A Multi-directional Meta-Learning Framework for Class-Generalizable Anomaly Detection

ArXiv.org · 2026-01-27

In this paper, we address the problem of class-generalizable anomaly detection, where the objective is to develop a unified model by focusing our learning on the available normal data and a small amount of anomaly data in order to detect the completely unseen anomalies, also referred to as the out-of-distribution (OOD) classes. Adding to this challenge is the fact that the anomaly data is rare and costly to label. To achieve this, we propose a multidirectional meta-learning algorithm -- at the inner level, the model aims to learn the manifold of the normal data (representation); at the outer level, the model is meta-tuned with a few anomaly samples to maximize the softmax confidence margin between the normal and anomaly samples (decision surface calibration), treating normals as in-distribution (ID) and anomalies as out-of-distribution (OOD). By iteratively repeating this process over multiple episodes of predominantly normal and a small number of anomaly samples, we realize a multidirectional meta-learning framework. This two-level optimization, enhanced by multidirectional training, enables stronger generalization to unseen anomaly classes.

PID‐Like Robust Control of Non‐Minimum Phase Robotic Manipulators

International Journal of Robust and Nonlinear Control · 2026-02-11

ABSTRACT This paper proposes an output‐feedback tracking controller for non‐minimum phase nonlinear systems with unknown uncertainties and external disturbances, where not all states are measurable, and the zero dynamics are unstable. The approach combines a backstepping‐based stabilizing state‐feedback law with a cascade extended high‐gain observer (EHGO): A fast observer estimates the measured output and its derivatives to generate a virtual output, while a slower observer estimates the unmeasured controlled output, its derivatives, and the lumped disturbance and uncertainty. For relative‐degree‐two systems, the controller exhibits a PID‐like control. This feature is rarely achieved for non‐minimum phase systems. Simulations of a wafer‐handling robot modeled as a flexible‐joint robot demonstrate accurate tracking, rapid disturbance rejection, smooth inputs, and improved robustness compared to existing methods.

A Multi-directional Meta-Learning Framework for Class-Generalizable Anomaly Detection

Open MIND · 2026-01-27

In this paper, we address the problem of class-generalizable anomaly detection, where the objective is to develop a unified model by focusing our learning on the available normal data and a small amount of anomaly data in order to detect the completely unseen anomalies, also referred to as the out-of-distribution (OOD) classes. Adding to this challenge is the fact that the anomaly data is rare and costly to label. To achieve this, we propose a multidirectional meta-learning algorithm -- at the inner level, the model aims to learn the manifold of the normal data (representation); at the outer level, the model is meta-tuned with a few anomaly samples to maximize the softmax confidence margin between the normal and anomaly samples (decision surface calibration), treating normals as in-distribution (ID) and anomalies as out-of-distribution (OOD). By iteratively repeating this process over multiple episodes of predominantly normal and a small number of anomaly samples, we realize a multidirectional meta-learning framework. This two-level optimization, enhanced by multidirectional training, enables stronger generalization to unseen anomaly classes.

Alleviating Community Fear in Disasters via Multi-Agent Actor-Critic Reinforcement Learning

arXiv (Cornell University) · 2026-04-09

During disasters, cascading failures across power grids, communication networks, and social behavior amplify community fear and undermine cooperation. Existing cyber-physical-social (CPS) models simulate these coupled dynamics but lack mechanisms for active intervention. We extend the CPS resilience model of Valinejad and Mili (2023) with control channels for three agencies, communication, power, and emergency management, and formulate the resulting system as a three-player non-zero-sum differential game solved via online actor-critic reinforcement learning. Simulations based on Hurricane Harvey data show 70% mean fear reduction with improved infrastructure recovery; cross-validation in the case of Hurricane Irma (without refitting) achieves 50% fear reduction, confirming generalizability.

Use of machine learning algorithms to predict optimal hospital length of stay

VTechWorks (Virginia Tech) · 2025-12-09

Problem: Hospitals often struggle to allocate beds, equipment, and staff efficiently, leading to unnecessary complications. Predicting a patient’s length of stay (LOS) early helps hospitals plan treatment, staffing, and bed availability more effectively. Both extremes of LOS carry risks: discharging too early can result in inadequate care and higher readmissions, while prolonged stays waste resources and increase costs. Solution: Optimizing LOS improves patient outcomes using machine learning, enhances operational efficiency, and reduces overall spending.

