SOLVING THE UNSOLVED - USING MACHINE LEARNING, DATA MINING AND KNOWLEDGE DISCOVERY TO MODEL A COMPLEX PRODUCTION PROCESS

Accurate, efficient and consistent production of products and materials is the cornerstone of successful industries. These processes represent a company’s revenue and its future. A tremendous amount of energy has been applied to developing and optimizing production processes. These include expert systems, simulation models and the like. Sometimes, these processes are so complex that they can not be modeled using conventional object orientated or procedural programming techniques. This paper concludes that machine learning technologies can reverse-engineer production processes with enough accuracy to develop useful simulation models. With the simulation model developed, adjustment to the production process can be optimized through simulation instead of expensive trial and error. The case we chose to analyze, i. e. the prediction of the properties of stabilized industrial process waste with respect to hydraulic conductivity, strength, and potential to leach stabilized contaminants, does not have a known solution. This process is also very expensive, often costing in the millions of dollars to implement, and it is extremely costly for the manufacturer to rework material that is out of specification. The resulting simulation models were calibrated and verified using laboratory data. The machine learning techniques evaluated consisted of genetic programming, simulated annealing, neural network/genetic algorithm technology, and a rule based classifier. Three methods, genetic programming, simulated annealing, and a neural network solved the highly non-linear production process. The rulebased techniques experienced difficulty with this particular case example, though it has been very successful modeling other data sets. The simulation models developed by machine learning were acceptable for the intended purpose. These machine learning techniques can be expected to solve many complex industrial processes thereby opening up new avenues to simulate and optimize production processes that were, until recently, either not practical or unsolvable.