Contributions to time series analysis, modelling and forecasting to increase reliability in industrial environments
The integration of the Internet of Things in the industrial sector is considered a prerequisite for achieving intelligence in a company. To obtain this, AI systems with analytical and learning capabilities are required for the optimisation of industrial processes.
This research work focuses on improving current approaches or proposing new ones to increase the reliability of AI solutions based on time series data and introduce them effectively in the industrial sector. This intervention is carried out in three different phases of the time series data life cycle by increasing data quality, model quality and error quality. A standardised definition of quality metrics for evaluation is proposed and these functionalities are collected in the R package dqts together with quality enhancement functions.
Furthermore, an exploration is made of the steps to follow in time series modelling from feature extraction and transformations, a correct cross-validation, the choice of the most appropriate prediction strategy and the application of the most efficient prediction model. The KNPTS method based on the search for patterns in the historical time series is presented as a convenient R package for estimating future data. Finally, the use of elastic time series similarity measures is proposed as an alternative to quantify the performance of a regression model and the importance of using appropriate classification metrics in unbalanced class problems is emphasised.
All these contributions have been validated in four industrial use cases with distinctive characteristics from four research fields: product quality in the agri-food sector, electricity consumption forecasting in households, porosity detection in additive manufacturing and machine diagnostics.