In this talk, we present a data mining based methodology for optimizing the outcome of a batch manufacturing process. Predictive data mining techniques are applied to a multi-year set of manufacturing data with the purpose of reducing the variation of a crystal manufacturing process, which suffers from frequent fluctuations of the average outgoing yield.  Our study is focused on specific defects that are the most common causes for scraping a manufactured crystal. A set of probabilistic rules explaining the likelihood of each defect as a function of interaction between the controllable features are induced using the Single-Target and the Multi-Target Information Network algorithms. The rules clearly define the worst and the best conditions for the manufacturing process, also providing a complete explanation of all fluctuations in the outgoing quality observed over the recent years. In addition, we show that an early detection of nearly the same predictive model was possible almost two years before the end of the data collection period, which could save many of the flawed crystals.  The talk provides a detailed description of the optimization process, including the decisions taken at various stages and their outcomes.  Conclusions applicable to similar engineering tasks are also outlined.  This is a joint work with Guy Danon, Sholomo Biderman, and Eli Miron.

Artificial Neural Networks (ANN) can create data-driven models, with no need of equations describing the system behavior. Since the early eighties, ANN modeling was used in hundreds of thousands applications in every fields of science, industry and commerce, including fault detection and predictive maintenance in industrial systems.

The proprietary algorithms used by me allow the easy training of high-dimensionality models, with hundreds to thousands inputs and response variables. Although ANN models are perceived as "black boxes", it is possible to analyze the resulting models and elicit new knowledge about the system behavior, such as the identification of the more relevant input variable, and the direction and magnitude of their influence on the system outputs.  ANN combined with Genetic Algorithm can be used for optimizing a system behavior by finding the best values for inputs. The presentation will include examples relevant to the electronic processing industry, both for identifying the reasons for manufacturing faults from existing data, and for incipient fault detection in real time, based on modeling the normal system behavior.

Ilan Shimshoni is the head of the Management Information Systems department at the University of Haifa. He received his B.sc. in mathematics from the Hebrew University in Jerusalem, his M.Sc. in computer science from the Weizmann Institute of Science, and his Ph.D. in computer science from the University of Illinois at Urbana Champaign (UIUC). Ilan was a post-doctorate fellow at the faculty of computer science at the Technion, from 1995--1998, and was a memberof the faculty of industrial engineering and management from 1998--2005. He joined the department of Management Information Systems at Haifa University in October 2005. In recent years Ilan’s research branched into the field of data analysis/ data mining, dealing with problems such as clustering, non-parametric statistics and algorithms for approximate nearest neighbor search. The initial application of these methods was for computer vision problems but the techniques can be used in many other fields.

Ran Wolff is faculty of the Management Information Systems department at University of Haifa, Israel. A graduate of the Technion -- Israel, he previously held post doctoral positions at the Technion and the University of Maryland in Baltimore County. His main fields of expertise are data mining in large-scale distributed environments: peer-to-peer networks, grid systems, and wireless sensor networks, and scalable privacy preserving data mining. ANN are composed of simple computing elements, "neurons" that are connected by variable weights that modify data passing in the ANN. The model is developed by repeated presentation of known inputs and outputs of a system to the ANN. An error reduction algorithm modifies the connection weights until the desired small error is found between the known system outputs and the ANN outputs.