A sparse approach to fault severity classification for gearbox monitoring

Fault detection and severity classification are critical to gearbox structural health monitoring. A common approach to fault severity classification is to identify the patterns associated with features extracted from raw sensor data that vary with fault deterioration. Since however features only represent partial information contained in the raw data, they may indicate different interactions as faults deteriorate. Consequently, data fusion using inappropriate features may lead to decreased classification accuracy. To address this problem, a signal-level classifier based on raw sensor data through sparse representation is presented in this paper. A library containing covariance matrices calculated from the raw data corresponding to different faults or levels of fault severity is first constructed. Classification of new observations is then performed through a sparse representation of the library. A sparse mapping vector, as the result of sparse representation, reveals the classification result. Experimental study performed on a gearbox test rig with multiple vibration sensors installed is conducted to evaluate the effectiveness of the developed sparse classifier. Results confirmed good performance of the developed data fusion method for gearbox fault severity classification.