Towards Real-time Noise Suppression for Acceleration Sensors on Resource-constrained Processors

The real-time, precise, and accurate data from the accelerometer sensor is essential for ensuring a reliable and robust structural health monitoring (SHM) system. However, the accelerometer data can be contaminated by various types of noise, interference, and artifacts, leading to potential data misinterpretation. To address the mentioned challenge, an efficient real-time noise suppression method for MEMS-based accelerometers is presented. The proposed approach is designed for resource-constrained enviroment and it is a combination of the Recurrent Neural Network (RNN) model with specialized preprocessing techniques. This combination not only enhances noise reduction capabilities compared to the traditional noise suppression approaches, but also optimizes computational efficiency compared to the existing learning-based methods, making real-time processing on micro-controllers feasible. The presented model is trained using data obtained from real-world tests and also from simulations where different levels of noise are added to the desired signals. Moreover, an experimental setup composed of higher-resolution accelerometer sensors and a shaking table is employed to ensure comprehensive noise scenarios. Notably, the model achieved SNR improvements of up to 16.97 dB in the presence of Gaussian noise. The performance and efficiency of the trained model prove its potential for enhancing the accuracy and precision of the MEMS-based acceleration sensors, especially for SHM applications.