Brought to you by EarthWeb
IT Library Logo

Click Here!
Click Here!

Search the site:
 
EXPERT SEARCH -----
Programming Languages
Databases
Security
Web Services
Network Services
Middleware
Components
Operating Systems
User Interfaces
Groupware & Collaboration
Content Management
Productivity Applications
Hardware
Fun & Games

EarthWeb Direct EarthWeb Direct Fatbrain Auctions Support Source Answers

EarthWeb sites
Crossnodes
Datamation
Developer.com
DICE
EarthWeb.com
EarthWeb Direct
ERP Hub
Gamelan
GoCertify.com
HTMLGoodies
Intranet Journal
IT Knowledge
IT Library
JavaGoodies
JARS
JavaScripts.com
open source IT
RoadCoders
Y2K Info

Previous Table of Contents Next


3.3. MACHINE LEARNING USING NEURAL NETWORKS

In order to code operational knowledge and use this knowledge for determining optimum strategies for machining operations, a feedforward neural network was studied. Based on the trained network, an optimization technique predicts the input conditions to be used by the machine tool to maximize a performance index under the appropriate operating constraints. The constraints are considered on the input variables due to machine limitations and upper bounds of some output variables in order to protect against factors such as excessive tool wear and tool breakage (Rangwala and Dornfeld, 1989). It was reported that for a turning operation, the neural network with 3 input, 4 hidden, and 4 output nodes can learn the mappings between input and output variables by observing the training samples. The input variables are the feed rate, depth of cut, and cutting velocity. The output variables predict the sensor outputs, such as cutting force, temperature, power, and surface finish.

The performance index is defined taking into account both goodness and efficiency. The former is reflected by some output variables kept sufficiently close to a desired value to ensure a good quality of cutting. The latter is represented using the material removal rate, which is simply the product of the three input variables. An optimization method using local computations based on the back propagation or an augmented Lagrangian method was proposed, although it gives a locally optimum set of input variables. The learning scheme can allow the partially trained network to predict an optimum set of input variables with the current state of knowledge, which is a crucial property required for real-time adaptive control. Usually, the training samples for neural networks, which are preset by the user, span the allowable range of the input variables. In the scheme that learns and synthesizes simultaneously, fewer samples need to be learned for only local knowledge in the vicinity of an optimum. In such cases, it is needed to guarantee the learning and synthesizing abilities using noisy sensor data in real manufacturing environments. In addition, it is required to assure the obtained result to be the global optimum when the synthesis algorithm adopts a gradient-based optimization.

3.4. GENETIC ALGORITHMS IN MANUFACTURING AND ENGINEERING

Traditional operations research approaches cannot resolve the long computation time needed for optimization with combinatorial complexity, so new techniques that assure the attainment of the global optimum with an allowable convergence speed have been investigated. Recently, genetic algorithms have been applied to optimization problems in a variety of manufacturing and engineering fields, such as the determination of optimum cutting conditions, tool selection scheduling, assembly process planning, and routing in traffic control of automated guided vehicles. At present, efficient codification methods to reduce the large search space of the genetic algorithm and implementation methods on parallel processing architectures for computational speed up are required.

3.5. PROCESS DIAGNOSIS USING PROBABILISTIC INFERENCE

The knowledge representation and reasoning with uncertainty is a critical issue for expert system development in manufacturing and process control. Production rules based on predicate calculus and classical binary logic have trouble taking into account competing or nondeterministic rules and imperfections in the measuring instruments. Several techniques have been proposed, such as Bayesian probabilities, fuzzy probabilities, Dempster-Shafer theory, etc. Influence diagrams provide an intuitive graphical framework for representing and combining evidence and can integrate dynamic sensor readings, statistical data, and subjective expertise in symbolic and numerical data structures. Probabilistic dependence between process variables in machining was represented using topological diagrams, and a polynomial-time symbolic-level algorithm for probabilistic inference and decision-making was applied to diagnostic reasoning, monitoring, and supervising control (Agogino, Srinivas, and Schneider, 1988). The system can determine the most likely process state from the observable sensor readings. Dynamic programming is used to optimize a parameterized utility function taking into account the cost trade-offs and assessed uncertainty of the possible failure states.

A machine troubleshooting expert system is being developed that determines the most efficient reasoning sequence in the diagnostic tree by a fuzzy multiattribute decision-making method. An inference engine controls the diagnosis process, considering the uncertainty with regard to the validity of a knowledge base rule and a user's response. A learning module trains the knowledge base by the failure-driven learning method from the past cases (Liu and Chen, 1995).

3.6. FAILURE DIAGNOSIS USING CASE-BASED REASONING

Case-based decision support systems can deal with ill-structured problems, for which mathematical models or formalized rules are nonexistent at present. An example in engineering fields is failure diagnosis of fractured parts in a factory, where due to the complexity of the metallographic features of the failed parts, there are no explicit rules or models to determine their failure mechanisms. A neural network with the back propagation algorithm is employed to learn the mapping between the cause sets and the result sets of verified cases based on metallographic features associated with the failure analysis problem to be solved (Gan and Yang, 1994).


Previous Table of Contents Next

footer nav
Use of this site is subject certain Terms & Conditions.
Copyright (c) 1996-1999 EarthWeb, Inc.. All rights reserved. Reproduction in whole or in part in any form or medium without express written permission of EarthWeb is prohibited. Please read our privacy policy for details.