基于深度卷积网络的液压复合故障智能诊断研究

Hydraulic composite fault diagnosis based on signal is a difficult diagnosis problem. The intelligent diagnosis based on online monitoring signal is of great significance as for making the equipment to have "self-sensing" ability and to ensure its safe, economical and reliable operation..Based on the track of the latest achievements in Artificial Intelligence and the depth analysis on the characteristics of hydraulic composite fault, this project will introduce a new intelligent diagnosis model with deep structure, which is deep Convolution Neural Network, abbreviated as DCNN..One of the project’s objectives is to try to use 2-D DCNN to realize the intelligent recognition of Hilbert time-frequency graph. The other is to integrate the several independent steps of intelligent diagnosis, including signal feature extraction and transformation, information fusion and pattern classification, into 1-D DCNN to achieve a higher level of intelligent fault diagnosis..The corresponding researches will focus on three issues. Firstly, the networks’ structure will be reconfigured to make 2-D DCNN more suitable for the identification of Hilbert time-frequency graph and to make 1-D DCNN to have a better performance on hydraulic vibrating signal analysis. Secondly, the training algorithm will be optimized based on the ideas of “self-adaptive label free learning” and “global optimization of parameters”. Thirdly, the results of training and layer-by-layer feature transformation will be deeply analyzed to have a thorough understanding of DCNN’s diagnosis mechanism, especially when used for hydraulic composite fault diagnosis..This project will explore a new approach for the intelligent fault diagnosis, to lay the foundations of model and algorithm of hydraulic fault diagnosis system with high intelligence and high reliability. It is also an innovative application of DCNN.

基于信号的液压复合故障是诊断领域的难点，基于在线监测信号的智能诊断对于使装备具备“自感知”能力，确保其安全、经济和可靠运行意义重大。.本课题追踪人工智能领域最新成果，基于液压复合故障特点，引入全新的、具有深度结构的智能诊断模型——深度卷积网络DCNN，尝试在结构重构和算法优化的基础上，实现二维DCNN对Hilbert时频图的智能识别；进而将信号特征提取、变换、融合与模式分类等环节集成在一维DCNN的深度结构中，实现更高智能水平的诊断。.研究的重点，一是面向诊断问题的DCNN重构，以及基于“无标签自适应学习”和“参数全局性调整”思想的算法优化，提升其包括诊断能力在内的综合性能；二是对训练结果和逐层特征变换结果展开分析，透彻了解DCNN的诊断机理。.本项目将为智能故障诊断探索全新的方法途径，为获得高智能、高可靠性的智能液压故障诊断系统奠定模型基础和算法基础，其同时也是对DCNN的创新性应用。

液压系统具有响应速度快、抗负载刚度大、功率密度大、低速稳定性好等优点，在无人化军用工程装备中应用广泛。随着无人装备数量的增多和服役时间的增长，能从累积的大量数据中提取出规律和特征，刻画数据的丰富内在信息，并能进行有效的分类和识别才是突破的关键。本深入开展了深度卷积网络DCNN（Deep Convolution Neural Network）在智能故障诊断中的应用研究，找到了一种能将海量的数据“内化”成知识，能自适应地进行特征提取、变换和融合的高智能算法，对于提升机械“自感知”的智能化水平，乃至促进人类对故障的认知，是非常有价值的。.在课题主要成果包括：①构建了以时域振动信号为输入的一维DCNN，和以时频图为输入的二维DCNN，实现了基于DCNN的特征提取和模式分类联合优化的智能诊断模型，并对模型结构、参数和算法进行了优化。②深入研究了卷积层的特征变换机理，包括：对卷积核进行可视化分析，对卷积核及信号的频率成分进行对比分析和相关性分析，发现了卷积核在训练中自适应学习到信号的频率特征这一机理；对同一层的不同卷积核进行相关分析和频率分析，发现了卷积核之间特征提取相互独立、协同优化这一机理。③基于上述研究成果，围绕基于振动信号的智能诊断问题的特点，提出了根据信号频率特征进行网络结构设计、参数配置和训练优化的新思路，并验证了该方法可较大幅度提升网络训练成功概率、收敛速度、模式识别正确率和模型运行效率。④创新性地提出了三次样条插值池化方法，改进了池化层，有效避免了传统池化方法导致的欠采样问题，提升了网络特征提取能力，降低了特征的维度。⑤研究了基于DCNN的自适应去噪问题及改进方法，并其与传统的小波阈值、奇异值分解等去噪方法进行了对比，发现了其在时变信号去噪中的独特优势，以及其对高、中、低频信号的过滤能力依赖于卷积层数量和卷积核尺寸这一机理，并由此给出了网络结构优化的参考性意见。