IJCNC 05

Edge-AI Enabled Fault Detection and Root Cause Analysis in Industrial Motors using Multimodal Sensor Data

Jayashree Kulkarni 1, Ramesh Kagalkar 2 and Nandini Sidnal 3
1,2,3 Department of Computer Science & Engineering,
1Graphic Era Deemed to be University, Dehradun, Uttarakhand, India,
2Nagarjuna College of Engineering and Technology, Bengaluru, Karnataka, India.
3Torrens University, Sydney, Australia

ABSTRACT

This article presents an Edge-AI enabled hybrid deep learning framework for real-time fault detection and root cause analysis in industrial motors using multimodal sensor data, containing vibration, temperature, and current signals. The proposed system leverages a CNN-LSTM model deployed on edge devices to accurately classify operational states into normal, minor, and major faults. A dataset comprising 15,000 labeled samples collected from real-world industrial setups and augmented through techniques such as SMOTE and time-warping was used for training and evaluation. In the implementation, the CNN component captures spatial patterns in sensor data, while the LSTM layer models temporal dependencies, enabling effective fault diagnosis. The proposed hybrid model achieved superior performance with 96.8% accuracy, 97.2% precision, 96.5% recall, and an F1-score of 96.8%, along with a low inference latency of ≤198 ms, demonstrating suitability for real-time edge deployment. Comparative analysis against CNN-only and LSTM-only models confirms the hybrid architecture’s advantage in fault sensitivity and prediction
reliability. Additional insights from confusion matrix analysis, ROC-AUC evaluation, and fault-wise performance metrics validate the model’s robustness. The system also incorporates TinyML-based optimizations and lightweight messaging for efficient edge computing, making it a scalable solution for predictive maintenance in Industry 4.0 applications.

KEYWORD
Edge-AI, CNN-LSTM, Predictive Maintenance, Industrial Motors, Fault Detection, Multimodal Sensors

1. INTRODUCTION
Industry 4.0 marks the transition from the traditional industrial operations that integrated cyberphysical systems, IoT, and Artificial Intelligence to create smart, connected production systems. This evolution would have improved monitoring, control, and optimization of manufacturing processes so that factories would have become more responsive, adaptive, and efficient. Edge computing, in this regard, has come up as an enabler for real-time analytics and decision making. As computational intelligence is brought close to the data source, edge computing kills latency and reduces the reliance on some bandwidth-heavy cloud infrastructure, thereby guaranteeing timely response for industrial applications where it is most needed mainly in, fault detection and system diagnostics. One of the most important use cases considered in smart industrial environments is predictive maintenance, i.e., predicting the deterioration of equipment prior to its failure. Predictive maintenance, unlike scheduled and reactive maintenance, continuously monitors sensor signals and applies intelligent algorithms that detect symptoms of impending faults, be it mechanical, thermal, or electrical. This way, unplanned downtime can be avoided, routine maintenance costs removed, and the usable life of motors, an example of an industrial asset, can be prolonged.

These modern systems use the latest tools and techniques, especially deep learning frameworks, to model complex relations in high-frequency sensor data. In this respect, hybrid architecture such as CNN combined with LSTM have been found to be highly effective. CNNs extract local and spatial features from sensor signals (such as vibration or current), while LSTMs study sequential dependency and temporal dynamics within time-series data so that temporal fault patterns can be recognized. Moreover, using multimodal sensor data-vibration, temperature, and electrical curren tincreases the system’s ability to detect faults and accentuate a failure’s diverse manifestations, whether mechanical, thermal, or electrical. Hence, this comprehensive data fusion ensures more straightforward fault detection, fewer false positives, and reliable operation in various situations. To provide the basis for a complete evaluation, a hybrid dataset has been constructed consisting of 15,000 labeled samples that were obtained from both real industrial motor logs as well as synthetic augmented data. The classes of faults include bearing damage, stator overheating, load imbalance, and normal operational states.

