Abstract: As a pivotal technology determining the reliability of industrial systems， anomaly detection occupies an indispensable position in modern manufacturing. With escalating demands for precision， automation， and generalization capabilities in production systems， the limitations of traditional methods have become increasingly apparent. In recent years， rapid innovations in deep learning and computer vision have catalyzed a multidimensional evolution in industrial anomaly detection. In the data dimension， detection targets have transitioned from single 2D texture images to 3D point clouds rich in geometric information； in the architectural dimension， model designs deeply integrate the long-range dependency modeling capabilities of Transformers with the local perception advantages of Convolutional Neural Networks （CNNs）； regarding information utilization， a progressive shift exists toward the deep fusion of multi-source heterogeneous information. Crucially， to address the intrinsic pain points of extreme class imbalance between positive and negative samples and the unpredictable nature of anomaly morphologies in industrial scenarios， mainstream methodologies have accelerated the paradigm shift from data-intensive fully supervised learning to more flexible and efficient mechanisms such as unsupervised reconstruction， embedding-based approaches， or weakly supervised learning. Despite the exponential growth of related research achievements， the existing literature remains fragmented， lacking a systematic theoretical induction and deep analysis from a unified perspective that covers both 2D and 3D full-modal technologies. This study aims to bridge this gap by comprehensively reviewing state-of-the-art detection algorithms. We delineate the technological evolutionary trajectory along multiple dimensions， including the characteristics of dataset construction， applicability of evaluation metric systems， and engineering application of mainstream open-source frameworks. Furthermore， we provide an in-depth discussion on urgent challenges such as logical anomaly recognition and lightweight model deployment， as well as future trends involving the empowerment of large models. We expect that this study will serve as a vital reference for promoting both academic theoretical innovation and practical engineering implementation in industrial anomaly detection by offering a comprehensive， systematic， and forward-looking survey.