Automated weigh-in-motion through vehicular telematics and machine learning

Abstract

Road safety is one of the major concerns in today’s world. Driving an overloaded vehicle causes various ill-effects to road safety. Various kinds of Weigh-in-motion (WIM) systems are used to control and reduce the impacts of overloaded driving. Existing WIM systems are either expensive or slower and influenced by various factors. Advancement in network connectivity and sensor devices led to the development of the Internet of Vehicles (IoV), a subfield of the growing Internet of Things (IoT). IoV is powered by Vehicular Telematics (VT), also known as flying car data. VT data is used by the transport industries for many reasons such as fleet management, insurance (pay as you drive), driving behaviour detection, and road anomaly detection. Intelligent Transportation System (ITS) uses both IoV and Machine Learning (ML) techniques to build an automated Artificially Intelligent (AI) transportation system. According to the Newtonians’ physics and literature, under certain conditions, the driving force needed by a vehicle to obtain a particular acceleration is influenced by the total weight of the vehicle. That implies that if driving force and other influencing parameters are known, we could infer the weight of a vehicle. VT data can be used to obtain many features, including the driving force. This dissertation discusses the effort taken to validate the idea of inferring the weight of a vehicle using VT and ML. This research involved designing and testing the prototype artefact. The Design Science Research (DSR) methodology was used in this research. The C-K design theory was used in this DSR. The application of C-K theory in DSR has shown the different dimension for approaching applied research. A pragmatist approach was used in the design and development of this research. According to the C-K design theory, with all the knowledge, K0, from literature and the laws of physics, we formed an initial concept C0: “A new WIM solution using VT and ML”, with the propositions p1: “faster”, p2: “economical/cheaper”, p3: “Ubiquitous”. The concept was tested by designing and developing the prototype (artefact). A backend to process VT data using ML was developed as a by-product of this research. We have tested several ML algorithms during the development stage, and an Artificial neural network (ANN) architecture of three hidden layers with 30 nodes in each layer has shown astounding performance with Accuracy = 0.945, R-Squared = 0.97, Adjusted R-Squared = 0.97, Mean Squared Error = 34.68, Residual Standard Error = 6.03. The ANN outperformed all other tested ML algorithms on the collected VT dataset. We can infer the weight using the smaller dataset obtained from the context of a small car. Results from small cars show the supports for the concept theory.