Time-Efficient Neural Architecture Search for Autonomous Underwater Vehicle Fault Diagnosis
Shicheng Pei, Huan Wang, Te Han
Abstract
An autonomous underwater vehicle (AUV) can replace a human to operate in a complex underwater environment, so it must have the ability of self-fault diagnosis. Existing deep learning-based diagnostic methods have achieved excellent performance, but designing effective neural network structures is a time-consuming and difficult task. Although neural network architecture search can automatically search effective neural network structures in a certain search space, neural architecture search (NAS) algorithms are usually slow and expensive; therefore, this article introduces a time-efficient NAS-based AUV fault diagnosis framework (TENAS-FD). TENAS-FD constructs a novel scoring algorithm that effectively gives a metric to characterize the performance of an untrained network. This metric is given based on the overlapping activation between data points in the untrained network with different inputs. This allows TENAS-FD to search for superior network architectures in seconds on a single graphics processing unit (GPU). Experiments were conducted on a real AUV dataset and showed that TENAS-FD can quickly obtain excellent network architectures for AUV fault diagnosis and has better diagnostic performance compared to hand-designing models.