Comparative Analysis of AI-Based Predictive Maintenance Solutions

Sammanfattning

Background: Industrial sectors face growing pressure to optimize maintenance processes and minimize downtime, as unexpected equipment failures can lead to sig- nificant safety risks, production losses, and increased operational costs. Despite the availability of large volumes of operational data from sensors and industrial systems, much of this data remains underutilized due to factors such as data silos, lack of ef- fective analytical tools, and challenges in processing complex data streams. AI-based predictive maintenance systems are now being introduced as a key component of the Industry 4.0 transformation, offering the potential to enhance fault detection, opti- mize maintenance schedules, and improve overall operational efficiency by leveraging advanced analytics and automation. However, challenges remain in selecting and im- plementing the most appropriate solutions. Objective: This thesis investigates the structure and functionality of AI-driven predictive maintenance solutions in real-world applications. The primary focus is a comparative analysis of six leading commercial systems, exploring their technical ca- pabilities, architectural approaches, and diagnostic features. The analysis specifically examines key aspects such as deployment models, data integration capabilities, and the sophistication of the embedded AI algorithms. To complement the system-level assessment and provide practical context, two case studies are conducted using open datasets: one based on sensor data (vibration) and one on image data to evaluate the real-world applicability and comparative performance of different AI methods for fault detection. Methods: The study follows the CRISP-DM framework, combining a qualitative system comparison with experimental data modeling. The system analysis draws from white papers and technical documentation of six commercial platforms: IBM Maximo Predict, Siemens MindSphere, ABB Ability Genix, Microsoft Azure IoT, SAP Predictive Asset Insights, and GE Vernova APM. The case studies apply a range of machine learning and deep learning models (e.g., Random Forest, LSTM, CNN, ResNet-18) to explore and compare the strengths and limitations of sensor- and image- based fault detection. Results: The system comparison highlights key differences in deployment mod- els, integration capabilities, and AI sophistication across vendors. While enterprise- oriented platforms prioritize interoperability with existing IT systems, operations- focused solutions offer tighter hardware integration for specific equipment. The case studies demonstrate that deep learning models (LSTM, GRU, ResNet-18) generally achieve superior fault detection accuracy compared to classical machine learning mod- els, but also underscore the critical importance of data quality and effective domain adaptation techniques. Furthermore, image-based detection is shown to be highly ef- fective for identifying surface defects, while sensor data analysis offers a broader range of diagnostic capabilities for internal and operational fault conditions. Conclusions: AI-based predictive maintenance offers significant potential for op- timizing industrial operations, but successful implementation requires careful align- ment between technical capabilities, data infrastructure, and organizational needs. 1The study emphasizes the importance of hybrid approaches that combine multiple data types and modeling techniques to maximize diagnostic coverage and robustness. Furthermore, the role of AI as a support tool for knowledge transfer—particularly in preserving tacit operational knowledge—emerges as a critical, yet often overlooked, benefit. By enabling structured knowledge capture and facilitating intergenerational learning, AI contributes not only to operational efficiency but also to long-term orga- nizational resilience. Keywords: Predictive Maintenance, AI Systems, Fault Detection, Comparative Analysis, Vibration Analysis, Image Classification, CRISP-DM, Industry 4.0, Knowl- edgeTransfer, TacitKnowledge, SECIModel,SystemArchitecture, MaintenanceKPIs