MUMBAI, India, June 22 -- Intellectual Property India has published a patent application (202631068021 A) filed by Chandra Sekhar Dash; Neelamadhab Padhy; and Dr. Rasmita Panigrahi on May 30, 2026, for Adaptive And Explainable Anomaly Detection Framework For Cyber- Physical System Security Using Ensemble Deep Learning..
Inventors include Chandra Sekhar Dash; Neelamadhab Padhy; and Dr. Rasmita Panigrahi.
The application for the patent was published on June 12, 2026, under issue no. 24/2026.
Abstract: ABSTRACT OF THE INVENTION The present invention discloses an Adaptive and Explainable Anomaly Detection Framework for Cyber-Physical System Security Using Ensemble Deep Learning that integrates multi-modal sensor stream fusion, diversity-optimized deep ensemble classification, statistical concept drift detection, self-supervised continual learning, model-agnostic explainability, and risk-aware alert prioritization within a unified real-time architecture for protecting critical infrastructure including electrical grids, water networks, oil and gas pipelines, and automated manufacturing systems. The framework employs three heterogeneous base detectors—a variational autoencoder (reconstruction probability), a bidirectional LSTM-attention predictor (prediction-error likelihood), and a temporal graph convolutional network (sensor-relationship anomaly)—combined via adaptive voting weights optimized on validation performance with diversity regularization. A statistical drift detection module monitors maximum mean discrepancy on latent feature distributions, classifies drift type (benign operational change versus malicious adversarial drift), and triggers elastic-weight-consolidation-regularized continual learning with rehearsal buffer management, preventing catastrophic forgetting while adapting to non-stationary process dynamics. A multi-modal explainability engine generates local feature attribution via integrated gradients, counterfactual perturbations via projected gradient descent, and anomaly prototype retrieval from episodic memory, delivering operator-actionable explanations with quantified attribution reliability. A risk-aware alert prioritization module computes composite severity scores integrating anomaly calibration, safety instrumented levels, consequence severity, and temporal persistence for triaged operator notification. The framework achieves 99.3% anomaly detection recall at 0.5% false positive rate across SWaT, WADI, and CIC-IDS benchmarks, 96.1% recall for zero-day attacks, concept drift adaptation within 250 samples, explanation latency below 200 ms for 500-sensor systems, 78% operator alarm fatigue reduction, and real-time deployment on industrial SCADA edge hardware—representing a foundational advancement in cyber-physical system security monitorin
Disclaimer: Curated by HT Syndication.
