Abstract: Assembly lines are affected by stochastic disturbance events,including machine failures and machine maintenance,resulting in the nonlinear relationship between the input and output of a production line.Given that the reliability distribution of machines is not deterministic,the classical performance evaluation methods struggle with this unreliability.For an unreliable assembly line,a Generative Adversarial Network(GAN) model based on probability density estimation was constructed,and a data-driven performance evaluation method was proposed.Specifically,firstly,building upon the preprocessing of machine failure data,two characteristic parameters including machine failure interval time and machine failure repair time were extracted.The probability density estimation method was introduced and the underlying distribution of the machine failure data was explored,which yielded the probability density function of the characteristic parameters.In addition,the converged GAN model was obtained through the adversarial training between the generator and discriminator of the GAN model,generating the data of machine failure interval time and machine failure repair time conforming to the reality machine reliability model.The operation logic of the production line and its components including assembly line body,buffer line body and disassembly line body was analyzed from three dimensions including state changing,event triggering and performance statistics.The simulation procedure of an unreliable assembly line was given,which provided the operational logic for the performance evaluation of the assembly line.Finally,a case study was conducted on an assembly line of new energy vehicles from Company M.The results showed that there was a 1.6% deviation between the output of the proposed method and the actual yield,and the maximum was no more than 5%,which verified the effectiveness of the proposed method.

Key words: unreliable assembly lines, performance evaluation, data-driven method, probability density estimation, generative adversarial network

摘要： 装配线受机器故障、机器维护等随机扰动事件的影响,使得产线投入产出呈现出非线性变化关系。此外,实际中的机器可靠性并不服从确定分布的假设,经典的性能评估方法难以适用。面向不可靠装配线,构建了基于概率密度估计的生成对抗网络模型,提出了一种数据驱动的性能评估方法。在机器故障数据预处理的基础上,抽取出机器故障间隔时间、机器故障修复时间两个特征参数,采用概率密度估计方法,挖掘了机器故障数据的潜在分布规律,获得了特征参数的概率密度函数。通过生成器和判别器的对抗训练,得到了收敛的生成对抗网络模型,输出符合实际机器可靠性模型的机器故障间隔时间和机器故障修复时间数据。进一步地,从状态变更、事件触发、性能统计等3个维度,剖析了装配、缓冲和拆解等线体及产线的运行逻辑,给出了不可靠装配线仿真流程,为产线性能评估提供了运行逻辑。对M公司新能源汽车总装线进行了实例分析,结果表明所提方法与实际产量相差1.6%,且最大偏差不超过5%,从而验证了所提方法的有效性。

关键词: 不可靠装配线, 性能评估, 数据驱动方法, 概率密度估计, 生成对抗网络