Using machine learning to detect small cell anomalies
Usage of small cells has proliferated to a point where most mobile operators run heterogeneous networks (HetNets) – networks built on a combination of different cell types and technologies.
According to Francis Miers, director at software lifecyle automation specialist, Automation Consultants, a contributing factor to this trend is environment: smalls cells typically supply network coverage in indoor public spaces and about 80% of today’s mobile data traffic is generated indoors.
Standard self-organising networks (SONs) can self-heal. However, SON technology was not built for HetNets, so operators using heterogenous small cell networks are reluctant to employ it. Any cell failures nonetheless need to be detected as quickly as possible to avoid creating a coverage gap, and so that neighbouring cells can kick in and carry the load, pending a repair.
Machine learning to the rescue
Mobile operators are often unable to know for sure if cells are working correctly. It’s not always possible for technicians back at base to detect failure. If the cells are ‘sleeping’, for example, then the network is reporting normally, but not broadcasting any radio. Nor can cell failure always be detected easily by looking at network traffic patterns. Unusual cellular activity might not always mean something significant. It could signify interference with another radio device, or it could be down to one of a hundred or more other factors that can cause a behavioural anomaly.
Each cell broadcasts huge amounts of low-level data in bits and bytes. Cellular data is numerical – so it’s suitable for ML techniques that focus on anomaly detection. For example, if you have a whole load of average points, ML can detect the outliers. Machine learning techniques can analyse the data in comparison to other, known behaviours. This enables operators to distinguish what is normal, such as a cell not currently being used, from what has gone wrong, such as a cell that has failed.
How does it work?
There are four steps to applying machine learning to small cell analysis.
- Collect the data
Quite a lot data is needed to draw effective conclusions. If an operator doesn’t think a cell is working properly, it might be that no-one is using that cell at certain times of the day. Distinguishing odd from normal behaviour requires lots of data for analysis.
The amount of data required would depend on the number of cells. With many cells, les data per cell is required because patterns emerge faster, but the overall volume of data required is still greater than with a smaller number of cells.
Types of data you should consider analysing include:
- Conventional log data
- Billing data to gain insight into customer habits
- Minimum Drive Test (MDT) data
Cross-correlating these different data types allows patterns to emerge faster than they would by analysing log data alone.
- Normalise the data
Telecoms data is naturally more structured than many other types of data but nonetheless, it needs to be in a normalised form that a data analysis tool can recognise. Normalisation means adjusting values that usually sit on different scales so that they fall on a common scale, such as a number between 0 and 1, percentage or fraction.
The learning period will depend on the amount of data you are analysing. The quantity of data described above would be processed for several months under supervision, to associate anomalies with root causes.
- Analyse in real time
Real-time analysis helps machine learning software compare cell behaviour with known failures, recognise patterns for outage prevention and build categories of cell behaviour. Each cell will fall into a category, for example: cells in busy intersections, cells in quiet corners that are there to provide full coverage, and cells in particular stores. By categorising cell types and their expected behaviours, operators are able to recognise patterns and respond to issues with greater speed and accuracy.
Results of a small cell machine learning project
A recently completed small cell ML project conducted for a shopping centre demonstrated its ability to analyse data and detect anomalous behaviour.
The data analysis showed that 1.3% of the studied small cell behaviour was abnormal or extremely different to standard expectations. Two cells were repeat offenders, constantly falling into ‘sleeping’ status and causing outages. This affected 28,739 subscribers who had to put up with lost or refused connections and a poor quality experience.
The type of mobile handset and operating system was found to influence the quality of a call. There was a significant decrease between the Call Setup Success Rate and the success rate for setting up a radio channel to carry a call (CS RAB Establishment Success Rate) when comparing the most used model of handset vs. the second most used model.
The machine learning techniques applied to this case study successfully predicted service degradation and identified cell outages and anomalies in real time to ensure a stronger, more reliable mobile network. Network faults were diagnosed quickly, ensuring immediate remedy.
Ultimately, the structured nature of cellular data makes it a natural candidate for ML. In areas of high density small cell activity, ML is the most effective way to maintain quality of service. And in the real world, where a difference in mobile signal strength will affect where consumers spend their hard earned money, ML made a big difference to footfall, rentals and, as a consequence, revenues, for the shopping centre customer described above.
The author is Francis Miers, director at software lifecyle automation specialist, Automation Consultants