Unsupervised Continual Anomaly Detection (UCAD)

• UCAD is an unsupervised framework that incrementally learns normal data patterns to detect anomalies in continuously evolving environments.
• It integrates methods like autoencoder-based scoring, coreset memory compression, and prompt-based adaptation to mitigate catastrophic forgetting.
• Benchmark results demonstrate UCAD's robust performance with high AUROC and AUPR, making it ideal for industrial inspection and dynamic auditing.

Unsupervised Continual Anomaly Detection (UCAD) is an emerging paradigm in machine learning that addresses the detection of abnormal events or data points in evolving, non-stationary environments, given only normal data for incremental training and without the use of labeled anomaly data. UCAD systems are required to assimilate new “normal” patterns as operational distributions evolve, while simultaneously resisting catastrophic forgetting of previously acquired knowledge, and to do so entirely unsupervised. This framework is increasingly critical in domains such as industrial manufacturing, dynamic auditing, and high-throughput quality control, where data distributions drift and ground-truth anomalies are scarce or unavailable (Hemati et al., 2021, Yang et al., 10 Nov 2025, Liu et al., 2024, Zhou et al., 10 Feb 2025, Zhou et al., 23 Mar 2026, Ren et al., 15 Dec 2025).

1. Formalization and Problem Context

UCAD is formally defined over a sequence of $K$ tasks or data streams $\{T_1, \dots, T_K\}$, where each task provides a (typically large) set of unlabeled “normal” data $X^t = \{x_i^{(t)}\}$. Anomaly labels are absent during training but may be available during evaluation for benchmarking. The task sequence reflects real-world, time-varying distributions: industrial lines with changing workpieces, financial systems with evolving journal types, etc. Catastrophic forgetting—whereby models tuned to new data degrade on old distributions—and the need for lightweight, real-time adaptivity are core challenges (Hemati et al., 2021, Yang et al., 10 Nov 2025, Liu et al., 2024).

Objective metrics include image- or instance-level AUROC and pixel-level AUPR for localization tasks, as well as an average forgetting measure (FM), defined across tasks as

$$FM = \frac{1}{K-1}\sum_{j=1}^{K-1} \max\left(a_j^{\text{max}} - a_j^{(K)}\right)$$

where $a_j^{(K)}$ is the accuracy/AUROC on task $j$ after learning all $K$ tasks (Zhou et al., 10 Feb 2025, Zhou et al., 23 Mar 2026).

2. Core Methodologies

UCAD research assembles mechanisms from deep unsupervised anomaly detection and continual learning. Foundational strategies include:

• Autoencoder-based Streaming: Early work (Hemati et al., 2021) employs deep autoencoders (AEs) with continual learning updates (EWC, replay) to compress high-dimensional transaction or feature data. The AE outputs a reconstruction loss for each sample that acts as an anomaly score.
• Coreset-based Memory Compression: Coreset strategies use fixed-size memory banks (embeddings obtained from a frozen or prompt-tuned backbone) updated via incremental $k$-center selection or dynamic replacement rules, ensuring global coverage of past “normal” patterns (Yang et al., 10 Nov 2025, Ren et al., 15 Dec 2025). A representative algorithm is as follows:

$\{T_1, \dots, T_K\}$6 (Yang et al., 10 Nov 2025)

• Prompt-based Continual Adaptation: Recent advances leverage learnable prompts (visual, textual, or multimodal) inserted into frozen backbones (ViT, CLIP) for task incremental learning. Prompt banks store task-specific adaptation vectors, keys for retrieval, and condensed normal prototype banks; only these banks are maintained, preventing raw data replay (Liu et al., 2024, Zhou et al., 10 Feb 2025, Zhou et al., 23 Mar 2026).
• Contrastive and Structural Losses: Structure-based contrastive learning (SCL/RSCL) builds feature compactness/robustness by exploiting external segmentation (e.g., SAM or Grounding DINO) to form semantically homogeneous regions and contrasting embeddings across regions (Liu et al., 2024, Zhou et al., 10 Feb 2025). Cross-modal contrastive terms further align vision and text representations.
• Meta-Learning for Sequence Adaptation: Bilevel “MAML”-style meta-learning can initialize model parameters and individual step sizes to optimize global retention and minimize forgetting across task sequences (Frikha et al., 2020).

