基于灰色系统理论和神经网络的呼吸机故障预测模型的建立与研究

Establishment and study on ventilator failure prediction model based on grey system theory and neural network

目的针对医疗设备故障有效数据样本少、模型预测精度低等问题,结合使用灰色系统理论与神经网络理论,建立两种基于灰色神经网络的呼吸机故障预测组合模型。方法介绍了GM(1,1)预测模型理论基础,根据最新信息优先原则结合重构背景值方法及建立GM优化的DI-GM。利用Levenberg Marquardt反向传播神经网络模型理论,建立并联式预测模型(PPM)和串联式预测模型(SPM)。使用福禄克Fluke117C数字万用表对南通市第一人民医院现行使用的某型号呼吸机马克涡轮供电电压进行采样,采样间隔5 min,采样时长120 min;同时收集并整理近年来呼吸机平均故障率作为研究数据集对4种模型进行建模、预测,并使用拟合度、均方根误差、自相关检验、残差序列异方差等对4种模型拟合和预测结果进行检验和评价分析。结果针对马克涡轮供电电压预测结果,PPM与SPM预测值均方根误差分别为2.230%、1.579%,平均绝对误差分别为0.323、0.225次,拟合度分别为0.793和0.890;针对呼吸机故障率的预测结果,PPM与SPM第9年、第10年预测值与实际值偏差分别为0.210、0.246台次和0.093、0.072台次;SPM对呼吸机马克涡轮供电电压与故障率预测效果更佳。结论灰色神经网络的呼吸机故障预测组合模型对马克涡轮供电电压与呼吸机故障率的预测精度优于单一预测模型,可为呼吸机及其关键性部件的预防性维护和更换提供参考,值得进一步探讨。

Objective To establish 2 kinds of combined models for ventilator failure prediction based on grey system theory and neural network theory,and solve the problem of small number of valid data samples for medical equipment failures and low accuracy of model prediction.Methods The theoretical basis of GM(1,1)prediction model was introduced,and the background value reconstruction method and DI-GM optimized by GM were established according to principle of the latest information priority.Based on Levenberg Marquardt back propagation neural network model theory,the parallel prediction model(PPM)and serial prediction model(SPM)were established.The Fluke117C digital multimeter was used to sample power supply voltage of Mark turbine of ventilator,the sampling interval was 5 minutes and sampling time was 120 minutes.The mean failure rate of ventilator in recent years was collected and sorted out as research data set to model and predict 4 models.The fitting,root mean square error,autocorrelation test and residual sequence heteroscedasticity were used to test and evaluate fitting and prediction results of 4 models.Results For prediction results of Mark turbine power supply voltage,the root mean square errors of predicted values of PPM and SPM were 2.230%and 1.579%,respectively,the mean absolute errors were 0.323 and 0.225 times,respectively,and fitting was 0.793 and 0.890,respectively.For prediction results of ventilator failure rate,the deviations between predicted value and actual value of PPM and SPM in the 9th and 10th years were 0.210,0.246 times and 0.093,0.072 times,respectively.The SPM was more effective in predicting power supply voltage and failure rate of ventilator turbine.Conclusion It is demonstrated that the combined model of ventilator failure prediction based on gray neural network is superior to single prediction model in predicting power supply voltage of Mark turbine and failure rate of ventilator,which could provide reference for preventive maintenance and replacement of ventilators and their key components,and is worthy of further discussion.