부산대학교 도서관

A method is suggested to ascertain the energy stored in the spring to identify whether a circuit breaker is operating stably. Initially, the current of the electromagnet and the displacement signal of the circuit breaker closing contact were obtained through a data acquisition device as the original signals; Subsequently, the original signals are decomposed using kernel regression residual decomposition (KRRD) to obtain multiple sub-signals. The training samples are then constructed using multidimensional signals combined by the sub-signals with the original signal. Then, key features from the signals are extracted by inputting the samples into the transformer auto-encoder (TAE) model for learning. The residual network (ResNet) classifier is employed to learn these features for predicting the labels associated with the spring mechanism’s energy storage condition. This method enhances the dataset by effectively decomposing the signal into high-frequency and low-frequency components, which surpasses traditional deep learning models. The high-frequency components in the decomposition layer significantly enhance features that are difficult to observe in the original signal. Meanwhile, the low-frequency components also retain the primary characteristics of the signal. It also achieves adaptive fusion of signals from multiple sensors and frequencies, significantly improving recognition accuracy. Specifically, compared to the baseline TAE-ResNet18 model, the KRRD TAE-ResNet18 model showed a 5.93% increase in accuracy.