Subject Ph. D.

Scientific context: Industry 4.0 marks a significant transformation in industrial systems through the integration of smart technologies such as the Internet of Things (IoT), artificial intelligence (AI), and cyber-physical systems (CPS/CPPS). A key component of this transformation is Predictive Maintenance (PdM) aiming to optimize maintenance strategies using data-driven methodologies. PdM leverages digitalization and data analytics to predict equipment degradation, optimize maintenance scheduling, allocate resources efficiently, reduce costs, enhance availability, and extend asset lifespan. This transition represents a crucial paradigm shift from reactive or scheduled maintenance to intelligent, data-driven maintenance strategy. Since the early 2010s, data-driven approaches have increasingly dominated prognostics, with Machine Learning (ML) methods widely adopted for PdM. However, recent trends indicate a shift toward Deep Learning (DL), as evidenced by a growing number of state-of-the-art reviews on the topic. The shift from conventional PdM methods to DL-based approaches is attributed to several key advantages. A fundamental strength of DL in PdM is its ability to extract and learn discriminative information from raw, complex, and multimodal data sources. Unlike traditional ML models that require predefined feature extraction, DL models autonomously discover relevant features, leading to improved fault detection, diagnostics, and prognostics. Additionally, DL can process large-scale heterogeneous datasets with minimal domain knowledge, making it particularly effective in industrial environments where vast amounts of sensor data are collected. However, despite these advantages, several challenges and limitations hinder the full adoption and effectiveness of DL in PdM applications. DL models frequently suffer from overfitting, particularly when trained on limited labeled datasets. This issue is critical in industrial settings, where acquiring high-quality labeled data is both costly and time-consuming. Furthermore, DL models require extensive computational resources, making their deployment in real-time industrial applications challenging. Another major limitation is interpretability⬔DL models operate as black boxes, making it difficult for maintenance engineers to trust or validate their predictions compared to physics-based or traditional statistical models, which offer explicit reasoning behind their outputs. This lack of transparency can slow the adoption of DL-based PdM solutions, particularly in industries where decision-making requires strong justifications and regulatory compliance. A major research challenge in PdM relies on capturing dynamic operating environments, that are influenced by both internal factors and external factors, into the PdM model. Traditional monitoring systems primarily focus on internal data, often neglecting the impact of external variables. This oversight can lead to misinterpretations of a machine's health state, resulting in inaccurate fauts detection and failure predictions, or lost maintenance opportunities. Furthermore, PdM models often fail to generalize across out-of-distribution conditions, such as changes in production processes or environmental shifts, which can degrade model performance. Existing DL models struggle to adapt to new operational scenarios, limiting their long-term usability. Another significant challenge is data scarcity. Many industrial environments lack sufficient labeled data to train robust models, necessitating alternative techniques such as transfer learning, synthetic data generation, and self-supervised learning. Additionally, PdM involves integrating data from multiple heterogeneous sources (e.g., vibration, sound, temperature), many existing DL models fail to fully exploit multimodal sensor data. This inability to effectively leverage diverse data sources limits predictive accuracy and holistic decision-making.

Objectives and expected contribitions: Therefore, the primary objective of this PhD is to develop a deep learning framework⬔including both models and training methodologies⬔for predictive maintenance. Several approaches will be explored to address the identified challenges:
⬢ Unsupervised and Self-Supervised Learning: Training DL models on large-scale industrial datasets to learn meaningful feature representations before fine-tuning on specific prognostic tasks. This approach enhances model robustness when labeled data is scarce.
⬢ Transfer Learning and Domain Adaptation: Investigating techniques that improve generalization across different machines and industries, addressing the challenge of OOD conditions.
⬢ Pre-training Deep Learning Networks: Utilizing unsupervised and self-supervised learning to extract generic, transferable representations from both labeled and unlabeled data. These representations will be fine-tuned for specific PdM applications, reducing dependence on large annotated datasets.
⬢ Multimodal Data Fusion: Integrating various sensor modalities (e.g., vibration, sound, temperature) to improve predictive accuracy and provide a holistic view of machine health.
⬢ Context-Awareness Mechanisms: Developing models that incorporate internal and external factors affecting machine health, improving anomaly detection and prognostics accuracy. By developing a generic and transferable DL framework, this research aims to create scalable predictive maintenance solutions that adapt across diverse industrial environments, overcome the limitations of current DL-based approaches, and enable more reliable, data-driven maintenance strategies. The integration of self-supervised learning, multimodal data fusion, and context-awareness will play a crucial role in improving model interpretability, adaptability, and efficiency in industrial applications. In addition, the work will leverage industrial datasets provided by Idemoov's customers to test and validate the proposed solution on reald-world operational data with practical industrial challenges.