| dc.description.abstract | This thesis, conducted in partnership with AutoStore, examines the potential of predictive maintenance (PM) in the lift systems of their automated storage and retrieval robots. In the context of Industry 4.0, PM becomes crucial in reducing operational costs and system downtime, essential to outperforming competitors. The research builds on previous studies, focusing on health indicators (HI) and remaining useful life (RUL) estimations from sensor-based timeseries data. Primarily, the investigation leverages synthetic data, produced through Simulink in a digital twin of the lift system, with further testing against real data. The results suggest that Long Short-Term Memory (LSTM) autoencoders and fast Fourier transformation analysis can proficiently determine a lift system's state of health from the synthetic data, with LSTM embeddings distance approach showing promising results when applied to real data. The research also delves into methodologies for RUL estimation, involving regressors, timeseries forecasting, and degradation- and similarity models, with gradient boosting regressor emerging as the most effective for RUL estimation. The implementation of a synthetic HI, which negates the need for expert knowledge, implies its transferability and versatility to similar systems. Given the limited availability of real data, it is recommended that AutoStore places a higher emphasis on collecting real data and establishing a comprehensive service record. This will facilitate further research and integration of a PM solution within AutoStore's current system. | |
