In this paper, the adaptation of spectral kurtosis technology is proposed, demonstrated and experimentally validated. Raw data signals were collected from a single-stage gearbox run in different combinations of speed and load, after which time synchronous averaging was used to leave the classical residual signal once meshing harmonics were removed. Each data file is split into many individual realisations based on the time taken for the time synchronous average to converge on stable values, after which the short-time Fourier transform is used to calculate the spectral kurtosis for each realisation. The effects of adapting spectral kurtosis technology parameters such as the resolution and threshold used in creating a Wiener filter are evaluated, showing the effects on the consistent frequency bands identified throughout the realisations. Taking a baseline set of processing parameters, the probability of correct diagnosis was calculated using a three-stage decision-making technique incorporating the k-nearest neighbour and cluster analysis methods. Adaptation of the spectral kurtosis technology is then shown to dramatically improve the probability of correct diagnosis, highlighting that each speed and load case requires different resolution and threshold values to return the optimal results.

In this paper, two novel consistency vectors are proposed, which when combined with appropriate machine learning algorithms, can be used to adapt the Spectral Kurtosis technology for optimum gearbox damage diagnosis in varying operating conditions. Much of the existing research in the field is limited to test apparatus run in constant and carefully controlled operating conditions, and the authors have previously publicised that the Spectral Kurtosis technology requires adaptation to achieve the highest possible probabilities of correct diagnosis when a gearbox is run in non-stationary conditions of speed and load. However, the authors’ previous adaptation has been computationally heavy using a brute-force approach unsuited to online use, and therefore, created the requirement to develop these two newly proposed vectors and allow computationally lighter techniques more suited to online condition monitoring. The new vectors are demonstrated and experimentally validated on vibration data collected from a gearbox run in multiple combinations of operating conditions; for the first time, the two consistency vectors are used to predict diagnosis effectiveness, with the comparison and proof of relative gains between the traditional and novel techniques discussed. Consistency calculations are computationally light and thus, many combinations of Spectral Kurtosis technology parameters can be evaluated on a dataset in a very short time. This study shows that machine learning can predict the total probability of correct diagnosis from the consistency values and this can quickly provide pre-adaptation/prediction of optimum Spectral Kurtosis technology parameters for a dataset. The full adaptation and damage evaluation process, which is computationally heavier, can then be undertaken on a much lower number of combinations of Spectral Kurtosis resolution and threshold.