Multi Sensor Predictive Maintenance Configurable System for Industry 4.0

Abstract

Modern Industry 4.0 plants require maintenance strategies that respond to actual machine condition rather than fixed schedules [1], [2]. Conventional reactive and preventive approaches either wait for failure or over-service equipment, both at significant cost. This paper presents a configurable Multi- Sensor Predictive Maintenance (PdM) system that unifies three strategies — scheduled, usage-based, and ML-driven predictive maintenance — within a single operational platform [3]. The system is validated across three heterogeneous machine types CNC machine, pump, and compressor — each instrumented with five machine-specific sensor modalities and connected via an ESP32 edge node that streams data to a cloud backend over MQTT [4]. For each machine type, a dedicated XGBoost regres- sion model is trained on an engineered feature set comprising raw sensor readings, lag-1 features, threshold-exceedance difference features, a composite health index, and temporal indicators, to estimate Remaining Useful Life (RUL) in days [5], [6]. Evaluated on machine-wise held-out test sets of 31,504 samples each, the models achieve MAE of 3.07, 3.41, and 2.60 days for the CNC, pump, and compressor respectively, demonstrating consistent and actionable prediction accuracy across diverse degradation profiles. Real-time Grafana dashboards and a three-tier alert system (Critical <3 days, Warning <10 days, Normal ?10 days) translate model outputs into prioritised maintenance actions for role-appropriate stakeholders [7].

Introduction

The paper presents a unified predictive maintenance (PdM) platform for Industry 4.0, designed to reduce costly unplanned failures on heterogeneous factory equipment such as CNC machines, pumps, and compressors. It moves beyond reactive and preventive maintenance by leveraging machine-specific sensor data and interpretable machine learning models to predict Remaining Useful Life (RUL).

Key Challenges Addressed:

• Traditional maintenance strategies are inefficient:
  • Reactive: waits for failure, causing costly downtime.
  • Preventive: replaces healthy parts, increasing costs.
• Factory machines are heterogeneous; a single PdM model cannot handle different degradation behaviors.
• Plant engineers need explainable outputs, which black-box deep learning models often fail to provide.

Proposed System Design:

1. Machine-Specific Multi-Modal Sensing: Five sensors per machine type (vibration, current, temperature, pressure, speed) with calibrated thresholds, connected to ESP32 edge nodes.
2. Unified Maintenance Strategies: Scheduled, usage-based, and ML-driven predictive maintenance coexist.
3. Per-Machine XGBoost Models: Dedicated gradient-boosted regressors per machine type for RUL prediction, interpretable via feature importance.
4. Cloud-Edge Architecture: MQTT-based streaming from edge to cloud, enabling real-time ingestion, inference, and alerts.
5. Real-Time Dashboards: Grafana dashboards provide role-specific visualizations and multi-channel alerting.

System Architecture:

• Physical Sensing Layer: Captures five sensor modalities per machine type, edge-timestamped.
• Edge Processing: ESP32 nodes handle noise filtering, message formatting, and MQTT publishing.
• Cloud Analytics: Processes data into features, runs machine-specific XGBoost models, and triggers alerts using three-tier (Critical, Warning, Normal) RUL thresholds.
• Application & Visualization: Grafana dashboards for operators, engineers, and management; integrates with CMMS for automatic work orders.
• Scalability & Security: Modular architecture allows onboarding new machines without backend redesign; TLS and RBAC ensure secure communication and access.

Predictive Analytics & ML Pipeline:

• Datasets: Timestamped multi-sensor readings for CNC, pump, and compressor machines; failure labels derived from threshold breaches.
• Feature Engineering: 18 features per machine including raw sensors, lag features, threshold-difference features, composite health index, and temporal features.
• Training & Testing: 80/20 machine-instance split to prevent leakage; separate XGBoost regression models per machine type.
• Inference & Alerts: RUL predictions inverse-transformed to days, classified into three alert states for prioritized maintenance planning.

Results:

• RUL Prediction Accuracy:
  • Compressor: MAE 2.60 days, RMSE 3.48 days
  • CNC: MAE 3.07 days
  • Pump: MAE 3.41 days
• Predictions align closely with actual RUL, enabling actionable maintenance prioritization.
• The platform effectively integrates heterogeneous machine data, interpretable ML, and real-time alerts to optimize maintenance scheduling and reduce unplanned downtime.

Contributions:

• A configurable, multi-strategy PdM platform validated on CNC, pump, and compressor machines.
• Machine-specific sensor instrumentation with threshold-aware feature engineering.
• Interpretable XGBoost RUL models suitable for real-time edge inference.
• Scalable deployment stack using MQTT, containerized microservices, and Grafana dashboards.

