Model Degradation Monitoring in GxP Manufacturing Analytics

Main Article Content

Rajesh Mannam

Abstract

However, for the highly regulated Good Practice manufacturing world, the success of predictive models is essential to the product quality and process compliance. In this study, the aim is to address the practical model degradation problem, in which the predictive capacity of the model becomes degraded over time, as a result of changes in the data distribution and changing environmental conditions. We use a large dataset of 315 instances of manufacturing performance, including sensor readings and batch throughput data and quality control checkpoints. The tools used for the research include automated pipeline drift detectors, statistical process control dashboards and are all part of a secure cloud-native ecosystem. We employ continuously measuring accuracy, precision and recall, and comparing these metrics to quality benchmarks. We have put in place real time monitoring, which has shown that the risks of non-conformance are greatly reduced if performance decay is detected in a pro-active manner. The results show how important it is to maintain tight connections between the use of operational data and re-training models. This research will be used as a foundation to maintain model efficacy in GxP environments where automated decision making is required to remain compliant with stringent regulatory requirements. Our results show that continual monitoring can effectively maintain the health of models, and also strengthen operational resilience, providing a sustainable way of scaling AI in manufacturing sectors.

Article Details

Section
Articles