Multi-feature-based parallel DarkNet53-GhostNet-SqueezeNet for supercapacitor remaining useful life classification prediction method

Abstract

Because of the complexity of the internal structure of supercapacitors, the aging information of supercapacitors is difficult to be captured fully. And the regression prediction methods for remaining useful life (RUL) exhibit errors. Instead, the classification divides several supercapacitor RULs into a life interval, which can avoid the loss caused by the regression prediction error. This study proposes a multi-feature-based parallel DarkNet53-GhostNet-SqueezeNet (PDGS) for a supercapacitor RUL classification prediction method. Classification methods are employed for the first time in predicting the supercapacitor RUL. To fully capture the features in the data and improve classification accuracy, this study selects three CNNs from multiple configured neural networks for feature extraction. The features of the three CNNs are then integrated and mapped by the fully connected layers to get more precise classification outcomes. PDGS accuracy is 13.66% higher than the best comparison result.