In short, the system provides a practical, Industry 4.0-ready PdM solution that improves reliability, reduces costs, and enables data-driven maintenance decisions across heterogeneous machinery.

Conclusion

This paper presented a configurable Multi-Sensor PdM system for Industry 4.0, validated across three heterogeneous machine types — CNC machine, pump, and compressor. By training dedicated XGBoost models per machine type on threshold-aware, lag-enriched feature sets, the system achieves MAE of 2.60–3.41 days across all three machines on held- out test sets of 31,504 samples each [5], [6]. The three-tier alert system and Grafana dashboards connect model output directly to actionable maintenance decisions, while the unified platform simultaneously supports scheduled, usage-based, and ML-driven maintenance strategies [4], [9]. Future work will pursue: large-scale longitudinal deploy- ment on a live factory floor to quantify actual reductions in downtime and maintenance cost; digital twin integration to simulate degradation scenarios and generate synthetic fault data for machines with limited failure history [1]; federated learning across multiple sites to share model improvements without exposing raw operational data [3]; hyperparameter optimisation per machine type to further improve model ac- curacy; expanded sensing (acoustic emission, flow rate) for richer fault observability; and CMMS integration for automatic work-order generation [13]. Together, these extensions move the framework toward fully autonomous, self-improving main- tenance for next-generation smart manufacturing [1], [3].

References

[1] T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” in Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, 2016, pp. 785–794. [2] J. Lee, B. Bagheri, and H. A. Kao, “A cyber-physical systems archi- tecture for industry 4.0-based manufacturing systems,” Manufacturing Letters, vol. 3, pp. 18–23, 2015. [3] G. A. Susto, A. Schirru, S. Pampuri, S. McLoone, and A. Beghi, “Machine learning for predictive maintenance: A multiple classifier approach,” IEEE Transactions on Industrial Informatics, vol. 11, no. 3, pp. 812–820, 2015. [4] P. J. R. Gama, E. Zio, and R. L. M. Silva, “Remaining useful life estimation: A review,” Mechanical Systems and Signal Processing, vol. 115, pp. 1–20, 2019. [5] U. Khan, D. Cheng, F. Setti, F. Fummi, M. Cristani, and L. Capogrosso, “A comprehensive survey on deep learning-based predictive mainte- nance,” ACM Transactions on Embedded Computing Systems, vol. 24, no. 4, pp. 1–42, 2025. [6] A. Al-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, and M. Ayyash, “Internet of things: A survey on enabling technologies, protocols, and applications,” IEEE Communications Surveys & Tutorials, vol. 17, no. 4, pp. 2347–2376, 2015. [7] A. Banks and R. Gupta, “Mqtt version 3.1.1,” OASIS Standard, 2014. [8] A. Rajkumar, M. Fuchs, and C. Busch, “Edge computing for predictive maintenance in industry 4.0,” IEEE Access, vol. 9, pp. 124 721–124 735, 2021. [9] F. Pedregosa et al., “Scikit-learn: Machine learning in python,” Journal of Machine Learning Research, vol. 12, pp. 2825–2830, 2011. [10] R. Yan, R. X. Gao, and X. Chen, “Wavelets for fault diagnosis of rotary machines: A review with applications,” Signal Processing, vol. 96, pp. 1–15, 2014. [11] T. Januska, J. Sedlacek, and J. Krejci, “Grafana for industrial monitoring and visualization,” in Proceedings of the IEEE International Conference on Industrial Informatics (INDIN), Helsinki, Finland, 2019. [12] P. Zhang, Y. Yan, and H. Gao, “Deep learning for smart manufacturing: Methods and applications,” IEEE Transactions on Industrial Informatics, vol. 17, no. 2, pp. 1245–1256, 2021. [13] A. Saxena and K. Goebel, “Turbofan engine degradation simulation data set,” NASA Ames Prognostics Data Repository, 2008. [14] Z. Zhao, Y. Wang, R. Yan, and J. Mao, “Machine health monitoring using local feature-based gated recurrent unit networks,” IEEE Transac- tions on Industrial Electronics, vol. 65, no. 2, pp. 1539–1548, 2018. [15] “Iso 13374-1:2019, condition monitoring and diagnostics of machines – data processing, communication and presentation,” ISO, Geneva, Switzerland, 2019.

Copyright

Copyright © 2026 Vaishnavi Borkar, Rahul Gone, Unmesh Kapure, Prajwal Gaikwad, Shubhangi Deshpande. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.