2. LITERATURE OUTLINE

The literature survey on AI-based fault detection systems highlights significant advancements in deep learning, edge computing, and IoT-enabled diagnostics across industrial applications. Researchers have developed various models for real-time fault diagnosis, predictive maintenance, and anomaly detection. Despite progress, challenges persist in data imbalance, model interpretability, real-time deployment, and scalability. This survey helps to identify research gaps and guides future innovation in intelligent, energy-efficient, and reliable fault detection frameworks.
Leite et al. reviewed Fault detection and diagnosis (FDD) strategies in Industry 4.0 environments, emphasizing real-time automation, data interoperability, and intelligent control systems. The study identified major limitations such as lack of data standardization, integration complexity, and insufficient real-time analytics for deployment [1]. Saeed et al. introduced deep learning methods for industrial equipment health diagnostics under scarcity scenarios, reporting better fault classification but issues with model generalization, energy efficiency, and scalability [2].

Kodumuru et al. discussed the integration of Artificial Intelligence and the Internet of Things in pharmaceutical manufacturing, proposing a smart synergy to enhance process optimization, quality control, and traceability; however, their framework faces challenges in regulatory compliance, system security, and high initial deployment costs [21]. Hassani and Dackermann conducted a systematic review on advanced sensor technologies for structural health monitoring and non-destructive testing, showcasing sensor versatility and data precision while identifying gaps in environmental robustness, real-time feedback, and calibration consistency [22]. Mohandas et al. surveyed incremental deep learning strategies for industrial defect detection, emphasizing continual learning adaptability, yet highlighting persistent issues like catastrophic forgetting, data imbalance, and lack of real-world datasets [23]. Janga et al. reviewed practical AI usage in Earth sciences via remote sensing, covering terrain fault detection and geospatial data interpretation, though noting constraints related to low-resolution imagery, limited ground truth validation, and model generalization [24].

Strantzalis et al. applied Edge-AI for operational state recognition in DC motors, enabling on device inference with reduced latency, yet encountering constraints in real-time accuracy under noisy conditions and hardware resource limitations [31]. Sabry and Amirulddin reviewed fault detection techniques in industrial robots and multi-axis machines, covering AI and traditional control-based methods, while noting shortcomings in dataset availability, system-specific tuning, and real-time adaptability [32]. An et al. proposed an efficient CNN-based edge solution for realtime motor fault detection, achieving fast and accurate classification but limited by hardware compatibility and model compression complexity [33]. De las Morenas et al. examined edgebased ML methods for electrical machine fault diagnosis, emphasizing local sensory data processing while reporting problems with update latency and memory constraints in low-power devices [34]. Tian et al. explored AI-based arc fault detection using entropy and cumulants, demonstrating high sensitivity to fault initiation but challenged by algorithmic complexity and scalability in DC circuit environments [35].

Alenizi et al. introduced a taxonomy of AI technologies in Industry 4.0, categorizing classification, optimization, and prediction tasks while identifying trustworthiness, lifecycle integration, and interoperability as critical challenges [71]. Chong Chen et al. reviewed digital twins in predictive maintenance and their interaction with machine learning, providing insights into system simulation, though difficulties remain in synchronization, twin fidelity, and implementation cost [72]. Mahesh et al. proposed a deep active learning framework for intelligent bearing fault detection, showing adaptability to diverse conditions, but suffering from annotation uncertainty and model convergence instability [73]. Khan et al. explored the role of IoT in Industry 4.0 adoption, focusing on real-time data collection and control integration while identifying gaps in device interoperability, data overload, and decision-making automation [74]. Su et al. presented the ICICOS framework for integrating cloud-edge AI in industrial automation, improving control reliability but facing orchestration delays and coordination complexities across layers [75].

2.1. Overview of Findings

presents a comparative analysis of ten selected research studies relevant to AI-based fault diagnosis in industrial systems. These studies were chosen based on their methodological innovation, dataset usage, practical relevance, and identified limitations. The analysis reveals that while techniques such as CNN, LSTM, and ensemble learning demonstrate high classification accuracy (e.g., [2], [55], [66]), they often suffer from issues like high computational cost, energy inefficiency, and limited generalization to real-world edge deployments. For example, study [2] utilizes a CNN-LSTM model for fault detection using vibration and current signals, achieving good accuracy but facing scalability and performance issues in low-resource environments. Similarly, [9] addresses wind turbine diagnostics with AI-based signal processing, yet suffers from noisy sensor data and limited dataset diversity. Study [27] explores an IoT–Edge–Cloudbased digital twin platform for real-time monitoringbut is hindered by latency jitter and high deployment costs in industrial testbeds. Moreover, [30] investigates neural network watermarking for IP protection in edge-AI models, which introduces robustness concerns under adversarial conditions.

