Industrial Analytics and Process Monitoring

We turn industrial measurements and operating histories into models of product quality, batch behaviour and equipment condition. Our work includes soft sensors, interpretable quality models, reconstruction of batch operations and predictive maintenance using limited labelled data.

The challenge

Manufacturing processes generate sensor histories, operating records and laboratory measurements. These sources capture different parts of a process, often at different frequencies, and important outcomes may only be measured at the end of a batch. Equipment degradation data present a related challenge: operating histories are abundant, but reliable failure labels are scarce.

We develop analytical methods that make these records useful for understanding operations and predicting outcomes. Our work addresses product quality, the reconstruction of batch activity and the use of unlabelled sensor histories for predictive maintenance.

Our approach

Identifying informative process measurements

A soft sensor estimates a property that is difficult or costly to measure directly from other available process measurements. Our batch-analysis work uses partial least squares to identify informative variables and time regions, followed by multiway modelling to predict final product quality. This helps reduce the complexity of industrial datasets and identify which parts of the operating history matter.

Developing interpretable soft sensors

We also develop interpretable soft sensors using symbolic regression alongside feature engineering and dimensionality reduction. The resulting equations connect quality indicators with process variables, allowing their relationships to be examined. Published studies evaluate this approach using industrial formulation datasets.

Reconstructing batch operating structures

Beyond quality prediction, our work reconstructs and groups process flow diagrams from historical batch records. This provides a way to compare how batches were operated and explore differences in their processing sequences.

Using unlabelled data for predictive maintenance

Our predictive-maintenance research addresses the scarcity of Remaining Useful Life labels. A semi-supervised framework combines temporal modelling with optimisation to infer labels for unlabelled trajectories, making more of the existing operational data available for model training.