Abstract
The pressure to achieve greater data rates from limited radio spectrum while reducing latency in 5G and soon 6G has resulted in the introduction of new and complex features into the network, such as mmWave and terahertz transmission, RAN slicing and reconfigurable intelligent surfaces. At the same time, to provide the necessary performance for vehicular networks and to support large numbers of users in urban areas, it is becoming necessary to reduce cell sizes and dramatically increase their numbers. Given all this, it seems inevitable that there will be a corresponding increase in the likelihood of parameter misconfiguration.
It has been recognised that these trends will require additional automated assistance to network operations centre staff. As a result many studies have now investigated how to apply machine learning techniques to this problem. Many of these approaches, however, are not transparent to their users, and can consume large amounts of computing resources, making them unsuitable for widespread local deployment.
Hence in this research work we have been concentrating on machine learning approaches with a small computing footprint and seeking to understand their operating characteristics, both to assist in design and to enable the network management team to be comfortable that they understand how the system operates. We have focused on two main areas of fault detection: transmission degradations and antenna tilt mismatches.
In Research Work 1 our aim was to improve on the state of the art in applying ML techniques to the cell coverage degradation detection problem. We showed that the RNN technique can provide greater detection sensitivity than the best of previous approaches, represented by the SVM technique, while at the same time eliminating the need for a separate hand-coded data dimensionality reduction stage. The operational detector can be efficiently implemented, as an embedded system if necessary, allowing it to be deployed locally in the RAN.
In Research Work 2 our aim was to understand how the RNN operates and what lessons there could be for future design. To do this, we built two parallel models to operate alongside the RNN and illuminate its inner workings. We showed that the gain in performance of the RNN by using more complex configurations was limited by the emergence of unwanted sidelobes in the probability density function of each RNN internal state. This leads to useful design insights; for local deployment in the RAN, for example, it could be appropriate to select a very simple RNN configuration, rather than increasing the complexity of the RNN in search of improved performance.
In Research Work 3 we considered mismatches between the designed downlink antenna electrical tilt and the actual tilt in the field. We have shown that we can detect such mismatches with unsupervised learning, using only reported data with no need for dedicated test user equipment or access to receiver signal processing. We have also shown that when training data is scarce, training performance can be improved by using federated learning to combine data from multiple sites. We have provided simple, intuitive visualisations to allow network operators to understand the system's operation. At the same time, our system is designed to be efficiently implemented in embedded systems distributed across the RAN.