A comprehensive bibliometric and topic-modelling evaluation of global scientific literature has revealed that anomaly detection research is rapidly expanding far beyond its traditional heartlands of network cybersecurity and financial fraud. Published in the journal Management and Marketing, the investigation systematically examined thousands of peer-reviewed articles to chart how algorithmic surveillance and deviation analysis operate across physical, biological, and engineering environments. For decades, foundational review papers, benchmark datasets, and high-visibility surveys focused almost exclusively on detecting malicious network intrusions and fraudulent electronic transactions. While those security-oriented and banking applications continue to generate substantial commercial interest, researchers found that non-security and non-financial implementations have surged in scholarly volume. Across the analyzed period from 2022 through 2025, scientific publications in these alternative sectors grew at an annual compound rate of 31.14 percent. This steady trajectory demonstrates that mathematical techniques designed to flag data points generated by unexpected probability distributions or irregular physical mechanisms are becoming central to real-world industrial automation, environmental tracking, and clinical diagnostics.
To overcome the historical skew that has long centered literature reviews on intrusion detection and credit risk, the research team constructed an empirical retrieval pipeline from the Web of Science Core Collection. An unconstrained baseline search for terms linked to anomaly, outlier, and novelty detection retrieved 23,653 documents, illustrating the vast scale of the broader discipline. To isolate previously neglected domains without completely divorcing the corpus from practical deployment contexts, the authors established a structured query incorporating three distinct negative thematic blocks: cybersecurity and digital defense, finance and fraud detection, and networking or communications terminology. This targeted filtering reduced the initial retrieval pool to 16,485 records. The corpus was subsequently narrowed to English-language journal articles published strictly between 2022 and 2025, yielding 5,154 raw candidate items. The investigators then deployed programmatic deduplication routines in Python based on Digital Object Identifiers and metadata records, while removing documents containing fewer than 120 characters across their combined titles and abstracts. This rigorous curation yielded a final analysis corpus comprising 4,362 substantive research papers.
Textual preparation of the curated collection followed a four-stage natural language processing workflow engineered to surface applied domain context rather than recurring algorithmic jargon. Raw titles and abstracts underwent normalization through character cleaning, lowercasing, URL deletion, and tokenization requiring a minimum token length of three characters. The authors enforced a multi-layered stopword strategy that eliminated standard grammatical terms alongside recurring methodological vocabulary, such as deep learning, transformers, autoencoders, and clustering. They also stripped core query anchors including anomaly, outlier, detect, and detector, preventing the algorithms from building trivial topics around ubiquitous baseline terminology. Morphological processing was conducted using the spaCy library without syntactic dependency parsers or named entity recognition modules, retaining only tokens classified as nouns, adjectives, or proper nouns. Phrase mining was subsequently executed using Gensim to extract domain-specific bigrams meeting minimum frequency and association thresholds, capturing meaningful compound concepts such as predictive maintenance and hyperspectral imaging before building the final vocabulary dictionary.
Unsupervised semantic mapping was executed using Gensim’s Latent Dirichlet Allocation algorithm across a bag-of-words dictionary filtered to retain terms appearing in at least 15 documents and no more than 70 percent of records. To determine the most informative structural resolution, the authors conducted a parameter grid search across candidate topic counts of 8, 10, 12, and 14, evaluating output using the standardized c_v topic coherence metric. The 12-topic configuration yielded the optimal coherence score of 0.4657, outperforming alternative cluster counts. Stability assessments across five random initialization seeds confirmed structural reliability, producing consistent overarching themes with an average Spearman rank correlation of 0.641 and an average Jaccard term overlap of 0.40. The resulting statistical model decomposed the corpus into nine distinct application-oriented domains and three general methodological toolkits, specifically ensemble classification, multivariate temporal modeling, and sparse or low-rank tensor representations that serve as transferable mathematical instruments across heterogeneous physical disciplines.
The study quantified topic prevalence through two complementary analytical perspectives: a dominant-topic assignment where each paper is categorized by its highest-probability cluster, and a continuous mixture-share calculation that tallies fractional topic memberships across the entire corpus. Under the mixture-share assessment, the single largest research concentration centered on behavioral and environmental patterns, which represented approximately 13.3 percent of the analyzed text. This expansive cross-domain category aggregates studies where anomalous behavior manifests as deviations from complex temporal baselines, ambient ecological shifts, or abnormal human activities. Close behind were energy systems faults and maintenance, capturing an estimated 11.6 percent share, and process control and industrial operations, accounting for 10.5 percent. When evaluated by singular dominant-topic assignment, these three practical concentrations together accounted for 37.7 percent of the entire scientific corpus, confirming that heavy engineering assets, industrial automation, and distributed sensing environments constitute the foundational pillars of non-security deviation modeling.
Tracking mixture-based topic trajectories from 2022 to 2025 revealed substantial structural shifts in the relative distribution of research effort across the nine application areas. Industrial visual inspection exhibited the steepest and most sustained upward momentum of any domain, expanding from a 5.50 percent share in 2022 to 9.99 percent in 2025, an overall absolute growth of 3.38 percentage points. Internet of Things edge-cloud and graph monitoring also experienced steady expansion, climbing from 7.31 percent to 8.60 percent, reflecting an absolute increase of 1.30 percentage points. Behavioral and environmental monitoring maintained its rank as the largest overall application area, edging upward from 12.50 percent to 13.45 percent. Conversely, energy systems fault diagnosis experienced a marked reduction of 2.14 percentage points over the study window despite a mild rebound in 2025. Industrial process control slipped by 1.12 percentage points to flatten near 10 percent, while healthcare and biomedical diagnosis registered a gradual decline of 1.14 percentage points.
