Design of an efficient spiking neural network for human activity recognition

• Main Author: Tan, Yee Leong
• Format: Thesis
• Language: English; English
• Published: 2024
• Online Access: http://eprints.utem.edu.my/id/eprint/27405/1/Design%20of%20an%20efficient%20spiking%20neural%20network%20for%20human%20activity%20recognition.pdf; http://eprints.utem.edu.my/id/eprint/27405/2/Design%20of%20an%20efficient%20spiking%20neural%20network%20for%20human%20activity%20recognition.pdf; http://eprints.utem.edu.my/id/eprint/27405/; https://plh.utem.edu.my/cgi-bin/koha/opac-detail.pl?biblionumber=123434

Human activity recognition (HAR) using Wi-Fi Channel State Information (CSI) has attracted significant interest as an alternative to conventional methods due to its potential to address human privacy concerns. While Long Short-Term Memory (LSTM) models have shown promising results in HAR, their resource-intensive nature and time-consuming computations limit their suitability for edge computing. The development of Spiking Neural Networks (SNNs) as a more power-efficient computational model presents a compelling alternative. However, a critical research gap exists as no prior study has explored the application of SNNs for time series data, particularly for Wi-Fi CSI analysis, within the context of the Industrial Revolution 4.0. This work addresses the research gap by proposing an SNNs model that involves preprocessing the CSI signals and encoding them into spike trains. The spike trains modulate the membrane potential at the postsynaptic neurons based on their respective weight values, enabled by the SpikeTiming-Dependent Plasticity (STDP) learning rule during the training process. The combination of these techniques enables accurate class prediction. Additionally, with different preprocessing methods and different values on the model’s parameters, SNNs models can achieve varying accuracy results. The application of the majority vote method to the outputs of divided signal segments ensures a robust final class prediction. Experimental results demonstrate that the proposed SNNs model achieves accuracy levels comparable to those of the LSTM model while significantly reducing computational memory usage by up to 70%. Remarkably, the SNNs model exhibits consistent performance even with smaller datasets and varying train-test ratios, showcasing its robustness in the face of limited training data. This memory-efficient and resilient nature positions SNNs as a viable solution for edge computing within the scope of the Industrial Revolution 4.0. In conclusion, this study introduces a pioneering application of SNNs for HAR using Wi-Fi CSI, highlighting the efficacy of spike trains and the STDP learning rule in enabling efficient computation and precise predictions. The demonstrated memory savings and robustness of the SNNs model underscore its potential to address the challenges associated with HAR while upholding privacy concerns and optimising resource utilisation in the era of the Industrial Revolution 4.0.