| Fatigue can lead to the decrease of alertness and work ability of ship OOW(officer on watch),which is a major cause of accidents such as grounding,collision,and landing on rocks.To avoid or reduce marine accidents caused by ship OOW fatigue and improve the safety of ship navigation,this thesis proposes a multi-feature fusion fatigue detection method based on smart wristband and video data and develops a ship OOW fatigue detection prototype system.The systemhas certain accuracy and robustness in fatigue detection and has potential application value onships.In this thesis,the main work is as follows:1.Aiming at the problems of difficult and intrusive physiological data collection in ship OOW fatigue detection,a fatigue detection method based on smart wristband data collection,data feature extraction and Logistic Regression is proposed.The smart wristbands are used to collect the heart rate and triaxial acceleration data of the ship OOW,and the original data are smoothed and denoised by Gaussian convolution and data conversion.The data features are extracted by calculating the variance or standard deviation in a fixed period.Based on the feature dataset of mixed heart rate and triaxial acceleration,the training and verification results of four machine learning methods are comprehensively compared,and the Logistic Regression method has the shortest processing time,and a good state classification effect index AUC(Area Under Curve)is 0.858,which could well classify the awake and fatigue states.2.To make up for the deficiency of only using smart wristband to detect fatigue,based on the deep learning method,in combination with fatigue state judgement method,the states of the eyes and the mouth in the video are identified to detect ship OOW fatigue.Firstly,based on the open-source dataset of personnel eyes and mouth state in various scenarios,the extended dataset is established by collecting and annotating the video data of navigation simulator and Changjiang 3 personnel operation.Then through MMDetection platform using Faster RCNN(Region Proposal Convolution Neural Network)algorithm for training and testing dataset.Comparing with MTCNN(Multi-task CNN)+ CNN and YOLOv5 algorithm,the overall average accuracy of this thesis’ s method to identify the states of eyes and mouth is higher and upto 0.885.Lastly,according to the PERCLOS(Percentage of Eyelid Closure over Time)and the single duration of eyelid closure and yawning,the fatigue can be determined comprehensively.After testing,the method of this thesis can better realize the real-time fatigue detection in the complex environment of the ship bridge,while the precision of fatigue detection in the simulator environment test is 90.91%,and the recall is 94.34%.3.For the practicability of ship OOW fatigue detection,a weighted decision mechanism is designed by integrating the fatigue detection method based on smart wristband data and video,and a robust multi-feature input fatigue detection system is established.The weight gain of the multi-feature fusion method is calculated by their respective fatigue detection accuracy,and the system control unit automatically adjusts its calculation method according to the feature input,which ensures the robustness of the system fatigue judgment when multiple features or single features are input.In addition,the ship OOW fatigue detection system is designed based on Python 3.9 and Py Qt 5 and tested on the navigation simulator.The results show that the system can meet the fatigue detection,alarm intervention and fatigue record under different features of data acquisition conditions. |
