基于深度学习的采煤机关键零部件剩余寿命预测.pdf

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剩余寿命预测 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 采煤机作为煤矿开采的主要设备之一， 拥有庞大的体型、 复杂的结构， 其关键零部件剩余寿命（remaining useful life，RUL）预测值受工作环境恶 劣、 操作空间狭窄等因素的影响难以准确获得， 使采煤机健康状态评估困难， 严重威胁煤矿安全生产及工作人员的生命安全。当前采煤机关键零部件可 靠性分析手段局限于基于软件、 数学模型等静态仿真的理论化分析， 未利用 监测数据进行挖掘和分析，导致分析结果片面、准确度差、效率低、智能化 程度滞后等缺陷。 研究和利用先进的理论与方法， 从煤机装备大数据中挖掘信息， 高效、 准确地识别装备的健康状态， 成为煤机装备健康监测领域面临的新问题。 结 合深度学习极强的非线性拟合能力优势，提出了基于深度学习的采煤机关 键零部件剩余寿命预测方法。依据监测对象实际退化趋势和监测数据特征， 将监测数据划分为全寿命周期和非全寿命周期两类数据，分别采用分类和 回归的思想构建深度神经网络剩余寿命预测模型，表征监测数据与剩余寿 命之间潜在的非线性映射关系， 并通过试验验证了模型的性能。 在模型研究 的基础上， 构建了采煤机关键零部件剩余寿命预测体系架构， 以采煤机摇臂 易损件作为实例对象，分析了构建的两种模型在关键零部件剩余寿命预测 方面的可行性。主要研究内容如下 （1）研究了采煤机各部件工作性能，分析了其关键零部件失效现象和 原因。 结合深度学习基本模型， 从结构、特征学习和反向参数优化三方面阐 述了各模型原理。提出了基于深度学习的采煤机关键零部件剩余寿命预测 方法。 （2）针对全寿命周期数据，提出基于分类思想的剩余寿命预测模型构 建方法。依据监测手段和监测数据特性，引进 3 sigma 准则去噪方法去除监 测数据中的粗大误差。 采用分层抽样手段获取训练集和测试集， 确保数据的 完整性。构建不含池化层的深度卷积神经网络（deep convolutional neural network，DCNN）预测模型，提高模型特征学习能力。经试验验证，该模型 具有高预测性和强泛化能力的优势。 （3）针对非全寿命周期数据，提出基于回归理念的剩余寿命预测模型 构建方法。采用快速傅里叶变换（fast Fourier trans，FFT）获取频域信 息。研究自编码器（auto-encoder，AE）自监督学习特性，提取输入的时域 和频域数据特征。通过在门循环单元（gated recurrent unit，GRU）网络隐藏 万方数据 太原理工大学硕士学位论文 II 层添加前向层，构建双向门循环单元（bidirectional gated recurrent unit，bi- GRU）预测模型，实现特征的双向学习；将 AE 提取的特征作为预测模型的 输入， 驱动 bi-GRU 预测剩余寿命。 经试验验证， 该模型具有准确预测能力。 （4）从模型结构、数据预处理和预测结果三方面对比了文章构建的两 种预测模型。 结合实际煤矿情况， 构建基于深度学习的采煤机关键零部件剩 余寿命体系框架。将构建的深度学习模型嵌入物联网数据分析层预测监测 对象剩余寿命， 研究了采煤机监测数据传输策略， 以采煤机摇臂易损件高速 区和低速区齿轮、 轴承为实例对象， 探讨了构建的两种模型在摇臂易损件剩 余寿命预测方面的可行性。 基于深度学习的采煤机关键零部件剩余寿命预测方法，通过深度学习 利用机械监测信号训练深度神经网络，其优势在于能够摆脱对大量信号处 理技术与预测经验的依赖， 克服了传统预测方法的缺陷， 完成特征的自适应 提取， 实现了自主学习和动态预测， 提高了预测结果的准确性和分析手段智 能化程度，为采煤机预测性维护策略的有效实施提供指导。 关键关键词词采煤机；深度学习；剩余寿命预测；DCNN；AE bi-GRU 万方数据 ABSTRACT III ABSTRACT As one of the main equipment of coal mining, shearer owns the large size and complex structure, and the remaining useful life RUL prediction value of its key parts is difficult to obtain accurately due to the influence of poor working environment, narrow operating space and other factors, which leads to the difficulty to predict shearer health status and seriously threatens the safety of coal mine production and the life safety of workers. At present, the of reliability analysis for the key parts of shearer is mainly limited to the theoretical analysis based on static simulation such as software and mathematical model, and the monitoring data is not used for mining and analysis, resulting in the defects such as analysis results unilateral, inaccurate, low in the analysis efficiency, and lagging behind in the intellectualization. Researching and using advanced theories and s to mine ination from the big data of coal machinery equipment, and efficiently and accurately identify the health status of equipment has become a new problem in the field of coal machinery equipment health monitoring. Combining the advantages of deep learning with extremely strong nonlinear fitting capabilities, a for the RUL prediction of shearer key parts based on deep learning is proposed. According to the actual degradation trend of monitoring objects and the characteristics of monitoring data, the monitoring data are divided into two types life cycle and non-life cycle data. The RUL prediction models based on deep neural network are constructed by using the concept of classification and regression respectively, which characterizes the potential non-linear mapping relationship between monitoring data and RUL. The perance of the model is verified by experiments. Based on the model research, the RUL prediction system architecture of the shearer key parts is built. Taking the vulnerable parts of the shearer rocker arm as an example, the feasibility of the two models in RUL prediction of the key parts is analyzed. The main research contents are as follows 1 The working perance of each part of the shearer is studied. The failure phenomena and reasons of its key parts are analyzed. For the basic model of deep learning, the principles of each model are described from three aspects structure, feature learning and reverse parameter optimization. The research 万方数据 太原理工大学硕士学位论文 IV of the RUL prediction for the shearer key parts based on the deep learning is put forward. 2 For the full life cycle data, a model is built based on the concept of classification to predict the RUL. According to the monitoring means and the characteristics of the monitoring data, the 3 sigma criterion denoising is introduced to remove the gross errors in the monitoring data. In order to ensure the integrity of the data, the of stratified sampling is applied to obtain the training set and test set. Then, a deep convolutional neural network DCNN prediction model without pooling layers is constructed to improve the features learning ability of model. The experimental results show that the proposed model has the advantages of high predictability and strong generalization. 