【作者】 臧朝伟；

【导师】 牛强；

【作者基本信息】 中国矿业大学 ， 计算机应用技术， 2025， 博士

【摘要】 在现代矿山生产中,矿井提升机作为核心运输设备,其安全稳定运行至关重要。然而,矿井内高湿度、高粉尘等恶劣环境,容易导致提升机核心轴承性能衰退并引发故障。实际生产中,信号采集受系统与环境噪声干扰,特征提取困难,且轴承故障样本稀缺,数据分布失衡,使传统智能诊断模型易过拟合,泛化能力与诊断精度不足。此外,复杂工况下多源异构信号相互耦合交织,加剧了特征空间的高维复杂性,显著提升了多变量故障诊断的特征提取与分类难度。针对上述问题,本课题围绕提升机故障诊断展开研究,深入剖析煤矿环境对轴承振动信号的影响,探索基于融合模型的信号去噪技术,研究单变量故障诊断方法以精准识别轴承故障,并提出融合多传感器信号的诊断策略,降低多变量故障诊断复杂度。本文的主要研究工作包括:(1)基于全局与局部特征融合的振动信号去噪方法针对提升机等机械设备振动信号易受噪声干扰、影响故障诊断准确性的问题,构建U-Net与Transformer融合的去噪模型。在传统U-Net架构中嵌入Transformer层,强化模型对振动信号全局与局部特征的捕捉能力,借助编码器-解码器结构实现端到端去噪。结合振动信号去噪模型与CWT-Res Net18模型,在有效滤除提升机振动信号噪声的同时,提升故障诊断准确率。通过在私有数据集及美国凯斯西储大学轴承数据集上采用多种评价指标评估,验证了该方法的优异性能;并于不同噪声背景数据集上开展抗噪实验,进一步证实其强大的抗噪能力。(2)基于单变量二维振动信号的提升机故障诊断方法针对现有轴承振动信号分析模型难以充分挖掘原始时间序列特征的问题,提出基于周期特征二维转换与卷积神经网络的故障诊断策略。通过将原始一维振动数据重构为二维张量,结合残差网络的高效学习特性与多尺度特征提取能力,构建Bi-FDNet故障诊断模型,并对模型的二维转换机制、核心T-Inception模块及网络架构进行详细设计。实验表明,该模型在私有数据集不同工况下均展现优异性能,显著提升故障诊断准确率。同时,通过特征可视化与消融实验,系统揭示了模型各组件的功能与协同机制。(3)基于多变量二维振动信号的提升机故障诊断方法针对多变量故障诊断中特征提取与分类的复杂性,提出了一种基于注意力时间卷积和长短时记忆网络的多源诊断框架,充分利用来源于不同传感器的轴承振动信号实现精确故障诊断。在时间维度上对多源传感信号进行建模,探索了多源信号之间的时间相关性,提出了时间相关的多源信号表示方法。通过拼接操作进行结合,设计了并行的具有注意力机制的时间卷积网络和卷积长短时记忆网络框架,能够捕获信号的局部和全局信息。将所提取的特征应用于两个多变量公开和私有数据集,通过在这些数据集上执行的故障分类任务,有力地验证了所提方法在多变量故障诊断中的高效性与准确性。(4)提升机全状态监测系统原型设计与应用为了解决煤矿提升机在复杂井下环境中的维护难题,设计并构建了一套全状态数据监测与智能故障诊断原型系统。该系统聚焦传统维护模式中存在的资源孤岛化、响应滞后性、故障溯源困难等关键问题,通过融合远程智能监测技术、多源数据融合分析及先进故障诊断算法,构建了设备运行状态实时监测、故障精准诊断及运维决策支持的全链条技术体系。实验验证表明,该系统可实现提升机全生命周期健康状态的数字化管控,为煤矿安全生产提供了理论依据与技术支撑,具有显著的工程应用价值与行业推广潜力。该论文有图42个,表35个,参考文献100篇。更多还原

