研究発表

水平2方向で剛性が異なる柱は幾何学的非線形性により水平変位でねじりトルクが生じる（Q–Δ効果）．2方向の並進モードの固有振動数の和または差がねじりモードの固有振動数と一致すると内部共振でねじれモードが励起する（Q–Δ共振）．本研究ではまず多層せん断型偏心建物のQ–Δ効果を定式化した．そして，Q–Δ共振条件を満足する階数20～60，偏心率0～0.15の一軸偏心建物モデルに長周期地震動を入力する時刻歴応答解析を行い，階数と偏心率が増大すると最大水平加速度が線形解析結果と乖離することを明らかにした．

The P–Δ effect is well known, in which an overturning moment emerges due to geometric nonlinearity when a gravity-loaded structure undergoes a large deformation. Similarly, when a column undergoes a large horizontal displacement, a torsional torque emerges to preserve the state of equilibrium. The speaker named this effect “Q–Δ effect.” Although the effect has not been considered in structural design, he has shown that periodical nonlinear restoring torque of the Q–Δ effect can produce a non-negligible torsional response in a structure during earthquakes owing to its internal resonance, named “Q–Δ resonance.” This talk will first explain the theory of the Q–Δ effect and the mechanism of the Q–Δ resonance. Then, he will present the results of shaking-table experiments using a reduced specimen and finite element analysis of a full-scale model of high-rise building under long period ground motions. This resonance phenomenon is a blind spot in seismic design, and given the mechanism of its occurrence, it may need to be considered in wind-resistant design as well. Researchers and engineers are expected to listen to this talk to widely promote proactive countermeasures.

Structural health monitoring (SHM) has attracted significant attention as a method for rapidly and efficiently assessing the damage state of buildings immediately after an earthquake. Application of SHM to houses could facilitate prompt evaluations of the risk of collapse from subsequent earthquakes, thereby reducing casualties caused by aftershocks and other seismic events. This study proposes a damage classification method for wooden buildings based on acceleration responses, using long short-term memory (LSTM), which is well-suited for capturing the long-term dependencies of time-series data. The analysis focuses on a specimen from the dataset of “Experimental Validation of Design Methods for Three-Story Wooden Frame Structures,” published by the National Research Institute for Earth Science and Disaster Resilience. A building model was constructed and analyzed using “wallstat,” a numerical analysis software for wooden structures. To generate training data for a damage classifier, response analysis of the model was conducted using 2,000 simulated seismic ground motions consisting of equal number of pulse and non-pulse types ground motions. The acceleration response of the first and second floors were converted input data for the damage classifier so that each input segment contains one second of acceleration responses. To prevent data imbalance caused by similar input segments with small variations in acceleration, a “lag” was introduced, ensuring that each input segment shifts by a specific number of data points relative to the previous one. In SHM, misclassifying an unsafe condition as safe poses a severe risk to building users, as users may overlook potential building collapse hazards leading to fatalities. To address this, training data near the threshold between safe and unsafe models were excluded from safe model datasets to reduce the possibility of misclassification. Three hyperparameters of the damage classifier, the number of LSTM units, the number of nodes in the first fully connected layer, and the learning rate, were optimized using the response surface method. This optimization process identified a combination of hyperparameter values that achieved high classification accuracy. A two-level central composite design was employed for the design of experiments. The proposed damage classifier demonstrated a high accuracy of 92.4%, marking a significant step toward practical implementation. This study highlights the effectiveness of LSTM for damage classification in wooden buildings and underscores the utility of the response surface method in optimizing machine learning models.

High-rise buildings can be significantly affected by external excitations such as earthquakes and wind and tuned mass dampers (TMDs) are widely used to address this issue. Conventional TMDs can only suppress a single vibration mode. Therefore, a bi-directional tuned mass damper (BTMD) has been proposed for buildings with different natural frequencies in the two horizontal directions. However, adjusting the natural frequency of a BTMD to accommodate changes and uncertainty in the natural frequencies of a building is challenging. This study proposes a novel BTMD named the hybrid tuned mass damper with rigid body swinging and horizontal spring vibration (RH-TMD) as a solution, which combines rolling motion of an auxiliary mass on a rail induced by horizontal spring and swinging motion of the rail and mass in the orthogonal direction. The natural periods can be adjusted by modifying the weight of the mass in rolling motion and the length of the pendulum in swinging motion. A rotary damper with an adjustable damping coefficient is adopted as a damping mechanism. First, numerical analysis using a 50-story model are conducted to examine the performance of the RH-TMD. The vibration characteristics, natural periods and damping coefficients of the RH-TMD in the two horizontal directions are optimized to minimize the root mean square response to white noise ground acceleration input. Its performance is compared with a conventional (simple) TMD with a natural period and damping coefficient optimized only in a single direction with respect to absolute acceleration and story drift responses to seismic ground motion. In addition, shaking table experiments are conducted using a small-scale single-layer specimen with a similarity ratio of 1:100, equipped with either an RH-TMD or a simple TMD to compare the responses to white noise ground acceleration and simulated seismic motions. The results of the numerical analysis and shaking table experiments validate the excellent performance of the RH-TMD in reducing absolute acceleration and story drift responses in the two horizontal directions.

