PBL Presentation
IoT-ML Based Predictive Maintenance System
Shrish Katrojwar, Himanshu Rokde, Siddhant Rajurkar
PBL Group No:
Sem: VII
Under the guidance of
Prof. Nisha Gongal
Symbiosis Institute of Technology, Nagpur Campus
1
Index
1. Problem statement 01
2. Objectives
3. Tools, Variables and parameters
4. Existing Processes
5. Result
6. Future Scope04
7. References
03
2
Problem Statement
01
Industries face unexpected machine failures Leads to:
• Costly downtime
• Production delays
• Increased maintenance cost 02
• Current methods (Reactive / Preventive) are in
efficient
• Need for a predictive approach that forecasts
failures in advance
03
3
Objectives
• Develop an AI-powered Predictive Maintenance
system
01
• Use IoT sensor data for real-time monitoring
02
• Train ML models to classify machine state → Normal
(0) or Failure (1)
• Build a dashboard to visualize:
a. Sensor data trends
b. Machine status alerts
03
• Reduce downtime and optimize maintenance schedule
4
Predictive Maintenance (PdM) Research Review
1. Traditional vs ML Approaches
01
• Early PdM: regression, time-series, threshold methods → limited for complex data.
• ML adoption: SVM, k-NN, ANN → better fault detection but issues: overfitting, imbalance, scalability.
2. Ensemble Learning in PdM
02
• Random Forest (RF): Robust, interpretable (feature importance), reduces overfitting.
• Gradient Boosting & XGBoost: High accuracy, handles imbalance, faster & regularized.
• LightGBM: Efficient on large IIoT data, fast training, high accuracy.
3. Evaluation Metrics & Interpretability
• Beyond accuracy: precision, recall, F1, ROC-AUC for imbalanced data.
• Confusion matrix → identify misclassifications.
4. Emerging Trends & Gaps
interpretability. 03
• Deep learning (CNN, LSTM, RNN) → strong performance but high computation & low
• Transfer & federated learning → knowledge sharing, privacy-preserving.
5
Tools, Variables & Parameters
Tools & Systems Used:
01
• Python, Pandas, NumPy, Scikit-learn
• XGBoost, LightGBM (ML models)
• Streamlit (Dashboard)
02
• IoT Sensors (Temperature, Vibration, Torque, RPM)
Key Variables:
• Air temperature [K]
• Process temperature [K]
• Rotational speed [rpm]
• Torque [Nm]
• Tool wear [min] 03
6
Existing Processes
Current Industrial Practices:
01
• Reactive Maintenance → Fix after failure
• Preventive Maintenance → Fixed schedule, often wasteful
02
• Condition Monitoring → Monitors but does not predict
Limitations:
• High downtime
• Resource wastage
• No early warning of failure
03
7
Solution Approaches (Models Tried)
Models Implemented: 01
• Logistic Regression → simple baseline, low accuracy
• Random Forest → ~98% accuracy, recall weaker
02
• XGBoost → ~97% accuracy, recall better (~66%)
• LightGBM (Best) → ~98.7% accuracy, best recall
(~69%)
Final choice: LightGBM (fast, scalable, reliable)
03
8
Results
01
Performance Metrics (Failure Detection – Class 1):
• Random Forest: Accuracy ~98%, Recall ~0.53, F1 ~0.66
• 02
XGBoost: Accuracy ~97%, Recall ~0.66, F1 ~0.76
• LightGBM: Accuracy ~98.7%, Recall ~0.69, F1 ~0.79
Conclusion: LightGBM is the best performer
03
9
Future Scope
01
• Connect real IoT sensors instead of simulation
• Deploy model on Cloud (AWS, Azure, GCP) for scalability
•
02
Extend to multi-class failure prediction (specific fault types)
• Add alert systems (SMS, Email, WhatsApp)
• Deploy on edge devices (Raspberry Pi/ESP32) for low-cost
industrial setups 03
10
References
01
• McKinsey & Company – Predictive Maintenance can save
15–30% maintenance costs
02
• AI4I Predictive Maintenance Dataset (UCI Repository)
• LightGBM Documentation:
https://lightgbm.readthedocs.ioXGBoost Documentation
https://xgboost.readthedocs.ioStreamlit Documentation
https://docs.streamlit.io
03
11
