Multi-type Anomaly (MulA) Detection

• MulA is defined as data anomalies manifesting in varied types, scales, and modalities, requiring specialized detection frameworks.
• Unified methodologies leverage Bayesian models, deep metric learning, and multi-modal fusion to overcome limitations of single-class detectors.
• Applications in industrial inspection and system monitoring demonstrate MulA's impact on enhancing segmentation, classification, and actionable anomaly detection.

A multi-type anomaly (MulA) refers to the phenomenon wherein data deviations manifest in multiple forms, categories, or modalities. Foundational MulA research identifies anomaly types spanning variable scales, data structures, relational complexities, and semantic classes. Recent methodologies emphasize unified detection, segmentation, and classification across diverse anomaly types, addressing the limitations of single-class, modality-specific, or label-dependent detection models.

1. Foundational Typologies and Dimensions of Multi-type Anomaly

Early and contemporary works establish that MulA is not a single phenomenon but a taxonomic landscape structured along several orthogonal dimensions. According to a comprehensive typology (Foorthuis, 2020, Foorthuis, 2021), five primary data-centric dimensions underpin anomaly characterization: data type (numerical, categorical, mixed), cardinality of relationship (univariate vs. multivariate), anomaly level (atomic/point vs. aggregate/group), data structure (independent, temporal, spatial, or relational), and data distribution (global, local, density-based, or cluster-wise). Across these, three broad groups and up to nine basic types (with dozens of subtypes) are enumerated—such as univariate extremes, rare class/categorical, mixed-type, multivariate peripheral, group-level trend anomalies, and distributional anomalies.

A key insight is that flexible frameworks must detect anomalies that may be visible only univariately (extreme value), emerge in local multivariate contexts (rare attribute combinations), or appear only at aggregate/group levels (e.g. a sequence whose statistics change over time but whose points are locally normal). This drives the need for technically rigorous, multidimensional MulA detection paradigms.

2. Multi-Modal, Multi-View, and Multi-Perspective Detection

Modern industrial and cyber-physical datasets often exhibit multi-view or multi-modal characteristics, where different sensors, data sources, or observation angles yield varied and complementary information. Several advanced models specifically address MulA from a multi-view or multi-modal perspective:

• Multi-view anomaly detection via probabilistic latent variable models (Iwata et al., 2014): Each instance is modeled as a collection of views generated from potentially multiple latent vectors, inferring inconsistency via nonparametric Bayesian mixture modeling. Non-anomalous data use a shared latent factor; anomalies reveal themselves through the necessity for multiple latent factors. This generative approach robustly detects anomalies manifesting only as inter-view inconsistencies.
• Deep structured cross-modal anomaly detection (Li et al., 2019): By mapping different modalities (such as image and text) into a consensus latent space and optimizing both “pull” (aligning consistent cross-modal pairs) and “push” (separating inconsistent pairs), this framework identifies cross-modal anomalies that are not apparent in any individual modality.
• Multi-source and multi-perspective approaches (Bogatinovski et al., 2021, Jakob et al., 2021, Li et al., 19 Dec 2024): Methods such as the MulSen-AD dataset and associated late-fusion classifier (Li et al., 19 Dec 2024), or joint LSTM embedding of logs and traces (Bogatinovski et al., 2021), illustrate high detection efficacy by integrating heterogeneous information channels. Empirical results show substantial performance gains (e.g., >95% AUROC for image-level multi-sensor detection) by combining RGB, 3D shape, and thermal properties.

3. Multi-Class and Defect-Type Anomaly Detection: Representation and Supervision

The migration from single-class anomaly detection to multi-class and multi-defect-type scenarios introduces requirements for maintaining discrimination, sample efficiency, and robustness to class imbalance:

• Multi-class and anomaly multi-classification frameworks (Singh et al., 2021, Liu et al., 9 Jun 2024, Heo et al., 4 Aug 2025): DeepMAD (Singh et al., 2021) shows that leveraging per-class contrastive encoders to learn compact, discriminative representations achieves higher AUC than naïvely combining one-class detectors or pooling all classes. HierCore (Heo et al., 4 Aug 2025) extends this idea by clustering semantic features in the absence of class labels, forming hierarchical memory banks that support both class-aware and class-agnostic thresholding during detection.
• Zero-/few-shot and defect-aware models (Sadikaj et al., 9 Apr 2025, Zhang et al., 5 Aug 2025): MultiADS (Sadikaj et al., 9 Apr 2025) demonstrates zero-shot, defect-type segmentation by aligning patch and text embeddings in CLIP’s joint feature space, producing per-type anomaly masks. ADSeeker (Zhang et al., 5 Aug 2025) exploits type-level annotation in the MulA dataset and knowledge retrieval from a curated, visual document base, achieving state-of-the-art zero-shot detection through hierarchical sparse prompts and retrieval-augmented reasoning.

