ML Alarm Suppression for Network Operations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Network operators face challenges in managing complex multi-layer networks with numerous alarms, requiring expertise to differentiate between important and non-actionable alerts, as existing solutions rely on manual intervention, static rules, and limited alarm correlation, failing to adapt to evolving networks and predict resolution efforts effectively.
Innovation Solution
The implementation of machine learning models using supervised and reinforcement learning techniques to automatically classify alarms as important or non-important, trained on historical data and NOC interactions, allowing for real-time suppression of non-actionable alarms while maintaining high recall of important ones.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If manual intervention and static rules are used to manage alarms, then network operators can handle alarms with expertise, but the system cannot adapt to evolving networks and requires continuous manual maintenance
Solution Approach 1:
The patent implements dynamic alarm suppression by training machine learning models on historical alarm data that continuously evolves with the network. The models adapt to changing network patterns and behaviors over time, transforming the static rule-based system into a dynamic one that learns and improves automatically as the network evolves.
Solution Approach 2:
The machine learning system performs self-service by automatically learning from historical data and making suppression decisions without requiring continuous manual intervention. The system trains on its own historical performance and adapts autonomously, reducing the burden of manual rule maintenance while improving adaptability.
2Reliability
If all alarms are displayed to network operators, then no important alarms are missed, but the volume of alarms becomes unmanageable and operators cannot efficiently identify critical issues
Solution Approach 1:
The patent extracts and removes non-actionable alarms from the display using machine learning-based suppression. The system identifies and filters out alarms that do not require operator attention while preserving all actionable and important alarms, thereby reducing the displayed volume without sacrificing reliability.
Solution Approach 2:
The system applies partial suppression by selectively filtering only the non-actionable portion of alarms while maintaining full visibility of actionable ones. This partial action approach optimizes productivity by reducing noise without excessive filtering that would compromise recall of important alarms.
3Productivity
If device-level alarm suppression is implemented, then non-important alarms can be reduced, but critical alarms such as link down, power failure, and service disruption cannot be suppressed
Solution Approach 1:
The patent introduces a machine learning-based intermediary layer between the network devices and operators. This intermediary intelligently evaluates each alarm using learned patterns from historical data, determining whether suppression is appropriate based on the alarm's actual impact, thereby overcoming the rigid limitations of device-level suppression rules.
Solution Approach 2:
The system changes the parameter of alarm suppression from a binary device-level decision to a nuanced, data-driven determination. By utilizing historical alarm data and machine learning models, the system dynamically adjusts suppression decisions based on multiple parameters including alarm patterns, network context, and historical resolution outcomes.
4Productivity
If alarm correlation techniques are used in NMS, then alarm count can be reduced, but the technique requires pre-configured correlation policies and cannot handle alarms without existing correlation rules
Solution Approach 1:
The patent performs preliminary action by pre-training machine learning models on extensive historical alarm data before deployment. This preliminary training enables the system to recognize patterns and relationships across diverse alarm types without requiring pre-configured correlation policies for each specific alarm scenario, thereby improving adaptability to new alarm types.
Solution Approach 2:
The machine learning model serves as a universal alarm correlation mechanism that can handle multiple alarm types and scenarios through a single unified approach. Unlike traditional correlation techniques that require specific policies for each alarm type, the ML model generalizes across different alarm patterns, making the system versatile and adaptable to new alarm types without additional configuration.
Data Source
Figure 1A~1B
Figure 2~3
Figure 4~5
AI summary
Systems and methods (100) include receiving (102) alarms from a network; utilizing (104) a machine learning model to classify the alarms as one of important and non-important; and displaying (106) the important alarms and suppressing display of the non-important alarms. The systems and methods can further include training (108) the machine learning model with historical alarm data that includes features related to an associated device and comments related to how a Network Operations Center (NOC) handles an associated alarm or group of alarms. The training (108) can be via supervised machine learning with the features used as labels or via reinforcement learning with the features used as a reward.