Prediction of the Most Fire-Sensitive Point in Building Structures with Differentiable Agents for Thermal Simulators

dataset

• Google Drive
  • drs3v6 is the unfiltered labeled dataset. It includes 3,000 structures and their gravity simulation results as well as 30 fire case simulation for each structure.
  • drs4v2 is the unfiltered unlabeled dataset. It includes 30,000 structures and their gravity simulation results.

requirements

• to generate dataset, you may need:
  • opensees: https://pypi.org/project/opensees/
  • pandas
• to run or train the neural networks, you may need:
  • pytorch
  • pytorch_geometric
• before running, rename the config-demo.json as config.json and edit the basedir as the absolute path of your working directory.

Dataset Generation

• Generate geometry by geometry_gen.py

• OpenSees Simulation by ops_simulation.py

  • for the saved .pt file, it is supposed to contain the following information

  attribute_name	meannning	comment
  x	node features: the coordinates and indices: $[x_i, y_i, z_i, h_i]$	dim: (N_node, 4)
  y	drift ratio for each node	dim: (N_node, 1)
  edge_index	the starting and ending nodes' number of each edge	dim: (2, N_edge)
  edge_attr	edge features: ['Es', 'Fy', 'b', 'length', 'h', 'gravity_param', 'direction_x', 'direction_y', 'direction_z']	dim: (N_edge, 9)
  model_number	model number	scaler, normal int, not torch tensor
  fire_number	fire number (corresponding to file)	scaler, normal int, not torch tensor
  fire_point	coordinates of the fire: $[x_0, y_0, z_0, h_0]$	dim: (4)
  fire_room_index	fire room indices: $[w_i, d_i, h_i]$	dim: (3)
  unit_lengths	unit width, unit depth, unit height, in mm	dim: (3)
  node_indices	node indices: $[w_i, d_i, h_i]$	dim: (N_node, 3)
  drift_ratio_node_ranking	ranking of each node by node's drift ratio, descending	dim: (N_node, 1)
  drift_ratio_ranking	ranking of the maximum drift ratio of each graph	scaler, torch.long
  max_drift_ratio	max of y	scaler tensor
  max_level	max of x[:,3]	scaler tensor

• Then in filter.py, the dataset was filtered by the maximum interstory drift ratio under gravity (also filter ou those obviously unreliable data).

• The scripts in data_generation folder now only present the generation of labeled dataset. Functions to generate unlabeled is a subset of that to generate labeled dataset.

Maximum Drift Ratio (MIDR) Predictor

• As the filenames indicate, there are strawman1, strawman2 and proposed scripts, they are for trainning the NN models
• Then the evaluation scripts are to evaluate these models together.

Most Fire-Sensitive Point (MFSP) Predictor

• mfsp_train.py to train a MFSP predictor and mfsp_eval.py to evaluate it.