Abstract

• A new feedback connection mechanism is proposed to determine the contribution of features. • The contribution of features to the final result does not need to be determined by the expert’s prior knowledge. • The proposed method can distinguish a priori features from other irrelevant knowledge. • The proposed method is superior to SOTA methods in determining the contribution of features. Data-driven intelligent methods to fusing feature with a priori information have made great progress in the recent times. And there are two limitations to fusion methods. Firstly, the structure of the fusion method cannot be precisely determined. Secondly, the contribution of the final result of the fusion method cannot be confirmed. In this paper, a novel feedback connection mechanism inspired by the traditional feedback control system is proposed to solve the above-mentioned limitations. The overall structure consists of three parts. The first part is a abstract feature extractor, which consists of a multilayer graph neural networks (GCNs). The second part uses memory augmentation of long short-term memory (LSTM) to achieve initial fusion of feature and re-extraction of abstract features. In order to combine the two abstract features more effectively, the third part of the mechanism of feedback recursive update of feature fusion weights is proposed. Our proposed method is first tested for damage detection of photovoltaic (PV) modules. Another experiment is the inverse problem of self-magnetic flux leakage (SMFL) detection with stress imbalance detection on a pipeline experimental platform. The result show that the proposed method can improve the detection efficiency and analyze the contribution of feature to the detection results.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call