Industrial Predictive Maintenance System: A Complete Practice from Reactive Repair to Intelligent Prediction

This article introduces an industrial thermal monitoring and predictive maintenance system developed for Yazaki Morocco. By integrating Modbus data collection, machine learning anomaly detection, soft sensing technology, and automated alerts through a three-layer architecture, it achieves a leap from reactive repair to intelligent prediction, providing a practical reference solution for the digital transformation of the manufacturing industry.

Traditional maintenance models in manufacturing have issues such as high risk of reactive repair and over-maintenance in regular servicing. This project, developed by students from ENSA Kénitra Engineering College in Morocco for Yazaki Morocco, addresses temperature anomalies in foam mold production lines caused by scale deposition, valve failures, etc., with the goal of transforming reactive response into proactive prevention.

The system adopts a three-layer architecture: Physical Data Collection Layer (Raspberry Pi 4 collects data from 12 temperature and 3 flow sensors via Modbus RTU, integrated with LED alerts); Data Processing and Machine Learning Layer (Isolation Forest for anomaly detection, Random Forest for fault classification, Ridge Regression and gray-box model for predictive maintenance); Display and Automation Layer (React frontend for real-time monitoring, FastAPI backend, InfluxDB storage, n8n for automated alerts).

The AMDEC risk assessment method is introduced, which calculates the Risk Priority Number (RPN) from three dimensions: severity, occurrence frequency, and detectability. It ranks 7 fault modes (e.g., valve low-position failure with RPN 180, heater resistance damage with RPN 160, etc.), helping to focus resources on handling key risks.

A gray-box model combining physical principles and data-driven approaches is used to indirectly estimate pipe scale thickness via heat conduction formulas. Without dedicated sensors, it only uses existing temperature and flow data to achieve quantitative monitoring of invisible issues.

Backend uses FastAPI, Uvicorn, PyModbus, scikit-learn, etc.; Frontend uses React18, TailwindCSS, Recharts; Deployment uses Docker containerization for n8n, with Raspberry Pi as edge nodes.

Progressive complexity (from monitoring to intelligent diagnosis), multi-dimensional data fusion (sensors + historical data + physical models), engineering methodology guidance (AMDEC), human-machine collaboration interface (LED + Web + automatic notifications).

The project demonstrates the practical application of AI in manufacturing and provides a complete reference implementation; future directions include exploring deep learning, digital twins, and expanding to more devices.