Top of page
ABSTRACT
An intelligent module has been developed for a continuous billet casting mould in which features are extracted from rapidly-changing data (~ 200 Hz). The system mimics how a human reads a graph or time-series chart and instantly recognizes aspects and trends about process operation and stability. A real-time expert system then uses this preprocessed data for intelligent supervisory or adaptive control. The tools for real-time filtering use elements of the new field known as Computational Intelligence. Mathematical primitives and heuristic processing allow rapid analysis to provide symbolic information. This paper describes the concept of CI as used to support real-time monitoring of continuous casting of steel billets.
1. INTRODUCTION
Intelligent real-time systems must make high-level decisions and diagnose unexpected events. They acquire data automatically, apply heuristic methods to interpret sensor readings and feed advice out to the process or up to the user via a friendly man-machine interface. Although, data are acquired directly, decision-making can be slow for effective action. Data must be filtered before passing to the knowledge base to ensure efficient processing.
In analyzing computer control of an industrial process, a hierarchy can be delineated: Level 0 - Process Instrumentation, Level 1 - Direct control, Level 2 - Supervisory control, Level 3 - Plant-Wide control and Level 4 - Enterprise (see Figure 1).
Figure 1. Control System Hierarchy
At the lowest level, instruments sense, monitor and manipulate process variables. Devices are connected to units such as single-loop controllers, PLCs or DCSs, which apply a combination of sequential or continuous-time logic. Collecting, presenting, and managing sensor data requires numerical methods. For a system to respond to external events fast, carrying out several activities simultaneously, the system must operate within a multitasking environment that can efficiently handle priority interrupts and multiple module actions.
A real-time system can be considered as an integrated computer system that responds fast enough to the interrupting external events to provide accurate and fast control or alarm action.
Usually, systems with responses of 10 ms to, perhaps, a few seconds are considered real-time depending on the process and/or circumstances.
The supervisory level is next. Connected to the primary control devices by network communication, the supervisory host maintains applications and databases. Relevant data are gathered from the low level systems at a relatively slow rate. The operating mode at this level can be called "pseudo" real-time defined as an integrated computer system that responds 'fairly' fast to external interrupts carrying out control actions at the speed of several seconds up to, perhaps, one minute.
"Pseudo" real-time provides fast, accurate control, but in a different domain from real-time. Clearly, such definitions are context-dependent.
Neither of the two upper levels need to operate in real- or "pseudo" real-time. However, there is much interest today, in applying Artificial Intelligence (AI) in process control. There are examples of non-real-time AI applications at level 3 and 4 for diagnosis and advice [1,2,3], but most interest in real-time AI is at level 2. The question addressed in this work is:
"Can AI techniques be introduced into level 1 to provide support for symbolic supervisory control?"
2. COMPUTATIONAL INTELLIGENCE
AI is the branch of computer science dealing with symbolic, non-algorithmic problem-solving. "Von-Neuman" - type computers are designed for fast, accurate number-crunching. So how can AI move into an intensive numerical computing environment?
The answer is found in a newly evolving paradigm -
Computational Intelligence (CI).
Coined by Bezdek in 1993, CI consists of "low level" knowledge in the style of the mind [4]. CI consists of "primitive" concepts, in the AI sense, supporting the beginning of symbolic knowledge. These "elements" are inputs to an AI structure that processes the symbols heuristically. "Primitives" are basic numerical operations: addition, subtraction, multiplication, division, and comparison that make up any complex structure to output "elements". Current hardware handles such structures efficiently and so, CI can be the support for AI methods. CI comprises Fuzzy Logic, Artificial Neural Networks and Genetic Algorithms.
This definition seems limiting, as CI should not rely only on pure arithmetic. Rules of thumb can aid the search for symbolic input for the high-intelligence level. This can increase performance speed, but usually creates error in the output. To bring AI into the lowest levels of the control hierarchy, CI modules must create "elements" very fast. By introducing AI approaches into a CI module, we cause certain error, but gain on speed. The key trade-off in real-time is always: accuracy versus processing speed.
Considering that error derives from using heuristics in CI, it is proposed that feedback from AI to CI can detect such error and provide interpretation. AI can test symbolic output from cooperating sensors and recognize, tune out or reduce the error (see Figure 2).
Figure 2. Error Detection in AI to Correct CI Module.
To assist AI in making rapid decisions intelligently, we believe that CI needs the following:
The approach resembles the hierarchy of human intelligence depicted in Figure 3.
