Fault prediction in the telephone access loop using a neural network

Abstract

Fault prediction is a vital component of proactive maintenance in the telephone access loop. Indication of line problems are found by regularly measuring line parameters such as resistance and voltage. The suggested technique uses preprocessed line measurements as input to a neural network. Line repair records are used as fault indications when creating the training and test data set for the neural network. The collected data is found to be inconsistent and noisy. This limits the achievable correctness of the results. The method uses multiple measurements of the same line. Using hidden layers in the neural network was not found to improve the results significantly. The results show that around 25 – 50 % of the predicted faults are later reported by the customers. Unfortunately, few faults can be predicted with a high correctness.