Multi-Modal LLM Anomaly Detection for Communications Networks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing anomaly detection methods in communication networks are unable to efficiently incorporate information about why an anomaly occurs and how to remedy it, as they typically focus on single modalities of data, limiting their ability to detect anomalies in complex systems with multiple potential causes.

Innovation Solution

A multi-modal large language model (LLM) is trained using multiple modalities of data, including telemetry, image, and external information, to provide granular and informative anomaly detection by transforming data into vector embeddings and fine-tuning the model with supervised learning, enabling it to identify the source of anomalies quickly.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If single modality data processing is used, then device complexity is reduced, but anomaly detection accuracy and information completeness deteriorate

Engineering Contradiction:
Improveanomaly detection accuracyVSAvoidmodel complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system segments different data types (telemetry, images, external information) into separate modalities, each processed by dedicated transformation modules before being integrated into the multi-modal LLM. This allows the model to handle complex multi-modal data while maintaining organized, modular processing for each data type.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The multi-modal LLM serves as a universal model that can process multiple types of data modalities (telemetry, images, external information) through a unified architecture. This single model performs anomaly detection across diverse data types, reducing the need for multiple separate models while improving detection accuracy.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Measurement precision

If multiple modalities of data are processed, then anomaly detection accuracy improves, but processing time increases

Engineering Contradiction:
Improveanomaly detection accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

Data from multiple modalities is transformed into vector embeddings before being input to the LLM. This preliminary transformation prepares the data in advance, enabling the model to process diverse data types more efficiently during anomaly detection without significant time penalty.

Inventive Principle:
Principle #10Preliminary action

3Loss of information

If comprehensive data transformation is performed, then information completeness improves, but computational complexity increases

Engineering Contradiction:
Improveinformation completenessVSAvoidcomputational complexity
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

Vector embeddings serve as an intermediary representation that bridges different data modalities (telemetry, images, external information) and the LLM processing architecture. This intermediate representation preserves information from diverse sources while enabling efficient processing through a unified computational pathway.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4617880A1Anomaly detection for communications networks
Publication Date: 2025.09.17 BRITISH TELECOM PLC
  • EP4617880A1 patent drawingFigure 1
  • EP4617880A1 patent drawingFigure 2
  • EP4617880A1 patent drawingFigure 3

AI summary

Some examples of training a large language model (LLM) to detect an anomaly comprise receiving a plurality of modalities of training data, each modality of training data comprising a different data type. The training data comprises labels indicating a presence of an anomaly. Each modality of training data is transformed with a respective transformation to obtain respective vector embeddings. The respective vector embeddings of the training data are input into a pre-trained LLM. Weights of the pre-trained LLM are fine-tuned using the training data via supervised learning such that the pre-trained LLM is able to detect anomalies in multi-modal data.