Simplify Telecom network operations through Generative AI and an advanced data platform

Telecom network operations are made up of several complex business processes that require going through terabytes of data daily to optimize the network, perform root cause analysis, etc. Such root cause analysis can take several hours, if not days, to complete. Generative AI can play a key role in all aspects of the network lifecycle. Engineers rely on manuals and documented processes when operationalizing network elements. Generative AI (genAI) can use this data and provide interactive guidance and prompts to speed up and simplify optimization tasks.

To get a full picture of the business, we must integrate data from two major silos in the world of telecommunications, including OSS and BSS data. OSS (Operational Support Systems) refers to the systems running the core telecom network, the radio network, the different network elements and others. BSS refers to the different business support systems that often deal directly with the customer, i.e., ERP, CRM and billing systems.

These systems have different focus areas: The OSS systems deal with the network parameters and outputs and often report the data in terms of system health, system utilization, etc., while the BSS systems deal with the customer data, customer complaints, billing history, etc. The disconnect comes when the network engineer says that their network has delivered a greater than 99.9% call connection rate, but the customer experience team comes back and reports, for example, that the 0.1% calls that dropped by the network were of their highest lifetime value customers, for example. This data disconnect starts a very long and often mundane process of running a root cause analysis on hundreds of terabytes of data across several OSS and BSS siloed systems.?

Let’s discuss how we can streamline this problem by collaborating between Amazon Web Services (AWS) , DigitalRoute and Snowflake .

The symptoms

The problem often starts with daily—yet complex—cases such as customers experiencing dropped call behavior over a specific location. The telecom operator starts receiving customer complaints over the bad experience.?

Now, when we plot the customer calls over a map of California based on how the calls ended, we see there are certain areas where drop rates are as high as 50%. This is clearly unacceptable. Let’s build this solution step-by-step to troubleshoot and isolate the faults.

What data is needed?

• We would need billing data, which also indicates how the call ended.
• We would need Open Cell ID data from Snowflake Marketplace. A database of all telecom cell towers in the world and their latitude/longitude.
• We would need US zip codes and their geospatial polygons, which can also be acquired from Snowflake Marketplace.?
• We would need some Points of Interest databases, which are also available in Snowflake Marketplace.
• We would need customer loyalty data to understand which of the highest-value customers were impacted by the network's poor performance.
• Lastly, we would need some OSS data from cell tower performance management data directly from the network elements.

What are the challenges?

• The performance management (PM) data is in a very complex format known as ASN.1. I have often referred to it as the Parquet format of the 1980s. Most databases cannot understand ASN.1 out of the box.?
• The PM data has no customer information; the primary key of this data set is the cell tower ID. So we would need to consider a time series join between the two data sets.
• The volume of data is massive and there are hundreds of parameters in the PM data.
• How do you quickly isolate the problem and identify the root cause of the issue?
• Highly skilled human resources are needed to help understand the data.

Solution flow

• DigitalRoute Usage Engine Private Edition captures the ASN.1 files from the network elements, parses it and sends it to Snowflake.
• The billing data is loaded from the DigitalRoute Usage Engine Private Edition mediation system in near real time to Snowflake via Snowpipe/Kinesis.
• The reference data is acquired from Snowflake Marketplace.
• Using AWS Generative AI services and the foundation model in Amazon SageMaker , the network engineer can chat with the service to narrow down the fault scenarios.

Solution in action

Amazon SageMaker Machine Learning platform uses a large language model (LLM) to formulate the SQL query. The LLM uses the Snowflake Data Cloud to provide context and the schema for the data. The generated SQL and resultant output are then used by the LLM to provide a relevant response. AWS Sagemaker can host a number of LLMs; this demo uses AI at Meta 's Llama 2 foundation model, and was tested with Anthropic 's Claude. We can ask questions in natural language and the LLM model will convert the text into a complex SQL which can be executed on the data model.

Conversational GIS Example

Here I will ask a simple question in English to the LLM. The LLM will reach out to Snowflake, understand the schema and convert the natural language to the appropriate SQL code. The application will also visualize the Geospatial information automatically on a map.

Conclusion

By integrating data from multiple complex telecom OSS and BSS sources into a genAI model, we can now solve problems in minutes rather than days and hours. This is why we believe Generative AI can truly be a productivity enhancer in the telecom world in solving these “needle in a haystack” problems. The future is bright and we expect to see genAI being adopted in telecom quite aggressively. The key is to ensure that organizations think of genAI as an application - and not as a data integration platform. Always bringing genAI applications to the data, rather than shipping the data to the application, allows telecoms to quickly and easily scale genAI across their ecosystem.

Learn more about how to bring Generative AI and LLM to data and come visit us at the AWS Next Level balcony, DTW Ignite! on Sept 19-21 in Copenhagen.