Neural Network Anomaly Detection in Telephony KPIs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current anomaly detection systems in telephony networks face challenges in efficiently detecting and mitigating anomalies across multiple dimensions due to data availability issues, overfitting, and the inability to effectively correlate intra-instance Key Performance Indicators (KPIs), requiring significant data for each model instance and failing to capture combined KPI anomalies.
Innovation Solution
A neural network-based system that uses a trained autoencoder to detect anomalies by inputting key performance indicator values and selector values, determining a score based on the neural network's output, and comparing it to a threshold, allowing for shared models across similar instances and separate models for distinct ones, while incorporating correlations among KPIs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate models are built for each instance of groupings (e.g., each trunk group, subscriber, or time period), then the system can capture specific traffic patterns for each instance, but the number of model instances explodes, leading to overfitting due to insufficient data for each individual model
Solution Approach 1:
The patent merges multiple separate instance models into a single unified neural network model that processes multiple instances simultaneously. Instead of training separate models for each trunk group or subscriber, the system uses one model that takes instance identifiers as input features and predicts anomalies across all instances, thereby reducing model complexity while maintaining detection accuracy.
Solution Approach 2:
The neural network model is designed to be universal across multiple instances by incorporating instance-specific features (such as trunk group ID, subscriber ID, time period) as input variables. This single model performs the function of multiple specialized models by adapting its predictions based on the specific instance being evaluated, eliminating the need for separate models while preserving instance-specific detection capabilities.
2Device complexity
If a single global model is used to model the entire network, then the system avoids the complexity of multiple models, but it requires either numerous small changes or a very large change to affect overall network traffic, reducing responsiveness to local anomalies
Solution Approach 1:
The patent applies local quality by incorporating instance-specific features into the global neural network model. Each instance (trunk group, subscriber, time period) has its own characteristic feature values that the model uses to make localized predictions. This allows a single global model to respond sensitively to local anomalies by adjusting its output based on the specific features of the instance being evaluated, without requiring separate models for each location.
3Measurement precision
If extensive data is collected for each model instance to train accurate separate models, then the system can achieve good detection accuracy for each instance, but the data requirement becomes unsustainable and training time increases significantly
Solution Approach 1:
The patent combines training data from multiple instances into a single unified training dataset for the neural network. Instead of requiring sufficient data to train separate models for each instance, the system pools data across all instances and uses instance identifiers as features. This allows the model to learn from the collective data of all instances while maintaining the ability to detect instance-specific anomalies, dramatically reducing the data requirement per instance.
4Device complexity
If traditional static thresholding approaches are used for anomaly detection, then the system is simple to implement, but it cannot effectively capture complex KPI relationships or detect combined KPI anomalies
Solution Approach 1:
The patent replaces the mechanical static thresholding system with a neural network-based intelligent detection system. Instead of using fixed thresholds that require manual configuration and cannot adapt to complex patterns, the system uses a trained neural network that automatically learns complex relationships between multiple KPIs from the data. The model processes multiple KPI inputs simultaneously and produces anomaly scores that reflect complex inter-KPI relationships, eliminating the need for manual threshold setting while capturing sophisticated anomaly patterns.
Data Source
AI summary
The present invention relates to communications methods and apparatus dynamically detecting and/or mitigating anomalies in communications systems/networks. An exemplary method of operating a system to detect traffic anomalies includes the steps of: inputting input values, said input values including i) a set of key performance indicator values and ii) one or more selector values corresponding to a first anomaly group classification instance into a trained neural network; operating the trained neural network to output values corresponding to the inputted set of key performance indicator values; determining a score for the first anomaly group classification instance based on the output values of the trained neural network corresponding to the key performance indicator values, each of said output values corresponding to one key performance indicator input value; and when the score is above a first threshold determining that an anomaly condition exists with respect to the first anomaly group classification instance.