Federated KPI Anomaly Detection for Communication Networks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional anomaly detection (AD) methods in communication networks rely on parametric models that require extensive parameter tuning and centralized training, which is resource-intensive and lacks sufficient labeled data for supervised learning, leading to inefficiencies and high costs.

Innovation Solution

A federated user-feedback driven AD system where a management entity receives and labels anomalous points from network elements (NEs), updates a global AD model with local parameters, and broadcasts configuration parameters to NEs for decentralized training and inference, utilizing unsupervised and supervised techniques to enhance detection efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If centralized training is used for AD models, then model training can be performed uniformly, but data transfer from each network entity to the management entity is required, increasing network load and processing time

Engineering Contradiction:
Improvemodel training efficiencyVSAvoiddata transfer time
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The patent segments the centralized training process into distributed local training at network elements and centralized aggregation at the management entity. Each network element performs local AD model training independently using its own data, then transfers only model parameters or gradients to the management entity for aggregation. This segmentation eliminates the need to transfer raw data while maintaining training effectiveness.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces an intermediary mechanism where network elements act as intermediate training nodes. Instead of directly transferring all data to the management entity, the intermediary elements perform preliminary processing and model training locally, then communicate condensed model information to the central management entity. This intermediary approach reduces network load and transfer time.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If supervised learning with labeled data is used for classifier models, then detection accuracy can be improved, but large amounts of labeled data are required, and manual labeling is expensive in terms of domain expertise

Engineering Contradiction:
Improveanomaly detection accuracyVSAvoiddata labeling complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent implements self-service through unsupervised and semi-supervised learning approaches where the AD models automatically learn anomaly patterns from unlabeled data without requiring manual labeling. The models perform self-training by identifying anomalies based on statistical deviations, clustering, or reconstruction errors, eliminating the need for expensive domain expert labeling while maintaining detection accuracy.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent changes the learning paradigm from supervised to unsupervised and semi-supervised learning, fundamentally altering the data requirements. Instead of requiring labeled anomalies, the models use parameter changes in unlabeled data (such as statistical moments, distribution shifts, or reconstruction errors) to detect anomalies, thereby avoiding the complexity of manual labeling.

Inventive Principle:
Principle #35Parameter changes

3Ease of manufacture

If parametric models are used for anomaly detection, then detection can be performed with established methods, but several parameters require fine-tuning to achieve efficient detection

Engineering Contradiction:
Improvedetection method implementationVSAvoidparameter tuning complexity
Core Design Contradiction:
Ease of manufactureVSDevice complexity

Solution Approach 1:

The patent implements feedback mechanisms where the AD models continuously monitor detection performance and automatically adjust parameters based on feedback from detected anomalies and system performance metrics. This closed-loop feedback enables adaptive parameter tuning without manual intervention, simplifying implementation while maintaining detection efficiency.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent transitions from static parametric models with fixed parameters to dynamic models that automatically adapt parameters based on incoming data characteristics and anomaly patterns. The models dynamically adjust their parameters in response to changing network conditions, eliminating the need for manual fine-tuning while maintaining ease of implementation.

Inventive Principle:
Principle #15Dynamics

4Reliability

If anomalies occur in a fraction of network elements, then the system remains mostly normal, but the required amount of labeled anomaly data may not be available to train classifier models

Engineering Contradiction:
Improveanomaly detection capabilityVSAvoidlabeled training data availability
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent merges data from multiple network elements to collectively train the AD models. By combining unlabeled data from numerous network elements, the system accumulates sufficient training data even when individual elements rarely exhibit anomalies. The federated learning approach merges local model updates from multiple elements, effectively pooling their collective experience to improve detection capability without requiring rare labeled anomalies from each element.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4238291B1Management entity, network element, and methods for supporting anomaly detection for communication networks
Publication Date: 2025.08.27 HUAWEI TECH CO LTD
  • EP4238291B1 patent drawingFigure 1
  • EP4238291B1 patent drawingFigure 2
  • EP4238291B1 patent drawingFigure 3

AI summary

The present disclosure relates to a Network Element (NE) and a management entity for supporting Anomaly Detection (AD) in a communication network. The NE is configured to determine, based on a local AD model, a point to be an anomalous point in Key Performance Indicator (KPI) time-series data, provide the anomalous point to the management entity, receive a label for the points of the KPI time series data from the management entity, and update the local AD model based on the received label. The management entity is configured to receive, from the NE, a point determined to be an anomalous point, and provide to the NEs, a label for the point. The management entity is further configured to receive, from the NE, a parameter of an updated local AD model of that NE, and update a global AD model based on the received parameter.