基于优化概率神经网络的飞机结冰情况识别

Aircraft icing classification using optimized probabilistic neural networks

针对飞机飞行过程结冰情况识别和预警需求,利用概率神经网络开展直接基于飞行状态的飞机结冰情况识别研究。首先利用双水獭结冰研究飞机的结冰数据,建立结冰飞行动力学模型,生成飞机在干净外形、中度和重度机翼结冰以及中度和重度尾翼结冰五种情况下的大量飞行仿真数据,作为分类神经网络训练和性能评估的基础;再基于大量飞行仿真数据分析各飞行状态变量受结冰影响的显著程度,选择受结冰影响较大的6个状态变量作为识别网络的输入;提出了一种传播参数优化策略对基于概率神经网络的结冰分类精度进行优化,在此基础上对6个状态输入的结冰分类结果展开对比分析,发现基于这6个受结冰影响较大的状态变量构建的优化概率神经网络,对于训练和评估数据均具有较高的分类精度,对于训练数据分类精度均可达到100%,对于评估数据,分类精度最高的网络正确率可达99.1%,效果最差的网络分类精度也超过85%;快变量对于评估数据分类精度较高,但对泛化性评估数据的分类精度偏低,慢变量则正好相反;基于快变量攻角α和慢变量位移xe的神经网络分类效果相对较好,若不考虑风场扰动的变化,基于攻角α的分类网络的最大误报概率不超过1.4%,若综合考虑风场扰动变化的情况,则采用基于xe的分类网络效果更好,对于五种结冰情况最大误报概率整体不超过10%。最后利用两种支持向量机方法与优化概率神经网络的分类精度进行对比,进一步验证该方法的性能。

This paper issues an aircraft icing classification problem by directly using the flight state variables and probabilistic neural networks(PNN) for meeting the needs of classifying and alarming icing situation during aircraft flight. A dynamic model for inflight icing is presented based on the icing experiment research data of the NASA Twin Otter airplane. The dynamic model simulates five situations, i.e., clean,moderate/severe wing icing, moderate/severe tailplane icing, to generate abundant flight data for the networks’ training and assessment. The influence of icing on different flight state variables is quantitatively analyzed, and the six variables significantly affected by icing are chosen for building the neural networks. A scheme for optimizing the network propagation parameter is presented to improve the classification accuracy, then the performance of the six state variables-based classification networks is analyzed and compared. The results indicate that PNNs built by the six state variables have relatively high accuracy for both the training and validation data. For the training data, the classification accuracy all reaches 100%, while for the validation data,the highest accuracy among all six nets is 99.1%, and the lowest accuracy is still above 85%. Fast variables have a higher accuracy for the validation data than slow variables, but a lower accuracy for the generalization assessment data, whereas the slow variables conduct oppositely. The networks built by the fast variable, angle of attack α and the slow variable, displacement xe perform relatively better than the others. Without the wind disturbance, the most considerable false alarming probability of α-based networks is below 1.4%. With the wind disturbance, the xe-based networks are more preferable, as the overall false alarming probability for all the five icing situations is below 10%. Finally, the performance of the optimized PNN is further validated by comparing the classification accuracy with that of two support vector machine methods.