Condition monitoring is usually performed over long periods of time when critical rotating machines such as wind turbine gearboxes are monitored. There are many potential signal processing and analysis techniques that can be utilised to diagnose the machine from the condition monitoring data, however, they seldom incorporate the available healthy historical data of a machine systematically in the fault diagnosis process. Hence, a methodology is proposed in this article which supplements the order-frequency spectral coherence with historical data from a healthy machine to perform automatic fault detection, automatic fault localisation and fault trending. This has the benefit that the order-frequency spectral coherence, a very powerful technique for rotating machine fault diagnosis under varying speed conditions, can be utilised without requiring an expert to interpret the results. In this methodology, an extended version of the improved envelope spectrum is utilised to extract features from the order-frequency spectral coherence, whereafter a probabilistic model is carefully used to calculate a diagnostic metric for automatic fault detection and localisation. The methodology is investigated on numerical gearbox data as well as experimental gearbox data, both acquired under time-varying operating conditions with two probabilistic models, namely a Gaussian model and a kernel density estimator, compared as well. The results indicate the potential of this methodology for performing gearbox fault diagnosis under varying operating conditions.