基于R语言BP神经网络瓦里安NovalisTx直线加速器MLC系统故障预测模型研究

摘要

Objective To construct and investigate the multi-leaf collimator (MLC) fault prediction model of Varian NovalisTx medical linear accelerator based on BP neural network.Methods The MLC fault data applied in clinical trial for 18 months were collected and analyzed.The total use time of accelerator,the quantity of patients per month,average daily working hours of accelerator,volume of RapidArc plans and time interval between accelerator maintenance were used as the input factors and the prediction of MLC fault frequency was considered as the output result.The BP neural network model of MLC fault prediction was realized by AMORE package of R language and the simulation results were validated.Results The model contained 3 layers of network to realize the input-output switch.There were 5 nodes in the input layer,13 nodes in the hide layer and 1 node in the output layer,respectively.The transfer function from the input layer to the hide layer selected the tansig function and purelin function was used from the hide layer to the output layer.The maximum time of training was pre-set as 150 in the designed model.Actually,111 times of training were performed.The pre-set error was 3％ and the actual error was 2.7％,which indicated good convergence.The simulation results of MLC fault applied in clinical trial for 18 months were similar to the actual data.Conclusions The BP neural network model realized by R language of MLC fault prediction can describe the mapping relationship between fault factors and fault frequency,which provides references for the understanding of accelerator fault and management of spare parts inventory.%目的 构建并研究瓦里安NovalisTx直线加速器MLC系统故障预测BP神经网络模型.方法 取加速器临床使用18个月MLC系统故障统计数据为研究对象,以加速器使用总时间、月治疗患者数量、日均开机工作时间、RapidArc计划数量及加速器保养后时间间隔为输入故障因素,以故障频次预测为输出结果,采用R语言AMORE包构建MLC系统故障预测BP神经网络模型并对其进行仿真验证.结果 模型采用3层网络实现输入输出转换,其输入层5个节点、隐层13个节点、输出层1个节点;输入层至隐层、隐层至输出层分别选用tansig、purelin传递函数;模型设定最大训练学习次数150次,实际使用111次,设定误差3％,实际误差2.7％,表明其收敛较好.该模型对18个月临床故障数据仿真验证结果表明预测数据与实际数据较为接近.结论 基于R语言BP神经网络故障预测模型实现了MLC系统故障因素与故障频次间映射关系描述,可为设备故障规律了解和备件库存管理提供参考.