A Bio-Inspired WASD Neural Network with Soft-Margin Fuzzy Inference for Human Activity Recognition in Assistive Technologies and Pattern Classification

Source abstract

Human Activity Recognition (HAR) in Assistive Technologies represents a crucial application framework, focusing on motion tracking adjustments to safeguard the daily autonomy of individuals with physical disabilities. However, traditional gradient-based learning setups frequently face significant limitations, such as poor convergence speed, a strong tendency to become trapped in sub-optimal local minima, and boundary data uncertainties. This study introduces a novel bio-inspired Weight and Structure Determination (WASD) network framework combined with a soft-margin three-membership fuzzy inference controller to bypass these intrinsic optimization flaws. Particularly, by incorporating the stochastic exploratory mechanisms of the Beetle Antennae Search (BAS) metaheuristic algorithm, we are able to resolve the optimal hidden layer neuron number, the power exponents of the activation function, and the internal bias profiles simultaneously. The connections and algebraic output vectors are directly resolved in a single step using the non-iterative direct weight-determination process, which is a sub-algorithm of WASD and lowers the total computational complexity. HAR motion tracking is evaluated as the primary domain to examine the potential applicability of the proposed model to mobility-support systems, while quantitative finance asset classification is framed as an essential secondary task to assess the network’s predictive generalizability under non-stationary conditions. Numerical experiments and statistical comparisons indicate that the hybrid network model achieves high precision and delivers competitive performance or outperforms multiple benchmark machine learning classifiers, presenting in this way a robust alternative for real-world binary decision tasks.