劉 ウェン

Liquefaction is a significant challenge in the fields of earthquake risk assessment and soil dynamics, as it has the potential to cause extensive damage to buildings and infrastructure through ground failure. During the 2011 Great East Japan Earthquake, Urayasu City in the Chiba Prefecture experienced severe soil liquefaction, leading to evacuation losses due to the effect of the liquefaction on roads. Therefore, developing quantitative predictions of ground subsidence caused by liquefaction and understanding its contributing factors are imperative in preparing for potential future mega-earthquakes. This research is novel because previous research primarily focused on developing predictive models for determining the presence or absence of liquefaction, and there are few examples available of quantitative liquefaction magnitude after liquefaction has occurred. This research study extracts features from existing datasets and builds a predictive model, supplemented by factor analysis. Using the Cabinet Office of Japan’s Nankai Trough Megathrust Earthquake model, liquefaction-induced ground subsidence was designated as the dependent variable. A gradient-boosted decision-tree (GDBT) prediction model was then developed. Additionally, the Shapley additive explanations (SHAP) method was employed to analyze the contribution of each feature to the prediction results. The study found that the XGBoost model outperformed the LightGBM model in terms of predictive accuracy, with the predicted values closely aligned with the actual measurements, thereby proving its effectiveness in predicting ground subsidence due to liquefaction. Furthermore, it was demonstrated that liquefaction assessments, which were previously challenging, can now be interpreted using SHAP factors. This enables accountable wide-area prediction of liquefaction-induced ground subsidence.

2016年4月熊本地震では，熊本県熊本地方を震源とするMw6.2の地震が発生し，その約16時間後に同地域を震源とするMw7.0の地震が発生した．本研究では，地震による建物の被害状況を効率的にかつ安全に把握する方法として，航空レーザ測量データを深層学習することによって建物被害検出モデルの構築を試みた．本震前後に収集した航空レーザ測量データに対し，深層学習のアルゴリズムの一つである畳み込みニューラルネットワーク（CNN）を適用し，ネットワーク構成を変えながら最良のモデルの検討を行った．その結果，正答率が90%を超えるモデルを構築することができた．

現在，日本の水道では年間2万件を超える漏水・破損事故が発生している．上水道管の漏水は，地上に流れ出す地上漏水と，地上には流れ出さず地下で流れている地下漏水の2種類に大別できる．地上漏水は人目に触れることから発見しやすいものの，地下漏水は漏水の状況を直接目視で確認できないため，早期発見のための技術開発が求められている．そこで本研究では，現在普及が進んでいるスマートメータを活用した水道管路のモニタリングを想定し，管網端部の水圧情報を使用した漏水位置予測に関する検討を行った．漏水シナリオや機械学習手法の異なる6つのモデルを構築し，その予測精度を比較した．水圧変化率，水圧変化量，管種情報を説明変数とし，LightGBMに基づき構築した漏水予測モデルが最も良好な結果を示した．

Road markings, including road lanes and symbolic road markings, can convey abundant guidance information to autonomous driving cars. However, recent works have paid less attention to the recognition of symbolic road markings compared with road lanes. In this study, a road-marking-segmentation dataset named the RMD (Road Marking Dataset) is introduced to compensate for the lack of datasets and the limitations of the existing datasets. Furthermore, we propose a novel multiscale attention-based dilated convolutional neural network (MSA-DCNN) to tackle the proposed RMD. The proposed method employs multiscale attention to merge the weighting outputs of adjacent multiscale inputs, and dilated convolution to capture spatial-context information. The performance analysis shows that the proposed MSA-DCNN yields the best results by combining multiscale attention and dilated convolution. Additionally, the proposed method gains the mIoU of 74.88%, which is a significant improvement over the existing techniques.