Analysis of Logs Produced by Push-to-Talk Radios and Infrastructure in Motorola Solution’s SmartConnect.
For decades radio communication has been confined to the boundaries of network coverage. In certain mission-critical communications for public safety organisations, it is crucial to stay connected even beyond the boundaries of the radio system. Motorola Solutions’ SmartConnect is a system optimised for reliability and excellent coverage to ensure connection between teams no matter where they are.
Land Mobile Radio (LMR) is the leading communication system for public safety, voice-centric, “Push-To-Talk” (PTT) mobile radio communications. P25 (Project 25), is a set of standards suited to public safety radio systems and first responders to address interoperability issues related to LMR radio communications. Long-Term Evolution (LTE) is a widely deployed broadband technology offering high-data-rate applications currently not supported in LMR. Both LMR and LTE provide value for public safety users, however, the greatest benefit is when both technologies are leveraged simultaneously. When LMR is unavailable, SmartConnect is automatically switching the P25 channel to the broadband LTE system without intervention.
The production system consists of infrastructure and directly connected radios. There are two kinds of data available. Both infrastructure and radios are being monitored and produce logs and metrics. In addition, there is a recorded history of customer documented issues of the production system, that can be correlated against the logs and metrics.
The goal of the thesis is to apply and evaluate various machine learning techniques to automate analysis of the ever-growing amount of logs and metrics produced by a production system consisting of two main actors - infrastructure and Push-ToTalk devices connected to the infrastructure. The target is to try to correlate the reported issues against the logs and metrics (e.g. caused by end-user loaded wrong certificate, system-admin applied a wrong configuration, dependency broke down) in order to automatically detect and possibly self-heal or help the end-user to overcome some of the production system problems.
Prior to applying machine learning to the problem, it is needed to preprocess the raw logged data from both the PTT devices and the infrastructure. The first problem our research tries to address is whether it is possible to use machine learning techniques at all to detect anomalies on the provided data.
Another part of the research is to try to analyse the current state of the logging to determine whether there are properties of the system that should be recorded further in order to improve the performance of our machine learning models or, conversely, there are data properties that could be omitted from the trace as they do not provide any valuable information and unnecessarily take up storage space. Exploratory analysis will be performed on the datasets in order to identify anomalous patterns and identify hypotheses to be further tested.
If there are informative patterns yielded by the exploratory analysis, we will propose different machine learning approaches to detect them and we will evaluate their performance in a series of experiments.
After achieving the thesis statement, there are possibilities of extending the project. One such extensions is applying machine learning techniques in order to characterise audio streams and classify different kinds of poor audio connection. Another possible research question is to determine how much data is required to log to predict an anomaly to effect the data size.
- Understand and analyse the structure of a real-world dataset.
- Apply data preprocessing techniques on raw log data from live production systems to prepare inputs for machine learning models.
- Build further on the grounding of principles acquired in the Machine Learning course, and to determine if those principles can be applied to achieve anomaly detection.
- Investigate the data quality requirements that would allow for anomaly detection.
- Ability to discuss the methodologies and theory of anomaly detection techniques in general as well as at an academic level.