STAMM: Soft sensor moniToring and mAintenance framework for Machine learning Models

Artificial intelligence (AI) and machine learning (ML) models as well as Digital shadows and Digital twins as well as Digital shadows and Digital twins have been, and are being, developed for a variety of applications, including industrial process supervision, bioprocess optimization, energy management, and environmental monitoring. In cases where models have the aim to classify, predict, or forecast a variable of interest then these models must become deployable in real-time so that their predictions can be used in decision making and any corrective actions can be taken to steer the system back on track in a timely fashion. However, once deployed these models must also undergo routine maintenance to ensure that they continue to perform adequately for the task at hand. Here, we present the Soft sensor moniToring and mAintenance framework for Machine learning Models (STAMM), a framework designed to support supervision, deployment, and upkeep of soft sensor models in industrial environments. STAMM integrates six key components: (1) data acquisition from sensors and actuators; (2) a time-series database; (3) a workflow management system; (4) a model repository; (5) real-time monitoring dashboards; and, (6) drift detection modules to facilitate the reliable execution and oversight of ML-based soft sensors. STAMM has been built and validated within a biotechnology project, using an industrial-scale fed-batch fermentation as the primary guiding use case. STAMM however, remains fully applicable to a broad range of industrial process scenarios.