Neural Network-Based Classification of Electric Vehicle Acceleration Pedal Signals: From Training to Microcontroller Deployment

In an effort to promote sustainable mobility, electric vehicles have emerged as a crucial innovation in the global transportation sector. This research, undertaken in conjunction with Brain Technologies and its innovative Evergrin project, investigates the use of artificial intelligence in electric vehicles, with a specific focus on the classification of accelerator pedal signals using neural networks. The research explores the application of Tiny machine learning (TinyML) in the automotive industry using the Raspberry Pi RP2040 as the microcontroller of choice. Key research questions include the ability of neural networks to detect anomalies in pedal signals, performance differences between TinyML models and neural networks, and the trade-off between model latency and accuracy on microcontrollers. Using throttle position data from an internal combustion vehicle's OBD-II system to produce the basic dataset, the study applies several neural network models, including convolutional neural networks (CNNs), long-short term memory (LSTM) and gated recurrent unit (GRU), highlighting their potential in time series classification. The results of the study demonstrate the strong error classification capabilities of neural networks, with all models achieving at least an accuracy of 0.92. The maximum accuracy of 0.96 was reached by the LSTM model with two channels of input data. While the performance of the TinyML models, obtained through conversion from TensorFlow to TensorFlow lite, is comparable to that of their CNN-based equivalents, the more sophisticated models, such as the LSTM, have problems due to the lack of adequate quantization techniques for their layers.  In particular, the accuracy of the LSTM model dropped to around 0.5. On the other hand, for the two-channel CNN model, the accuracy is the same for both the TensorFlow and TensorFlow lite versions at 0.95. The complexity of implementing neural networks on microcontrollers was a significant obstacle, especially when considering the safety requirements of the automotive industry. Even the simplest models required almost two seconds for inference, significantly longer than the safety criterion of around 100 ms; furthermore, advanced models such as LSTM and GRU exceeded the memory capacity of microcontrollers such as the RP2040, making it impossible to implement said models on them. Given the promising potential for integrating these technologies into the automotive industry, it is believed that neural networks must be implemented on platforms with computational capacity beyond that of microcontrollers for optimal performance and feasibility. In conclusion, this research emphasises the central role that AI could play in the automotive industry, particularly in the classification of automotive signals. It aims to highlight the need to improve current technologies for optimising and converting machine learning models.