基于改进D-S证据理论的航空发动机转子故障决策融合诊断研究

针对单一传感器的测量信息难以准确、全面地反映航空发动机转子、轴承和齿轮的工作状况，进而造成振动故障诊断难度大的问题，提出安装多个振动传感器组成传感器网络，建立基于多传感器信息的发动机转子故障决策融合诊断系统。由于多传感器系统不可避免地会存在各传感器信息不一致、信息冲突的情形，因此针对该融合诊断系统的信号测量、信息预处理、特征提取、故障诊断及决策融合5个环节，重点研究了决策融合环节的Dempster-Shafer（D-S）证据决策融合方法存在的冲突证据融合失效问题。通过分析原因，从避免“一票否决”现象和证据加权平均两个方面进行改进，提出了改进D-S证据融合方法，并应用于航空发动机转子的模拟故障决策融合诊断中。结果表明基于D-S证据理论对3个传感器的单一诊断结果进行决策融合，能得到比任一单个传感器更准确、可靠的结果；而改进D-S证据融合方法由于能在一定程度上克服冲突证据融合带来的失效问题，且能同时兼顾处理好非冲突证据的融合，故其对于证据冲突和非冲突情形都取得了较好的融合效果，因此总的分类正确率要高于常规D-S算法和PCR5算法。

Research of Decision Fusion Diagnosis of Aero-engine Rotor Fault Based on Improved D-S Theory

As the information measured by a single sensor can not reflect the working status of aero-engine rotors, bearings and gears accurately and completely, it is difficult to make vibration fault diagnosis based on it. In an attempt to solve this problem, several sensors are used to establish a sensor network. Thus, an aero-engine rotor decision fusion diagnosis based on multi-sensor information is proposed in this paper. However, information inconformity and conflict of different sensors is inevitable in a multi-sensor system, which is composed of five segments: signal measurement, signal pretreatment, feature extraction, fault diagnosis and decision fusion. A focused study is conducted on conflict evidence fusion failing of the Dempster-Shafer (D-S) evidence decision fusion method in the decision fusion segment. Through a cause analysis, improvement is proposed to avoid the "one ticket veto" phenomenon and ameliorate the weighted average evidence process. Based on the improvement, an improved D-S evidence fusion method is put forward, which is applied to the decision fusion diagnosis of a simulated aero-engine rotor vibration fault. The result shows that, when the diagnosis result of all sensors is conducted to realize decision fusion using the D-S evidence theory, the fusion result is more accurate and reliable than any single sensor result. The improved D-S evidence fusion method can overcome the failing brought by conflict evidence fusion. Consequently, a better fusion result can be achieved in both the evidence conflict and non-conflict situation. And holistic classification accuracy is higher than general D-S algorithm and PCR5 algorithm.