Effective condition monitoring is critical for preventing costly machine failures. Vibration analysis is one of the most widely adopted condition monitoring approaches. It enables early fault detection by capturing subtle dynamic changes caused by misalignment, imbalance, and bearing wear. Traditional techniques rely on signal processing in the time and frequency domains, and experts manually track individual condition indicators to identify fault trends. The manual tracking of condition indicators becomes impossible when monitoring large fleets of complex machines, such as wind turbines. This research proposes a novel approach to consolidate multiple condition indicators into a single high-level health indicator to simplify the monitoring process. A physics-informed multivariate autoencoder models the machine's normal behaviour using vibration-based condition indicators computed from the vibration signal measured during healthy operation. The non-linear model incorporates operating conditions from vibration condition indicators, without using SCADA as input. It identifies faults by detecting deviations from the established normal behaviour. The proposed method is validated on NASA’s Intelligent Maintenance Systems bearing dataset and a multi-year offshore wind farm dataset with confirmed fault cases. Validation on a wind turbine drivetrain dataset demonstrated that the proposed method detects 100 % (19/19) of labelled fault cases. The model achieves 97 % balanced accuracy, and threshold optimisation further reduces false positives to 1 case out of 91 healthy cases, while maintaining high diagnostic sensitivity with only a single false negative (missed fault alarm). Results demonstrate that the proposed method reliably accommodates diverse condition indicators, effectively detects faults, and reduces the time required for condition monitoring analysis.