基于深度学习的采煤机关键零部件故障诊断.pdf
万方数据 万方数据 万方数据 万方数据 硕士学术学位、硕士非工程类专业学位 学位论文答辩信息表 论文题目 基于深度学习的采煤机关键零部件故障诊断 课题来源* 省科技厅项目 论文答辩日期 2020 年 6 月 11 日 答辩秘书 谢嘉成 学位论文答辩委员会成员 姓名 职称 博导/硕导 工作单位 答辩委员会 主席 李秀红 教授 博导 太原理工大学 答辩委员 1 庞新宇 副教授 硕导 太原理工大学 答辩委员 2 崔红伟 副教授 硕导 太原理工大学 *课题来源可填国家重点研发计划项目、国家自然科学基金项目、国家 社科基金项目、教育部人文社科项目、国家其他部委项目、省科技厅项 目、省教育厅项目、企事业单位委托项目、其他 万方数据 摘 要 I 摘 要 采煤机作为煤炭开采的重要机械设备, 其结构复杂、 可靠性高、 功率大, 并且易于维护。 其中, 采煤机的设备安全和稳定运行对于保证煤炭的开采效 率具有重要意义。由于采煤机长期处在电磁干扰严重和潮湿的恶劣工作环 境,经常会出现某些关键零部件的故障,例如在摇臂传动系统中,齿轮易发 生磨损、断裂和剥落等故障。这些故障的发生容易导致生产效率下降,甚至 出现人员伤亡事故。 因此, 对采煤机的关键零部件进行故障诊断尤为重要。 目前,传统的故障诊断方法常采用时域分析、频域分析和 BP 神经网络等对 采煤机关键零部件进行故障识别。 但是, 传统的故障诊断方法大多通过少量、 小规模数据分析实现, 在面对采煤机多工况交替、 故障信息不明的海量信号 时,存在分析结果片面、准确率差、效率低和智能化程度滞后等问题。针对 上述问题,本文以采煤机截割部摇臂传动系统中高速区的齿轮为研究对象, 分别构建一维卷积神经网络 (1D-CNN) 和预激活深度残差网络 (PA-DRN) 的深度学习算法模型,并利用上述模型对采煤机摇臂齿轮的故障诊断进行 研究,研究内容主要包括以下三个方面 (1)基于深度学习中卷积神经网络的基本理论,构建适应于原始时域 信号的 1D-CNN 故障诊断模型,通过 dropout 策略和批量归一化(BN)对 模型结构进行优化。 利用 CWRU 轴承数据对提出的 1D-CNN 模型进行验证 实验。 将轴承的振动信号进行数据增强处理,并将其划分为训练集、 验证集 和测试集。利用构建的 1D-CNN 模型对一维轴承振动信号样本进行训练、 验证和测试,通过与深度神经网络(DNN) 、稀疏自编码器(SAE)和深度 置信网络(DBN)模型做对比,结果表明构建的 1D-CNN 模型在轴承上的 综合识别率达到了 100。 (2)将构建的 1D-CNN 模型应用于采煤机摇臂齿轮的故障分类。首先 利用采煤机摇臂加载试验台和加速度传感器模拟齿轮的不同故障,并收集 其振动信号。然后,通过 1D-CNN 模型分析模型不同训练参数对摇臂齿轮 分类效果造成的影响。最后,通过 t-SNE 降维可视化技术和基于混淆矩阵 的评估方法分析 1D-CNN 模型在采煤机摇臂齿轮上的分类性能。实验结果 表明 1D-CNN 模型有效的提高了采煤机摇臂齿轮的分类精度,其分类精度 达到 98.60。 (3)通过优化残差学习模块构建 PA-DRN 故障诊断模型。一方面,利 用 CWRU 轴承数据对 PA-DRN 模型进行实验验证, 通过混淆矩阵可视化和 万方数据 太原理工大学硕士学位论文 II 泛化性能验证实验证明了 PA-DRN 模型的有效性;另一方面,将 PA-DRN 故障诊断模型应用于采煤机摇臂齿轮的故障识别。通过混淆矩阵的评估方 法、分类过程可视化和模型对比分析,证明 PA-DRN 模型在采煤机摇臂齿 轮的故障识别上具有良好的分类效果,其分类精度达到 99.07。 关键词关键词采煤机;故障诊断;深度学习;一维卷积神经网络;深度残差网络 万方数据 ABSTRACT III ABSTRACT As an important mechanical equipment in coal mining, shearer has complex structure, high reliability, high power and easy maintenance. Among them, the safe and stable operation of shearer equipment is of great significance to ensure the mining efficiency of coal. However, because the shearer is in a bad working environment with serious electromagnetic interference and humidity for a long time, some key parts often fail. For example, in the shearer rocker transmission system, the gear is easy to wear, break and crack. However, the occurrence of these failures is likely to lead to the decline of production efficiency and even casualties. Therefore, the fault diagnosis of the key parts of the shearer is particularly important. At present, the traditional fault diagnosis s such as time domain analysis, frequency domain analysis and BP neural network are often used to identify the faults of the key parts of the shearer. However, traditional fault diagnosis s are mostly realized by small-scale data analysis. In the face of a large number of signals with multiple working conditions alternation and unknown fault ination, there are some problems such as one-sided analysis results, poor accuracy, low efficiency and lagging intelligence. In view of the above problems, this paper takes the high speed gear of the shearer rocker arm cutting part as the research object, and adopts the deep learning algorithm model of one-dimensional convolution neural network 1D-CNN and pre- activation deep residual network PA-DRN to study the gearing fault diagnosis of the shearer rocker arm. The main research contents include the following three aspects 1 1D-CNN fault diagnosis model suitable for the original time domain signal is constructed based on the basic theory of convolutional neural network in deep learning, The structure of the model is optimized by dropout strategy and batch normalization BN. The proposed 1D-CNN model is validated by CWRU bearing data. Firstly, the vibration signal of bearing is processed by data enhancement, and it is divided into training set, validation set and test set. Then, the 1D-CNN model is used to train, verify and test the one-dimensional vibration signal samples of bearing. By comparing with deep neural networkDNN, sparse autoencoderSAE and deep belief networkDBN models, the results show that 万方数据 太原理工大学硕士学位论文 IV the comprehensive recognition rate of the 1D-CNN model on the bearing is 100. 2 The 1D-CNN model is applied to the gearing fault classification of the shearer rocker arm. Firstly, the loading test-bed of the shearer rocker arm and the acceleration sensor are used to simulate the different faults of the gears, and the vibration signals are collected. Then, 1D-CNN model is used to analyze the influence of different training parameters of the model on the gearing classification effect of rocker arm. Finally, the gearing classification perance of 1D-CNN model on the shearer rocker arm is analyzed by t-SNE and uation based on confusion matrix. The experimental results show that the 1D- CNN model effectively improves the gearing classification accuracy of the shearer rocker arm, which reaches 98.60. 