健康流形空间中机电部件性能衰退预测的典型问题研究

To address the current three typical issues existing in the performance degradation prediction of electromechanical components, i.e., the adaptability of performance degradation prediction under variable operating conditions, the effectiveness of performance degradation prediction with only truncated dataset available, and the accuracy of long-term performance degradation prediction, this proposal proposes four promising research points for performance degradation prediction of electromechanical components as follows: (1) feature extraction and representation of performance degradation within the high-dimensional manifold space (i.e., health manifold space, HMS); (2) performance degradation prediction within the HMS under variable operating conditions; (3) performance degradation prediction within the HMS with only truncated dataset available; and (4) long-term performance degradation prediction within the HMS. .With the premise of “Presence of objective evolutionary laws of a specific object”, we aim to effectively combine the existing theories and methodologies of prognostics with the modern differential geometry and physics, as well as the ideas such as ’multidisciplinary features reflect the objective evolutionary laws of a specific monitored object comprehensively’; ‘multiple parameters information fusion may highlight the objective evolution laws within the HMS (high-dimensional geometrical manifold space)’; and ‘traditional prognostic methods may be optimized and improved to achieve a high accuracy’. The proposed methodologies will be specifically applied to the typical components of rotating machineries, closed-loop control systems, and storage batteries for verification and validation. Thus, the project is expected to explore and reveal the accurate performance evolution laws as comprehensively as possible on the basis of the existing available data, and further enhancing the capabilities to realize an integrated solution to the above three typical issues in performance degradation prediction of electromechanical components.

针对当前机电部件性能衰退预测存在的“变工况条件下性能衰退预测的适应性”、“截断数据下性能衰退预测的有效性”、“中长期性能衰退预测的准确性”三类典型问题，基于“事物客观演化规律存在”的前提，本申请在现有性能衰退预测理论和方法研究基础上，结合现代微分几何、物理学理论与“系统多学科特性能从不同侧面反映被监测对象的客观演化规律、健康流形空间（高维几何空间）中的多特性参数信息融合可能凸显系统的客观演化规律、优化和改进传统的性能衰退预测方法以提高预测精度”思想，在高维流形空间（健康流形空间）中，依次开展：（一）机电部件性能衰退特征参数提取与表征、（二）机电部件变工况条件下性能衰退预测、（三）机电部件不完全/截断数据条件下性能衰退预测、（四）机电部件中长期性能衰退预测等研究。从而在已有数据条件下，尽可能客观全面地揭示性能退化规律，力争在统一的几何框架下，综合解决机电部件性能衰退预测的上述三个典型问题。

机电系统日趋复杂、运行环境多样，由此带来了故障的多发性、致命性、随机性、交联性等问题，导致了复杂系统维护面临着更加严峻的挑战。如何根据系统的实际健康状态合理地制定维护计划，防止设备和产品因故障而失效，已成为降低运作成本、提高运行安全、生产效率和市场竞争力的重要手段。然而，作为解决上述问题的核心关键技术——性能衰退预测，由于机电系统及部件演化规律认识的局限性、运行环境的多变性、预测方法使用条件的依赖性等因素，使得机电系统性能衰退预测仍存在诸多问题亟待解决。.本研究针对当前机电部件性能衰退预测存在的“变工况条件下性能衰退预测的适应性”、“截断数据下性能衰退预测的有效性”、“中长期性能衰退预测的准确性”等三类问题，借助于流形学习理论、深度学习、随机深度认知、几何形态学、集成学习技术，以及其他智能化技术方法，深入研究了：1）健康流形空间中，健康状态及演化过程的描述与几何度量技术研究；2）变工况条件下健康流形空间中性能衰退预测技术研究；3）完全截断数据条件下健康流形空间中性能衰退预测技术研究；4）流形空间下剩余寿命预测技术（中长期预测）研究。.本研究结合一系列相关典型机电产品对各项研究内容及关键技术进行了测试验证，结果表明：基于上述研究工作，实现了多维度故障/衰退特征自提取、特征参数表征、流形空间的构建、流形空间中衰退规律的随机深度认知（仅利用50%衰退数据，可实现高于97%的预测精度，且方差控制在0.518%范围内）、流形空间中截断数据的有效利用（流形空间中截断数据的生存概率寿命预测整体精度不低于92%）；结合几何图形学和智能化算法，实现了流形空间中的中长期预测（性能退化预测精度96%以上），综合多预测模型的优势，开展了流形空间中集成学习预测技术研究，实现了寿命预测。