TL;DR
This paper introduces a novel two-step calibration method to improve the accuracy of deformation reconstruction using the inverse finite element method (iFEM), particularly when dealing with limited strain measurement data. By separating elastic displacement errors and then expanding the available data using NURBS curves, the method significantly enhances the robustness and accuracy of self-structuring fuzzy networks (SSFNs) for calibration.
What Problem Does This Paper Address?
Structural health monitoring and the development of intelligent structures heavily rely on accurately reconstructing deformation fields from strain measurements. The inverse finite element method (iFEM) is a powerful tool for this, but its accuracy can be compromised by errors in strain measurements and limitations in the calibration process, especially when dealing with small sample sizes of training data. Specifically, the paper highlights two key issues:
- Projection Displacement: In multi-element structures, the overall deformation includes both elastic displacement (strain-dependent) and projection displacement (due to rotation of earlier elements). As you move towards the end of a structure, the elastic displacement becomes a smaller proportion of the total, making direct calibration challenging.
- Small Sample Size: The calibration ability of methods like self-structuring fuzzy networks (SSFNs) is often weak when trained on limited data, leading to reduced reconstruction accuracy.
Methods
The paper proposes a "two-step calibration method" to address these issues:
- Step 1: Deformation Separation and Error Distribution:
- The elastic displacement is first separated from the total displacement using iFEM principles.
- Errors derived from measured and reconstructed displacements are then distributed to the degrees of freedom (DOFs) of nodes. This involves calculating elastic displacement errors and then distributing these errors to the six kinematic variables (translations and rotations) of the nodes.
- A root mean square error (RMSE) is used to evaluate the accuracy of this calibrated displacement.
- Step 2: Sample Extension and SSFN Calibration:
- Sample Expansion: To overcome the limitation of small sample sizes, the Non-uniform Rational B-spline (NURBS) curve method is employed. The DOFs of nodes, identified in Step 1, are used as knots to generate NURBS curves. This process interpolates and expands the limited dataset, creating a richer set of training data.
- SSFN Calibration: The expanded dataset is then used to train a Self-Structuring Fuzzy Network (SSFN). The SSFN learns the relationship between the measured strain and the corrected displacement (obtained after Step 1). The SSFN algorithm iteratively adjusts its membership functions and rules to minimize system error (RMSE) and ensure completeness criteria are met. This adaptive process allows the SSFN to become more robust and accurate.
Datasets
- Experimental Data: The study utilizes experimental data from a three-element frame structure. Strain data was collected using Fiber Bragg Grating (FBG) sensors, and displacement data was captured by a three-dimensional dynamic displacement measurement instrument (NDI Optrotrak Certus).
- Training and Verification Data: Data from 8 working conditions (Table 3, weights 1.5 kg to 12 kg) were used for training the SSFN. Data from another 8 working conditions (Table 3, weights 6 kg to 11.5 kg) were used to verify the calibration accuracy.
- Data Expansion: The NURBS method was used to expand the initial 8 sets of training data into 251 sets, significantly increasing the sample size for SSFN training.
Key Insights and Ideas
- Error Decomposition is Crucial: Separating elastic displacement and then distributing the errors to nodal DOFs is a fundamental step that isolates the components needing calibration.
- NURBS for Data Augmentation: Using NURBS curves to interpolate and expand sparse strain and error data is an effective strategy to overcome the "small sample size" problem, which is a common bottleneck in machine learning-based calibration.
- SSFN Adaptability: The self-structuring nature of the SSFN, with its ability to add membership functions and adapt rules, makes it well-suited for learning complex relationships and improving calibration accuracy iteratively.
- Element-wise vs. Whole Structure Calibration: The paper suggests that both individual element calibration and whole-structure calibration yield improvements, but the two-step method is particularly effective.
Key Results
The proposed two-step calibration method demonstrated significant improvements in deformation reconstruction accuracy:
- Reduced Errors:
- For the first element, the maximum RMSE in the x-direction was reduced from 0.35 mm to 0.18 mm.
- The maximum RRMSE in the y-direction was reduced from 12.62% to 5.57%.
- The maximum RRMSE in the z-direction was reduced from 13.77% to 7.68%.
- Similar reductions were observed for the second element.
- Improved Accuracy with Expanded Data: Figure 8 clearly shows that using the "extended calibration" (with NURBS data expansion) resulted in significantly lower displacement errors across all directions (x, y, and z) compared to "initial calibration" (using only the original small dataset).
- Robustness: The method provides an effective solution for scenarios with limited strain measurement data, making iFEM more practical for real-world applications.
Limitations and Future Work
- Limitations:
- The paper acknowledges that sensor installation and measurement errors are inherent and affect accuracy.
- While the method improves accuracy, it doesn't eliminate errors entirely.
- Future Work:
- The paper implies that further optimization of the SSFN structure and training parameters could lead to even better results.
- The effectiveness of the method on more complex structures or under dynamic loading conditions could be explored.
- Investigating the impact of different types of sensor errors on the calibration process might be beneficial.
Summary updated