Studies like [60] and [66] tackle class imbalance and resource constraints, proposing ensemble learning and edge-optimized deep learning respectively. Notably, [80] emphasizes the importance of explainable AI (XAI) for operator trust, although it faces trade-offs between model interpretability and response time.

3. OVERVIEW OF THE SYSYTEM

The proposed system is a comprehensive Edge-AI-enabled fault detection framework designed to operate autonomously within industrial environments. It integrates multimodal sensor data acquisition, real-time inference on edge devices, and lightweight communication protocols to enable accurate and low-latency fault classification in industrial motors. This modular system is specifically tailored for predictive maintenance by minimizing reliance on cloud infrastructure and ensuring continuous operation even in resource-constrained setups.

Key Components and Functional Overview

HVAC systems for flexible deployment. Also, the design allows for expansion with more sensors or the incorporation of more advanced AI models for increasing functionality. By integrating multimodal sensors with edge-compatible deep learning, the system guarantees effective ondevice fault identification and real-time notifications through light-weight communication protocols, while ensuring minimum dependency on cloud infrastructure.

The system shown in figure 1 is an Edge-AI-based predictive maintenance system for industrial motors based on multimodal sensor data. It consists of three physical sensors: vibration, temperature, and current mounted on the motor to monitor continuously mechanical, thermal, and electrical parameters respectively. These sensor readings are processed in realtime by an AI-powered IoT edge device, which resolves major issues like operational latency, low-latency processing, and data timeliness. The edge device carries out local inference with a hybrid CNNLSTM deep learning model, which integrates convolutional neural networks (CNN) for spatial feature extraction from sensor readings and long short-term memory networks (LSTM) for temporal dependency analysis to identify changing faults. This enables the system to detect faults and determine root causes on the device itself, with much less reliance on cloud processing. When a fault is detected, the system categorizes it as minor, major, or normal and triggers alerts through light protocols such as MQTT or HTTP to a central dashboard. The system architecture is predictive maintenance-capable with conditionbased monitoring and timely alert generation, leading to lowered maintenance expenses (up to
30%), unplanned production downtimes’ avoidance, and enhanced system energy efficiency. This modular configuration is scalable between various types of motor-driven equipment and extendable to multiple industrial applications, including water pumping, HVAC, and conveyor systems, with optional cloud connectivity for long-term trend analysis or model updating.

3.1. System Flow Diagram
Figure2 illustrates the methodology and system design for an Edge-AI-enabled fault detection framework in industrial motors. Each block in the flow represents a critical stage in the processing pipeline from raw data acquisition to real-time alerting. Below is a detailed description of each step:

4. MATHEMATICAL MODELING

To formalize the operation of the proposed Edge-AI-based fault detection system, we define the mathematical constructs governing sensor fusion, feature extraction, fault classification, and decision-making.

5. RESULT ANALYSIS

The system’s performance was evaluated using multiple key metrics: accuracy, defined as the ratio of correctly predicted outcomes to the total number of predictions; precision, which measures the proportion of true fault detections among all predicted positives; recall (or sensitivity), indicating the proportion of actual faults correctly identified by the model; and the F1 score, calculated as the harmonic mean of precision and recall to ensure balanced performance on imbalanced datasets. In addition, inference time, which is the average time taken per sample for classifying on edge devices, was tracked to ensure real-time deployment applicability. All the metrics were calculated across the three classes of faults (normal, minor, and major) to assess robustness and consistency.

The system consistently demonstrated high accuracy and low latency across all test cases shown in Table 2. The average inference time was ≤200ms, meeting the requirements for real-time edge processing. This table summarizes the performance of the CNN-LSTM fault detection system across four key motor conditions. The model achieved high accuracy and F1 scores (above 94%) for all fault types. Inference time remained under 200 ms, ensuring suitability for real-time edge deployment. Notably, the system showed the highest precision and recall for normal and stator overheat cases.

5.1 Graphical Representation of key Visualizations:

Here are the key visualizations generated to support the experimental evaluation, along with the
corresponding key points.

