基于泛函逼近的旋转机械性能衰退预测方法研究

It is well known that the time accumulation effect severely influences the performance deterioration of the complex rotary machinery in practical engineering. To avoid the fault of the equipment caused by the performance deterioration, it is very important to utilize the abundantly accumulated equipment performance historical data to predict its performance deterioration trend. However, the existing performance prediction methods can not reflect the problem of the time accumulation effect actually implied in the performance data. To solve this problem, this project will lay a strong emphasis on the study of the time series prediction method based on the theory of function approximation, and then an equipment performance deterioration prediction model is developed based on the proposed time series prediction method. The input of the proposed equipment performance deterioration prediction model can be time-varying continuous functions. Thus, the calculation of the proposed equipment performance deterioration prediction model based on the functional regression analysis method is very complex, and it is difficult to adopt the proposed equipment performance deterioration prediction model to practical engineering with this limitation. It has been proved that the process neural network can be approximate any continuous functional at any accuracy degree. To simplify the calculation complexity, the process neural network is utilized to calculate the proposed rotary machinery performance deterioration prediction model, and a novel learning algorithm based on the convolution computation for the process neural network is developed. Finally the research findings of this project are utilized to predict the performance deterioration of some helicopter transmission system, and the correctness of the theoretical predicted results is verified through an accelerated degradation test of the above transmission system, and thus the research objective is to provide some theory and technique supports to the complex rotary machinery for the maintenance support.

工程实践表明，复杂旋转机械的性能衰退主要是由时间累积效应造成的。为避免由性能衰退引起的装备故障，利用积累的装备性能历史数据预测其性能衰退的趋势，具有重要的意义。针对现有性能预测方法难以反映性能数据中实际隐含的时间累积效应的问题，本课题将在深入研究泛函逼近理论的基础上，提出一种基于泛函逼近的时间序列预测方法，并在该方法的基础上建立性能衰退预测模型。基于泛函逼近的性能衰退预测模型的输入是与时间相关的连续函数，因而其基于泛函回归分析的模型求解非常复杂，工程实用性不强，而过程神经网络能够以任意精度逼近任意连续泛函，因此本课题拟以过程神经网络为逼近算子，研究新的基于卷积运算的过程神经网络学习算法，对建立的旋转机械性能衰退预测模型进行求解。将课题的理论研究成果应用于直升机传动系统的性能衰退预测中，并通过该传动系统加速退化试验验证理论预测结果的正确性，以期对复杂旋转机械的维修保障提供理论依据和技术支持。

针对旋转机械维修保障和维修理论的发展需求，对旋转机械性能衰退预测理论展开研究。将函数型数据分析方法引入到旋转机械的振动数据预处理，并应用卷积和神经网络相结合的人工智能方法提高旋转机械的性能预测精度，建立了一套以过程神经网络作为逼近算子的旋转机械性能预测方法，该方法能够有效克服现有旋转机械性能预测方法中存在的难以处理高频性能数据、间隔采样缺失而导致数据隐含信息丢失及预测精度低等问题。课题的研究丰富了学科在机械设备状态监测与故障诊断的研究方向，为企业的设备维修和健康管理提供理论依据和技术支持，完善了机械装备维修的理论，具有重要的应用价值。