5:15 PM - 6:45 PM
[HDS09-P15] Building a machine learning model for finite element simulation of damage estimation in truss bridge model
Keywords:Machine Learning, Bridge, Finite Element Method, Surrogate Model, Damage Identification
Maintaining infrastructure structures such as bridges and evacuation shelter buildings is extremely important for developing plans for transporting supplies and evacuating people when a disaster strikes. However, the combined effects of aging infrastructure and declining populations have escalated the costs associated with their maintenance and management. To address this, utilizing machine learning (ML) models has been proposed as a means to reduce these costs [1]. For instance, ML models have been used to detect the crack distribution on the concrete surface [2] [3] [4]. Also, the ML model is used to interpret the damage levels in the bridge members using photo data taken during the maintenance work [5]. While machine learning models are being actively used in the civil engineering field, there is an issue with preparing a sufficient number of high-quality training data.
In light of this context, we designed a method that harnesses extensive, high-quality simulation data to train ML models and have conducted finite element analysis (using COMSOL Multiphysics) for various cases (360,000 cases) to estimate the damage status and displacement distribution on bridge members [6]. In general, the CPU time of ML models is much shorter than that of simulations such as finite element analysis. Therefore, ML models trained with the simulation results created in [6] can output approximate values of the simulation results at high speed, resulting in the construction of a surrogate model corresponding to the finite element analysis for the damage area estimation. In addition, since machine learning models are expected to work without high-performance computing resources, the use of machine learning models enables damage estimation analysis to be conducted during fieldwork, thus contributing to saving manpower.
The objectives of this study are (1) to construct an ML model using the simulation results generated in [6] as training data, and (2) to clarify the accuracy of the constructed ML model. The machine learning model was developed using an automated machine learning library to minimize programming effort and enhance automation. Currently, the model exhibits limited accuracy in predicting damage to vertical members. This limitation is considered to stem from the relatively minor influence that vertical members exert on the displacement distribution within the bridge model.
References
[1] Tokyo Metropolitan Government: Roadmap for the Social Implementation of Digital Twin First Edition, https://info.tokyo-digitaltwin.metro.tokyo.lg.jp/roadmap/ (accessed Feb 1, 2024).
[2] Pang-jo Chun and Atsushi Igo: Crack detection from image using random forest, Journal of Japan Society of Civil Engineers, Ser. F3 (Civil Engineering Informatics), Vol.71, No.2, pp.I_1-I_8, 2015 (in Japanese).
[3] Suguru Yokoyama and Takashi Matsumoto: Development of an automatic detector of concrete surface deteriorations using deep learning and implementation of web system, Journal of Japan Society of Civil Engineers, Ser. A2 (Applied Mechanics (AM)), Vol.73, No.2, pp.I_781-I_789, 2017 (in Japanese).
[4] Kosuke Aoshima, Satoshi Nakano, Kohei Tokunaga, Hideaki Nakamura: Damage detection of concrete surface using anomaly detection method by deep learning, Journal of Japan Society of Civil Engineers, Ser. A2 (Applied Mechanics (AM)), Vol.75, No.2, pp.I_559-I_570, 2019 (in Japanese).
[5] Tatsuya Suzuki and Mayuko Nishio: Application of deep learning to damage level determination of structural members in the bridge inspection, Journal of Japan Society of Civil Engineers, Ser. F3 (Civil Engineering Informatics), Vol.75, No.1, pp.48-59, 2019 (in Japanese).
[6] Hidetaka Saomoto and Takashi Miyamoto: Generating machine learning datasets on damage identification using finite element bridge model, Journal of Japan Society of Civil Engineers, Ser. A1 (Structural Engineering & Earthquake Engineering (SE/EE)), Vol.78, No.4, pp.I_10-I_21, 2022 (in Japanese).
In light of this context, we designed a method that harnesses extensive, high-quality simulation data to train ML models and have conducted finite element analysis (using COMSOL Multiphysics) for various cases (360,000 cases) to estimate the damage status and displacement distribution on bridge members [6]. In general, the CPU time of ML models is much shorter than that of simulations such as finite element analysis. Therefore, ML models trained with the simulation results created in [6] can output approximate values of the simulation results at high speed, resulting in the construction of a surrogate model corresponding to the finite element analysis for the damage area estimation. In addition, since machine learning models are expected to work without high-performance computing resources, the use of machine learning models enables damage estimation analysis to be conducted during fieldwork, thus contributing to saving manpower.
The objectives of this study are (1) to construct an ML model using the simulation results generated in [6] as training data, and (2) to clarify the accuracy of the constructed ML model. The machine learning model was developed using an automated machine learning library to minimize programming effort and enhance automation. Currently, the model exhibits limited accuracy in predicting damage to vertical members. This limitation is considered to stem from the relatively minor influence that vertical members exert on the displacement distribution within the bridge model.
References
[1] Tokyo Metropolitan Government: Roadmap for the Social Implementation of Digital Twin First Edition, https://info.tokyo-digitaltwin.metro.tokyo.lg.jp/roadmap/ (accessed Feb 1, 2024).
[2] Pang-jo Chun and Atsushi Igo: Crack detection from image using random forest, Journal of Japan Society of Civil Engineers, Ser. F3 (Civil Engineering Informatics), Vol.71, No.2, pp.I_1-I_8, 2015 (in Japanese).
[3] Suguru Yokoyama and Takashi Matsumoto: Development of an automatic detector of concrete surface deteriorations using deep learning and implementation of web system, Journal of Japan Society of Civil Engineers, Ser. A2 (Applied Mechanics (AM)), Vol.73, No.2, pp.I_781-I_789, 2017 (in Japanese).
[4] Kosuke Aoshima, Satoshi Nakano, Kohei Tokunaga, Hideaki Nakamura: Damage detection of concrete surface using anomaly detection method by deep learning, Journal of Japan Society of Civil Engineers, Ser. A2 (Applied Mechanics (AM)), Vol.75, No.2, pp.I_559-I_570, 2019 (in Japanese).
[5] Tatsuya Suzuki and Mayuko Nishio: Application of deep learning to damage level determination of structural members in the bridge inspection, Journal of Japan Society of Civil Engineers, Ser. F3 (Civil Engineering Informatics), Vol.75, No.1, pp.48-59, 2019 (in Japanese).
[6] Hidetaka Saomoto and Takashi Miyamoto: Generating machine learning datasets on damage identification using finite element bridge model, Journal of Japan Society of Civil Engineers, Ser. A1 (Structural Engineering & Earthquake Engineering (SE/EE)), Vol.78, No.4, pp.I_10-I_21, 2022 (in Japanese).