Online Tech Talk - GPU-Free Real-Time Utility Asset Anomaly Detection
The San Francisco Bay Area chapter of the IEEE Computer Society invites to our free and open Virtual Tech Talks (no IEEE membership required):
Speaker: Lakshmana Rao Koppada (Connect on LinkedIn)
Title: GPU-Free Real-Time Utility Asset Anomaly Detection with IBM Granite TSPulse-R1
Abstract: Utility enterprises operate geographically distributed critical assets—such as pumps, tanks, flow meters, and substations that demand continuous health monitoring to avert service disruptions, environmental hazards, and substantial economic losses. Conventional centralized cloud-based Enterprise Asset Management systems incur high latency, excessive bandwidth consumption, and reliance on GPU acceleration, thereby impeding real-time response in connectivity-constrained environments. This paper introduces a GPU-free edge-to-cloud architecture for real-time anomaly detection that exploits the compact IBM Granite TSPulse-R1 time-series foundation model (approximately 1 million parameters). The framework executes lightweight CPU-only inference on edge gateways and transmits only concise anomaly summaries via Message Queuing Telemetry Transport (MQTT), realizing over 99.9% bandwidth reduction relative to raw data transfer. A novel sensor-weighted reconstruction error scoring mechanism prioritizes the most discriminative sensors, enhancing zero-shot multivariate detection performance. Rigorous evaluation on the Water Treatment dataset, a real-world industrial control system benchmark featuring 51 sensors and a 12.53% anomalous timestep ratio, demonstrates an average inference latency of 710.3 ms per 512-timestep chunk, a ROC-AUC of 0.717, and tolerance-adjusted recall exceeding 0.945 at ±20 timesteps. Scalability analysis reveals near-linear latency growth with increasing sensor counts and chunk sizes, confirming feasibility on resource-constrained edge devices. The proposed architecture delivers a scalable, cost-effective, and resilient predictive maintenance solution for critical infrastructure, providing a practical GPU-independent alternative to traditional cloud-centric paradigms while supporting robust operation in distributed utility networks.
Bio: Lakshmana Rao Koppada (IETE Fellow, Senior Member IEEE, Professional Member of BCS) is a Digital Transformation Leader and Technical Architect at PwC with over 16 years of experience delivering large-scale enterprise modernization programs. Specializing in IBM Maximo, cloud platforms, and Red Hat OpenShift, he integrates AI/ML, IoT, and predictive analytics to improve asset reliability and operational efficiency. He has led global digital transformation initiatives for Fortune 500 organizations across utilities, pharmaceuticals, manufacturing, energy, and data center industries. A published researcher and certified IBM Maximo/MAS architect, he actively contributes to international technology conferences and innovation initiatives in AI-driven enterprise systems.