A novel fault diagnosis method under limited samples based on an extreme learning machine and meta-learning

Abstract

Background

Traditional fault diagnosis methods typically rely on an adequate number of fault samples. Yet, in industrial processes, the availability of faulty samples is often limited, accompanied by high sample collection cost. In this case, traditional methods prove ineffective for accurate diagnosis.

Methods

To solve this issue with limited samples, a novel fault diagnosis method, incorporating Extreme Learning Machine (ELM) and Meta-learning, is proposed. This method comprises two stages: Meta-learning optimization and top-model classification. In the first stage, the Model-Agnostic Meta-Learning framework is adopted to extract gain valuable model parameters from the available faulty data, yielding in the optimization of the initial weight and bias of the network model to obtain the reconstructed ELM. This enhancement significantly bolsters the ELM's parameters optimization capability, especially in scenarios with limited fault samples. Consequently, in the second stage, the reconstructed ELM model is deployed for effective fault diagnosis.

Significant Findings

The proposed method has proven successful in the real-time diagnosis of electric submersible pump, specifically addressing the challenging issue of ineffective fault diagnosis due to the limited fault samples in condition monitoring. The results showcase a 30 % classification improvement compared to ELM and an 8 % enhancement over PSO-ELM, FOS-ELM, LE-ELM, and Siamese Nets.