A Parzen window based semi-supervised fuzzy c-means (PSFCM) clustering algorithm was presented. The initial clustering centers of fuzzy c-means (FCM) were determined with training samples. The membership iteration of FCM was redefined after the membership degrees of testing samples relatively to each state were calculated using Parzen window. Two typical faults of gear box were simulated through the gear box bed in order to acquire the lubricant samples. Concentration of Fe, Si and B, which were the representative elements, was selected as the three-dimensional feature vectors to be analyzed with FCM and PSFCM clustering methods. The clustering results were that the correct ratio of FCM was 48.9%, while that of PSFCM was 97.4% because of integrating with supervised information. Experimental results also indicated that it can reduce the dependence of the experience and lots of faults data to introduce PSFCM into oil atomic spectrometric analysis. It was of great help in improving the wear faults diagnosis ratio.