mercy456-coll/Refrigeration_Fault_Prediction

Random Forest + XGBoost for refrigeration fault detection. 1.87M sensor readings | 13 fault classes | 77% accuracy | 99%+ recall on critical failures. Built for affordable predictive maintenance in Nigeria's cold chain. # Refrigeration Fault Prediction A machine learning model for predictive maintenance in cold storage systems using fault detection across 13 fault types. Demonstrates real-world industrial ML application with focus on data-driven maintenance strategies. **Repository Contents:** This repo contains only the Jupyter Notebook with the full training pipeline and this README. The trained model is not saved here. You can train and save the model yourself using the notebook. ## Problem Refrigeration system failures in food cold storage infrastructure lead to significant losses—spoiled inventory, operational downtime, and supply chain disruption. This project builds a predictive model to identify faults before critical failures occur. **Context:** Developed for the APWEN Queen's Engineering Challenge 2026 (Distributed AI-Enabled Predictive Maintenance for Food Cold Storage Infrastructure) ## Model Performance | Metric | Value | |--------|-------| | **Accuracy** | 77% | | **Dataset Size** | ~1.87M sensor readings | | **Fault Types** | 13 different fault classes | | **Approach** | Random Forest Classification | ## Dataset - **Source:** Simulated Refrigerator Fault Diagnosis Dataset (Kaggle) - **Size:** Approximately 1.87 million sensor readings - **Features:** Temperature sensors, pressure sensors, compressor cycles, humidity, runtime statistics - **Target Classes:** 13 distinct fault types (including normal operation) - **Domain:** Food cold storage systems (agricultural/food supply chain) ## Approach ### Data Preprocessing - Aggregated high-frequency sensor streams into meaningful features - Handled missing values and sensor noise - Normalized numerical features for model compatibility - Balanced fault class representation for training stability ### Model Architecture **Random Forest Classifier** - **Estimators:** 100 decision trees - **Max Depth:** Optimized to prevent overfitting on imbalanced classes - **Feature Importance:** Identifies which sensors are most …