🌱 Algae Health Monitoring using CNNs on Multivariate Sensor Data 📌 Project Overview
This project implements an end-to-end machine learning pipeline to monitor algae system health using multivariate time-series sensor data. The system identifies unhealthy operating conditions early and prioritizes high-risk periods for intervention.
🎯 Business Problem Industrial algae cultivation systems generate continuous sensor data (oxygen, PAM fluorescence, density). Manual monitoring is inefficient, and failures are often detected too late.
Goal: Automatically detect unhealthy system behavior while minimizing missed failures.
🧠 Solution Approach Converted raw sensor streams into sliding time windows
Labeled windows using domain-informed voting logic
Trained a 1D CNN to learn temporal health patterns
Evaluated using imbalance-aware, deployment-safe metrics
🔍 Key Features Sliding-window time-series modeling
Robust preprocessing for noisy sensor data
CNN-based temporal feature learning
Risk-based evaluation (Recall@Top-K)
Overfitting-safe validation strategy
📊 Evaluation Metrics PR-AUC (Primary): Measures unhealthy detection quality under imbalance
ROC-AUC (Secondary): Overall separability
Recall@Top-K: Operational usefulness when resources are limited
High Recall@Top-K means most unhealthy states are detected by inspecting only a small fraction of time windows.
🛠 Tech Stack
Python
NumPy, Pandas
Scikit-learn
TensorFlow / Keras
CNNs for time-series
🚀 Use Cases
Algae cultivation monitoring
Industrial sensor health detection
Early warning systems
Predictive maintenance pipelines
Challenges faced (and how they were handled) ⚠ Challenge 1: Severe class imbalance
Healthy dominates unhealthy
Accuracy and ROC-AUC misleading
✅ Solution:
PR-AUC
Recall-focused evaluation
Risk-based metrics
⚠ Challenge 2: Sliding-window leakage
Overlapping windows inflate metrics
✅ Solution:
No threshold tuning
Conservative evaluation
Emphasis on ranking quality, not absolute scores
⚠ Challenge 3: Sensor noise & NaNs
Density and PAM signals noisy
Scaling caused NaNs initially
✅ Solution:
RobustScaler
Explicit NaN checks
Window-level aggregation
⚠ Challenge 4: Day-night bias
Oxygen naturally oscillates
Static thresholds fail
✅ Solution:
Global mean comparison
Window-based voting
CNN learns temporal context
✅ Key Takeaway The model is not optimized for perfect accuracy but for early and reliable detection of unhealthy states under real operational constraints.