The rapid acceleration of industrial visual inspection directly mirrors the operational integration of automated computer vision pipelines within manufacturing quality assurance. High-impact literature within this topic emphasizes benchmark datasets and deep learning architectures capable of identifying both localized surface blemishes and complex structural defects in components such as photovoltaic cells, textiles, wood, and steel. Rather than merely spotting simple surface scratches, contemporary inspection models are increasingly trained to detect logical anomalies, such as misplaced components or incorrect sub-assemblies. Conversely, the relative contraction observed in energy systems faults and predictive maintenance does not imply diminished operational relevance. Instead, highly cited papers indicate that foundational methodologies—specifically semi-supervised models trained on vibration signals and healthy operating baselines—have reached technical maturity. As vibration analysis and multi-sensor fusion become routine predictive maintenance standards, novel contributions are increasingly integrated into broader operational frameworks, potentially shifting publication classification toward generalized process control.
To evaluate whether the twelve text-derived topics genuinely reflected distinct operational domains rather than computational artifacts of the modeling software, the researchers performed an external validation against official Web of Science journal subject categories. This categorical cross-referencing demonstrated robust disciplinary alignment with the discovered topics. Hyperspectral remote sensing, which commanded approximately 6.1 percent of the topic mixture, exhibited dominant associations with Remote Sensing as well as Imaging Science and Photographic Technology. Industrial visual inspection was tightly aligned with Artificial Intelligence and Electrical Engineering categories, reflecting its computer-vision foundation. IoT edge-cloud and graph monitoring clustered heavily within Telecommunications, underscoring its reliance on distributed communications infrastructure, while energy systems maintenance demonstrated pronounced links to Instrumentation and Multidisciplinary Engineering, verifying that Latent Dirichlet Allocation effectively captured authentic disciplinary boundaries.
The researchers further corroborated the intellectual topography of the field by conducting source-level bibliographic coupling in R across the top 113 publishing venues, which represented roughly 60 percent of total paper output. Applying Louvain community detection to the weighted citation network revealed four distinct publication communities consisting of 68, 32, 10, and 3 journals respectively. The largest venue cluster constituted a broad publishing core encompassing applied artificial intelligence and general engineering periodicals. The second community clustered tightly around specialized remote sensing and earth observation journals, such as IEEE Transactions on Geoscience and Remote Sensing. The third venue cluster grouped image processing, machine vision, and video analysis publications, directly aligning with visual quality control and automated surveillance. A fourth, highly compact community was formed by specialized machine learning journals. Crucially, the analysis revealed that remote sensing and industrial vision venues remain publication-isolated despite sharing profound methodological overlaps in high-dimensional imaging analysis.
The authors underscored several methodological boundaries and governance considerations that must frame the interpretation of their findings. Restricting the source collection to English-language journal articles within the Web of Science Core Collection inevitably excluded relevant studies indexed in Scopus, conference proceedings, preprints, or non-English outlets. Furthermore, while screening confirmed the effectiveness of negative search terms, the retrieval exclusions could potentially omit legitimate non-security networking studies. The authors emphasized that topic modeling treats all documents equally, whereas citation distributions within each cluster remain heavily skewed toward a small number of landmark benchmarks and critical reviews. Moving forward, the researchers highlighted pressing governance and operational challenges that accompany real-world deployment across non-security domains, including alarm fatigue caused by excessive false positives in high-throughput industrial plants, data privacy violations in ambient human tracking, and diagnostic risks stemming from opaque deep learning architectures trained solely on unlabelled normal operating conditions.
Subject of Research: Bibliometric and topic-modelling analysis of anomaly detection applications beyond cybersecurity and finance
Article Title: Mapping the recent landscape of anomaly detection beyond cybersecurity and finance: a Web of Science bibliometric and topic-modelling review
Article References: Albu, O., Chinie, C., & Berbece, D.-C. (2026). Mapping the recent landscape of anomaly detection beyond cybersecurity and finance: a Web of Science bibliometric and topic-modelling review. Management & Marketing, 21(1), Article 6. https://doi.org/10.1007/s44491-026-00006-9
Image Credits: AI Generated
DOI: 10.1007/s44491-026-00006-9
Keywords: anomaly detection, topic modelling, latent dirichlet allocation, bibliometric analysis, industrial visual inspection, behavioral pattern recognition, energy systems maintenance, process control, IoT edge-cloud monitoring, hyperspectral remote sensing, healthcare diagnostics, video surveillance, bibliographic coupling, web of science, sensor networks, predictive maintenance, automated quality assurance, unsupervised learning
Cite Scienmag News
Hailey Crawford. (November 19, 2026). Study Maps Anomaly Detection Beyond Security With Visual Inspection Surging Fastest. Scienmag. https://scienmag.com/study-maps-anomaly-detection-beyond-security-with-visual-inspection-surging-fastest/
Hailey Crawford. "Study Maps Anomaly Detection Beyond Security With Visual Inspection Surging Fastest." Scienmag, 19 November 2026, https://scienmag.com/study-maps-anomaly-detection-beyond-security-with-visual-inspection-surging-fastest/. Accessed 19 November 2026.
Hailey Crawford. "Study Maps Anomaly Detection Beyond Security With Visual Inspection Surging Fastest." Scienmag. November 19, 2026. https://scienmag.com/study-maps-anomaly-detection-beyond-security-with-visual-inspection-surging-fastest/








