Automatic Optimization of One-Dimensional CNN Architecture for Fault Diagnosis of a Hydraulic Piston Pump Using Genetic Algorithm
Oybek Eraliev, Kwang‐Hee Lee, Chul-Hee Lee
Abstract
A hydraulic piston pump is an essential component of a hydraulic transmission system and is extensively used in contemporary industrial settings. Therefore, fault diagnosis of piston pumps is a crucial topic in the engineering field. The convolutional neural network (CNN) is currently the most popular deep neural network model and has been successfully employed for fault detection and other tasks. The design and hyperparameter settings of CNNs significantly affect the overall diagnosis performance. In this study, a genetic method is proposed that can quickly investigate a specific set of potentially viable one-dimensional CNN (1D-CNN) architectures while also optimizing their hyperparameters for a fault detection task of an axial hydraulic piston pump. The proposed model is automatically designed based on a direct connect 1D-CNN block, which is another contribution of this study. The proposed method is evaluated on the raw sound signal dataset of an axial hydraulic piston pump without any signal pre-processing techniques. The experimental results demonstrate that the proposed method outperforms several well-known deep learning (DL) models in terms of fault diagnosis performance. Additionally, the suggested method uses significantly less computational power to determine the best 1D-CNN structures than most peer rivals.