Fault diagnosis of rotating machinery based on a new hybrid clustering algorithm

A new hybrid clustering algorithm based on a three-layer feed forward neural network (FFNN), a distribution density function, and a cluster validity index, is presented in this paper. In this algorithm, both feature weighting and sample weighting are considered, and an optimal cluster number is automatically determined by the cluster validity index. Feature weights are learnt via FFNN based on the gradient descent technique, and sample weights are computed by using the distribution density function of a sample. Feature weighting and sample weighting highlight the importance of sensitive features and representative samples, and simultaneously weaken the interference of insensitive features and vague samples. The presented algorithm is described and applied to the incipient fault diagnosis of locomotive roller bearings. The diagnosis result demonstrates the superior effectiveness and practicability of the algorithm, and shows that it is a promising approach to the fault diagnosis of rotating machinery.