撞击密度和子空间聚类在AGC伺服缸功能精度诊断中的应用

Functional-accuracy pattern recognition is playing an important role in the functional-accuracy maintenance of automatic gauge control (AGC) hydraulic cylinder. As the classification of functional-accuracy related patterns cover more set of equipments’ subtle working conditions with high similarities, which bring even tricky restrictions to the classifier. The project focuses on the pattern classification of functional-accuracy maintenance oriented equipment conditions. By integrating the genetic algorithm with the distributed neural network, an integrated feature selection optimiser will be designed to optimise the spaces-spanned feature selection. Based on the modified hits density model, the image based profiles regarding to the conditions being recognised will be constructed for visualisation and feature extraction purposes. As the precision of image based profile is constrained by the pixel size as well as the image intensity, which dominates by the majority of signal conditioned by the feature granularities. With each dimension of hits density image profile can be considered as a vector with linear relationship to the original when its being projected onto the principal component plane, the project will proposes a force analyse model, thereby finding the optimal feature granularity based on the resultant force of the vectors. In order to get accurate classification performance, the subspace clustering based classification technique will be taken into account. In terms of providing a coarse classification result to identify the general working conditions of the asset under examination, the backbone subspaces related to the working conditions aided by the prior knowledge will be provided. In terms of further recognising the even accurate statuses under various working situations, as well as to build the status development trajectory to the equipment, a generalized principal component analysis based methodology will be used in order to find the appropriate number and dimensionality of subspace, a homogeneous polynomial that passes through all the subspaces will be derived to present a clue to the AGC hydraulic cylinder condition development from health to failure. The project points out a force based feature granularity optimisation methodology as well as a functional-accuracy development trajectory constructing scheme based on subspace clustering, which may triggers a pioneering technique to the value interval optimisation in the intelligent modelling, as well as a new trend based strategy to evaluate the function-accuracy based degradation and failure.

AGC伺服缸的功能精度诊断对产品质量的管控具有重要意义。但由于功能精度识别涉及的状态及运行精度变量较多，且相似度较大，这就使仅利用面向故障管理理念的模式识别手段实现功能精度的识别面临较大挑战。为此，本研究面向具体的待识别功能精度状态，首先采用遗传算法与分布式神经网络融合，构造功能精度特征优化搜索器，实现特征的跨平面优化搜索与组合。进而基于扩展的撞击密度模型，构造各功能精度状态的可视化形态，鉴于值域约束对撞击密度模型的决定作用，将象素亮度作为向量空间中的力学矢量考虑，通过受力模型优化撞击密度图像的值域约束。后将表征各状态的撞击密度图像通过子空间聚类的方式进行分类识别，并通过建立子空间之间的连通，完成性能演化趋势的构建并实现功能精度的诊断。本研究所提出的基于向量空间力学分析实现值域优化和基于子空间连通构造设备性能演化趋势的方法，可有效解决高相似度状态的特征提取和功能精度状态的识别问题。

AGC伺服缸的功能精度诊断对产品质量的管控具有重要意义。但由于功能精度识别涉及的状态及运行精度变量较多，且相似度较大，这就使仅利用面向故障管理理念的模式识别手段实现功能精度的识别面临较大挑战。为此，本项目以轧钢机 AGC 控制系统的压下伺服油缸在正常、振动状态下获取的声发射信号为具体数据来源模拟重载、往复装备的性能演化。首先采用当下广泛应用的主成分分析法对不同状态的声发射信号进行特征降维分析并实现不同状态的可视化以及分类识别。其次，研究基于扩展的撞击密度热图实现精度状态可视化识别, 后将表征各状态的撞击密度图像通过子空间聚类的方式进行功能精度的诊断。最后，采用小波分析法作为辅助分析和实验结果的交互验证，完成声发射信号的能量特征值提取进而实现不同状态的分类识别。结果显示，本项目所采用的方法均能一致有效的区分出具体的待识别功能精度状态。本项目在经典智能诊断流程的统一框架下兼顾信号特征的可视化及功能精度状态的精准辨识，为解决重载往复装备的功能精度诊断问题提供一种全新的基础研究思路，具有一定的理论意义和实际指导价值。