船舶机舱协作式模拟训练智能评估

Intelligent assessment for collaborative simulation training in ship engine room

基于人-机-环境系统工程理论,研究了船舶机舱模拟训练评估问题,建立了船舶的角色-任务-资源系统协作训练模型,提出了基于专家系统和机器学习的智能评估方法,并用于改进三维机舱协作训练系统;建立了船舶机舱模拟训练情景知识库、任务知识库和评估指标体系,结合专家经验和评估规范提取评估规则,采用遗传算法优化后的权重对数据进行多重模糊综合评判;针对实际的评估问题构造合适的机器学习网络结构,利用BP神经网络和深度学习算法具备的自学习优势,以简化运算步骤;将评估指标数据进行归一化处理后作为评估模型的输入数据,将评估结果作为目标数据,采用稀疏自动编码器对大量数据样本进行特征变换,深入学习样本特征并用于分类评估,经反复训练后得到较好的评估模型;对基于专家系统和机器学习的智能评估方法进行了对比分析。分析结果表明:经遗传算法优化的评估结果误差明显较小,平均绝对误差为0.761分,平均相对误差为0.983%,均方误差为0.938分,最大误差为2.263分,最小误差为0.248分;对于较简单任务的评估,基于机器学习评估的最大绝对误差为3.521分,最小绝对误差为0.304分,较好的深度学习评估网络的所有指标的评估误差均可小于1分。

The assessment for simulation training in ship engine room was studied based on the theory of man-machine-environment system engineering.The system model called role-missionresource for collaborative training was constructed,and the intelligent assessment methods based on expert system and machine learning were proposed to improve the three dimensional engine room collaborative training system.The scene knowledge base,mission knowledge base and evaluation index system of simulation training in ship engine room were established.The assessment rules were extracted according to the expert experience and the evaluation criteria.The data were evaluated by using multiple fuzzy comprehensive evaluation and the weights were optimized by genetic algorithm.The structures of machine learning network were constructed appropriately according to the practical assessment problem.The computing steps were simplified by using the advantage of self-learning of BP neural network and deep learning algorithms.The normalized evaluation index was used as the input of evaluation model,and the evaluation resultswere used as the target data.The features of mass sample data were transformed by using sparse autocoder.Sample features were learned deeply and used for classification evaluation,thus the better evaluation model was achieved after repeated training.A comparative analysis of the intelligent assessment methods based on expert system and machine learning was conducted.Analysis result indicates that after the optimization of genetic algorithm,the assessment result error is smaller apparently.The average absolute error is 0.761 point,the average relative error is 0.983%,the mean square error is 0.938 point,the maximum error is 2.263 point,and the minimum error is 0.248 point.For the evaluation of simpler missions,the maximum absolute error of assessment result based on machine learning is 3.521 point,and the minimum absolute error is 0.304 point.The assessment errors of all indexes in the better assessment network based on deep learning are less than 1point.