Akademik Veri Yönetim Sistemi
INTERNATIONAL TECHNOLOGICAL SCIENCES AND DESIGN SYMPOSIUM (ITESDES 2022), Giresun, Türkiye, 2 - 05 Haziran 2022, ss.10-11
It is obvious that the quality expectations of the telecom services have been increasing in recent decades. In addition, a continual future growth in the number of the users is foreseen. Service interruption in telecommunication can lead to disruptions in individual and public usage. Therefore, fault management is becoming increasingly important. Faults are responsible for making it difficult for a system to maintain its normal functioning and they manifest themselves in deviations from the planned operational routine. In common applications, every error that occurs is logged as system history and the associated alarm texts specify where the malfunctioning is. Fault detection approaches try to identify such misbehavior in running systems. Real time fault prediction aims to predict emerging failures before they occur at runtime. Such approaches encompass not only the classification of the misbehavior, but also the prediction of time to the faults that may occur in a future period. The data used in the study were obtained from 531 different network devices. Alarm data per device are sorted in ascending order according to the occurrence time. A time series is defined by processing the time difference between consecutive alarm data. Therefore, the fault prediction is transformed into a time series prediction problem. The events which occur within a lower period than a predetermined threshold are filtered out from the extracted time series. Similarly, the fault types are encoded as distinctive IDs and converted into time series. Hence, two different LSTM models are created per device: one for determining whether a fault will emerge in the near future, and one for classification of the imminent fault according to predefined categories. Since each device has unique failure time model, independent LSTM networks are created for each device. The parameters of each LSTM structure are optimized with the Particle Swarm Optimization approach.