3. Key Architectural Elements

UCAD systems integrate several essential architectural and algorithmic components, adapted to the specifics of the streaming environment:

Component Category	Example Techniques/Modules	References
Memory management	Coreset (incremental $k$-center, FPS), task-prototype bank	(Yang et al., 10 Nov 2025, Ren et al., 15 Dec 2025)
Adaptation mechanism	Prompt banks (visual, text, multimodal), EWC, replay buffer	(Liu et al., 2024, Zhou et al., 10 Feb 2025, Hemati et al., 2021)
Feature extractor	ViT, CLIP, MobileNetV3 (for edge), BERT (text modality)	(Yang et al., 10 Nov 2025, Ren et al., 15 Dec 2025, Zhou et al., 23 Mar 2026)
Anomaly scoring	Nearest-neighbor on prototypes, AE reconstruction loss	(Hemati et al., 2021, Yang et al., 10 Nov 2025, Liu et al., 2024)
Losses	Reconstruction, contrastive, cross-modal, regularization	(Zhou et al., 10 Feb 2025, Liu et al., 2024, Zhou et al., 23 Mar 2026)

A full pipeline may entail frozen backbone feature extraction, mid-layer patch sampling (by farthest-point or coreset selection), prompt injection (prefix tuning), and one or more learned prototype banks for each task. At inference, task keys index the correct prompt/prototype set, and anomaly scores are derived by nearest-neighbor or contrastive matching.

4. Mitigating Catastrophic Forgetting

UCAD approaches are fundamentally evaluated on their ability to prevent catastrophic forgetting: the degradation in detection performance on earlier tasks after learning subsequent data. Major mitigation strategies include:

• Unified vs. Fragmented Memory: Methods holding a single, global coreset (e.g., CADIC) empirically avoid fragmentation and maintain low forgetting ($FM{<}0.02$) (Yang et al., 10 Nov 2025).
• Prompt-based Modularity: Task-identity keys linked with split prompt sets afford sublinear memory growth while permitting precise per-task adaptation (Liu et al., 2024, Zhou et al., 10 Feb 2025, Zhou et al., 23 Mar 2026).
• Contrastive Compaction: SCL/RSCL compact normal representations and decorrelate structurally distinct regions, yielding low “drift” in patch-space and diminished memory bloat (Liu et al., 2024, Zhou et al., 10 Feb 2025).
• Meta-learned Update Schedules: Task-specific, parameter-wise learning rates can isolate layers more prone to forgetting and suppress overfitting, as shown in ARCADe’s $\{T_1, \dots, T_K\}$0-distribution analysis (Frikha et al., 2020).

For all high-performing methods, empirical $\{T_1, \dots, T_K\}$1 remains below 0.02 on benchmarks, with MTRMB and CMPMB architectures reporting additional robustness due to multimodal fusion (Zhou et al., 10 Feb 2025, Zhou et al., 23 Mar 2026).

5. Multimodality and Adaptive Fusion

The inclusion of multimodal information—especially vision and language—enhances the expressiveness of “normality” and fortifies anomaly detection in complex scenes (Zhou et al., 23 Mar 2026, Zhou et al., 10 Feb 2025). The CMPMB framework maintains, for each task, keys, visual/text prompts, and feature banks. Defect-Semantic-Guided Adaptive Fusion (DSG-AFM) normalizes and fuses visual and text-based anomaly cues:

• Adaptive Normalization Module (ANM): Scores from different modalities are squashed and thresholded by adaptive sigmoidal functions with learnable parameters to optimally separate normal/anomaly.
• Dynamic Fusion Strategy (DFS): Fused anomaly maps $\{T_1, \dots, T_K\}$2 leverage the strengths of both visual ($\{T_1, \dots, T_K\}$3) and textual ($\{T_1, \dots, T_K\}$4) streams, with $\{T_1, \dots, T_K\}$5 tuned for detection/localization tradeoff (Zhou et al., 23 Mar 2026).

