Health Status Prediction for Nonstationary Systems Based on Feature Decoupling of Time Series

This article proposes a nonstationary time series forecasting model based on feature decoupling (NTFFD), which effectively enhances the efficiency of health status assessment and prediction for nonstationary systems in the field of instrumentation and measurement (I&M), while reducing resource consumption in practical industrial applications. Current online prediction methods, despite their advancements, often necessitate complex network architectures and update mechanisms, which hinders efficiency and practical application. The NTFFD model transcends the limitations of existing methods by constructing simple multilayer perceptron (MLP) predictors to independently extract gradual and abrupt features from time series data. It employs a feature decoupling approach to analyze the time series separately and then dynamically weighs and integrates them through reinforcement learning. Consequently, significant improvements are achieved in both training speed (a 3270% increase) and memory efficiency (a 91% reduction). The NTFFD is particularly suited for applications like real-time monitoring in high-speed maglev train’s traction drive control systems (TDCSs), demonstrating its potential to advance the I&M field.