MUMBAI, India, Aug. 31 -- Intellectual Property India has published a patent application (202641101799 A) filed by Madhankumar C; Mrs. A. Eenaja - Vignan'S Institute Of Management And Technology For Women; Mr. R. Krishna Naik - Vignan'S Institute Of Management; Mr. V. Srikanth Kumar - Vignan'S Institute Of Management And Technology For Women; Mrs. G. Krishna Keerthana - Vignan'S Institute Of Management And Technology For Women; and Mrs. P. Pavani - Vignan'S Institute Of Management And on August 24, 2026, for On-Device Continual-Learning System For Industrial Anomaly Detection Without Catastrophic Forgetting.
Inventors include Mrs. A. Eenaja - Vignan'S Institute Of Management And; Mr. R. Krishna Naik - Vignan'S Institute Of Management; Mr. V. Srikanth Kumar - Vignan'S Institute Of; Mrs. G. Krishna Keerthana - Vignan'S Institute Of Management And Technology For Women; and Mrs. P. Pavani - Vignan'S Institute Of Management And Technology For Women.
The application for the patent was published on August 28, 2026, under issue no. 35/2026.
Abstract: On-Device Continual-Learning System for Industrial Anomaly Detection Without Catastrophic Forgetting ABSTRACT The present invention relates to an On-Device Continual-Learning System for Industrial Anomaly Detection Without Catastrophic Forgetting, designed to enable industrial edge devices to continuously learn new operational patterns and detect previously unseen anomalies while preserving previously acquired knowledge. Conventional industrial anomaly detection systems generally depend on periodically retraining centralized machine- learning models using large datasets, resulting in increased communication overhead, computational requirements, deployment delays, and potential loss of previously learned knowledge when models are updated with new operating conditions. The proposed invention provides an on-device continual-learning architecture comprising an industrial sensor acquisition module, edge computing device, data preprocessing module, feature extraction module, anomaly detection module, novelty detection module, memory management module, continual-learning engine, knowledge preservation module, adaptive model update module, confidence evaluation module, and alert generation interface. The system receives real-time industrial data from sensors and equipment, including vibration, temperature, pressure, current, voltage, acoustic signals, rotational parameters, and other machine-operational measurements. The data preprocessing module performs noise filtering, normalization, segmentation, and feature preparation. The feature extraction module generates temporal, statistical, spectral, and learned representations from the sensor data. The anomaly detection module establishes an adaptive representation of normal industrial operating behavior and identifies deviations from the learned patterns. The continual- learning engine enables the edge device to incrementally update the anomaly detection model when new operating patterns or anomaly characteristics are observed. To prevent catastrophic forgetting, the knowledge preservation module maintains previously learned information using a combination of replay memory, representative sample selection, knowledge distillation, parameter regularization, protected parameter storage, or feature-space consistency constraints. The memory management module selectively stores informative historical samples or compact representations rather than retaining the complete historical dataset. The novelty detection module determines whether newly observed patterns represent known operating conditions, previously learned anomalies, or previously unseen conditions. Based on novelty and confidence scores, the system determines whether an observed pattern should be used for incremental model adaptation. Low-confidence or uncertain samples may be temporarily retained for further evaluation rather than immediately modifying the model. The adaptive model update module performs incremental learning directly on the industrial edge device. Model updates may be controlled according to computational capacity, available memory, energy condition, update frequency, and detected concept drift. The system may selectively update model parameters while protecting previously important parameters from excessive modification. The invention further incorporates a concept-drift detection mechanism to identify changes in industrial operating conditions caused by equipment aging, maintenance, process modifications, environmental variations, production changes, or sensor characteristics. When significant drift is detected, the continual-learning engine updates the model using newly acquired information while retaining historical knowledge through the knowledge preservation mechanism. The system generates real-time anomaly scores and may classify detected conditions into normal operation, known anomaly, emerging anomaly, or unknown anomaly categories. Alerts may be transmitted locally to an industrial controller or operator interface, while selected information may optionally be synchronized with a centralized monitoring system. The proposed invention enables real-time, privacy-preserving, communication-efficient, and adaptive industrial anomaly detection directly at the edge, reducing dependence on continuous cloud connectivity and centralized retraining. By combining on-device continual learning with memory-efficient knowledge preservation and adaptive anomaly detection, the system maintains previously learned knowledge while adapting to evolving industrial environments. The invention is applicable to predictive maintenance, manufacturing equipment monitoring, industrial robotics, motors, pumps, turbines, compressors, production lines, power systems, process industries, and other Industrial Internet of Things (IIoT) environments requiring continuous and reliable anomaly detection.
Disclaimer: Curated by HT Syndication.
