Young-Sang Kim

Robust design of multilayer feedforward neural networks: an experimental approach

Abstract Artificial neural networks (ANNs) have been successfully used for solving a wide variety of problems. However, determining a suitable set of structural and learning parameter values for an ANN still remains a difficult task. This article is concerned with the robust design of multilayer feedforward neural networks trained by backpropagation algorithm (Backpropagation net, BPN) and develops a systematic, experimental strategy which emphasizes simultaneous optimization of BPN parameters under various noise conditions. Unlike previous works, the present robust design problem is formulated as a Taguchis dynamic parameter design problem, together with a fine-tuning of the BPN output when necessary. A series of computational experiments are also conducted using the data sets from various sources. From the computational results, statistically significant effects of the BPN parameters on the robustness measure (i.e., signal-to-noise ratio) are identified, based upon which an economical experimental strategy is derived. It is also shown that fine-tuning the BPN output is effective in improving the signal-to-noise ratio. Finally, the step-by-step procedures for implementing the proposed approach are illustrated with an example.

Robust design of artificial neural network for roll force prediction in hot strip mill

In the steel industry, a vast amount of data are gathered and stored in databases. These data usually exhibit high correlations, nonlinear relationships and low signal to noise ratios. Artificial neural networks (ANN) are known to be very useful for such data. However, selecting a suitable set of ANN parameter values is difficult even for an experienced user. This article proposes an experimental approach for determining ANN parameters in a robust manner for predicting the roll force in a hot strip mill process. Four design variables and two noise variables are included in the experiment, a full factorial design is adopted for the design matrix to estimate all main and two factor interaction effects, and the signal-to-noise (SN) ratio is used as a performance measure for achieving robustness. In the second experiment, only a fraction of the full factorial design is used as the design matrix and the results are compared with those from the full factorial experiment in terms of prediction accuracy. Experimental results show that the learning rate is the most significant parameter in terms of the SN ratio. The proposed method has a general applicability and can be used to alleviate the burden of selecting appropriate ANN parameter values.

Analysis of sedimentation/consolidation by finite element method

Abstract A low concentration of solid suspension material is transformed into a soil layer through two stages of sedimentation and consolidation, occurring simultaneously. During sedimentation in the uppermost part of the problem domain, in the rest of the domain, the soil particles form a continuous structure, the ‘Terzaghi soil’ so called. This soil structure consolidates under its own weight and the weight of the upper part. Because of the formation of this structure, a discrete boundary separating sedimentation from suspension is formed. On this boundary, termed a ‘shock boundary’, kinematic shocks occur. In order to analyze the phenomena of sedimentation/consolidation by finite element method, in this study, the sedimentation was considered as a solid flux problem from the standpoint of the Eulerian coordinate system; while the consolidation was considered as a fluid flux problem from the standpoint of Lagrangian coordinate system, with keeping the shock boundary between them. From this basic understanding, a program for one-dimensional analysis of sedimentation/consolidation was developed using the finite element method. In the program the governing equations were separately appliedxa0— continuity equation only for sedimentation; and continuity and force equilibrium equations for consolidation. Nevertheless, the interaction between them was continuously accounted for by keeping track of the shock boundary. From a comparison of the results obtained from the program with those from two previous studies, the performance of the developed FEM program was validated. To apply this program to real dredging fields, two data sets obtained from sedimentation tests for Korean marine clay were analyzed.

A hybrid model of partial least squares and artificial neural network for analyzing process monitoring data

Due to the advancement of data acquisition technology, a vast amount of process monitoring data can be easily gathered at most manufacturing sites. However, analyzing such data is difficult in that they usually consist of many variables correlated with each other. The partial least squares (PLS) method or artificial neural network (ANN) is known to be useful for analyzing such process monitoring data. In the article, a hybrid model of PLS and ANN is developed for increasing prediction performance, reducing the training time, and simplifying the ANN structure for analyzing process monitoring data. Computational results indicate that the proposed hybrid approach is a promising alternative to the usual PLS or ANN for analyzing process monitoring data. The proposed approach also results in a simpler optimum structure and can be generally trained faster than the ordinary ANN.

Development of a Framework for Analyzing Process Monitoring Data with Applications to Semiconductor Manufacturing Process

A semiconductor manufacturing process consists of hundreds of steps, and produces a large amount of data. These process monitoring data contain useful information on the behavior of a process or a product. After semiconductor fabrication is completed, dies on a wafer are classified into bins in the EDS (Electrical Die Sorting) process. Quality engineers in semiconductor industry are interested in relating these bin data to the historical monitoring data to identify those process variables that are critical to the quality of the final product. Data mining techniques can be effectively used for this purpose. In this article, a framework for analyzing semiconductor process monitoring and bin data is developed using the data mining and other statistical techniques.

Experimental Determination of Uncertainty Importances for Monotonic Systems

A system is called monotonic if its output (i.e., performance characteristic) is either increasing or decreasing with respect to any of the input variables (i.e., system parameters). For such systems, an experimental method is introduced to assess the effect of the uncertainty of an input variable on the uncertainty of the output, and thus, to determine relative criticalities of input uncertainties. An analysis method is also introduced to predict the percentage reduction in the uncertainly of the output when the uncertainty of an input or of a group of two inputs is reduced. The proposed method is computationally economical, especially for large-sized problems.