随着工业自动化和数据科学技术的发展，工业生产过程中产生了海量的设备状态监测数据，其中包含了丰富的设备运行性能、动态趋势、异常征兆等相关信息。本文致力于通过深度学习和时间序列分析方法对这些设备运行信息进行挖掘，针对工业设备数据数量繁多、时序关联性强、多模态、工况多变的特性，建立基于状态监测数据的设备状态识别模型，实现设备运行状态的异常监测和早期预警，为工业设备运行状态监测提供一套通用的解决方案。论文所作研究将对工业系统的实时智能监测、维修运行策略优化和安全经济运行有重要的积极意义。论文在全面深入分析基于数据驱动的设备状态识别方法的发展现状、技术特点和在工业领域应用的难点的基础上，提出了结合时间序列分析和深度学习模型进行设备状态识别的技术路线：针对工业数据的强时序关联性，提出使用时间序列分析方法提取时序特征；针对复杂工业系统的非线性和多模态特性，提出使用深度学习模型自主学习并抽象深层特征，提高设备状态识别的准确性。为解决工业过程的典型时间序列数据处理中的平移、缩放以及异常噪声问题，论文通过改进卷积神经网络结构，采用卷积核局部视野的思想，设计子序列特征提取层，对时间序列的子序列间的相似性进行特征提取，之后通过网络迭代计算得到两条时间序列的相似性距离。在相似性度量模型的基础上，进一步采用K最近邻算法构建设备的状态识别模型，在基于时序相似性度量的设备工况分类中表现出良好的性能。为了解决工业数据的强时序关联性、样本不均衡等难题，论文通过集成基于动态时间规整和模糊隶属度的时序变化特征提取方法、基于卷积神经网络的平稳参数特征提取方法、基于改进平方对数损失函数的逻辑回归方法，建立针对时序动态变化设备的状态识别框架，在关键参数时序变化特征提取和故障小样本工况识别等方面表现出优异的性能。针对工业过程中存在的多变工况特性和参数时变性，论文基于长短时记忆网络建立设备关键参数的重构模型，基于残差分析提出自适应阈值方法的健康度指标设计，形成设备状态在线识别框架，在平衡误报警与漏报警、早期故障征兆识别等方面有良好表现。

With the increasing automation of industry and rapaid devoploment of data science technology, massive equipment condition monitoring data has been produced in the industrial production process. Information related to equipment operating performance, dynamic trends and abnormal symptoms is usually contained in these data. This dissertation is devoted to mining these device information through time series analysis and deep learning methods. Industrial data has characteristics of large quantity, strong temporal correlation, multi-modality and variable working conditions. This dissertation builds models of equipment status identification based on condition monitoring data, which has ability to carry out abnormal monitoring and fault early warning of equipment and provides a set of general solutions for industrial equipment condition monitoring. This research will have positive significance for real-time intelligent monitoring of industrial systems, optimization of maintenance strategies, and the improvement of opreational safety and efficiency.Based on a comprehensive and in-depth analysis of the development status, technical characteristics and difficult challenges of data-driven equipment state identification methods, this research presents a technical route for industry equipment state identification by combining time series analysis and deep learning models: For the strong temporal correlation of industrial data, time seires analysis is adpoted to extract temporal characteristics. In view of the non-linear and multi-modal characteristics of complex industrial systems, deep learning models are adpoted to automatically learn and abstract deep features to improve the accuracy of equipment state identification.In order to solve the problems of translation, scaling, and abnormal noise in typical time series data of industrial processes, this research improves the structure of the convolutional neural network by using the idea of partial field of the convolution kernel to design a subsequence feature extraction layer. Feature extraction is based on the similarity between segment sequences and the similarity distance between two time series is calculated through network iteration. On the basis of the similarity measure model, the K-nearest neighbor algorithm is further adopted as equipment state identification model, which shows good performance in the classification of equipment operating conditions.In order to solve the problems of temporal correlation and sample imbalance of industrial data, a sequential dynamic equipment state identification framwork is proposed in this research, through the integration of time-varying feature extraction methods based on dynamic time warping and fuzzy membership, stationary parameter feature extraction methods based on convolutional neural networks, and improved logistic regression method based on squared log loss function. The framework shows excellent performance in terms of time-varying feature extraction of key parameters and identification of small sample failure conditions.Aiming at the variable conditions and time-varying parameters in the industrial process, this research builds a reconstruction model of the key parameters of equipment based on long short term memory networks. An adaptive threshold method based on residual analysis is proposed to design the health index of equipment condition monitoring. The equipment state on-line identification framework has a good performance in balancing false alarms and missed alarms and early failure symptoms identification.