The application of artificial intelligence to automate condition monitoring has gained a lot of interest. Machine learning has simplified the condition monitoring task. Vibration-based condition monitoring is most commonly used to monitor the health of mechanical rotating components. Each mechanical component generates a characteristic vibration signal with distinct frequencies and amplitude based on its mechanical properties and operating conditions. The health of mechanical components can be monitored by analysing their vibration signatures. Faults such as misalignment, unbalance, or damage introduce unexpected frequencies into the vibration signature. The envelope and autopower spectra highlight characteristic and fault-related frequencies in the vibration signal. However, continuous manual monitoring of both spectra for each component in a large fleet of complex machines where each machine has numerous components is not feasible. Nonetheless, identifying unexpected frequencies in the spectra can improve the efficiency of health monitoring for mechanical components. The proposed spectral autoencoder method is a normal behaviour machine learning method that learns the characteristic frequencies of components from the envelope and autopower spectra calculated from the vibration signal measured during the healthy operation of the mechanical component. The trained model on healthy operational data of the mechanical component can detect unexpected fault frequencies introduced in the vibration signal as a result of a fault. It generates alarms based on the number of anomalous frequencies detected in the signal, and it provides detailed information about each anomalous frequency to experts for further analysis. The proposed method is validated using NASA's Intelligent Maintenance Systems (IMS) bearing dataset and offshore wind farm gearbox vibration data collected over multiple years, with labelled fault cases. The offshore wind farm data includes labelled fault cases for planetary gear stage faults, generator, and high-speed stage failure. This approach streamlines the health monitoring of mechanical components by generating high-level alarms based on fault frequencies and offering detailed analysis of unexpected fault frequencies. Future work aims to incorporate operating conditions for more reliable condition monitoring of machines operating under variable conditions, such as wind turbines.