On MVTec AD and VisA datasets, multimodal prompt-based UCAD yields substantial absolute gains in both AUROC and AUPR, including +14.8% AUPR improvement over visual-only methods (Zhou et al., 23 Mar 2026).

6. Experimental Protocols and Quantitative Benchmarks

UCAD methods are systematically evaluated on sequential benchmarks such as MVTec AD and VisA, with sequential per-category task increments, using only normal data for training. Protocols enforce strict task ordering and zero raw-data replay. Baselines include single- and joint experience learners, AE-only detectors, naive replay, and state-of-the-art continual learning (Hemati et al., 2021, Liu et al., 2024).

Key results (MVTec AD, final task):

Method	AUROC (image)	AUPR (pixel)	FM
UCAD (visual, prompt-CL)	0.930	0.456	0.013
CADIC (k-center coreset)	0.972	0.584	<0.02
MTRMB (multimodal)	0.941	0.468	0.016
CMPMB (multimodal)	0.974	0.604	0.009
ARCADe (meta-learned)	≈0.94–0.96†	—	—

†Depends on dataset/task sequence; see (Frikha et al., 2020).

UCAD with adaptive (multimodal) prompt memory consistently reports state-of-the-art results in both detection accuracy and retention.

7. Practical Implications, Limitations, and Future Directions

UCAD pipelines have demonstrated significant improvements in domains such as continuous auditing (Hemati et al., 2021), industrial visual inspection (Liu et al., 2024, Yang et al., 10 Nov 2025, Zhou et al., 23 Mar 2026), and edge-AI manufacturing (Ren et al., 15 Dec 2025). On-device implementations using lightweight architectures (MobileNetV3 with k-center coreset) provide >12% AUROC gain and >80% memory reduction over traditional retrain+replay solutions (Ren et al., 15 Dec 2025).

Key limitations include:

• Memory Bank Growth: Even with coreset/prototype selection, bank size grows with task count; further prototype condensation and importance weighting are areas for research (Zhou et al., 23 Mar 2026, Zhou et al., 10 Feb 2025).
• Modality Scalability: Current approaches mainly fuse vision and language; extending to richer modalities (e.g., time series, 3D, sensors) remains open.
• Forgetting Tradeoff: Multimodal adaptation may marginally increase forgetting rates; regularization or efficient rehearsal-free stabilizers are prospective solutions (Zhou et al., 23 Mar 2026).
• Absence of explicit drift detection: Most approaches rely on continual learning itself to maintain stability under distribution shift, though active drift detection is rarely implemented directly (Hemati et al., 2021).

Research directions include integrated drift detectors, scalable cross-modality knowledge banks, efficient on-device bank compression, and open-set or semi-supervised extensions. The research trajectory indicates growing capabilities for robust, label-free monitoring of complex systems under continuous evolution.

References:

• (Hemati et al., 2021): Continual Learning for Unsupervised Anomaly Detection in Continuous Auditing of Financial Accounting Data
• (Yang et al., 10 Nov 2025): CADIC: Continual Anomaly Detection Based on Incremental Coreset
• (Liu et al., 2024): Unsupervised Continual Anomaly Detection with Contrastively-learned Prompt
• (Zhou et al., 10 Feb 2025): Multimodal Task Representation Memory Bank vs. Catastrophic Forgetting in Anomaly Detection
• (Zhou et al., 23 Mar 2026): Exploring Multimodal Prompts For Unsupervised Continuous Anomaly Detection
• (Ren et al., 15 Dec 2025): On-Device Continual Learning for Unsupervised Visual Anomaly Detection in Dynamic Manufacturing
• (Frikha et al., 2020): ARCADe: A Rapid Continual Anomaly Detector