Comparison of FDA-based and PCA-based features in fault diagnosis of automobile gearboxes

• Use of Fisher Discriminant Analysis (FDA) which is optimum for the fault diagnosis system.
• Our experimental setup is a routine case in industrial applications.
• Fault diagnosing based on only one observation for a complex gearbox.
• Introducing two powerful and applicable classifiers for fault diagnosis purposes.
• Accessing to high fault classification rates for a group of nearly similar gears faults.

Several advantages for machine condition monitoring and fault diagnosis, such as reducing maintenance costs, improving productivity and increasing machine availability, have been formerly reported. The gearbox is one of the most popular machines in the world. The importance and need of this machine is clear; so, fault diagnosis of them is a core research area in the condition monitoring field. This paper presents an intelligent method to diagnose a kind of automotive multi-speed gearbox, operating in constant speed, using the vibration signal. In this research, the studied gears are located on the main input shaft which is supported with a tachometer sensor. Continuous wavelet transform (CWT) is applied to vibration signals of individual revolution cycles of input shaft; next, the continuous wavelet coefficients (CWC) are evaluated for some different scales. To prevent the curse of dimensionality problem, the Fisher discriminant analysis (FDA) is applied to this set of features. The fault diagnosis results are compared to the formerly introduced feature extraction approach, the Principal Component Analysis (PCA). As the classifier, the Gaussian mixture model (GMM) and K nearest neighbor (KNN) are individually examined, and the final classification performances are compared. Various faults are introduced and studied in this research. Faults are applied to the healthy gears, in a controlled way, to make possible to investigate them exactly. The experimental results show that the adoption of FDA diagnosis method leads to higher accuracy and less cost monitoring system for a multi-speed gearbox in comparison to the PCA-based feature extraction method for the both implemented classifiers.