Department of Electrical, Electronics and Computer Engineering (BK21Four), University of Ulsan, Ulsan 44610, Korea.
School of Computer Science and Engineering, Soongsil University, Seoul 06978, Korea.
Sensors (Basel). 2020 Dec 22;21(1):18. doi: 10.3390/s21010018.
Gearbox fault diagnosis based on the analysis of vibration signals has been a major research topic for a few decades due to the advantages of vibration characteristics. Such characteristics are used for early fault detection to guarantee the enhanced safety of complex systems and their cost-effective operation. There exist many fault diagnosis models that have been developed for classifying various fault types in gearboxes. However, the classification results of the conventional fault classification models degrade when they are applied to gearbox systems with multi-level tooth cut gear (MTCG) faults operating under variable shaft speeds. These conditions cause difficulty in discriminating the gear fault types. Due to the improved computational capabilities of modern systems, the application of deep neural networks (DNNs) is getting popular in a variety of research fields, such as image and natural language processing. DNNs are capable of improving the classification results even when addressing complex problems such as diagnosing gearbox MTCG faults. In this research, an adaptive noise control (ANC) and a stacked sparse autoencoder-based deep neural network (SSA-DNN) are used to construct a sensitive fault diagnosis model that can diagnose a gearbox system with MTCG fault types under varying shaft rotation speeds, despite its complicatedness. An ANC is applied to gear vibration characteristics to remove a significant level of noise along the frequency spectrum of vibration signals to fix the most fault-informative components of each fault case. Next, the autoencoder learns the gear faults characteristic features from these fault-informative components to separate the fault types considered in this study. Furthermore, the implementation of the SSA-DNN is substituted for feature extraction, feature selection, and the classification processes in traditional fault diagnosis schemes by high-performance unity. The experimental results show that the proposed model outperforms conventional methodologies with higher classification accuracy.
基于振动信号分析的齿轮箱故障诊断由于振动特性的优势，几十年来一直是一个主要的研究课题。这些特性可用于早期故障检测，以保证复杂系统的增强安全性和具有成本效益的运行。已经开发了许多故障诊断模型，用于对齿轮箱中的各种故障类型进行分类。然而，当应用于在变速轴速下运行的具有多级切齿齿轮 (MTCG) 故障的齿轮箱系统时，传统故障分类模型的分类结果会降低。这些条件使得难以区分齿轮故障类型。由于现代系统计算能力的提高，深度神经网络 (DNN) 的应用在图像和自然语言处理等各种研究领域越来越流行。即使在处理诊断齿轮箱 MTCG 故障等复杂问题时，DNN 也能够提高分类结果。在这项研究中，自适应噪声控制 (ANC) 和基于堆叠稀疏自动编码器的深度神经网络 (SSA-DNN) 用于构建敏感故障诊断模型，即使在变速轴转速下，该模型也可以诊断具有 MTCG 故障类型的齿轮箱系统，尽管其复杂性很高。ANC 应用于齿轮振动特性，以去除振动信号频谱上的大量噪声水平，以固定每个故障情况下最具故障信息量的组件。接下来，自编码器从这些故障信息量组件中学习齿轮故障特征，以分离本研究中考虑的故障类型。此外，SSA-DNN 的实现通过高性能的统一性替代了传统故障诊断方案中的特征提取、特征选择和分类过程。实验结果表明，所提出的模型具有更高的分类精度，优于传统方法。
