ML-Based Physical Impairment Detection in Subscriber Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Subscriber networks face challenges in locating and monitoring physical impairments, such as bad splices, which can vary over time and affect other lines due to vectored system coupling, making it difficult to detect and diagnose without causing service disruptions.
Innovation Solution
The use of machine learning techniques to dynamically detect physical impairments by analyzing primary per-tone data like XLIN, SNR, QLN, and secondary information, employing a prediction model that uses stacked generalization and feature subset ensembles to predict impairment likelihood without requiring all data points, allowing for self-healing processes that maintain network uptime.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional dual-ended line testing (DELT) is used to detect physical impairments, then impairment detection capability is improved, but service disruptions occur
Solution Approach 1:
The system performs preliminary actions by continuously collecting and analyzing per-tone data in the background during normal network operation. Machine learning models are trained and updated with historical data, enabling the system to detect impairments proactively before they cause service failures, thus avoiding the need for disruptive testing
Solution Approach 2:
The patent introduces machine learning models as intermediary components that analyze per-tone data to detect physical impairments. These models act as mediators between the network data and impairment detection, enabling continuous monitoring without requiring traditional disruptive testing methods
2Measurement precision
If per-tone data is collected continuously for impairment detection, then detection accuracy is improved, but data complexity and processing difficulty increase
Solution Approach 1:
The patent segments the high-dimensional per-tone data into multiple feature subsets, each representing different aspects of network conditions. Machine learning models are trained on these segmented features independently, then combined to make final impairment predictions. This segmentation reduces processing complexity while maintaining detection accuracy
Solution Approach 2:
The system transforms raw per-tone data into derived parameters and features that are more suitable for analysis. By changing the parameter representation from raw high-dimensional data to meaningful features, the system reduces processing complexity while preserving or enhancing detection capability
3Ease of operation
If visual inspection is used to locate bad splices, then simplicity of method is maintained, but detection capability deteriorates due to time-variability and environmental conditions
Solution Approach 1:
The patent introduces machine learning models as intermediaries that automatically analyze per-tone data to detect and locate physical impairments. This automated intermediary replaces manual visual inspection, providing continuous monitoring capability that overcomes the limitations of human inspection while maintaining operational simplicity through automated processing
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
In some implementations, per-tone data for a line of a subscriber network and data indicating a set of one or more scores is obtained. Each score included in the set of scores indicates a conditional likelihood that the line has a type of impairment with respect to a different feature subset ensemble. The per-tone data and the data indicating the set of one or more scores is provided as input to a model. The model is trained to output, for each of different sets of feature subset ensembles, a confidence score representing an overall likelihood that a particular line has a physical impairment. Data indicating a particular confidence score representing an overall likelihood that the line has the physical impairment is obtained.