Deep Neural Network Classifier for Communication Impairment Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep neural network classifiers for detecting impairments in communication networks are affected by modem effects not included in their training sets, leading to inefficiencies and a need for costly manual updates.
Innovation Solution
A method that involves obtaining measured loop frequency responses, extracting modem parasitic effects, clustering them into groups, determining modem signatures, and creating an augmented training data set by subtracting and adding these signatures to original data, which is then used to train a deep neural network classifier.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a deep neural network classifier is trained on a limited training set, then the training process is faster and requires less data, but the detection reliability deteriorates when encountering modem effects not included in the training set
Solution Approach 1:
The patent applies preliminary action by extracting and clustering modem parasitic effects from available data before training the neural network. By pre-identifying and organizing M distinct modem effect patterns through clustering analysis, the system prepares a comprehensive set of interference characteristics in advance. This allows the augmented training set to proactively cover various modem effects without requiring exhaustive real-world data collection, thus maintaining detection reliability while avoiding the need for extremely large training datasets.
Solution Approach 2:
The patent uses copying by creating augmented training samples through synthetic generation. For each original training sample, the system generates M augmented versions by adding different clustered modem parasitic effects. This copying approach with deliberate variation allows the neural network to learn robust feature representations that generalize to unseen modem effects, improving detection reliability without proportionally increasing the amount of real measurement data required.
2Measurement precision
If manual feature extraction algorithms are adapted to account for every modem effect, then detection accuracy improves, but the device complexity and development cost increase significantly
Solution Approach 1:
The patent applies the extraction principle by separating modem parasitic effects from the overall frequency response measurement. The system explicitly extracts M clustered modem effect patterns as distinct components from the measured data. This extraction allows the neural network to learn to identify and compensate for these specific interference patterns automatically, achieving high detection accuracy without requiring complex manually-crafted feature extraction algorithms for each modem effect.
Solution Approach 2:
The patent implements self-service by enabling the deep neural network to automatically learn and adapt to various modem effects through the augmented training process. Instead of requiring subject-matter experts to manually configure and update algorithms for each new modem effect, the system allows the neural network to self-adjust its detection capabilities by learning from training samples that include diverse modem parasitic effects. This reduces algorithmic complexity and eliminates the need for manual algorithm adaptation.
3Reliability
If the training set is expanded to include all possible modem effects, then robustness against parasitic frequency distortion improves, but the data collection and processing time increase
Solution Approach 1:
The patent applies partial action by generating a targeted augmented training set that focuses specifically on the M most significant clustered modem effect patterns rather than attempting to include every possible variation. This selective approach provides sufficient robustness against common parasitic distortions while avoiding the excessive time cost of collecting and processing exhaustive real-world data for all conceivable modem effects. The clustering step identifies the essential patterns that need to be covered.
Solution Approach 2:
The patent uses parameter changes by systematically varying the modem parasitic effect parameters in the augmented training samples. For each original training sample, M different versions are created with different clustered modem effects added, creating diversity in the training data through parameter manipulation rather than requiring equivalent diversity through extensive data collection. This accelerates the training process while maintaining robustness.
Data Source
Figure 1~2
Figure 3~5
AI summary
The present invention discloses a method for training a deep neural network classifier for use in detecting impairments in a communication network, the communication network comprising a plurality of communication lines, the method comprising: a) obtaining a measured loop frequency response for the respective communication lines comprised in the communication network, wherein, the loop frequency response is measured by a modem over the respective communication line; b) extracting a modem parasitic effect for the respective communication lines from the measured loop frequency response of the respective communication line; c) clustering the modem parasitic effect of the respective communication lines into a first number M of groups and determining a modem signature for respective ones of the first number M of groups; d) obtaining an original training data set comprising a second number N of loop frequency responses corresponding to a first set of communication lines and associated information on the presence of impairment on the first set of communication lines; e) determining an augmented training data set comprising a third number of loop frequency responses and associated information on the presence of impairment based on the original training data set and the modem signature for respective ones of the first number of groups, wherein, the third number is not greater than N*M; f) training the deep neural network classifier using the augmented training data set.