Incremental Learning-Based Quantitative Crack Detection Using Prioritized Experience Replaying and Layered Importance Sampling

The generation and propagation of blade crack directly affect the stable operation of compressor. To quickly and accurately detect crack, an incremental learning (IL)-based quantitative crack detection method is proposed through historical and current vibration samples. The purpose of the proposed method is to reasonably replay prioritized experience and adaptively adjust sampling proportion. Specifically, the adaptive layered importance sampling (ALIS) method is proposed to select representative historical samples by hierarchically assessing its importance. Besides, the sampling proportion is adjusted adaptively based on detection performance and nonlinear mapping with inverse Tanh function. Taking vibration data with 0–5-mm crack as historical samples, cracks from 6 to 10 mm can be quantitatively detected with high efficiency and accuracy through the proposed approach. The effectiveness of the IL-based quantitative crack detection method is verified by the compressor experiment under five working conditions, which can obtain the average value of 0.2866. Moreover, the effectiveness is examined and verified through the analysis of experimental data.