Modelling time-varying delays in networked automation systems with heterogeneous networks using machine learning techniques

Time-varying delays affect the performance and reliability of networked automation systems (NAS). Recent trend to use wired and wireless networks within NAS induces network delays that vary depending on many factors such as loading, sharing, length of the channel, protocol, and so on. As these factors are inherently time-varying, developing analytical models capturing the effect of all these parameters is complex. This investigation presents a methodology that combines experiments with machine learning techniques to model time-varying delays in networked automation systems integrated with heterogeneous networks. Experiments are conducted on NAS by varying the factors that influence delays and time stamping obtained using Wireshark are used to compute the delay. The data collected on the factors influencing the delays and the corresponding delay values are used to model the delays. In data-mining techniques, the accuracy of the estimates varies with the number of computing elements in the hidden layer and selecting them using trial-and-error approach is cumbersome. The minimum resource allocation network (MRAN) over comes the short-coming as it decides the number of computing elements (neurons) in the hidden layer using error thresholds and pruning strategy. The data collected from the experiment is the input training set to the MRAN. Once trained, the MRAN model gives a functional representation relating the factors affecting delays and the estimated delay for a given network condition. During testing, MRAN estimates are validated using error measurements. Results show that the MRAN delay model can capture delays with good accuracy and can be used a tool to assist design decisions on engineering automation systems with heterogeneous networks. The proposed model gives a framework to model time-varying delays as a function of factors influencing them and can be modified to include any number of parameters. This is a significant benefit against existing models in literature that capture the delays only for particular conditions.