Airborne Flight Control with Consensus Learning Against Parameter Drift
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Adaptive neural networks in safety-critical systems, such as flight control computers, suffer from parameter drifting due to asynchronous learning in redundant architectures, leading to misaligned weights and potential system instability.
Innovation Solution
A computer-implemented method for adaptive control in airborne flying control systems uses a consensus learning mechanism to align the weights of artificial neural networks across multiple computers by averaging their outputs and incorporating a consensus term in the weight update law, ensuring stability and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If online learning is implemented in redundant flight control computers, then the system adaptability and ability to handle unknown dynamics improve, but parameter drifting occurs causing weight misalignment between computers
Solution Approach 1:
The patent implements a feedback mechanism where each computer's weight update is influenced by the average output of all computers. The modified learning law includes a feedback term that compares individual computer output with the system average, automatically correcting deviations and preventing parameter drifting while maintaining online adaptability
Solution Approach 2:
The patent merges the learning processes of multiple redundant computers by introducing a consensus term based on the average output of all computers. This combining approach ensures that all computers converge to the same weight values, maintaining parameter consistency across the redundant system while preserving online learning capabilities
2Productivity
If asynchronous learning is allowed in redundant computers, then each computer can independently adapt to changing conditions, but weight misalignment and system instability occur
Solution Approach 1:
The feedback mechanism continuously monitors each computer's output deviation from the average and applies corrective adjustments to the weight update process, ensuring that asynchronous learning does not lead to weight misalignment while preserving the benefits of parallel independent adaptation
Solution Approach 2:
The patent creates an equipotential state among all redundant computers by using the average output as a reference level. The modified learning law ensures all computers converge to this common reference, eliminating weight misalignment while allowing simultaneous independent learning operations
3Speed
If local learning is performed in each redundant computer, then response time improves, but parameter drifting leads to voting failures in safety-critical systems
Solution Approach 1:
The feedback mechanism ensures that local learning in each computer maintains consistency with the overall system behavior by continuously adjusting weights based on the average output, preventing parameter drifting that would cause voting failures while preserving fast local response capabilities
Solution Approach 2:
The patent enforces homogeneity among all redundant computers by using the consensus-based learning law that drives all computers to produce identical weight values and outputs, ensuring voting reliability while maintaining the speed benefits of distributed local learning
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
Computer implemented method for adaptive control in an airborne flying control system, the method being indicated for avoiding the drifting phenomenon occurring when different computers (1) with artificial neural networks (ANNs) are used to learn and adapt the algorithm for determining a particular aircraft related variable during a flight, the method comprising, at each time step of determination of the aircraft related variable, feeding each computer (1) with the mean value of the aircraft related variable calculated by all of the computers (1) of the airborne flying control system at the previous time step.