A Generic Intelligent Bearing Fault Diagnosis System Using Compact Adaptive 1D CNN Classifier
A Generic Intelligent Bearing Fault Diagnosis System Using Compact Adaptive 1D CNN Classifier
复制标题
DOI：
10.1007/s11265-018-1378-3
复制
发表时间：
2019-02-01
影响因子：
1.8
通讯作者：
Kiranyaz, Serkan
中科院分区：
文献类型：
--
作者：
Eren, Levent;Ince, Turker;Kiranyaz, Serkan
及时、准确的轴承故障检测和诊断对于工业系统可靠、安全运行具有重要意义。在本研究中，广泛研究了采用紧凑型自适应一维卷积神经网络（CNN）分类器的通用实时感应轴承故障诊断系统的性能。在文献中，尽管许多研究开发了用于检测轴承故障的高精度算法，但其结果通常仅限于相对较小的训练/测试数据集。与通常将特征提取、特征选择和分类封装为不同模块的传统智能故障诊断系统相反，所提出的系统直接采用原始时间序列传感器数据作为输入，并且可以通过适当的训练有效地学习最佳特征。基于 1D CNN 的方法的主要优点是 1) 其紧凑的架构配置（而不是复杂的深度架构）仅执行 1D 卷积，使其适合实时故障检测和监控，2) 其经济高效且实用的实时硬件实现，3) 无需任何预先确定的变换（例如 FFT 或 DWT）、手工特征提取和特征选择即可工作的能力，以及 4) 能够以有限的尺寸提供分类器的有效训练。训练数据集和有限次数的BP迭代。通过将基于一维 CNN 的故障诊断方法应用于两个常用的基准真实振动数据集，并将结果与其他竞争的智能故障诊断方法进行比较，验证了该方法的有效性和可行性。
Timely and accurate bearing fault detection and diagnosis is important for reliable and safe operation of industrial systems. In this study, performance of a generic real-time induction bearing fault diagnosis system employing compact adaptive 1D Convolutional Neural Network (CNN) classifier is extensively studied. In the literature, although many studies have developed highly accurate algorithms for detecting bearing faults, their results have generally been limited to relatively small train/test data sets. As opposed to conventional intelligent fault diagnosis systems that usually encapsulate feature extraction, feature selection and classification as distinct blocks, the proposed system takes directly raw time-series sensor data as input and it can efficiently learn optimal features with the proper training. The main advantages of the 1D CNN based approach are 1) its compact architecture configuration (rather than the complex deep architectures) which performs only 1D convolutions making it suitable for real-time fault detection and monitoring, 2) its cost effective and practical real-time hardware implementation, 3) its ability to work without any pre-determined transformation (such as FFT or DWT), hand-crafted feature extraction and feature selection, and 4) its capability to provide efficient training of the classifier with limited size of training data set and limited number of BP iterations. Effectiveness and feasibility of the 1D CNN based fault diagnosis method is validated by applying it to two commonly used benchmark real vibration data sets and comparing the results with the other competing intelligent fault diagnosis methods.