The malfunction of critical equipment in industries, such as gearbox in power generator enterprise, may present a significant environmental and financial risk. The motivation of this thesis comes from the requirements of a local company, the Companhia de Electricidade de Macau, S.A. (CEM), which was established in 1972 and has become the major supplies of electricity in Macau. The important equipment is Turbo Compound System (TCS) in the Coloane Power Station. Diagnosis of potential faults in TCS of Macau power station is the key of ensuring electrical power supply to consumers. Development a proper monitoring and fault diagnosis technique to prevent malfunction and fault of machine during operation is necessary. The facility fault in rotating gearbox machinery can be resulted by various failure types, including single component level errors, structural failure in system level, and compounded failures with multiple errors coexisted simultaneously. At the main time, the acquisition of various gearbox failure patterns from real power generator for training the diagnosis system is impractical. This thesis proposes a machine learning method based condition monitoring and fault diagnosis methodology to monitor the status of machinery in real time and avoid the uncontrolled failure. To analyze the TCS system in laboratory environment, a simulated power transmission system with concentration on gearbox subsystem has been developed, which could be adjusted to run on nine failure conditions. To preprocess the raw fault vibration signal which is in high dimensional scale and with noise, a two stage data preprocessing method is proposed. Firstly, the features of raw data are extracted through the wavelet packet transform and time-domain statistical features computation. To further optimize the feature set, feature selection via sensitivities ranking is carried out using two approaches: the compensation distance evaluation technique (CDET) for optimal features selection, and kernel principal components analysis (KPCA) to obtain principal component features. The point distributions of various failure types in each stage demonstrate the effect on dimension reduction and accuracy improvement. The preprocessed features are analyzed by intelligent classifiers. The thesis compares the performance of support vector machines with artificial neural networks. To examine the performance of intelligent classifiers, two case studies are carried out by combining different preprocess and classifier, where the KPCA output is used by SVMs and CDET is linked with ANNs. The experimental results show that the both combination models could reach high diagnosis accurate rate for simultaneous fault. It also found that the WPT+TDSF+CDET+ANNs framework is robust to be less sensitive to the environment noise when the training and testing are carried out in different date acquired in different time. It is believed that the proposed method has the potential to be applied in related industries with gearbox system.