Detection of faults in a gearbox is a first and foremost step before diagnostic and prognostic operations are performed. This paper proposes a novel gearbox fault detection and feature extraction technique. The proposed method adaptively filters the vibration signals emanating from a gearbox. A bandpass filter is designed and optimized through particle swarm optimization (PSO) to maximize kurtosis as an objective function. Gearbox health-related features are extracted from the filtered signals using second-order transient analysis. The method is validated on experimental data collected from a running gearbox in healthy and faulty conditions. The proposed method has successfully identified the faulty conditions inside the gearbox.