High-precision transportation status monitoring of large electromechanical equipment, including structural health monitoring and fault monitoring during transport, remains a significant challenge. Extracting key features from non-stationary signals is particularly difficult under complex operating conditions and environments with high noise levels. To overcome the limitations of traditional methods in feature extraction, high-dimensional feature classification, and stability, this study proposes a key information classification approach integrating hybrid feature extraction with a multi-level deep network (MLDN). The proposed method first decomposes the original signal using an improved feature mode decomposition and reconstructs key components of the original signal based on the multidimensional criterion SKEC, incorporating kurtosis, energy entropy, and correlation coefficient. Subsequently, hybrid features are constructed by integrating the upper envelope, moving kurtosis, and moving root mean square. Deep spatiotemporal features are then extracted using a convolutional neural network-bidirectional long short-term memory architecture within an MLDN. Finally, the relevance vector machine embedded in the MLDN is employed to achieve high-precision and sparsity-enhanced state feature classification. Experimental results revealed that the proposed method exhibited superior recognition performance and strong noise robustness under complex, non-stationary environments, achieving a key information classification average accuracy of 99.08%.
Xu et al. (Wed,) studied this question.