Abstract: In view of the insufficient label information of bearing original vibration data, this made it difficult to model and analyze for it. Aiming at these problems, taking the intelligent maintenance system(IMS) data set as the research object, an unsupervised fault diagnosis of bearings method was proposed, which was based on the seasonal-trend decomposition procedure using loess and one-dimensional deep convolution auto-encoder network(STL-1DDCAE).Firstly, the nonlinear characteristics of bearing normal operation data were mined by one-dimensional deep convolution auto-encoder network, and the reconstruction error of healthy samples was obtained. Then, the reconstruction error signal of healthy samples was fitted by means of probability distribution, and its parameters of orthonormal distribution was calculated. Finally, seasonal-trend decomposition procedure using loess(STL)was used to analyze the reconstruction error curve of bearing, and the fault time of bearing was determined by using the component of trend term. The results show that this method can fully extract the bearing fault characteristics, and determine the sample critical threshold adaptively to avoid the high misjudgment rate of bearing abnormal state; and identify the time stamps of three bearing abnormal signals as 760,1 780 and 1 700 accurately. Then, labels of health status, inner ring fault, outer ring fault and rolling element fault can be added to the bearing data according to the anomaly detection time point, so as to realize data labeling processing.

Key words: mechanical operation and maintenance; fault diagnosis; abnormal signal detection; reconstruction error; seasonaltrend decomposition procedure using loess(STL); onedimensional deepconvolution auto-encoder(1DDCAE)