TL;DR
This paper introduces a modified Electromagnetism-Like Mechanism (EM) algorithm designed to improve the accuracy and efficiency of structural model identification using modal data. The enhanced algorithm demonstrates superior performance compared to the original EM and Particle Swarm Optimization (PSO) algorithms on benchmark functions and successfully identifies parameters for both numerical and experimental structural models, even with noisy and incomplete data.
What Problem Does This Paper Address?
Structural system identification is crucial for applications like structural health monitoring and damage detection. It involves finding optimal model parameters (like mass and stiffness) that best match the system's measured response. Traditional methods often require gradient information and good initial guesses, which can be problematic with incomplete or noisy data. While heuristic algorithms like EM have shown promise, the original EM algorithm has limitations in its local search, charge/force calculations, and particle movement rules, hindering its accuracy and convergence speed. This paper aims to overcome these limitations by proposing a refined EM algorithm specifically for structural model identification.
Methods
The core of this paper is the development of a modified Electromagnetism-Like Mechanism (EM) algorithm. The authors introduce several key improvements to the original EM algorithm:
- New Local Search Method: Instead of a simple linear search, a more sophisticated method is proposed that branches from the current best particles within a certain radius, generating sub-points to explore the solution space more effectively.
- Simplified Charge Calculation: The original charge calculation formula, which involved complex exponential functions, is replaced with a simpler linear function. This aims to improve computational efficiency, especially in later iterations.
- New Particle Movement and Updating Rules: Particles are sorted, and their movement is governed by the attraction to the next best particle and repulsion from the next worst particle. A new rule is introduced to find an alternative location if the current movement doesn't lead to a better solution, using the center of four surrounding particles.
- Enhanced Exploration: In the early iterations, a portion of the worst particles are randomly regenerated to improve exploration and help the algorithm escape local minima.
The modified EM algorithm is then applied to structural model identification, where the objective is to minimize the discrepancy between predicted and measured modal parameters (frequencies and mode shapes).
Datasets
The paper evaluates the modified EM algorithm using:
- Six well-known benchmark functions: These are standard test functions used to assess the performance of optimization algorithms. The paper provides details on their names, search ranges, and accepted accuracy.
- A numerical truss model: A 21-bar planar truss model is used to simulate structural parameter identification under various conditions of noise contamination (1% to 10% for frequency, 5% to 10% for mode shape) and information incompleteness (varying numbers of measurements and modes).
- An experimental shear-building model: A three-story shear-type structure is used to validate the algorithm's performance on real-world data.
Key Insights and Ideas
- The power of refined heuristics: The study demonstrates that even well-established heuristic algorithms like EM can be significantly improved with targeted modifications, leading to better performance in complex optimization tasks.
- Balancing exploration and exploitation: The new local search and particle movement strategies seem to strike a better balance between exploring the search space and exploiting promising regions, leading to faster convergence and higher accuracy.
- Robustness to noise and incompleteness: The modified EM algorithm shows remarkable resilience to noisy modal data and incomplete measurements, which are common challenges in real-world structural identification problems. This is a critical finding for practical applications.
- Information incompleteness is a major challenge: The numerical truss model results highlight that while noise is a factor, the lack of measurement data (fewer sensors or fewer modes) has a more significant impact on the uncertainty of the identified parameters.
Key Results
- Benchmark Functions: The modified EM algorithm outperformed the original EM algorithm and the PSO algorithm in terms of accuracy on six benchmark functions. It also achieved a significantly faster convergence rate than both. The average computation time was comparable to PSO and much faster than the original EM.
- Numerical Truss Model: The modified EM algorithm successfully identified the stiffness parameters for the truss model, with results agreeing well with preset values. Even with up to 10% frequency noise, 10% mode shape noise, and incomplete measurements (7 measurements and 5 modes), the identification was effective. The standard deviations of the identified parameters increased with higher noise and incompleteness, particularly for measurement incompleteness.
- Experimental Shear-Building Model: The algorithm accurately identified the stiffness parameters of the three-story shear-building model. The updated modal frequencies and mode shapes were found to be closer to the experimental results than those of the nominal model, validating the algorithm's effectiveness.
Limitations and Future Work
The authors acknowledge several areas for future research:
- Hybridization with other algorithms: For heuristic algorithms, their stochastic nature can lead to some randomness in results. Combining the modified EM with traditional algorithms that provide gradient information or other heuristic algorithms could further eliminate randomness and improve robustness.
- Quantifying uncertainties: The study identified uncertainties in the identified model parameters due to noise, model errors, and incomplete information. Developing methods to quantify these uncertainties is an important next step.
- Addressing model unidentifiability: For cases with very limited information, models might become unidentifiable (multiple solutions). Bayesian inference methods are suggested as a potential approach to handle such situations.
- Sensor placement optimization: The number and arrangement of sensors significantly impact identification accuracy. Optimizing sensor placement is crucial for obtaining sufficient and reliable data.
- Expanding application fields: The authors suggest exploring the algorithm's use in other optimization problems and further developing hybrid optimization algorithms.
Summary updated