3 The PA-DRN fault diagnosis model is constructed by optimizing the residual learning module. On the one hand, the PA-DRN model is validated by CWRU bearing data, and the effectiveness of PA-DRN model is proved by confusion matrix visualization and generalization perance validation experiments. On the other hand, the PA-DRN fault diagnosis model is applied to the gearing fault identification of the shearer rocker arm. Through the uation of confusion matrix, the visualization of classification process and the comparative analysis of models, it is proved that PA-DRN model has a good classification effect on the gearing fault identification of shearer rocker arm, and its classification accuracy reaches 99.07. Key Words Shearer; Fault Diagnosis; Deep Learning; One-dimensional Convolution Neural Network; Deep Residual Network 万方数据 目 录 V 目 录 摘 要 ......................................................................................................................................... I ABSTRACT ............................................................................................................................. III 第一章 绪论 ............................................................................................................................ 1 1.1 研究目的及意义 ........................................................................................................... 1 1.2 国内外研究动态 ........................................................................................................... 2 1.2.1 故障诊断技术 ........................................................................................................ 2 1.2.2 深度学习故障诊断方法 ........................................................................................ 4 1.2.3 采煤机故障诊断技术 ............................................................................................ 6 1.3 研究内容 ....................................................................................................................... 7 1.4 技术路线 ....................................................................................................................... 8 第二章 采煤机结构及故障分析 .......................................................................................... 11 2.1 引言 ............................................................................................................................. 11 2.2 滚筒式采煤机结构 ..................................................................................................... 11 2.2.1 滚筒式采煤机组成 .............................................................................................. 11 2.2.2 滚筒式采煤机截割部 .......................................................................................... 12 2.3 滚筒式采煤机故障分析 ............................................................................................. 15 2.3.1 滚筒式采煤机故障及故障原因 .......................................................................... 15 2.3.2 截割部常见故障及故障原因 .............................................................................. 17 2.4 本章小结 ..................................................................................................................... 18 第三章 卷积神经网络理论基础 .......................................................................................... 19 3.1 引言 ............................................................................................................................. 19 3.2 卷积神经网络基本结构 ............................................................................................. 19 3.2.1 卷积层 .................................................................................................................. 20 3.2.2 池化层 .................................................................................................................. 22 3.2.3 全连接层 .............................................................................................................. 22 3.3 卷积神经网络技术特点 ............................................................................................. 23 3.3.1 激活函数 .............................................................................................................. 23 3.3.2 目标函数 .............................................................................................................. 24 3.3.3 优化算法 .............................................................................................................. 24 万方数据 太原理工大学硕士学位论文 VI 3.4 卷积神经网络反向传播 ............................................................................................. 26 3.5 本章小结 ..................................................................................................................... 27 第四章 基于时域振动信号的 1D-CNN 故障诊断模型 ..................................................... 29 4.1 引言 ............................................................................................................................. 29 4.2 1D-CNN 故障诊断方法的设计 ................................................................................. 29 4.2.1 1D-CNN 故障诊断模型构建 .............................................................................. 