Bridge Internal Response Reconstruction via Hybrid Physics-Data Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for reconstructing internal responses of bridges face challenges due to the difficulty in establishing accurate mechanism models and the reliance on incomplete data, leading to inaccurate predictions.
Innovation Solution
A dual physically-driven and data-driven method that uses sensors to collect acceleration responses, constructs a finite element model, and employs a convolutional neural network (CNN) to predict internal node responses, incorporating physical constraints and loss functions to improve accuracy and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If physics/formula-based modeling is used to reconstruct internal response, then the method can work with limited data, but the prediction accuracy deteriorates due to difficulties in establishing complete and accurate mechanism models
Solution Approach 1:
The patent merges physics-based modeling and data-driven machine learning into a hybrid framework. The physics model provides structural constraints and relationships, while the neural network learns from data to compensate for model inaccuracies. This combination allows the system to work with limited data while maintaining high prediction accuracy for internal responses.
Solution Approach 2:
The patent creates a composite modeling approach by integrating two different modeling paradigms (physics-based and data-driven) into a unified framework. This composite model leverages the strengths of both approaches: the interpretability and constraint satisfaction of physics models, and the adaptive learning capability of neural networks.
2Productivity
If data-driven machine learning is used to reconstruct internal response, then model training is efficient with vast data, but prediction accuracy deteriorates due to incomplete data coverage and inability to generalize beyond training data
Solution Approach 1:
The patent transforms the pure data-driven approach by incorporating physics-based parameters and constraints into the loss function and model architecture. This allows the neural network to learn from data efficiently while being guided by physical laws, improving generalization capability beyond the training data distribution.
Solution Approach 2:
The patent implements feedback mechanisms where the physics model provides constraints and guidance to the neural network during training and prediction. The loss function incorporates physics-based error terms that continuously guide the learning process toward physically consistent solutions, improving accuracy even for unseen data.
3Measurement precision
If sensors are placed at internal positions to directly measure response, then measurement accuracy is improved, but device complexity and measurement difficulty increase due to accessibility issues
Solution Approach 1:
The patent uses a neural network as an intermediary to infer internal responses from external measurements. Instead of placing sensors directly at internal positions, the system uses easily accessible external sensors and lets the trained neural network compute the internal responses, avoiding the complexity of internal sensor installation while maintaining measurement accuracy.
Solution Approach 2:
The patent creates a virtual copy of the internal response through the trained neural network model. Rather than physically measuring internal responses with sensors, the system learns a mapping from external to internal responses and uses this learned model to predict internal states, effectively copying the measurement capability without physical intrusion.
Data Source
AI summary
Disclosed is a dual physically-driven and data-driven method for reconstructing internal response of a bridge. The method includes: obtaining acceleration response by an acceleration sensor under an action of an unknown load of the bridge; embedding a physical logic into a neural network based on a frequency response function; putting a physical formula and corresponding boundary conditions and initial conditions into a loss function as penalty terms, and limiting a space of a feasible solution accordingly; and training a neural network model, and predicting acceleration of an unknown point by inputting an acceleration response set of a known point obtained by the sensor into the network. The formula is solved by converting direct solving of a control formula into optimization of the loss function, such that the problems that the internal response of the bridge is difficult to measure and excessively depends measured data can be effectively solved, and accuracy and robustness of internal response prediction of the bridge can be improved.