A Drift Detection Prototype for Quality Control in Manufacturing

Drift detection is a critical factor in sustaining the performance of Computer Vision (CV) models in modern manufacturing, where it ensures the continuous achievement of high efficiency, quality, and innovation. CV models efficiency is more likely to reduce over time due to environmental changes, equipment wear and tear, or alterations in product specifications. Moreover, the collection of labeled data, essential for monitoring and adjusting these models, is often hindered by constraints such as time, cost, and the lack of domain experts for annotation. Addressing this gap, our research introduces an innovative unsupervised drift detection technique capable of identifying shifts in data distribution without the need for labeled data. This method proactively notifies operators of any decline in data quality that breaches a predetermined safety threshold, thereby preserving the operational integrity of CV applications. Our experimental results confirm the method’s efficiency in detecting environmental shifts, changes in product specifications, and an increase in the production of defective parts, highlighting its significant potential to enhance quality control in manufacturing processes through reliable drift detection.