Flow lines are frequently used to organize the mass production of physical goods in manufacturing. Such a line consists of serially arranged stations that are designed and equipped to perform dedicated tasks. The product units flow through the line to receive a series of operations at those stations. In a deterministic setting, the slowest station is the system's bottleneck and determines its throughput or production rate (measured in product units per time unit). However, in reality processing times are often stochastic, e.g., because of machine failures. In this case, to avoid blocking and starving, costly buffers can be installed between the stations to limit the propagation of failures up- and downstream of the system. In practice, discrete-event simulation is often used to estimate the production rate of a given (planned) flow line configuration. As an alternative, extremely fast approximate analytical methods have been developed to estimate the production rate of stochastic flow lines without using discrete-event simulation. We use such an analytical method to create and evaluate a large number of hypothetical flow lines and then train an artificial neural network to predict the production rate of flow lines which have not yet been analyzed before. We present first results from a systematic study of this new approach for flow line performance evaluation.
Bienvenue à tous!