Figure 3. Biological versus Machine Intelligence.
Biological Intelligence consists of manipulating symbols supported by low-level numerical processing to generate belief in a particular symbol. This organization is mirrored in the arrangement of AI with CI to form the basis for rapid problem-analysis. Like humans, conventional computers work better when tasks are divided between symbolic and numerical analysis.
Within Biological Intelligence, the ability of autistic savants to carry out rapid and accurate data calculations, musical recall, etc., are examples of how the human brain can perform unusually accurate real-time computation. Whether output is intelligent or not is determined by those who interact with such exceptional people. Perhaps they use fuzzy sets with very broad support characteristics (see Figure 4). While savants may be able to tell the day of the week for a particular date, they rarely understand the significance of the date in question.
Figure 4. Normal and "Savant" Fuzzy Sets.
3. APPLICATION OF CI IN REAL-TIME
CI can be applied for many real-time tasks:
3.1 A Real-Time Supervisory Application
We have developed an intelligent SCADA system for continuous casting of steel billets which relies on the use of a CI module for real-time data processing.
Continuous casting, involves pouring molten steel into a water-cooled copper mould. The semi-solid strand is pulled from the mould by rotating pinch-rolls. Rotational or casting speed, is linked to mould metal level through a standard PID loop. Level changes are reflected by changes in casting speed. Metal level varies appreciably when turbulent conditions exist from "ropey" stream conditions. The machine oscillates to strip the solidifying shell from the mould wall. Displacement is usually sinusoidal but when sticking/ binding occurs, these signals are distorted.
The following variables are monitored:
The sensors and devices comprise the following:
Our project objective has been to create the "Intelligent" Mould to monitor and control the process. ProcessVision, a SCADA development tool from Comdale Technologies, was used to build the application. The multi-tasking nature of a real-time ProcessVision system is depicted in Figure 5. Each module interacts with others in a true multi-tasking fashion with appropriate interrupts and priority scheduling as required.
Figure 5. Typical Configuration of ProcessVision.
High-level supervisory decision-making occurs in the inference engine, the knowledge base and the explainer module. ProcessVision (PV) operates under QNX, a UNIX-based truly-distributed real-time multitasking operating system running on a PC. To accomplish rapid data acquisition within the SCADA, we installed a DAS-20 plug-in board from Keithley-Metrabyte.
3.2 Building the CI module
Our research group at UBC has been working to interpret patterns from sensor responses from many field trials conducted over the past 20 years. Specific curve shapes from thermocouple time responses (temperature peaks, drops, etc.) are related to specific billet defects. These correlations make up the knowledge base which detects surface defects - bleeds/laps and depressions.
When a temperature drop and rise propagates down the mould, a defect is indicated. The drop to base temperature ratio together with upset duration defines the significance and extent of a defect. These inputs are used to predict defects. A data acquisition rate of 20 Hz can capture all important features of these traces. The "pseudo" real-time Expert System (ES) is not designed to process data intelligently at this rate, so, application of a CI module was a necessity.
A multi-threading 'C' program was written for the CI module. The objective was to acquire data without delay while processing data. The two main threads are distinguished in Figure 6.
Figure 6: Structure of the "Intelligent" Mould.
The first program, main task, runs in an infinitive loop to execute the following functions:
The processing task receives data from the main task and sequentially filters inputs from each channel. Up to 5 functions per channel are applied:
average - calculate average(s) over specified number of points and feed PV with key-word-triplet(s).
minmax - look for minimum and maximum values and pass to PV as two key-word-triplets.
storedata - place collected data in volts, into a file.
compare - combine data from 2 channels to calculate negative strip time. We use inputs from an LVDT and the metal level sensor, at 200 Hz.
valley - this is an example of shape recognition and feature extraction that finds a "valley" shape in the data. The algorithm uses prior knowledge to direct the search, and a "window" technique to locate the minimum, left and right maximums. The search is set up by the TI signal from the AI module. The function can recognize 5 valleys in the data table and pass 20 key-word-triplets to PV as shown in Table 1.
calibration - convert input volts to actual values.
The main task collects data while the processing task filters recorded inputs. When the processing task finishes its routines, it "kills" itself. The main program recreates the processing task when it completes another acquisition cycle. The multi-threading design of the driver provides continuous data acquisition and processing of signals at upto 400 Hz.
Connection to PV uses the Comdale Third-Party Interface Library, communicating with the point database using keyword triplets. Table 1 shows typical TC signal procsessing output.
