Neural networks and quality control when the process dynamics are unknown

In a complex production process the underlying structure of the process is often unknown to the operations managers. Hence, identification of the source of variations and variation reduction are difficult and time consuming tasks. Under the assumption that the process design is capable of producing the product with the target value, this paper employs the adaptive control approach proposed by Spall and Cristion (1992) in reducing deviation of output value around its target value. The goal is to examine the performance of the above control approach for a production process whose underlying structure is unknown and may change during the production process. The study assumes that: operations managers first employ statistical monitoring steps to identify the operating factors that have significant effect on the output value. Then they employ the above adaptive control to readjust the level of those factors to reduce the deviation of output. The controller of the adaptive control approach is a feedforward neural network (NN). The approach uses simultaneous perturbation stochastic approximation to estimate the weights of the neural network. The NN controller does not need knowledge of the underlying dynamics of the system and does not require that one build a model for the dynamics. Further, it accounts for the process noise. The findings of the simulation study provide encouraging news for practitioners