Artificial neural network and dataset optimization for implementation of linear system models in resource-constrained embedded systems.

On-board training of artificial neural network (ANN) is important in instances where real time data are required for model training. Provision of on-board intelligence enables the developed systems to self-recalibrate and enhances their efficiencies. In this work, investigations have been performed to determine optimized parameters of ANN model for linear systems. The performance parameters that is, model parameters, memory requirements, accuracy and processing time are chosen by considering the model to be installed on commercially available microcontrollers that have very limited on-board memory. Minimum data requirements for training ANN models of linear systems are also explored for better performance. All dataset ranges are normalized in order to exclude the effects of range differences. It is shown that for linear systems, 1-3-1 architecture produces best results against <= 100 data points when Bayesian Regularization (BR) training function is used along with Log Sigmoid Activation function. Simulations for 1-3-1 architecture are then performed for datasets having 10, 25, 50 and 100 data points. The results show that training with 25 data points produces over-all better performance than other datasets. A large dataset utilizes more training time and memory whereas a smaller dataset produces relatively lesser accuracy. The effects of clustered data and uniformly distributed data are also explored. It is found that total epochs in case of clustered data are significantly higher than uniformly distributed data. The combination of these optimized parameters that is, 1-3-1 architecture, with BR and Log function, for <= 100 data points can be used for the development and implementation of linear components or systems in resource-constrained embedded systems.