Table 1. CI output for one sensor for one scan cycle
The system tracks 4 thermocouples (TC). The AI module uses Fuzzy Logic to interpret the temperature drop and span interval. Together with the number of valleys from each TC, a degree of belief (DoB) is established that a bleed/lap or depression has occurred.
4. RESULTS
The "Intelligent" Mould was implemented and tested at two Canadian mini mills in November 1994 and March 1995. The system was used for real-time monitoring in each case. For the first time, negative-strip-time can be followed on the screen in real-time. Correlations between metal-level fluctuations, casting speed and billet defects were clearly evident.
This has helped us to associate ropey streams and a turbulent meniscus with on-line TC responses. Depression defects were predicted and a high degree of mapping between predicted and actual defects was obtained. TC responses obtained during the March 1995 plant trial are shown in Figure 7. Two temperature drops were detected for the period shown. The Degrees of Belief that a defect occurred were 98 and 97 respectively.
Figure 7: Results from March 20th plant trial (Heat 351)
The system correctly ignored other "apparent" drops. The CI module can be setup to include or exclude temperature depressions based on the desired degree of surface defect detection. Correlation with surface measurements was virtually perfect. Examination of the billet cast during this period proved the predictions correct: depressions were obvious on the billet surface. A mathematical model to predict temperature drops and durations based on depression measurements shows high correlation with the on-line detection system. The future plan for the "Smart Mould" is to use these defect predictions to provide an on-line quality rating for each cast billet.
To create a solid foundation for a real-time Quality Control System, we examined a number of billets for which detailed temperature trends were recorded. Surface defects were measured according to their distribution, position, and surface depth. These served as input to a 2-dimensional mould heat-transfer model. The depth and distribution of billet defects were translated into upsets in mould heat flux, assuming that the deepest depression represented about a 70% upset in heat extraction. Output from the model was obtained for several thermocouples located around and down the mould. Model output was plotted concurrently with real sensor data obtained during casting, as depicted in Figure 8.
Figure 8 shows a clear correlation between these trends - one for the model temperature profile based on depression measurements and the second for the direct measurements taken from the thermocouple. The model output has the same general curve shape as the real thermocouple-time response. The differences in the peak and drop values are a result of assuming a 70% upset. For this particular case, 80% gave a better fit. The model assumes a constant casting speed, although speed varies appreciably during casting. Fluctuating casting speed also causes a slight shift in the thermocouple data. The analysis establishes a clear correlation with existing defects on the billets, sensor data, and our mould heat-flux model. We believe this gives us the basis for direct measurement of billet defects at the time of formation.
5. CONCLUSION
A new paradigm known as Computational Intelligence is evolving in which intelligent numerical manipulation forms the underpinning of successful real-time AI. Parallel with Biological Intelligence is evident from man's ability to process numbers rapidly when necessary or when specialized tasks are required.
The CI module acquires high-frequency data rapidly using AI techniques that allow monitoring and control of a continuous casting process. Field trials have confirmed the method can be applied to predict bleeds/laps and depressions. Our efforts are continuing to develop fast-responding modules to give sufficient accuracy for high-level AI processing.
Effective real-time analysis of billet quality will lead to a useful performance criterion to give on-line analysis of cast product quality. The ultimate goal of such a system will be to provide on-line billet quality measurement. Both surface and internal defects will be included in this system and an adaptable billet quality index will be derived.
Acknowledgment
We wish to thank other members of the ConCast research group at UBC who provided considerable experience and advice. Support from NSERC and our sponsor companies - Alta Steel, Manitoba Rolling Mills, Comdale Technologies, Accumold and Hatch Associates is gratefully acknowledged.
References
1. S. Kumar, J. Meech, I. Samarasekera, K. Brimacombe, "Knowledge Engineering an Expert System for Quality Problems in Continuous Billet Casting of Steel Billets", Iron & Steelmaker, 1993, 20(9),
29-36.
2. K. Otsuka et al,, "Expert System for Blast Furnace Operation", Sumitomo Search, 1992 , 50, 43-50.
3. R. Edwards & A. Mular, "An Expert System Supervisor of a Flotation Circuit", CIM Bulletin., 1992, 69-76.
4. J.C. Bezdek, "What is Computational Intelligence?", Computational Intelligence - Imitating Life, 1994, IEEE World Congress on CI (WCCI), 1-12.
Return to the publications index