Physics-Constrained Anomaly Detection

16 concepts14 papers1 proposal

The Limits of Pure Statistics

Standard machine learning models excel at finding statistical anomalies—data points that deviate significantly from a learned distribution. However, in complex engineering and physical systems, this reliance on pure statistics creates a dangerous blind spot: what happens when an anomaly is statistically plausible but physically impossible?

Consider a sophisticated cyber-attacker executing a False Data Injection Attack (FDIA) on an industrial plant or power grid. By carefully manipulating sensor readings, the attacker can forge data that mimics normal operational fluctuations. As noted by Zheng et al., purely data-driven methods operating in isolation from physical constraints struggle to separate these malicious manipulations from legitimate technical fluctuations. Similarly, Mushtaq et al. demonstrate that conventional data cleaning techniques frequently miss "hidden outliers" that fall perfectly within valid statistical ranges but flatly contradict the physical principles of the system.

To catch these stealthy deviations, a growing body of research anchors anomaly detection in the immutable laws of nature—such as conservation of energy, fluid dynamics, and mass balances. By treating physical laws as an uncheatable baseline, these systems can identify anomalies not by how unusual they look, but by whether they violate reality.

The Physical Layer as an Uncheatable Baseline

Why are physical constraints such an effective filter? While an attacker or a failing component can easily manipulate data packets or statistical distributions, they cannot alter the underlying physics of the physical infrastructure.

In modern smart grids, for instance, attackers can attempt to bypass traditional IT filters by injecting synthetic data that mimics real fault conditions. However, Mashima et al. explain that physics-based monitoring can expose these stealthy attacks by verifying whether reported measurements adhere to fundamental principles like Kirchhoff's laws.

A powerful application of this concept involves "zero-injection nodes"—junctions in a power grid where no power is added or consumed, meaning the net current entering the node must strictly equal zero. As shown by Chu et al., these nodes serve as physical anchors for data integrity. If an attacker fails to manipulate every single sensor within the exact topological neighborhood of a zero-injection node, Kirchhoff's Current Law is violated, immediately exposing the physical inconsistency.

   [Sensor A] ---> ( Zero-Injection Node ) ---> [Sensor B]
                        Input = Output
             (If A and B disagree, physics flags an anomaly)

This paradigm shift requires moving away from network packet analysis and focusing directly on the physical layer. As Kravchik demonstrates, analyzing raw sensor and actuator data directly captures deviations in the system's physical state, catching both equipment malfunctions and cyberattacks. Khaw et al. reinforce this by showing that processing raw time-series current and voltage measurements allows security systems to identify malicious injections based on their actual physical impact, distinguishing genuine grid faults from synthetic ones.

Distinguishing Demand from Structural Failure

Physical constraints do more than stop attackers; they also drastically reduce false positives caused by unusual but legitimate behavior. In a water distribution network, a sudden spike in consumer demand can easily look like a pipe leak to a purely statistical model because both cause a drop in pressure.

By incorporating fluid dynamics—specifically principles like head loss and friction—Lindstrom et al. show how a model-based approach can accurately isolate structural failures. Because a real leak obeys different fluid dynamic equations than a simple increase in consumer draw, the physical model correctly distinguishes between operational fluctuations and structural failures.

The same logic applies to complex maritime systems. In a framework developed by Guo et al., operational data from ships is evaluated against domain-specific physical models. By checking if reported power consumption aligns with known constraints like calm-water resistance, the system can flag subtle sensor drifts or errors that would otherwise appear statistically valid.

Hybrid Architectures: Merging ML with Physical Safety Nets

If physics provides the ultimate truth, why use machine learning at all? Machine learning offers unparalleled speed, flexibility, and the ability to process high-dimensional data without solving complex differential equations in real time. The goal of modern research is to build hybrid architectures where machine learning handles the heavy lifting, while physics serves as a safety net.

A clear example of this is the SPEC framework introduced by Kang et al.. They wrap a standard machine learning estimator in a dual-mode framework:

  1. Fast Prediction: The ML model generates a rapid prediction based on input data.
  2. Physical Verification: The prediction is passed to a physical forward model that computes reconstruction and feasibility errors.
  3. Optimization Loop: If the prediction violates physical boundaries or fails to match expected physical signatures, the system triggers an optimization-based correction loop to pull the prediction back to physical reality.
[Input Data] ---> [ ML Estimator ] ---> [ Physical Forward Model ]
                                                 |
                                       Does it violate physics?
                                        /                  \
                                    (No)                  (Yes)
                                    /                        \
                            [Accept Output]          [Optimization Loop]
                                                              |
                                                     (Correct to Reality)

This integration is especially vital for detecting "unknown-unknowns"—unforeseen operational scenarios that data-driven models have never encountered in their training data. As Maity et al. argue, integrating physics-guided process models provides a robust baseline for measuring deviations, capturing critical safety failures that black-box models miss. Conversely, the absence of these constraints in multi-scale time-series architectures often leads to excessive false positives in industrial settings, a gap that Hu et al. note must be bridged by embedding domain-specific physical boundaries directly into forecasting models.

Extending Physical Expectations to Visual Scenes

The intuition behind physics-constrained anomaly detection extends far beyond industrial control loops and power grids; it is increasingly vital in computer vision. Standard video anomaly detection focuses on semantic context (e.g., detecting a bicycle on a pedestrian walkway). However, as Kart et al. point out, these models fail when confronted with mechanical or physical irregularities, such as a malfunctioning pendulum or unnatural motion trajectories, because the visual patterns of security contexts are fundamentally decoupled from the laws governing mechanical consistency.

To address this, researchers are embedding basic physical expectations directly into visual models. For example, the POD-Net architecture developed by Du et al. pairs a visual generative model with a physics-informed dynamics model. By learning physical expectations, the network can detect anomalies that violate basic laws of nature, such as an object suddenly teleporting or disappearing behind an occlusion.

In a similar vein, Li et al. demonstrate that physics-inspired models can act as powerful feature extractors for crowd analysis. By quantifying internal crowd pressure and interaction forces, the system can monitor deviations from normal physical interaction patterns, detecting dangerous anomalies like stampedes or panic before they become visually obvious to standard surveillance systems.

Key Foundational Work

Unobservable False Data Injection Attacks against PMUs: Feasible Conditions and Multiplicative Attacks (2018) — contributed 2 concept elements to this direction.
Detecting Cyber Attacks in Industrial Control Systems Using Convolutional Neural Networks (2018) — contributed 2 concept elements to this direction.
Crowded Scene Analysis: A Survey (2015) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_physics_informed_anomaly_detection_b3a5cc10,
  title = {physics-informed anomaly detection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physics-informed-anomaly-detection/rdr_b3a5cc10b3524cbdba32d47b379637ef},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_physics_informed_anomaly_detection_b3a5cc10_v1,
  title = {physics-informed anomaly detection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physics-informed-anomaly-detection/rdr_b3a5cc10b3524cbdba32d47b379637ef/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake