基于代价敏感主动学习的氧化铝蒸发过程故障检测(英文)

Fault detection using modified cost-sensitive active learning for alumina evaporation process

针对氧化铝蒸发过程故障检测中标注者不切实际的假设和控制参数难以确定问题,提出改进的代价敏感主动学习方法。给出了代价敏感主动学习形式化描述和放松了标注者不切实际的假设。为了提高分类精度和减少标注代价,该方法结合粒子群优化和代价敏感主动学习。利用连续的粒子群优化代价敏感主动学习的控制参数,该参数用于最大化未标注样本的信息度和最小化标注代价。将所提出的方法应用于氧化铝蒸发过程故障检测,实验结果表明,该方法能正确地选择控制参数,有效地减少了误分类代价和标注代价,提高了故障检测率。

Aiming at the unrealistic assumptions of oracles and the difficult determination of control parameter in fault detection for alumina evaporation process,a modified cost-sensitive active learning（MCAL）method is proposed.The formal description of MCAL is given and unrealistic assumptions of oracles are relaxed.MCAL hybridizes particle swarm optimization（PSO）and cost-sensitive active learning（CAL）to improve the classification accuracy with multi-oracles and reduce the cost of labeling instances sampled.This optimization mechanism employs the continuous-valued PSO to optimize the control parameter to maximize the value of information of instance and minimize the cost of oracle.MACL is applied on the benchmark of alumina evaporation process for fault detection.Experimental results show that MCAL correctly selects the control parameter,obtains low misclassification cost,reduces labeling costs and increases the rate of fault detection.