MUMBAI, India, Oct. 12 -- Intellectual Property India has published a patent application (202641119105 A) filed by Mr. Venkatnarayanan C; Dr. Sivasundar M; Dr. Shobana D; Dr. Anni Princy B; Ms. Saranya Raj S; and Ms. Shakkthi Sri S on October 05, 2026, for Intelligent Ai-Vlsi Device For Predictive Battery Health Management In Elec-Tric Vehicles Using Deep Learning.
Inventors include Mr. Venkatnarayanan C; Dr. Sivasundar M; Dr. Shobana D; Dr. Anni Princy B; Ms. Saranya Raj S; and Ms. Shakkthi Sri S.
The application for the patent was published on October 09, 2026, under issue no. 41/2026.
Abstract: Intelligent AI-VLSI Device for Predictive Battery Health Management in Elec-tric Vehicles Using Deep Learning Abstract The present invention disclosed an intelligent artificial intelligence very large scale integration device for predictive battery health management in electric vehicles using deep learning. The device comprises a compact semiconductor die that integrates analog measurement front ends, digital conditioning logic, a dedicated neural inference array, nonvolatile parameter memory, and a vehicle-side communication interface in a form factor to fit within a traction battery enclosure or associated control housing. Continuous electrical and thermal observations from individual cells and grouped modules are converted to features capturing instantaneous operating stress and long-term degradation on the device, yielding health estimates at the battery rather than a remote compute resource. Voltage, current, and temperature signals are acquired through protected input channels that reject switching noise of traction inverters and direct-current converters. Optional impedance excitation circuitry imposes a low-amplitude probe and samples the response, providing an alternate signature of electrode and electrolyte condition. The conditioned samples are buffered in on-chip memory and formed into time-aligned windows capturing both fast transients and slow drift. A calibration store retains cell-specific offsets determined at manufacture and updated in service, allowing pack variation to be subsumed into the learned representation. A deep learning engine implemented in the very large scale integration fabric executes a trained neural model, whose layers progressively extract hierarchical patterns from the windowed measurements. Early layers focus on local shape of voltage relaxation, current pulses, and thermal rise; later layers combine these with accumulated charge throughput, temperature exposure, and prior health estimates to infer present capacity retention, internal resistance growth, and remaining useful service interval. Model weights are stored in nonvolatile memory and may be replaced by an authenticated update without altering the surrounding analog circuitry. Inference is gated to complete a full health evaluation within a vehicle control cycle, with a lighter monitoring pass running continuously to detect abrupt divergence from expected behavior. Predicted health quantities are compared against configurable thresholds that encode chemistry, duty cycle, and safety policy. On capacity loss, resistance rise, thermal imbalance, or anomalous trajectory exceeding a limit, the device asserts a graduated indication to the vehicle controller ranging from maintenance advisory to request for power derating or isolation of an affected module. The same indications are formatted for a service tool, allowing a technician to access trend history without pack dismantling. Computation being local, prediction continues in loss of external connectivity and is not compromised by intermittent transmission of raw cell measurements. The integrated arrangement reduces latency versus a split architecture that separates sensing and learning, minimizes exposure of detailed operating records, and scales the predictive function from a single module controller to a pack-level supervisor by replication of the same device. Fabrication in a very large scale integration process enables the neural array, memory, and analog front end to share a common substrate or form a tightly coupled multi-die stack, with interconnect parasitics low enough to permit stable impedance measurement and deterministic inference timing. The invention thus provides an embedded means of anticipating battery degradation in electric vehicles and acting on such anticipation in time to avoid operational and maintenance consequences.
Disclaimer: Curated by HT Syndication.
