基于聚焦注意力机制的飞行动作识别方法

Flight Action Recognition Method Based on Focus Attention Mechanism

飞行动作识别对飞行训练质量评估、提升飞行员的驾驶技术具有重要意义。在飞行动作序列数据中,某时刻飞行状态数据及该状态对应的某些飞参数据对飞行动作识别有非常重要的贡献,但传统的注意力机制只关注上层状态特征的贡献值,而忽略了下层特征的影响。为有效提取飞参数据的关键特征表示,提出聚焦注意力(FA)机制,扩展了传统注意力机制,进一步聚焦注意力,学习其下层特征对上层的贡献。同时,本研究将FA机制扩展到BiLSTM网络中,提出FA-BiLSTM网络模型,该模型不仅重点关注飞行动作序列中关键时刻的飞行状态数据,还可学习该飞行状态中关键的飞参数据。实验表明,该方法有效提升了飞行动作识别的准确率,加权平均准确率达到了94%。

Flight action recognition is of great significance to evaluate the quality of flight training and improve the pilot's driving skills.In the flight action sequence data,the flight state data at a certain time and some flight parameter data corresponding to the state make a very important contribution to flight action recognition,but the traditional attention mechanism only focuses on the contribution value of the upper-layer state features and ignores the influence of the lower-layer features.In order to effectively extract the key feature representation of flight parameter data,the Focus Attention(FA)mechanism is proposed,which extends the traditional attention mechanism,further focuses attention and learns the contribution of its lower-layer features to the upper layer.At the same time,the FA mechanism is extended to the BiLSTM network and the FA-BiLSTM network model is proposed.The model not only focuses on the flight state data at the critical time in the flight action sequence,but also can learn the key flight parameter data in the flight state.Experiments show that this method effectively improves the accuracy of flight action recognition,and the weighted average accuracy reaches 94%.