1. Accuracy per Fault Type
2. Precision, Recall, F1 Score Comparison
3. Inference Time per Fault Type
4. Confusion Matrix 5. ROC Curve
5. Baseline Model Comparison:

Shows the efficiency of edge-based predictions in milliseconds.The bar graph shown in Figure 4a illustrates the average time (in milliseconds) taken by the CNN-LSTM model to infer each fault category on the edge device. The model achieves the lowest inference time for the “No Fault” condition at around 180 ms, indicating efficient recognition of normal motor behavior. “Stator Overheat” follows closely with approximately 185 ms, showing that thermal fault detection does not significantly impact latency. “Bearing Fault” inference takes slightly more time at about 190 ms, dueto complex vibration signature analysis. The highest inference time is observed for “Load Imbalance” at 200 ms, likely because current fluctuations are harder to classify and require deeper temporal analysis, slightly increasing model processing time. Overall, all classes are inferred under 210 ms, meeting realtime edge AI constraints.

CONCLUSION

This research proposes an efficient Edge-AI-based approach for real-time industrial motor fault detection using a hybrid CNN-LSTM architecture for multimodal sensor data such as vibration, temperature, and current signals. The proposed research study is shown to perform with high accuracy and low latency, indicating applicability for deployment in resource limited edge environments. Through enabling on-board inference, predictive analytics, and smart alerting mechanisms, the framework significantly reduces equipment downtime and maintenance costs. Further, the combination of real-time monitoring and machine learning based diagnostics results in prompt fault identification, even in dynamic operating conditions. The framework is compliant with the fundamental goals of Industry 4.0 by providing a scalable, autonomous, and energy-efficient solution for smart fault diagnosis. It enables industries to move from reactive to proactive maintenance practices, hence ensuring operational reliability and sustainability.

CONFLICT OF INTEREST

The authors declare no conflict of interest.

REFERENCES

[1] Leite, D.; Andrade, E. Rativa, D., & Maciel, A. M. A. Fault Detection and Diagnosis in Industry 4.0: A Review on Challenges and Opportunities. (2025). Sensors, 25(1), 60.
[2] Saeed, A., A.; Khan, M.; Akram, U. Deep learning-based approaches for intelligent industrial machinery health management and fault diagnosis in resource-constrained environments. Sci Rep15, 1114 (2025).
[3] Habyarimana, M.; Adebiyi, A. A. A Review of Artificial Intelligence Applications in Predicting Faults in Electrical Machines. Energies. (2025). 18(7), 1616.
[4] Altaf, S., Al-Anbuky, A., & Gheitasi, A, Enhancing Fault Detection in Distributed Motor Systems Using AI-Driven Cyber-Physical Sensor Networks. Engineering Proceedings, (2024), 82(1), 78.
[5] Akhyar Akhyar.; Mohd Asyraf Zulkifley.; Jaesung Lee.; Taekyung Song.; Jaeho Han.; Chanhee Cho.; Seunghyun Hyun.; Youngdoo Son.; Byung-Woo Hong.; Deep artificial intelligence applications for natural disaster management systems: A methodological review, Ecological Indicators. (2024), Volume 163,2024, 112067, ISSN 1470-160X,
[6] Chen L.; Xia C.; Zhao Z.; Fu H.; Chen. Y. AI-Driven Sensing Technology: Review, Sensors, (2024), 24(10), 2958.
[7] Del-Coco M.; Carcagnì P.; Oliver S. T.; Iskandaryan D.; Leo M. The Role of AI in Smart Mobility: A Comprehensive Survey. Electronics, (2025), 14(9), 1801.
[8] Ucar A.; Karakose M.; Kırımça, N. Artificial Intelligence for Predictive Maintenance Applications: Key Components, Trustworthiness, and Future Trends. Applied Sciences,
(2024), 14(2), 898.
[9] Alagha N.; Khairuddin A. S. M.; Haitaamar, Z. N.; Al-Khatib O.; Kanesan J. Artificial Intelligence in Wind Turbine Fault Detection and Diagnosis: Advances and Perspectives. Energies, (2025), 18(7), 1680.

