A Dynamic Interactive Fusion Model for Extracting Fatigue Features Based on the Audiovisual Data Flow of Air Traffic Controllers

Abstract

Fatigue among air traffic controllers is a factor contributing to civil aviation crashes. Existing methods for extracting and fuzing fatigue features encounter two main challenges: (1) the low accuracy of traditional single-mode fatigue recognition methods, and (2) disregarding multimodal data correlations in traditional multimodal methods for feature concatenation and fusion. This paper proposes an interactive algorithm for the fusion and recognition of multimode fatigue features that combines multihead attention (MHA) and cross-attention (XATTN) which are based on an improved speech and facial fatigue recognition model. First, an improved conformer model which combines a convolutional module with a transformer encoder is proposed using the radiotelephony communication data of controllers by employing the filter bank method for extracting profound speech features. Second, facial data of controllers are processed via pointwise convolutions employing a stack of inverted residual layers, which facilitates the extraction of facial features. Third, speech and facial features are fuzed interactively by combining MHA and XATTN, which achieves high accuracy of recognizing the fatigue state of controllers working in complex operational environments. A series of experiments were conducted with audiovisual data sets collected from actual air traffic control (ATC) missions. Comparing with four competing methods for fuzing multimodal features, the experimental results indicate that the proposed method for fuzing multimode features achieved a recognition accuracy of 99.2%, which was 8.9% higher than that for a speech single-mode model and 0.4% higher than that for a facial single-mode model.