Deep, Generative, and Explainable Models for Anomaly Detection in Smart Manufacturing

This chapter explores the advances brought by deep and generative models in anomaly detection for smart manufacturing. It begins by discussing the limitations of classical supervised and unsupervised approaches when faced with high-dimensional, non-stationary, and data-scarce industrial environments. We then examine how autoencoders, variational autoencoders, generative adversarial networks (GANs), and transformers provide powerful mechanisms for modeling normal behavior, generating synthetic data, and capturing long-range dependencies in multivariate time series. Transfer learning is highlighted as a key enabler of adaptability, allowing knowledge to be reused across processes and equipment while reducing the need for costly retraining. Beyond algorithmic performance, the chapter emphasizes the importance of explainability and human-centered AI as prerequisites for trustworthy deployment, ensuring that models remain transparent and actionable in real-world production settings. We also review emerging evaluation metrics tailored to time series anomaly detection, which account for temporal adjacency, event duration, and detection latency dimensions often overlooked by classical measures. By integrating deep learning, generative modeling, transfer learning, and explainable AI, this chapter provides a holistic perspective on next-generation anomaly detection systems, supporting predictive maintenance, quality assurance, energy efficiency, product traceability, and resilient manufacturing in the era of Industry 5.0.