[10] Daniel Fährmann; Laura MartínLuis Sánchez.Naser Damer.Anomaly Detection inSmart Environments: A Comprehensive Survey, (2024).
[11] Zixuan Zhang.; Luwei Wang.; Chengkuo Lee. Recent Advances in Artificial Intelligence Sensors, (2023).
[12] Aguayo-Tapia, Avalos-Almazan, Rangel-Magdaleno J. Entropy-Based Methods for Motor Fault Detection: A Review. Entropy, (2024), 26(4), 299.
[13] Ahmed, S.F.;Alam, M.S.B.; Hassan, M.; Deep learning modelling techniques: current progress, applications, advantages, and challenges. Artificial Intelligence Rev56, 13521–
13617 (2023).
[14] Uçar Ayşegül.; Karakose Mehmet.; Kırımça Necim. Artificial Intelligence for Predictive Maintenance Applications: Key Components, Trustworthiness, and Future Trends. Applied Sciences. (2024),14, 898.
[15] Moosavi S.; Farajzadeh-Zanjani M.; Razavi-Far.; Palade V.; Saif M. Explainable AI in Manufacturing and Industrial Cyber–Physical Systems: A Survey. Electronics, (2024),
13(17), 3497.
[16] Marium Jalal.; Ihsan Ullah Khalil.; Azhar ul Haq.; Deep learning approaches for visual faults diagnosis of photovoltaic systems: State-of-the-Art review, Results in Engineering, sciencedirect.com, (2024), Vol. 23,2024,102622.
[17] Gültekin Ö.; Cinar E.; Özkan K.; Yazıcı A. Real-Time Fault Detection and Condition Monitoring for Industrial Autonomous Transfer Vehicles Utilizing Edge Artificial Intelligence. Sensors, (2022), 22(9), 3208.
[18] Miller T.; Durlik I.; Kostecka E.; Borkowski P.; Łobodzińska A. A Critical AI View on Autonomous Vehicle Navigation: The Growing Danger. Electronics, (2024), 13(18), 3660.
[19] Jouini O.; Sethom K.; Namoun A.; Aljohani N.; Alanazi M. H.; Alanazi M. N. A Survey of Machine Learning in Edge Computing: Techniques, Frameworks, Applications, Issues, and Research Directions. Technologies, (2024), 12(6), 81.
[20] Visconti P.; Rausa G.; Del-Valle-Soto C.; Velázquez R.; Cafagna, D.; De Fazio, R. Machine Learning and IoT-Based Solutions in Industrial Applications for Smart Manufacturing: A Critical Review. Future Internet, (2024), 16(11), 394.
[21] Kodumuru R.; Sarkar S.; Parepally V.; Chandarana, J. Artificial Intelligence and Internet of Things Integration in Pharmaceutical Manufacturing:ASmart Synergy. Pharmaceutics, (2025), 17(3), 290.
[22] Hassani S.; Dackermann U. A Systematic Review of Advanced Sensor Technologies for Non-Destructive Testing and Structural Health Monitoring. Sensors, (2023), 23(4), 2204.
[23] Mohandas R.; Southern M.; O’Connell E.; Hayes, M. A Survey of Incremental DeepLearning for Defect Detection in Manufacturing. Big Data andCognitiveComputing, (2024), 8(1), 7.
[24] Janga B.; Asamani G. P.; Sun Z.;Cristea N. A Review of Practical AI for Remote Sensing in Earth Sciences. Remote Sensing (2023), 15(16), 4112.
[25] Melo A.; Câmara M. M.; Pinto, J. C. Data-Driven Process Monitoring and Fault Diagnosis: A Comprehensive Survey. Processes, (2024), 12(2), 251.
[26] Ferraz Júnior F.; Romero R. A. F.;Hsieh, S.-J. Machine Learning for the Detection and Diagnosis of Anomalies in Applications Driven by Electric Motors. Sensors, (2023), 23(24), 9725.
[27] Crespo-Aguado M.; Lozano R.; Hernandez-Gobertti, F.; Molner N.; Gomez-Barquero D. Flexible Hyper-Distributed IoT–Edge–Cloud Platform for Real-Time Digital Twin Applications on 6G-Intended Testbeds for Logistics and Industry. Future Internet, (2024), 16(11), 431.
[28] Mercorelli P. Recent Advances in Intelligent Algorithms for Fault Detection and Diagnosis. Sensors, (2024), 24(8), 2656.