Koopman-based modeling and control of motion systems with hysteresis dynamics using partial state feedback

Mechatronics · 2025-05-19 · 4 citations

Koopman operator theory represents nonlinear dynamical systems as linear systems in an extended state-space. By selecting observable functions composed of derivatives and functions of derivatives derived from the system output, it is possible to model the system without requiring knowledge of its internal states. The Koopman observable functions are iteratively refined to achieve close alignment with the original system dynamics. The resulting linear model in the extended space is then incorporated into a linear quadratic tracker (LQT) framework, enabling the system output to track a desired reference signal. The proposed method is demonstrated on a mechanical motion system with Bouc–Wen hysteresis, where the Koopman-based model and LQT provide robust control of the nonlinear system and can track a smooth trapezoidal step-scan trajectory for wafer scanner machines. The controller was also compared to a traditional PID controller, showing improved performance over the trajectory. Furthermore, simulation results demonstrate the controller’s robustness against varying initial conditions, parameter uncertainties, and external disturbances.

A Koopman-Based Digital Twin Approach for Fault Detection in Cable Slab Dynamics of Wafer Scanners

IEEE Sensors Journal · 2025-05-20 · 1 citations

This paper presents a higher-dimensional state-space model that employs Koopman operators to predict and monitor the nonlinear dynamics of a cable slab in wafer scanners or other precision motion systems. Developed as a digital twin, our approach not only approximates the complex behavior of the cable slab but also facilitates early fault detection by comparing predicted dynamics with real-time sensor data. To address the absence of a definitive analytical model for cable slab dynamics, we systematically evaluate various Koopman observable functions to minimize tracking errors. The resulting model achieves an error margin of approximately ±1% over the specified motion range and demonstrates robust performance in predicting untrained, acyclic randomized cable slab motion and reaction forces. The system was also compared against the state-of-the-art neural network model for the cable slab, showing a reduction in reaction force error prediction of 75.4%. A sensor noise fault with a signal-to-noise of 20 was detected in 0.35 s using only the reaction force measurement. A position-based noise fault (signal-to-noise of 30) was detected in 0.04 s using position-based fault detection. A hardware malfunction fault test is also performed by removing some tubing material inside the cable slab and is detected in 0.38 s through position-based fault detection with the Koopman model. Experimental results confirm that the Koopman-based framework effectively approximates the nonlinear dynamics and detects faults in the cable slab, thereby enhancing the reliability of precision motion systems.

An output feedback control based on extended high-gain observer for active vibration control systems with nonlinear Galfenol-based actuators

International Journal of Control · 2025-06-25

Output Feedback Decoupling Control of Deformable Mirrors for Adaptive Optics Applications

2025-07-08

The deformable mirror (DM) is a critical component of adaptive optic systems. However, controlling and modeling it is challenging due to its nonlinear, coupled, and position-dependent dynamics, which arise from the interaction of actuators and mechanical coupling across multiple axes. This paper presents a decoupled control strategy for the DM that addresses the challenges of unknown system dynamics and position dependency. The proposed approach incorporates a proportional controller as an inner control loop and an output feedback controller as an outer control loop, combining state feedback with an extended high-gain observer (EHGO). The controller design does not require prior knowledge of the mirror’s dynamics and can be implemented independently for each actuator and axis. Simulations were conducted to evaluate the controller’s performance, demonstrating its effectiveness in tracking desired motions along both the X and Y axes, even in the presence of unknown dynamics.

Beyond Marginals: Learning Joint Spatio-Temporal Patterns for Multivariate Anomaly Detection

ArXiv.org · 2025-09-18

In this paper, we aim to improve multivariate anomaly detection (AD) by modeling the \textit{time-varying non-linear spatio-temporal correlations} found in multivariate time series data . In multivariate time series data, an anomaly may be indicated by the simultaneous deviation of interrelated time series from their expected collective behavior, even when no individual time series exhibits a clearly abnormal pattern on its own. In many existing approaches, time series variables are assumed to be (conditionally) independent, which oversimplifies real-world interactions. Our approach addresses this by modeling joint dependencies in the latent space and decoupling the modeling of \textit{marginal distributions, temporal dynamics, and inter-variable dependencies}. We use a transformer encoder to capture temporal patterns, and to model spatial (inter-variable) dependencies, we fit a multi-variate likelihood and a copula. The temporal and the spatial components are trained jointly in a latent space using a self-supervised contrastive learning objective to learn meaningful feature representations to separate normal and anomaly samples.