Increasing market competitiveness is driving manufacturers to continuously adapt and optimize their production processes. Under these conditions, even marginal improvements in production asset settings can yield substantial financial benefits. However, the continuously evolving market requirements shorten the lifecycle of AI-enhanced digital assets, leading developers to primarily rely on anomaly detection-based solutions. The currently available anomaly detection models are designed to detect drastic changes in the system, such as failure and leakage, ignoring the subtle differences between ideal and suboptimal production settings. This paper proposes a novel contextual anomaly detection framework that utilizes the decision scores of a classification model trained on the reconstruction errors of an autoencoder. Contextual information from the signal reading batches of the physical asset is integrated into the classification model to improve the interpretability of the reconstruction errors. Experimental results from a production line at ABB Schweiz AG Production Plant Schaffhausen demonstrate that the proposed method achieves zero false positives and false negatives, outperforming state-of-the-art solutions.