基于深度信念网络的轮机实操智能评估方法

Intelligent assessment method of marine engine-room operation based on deep belief networks

结合我国推进船员实操评估电子化、智能化的任务,提出基于深度信念网络的轮机实操智能评估方法.针对轮机实操评估的特点,给出了确定网络层次结构的具体方法.在提取大量的实操数据作为训练数据的基础上,通过逐层贪婪训练算法对限制玻尔兹曼机逐层训练,最后经BP算法对网络微调后形成评估模型.在仿真实验中,分别对带回归模块的深度自编码网络、BP神经网络和该模型的预测效果进行对比验证.结果表明,该模型评估效果比较客观、公正,评估误差最小,且避免了多层神经网络过早陷入局部最优的问题.

The intelligent assessment method for marine engine-room operation based on deep belief networks was proposed according to the task of promoting the electronic and intelligent assessment for seafarers in China. According to the characteristics of assessment for practical engine-room operation,the method for determining the hierarchical network structure was provided. Based on the large amounts of extracting practical operation data as the training data,the restricted Boltzmann machines were trained by greedy training algorithm. Finally,the assessment model was generated by using BP algorithm for network fine-tuning. In the simulation experiments,the prediction results of the deep auto-encoder network,BP neural network and the proposed model were verified comparably. Results show that the proposed model is objective and impartial,and the error is the minimum,and the problem that multilayer neural networks fall into local optimum is avoided.