In-Vehicle Acoustic Event Detection Model Based on Deep Neural Network

The trend towards intelligence is prominent in the modern automobile industry, leading to a continuous increase in vehicle computing power. The incorporation of artificial intelligence into the vehicle cabin is expected to significantly enhance user experience. Sound, as a medium, holds the potential to offer a plethora of valuable vehicular information. Prompt identification of anomalous sounds within the vehicle can preemptively identify potential safety risks and contribute to overall vehicular safety. In this study, we propose a neural network-based approach to monitor certain irregular events within the vehicle. The model training utilized recorded in-car data. The dataset content encompasses various abnormal sound, including knocking sounds, pet vocalizations, etc. Additionally, data augmentation was performed using the log-Mel spectrogram transform and SpecAugment method. The model classifies them through a neural network, and the mixup method was utilized. We used three models, all of which are desgined based on convolutional neural network architecture. In the result, the structure of Deep Space Separable Distillation Block reaches an accuracy of 99.66%.