An Automated Approach for Detection of Localized Gear Tooth Faults Based on Empirical Mode Decomposition

Maintaining smooth operation of gearboxes represents a fundamental consideration when ensuring the continuous reliable functioning of many pieces of industrial rotating machinery. Very often the inability of detecting an incipient gear fault in advance may result in costly down time as production is lost and additional costs are incurred in order to remedy the failure, not to mention the potential hazard people operating in the vicinity of the shaft line are exposed to. Insurance companies are well aware of the potential risks related to catastrophic failure of machinery and often demand condition monitoring systems to be installed on site. This in turn motivates research aimed at developing robust techniques for monitoring the condition of rotating machinery, in which gearboxes play an important role. Empirical Mode Decomposition is a signal processing technique which performs adaptive signal decomposition and results in a generation of a number of empirical modes, known as the Intrinsic Mode Functions (IMF). The decomposition process can be thought of as a form of adaptive filtering resulting in a number of waveforms (IMFs) each having a different composition of frequencies determined on the basis of the strict criteria of the method. EMD is a relatively new concept in the field of condition monitoring of rotating machinery, yet is has already been used in a number of different applications. One of the limitations of the technique which prevents it from being applied for condition monitoring purposes in a fully automated way is the selection of the optimal IMF, which often requires input from an expert in the field. This paper proposes a method for automatic selection of the IMF which contains the information most relevant for condition monitoring of localized gear tooth faults. The method is based on a criterion set in the frequency domain of the Time Synchronously Averaged (TSA) signal. The results show that the proposed approach may be more robust in extracting impulsive content from the signal, when compared to the raw TSA signal.