Automatic Concrete Slump Prediction

This repository contains the code and sample data for the paper “Automatic concrete slump prediction of concrete batching plant by deep learning”

The work uses deep convolutional neural networks (e.g., VGG, ResNet, DenseNet, MobileNet, ShuffleNet) to analyze images from the discharge hopper of a ready-mix concrete batching plant and classify them into discrete slump classes. This enables automatic, continuous monitoring of concrete quality directly at the plant.

Data Availability

The full industrial dataset used in the paper cannot be released due to company and project confidentiality policies.
To help users understand and test the pipeline, we include only small sample image sets under sample_data/. These samples illustrate the data format and typical visual appearance of different slump classes but are not sufficient to reproduce the full quantitative results reported in the paper.

Trained model weights corresponding to the paper’s experiments will be released soon.

Environment Setup (Conda, Python 3.9)

We recommend using a Conda environment with Python 3.9 and installing the dependencies from requirements.txt.

1. Create a Conda environment

   conda create -n concrete_slump_prediction python=3.9

2. Activate the environment

   conda activate concrete_slump_prediction

3. Install dependencies

   pip install -r requirements.txt

Note: The requirements.txt file was exported from a working environment for the experiments. Versions and CUDA support (e.g., for torch, torchvision) may need minor adjustments depending on your local system and GPU drivers.

Basic Usage

• Training (example):

  python train.py \
      --model_name vgg

  • --model_name: Backbone architecture (vgg, resnet, densenet, mobilenet, shufflenet).

• Testing / Evaluation:

  The scripts test.py, test_multiple.py, and test_temporal.py show different evaluation scenarios (e.g., multiple models, temporal splits). Please refer to their arguments and adapt the dataset paths to your local data.

• Visualization (CAM):

  Use CAM.py together with trained checkpoints to visualize discriminative regions that drive the model’s predictions for different slump classes.

Citation

If you use this code, sample data, or ideas from the paper in your research, please cite:

@article{Idrees2024,
  title   = {Automatic concrete slump prediction of concrete batching plant by deep learning},
  author  = {Idrees, Sarmad and Nugraha, Joshua Agung and Tahir, Shafaat and Choi, Kichang and Choi, Jongeun and Ryu, Deug-Hyun and Kim, Jung-Hoon},
  journal = {Developments in the Built Environment},
  volume  = {18},
  pages   = {100474},
  year    = {2024},
  publisher = {Elsevier},
  doi     = {10.1016/j.dibe.2024.100474}
}