Multivariate anomaly detection on sensor streams requires methods that handle inter-variable correlation, not just per-channel thresholds, because no single variable may breach its univariate limit while the joint state is already faulted [S1]. The 2024-2025 literature groups these methods into three families: statistical/distance-based, classical machine learning, and deep reconstruction models [S2][S5].
Public benchmarks for multivariate sensor anomaly detection (SWaT, WADI, MSL, SMAP, SMD) remain the standard testbeds in 2025 papers, and a Gramian Angular Field plus adversarial-generative model posted an 11.5% F1 gain on the high-dimensional WADI dataset versus the cited baseline [S4]. Reconstruction-based deep models are now the most frequently proposed family in the 2024-2025 survey literature [S5].
Statistical and Distance-Based Methods: When They Still Win
Hotelling's T² test and Mahalanobis distance are the canonical multivariate statistical detectors, both flagging points that deviate from the joint mean when data is approximately multivariate Gaussian [S1]. An Elliptic Envelope fit can supply the covariance estimate, and the same paper documents two Gaussian clusters with a small outlier population recovered cleanly using Mahalanobis scoring [S1]. These methods run in milliseconds on a laptop and need no training labels, which keeps them in scope for data loggers feeding low-rate, well-behaved plant signals. They fail when the data is non-Gaussian or very high-dimensional, which is precisely where the 2024-2025 deep-learning literature has concentrated its gains [S1][S2].
Classical Machine Learning: Isolation Forest, OCSVM, and Friends
Isolation Forest remains the most cited tree-based multivariate detector: anomalies isolate with fewer random splits, so a short average path length across many trees scores them as outliers [S1]. One-Class SVM and Local Outlier Factor (LOF) fill the same role for boundary-based and density-based scoring respectively, and they handle non-Gaussian clusters better than Mahalanobis distance does [S6]. These methods scale to thousands of features when feature engineering is reasonable, and they are the practical default when only a few thousand labelled-or-unlabelled multivariate windows are available from a flow sensor or gas detection array. Their hard limit is that they treat each window as an i.i.d. point and ignore temporal order, which is why 2025 reviews list them as a useful baseline, not a frontier method [S2][S5].
Reconstruction-Based Deep Learning: The 2024-2025 Dominant Family
Reconstruction-based models train an autoencoder, variational autoencoder, or GAN to reproduce normal windows; the reconstruction error itself becomes the anomaly score, and no anomalous labels are required [S5]. A 2025 systematic review notes these methods are simpler, computationally efficient, and avoid the data-hunger and context-shift problems that hurt forecasting-based and representation-based models respectively [S5]. The same review cites Zamanzadeh Darban et al. 2024 (ACM Computing Surveys 57(1):1-42) as the canonical taxonomy that classifies deep TAD into forecasting-, reconstruction-, and representation-based branches, with reconstruction methods emerging as the most widely used in practice [S5].
For high-dimensional control-equipment streams, the most cited 2024 architecture encodes 1-D time series into 2-D images via Gramian Angular Fields, then trains a GAN to flag pixel-level anomalies; the authors report an 11.5% F1 improvement on WADI versus the prior baseline, and they test on five open datasets [S4]. This image-encoding route is one branch in a broader 2025 taxonomy that also covers LSTM-AE, Transformer-AE, and diffusion-based reconstructors for multivariate time series [S2][S5].
Comparison: Statistical vs Tree-Based vs Reconstruction-Based Deep
On four practical decision criteria, the families rank as follows. Training data requirement: statistical and Isolation Forest need near-zero to a few thousand unlabelled windows; reconstruction deep models need tens of thousands of normal windows [S2][S5]. Interpretability: Mahalanobis distance and Isolation Forest path-length are directly auditable; autoencoder residuals are not, although Gramian Angular Field heatmaps are visualisable at the pixel level [S1][S4]. Scalability to 50+ sensor channels: statistical methods degrade, Isolation Forest holds up, reconstruction autoencoders and GAN-encoder hybrids are the demonstrated winners, with the cited WADI result on a high-dimensional benchmark as evidence [S4]. Suitability for streaming displacement sensor or inductive sensor data: streaming inference is comfortable for all three; only the reconstruction models natively model temporal order without bespoke feature windows [S5].
Application Domains and 2025 Case Studies
Water quality monitoring is a fast-growing application: a 2025 Environmental Modelling & Software paper uses multivariate functional data analysis to detect anomalies in water-resources sensor streams, highlighting that most field datasets are unlabelled and require unsupervised models [S3]. Industrial IoT remains the largest single driver, with 2024-2025 work specifically targeting control-system equipment where multivariate interlock behaviour matters more than per-channel thresholds [S4]. Cybersecurity, financial fraud, and healthcare vitals are the other three domains consistently named across both 2024 and 2025 reviews [S1][S2].
Limitations, Open Problems, and What to Track Next
Forecasting-based deep TAD needs large labelled anomaly corpora, which most plants cannot supply, and its accuracy drops on noisy or non-stationary streams [S5]. Representation-based methods suffer context shift when operating regimes change (e.g. product-grade changeover on a process line), so reconstruction-based hybrids are usually the safer pick [S5]. The 2024 image-encoding GAN approach still depends on five open datasets and has not yet been independently replicated on closed industrial streams [S4]. For a related spec-side decision on the sensor hardware feeding these pipelines, the Contact vs Non-Contact Displacement Sensors spec guide covers the upstream measurement choice that drives downstream anomaly model performance.