Using Expert Knowledge to Generate Data for Broadband Line Prognostics Under Limited Failure Data Availability ?

摘要

Due to exposure to the driving rain, water ingress can cause faults in electrical joints, junctions and distribution points in broadband lines. Over time, faulting behaviour may grow in magnitude eroding the electrical capability of these lines causing degradation of broadband service. Developing effective data-driven models for broadband line prognostics remains a challenge due to the limited failure data availability in the telecommunications industry. In order to address this problem, we present a technique for generating failure data that realistically reflect the behaviour of degrading broadband lines. To this end, we use the conditional generative adversarial network and more importantly, we control and direct the failure data generation process using expert knowledge on the water ingress failure cause. The proposed technique is evaluated using a real-world case study involving the time-to-failure prediction of two types of broadband lines in a south-west city in England. The prognostics performance is measured using the Kappa statistic and F-score. Benchmark performance is obtained using Random Oversampling, Synthetic Minority Oversampling and Adaptive Synthesis which can be used to oversample failure data by duplicating existing failure data or randomly generating data. Among these techniques, Random Oversampling achieved the best prognostics performance. It is shown that the proposed technique outperforms Random Oversampling technique by a large margin. More specifically, it increased the prognostics performance by 33% (Kappa statistic) and 25% (F-score) for Asymmetric Digital Subscriber Lines, and 17% (Kappa statistic) and 13% (F-score) for Very High Bitrate Digital Subscriber Lines compared to the Random Oversampling technique.