LatAD — Physically-Grounded Anomaly Detection for Massively-Multimodal CPS Normal Data

Read the paper: https://apartsinprojects.github.io/LatAD/ (rendered from index.html)

A proof-of-concept study on anomaly detection for predictive maintenance in transportation cyber-physical systems (CPS). The premise: the hard problem is not characterising anomalies (which are rare and unrepresentative) but modelling normal behaviour, which in CPS is Massive, Implicit, Imbalanced, Multimodal (MIIM) — a very large number of unlabeled operating modes whose frequencies span orders of magnitude. That structure induces three detector errors, and the framework here targets each one.

The framework

A joint variational latent model with a Gaussian-mixture prior (VaDE) discovers the operating modes; on top of it four components each attack one MIIM failure:

	Component	Targets	Mechanism
C1	Physically-guided anomaly generation	calibration / ranking without labels	shuffle channels across modes → oracle-filter → tune difficulty (poc/generate.py)
C2	Basin-of-attraction test	false positive on rare-but-valid modes	latent gradient-descent; rare-valid = one stable basin, pocket = split (poc/component2.py)
C3	Cluster-based ensemble	false negative on over-interpolated pockets	per-mode reconstruction experts and a supervised mode-subset ensemble trained on C1 anomalies (poc/component3.py)
C4	Mode-conditional risk control	systematic per-mode false alarms	per-mode worst-case FPR + mode-conditional thresholds (poc/component4.py)

Plus a whitened (Ledoit-Wolf Mahalanobis) reconstruction score in poc/models_vade.py.

The synthetic benchmark (poc/data.py) models CPS realism: bounded modes (truncated supports from control/physical limits), heterogeneous per-channel noise, precision, and quantization, and heavy-tailed mode imbalance — properties under which raw-distance baselines (LOF) degrade while structure-aware, whitening-based detection holds.

Headline results (synthetic MIIM benchmark)

• Joint latent+clustering foundation AUROC 0.94, full framework 0.997, vs 0.74–0.87 for Isolation Forest / LOF / AutoEncoder.
• On physically-valid dependency-violation faults, Isolation Forest and a single AutoEncoder are blind (TPR 0.00 @ 5% FPR — they score these faults as more normal than normal); the framework detects 0.92.
• C2 lowers rare-mode false-positive rate 0.93 → 0.69 (in its dedicated configuration).
• C3: a single global reconstruction model scores AUROC 0.451 on pockets (over-interpolates); the per-mode ensemble scores 0.996.
• Whitening is neutral on this isotropic-noise synthetic (expected to help on real correlated-channel data — not yet demonstrated).

First real-CPS validation: on HAI (59 sensors, multi-process) the framework leads on the imbalance-appropriate metrics — AUPRC 0.77 / TPR@5%FPR 0.77 vs AutoEncoder 0.75/0.73, LOF 0.74/0.71, Isolation Forest 0.40/0.53. On near-unimodal SKAB all methods are weak and it's only competitive — the contrast supports the thesis (mode machinery helps where normal is genuinely multimodal). WADI/SWaT pending.

Full ablation, tables, and honest limitations are in the paper.

Run

python poc/run_seeds.py --seeds 5     # main benchmark, baselines vs foundation + C2
python poc/eval_generated.py          # C1 generated faults: standard detectors are blind
python poc/component3.py              # C3: single-global vs cluster ensemble on pockets
python poc/ablation.py --seeds 3      # each component, separately and combined
python poc/make_figures.py            # regenerate results/figure.png
python poc/run_real.py hai            # real CPS: HAI (multi-process) — framework leads
python poc/run_real.py skab           # real CPS: SKAB (near-unimodal) — competitive

Requires: torch, scikit-learn, pyod, numpy, scipy, matplotlib (CUDA optional).

Data

• Synthetic MIIM generator (poc/data.py) — controlled, ground-truth modes.
• Real CPS datasets are not committed (size / redistribution). SKAB and HAI 20.07/21.03 are fetched locally under poc/datasets/ (git-ignored); WADI/SWaT are access-gated via the iTrust request form. A dataset-agnostic loader stub (load_wadi) is the single switch point.

Status

Preliminary proof of concept on synthetic data. Real-testbed validation (WADI/SWaT/HAI) is the next step. See the paper's Limitations section for the honest scope. Derived from the project executive summary.