0.1
Dataset 2 Normal 99.8 0.2
Dataset 3 Normal 99.4 0.1
Figure 5 Experiments' specificity with 4 occ tested methods, and 3 datasets (only normal datasets used)
The normal conditions include the following process behavior: < 5% waste, output of > 1000 kg/h, fast production with > 100 m/min, energy efficiency of < 0.6 kWh/kg, low maintenance costs, best roll quality with 0 incidents, and no additional human resources necessary in contrast to abnormal states. In comparison, the abnormal conditions lead to the opposite results, with low productivity due to high rework and capacity problems for the divisions.
The requirement in the polymer film industry for efficient control through a recognition system is 90% for the specificity rate, which defines the error of the target class to become an outlier; this specificity is similar to that of the optical control systems for defect detection, set up in cooperation with clients. All Optimistic concurrency control (OCC) methods achieved more than 90% specificity. The elapsed time for recognition is 13 s for one machine, and about 69 s for seven more machines, depending on the number of attributes to be analyzed. In the following, the next step new datasets were acquired with OCC methods and a second classification approach examined, the trajectory recognition.
In Figure 6 the sequence of condition states is visualized within a multi-dimensional scaled view.
Fig. 6: Trajectory Approach on mapped view
The methods (neural networks, support vector machines, k-nearest-neighbor classifier, naive Bayes, automated network search (MLP), boosted trees, NOVClass) were investigated off-line. New datasets were extracted from the previous novelty investigation for a trajectory analysis, associating datasets in a time chain with the previous staged novelty detection, with two classes with similar amounts of datasets.
The investigated features deduce machine properties from the extrusion stage. The best settings for the normal and abnormal trajectory directions, displayed as green (left direction) and red (upward direction) part within the visualization, were achieved by the boosted trees method (number of trees = 186) and, for the examined data (tree size = 3), are presented in Figure 7. Boosted trees achieved the best specificity rate for the acquired datasets with 100% for normal datasets and 99.7% for abnormal datasets. The accuracy for normal datasets was low due to the different types of trajectories for normal conditions and their starting behavior, which is very similar to the abnormal datasets; however, the abnormal dataset results were in the focus of this case.
Condition Mean in STD in Class [%] [%]
1 Neural Networks (RBF)
Dataset4 DatasetS
Normal Abnormal
88.4 92.2
0.2 0.2
Support Vector Machines Dataset4 (RBF) DatasetS
Normal Abnormal
84.0 94.6
0.2 0.6
3 ¿-Nearest-Neighbor Classifier (fc = 5)
Dataset4 DatasetS
Normal Abnormal
80.4 94.0
1.6 2.0
Naïve Bayes
Dataset4 DatasetS
Normal Abnormal
61.2 94.1
2.3 2.0
5 Automated Network Search (MLP)
Dataset4 DatasetS
Normal Abnormal
95.8
0.2 0.1
Boosted Trees
Dataset4 DatasetS
Normal Abnormal
100 99.7
0.0 0.7
Dataset4 DatasetS
Normal Abnormal
92.2 97.1
1.2 1.6
method, dividing the set of data into subsets best fitting to the target class with a low amount of settings.
One approach for off-line analysis of the quality of process sensor property prediction with a trajectory classification system based on a previous OCC analysis of machine sensor data was improved by additional classification methods. After best adjustment of the classifiers to the datasets, the methods should be implemented into a novel design graphical user interface (GUI), with e.g. Matlab& Simulink, supporting the online process control by the workers, besides optical control systems and laboratory results [7].
Interactive Visualization and Recommendation System
According to the investigated approaches and the worker interview, the methods were integrated into a graphical user interface on the laboratory machine, to permanently visualize the results at the machine terminal computer for direct monitoring, which has not been established before. Therefore, the ideas from a workers' feedback were applied. The realization within a process monitoring environment was accomplished according to the system architecture described in the end.
The layout for the software, constructed as shown in Figure 8, includes the previously described settings, such as the data acquisition part, the classification methods, the software interface, the processing computer hardware, the visualization elements, and the message system.
Figure 7: Experiments with 7 learning methods, and 2 datasets (normal and abnormal)
All repeatedly achieved mean results after ten runs were monitored. Automated network search (MLP) and boosted trees achieved more than 90% accuracy for abnormal datasets, which corresponds to the accuracy of optical control systems for visual defect detection. The boosted trees, in comparison to neural networks, run efficiently on large datasets, are robust to noise and can be extended to unsupervised unlabeled data approaches. Mathematically, it is a gradient
Fig. 8: Layout for Standalone Application
The data acquisition and processing parts strongly depend on the performance ratio of the storage system and the computer hardware used [8]. In-Memory storage technology and a 64-bit computer system are recommended to achieve fast acquisition and recognition results, visualized and transferred within 1 min for at least 200 attributes such as pressure, temperature, speed, output, order number, recipe, and winder settings, and calculated attributes such as pressure gradient and amplitude, which are mostly double and string dataset types.
