Method Study in Time-Series Reconstruction for Machine Systems : Application of Artificial Intelligence for the Reconstruction of High-Resolution Signal Data in Machine Systems

Sammanfattning

Time-series reconstruction aims to infer sequential data from limited or degraded inputs, with broad relevance across scientific, industrial, and data-driven domains. Legacy machine systems, such as wheel loaders, are equipped with hardware that stores data in coarse categories without accurate measurements and sequential order. Potential downstream use cases and future innovations in the field require a level of data accuracy currently not achieved. Machine learning (ML) offers a potential avenue for time-series reconstruction by estimating high-resolution sequences from such coarse inputs, without requiring replacement of the underlying foundations of machine systems, thereby reducing industry costs. Previous work lacks broad analysis in time-series reconstruction from non-sequential data in machine systems. This thesis investigates the feasibility of reconstructing high-resolution time-series data from coarse, non-sequential cell-based machine log data using a deterministic and a generative neural network model. Specifically, a recurrent neural network (RNN) and a conditional variational autoencoder (VAE) were implemented and compared to a baseline that samples values uniformly within predefined interval bins. The results demonstrated that both RNN and VAE models struggled to outperform the baseline in multiple quantitative metrics, including log-likelihood. Although the ML models produced smoother sequences, these outputs did not align with the expected distributions. Based on these observations, there is little to no justification for applying ML models of this nature to the given reconstruction problem, highlighting key limitations in applying neural network models for vector-to-sequential reconstruction tasks. Major methodological refinements or hybrid approaches are needed to achieve plausible and reliable results for this problem.