The following table summarizes representative multi-type anomaly paradigms, their coverage, and supervision assumptions:

Approach/Framework	Multi-type Axis	Supervision/Label
Probabilistic Latent	Multi-view, modality	Unsupervised
DeepMAD/HierCore	Multi-class, unified	Label-agnostic / weak
MultiADS/ADSeeker	Defect-type & reasoning	Zero-/few-shot
MulSen-AD/M²AD	Multi-sensor, multi-system	Unsupervised

4. Learning Paradigms and Technical Solutions

A wide spectrum of method families address the MulA challenge, each chosen according to anomaly structure, available data, and application constraints:

• Nonparametric Bayesian models (Iwata et al., 2014): Probabilistic models with Dirichlet process priors adaptively learn the number of latent components per instance, offering flexibility for unknown class or view structure.
• Deep metric learning and clustering (Singh et al., 2021, Dong et al., 23 Aug 2024, Heo et al., 4 Aug 2025): Class-specific encoders with supervised contrastive loss (DeepMAD), multi-normal prototype learning with deep clustering and contrastive regularization (Dong et al., 23 Aug 2024), or semantic clustering without explicit labels (HierCore) allow for compact representations robust to class imbalance and unlabeled anomaly contamination.
• Multi-head/multi-hypotheses architectures (Nguyen et al., 2018): Multiple-hypothesis autoencoders learn to locally fit the multi-modal support of normal data, supporting diverse and rare anomaly types; adversarial loss with a discriminator enforces mode coverage, reducing artificial artifact generation.
• Feature, label, and view fusion (Li et al., 19 Dec 2024, Jakob et al., 2021, Yu et al., 30 Apr 2025): Early and late fusion methods, semi-frozen encoders with prior enhancement (Yu et al., 30 Apr 2025), and anomaly amplification modules facilitate the coherent integration of multiple data perspectives.
• Language–vision and multi-model fusion (Lee et al., 13 Jun 2025, Sadikaj et al., 9 Apr 2025, Zhang et al., 5 Aug 2025): Models such as CLIPFUSION (Lee et al., 13 Jun 2025), MultiADS, and ADSeeker combine discriminative image-text alignments with generative diffusion model features or knowledge retrieval, capturing both global semantic context and local defect detail.

5. Evaluation Protocols and Empirical Findings

Growing availability of multi-type datasets and benchmarks enables systematic validation of MulA frameworks under real-world, multi-class, and multi-modal scenarios:

• Dataset diversity: Modern MulA benchmarks, such as MulA (72 defect types, 26 categories) (Zhang et al., 5 Aug 2025), MulSen-AD (across RGB, infrared, point cloud; 15 products, 14 anomaly types) (Li et al., 19 Dec 2024), and MvTec-AD/3D, contain comprehensive annotations spanning scale, type, and modality.
• Metric selection and scenario coverage: Performance is typically reported using AUROC, AUPRO (pixel-wise/region overlap), F1, and precision–recall. HierCore’s robustness is evaluated across four scenarios (training and evaluation with/without class labels), emphasizing threshold adaptability (Heo et al., 4 Aug 2025).
• Empirical results: Multi-modal and multi-view fusion yields consistent AUROC gains (~1–10% absolute increase) compared to single-modal baselines. Unified models that avoid per-class retraining (e.g., MAAE, MultiADS, HierCore) exhibit both scalable efficiency and comparable or superior detection/localization scores in industrial image benchmarks. Multi-normal prototype and contrastive clustering approaches (Dong et al., 23 Aug 2024) are shown to improve precision–recall, especially under anomaly contamination and imbalanced class statistics.

6. Applications, Implications, and Future Research

The goal of MulA research is the reliable discrimination, segmentation, and categorization of diverse anomalies under industrial, cyber-physical, and data science settings:

• Industrial defect inspection: Unified, defect-type–aware detection and segmentation (e.g., MultiADS, ADSeeker, MAAE) enable detailed reasoning and targeted quality control actions, reducing false positives and improving actionable explainability (Sadikaj et al., 9 Apr 2025, Zhang et al., 5 Aug 2025, Liu et al., 2023).
• System monitoring and maintenance: Frameworks such as M²AD (Alnegheimish et al., 21 Apr 2025) and MulSen-AD (Li et al., 19 Dec 2024) show efficacy for predictive maintenance and anomaly detection in distributed infrastructure, highlighting the importance of calibrated, sensor-aware scoring.
• Medical and scientific imaging: Multi-modal fusion and multi-type segmentation approaches generalize to disease identification, experimental quality control, and other scientific imaging applications.
• Scalability and deployment: Unification paradigms (single models for MulA), hierarchical memory banks, and training-free models (CLIPFUSION) reduce computational cost, support practical thresholding, and facilitate plug-and-play deployment.
• Outlook: Key directions include improving knowledge-based reasoning and cross-domain adaptation (as in ADSeeker), enhancing few-shot and zero-shot learning for anomalous defect types under data scarcity, and developing richer, application-tailored benchmarks that capture the full variety of MulA phenomena.

7. Conceptual and Methodological Significance

Multi-type anomaly detection research has converged on several unifying principles:

• Effective MulA algorithms must jointly address within-class heterogeneity (multi-modal normality), anomaly diversity (semantic and spatial/temporal structure), and regulative flexibility (label presence/absence, adaptable thresholding).
• Theoretical developments—such as Bayesian nonparametrics for latent allocation, contrastive learning for compactness, and score calibration for sensor/system heterogeneity—have informed practical, scalable solutions.
• Through typological frameworks, functional evaluation is now possible for diagnosing algorithmic coverage over the “anomaly type spectrum” rather than relying solely on single-class benchmarks or black-box scores.
• A plausible implication is that continued cross-fertilization between probabilistic modeling, deep representation learning, and knowledge-augmented reasoning will be central to advancing the MulA field, especially for integrating anomaly type identification and actionable insights at scale.