【Abstract】 In modern mine production,the mine hoist,as a core transportation equipment,plays a crucial role in ensuring safe and stable operation.However,the harsh environment inside the mine,such as high humidity and high dust content,leads to the performance degradation of the core bearings of the hoist and subsequently causes various faults.In actual production,the signal acquisition is interfered with by both system noise and environmental noise,making it difficult to extract features.Moreover,the scarcity of bearing fault samples further exacerbates the imbalance in data distribution,causing traditional intelligent diagnostic models to be prone to overfitting,with insufficient generalization ability and diagnostic accuracy.In addition,the interweaving of multi-source signals under complex working conditions further increases the difficulty of feature extraction and classification in multi-variable fault diagnosis.To address the above issues,this project focuses on the fault diagnosis of the hoist.It deeply analyzes the impact of the coal mine environment on the bearing vibration signals of the hoist,explores signal denoising techniques based on fusion models,studies univariate fault diagnosis methods to identify bearing faults accurately,and proposes a diagnostic strategy that fuses the bearing vibration signals from different sensors to reduce the complexity of multi-variable fault diagnosis.The main research work of this thesis includes:(1)Method on Vibration Signal Denoising based on the Fusion of Global and Local FeaturesAiming at the problem that the vibration signals of mechanical equipment such as hoists are vulnerable to noise interference,which affects the accuracy of fault diagnosis,a denoising model integrating U-Net and Transformer is constructed.The Transformer layer is embedded in the traditional U-Net architecture to enhance the model’s ability to capture the global and local features of vibration signals,and the encoder-decoder structure is used to achieve end-to-end denoising.By combining the vibration signal denoising model with the CWT-Res Net18 model,while effectively filtering out the noise in the hoist vibration signals,the accuracy of fault diagnosis is improved.Through evaluation with multiple evaluation indicators on the private dataset and the bearing dataset of Case Western Reserve University in the United States,the excellent performance of this method is verified.Moreover,anti-noise experiments are carried out on datasets with different noise backgrounds,further confirming its strong anti-noise ability.(2)Method on Fault Diagnosis of Hoist based on Univariable Two-dimensional Vibration SignalAiming at the problem that existing models for analyzing bearing vibration signals are complicated to fully exploit the features of the original time series,a fault diagnosis strategy based on two-dimensional transformation of periodic features and convolutional neural networks is proposed.By reconstructing the original one-dimensional vibration data into two-dimensional tensors,and combining the efficient learning characteristics and multi-scale feature extraction capabilities of the residual network,the Bi-FDNet fault diagnosis model is constructed.Moreover,the two-dimensional transformation mechanism,the core T-Inception module,and the network architecture of the model are designed in detail.Experiments show that the model demonstrates excellent performance under different working conditions in private datasets,significantly improving the accuracy of fault diagnosis.Meanwhile,through feature visualization and ablation experiments,the functions and collaborative mechanisms of each component of the model are systematically revealed.(3)Method on Fault Diagnosis of Hoist based on Multivariable Two-dimensional Vibration SignalAiming at the complexity of feature extraction and classification in multi-variable fault diagnosis,a multi-source diagnostic framework based on attention temporal convolution and long short-term memory networks is proposed.This framework makes full use of the bearing vibration signals from different sensors to achieve accurate fault diagnosis.The multi-source sensing signals are modeled in the time dimension,and the temporal correlations among the multi-source signals are explored.A temporal correlation-based representation method for multi-source signals is proposed.Through concatenation operations,a parallel framework of a temporal convolutional network with an attention mechanism and a convolutional long short-term memory network is designed,which is capable of capturing both local and global information of the signals.The extracted features are applied to two multi-variable public and private datasets.By performing fault classification tasks on these datasets,the effectiveness and accuracy of the proposed method in multi-variable fault diagnosis are strongly verified.(4)Prototype Design and Application of Hoist Full-State Monitoring SystemTo address the maintenance challenges of coal mine hoists in complex underground environments,a prototype system for full-state data monitoring and intelligent fault diagnosis has been designed and constructed.This system focuses on key issues existing in traditional maintenance models,such as resource isolation,lagging response,and difficult fault tracing.By integrating remote intelligent monitoring technology,multi-source data fusion analysis,and advanced fault diagnosis algorithms,a full-chain technical system for real-time monitoring of equipment operation status,accurate fault diagnosis,and operation and maintenance decision support has been established.Experimental verification shows that this system can achieve digital control of the health status throughout the life cycle of the hoist,providing a theoretical basis and technical support for the safe production of coal mines.It has significant engineering application value and great potential for industry-wide promotion.The thesis has 42 figures,35 tables,and 100 references.更多还原