建物の地震応答データによって描かれる復元力－層間変形角関係曲線（履歴ループ）は，建物の健全度合いによって形状が変化する．本研究ではこの性質を活用し，3次元の畳み込みニューラルネットワーク（CNN）を用いて時間軸方向を考慮した履歴ループ画像をもとに将来の地震動による倒壊の危険性を判別する方法を提案した．提案手法と既往研究の2次元CNNを用いた手法と正解率および危険モデルの再現率を比較し，有効性を検証した．その結果，提案手法は作成する画像サイズや時間軸方向の分割数を適切に調整することで2次元CNNよりも正解率やロバスト性を向上できることがわかった．

This study investigates the design of a linear quadratic regulator (LQR) for a seismic control device to improve the resilience of a medical facility. The functionality of a medical facility is quantified based on a fault tree. A frame-work to evaluate the recovery curve of a medical facility is constructed based on the fault tree and recovery-time probability models for damaged medical devices. Furthermore, the approach to design parameters for an LQR control system is investigated to minimize the time required to recover up to the de-fined functionality level after the onset of an earthquake. A case analysis is conducted using a model of a ten-story hospital building with a control device installed on its top floor, and the resulting expected recovery time is evaluated. It was confirmed that the proposed method could reduce the recovery time even for the different input ground motion from that used in a design stage.

The Data Science Working Group of the Numerical Shaking Table Committee of the National Research Institute for Earth Science and Disaster Resilience is working on the development of structural health monitoring technology using artificial intelligence. In the 2016 Kumamoto earthquakes, some residents returned to their wooden houses that had been severely damaged by the strong ground motion, and the houses collapsed due to the following main shock causing human casualties. This kind of disaster should be reduced if they could understand the remaining performance of the house using a structural health monitoring system. Deep neural networks have been reported to achieve high accuracy in various fields in pattern recognition and evaluation of regression equations with a large amount of training data. The authors have also taken the same approach of preparing big data for training by conducting time history response analysis with many input earthquake motions and damage patterns in advance. However, the stiffness and strength of the members assumed at the design stage do not necessarily match the actual values of the completed building. Epistemic uncertainty exists, and it is desirable to recognize damage patterns and evaluate residual performance with features that are robust to such variations. In this study, we propose a method of generating training data with variability based on the results of elemental experiments conducted in conjunction with full-scale wooden building experiments in the E-Defense experiment, and constructing training data by providing input seismic motions of appropriate intensity so that the number of training data with and without damage is balanced.

Structural health monitoring is effective for quick and efficient recognition of earthquake damage to a building. However, its applications to wooden buildings are still limited. One of the challenges in predicting the response of a wooden building is that the variance in its material property considerably affect its response. To handle this characteristic, we propose a method using global sensitivity analysis based on polynomial chaos expansion to select dominant material parameters in the variability of a quantity of interest. The method enables us to reduce the computational cost in multiple time-history response analysis, which is necessary to construct a damage classifier that takes the variance of material property into account by machine learning.

Structural health monitoring (SHM) is useful for quick and efficient damage recognition of a building under seismic excitation. It has the potential to replace the emergency safety evaluation, which is conducted after major earthquakes in purpose to prevent a secondary disaster from aftershocks. SHM systems are popular for large-scale civil structures in the world, but their application to low-rise buildings such as wooden houses are still limited. One of the challenges in predicting the response of a wooden structure is that the effect of variance in its material property is considerably significant. To solve this problem, we propose a method using global sensitivity analysis with Sobol’ indices based on polynomial chaos expansion (PCE) to find dependent parameters. By considering only material parameters with a high impact on the variability of the quality of interest, we can reduce computational cost in time history response analyses. PCE is a popular technique in the field of uncertainty quantification, which enables us to model the response of a dynamical system with uncertainty. PCE is also known to be highly compatible with global sensitivity analysis, which allows to identify the dependent input variables that contribute most to the variability of the computational model response. Based on the results of a full-scale three-story wooden building experiments and its member experiments conducted by E-Defense of National Research Institute for Earth Science and Disaster Resilience, we performed three-dimensional structural response analyses of a model of the specimen using a numerical collapsing simulation program, wallstat. As input ground motions, we used design ground motion specified by the Building Standard Law with an amplification factor of 120%. Multiple models were prepared with a varied value of material properties to consider the epistemic uncertainty. As a result, the sensitivity analysis clarified that the variance of Young's modulus in columns and beams had a major contribution to the variance of the maximum inter-story drift angle of the model.