Evaluation of Machine Learning approaches for resource constrained IIoT devices

Resource-constrained devices such as sensors, industrial controllers, analyzers etc., mostly contain limited computational capacity and memory. They are largely deployed in all industries and have been generating a huge amount of data. This data is sent to the cloud servers where various Machine Learning (ML) algorithms are applied to perform the analysis or prediction as per the application. In this process, communication requires bandwidth and time. Since the data is sent into the network, the privacy of the data is not guaranteed. Cloud servers consume a huge amount of power. To reduce these cost factors, the machine learning models are compressed and optimized such that they can fit and run in small footprint devices. The Federated Learning (FL) approach at the edge device level promises to address the data privacy and bandwidth related issues. Since it is a decentralized learning method across a set of devices, the performance of the model also improves. This paper describes and evaluates the machine learning algorithms with various compression methods suitable for resource-constrained IIoT devices and federated learning approach, particularly for time series data applications. Simulation results show that FastGRNN algorithm gives the least model size compared to the traditional RNN algorithms for time series.