In sum, what to know:
–Machine learning and large language models could help operators connect messaging failures, service degradation and user experience — but only if they first solve the challenge of integrating data across network silos.
-Telecom systems engineer and independent researcher Srinivasa Maringanti shares insights from building AI models that recognize data signals which indicate network instability, and analyze data on a more granular geospatial basis to discover patterns that are hidden by aggregate performance metrics.
One of the ways that network operators want to use artificial intelligence is to improve network operations and customer experience. Telecom networks being what they are, you could turn AI models loose to analyze nearly infinite amounts of data—and spend a lot of money doing it. But that doesn’t mean that models will recognize patterns that actually have meaningful effects on network experience.
Srinivasa Chary Maringanti, a telecoms systems engineer and independent researcher, is working on using overlooked network signals that could help AI recognize network instability. His approach centers around something fundamental to telecom networks, that typically, customers don’t even notice: Fallbacks. Specifically, the times when a text message or a voice call falls back to older technology.
When a Rich Communication Services (RCS) message falls back to SMS text, it doesn’t look like a network problem to the person sending it—after all, the message still arrives. But the fact that the fallback has happened can offer an important clue about the underlying health of the mobile network.
“In the case of RCS and fallbacks, or VoLTE calls and fallbacks to traditional systems, it gives us a lot of signals on what’s actually happening in the network,” Maringanti explained. Together with signaling and transport data, that can help identify patterns that could indicate emerging congestion or network instability, he added. But these signals are often overlooked, because they don’t result in service disruptions for customers, and they are buried deep in highly granular data. “All these signals are at message level or a call level, it’s humanly impossible to monitor or identify,” Maringanti said. That’s where AI can potentially be of use.
Maringanti has more than a decade of network engineering experience across the Radio Access Network, core, and O-RAN architectures, and currently works with telecom infrastructure and applying AI and ML to network reliability issues at a major U.S. telecom operator. He is pursuing AI research as part of a personal project to independently develop AI models that connect network performance data with highly granular geospatial information.
Maringanti’s work illustrates the challenge for network operators: how to extract actionable intelligence from enormous volumes of data that are often silo’d by service, technology or operational team. That’s an issue that the broader telecom ecosystem is also trying to figure out: It seems clear that the right network data could be analyzed by AI and reveal insights that might otherwise get missed. But how do you get there?
RCS fallback as a signal of network instability
Maringanti’s approach of using RCS fallbacks as a sign of impending network instability illustrates some of the fundamental complexities in correlating data from telecom networks, and putting it into a form that AI models can use.
Back to using SMS as a signal for network instability: Because SMS/MMS are legacy, non-IMS cellular technologies, they don’t have a direct correlation to RCS, Maringanti explained—so in order for AI models to understand RCS fallback as a signal, humans have to establish one-to-one mapping between RCS and SMS/MMS that the models can recognize. Maringanti is developing models to establish that relationship, using available open-source data to explore how RCS events can be mapped to SMS and MMS fallbacks. There is nuance to what can be learned from these signals: Degradation on the IMS network will cause issues with calls or RCS messages, but failures at the IMS level don’t necessarily mean that the infrastructure or network system is failing, Maringanti pointed out. However, multiple message types failing at the same time could indicate a larger issue.
Fallbacks should be relatively uncommon when a service is performing well—so an operator could have a largely positive customer-experience picture while still accumulating a small number of under-the-radar failures among the billions of messages sent over telecom networks (more than 2.2 billion SMS and MMS messages alone, according to CTIA’s most recent annual survey). At the individual-message level, those fallbacks may be difficult to decipher and occur too rarely for current monitoring systems to register; and the point isn’t to count failed messages to reach a threshold that triggers an alarm. Instead, if AI can discern a pattern from many events at scale, it could help to reveal recurring issues associated with particular locations or service conditions.
Looking beyond the averages of network performance
Network operators and benchmarking companies routinely examine performance at a market or city level. That broad view has its uses, but it can obscure problems affecting a relatively small area, or a small group of customers. As most people have experienced, a city can have strong coverage performance on an overall basis, while users in a particular building, street or neighborhood still have persistent connectivity problems. This is the downside of reliance on average performance across a given geography: outliers effectively get erased.
Operators typically look at performance in the context of a relatively high level, because of the scale of the data involved. They could get more granular—but it would mean sifting through billions of records and terabytes of data, Maringanti pointed out. “Without AI and ML algorithms, it would be humanly impossible,” he said. “So it’s not a matter of, we can’t go any further [in granularity]. It’s the problem of how to handle such enormous amounts of data.”
He’s working on building AI models in this area as well: Models that use geospatial analysis to identify localized network-performance patterns that can be hidden when conditions are viewed through those broad geographic averages. His approach here is based on an open-source geospatial indexing system, which divides the globe into hierarchical hexagonal cells, or hexbins. The largest such “hexbin” cell could represent continent-level scale, but at finer resolutions, the cells can represent progressively smaller geographic areas—city-level, yes, but potentially, down to neighborhood or street level as well.
In principle, this framework could apply across different services (not just RCS and messaging), Maringanti said, relying on AI to help identify patterns in a specific geographic location that would be difficult to see when each service is monitored separately and with market-level aggregate data.
The data-integration challenge comes first
Maringanti’s research shows that developing and operationalizing AI-based solutions is much more than just a matter of connecting a model to a network database and asking it to find problems. Correlation and causation in complex, dynamic networks is a tricky thing. And perhaps even more fundamentally, in order to utilize AI, you have to deal with fragmented data systems that have already been barriers to efficient telecom data use for more than a decade.
Messaging, voice, data, and various apps are all in silos, Maringanti said, and when issues arise, a network expert looks at the related data for the known issue. Identifying the signs and cause of disruptions across services may require engineers from several specialties to sit down together—which usually only happens when they are trying to fix something. “But when you start integrating systems together and have a LLM monitor them, that would give us better signals,” Maringanti said. “The biggest challenge here is building the integration of these systems together before an LLM even looks into it.” His concept to deal with this is to build a common identifier that links all the services available at a particular level of data granularity, then map the resulting information to the geographic hexbins—which then become the units that the AI models monitor, rather than the billions of individual records.
It’s a balancing act: Disparate data sources need to be integrated into data streams that AI can digest, with the option to get finer granularity than is currently commonly used; but at the same time, turning LLMs loose to seek patterns at the individual record level across massive data sets could end up being the equivalent of a very expensive wild goose chase, if the models don’t know what to look for. To cut down on operational costs, MMaringanti envisions advanced AI models sharing analysis duties: machine-learning algorithms can operate less expensively across large data sets, while the LLMs do the work of interpreting patterns in context and connecting information from different systems.
Putting the ‘text’ in context for AI
Finding a potentially interesting correlation that could be the basis for smarter AI detection of network instability, is only the beginning. The broader objective is to develop operational AI that converts distributed network data into traceable and actionable intelligence to help network teams maintain more reliable network operations. But telecom networks are constantly changing environments, and a model that encounters an unfamiliar combination of events may lack the context required to interpret it correctly—so human expertise is necessary. Helping AI to find the right signals in all of the data noise is a series of small steps, and requires substantial validation. The work is in its early stages, but it is developing and changing rapidly—just as the models themselves are.
The key, Maringanti said, is to build the systems that give AI the right context before expecting it to deliver reliable results. That context, in his view, starts with texts—RCS messages and SMS fallbacks, that could end up communicating messages not just between people, but about network health.
Table of Contents
