An on-line tracking self-learning algorithm for fuzzy basis function (FBF) neural network classifier is proposed in this paper.  Based on the previous possibility distribution of the clusters,  which is kept within the sample mean and covariance matrix with forgetting factor,  a strategy for constructing the target output of the new training sample set is given.  With the new sample set the FBF network can be trained to track the variable clustering boundary.  Meanwhile,  a recursive algorithm for computing the sample mean and covariance matrix with forgetting factor is also proposed to overcome the difficult of storing the vast old training samples.  The proposed method is used for fault recognition of the rotating machinery,  and the results show that it is feasible and effective.