Signal fusion-based deep fast random forest method for machine health assessment

The health assessment of machine is of great significance in modern industry. Vibration and acoustic signals collected from machine have been extensively and effectively applied in these fields. In view of the existing deficiencies, this paper reports a deep fast random forest (DFRF) fusion technique for machine running condition analysis by replacing the shallow learning theories with the deep learning ones and applying the vibration and acoustic measurements for analysis simultaneously. The fault-sensitive statistical parameters of the vibration and acoustic signals are firstly extracted with the help of wavelet packet transform (WPT), and then two deep belief networks (DBNs) were constructed to develop the deep representations of the WPT features. Finally, the fast random forest (FRF) is suggested as an information fusion-based classification tool for the fusion of the two kinds of deep features and identifying the type of abnormal conditions. To validate the present DFRF method, the gearbox running status analysis experiments containing 9 different conditions were executed, a series of experimental analysis and the comparisons with the peer approaches, such as K-nearest neighbor (KNN) and support vector machine (SVM), proved that the proposed method is more effective and reliable for machine health assessment.