29 4.2.2 基于 dropout 策略和批量归一化(BN)的模型优化 ...................................... 31 4.2.3 数据增强技术 ...................................................................................................... 33 4.2.4 故障诊断算法流程 .............................................................................................. 34 4.3 1D-CNN 模型验证实验 ............................................................................................. 35 4.3.1 CWRU 试验台介绍 ............................................................................................. 35 4.3.2 轴承数据集选取 .................................................................................................. 36 4.3.3 训练结果分析 ...................................................................................................... 38 4.3.4 模型对比分析 ...................................................................................................... 39 4.4 实例测试与分析 ......................................................................................................... 42 4.4.1 采煤机摇臂加载试验台数据采集 ...................................................................... 42 4.4.2 模型训练参数分析 .............................................................................................. 46 4.4.3 基于混淆矩阵的摇臂齿轮故障分类性能评估 .................................................. 49 4.4.4 摇臂齿轮故障分类过程可视化 .......................................................................... 51 4.5 本章小结 ..................................................................................................................... 52 第五章 基于时域振动信号的 PA-DRN 故障诊断模型 ..................................................... 53 5.1 引言 ............................................................................................................................. 53 5.2 PA-DRN 故障诊断方法的设计 ................................................................................. 53 5.2.1 残差学习模块基本原理 ...................................................................................... 53 5.2.2 残差学习模块优化 .............................................................................................. 55 5.2.3 PA-DRN 故障诊断模型构建 .............................................................................. 57 5.3 PA-DRN 模型验证实验 ............................................................................................. 58 5.3.1 轴承数据集选取 .................................................................................................. 58 5.3.2 实验结果分析 ...................................................................................................... 59 5.3.3 泛化性能实验验证 .............................................................................................. 61 5.4 实例测试与分析 ......................................................................................................... 61 5.4.1 摇臂齿轮数据集选取 .......................................................................................... 61 万方数据 目 录 VII 5.4.2 实验结果分析 ...................................................................................................... 62 5.4.3 摇臂齿轮故障分类过程可视化 .......................................................................... 64 5.4.4 模型对比分析 ...................................................................................................... 65 5.5 本章小结 ..................................................................................................................... 66 第六章 结论与展望 .............................................................................................................. 67 6.1 研究结论 ..................................................................................................................... 67 6.2 进一步工作展望 ......................................................................................................... 67 参考文献 .................................................................................................................................. 69 攻读学位期间发表的学术论文 .............................................................................................. 75 致 谢 .................................................................................................................................... 797 万方数据 太原理工大学硕士学位论文 VIII 万方数据 绪论 1 第一章 绪论 1.1 研究目的及意义 目前,煤矿企业安全问题已经成为制约煤炭工业发展的突出问题之一。预防和控制 煤矿事故的发生,促进煤矿安全稳定生产形势的根本好转已经成为我国急需解决的重大 问题[1]。 在煤矿企业的实际生产中, 采煤机作为综采生产的核心设备,承担着主要的采煤 任务。 它是一个集机械、 电气和液压为一体的大型复杂系统[2]。 由于采煤机工作环境差, 自身的构造组成较复杂,直接导致了采煤机在运行过程中,经常发生各种类型的故障, 其运行的安全可靠性直接影响采煤机的安全生产和正常运行。这给企业造成了巨大的经 济损失和人员伤亡,严重影响了企业的正常发展。 从上世纪 60 年代到现在,国内外学者对采煤机故障诊断技术的研究越来越多,目 前针对采煤机的故障诊断方法主要包括温度压力监测诊断法、 专家系统法、 支持向量机、 BP 神经网络等传统方法,虽然这些传统的故障诊断方法在采煤机上取得了良好的效果, 但是仍存在一些缺陷 (1)故障分辨率较低。在采煤机的实际运行中,由于采煤机结构比较复杂,测试信 号和检测位置受到了限制,影响了测试结果的准确性,从而导致故障诊断结果具有模糊 性,容易使采煤机对实际发生的故障类型出现误判现象。 (2)诊断效率较低。诸如支持向量机和 BP 神经网络故障诊断方法,均需要对收集 的采煤机监测数据进行人工提取特征,然后对提取的特征根据一定的标准择优选择,最 后将选择的特征输入到分类器实现故障的识别。其中,特征提取、特征选择和分类器分 类三个过程单独进行,使得诊断耗费时间较长,从而降低了采煤机的故障诊断效率。 (3)对不确定信息的处理能力较差。在采煤机的诊断系统中通常存在大量的不确 定信息,这些信息可能是随机的,