The real-time monitoring of trajectories and conditions was implemented as a standalone application close to the manufacturing process. A monitoring interface at the machine terminal supported the worker with graphical evaluations and recommendations for process intervention.
The complete laboratory setting was transferred to
terminal computers directly at the machine. Besides monitoring, the terminal interfaces are mostly used for manufacturing execution system inputs, specification modifications, or communication. All terminals are directly linked to a machine network, which operates in parallel to the business network as it is shown on Figure 9.
A total view gives information on all machine condition states by traffic light and/or by sending recommendations via e-mail or SMS if the current dataset classification result changes from normal to abnormal conditions: green (Good operation mode), yellow (Warning! Critical condition!), and red (Shut down machine). The operating condition is adjustable by changing the temperatures (170-200 °C) in the extrusion zones, instigating lower settings for the speed or the torque, increasing the dosing, or decreasing the percentage of the returned re-granulate by the worker. At the machine terminals, the staff is able to monitor the current process data and receives recommendations in case the condition state changes in the next few minutes (Figure 10).
Figure 9: Complete Setup for Computer-Aided Monitoring & Recommendation System
\mondi
Figure 10: Graphical User Interface for Monitoring & Recommendation
System Architecture
The system implementation overview with regard to the previously explained particular Layers, the Acquisition Layer (A), the Monitoring & Recommendation Layer (B), and the Adaptation Layer (C), are currently applicable to the mentioned object (machine).
The conceptual design with the current implementations marked in green is displayed on the left; on the right side, the detailed architecture with the also green-marked currently implemented layers and the grey-shaded future approaches is presented.
In this case (Figure11), first regarding the Acquisition
Layer (A), a Smart Production and Digital Factory has to be established. The acquired 21,900 datasets from multiple PLCs of eight extrusion machines are then transferred to the Oracle database through the TCP/IP local area network and queried by the software tool Matlab for further investigation. This part describes the network and processing structures from Sections 2 and 3 that surround the machinery of the plant. The realized software workspace settings offer update functions for cyclic reloading of new datasets. Such direct connections form the block gateway from (A) to the Monitoring & Recommendation Layer (B), with transferred information.
Figure 11: System Architecture
The following discussion reflects the advantages and disadvantages of the current approach.
The material waste reduction as the main target was achieved, although only one specific problem was regarded and solved. The investigation should be extended to more particular critical cases from the polymer production process, to demonstrate the general validity of the system in predicting the occurrence of unknown conditions.
Regarding the sensor data points, the acquired information was not tested with regard to sensor drift effects and quality degradation. The diagnostic procedures were not investigated in this work; such an investigation could prevent system failures due to low data quality.
The restriction to 6 dimensions out of 160 at an early processing stage (2 years ago) is questionable due to the limited time range of the acquired datasets for feature reduction. The recognition rate depends on the datasets used for training; therefore, a cyclical training loop verification should perhaps be additionally integrated in order to recalibrate and dynamically adapt the complete system architecture by extension of the training cases. Further on, the threshold setting for the boundary zone does not exactly display the class separation in the regarded dimension. Therefore, the trajectory acquisition by equal squared segmentation without investigation of other possible approaches could weaken the result stability.
References
1. Firm Description. OCS Optical Control Systems GmbH 2015. URL: http://www.ocsgmbh.com/ (дата обращения 01.01.2015)
2. Программный комплекс для автоматизированной обработки измерений и исследования качества полимерного материала: свидетельство об официальной регистрации программы для ЭВМ 2008612454 Рос. Федерация. № 2008611543; заявл. 11.04.08; зарег. 20.05.08
3. Чистякова Т.Б., Садыков И.А., Колерт К., Иванов А.Б. Методы кодирования и идентификации упаковок фармацевтической продукции для защиты от фальсификации // Информационные технологии. 2011. № 7. С. 52-57.
4. Kohlert C., Schmidt B., Schnabel A., Michels F., Schaaf M., Razigraev A., Chistyakova T. GefärbtepolymereFormkörper, Verfahren und VorrichtungzurHerstellung der Formkörper: pat. DE10201113543; Klöck-nerPentaplastGmbH; заявл. 15.09.11 ;опубл. 21.01.13.
5. König A., Gratz A.Advanced Methods for the Analysis of Semiconductor Manufacturing Process Data. Heidelberg: Springer, 2005. 262 S.
6. Tax D. M. J. Data Description Toolbox.: Pattern Recognition Laboratory // TUDelft. 2013.URL: http:// prlab.tudelft.nl/david-tax/dd-tools.html (дата обращения 01.01.2015)
7. Firm Description: The MathWorksInc // The-MathWorks Inc. 2013. URL: http://de.mathworks.com/ (дата обращения 01.01.2015)
8. Oracle Database Concepts.: Oracle Corporation // Oracle Corporation. 2015. URL: http://docs.oracle.com/ cd/E11882 01/server.112/e25494.pdf (дата обращения 01.01.2015)
