Deep Multi-Attention Network with Earth Mover's Distance for Few-Shot fault Diagnosis of Guardrail Production Mold Platform

The structure of the guardrail production mold platform is a typical rotor system. Rotor system fault diagnosis based on few-shot learning aims to solve the problem of sample scarcity in practical engineering. Many methods based on metric-learning perform few-shot fault classification by learning a metric function. However, they have the following problems: (1) The method of using local descriptors to perform metric-learning does not consider the impact of irrelevant regions. In their method, relevant regions and irrelevant regions have the same importance. (2) Most few-shot fault diagnosis methods pay little attention to the problem of insufficient sample extraction, resulting in the failure to maximize the use of limited fault samples. To solve the above problems, deep multi-attention network with Earth Mover's Distance (EDMN) is proposed for few-shot intelligent fault diagnosis of rotor system. This method adopts a best matching between fault samples to solve the problem of few-shot fault diagnosis. Firstly, a deep network using multi-attention mechanism is applied to extract effective fault features of rotor system. Then, Earth Mover's Distance is used as a metric to calculate the structural distance between fault features. In order to generate different weights for different local descriptors, a cross-reference method is introduced to generate weights. We carried out a number of experiments to verify the effectiveness of the proposed method. The results show that our method has certain advantages over other few-shot learning fault diagnosis methods.