Driver Behavior Classification in Electric Vehicles
| dc.contributor.author | COMUNI, FEDERICA | |
| dc.contributor.author | MÉSZÁROS, CHRISTOPHER | |
| dc.contributor.department | Chalmers tekniska högskola / Institutionen för data och informationsteknik | sv |
| dc.contributor.examiner | Haghir Chehreghani, Morteza | |
| dc.contributor.supervisor | Åkerblom, Niklas | |
| dc.date.accessioned | 2021-07-06T06:15:23Z | |
| dc.date.available | 2021-07-06T06:15:23Z | |
| dc.date.issued | 2021 | sv |
| dc.date.submitted | 2020 | |
| dc.description.abstract | Studies have shown that driving style affects the energy consumption of electric vehicles, with aggressive driving consuming up to 30% more energy than moderate driving. Therefore, modeling of aggressive driving can provide a more precise estimation of the energy consumption and the remaining range of a vehicle. This study proposes driver behavior classification on vehicle-based measurements through several deep learning models: convolutional neural networks, long short-term memory recurrent neural networks, and self-attention models. The networks have been trained on two naturalistic driving datasets: a labeled dataset generated from a test vehicle on-site at Volvo Cars and unlabeled data collected from co-development Volvo Cars vehicles. The latter dataset has been annotated following rules and driving parameters quantifying the aggressiveness of driving style. The implemented models achieve promising results on both datasets, with the one-dimensional convolutional neural network yielding the highest test accuracy throughout experiments. One of our contributions is to use self-attention and deep convolutional neural networks with joint recurrence plots, which are appropriate for longer sequences because they bypass sequential training. The study also explores several active learning techniques such as uncertainty sampling, query by committee, active deep dropout, gradual pseudo labeling, and active learning for time-series data. These techniques showed variable results, with uncertainty sampling performing consistently better than random sampling. This study confirms the effectiveness of machine learning models in classifying driver behavior. It also shows that active learning can considerably decrease the need for training data. | sv |
| dc.identifier.coursecode | MPALG | sv |
| dc.identifier.uri | https://hdl.handle.net/20.500.12380/303630 | |
| dc.language.iso | eng | sv |
| dc.setspec.uppsok | Technology | |
| dc.subject | Aggressive driver behavior | sv |
| dc.subject | Driver behavior classification | sv |
| dc.subject | Self-attention | sv |
| dc.subject | Recurrence plots | sv |
| dc.subject | active learning | sv |
| dc.subject | Active deep dropout | sv |
| dc.subject | Gradual pseudo labeling | sv |
| dc.title | Driver Behavior Classification in Electric Vehicles | sv |
| dc.type.degree | Examensarbete för masterexamen | sv |
| dc.type.uppsok | H |