A Supervised Learning DNN-Based Ultrasonic Distance Measurement System Using Sinusoidal Activation Function

Ultrasonic distance measurement systems are relatively simple and inexpensive, making them widely used across various industries and applications. However, in environments with narrow walls or obstacles around the target, accurate distance measurement can be challenging due to the beam-width characteristics of ultrasound, which cause reflection signals from surrounding obstacles or walls. This paper proposes a new deep learning signal processing technique that analyzes the waveform of ultrasonic reflection signals to accurately measure the distance to the target in such environments. We use a two-stage parallel structure supervised learning autoencoder model and propose a new activation function combining a sine function with Leaky_ReLU with an alpha value of 0.008. This new activation function shows improved RMSE compared to existing activation functions such as ReLU and Leaky_ReLU, and experimental results demonstrate higher target prediction accuracy in tests. Ultimately, by measuring the distance based on the model's target prediction data, the system can accurately predict the target's location even in environments with narrow walls or obstacles.