迁移学习辅助基于机器学习的传输质量评估

Transfer Learning Assisted Transmission Quality Evaluation based on Machine Learning

当光纤链路的数据分布发生变化时,机器学习评估链路传输质量需要重新搜集数据并重新训练,这个过程是耗时的、复杂的。迁移学习直接将以前学到的知识应用到现在的任务,需要更少的数据。因此,文章提出在具有相关性的光通信系统中,使用两种迁移学习方式辅助基于机器学习的多分类器,仿真结果表明,迁移学习结合微调技术的机器学习多分类器比直接迁移的机器学习多分类器多分类指标分数提高0.25以上,减少了样本不均衡的影响,每个类别都具有高性能,证实迁移学习结合微调技术的机器学习多分类器能够减小数据集,降低了搜集数据集的成本,提高了光纤链路传输质量评估问题的效率。

When the data distribution of optical fiber link changes, machine learning can be used to evaluate the transmission quality of link, which needs to recollect data and retrain. This process is time-consuming and complex. Transfer learning directly applies the previously learned knowledge to the current task, which requires less data. Therefore, it is proposed to use two transfer learning methods to realize multi classifier based on machine learning in the correlated optical communication system. The simulation results show that the multi classifier of machine learning combined with transfer learning and fine-tuning technology improves the multi classification index score by more than 0.25 compared with the direct transfer machine learning multi classifier. It is also shown that the impact of sample imbalance is reduced, and each category has high performance. It is proved that the machine learning multi classifier combined with transfer learning and fine-tuning technology can reduce the size of data set, reduce the cost of collecting data set, and improve the efficiency of transmission quality evaluation process in optical fiber link.