Optical Launch Power Neural Network Updates Without Traffic Interruptions

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

In optical fiber networks with multiple neural networks, training each network requires significant data transfer, leading to high bandwidth consumption and potential traffic interruptions.

Innovation Solution

A distributed architecture with a master node and slave nodes, where the master node trains a neural network to compute optical launch power and sends updated parameters to slave nodes via the optical fiber network without interrupting traffic, using a centralized training and analytics entity to manage parameter updates.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If neural networks are trained at each slave node, then each node can independently compute optimal optical launch power, but bandwidth consumption increases significantly and traffic interruptions occur

Engineering Contradiction:
Improveoptical launch power computation accuracyVSAvoidbandwidth consumption
Core Design Contradiction:
Measurement precisionVSLoss of energy

Solution Approach 1:

The patent extracts the training process from distributed slave nodes and concentrates it at a central master node. Only trained parameters (weights and biases) are transmitted to slave nodes, rather than transmitting full training datasets. This extraction of the computationally intensive training operation from multiple nodes resolves the bandwidth consumption problem while maintaining computation accuracy at slave nodes through parameter updates.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The neural network training system is segmented into two distinct functional parts: a master node responsible for training and parameter generation, and slave nodes responsible for inference and local optimization. This segmentation allows the training workload and data requirements to be isolated at the master node, while slave nodes receive only lightweight parameter updates, thereby reducing overall bandwidth consumption.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If neural network parameters are updated across the network, then optical launch power computation is improved, but network traffic may be interrupted

Engineering Contradiction:
Improveoptical launch power computation accuracyVSAvoidnetwork traffic continuity
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The master node performs training and generates updated parameters in advance, before slave nodes need to apply them. This preliminary action at the master node allows slave nodes to receive pre-computed parameters without needing to interrupt their operations for real-time training, thus maintaining traffic continuity while improving computation accuracy.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

Instead of transmitting and executing full training processes at each slave node, the patent creates and transmits copies of trained parameters (weights and biases) from the master node to slave nodes. These parameter copies enable slave nodes to improve their local computation accuracy without executing disruptive training routines, thereby maintaining network traffic continuity.

Inventive Principle:
Principle #26Copying

3Measurement precision

If all neural networks are trained independently, then each node has optimal local performance, but device complexity and training time increase

Engineering Contradiction:
Improveneural network performanceVSAvoidtraining system complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent merges the training functions of multiple slave nodes into a single centralized training process at the master node. This consolidation reduces device complexity by eliminating redundant training infrastructure at each slave node, while still enabling all nodes to achieve optimal performance through shared parameter updates from the centralized training system.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4422090B1Method and system for updating parameters of neural networks in an optical fiber network
Publication Date: 2026.03.11 NOKIA SOLUTIONS & NETWORKS OY
  • EP4422090B1 patent drawingFigure 1
  • EP4422090B1 patent drawingFigure 2
  • EP4422090B1 patent drawingFigure 3A~3B

AI summary

An optical fiber network comprises network nodes including a master node and slave nodes of the master node. The master node has electronic processing hardware configured to obtain parameters of a neural network to compute an optical launch power for an optical fiber link. The master node is configured to send the parameters via the optical fiber network, to some of the slave nodes. Each one of the slave nodes includes electronic processing hardware to support a neural network for computing an optical launch power of one or more fiber links ending at the one of the slave nodes. The recipient slave nodes are configured to update the neural network supported therein in response to receiving the parameters sent by the master node.