Interpretable Prognostics with Concept Bottleneck Models

Florent Forest1    Olga Fink1
1 Intelligent Maintenance and Operations Systems (IMOS), EPFL, Lausanne, Switzerland

Source code for the implementation of the paper "Interpretable Prognostics with Concept Bottleneck Models".

Getting started

Install the dependencies:

pip install -r requirements.txt

Training the RUL prediction models

Train the RUL prediction models on the N-CMAPSS datasets using following command:

python cem/cmapss_train.py \
    --data-path /path/to/data/N-CMAPSS/ \
    --output-dir output/output-dir \
    --batch-size 256 \
    --model-type MODEL \
    --concepts "['Fan', 'HPC', 'HPT', 'LPT']" \  # Set the desired degradation concepts
    --train-ds "['01', '04', '05', '07']" \
    --train-units "[1, 2, 3, 4, 5, 6]" \
    --test-ds "['01', '04', '05', '07']" \
    --test-units "[7, 8, 9, 10]" \
    --downsample 10 \
    --RUL flat \  # 'flat': constant RUL before fault onset. 'linear': linear RUL from first cycle.
    [--emb-size 16 \]  # for CEM only
    [--boolean_cbm \]  # for Boolean CBM only
    [--extra-dims 60 \]  # for Hybrid CBM only

where MODEL can be one of the following model types:

Model	MODEL	Description	Interpretablility
CNN	cnn	CNN regressor	❌ (black-box)
CNN+CLS	cnn_cls	CNN regressor + classification head	❌ (black-box)
CBM	cnn_cbm	Concept Bottleneck Model (boolean, fuzzy or hybrid depending on parameters)	✅
CEM	cnn_cem	Concept Embedding Model	✅

Evaluation

Evaluate trained RUL prediction models using following command:

python cem/cmapss_eval.py \
    --checkpoint output/output-dir/checkpoints/last.ckpt \  # Path to trained weights
    --data-path /path/to/data/N-CMAPSS/ \
    --output-dir output/ \
    --batch-size 256 \
    --model-type MODEL \
    --concepts "['Fan', 'HPC', 'HPT', 'LPT']" \  # Set the desired degradation concepts
    --train-ds "['01', '04', '05', '07']" \
    --train-units "[1, 2, 3, 4, 5, 6]" \
    --test-ds "['01', '04', '05', '07']" \
    --test-units "[7, 8, 9, 10]" \
    --downsample 10 \
    --RUL flat \  # 'flat': constant RUL before fault onset. 'linear': linear RUL from first cycle.
    [--emb-size 16 \]  # for CEM only
    [--boolean_cbm \]  # for Boolean CBM only
    [--extra-dims 60 \]  # for Hybrid CBM only
    [--interventions \]  # enable test-time interventions

For evaluating the concept alignment score (CAS), use the script cem/cmapss_eval_cas.py in a similar way. Note that this computation is very slow.

For test-time interventions, we use the strategy for prognostics described in the paper by default. Different intervention strategies can be implemented in cem/utils/eval_with_interventions.py.

Evaluation will produce many figures and CSV files in the output directory; feel free to comment out some outputs in cem/cmapss_eval.py if these are not useful for you.

Acknowledgements

Code for CEM and CBM borrowed from cem, thanks to the original authors!