Intelligent process quality control and tool monitoring in manufacturing systems
The work presented is best characterized as an investigation of neural networks for effective process quality control and monitoring in automated manufacturing systems. The research addresses two basic questions. The first question is whether neural networks have the potential to "identify" cause-effect relationships associated with advanced manufacturing systems to achieve real-time quality control? The second question is whether it is possible to use neural networks to develop effective reliability based real-time tool condition monitoring models for manufacturing systems? Both multilayer feedforward perceptron networks and radial basis fiinction networks are used in novel configurations to achieve real-time process parameter design. The models developed are capable of monitoring process performance characteristics of interest by building empirical based relationships to relate the process response characteristics with controllable and uncontrollable parameters, simultaneously. Using these empirical models and the levels of the uncontrollable parameters obtained through sensors, the quality controller provides levels for the controllable parameters that will lead to the desired levels ofthe quality characteristics in real-time. In general, the quality controller models were able to provide levels for the controllable variables that resulted in the desired process quality characteristics. Test results are discussed for several simulated production processes. A validity index neural network based approach was developed to automate the toolwear monitoring problem. In contrast to the contemporary approaches that basically deal with a classification problem, classifying a given tool as either fresh or worn, the model derived from radial basis function networks predicts the conditional probability of tool survival in accordance with the traditional reliability theory, given a critical performance plane, using on-line sensory data. In general, the radial basis fimction networks performed extremely well in time-series prediction, when tested on actual data collected from a drilling process. The validity index neural network is extended to arrive at the desired conditional tool reliability.