设备故障评估新指标及基于ARMA的预测系统

Innovative metrics for equipment failure evaluation and prediction system based on ARMA model

设备故障停机时间受生产调度的影响较大,不能真实反映设备的自身性能,且具有很强的随机性和波动性,不适于直接用来进行自回归移动平均(auto-regressive moving average,ARMA)建模。针对此问题,提出一种设备故障评估指标——设备不可用度,将设备故障停机时间转换为设备不可用度,通过异常点替代和数据平稳化等两种数据预处理,建立零均值平稳随机序列进行ARMA建模,并把预测结果转换为设备在一定时间内的故障发生概率。在某半导体芯片封装测试工厂的试验结果表明该方法能以70%的精度预测设备状态,在一个班(12 h)里设备不可用度平均降低2.62%,设备故障停机时间平均减少14.8 min。

The equipment failure down time（or the failure rate） in a certain period of time（such as 12 hours）,which is greatly influenced by the production planning and scheduling in semiconductor manufacturing factory,can not reflect the true equipment performance.Moreover,the data series of down time is not suitable for being directly used for auto-regressive moving average（ARMA） modeling because it has very strong randomness and undulatory property.An innovative metrics for equipment failure evaluation,named equipment unavailability（EU）,is proposed according to this problem.When building an ARMA model,the equipment failure down time is firstly transformed to EU.Then the data is converted into stationary random sequence by outliers＇ replacing,and thirdly the trend term of data is removed by using an improved moving average algorithm.So the zero-mean stationary random sequence is available to build the ARMA model.The forecasting result is transformed into equipment failure probability in a certain period of time at last.The process of data pretreatment,modeling,forecasting and result transforming is realized to a software application system by using VS.NET.The application in a chipset assembly and test factory shows that the method can predict machine status with the accuracy of 70%,the equipment downtime is average reduced by 14.8 minutes and the machine unavailability is average reduced by 2.62% in one shift（12 hours）.