In this study, I developed a novel GNN-LSTM-based fusion model framework. It can predict the nonlinear responses history of acceleration, velocity, and displacement for each floor of any SMRF structure between 4 and 7 stories in height.

In addition, two LSTM-optimized learning strategies, including a packing padded sequences(PPS) strategy and a sequences compression(SC) strategy, are also proposed in this project. The experimental results of this study demonstrated that the PPS learning strategy successfully reduced the noise in the sequence data, improving the stability of the training process and the model’s predictive ability, and the SC learning strategy significantly short- ened the training time with only a minor reduction in model accuracy while retaining the complete information of the original sequences in the compressed sequences.

Finally, the two-dimension graph embedding empowered by t-sne algorithm and graph attention distribution within the GAT network were visualized. The graph embeddings of the structures were highly correlated with their first modal periods. Moreover, the distribution of attention weight on each floor shows symmetry, implying that the GAT model effectively captures the inherent symmetries in the geometry and vibration shape of the building. This observation aligns well with the rigid diaphragm assumption applied to each floor of the numerical structural models. The combined results suggest that the message-passing mechanism acquired by the GAT during training aligns with physical principles.

• Linux, CUDA>=11.3
• Python>=3.9.7
• pyg>=2.0.3
• pytorch>=1.10.2

Anaconda is recommended to use to create a conda environment:

# Optional. If it doesn't work, just make sure you have installed the above packages.
conda env create -f ./environment.yml

bash download.sh

Following the checkData.ipynb to see the detail of the dataset

In this repository, I provide the best GAT model for inference. Follow the steps bellow to predict Acceleration, Velocity, Displacement dataset.

• Acceleration Dataset

python inference.py --output_dir ./Inference/Acceleration --response_type Acceleration

• Velocity Dataset

python inference.py --output_dir ./Inference/Velocity --response_type Velocity

• Displacement Dataset

python inference.py --output_dir ./Inference/Displacement --response_type Displacement

1. create folder:

mkdir -p ./Results/GCN_LSTM
mkdir -p ./Results/GAT_LSTM

2. Go to train_GCN_LSTM.arg.py or train_GAT_LSTM.arg.py. Then set the training enviornment and learning target
   • --pack_mode: PPS strategy
   • --compression_rate: SC strategy. Feel free to try 10, 20, 40
   • --response_type: Acceleration or Velocity or Displacement
   • --n_layers: # of LSTM layer
3. Train the model

# use GCN as aggregation function
python train_GCN_LSTM.py

# use GAT as aggregation function
python train_GAT_LSTM.py

4. Test the model

# use GCN as aggregation function
python test_GCN_LSTM.py --output_dir <folder path of target GCN_LSTM model>

# use GAT as aggregation function
python test_GAT_LSTM.py --output_dir <folder path of target GAT_LSTM model>