基于改进卷积神经网络的滚动轴承故障诊断模型

A Fault Diagnosis Model for Rolling Bearings Based on Improved Convolutional Neural Network

摘要: 针对传统故障诊断方法在滚动轴承实际工况下表现出的问题，本文给出了一个全新的滚动轴承故障诊断模型，即基于GAF-ICNN的新模型。该模型的核心理念是将一维振动信息转换为带有时间关系的二维特征图像，并采用格拉姆角场(GAF)的编码方式进行。这些特征图像被输入卷积神经网络(CNN)，用于自动特征提取和故障诊断。我们将原始数据的归一化预处理从传统的批归一化(Batch Normal-ization, BN)算法改为组归一化(Group Normalization, GN)算法，以提高模型的诊断性能。我们将凯斯西储大学的滚动轴承数据中的不同故障类型的数据进行分类，然后将数据用来验证我们的模型，证实了我们现在所认为的GAF-ICNN模型的有效性和优越性。我们通过采用对控制Mini-batch Size的变化和监控不同数据的规模大小等特点进行设计的GAF-ICNN模式进行了一般化稳定性检验，并开展了与传统智能算法的比较分析。研究结果表明，相比于其他常规的故障判定方式，我们所提供的模式在控制数据集较小和Mini-batch Size变化的情况下，对滚动轴承故障判定方式有着较良好的一般化稳定性和识别有效性。

Abstract: In response to the challenges posed by traditional fault diagnosis methods in practical operating conditions of rolling bearings, this paper presents a novel fault diagnosis model for rolling bearings, namely the GAF-ICNN-based model. The core concept of this model involves transforming one-dimensional vibration information into two-dimensional feature images with temporal rela-tionships, using Gramian Angular Field (GAF) encoding. These feature images are fed into a Convo-lutional Neural Network (CNN) for automatic feature extraction and fault diagnosis. Notably, we re-place the conventional Batch Normalization (BN) algorithm for normalizing raw data with the Group Normalization (GN) algorithm to enhance the diagnostic performance of the model. We classify data from different fault types in the rolling bearing dataset from Case Western Reserve University and employ this data to validate our model, confirming the effectiveness and superiority of the proposed GAF-ICNN model. We conduct a generalization stability test on the GAF-ICNN model, designed to control variations in Mini-batch Size and monitor different data set sizes. Additionally, we perform a comparative analysis with traditional intelligent algorithms. The research results indicate that, compared to other conventional fault detection methods, the proposed model demonstrates favora-ble generalization stability and recognition effectiveness in scenarios with limited dataset size and varying Mini-batch Size, highlighting its robustness in rolling bearing fault diagnosis.