With more widespread deployment of 5G base stations, along with the rise in the types of uplink interferences in telecommunication systems, multi-network coordination becomes increasingly complex. Simultaneously, there has been a rapid growth in interference caused by external telecommunication devices, including the latest smart devices.
Frequent occurrences of single-typed and compound-typed disruptions have a greater impact on user experience. Currently, the existing 5G interference identification system struggles with issues of high operation and maintenance manpower costs, low efficiency and low accuracy of identification. It is the utmost importance to distinguish accurately and promptly between types of network interferences to increase the efficiency of operations and maintenance, as well as problem resolving effectiveness encountered by the existing networks.
China Mobile Research Institute introduced the innovative two-way dynamic feature matching algorithms – FETTrans – for identifying single-typed or compound-typed wireless network interferences precisely in existing networks. For the first time, this algorithm puts forward the mechanism of bidirectional dynamic features matching to enhance the computing performance of multi-label classification issues. It also introduced a multi-head attention mechanism to improve its general ability and designs an end-to-end parallel network structure to increase computing efficiency and to lower deployment difficulty of the existing networks. The above characteristics are practical and inspiring solutions to the classification issue.
This algorithm decomposes the input signals to acquire multi-dimensional input features (Data-Tokens), and it completes a bidirectional data feature training in the Dynamic Adaptation module simultaneously with the structured multi-dimensional output codes (Key-Tokens), accomplishing the tasks of classifying complex disruptions. The research paper “FETTrans: Analysis of Compound Interference Identification Based on Bidirectional Dynamic Feature Adaptation of Improved Transformer” based on this algorithm has been accepted by SCI journal “IEEE ACCESS”.
The algorithm has been verified in multiple scenarios in the existing networks of five provinces in China, with over 90% accuracy in identified classification, 92% in mAP and more than 90% in identifying interference scenarios. Furthermore, this algorithm has high universality and operating efficiency. In the GPU and CPU hardware deployment environment, it can recognize 8,000 and 1,100 cell interference types per second respectively, greatly improving the accuracy of identifying disruption, operating costs and efficiency.
Moving forward, China Mobile will commence research on interference source tracking to optimize user perceptions, providing new ideas and methods for the smart optimization of wireless networks.
