Stall Detection with a 1D Convolutional Neural Network

University project — a graded individual assignment from my studies at TU Berlin. This repository is published for demonstration purposes only. See Disclaimer & License — no permission is granted to use or copy any of its contents.

Task

Binary classification of univariate sensor time series: each sample is a signal of 1,400 time steps, labeled either stable (0) or stall (1).

• 548 labeled samples, with only ~17% belonging to the stall class (strong class imbalance)
• Evaluation metric: F1 score
• In addition to the trained model, the assignment required reporting an estimated F1 score with an uncertainty margin for unseen data, plus the chosen decision threshold — the submission was then graded against a hidden hold-out set

Approach

1. Data exploration — dimensions, class distribution, visual comparison of stable vs. stall signals
2. Stratified 5-fold cross-validation — within each fold, an inner 80/20 train/validation split is used for early stopping and threshold selection; the outer fold serves as an untouched test set
3. Class imbalance handling — class weights computed per training fold
4. Regularization / training control — EarlyStopping (restoring best weights) and ReduceLROnPlateau callbacks, Dropout, BatchNormalization
5. Decision-threshold optimization — thresholds from 0.01 to 0.99 are swept on validation predictions to maximize F1, in two variants:
   • cross_validation.py: a threshold per fold, with the median reported
   • cross_validation_pooled_threshold.py: one global threshold chosen from the pooled out-of-fold validation predictions, then evaluated on the outer test folds
6. Stability check — re-runs with different random seeds
7. Final model — trained on all data with an 80/20 train/validation split (train_final_model.py)
8. Project documentation as a 90s overview video presentation/project_presentation.mp4

To avoid data leakage, the Normalization layer is adapted only on the respective training split in every setting.

Model architecture

A compact 1D CNN (Keras / TensorFlow):

Input (1400, 1)
→ Normalization (adapted on training data only)
→ Conv1D(16, kernel 7) → BatchNorm → ReLU → MaxPool(2) → Dropout(0.2)
→ Conv1D(32, kernel 5) → BatchNorm → ReLU → MaxPool(2) → Dropout(0.2)
→ Conv1D(64, kernel 3) → BatchNorm → ReLU
→ GlobalAveragePooling1D
→ Dense(32, ReLU) → Dropout(0.3)
→ Dense(1, sigmoid)

Results

Metric	Value
Estimated F1 on unseen data (cross-validation)	0.90 ± 0.10
Chosen decision threshold	0.51
F1 on the hidden hold-out set (grading)	1.00 (no misclassifications)

The hold-out result lies inside the reported interval; the margin was chosen conservatively because of the F1 variance across folds (± 0.06) and the high variance of the fold-optimal thresholds (± 0.3 around a mean of 0.51).

Identified possible improvements (see presentation): a smaller architecture reaching the same F1 with fewer parameters, and RepeatedStratifiedKFold for a more robust performance estimate.

Repository contents

Path	Description
src/data_overview.py	Data exploration: shapes, class distribution, example signal plots
src/data_preprocessing.py	Input shape preparation and naive baseline
src/cross_validation.py	Stratified 5-fold CV with per-fold threshold optimization
src/cross_validation_pooled_threshold.py	Stratified 5-fold CV with one global threshold from pooled out-of-fold predictions
src/train_final_model.py	Trains the final model on all data and saves it
presentation/project_presentation.mp4	Short video presentation of approach and results
requirements.txt	Python dependencies (Python 3.12, TensorFlow 2.19, Keras 3.11)

Dataset — intentionally not included

The dataset was provided by the instructors and is not part of this repository and will not be shared. The scripts are therefore not runnable as-is.

Disclaimer & License

• This repository documents my own coursework from my studies at TU Berlin. It is not affiliated with or endorsed by TU Berlin.
• All rights reserved. No permission is granted to use, copy, modify, distribute, or build upon any content of this repository — see LICENSE.
• In particular, submitting this work (in whole or in part) as your own in any academic context is prohibited and constitutes plagiarism.