基于稀疏优化的大口径光学元件超精密磨削机床状态监测

In order to improving the reliability and precision maintenance of ultra-precision large-aperture optics grinding machine, monitoring techniques for grinding machine healthy and grinding process are studied. The major contents of the project are as follows. Firstly, interior, built-in and built-out sensors are overall considered to construct monitoring system. In order to simplify the system, an optimization model with regularization is built to decide sensor location. Secondly, state signals are abstracted by a deep neural network to high-level features, which are employed to monitoring grinding machine conditions. The training mode and test criterion of the network are studied considering unbalance samples of grinding machine running. Thirdly, feature learning methods are put forward aiming to long time series. Conditions of grinding interference are sparsely expressed by the learned features. Labeled and un-labeled samples are used to build semi-supervised regression model for predicting grinding wheel residual life. In conclusion, sparsity and optimization are the keys of the project. Concise system hardware, and sparse features learned from interior nature structure of origin data, and decisions by optimization model instead of manual intervene can ensure high reliability monitoring. The research is helpful to improve the intelligence of grinding machine and provide generalized feature learning methods for machining process monitoring.

为提升大口径光学非球面精密与超精密磨削机床的可靠性和精度保持性，开展磨床健康监测与磨削性能优化的研究工作。采用内部、内置和外置传感器相结合的监测方案，结合整机有限元仿真分析提出测点分布的正则稀疏优化模型，实现最小数量传感器监测系统；建立运行状态信号深度神经网络模型，基于自动编码实现多维原始信号的逐层抽象压缩，提出针对不平衡样本的模型训练与检验准则，结合磨床健康运行大数据实现模型进化。实现磨床健康监测；针对磨削干涉过程信号，聚焦原始信号本质结构，提出特征自学习方法实现复杂特征稀疏化表达，建立半监督回归模型，实现砂轮剩余寿命评价。项目拟通过监测硬件、数据构成、特征表达等方面的精简与稀疏化，完全借助优化模型避免人工干预监测的盲目性和不确定性，基于磨床监测大数据提升大口径光学非球面磨削机床智能化。课题研究成果将为加工状态监测领域的多维快变长序列信号特征自学习研究提供系统的理论指导。

激光核聚变装置等国家重大工程及国防尖端科技对高精度大口径光学非球面的需求量巨大，稳定质量高效率加工成为制约其发展的瓶颈问题。为此，针对大口径光学元件磨削阶段这一主要的工艺过程，开展了装备监测系统与监测方法研究，项目搭建了高性能通用监测系统和低成本嵌入式监测系统，可以适应不同监测成本需求，建立以外部传感和内部数据相结合的监测模式，实现了监测系统和数控系统的双向数据传输，并基于OPC-UA通讯协议将机床及其本地监测系统包装成统一的对外接口，便于机联网拓展。另外，通过对比研究，有选择地在监测系统集成了多种机器学习和深度学习算法，可以用于机床运行异常识别和加工过程监测。决策方法方面，研究基于线性判别分析的高维声发射特征稀疏表征，以此为基础提出了具有普适性和高鲁棒性的监测策略，实现了砂轮性能退化在线精准评估，研究并实现了针对一维数据的循环神经网络和卷积神经网络两种深度学习框架，深入探讨了模型参数选择及中间层特征可解读性问题，明确了深度学习对于砂轮微弱退化的识别能力。