3 For the non-full life cycle data, an RUL prediction model is built based on the concept of regression. Fast Fourier trans FFT is applied to obtain the frequency doamin ination. The data features of time domain and frequency domain are get by self supervised learning of auto-encoder AE. The bidirectional gated recurrent unit bi-GRU prediction model is constructed by adding forward layers in the hidden layer of gated recurrent unit GRU, which realizes the bidirectional learning of data features, the extracted features by AE are regarded as the of prediction model, driving bi-GRU to predict the RUL. It is verified by experiments that the model has the ability of accurate prediction. 4 The two proposed prediction models in this paper are compared from structures, data preprocessing and prediction results. Combined with the actual situation of coal mine, the system framework of RUL prediction scheme for shearer key parts based on deep learning is constructed. Then, the proposed deep learning models are embedded into the data analysis layer of the Internet of things IoT to realize the RUL prediction. The data transmission strategy of shearer monitoring is studied. Taking the gears and bearings in the high-speed and low- speed areas of the vulnerable parts of the shearer rocker arm as examples, the feasibility of the two models in predicting the RUL of the vulnerable parts of the rocker arm is discussed in this paper. The of RUL prediction for the shearer key parts based on the deep learning is used to train the deep neural network by the mechanical monitoring signals. The advantage of deep learning is that it can get rid of the dependence on a large number of signal processing technology and prediction experience, 万方数据 ABSTRACT V overcome the shortcomings of traditional prediction s, complete the adaptive extraction of features, realize self-learning and dynamic prediction, improve the prediction accuracy of the prediction results and the intelligent level of the analysis means, and provide guidance for the effective implementation of the predictive maintenance strategy for the shearer. Key Words Shearer; Deep Learning; Remaining Useful Life Prediction; DCNN; AE bi-GRU 万方数据 太原理工大学硕士学位论文 VI 万方数据 目 录 VII 目 录 摘 要 ..................................................................................................................................... I ABSTRACT ............................................................................................................................. III 目 录 ................................................................................................................................ VII 第一章 绪论 ............................................................................................................................ 1 1.1 研究目的及意义 ........................................................................................................... 1 1.2 国内外研究现状 ........................................................................................................... 2 1.2.1 采煤机关键零部件可靠性 .................................................................................... 2 1.2.2 基于深度学习的机械零件剩余寿命预测 ............................................................ 4 1.2.3 存在的问题 ............................................................................................................ 7 1.3 研究内容 ....................................................................................................................... 8 1.4 论文结构 ....................................................................................................................... 9 第二章 基于深度学习的采煤关键零部件剩余寿命预测理论基础 .................................. 11 2.1 引言 ............................................................................................................................. 11 2.2 采煤机关键零部件失效原理 ..................................................................................... 11 2.2.1 采煤机结构 .......................................................................................................... 11 2.2.2 采煤机摇臂结构 .................................................................................................. 13 2.2.3 采煤机摇臂关键零部件失效形式 ...................................................................... 14 2.3 深度学习基本理论 ..................................................................................................... 15 2.3.1 监督学习模型 ...................................................................................................... 15 2.3.2 无监督学习模型 .................................................................................................. 24 2.4 基于深度学习的采煤机关键零部件剩余寿命预测技术路线 ................................. 28 2.5 本章小结 ..................................................................................................................... 