An Analog Neural Network for Estimating Sea State or Wave Height from Inertial Sensor Data

This research introduces an alternative approach to design low-power analog artificial neural networks with a power consumption of only 532nW. The proposed network consists of sigmoid activation function circuit, tanh approximation circuit along with a current comparator. In order to validate the behavior as an analog classifier, a real-time vessel dataset is used, achieving an accuracy of 86.50% in predicting true waves from inertial sensor data. Moreover, a comparative assessment among other analog classifiers is conducted, all trained on the same dataset. These models underwent training via a software-based implementation. The architecture was implemented using the TSMC 90nm CMOS process and simulated using the Cadence IC Suite.