Equipment failures in industry can lead to significant economic losses and production downtime. In this context, machine learning-based anomaly detection algorithms play a key role in early fault identification. Among these, autoencoders (AE) have emerged as a powerful tool: when trained on normal operating data, they reconstruct sensor signals and quantify deviations through reconstruction error. However, selecting an effective anomaly score remains challenging. Highly sensitive metrics may trigger false alarms due to minor fluctuations, even when such variations are well within acceptable operating ranges. Conversely, even well-calibrated anomaly scores may fail to identify truly critical events without contextual understanding. To address this, we propose a framework that combines deep learning with expert knowledge by embedding predefined threshold values into the anomaly detection process. Our method, applied to the monitoring of drive pumps, uses an Long Short Term Memory (LSTM)-based AE to reconstruct sensor signals. The reconstruction error for each sensor is then weighted according to its proximity to both its nominal operating point and a domain-specific threshold. These weighted errors are combined and passed through a sigmoid function to produce a normalised performance index, ranging from 0 (normal) to 1 (critical anomaly). By incorporating threshold information directly into the metric, the proposed method suppresses false positives during normal operation and mitigates the risk of overestimating anomalies in the absence of fault data. This hybrid approach leverages the strengths of data-driven models while anchoring them in technical expertise, enabling more robust and interpretable alerts for predictive maintenance.