Knowledge-Distilled Temporal Convolutional Networks for Transportation Mode Detection Using Edge-enabled Consumer Smartphone Sensors

Nowadays, smartphones and their built-in sensors not only have become ubiquitous but also serve as invaluable tools for collecting a diverse array of data. When we navigate through the complexities of Transportation Mode Detection (TMD), two key challenges emerge as: a) the need to balance computational efficiency with model accuracy, and b) the importance to capture temporal dependencies in time-series data for accurate mode classification. To address these issues, we propose TRTCN, a novel algorithm designed to transportation mode detection via edge-enabled consumer smartphone sensors. This algorithm employs a Temporal Convolutional Network (TCN) and integrates data from lightweight sensors housed within the smartphone. In our proposed model, we employ knowledge distillation to construct a framework that comprises both a teacher model and a student model. The student model, represented as a Convolutional Neural Network (TRTCN-CNN), can help reduce the number of parameters required for the model, thereby decreasing the training time. A multi-Head attention mechanism is also incorporated to fine-tune the accuracy of the model’s accuracy. The student model has exceptional performance, with an average precision of 88.32% on the SHL 2018 dataset and an astonishing 96.04% on the SHL 2021 dataset. Our work shows that the transportation mode detection can be enhanced via the existing edge-enabled consumer smartphone sensors.