Noise-resilient feature selection for accelerometer-based guyed tower monitoring

Abstract

Guyed towers are vulnerable to environmental hazards that can lead to the collapse of transmission lines, jeopardizing essential services. Relaxed cables represent a critical condition that can result in structural failure. A Structure Health Monitor (SHM) is an Internet of Things (IoT) application that relies on an accelerometer to measure cable-stayed vibrations. This data is then inputted into a machine learning algorithm to identify relaxed cables. We incorporated a feature engineering step to enhance machine learning inference and mitigate uncertainty from raw accelerometer channels. However, there are concerns regarding high dimensionality and complexity arising from more irrelevant and correlated features. Additionally, sensors from IoT applications can introduce various types and magnitudes of noise. eXplainable Artificial Intelligence (XAI) approaches enable feature importance ranking and select more relevant and influential features. We experimented with traditional feature selection and XAI approaches to feature importance rankings. Our results show the robustness of features selected by XAI approaches compared to traditional accuracy. The baseline, which includes all features, achieved an accuracy of 96%, while the performance of machine learning algorithms under noise varied between 87% and 98%, closer to the baseline.