29 第三章 基于 DCNN 的剩余寿命预测模型 ......................................................................... 31 3.1 引言 ............................................................................................................................. 31 3.2 基于 DCNN 的剩余寿命预测流程 ............................................................................ 31 3.3 数据预处理 ................................................................................................................. 32 3.3.1 数据去噪 .............................................................................................................. 33 3.3.2 数据特征提取 ...................................................................................................... 33 万方数据 太原理工大学硕士学位论文 VIII 3.3.3 数据集划分 .......................................................................................................... 37 3.4 DCNN 预测模型构建 ................................................................................................ 38 3.4.1 DCNN 模型结构 ................................................................................................. 38 3.4.2 DCNN 模型参数设置 ......................................................................................... 39 3.4.3 预测模型评估指标 .............................................................................................. 40 3.5 DCNN 模型预测性能验证实验 ................................................................................ 41 3.5.1 DCNN 模型预测精度验证 ................................................................................. 42 3.5.2 DCNN 模型泛化能力验证 ................................................................................. 51 3.5.3 分层抽样有效性验证 .......................................................................................... 55 3.5.4 实验结果与分析 .................................................................................................. 57 3.6 本章小结 ..................................................................................................................... 57 第四章 基于 AE bi-GRU 的剩余寿命预测模型 ................................................................ 59 4.1 引言 ............................................................................................................................. 59 4.2 基于 AE bi-GRU 的剩余寿命预测流程 ................................................................... 59 4.3 数据预处理 ................................................................................................................. 60 4.3.1 数据去噪 .............................................................................................................. 60 4.3.2 时域和频域特征选择 .......................................................................................... 61 4.3.3 基于 AE 的特征提取及数据集划分 .................................................................. 61 4.4 bi-GRU 预测模型构建 ............................................................................................... 62 4.4.1 GRU 模型结构 .................................................................................................... 62 4.4.2 bi-GRU 模型结构 ................................................................................................ 63 4.4.3 bi-GRU 模型参数设置 ........................................................................................ 63 4.5 AE bi-GRU 剩余寿命预测性能验证实验 ................................................................ 65 4.5.1 AE bi-GRU 预测效果验证 ................................................................................. 65 4.5.2 AE 对模型预测结果的影响 ............................................................................... 69 4.5.3 实验结果与分析 .................................................................................................. 70 4.6 本章小结 ..................................................................................................................... 70 第五章 基于深度学习的采煤机关键零部件剩余寿命预测分析 ...................................... 71 5.1 引言 ............................................................................................................................. 71 5.2 采煤机关键零部件剩余寿命预测体系框架 ............................................................. 71 5.2.1 体系框架 .............................................................................................................. 71 万方数据 目 录 IX 5.2.2 基于物联网的采煤机数据传输策略 .................................................................. 72 5.2.3 基于深度学习的采煤机关键零部件剩余寿命预测流程 .................................. 73 5.3 采煤机零件剩余寿命预测深度学习模型对比 ......................................................... 74 5.3.1 DCNN 和 AE bi-GRU 模型结构对比 ................................................................. 74 5.3.2 DCNN 和 AE bi-GRU 模型数据预处理对比 ..................................................... 75 5.3.3 DCNN 和 AE bi-GRU 模型预测效果对比 ......................................................... 75 5.3.4 对比结果分析 ..........................................................................................