Link to Hackathon Website: https://smartfactory-owl.de/eventer/ai-challenge-days-2023/
Predict the temperature at position 6 of the chocolate melting and cooling machine to replace the current sensor with a predictive model to reduce costs.
After an exploratory data analysis (EDA), we decided to train two separate regression models. One day and one night cycle model. In addition to the day/night cycle split, we added hours and minutes as features. The use of a standard scaler was also very important since the sensor data scales are very different. We then performed PCA, where the number of principal components was chosen based on the PVE/CPVE plot, as 3 components already explained <98% of the variance. The hackathon was evaluated based on the mean absolute error (MAE). We decided to use an LSTM for this regression task because it gave the best MAE results.
- Mean Absolute Errors (MAE):
- Day MAE: 0.071-0.105
- Night MAE: 0.01
- Overall MAE: 0.041-0.058
- Python, Pytorch
- PCA, MDS, StandardScaler, PVE/CPVE
- Ridge Regression, Lasso Regression, XGBoost, Support Vector Regression, Long short-term memory NNs (LSTM), Linear Regression
- Model Evaluation Dataset Splits: (week 1-4 excl. testset/week 5)
- Full Data
- Train on full data and evaluate on Testset
- Holdout Split 1
- Train on (Tuesdays, Wednesdays) | Evaluate on (Thursdays)
- Holdout Split 2
- Train on (Wednesdays, Thursdays) | Evaluate on (Tuesdays)
- Holdout Split 3
- Train on (Tuesdays, Thursdays) | Evaluate on (Wednesdays)
- Single Holdout 1
- Train on (Tuesdays) | Evaluate on (Wednesdays, Thursdays)
- Single Holdout 2
- Train on (Wednesdays) | Evaluate on (Tuesdays, Thursdays)
- Single Holdout 3
- Train on (Thursdays) | Evaluate on (Tuesdays, Wednesdays)
- Testset (DO NOT USE)
- Week 5 Data from tuesday, wednesday, thursday