The paper investigates the applicability of time series models for processing data obtained from a customized crack-propagation sensor. Because the sensor records a variable and noise-affected waveform, the study focuses on models capable of forecasting signals composed of both trend and stochastic components. Adaptive, analytical, and autoregressive approaches were examined, with particular attention to their suitability for short, non-stationary sequences typical of fatigue-related measurements. Based on the statistical characteristics of the sensor output during crack growth, the ARIMA model was selected for further analysis and algorithm development. The forecasting performance of ARIMA was evaluated for different parameter configurations by comparing the range and variability of the base and predicted data. Initial tests using first-order parameters produced unsatisfactory results, with high variance observed in both raw and modeled signals. Therefore, model parameters were optimized using the aicbic function, and the analyses were repeated. For the selected datasets, variance reduction by 3–4 orders of magnitude was achieved, demonstrating a substantial improvement in prediction stability. The presented results confirm that the proposed methodology is effective for processing complex sensor signals and highlight the broader significance of applying statistically grounded time series models in structural health monitoring. The study introduces an innovative framework for evaluating fatigue-related sensor data and establishes a reliable baseline for future predictive methods.
Kurnyta-Mazurek et al. (Thu,) studied this question.