Most mobile operators run heterogeneous networks (HetNets), combining various cell types and technologies. The use of small cells in these networks has grown hugely in recent years, partly because smalls cells supply network coverage for indoor public spaces. Indoor spaces account for roughly 80% of today’s mobile data traffic.
Small cells and HetNets are thus vital to modern mobile communications, but while standard self-organising networks (SONs) can self-heal, this does not work well with HetNets. As such, operators cannot rely on SON technology with such networks, but mobile operators that use HetNets still need to detect cell failures so the network can react appropriately. And if they cannot use standard SON technology to prevent coverage failures, what can they do?
This is where machine learning can provide a solution.
Detecting small cell failures with machine learning
It can be immensely difficult to detect small cell failures remotely, just from warnings and logs. For example, if a cell is ‘sleeping’, it will not broadcast information. And a ‘sleeping’ cell will look almost identical to a healthy one. So, even if the network is reporting normally, a technician back at base may not be able tell the difference between the two.
Similarly, we cannot rely on network traffic patterns to identify cell failures. An unusual pattern may indicate a problem, but it could also be interference from another radio device, or one of hundreds of other variables that could produce abnormal signals.
Each cell in a network, however, broadcasts huge amounts of low-level data, such as logs and other monitoring traces, which are produced in sufficient volume to be used as input to machine learning systems. If we feed these data into a machine learning program, we can teach it to recognise ‘normal’ network behaviour, and thus recognise when the system departs from the normal. In other words, we can train it to identify abnormalities.
Machine learning and small cell analysis: How it works
How does this work in practice? It may help to break the process down into four steps:
- Collect the data
The first step is to collect ‘training data’. This is the information you will use to ‘teach’ the machine learning program to identify failed cells. How much data you need for this depends, to a great extent, on the number of cells in the network. As a rule, larger networks will want more data than small ones. But networks with more cells also require less data per cell because patterns emerge more quickly.
In either case, this initial data set should be large enough to offer a representative example of network behaviour, including both normal performance and abnormalities. You can enhance this process by cross-correlating – i.e. analysing simultaneously – different data types, such as:
- Conventional log data
- Billing data, which gives further information on customer habits
- Minimisation of drive test (MDT) data. (MDT data is data collected by driving or otherwise moving around the area of the mobile network and measuring network performance. It is done in such a way as to minimise the amount of driving required.)
This correlation process speeds up the emergence of patterns in the data, helping the computer to ‘learn’ from it more quickly.
- Normalise the data
Different types of data have different attributes, which we measure in different ways. Therefore, before a computer can analyse different types of data together, we need to rescale it all so the software can make sense of it. This is known as ‘normalisation’.
This step is crucial to any machine learning project, and, helpfully, many modern software packages include features to aid data pre-processing, including normalisation.
- Help the computer learn from the data
When the data set is ready, the learning will begin. To simplify, this means feeding your data into a machine learning algorithm and asking it to perform a task (e.g. identify failed small cells in a network).
The results may not be perfect at first, but this allows you to identify errors and root out the causes before running another test. And by doing this iteratively, it will improve each time. How long this process takes will depend on the amount of data you are analysing, and how difficult the patterns are to detect, but if patterns exist, the right machine learning techniques should be able to detect them.
- Analyse in real time
Eventually, you will have a machine learning algorithm that has learned by being trained on very large volumes of data. Once trained, it will be able to analyse cell behaviour in real time, not only flagging suspected anomalies, but also further improving its efficacy by learning on the continuous stream of real-time data. One aspect of this is that it will be able to categorise further the cell types and their characteristic behaviours. Each cell will normally fall into a recognisable category, such as cells in busy intersections, cells in quiet corners that help to provide full coverage, and cells in stores. And by learning what different cell types usually do, the software can improve the accuracy of its diagnosis of failures.
Any problems detected are flagged for engineers. Machine learning allows operators to respond to issues with ever greater speed and accuracy, and reduces the amount of routine diagnostic checks, which are performed without any evidence of there being a problem.
Small cell machine learning in action
We can see how the above works by looking to a recent small cell machine learning project conducted for a shopping centre.
After gathering the training data and refining the machine learning model, analysis showed that 1.3% percent of the studied small cell behaviour differed from expectations. Two cells were found to be repeat offenders, constantly entering ‘sleep’ status and causing outages. As a result of these failures, almost thirty thousand subscribers experienced lost or refused connections.
The study also found that the user’s mobile handset and operating system could impact the quality of a call, with a significant decrease in the Call Setup Success Rate and the CS RAB Establishment Success Rate between the most and second most used handsets. (CS RAB, ‘conversational speech radio access bearer’, is a channel via which speech communication can take place).
The machine learning techniques used allowed operators to predict service degradation, cell outages and anomalies in real time. Network faults were thus quickly diagnosed, ensuring quick remedy and a stronger, more reliable mobile network.
And with access to mobile data increasingly vital for consumers and businesses, the shopping centre saw increased footfall, rentals and revenues from using the machine learning system.
Mobile network maintenance going forward
Many industries are now embracing machine learning, and the structure of cellular data makes it ideal for this kind of analysis. In the case study above, for instance, we see how machine learning can play a crucial role in maintaining small cell mobile networks.
In fact, in areas of high-density small cell activity, machine learning is one of the most effective ways to maintain a high-quality service. In a world where signal strength affects where consumers spend their money, this is not something mobile operators can ignore.
Interested in hearing industry leaders discuss subjects like this? Attend the co-located IoT Tech Expo, Blockchain Expo, AI & Big Data Expo, and Cyber Security & Cloud Expo World Series with upcoming events in Silicon Valley, London, and Amsterdam.
