Gearboxes are a critical component in many types of machinery and are also one of the most heavily stressed. Condition monitoring provides early indications of damage and can reduce unexpected failures, while allowing operators to tailor maintenance strategies. Vibrational analysis is a key principle for the diagnosis of gearboxes and other rotating machinery, with Spectral Kurtosis techniques proven to be effective. There are several limitations often present in research, including the gearboxes being operated in a constant, singular combination of operating parameters and damage often being unrealistically large – combining to create a set of conditions rarely seen outside of academia. Limiting diagnostic capabilities of condition monitoring systems to a single set of speed and load values can drastically reduce the effectiveness and practicality of these processes. Gearboxes operated in industrial settings typically show early signs of damage long before tooth loss is likely, e.g., micro-pitting, so detection of these low level naturally formed damage mechanisms is more relevant but also more complex, requiring increasingly sensitive diagnostic techniques. Changing operating conditions are known to create significant complexities for damage diagnosis, with speed and load variations causing modifications to the vibrational signal that manifest similarly to changes caused by the presence of damage. This study focuses on modification and adaptation of already well understood and trusted techniques to work more effectively in changing operating environments. This research optimises the Spectral Kurtosis technologies to varying conditions of speed and load, enhancing the probability of correct diagnosis of gearboxes exhibiting low levels of naturally formed pitting damage. A literature review and theoretical analysis are presented before the Spectral Kurtosis technology and adaptation process is demonstrated, proving that correct selection of processing parameters brings significant diagnostic gains. The computationally heavy and time-consuming nature of this optimisation is highlighted, leading to the proposal of two novel consistency vectors that are used as inputs for a machine learning regression process which performs the adaptation in a faster, computationally lighter manner. A full methodology is detailed, with the novel changes highlighted and their benefits discussed. Experimental validation of the novel techniques is presented, utilising data generated at the Newcastle University Gear Research Centre from an instrumented back-to-back gearbox test rig. Data was collected in four combinations of speed and load, first in an undamaged state, then again in each combination once natural pitting damage had begun to form. A detailed analysis was undertaken to ascertain the full impact that changing operating conditions have on the damaged and undamaged data, proving that in both time and frequency domains these changes are indistinguishable from those related to damage condition. Two different benchmark diagnostic results were produced using commonly selected Spectral Kurtosis technology parameters, before parameter adaptation was undertaken in stages, ending with full technology optimisation. The novel techniques showed significant diagnostic gains in all sets of operating conditions. Two novel consistency vectors were developed and combined with regression modelling to optimise the SK technology in a computationally light manner with significant time saving over traditional techniques.
|Date of Award||6 Feb 2023|
|Supervisor||Len Gelman (Main Supervisor) & Andrew Ball (Co-Supervisor)|