W. Giegerich, Dennis Federoshin, Philip Gentz, Livio Forte III, Ph.J. Schneider, K.W. Oh
University at Buffalo,
United States
Keywords: magnetic field characterization, machine learning, classification, vehicle classification, electric vehicle, magnetometer
Summary:
This research focuses on the development of two machine-learning models that separately classify electric vehicles’ model types and drivetrain types based on their magnetic field signature. The proprietary Virtual Lift™ imaging system is equipped with a mobile phone that contains a magnetometer [1]. This MEMS magnetometer is used to record vehicles' magnetic field distortion / signature as they drive over the Virtual Lift. Through analysis of over a thousand electric vehicle magnetic field signatures, signal trends were identified across vehicles based on their model and drivetrain types. Two separate machine-learning models were developed to characterize and classify electric vehicles within these categories. Both achieved overall accuracies greater than 85%. Prior Work: The classification and quantification of different vehicle types such as trucks, vans, sedans, tractor trailers, etc. on public roads is important for transportation efficiency. Methods to accomplish this have been developed over the past several decades and to do so, the measurement / assessment of magnetic field disturbances from vehicles has been used [2]. This work aims to prove the Virtual Lift imaging system’s ability to accomplish this and further identify more detailed characteristics of the vehicle being sensed, including their model and drivetrain. Methods: To achieve full functionality of the Virtual Lift, a vehicle must drive over the entire system for the full length of the vehicle. During this time the Virtual Lift records the magnetic field signature of the vehicle, doing so through utilizing a three-axis magnetometer inside the mobile phone of which it is equipped. Figure 1 shows examples of magnetic field signatures sensed and recorded with the Virtual Lift. The orientation of the three-axis signals with respect to the vehicle is shown in figure 2. A dataset of roughly 1,300 electric vehicle magnetic field signatures was curated and labeled with their corresponding vehicle models and drivetrains (Table 1). Two custom built 3 layer convolutional neural networks were then developed and trained on approximately 80% of the curated dataset. Five classes were established for the CNN classifying vehicles’ models and three for the CNN classifying vehicles’ drivetrains (Table 1). Experimentation and Results: To test the performance of the models, a validation dataset (20% of total dataset) was curated to validate the models’ accuracies (Table 1). The CNN developed to classify vehicle models achieved an overall micro-average accuracy of 87.8% and the drivetrain classification CNN achieved an overall micro-average accuracy of 85.2%. With the low amount of training data, the models’ accuracies prove the Virtual Lift’s ability to record unique and detailed magnetic field signatures and that the developed CNNs could be used to classify electric vehicles’ models and drivetrains. More vehicle magnetic field data gained by the use of the Virtual Lift will provide more training data for future model development, helping to improve accuracies dramatically.