VoltPulse AI

Edge-AI battery management platform uniting SAE J1939 CAN telemetry, Gaussian Process uncertainty forecasting, sub-ms thermal runaway detection, and an industrial SCADA digital twin.

Comment

💡 Inspiration

Battery Energy Storage Systems (BESS) and Electric Vehicle (EV) battery packs are the fundamental lifeblood of the global clean energy transition. However, catastrophic thermal runaways, sudden capacity collapse, and sensor telemetry packet loss remain severe operational threats:

  1. The "Silent Micro-Short" Hazard: Catastrophic fires in multi-megawatt grid battery installations and EV packs originate from dendritic separator micro-shorts 100+ cycles before visible smoke. Conventional BMS thresholding detects heat only after irreversible exothermic decomposition has already begun.
  2. Context & Telemetry Amnesia: In real-world IoT networks, cellular dropouts and edge buffer delays cause battery telemetry to arrive out of order (e.g. cycle 140 arrives after cycle 180 has been processed). Naive cloud forecasting systems silently overwrite state or fail to reconcile retroactive degradation.
  3. Black-Box Point Predictions: Traditional neural network estimators produce single deterministic State-of-Health (SOH) numbers without statistical confidence intervals, risking dangerous over-utilization when sensor telemetry is sparse.

We built VoltPulse AI to bridge deep embedded hardware engineering (automotive CAN-bus J1939 & Modbus TCP protocols, 16-cell series pack monitoring, active balancing) with physics-informed Gaussian Process Regression (GPR), sub-millisecond thermal runaway prediction, and an industrial SCADA digital twin.

🚀 What It Does

VoltPulse AI unifies four mission-critical subsystems into a unified edge-to-cloud resilience platform:

  1. Embedded Hardware BMS & Protocol Streamer (SAE J1939 CAN & Modbus TCP):

    • Emulates an automotive/grid-tier BMS MCU streaming 10Hz cell-level telemetry (16-cell series string with thermistors).
    • Encodes standard 29-bit CAN arbitration IDs (0x18F00100 Pack Summary, 0x18F00200-0x18F00500 Cell Voltages at 1mV resolution, 0x18F00600 Thermistors, 0x18F00700 Safety Interlocks).
    • Provides a 16-bit Modbus TCP register map (40001–40025) for industrial PLC and SCADA integration.

  2. Physics-Informed Gaussian Process Degradation Forecaster with 95% Bayesian Uncertainty:

    • Evaluates multi-kernel families (Matérn 5/2, Matérn 3/2, RBF, Rational Quadratic, ARD Composite) on a temporal cross-validation split.
    • Computes exact predictive mean $\hat{y}*$ and predictive variance $\mathbb{V}[y*]$ using Cholesky factorization $\mathbf{L}\mathbf{L}^T = \mathbf{K} + \sigma_n^2 \mathbf{I}$ with an adaptive diagonal jitter ladder ($10^{-10} \to 10^{-4}$).
    • Generates 95% Bayesian confidence ribbons ($\pm 1.96\sigma$) and Remaining Useful Life (RUL) cycle estimates to 80% End-of-Life (EOL).

  3. Sub-Millisecond Early-Warning Thermal Runaway & Micro-Short Detector:

    • Analyzes temperature gradients ($dT/dt$), voltage collapse rates ($dV/dt$), and pack thermal divergence in sub-millisecond execution.
    • Flags separator micro-shorts and triggers automated high-voltage contactor relay cutoff within 54.6 microseconds.

  4. Deterministic Late-Telemetry Reconciler:

    • Reconstructs historical telemetry timelines in chronological event order when delayed IoT packets arrive.
    • Re-evaluates GPR kernels and computes parameter diffs ($\Delta\text{SOH}$, uncertainty shift, RUL adjustment) with complete audit logging.

  5. Industrial Cyber-SCADA Digital Twin Operations Center:

    • Real-time 16-cell thermal matrix with active balancing indicators.
    • Interactive Electrochemical Impedance Spectroscopy (EIS) Nyquist plot ($10\text{ kHz} \to 10\text{ mHz}$) separating Ohmic $R_s$, charge transfer $R_{ct}$, and Warburg diffusion.
    • Live CAN-bus hexadecimal stream terminal, 1-Click Judge Tour, and Web Speech API audible emergency broadcaster.

🛠️ How We Built It

  • Hardware & Protocol Layer: Python 3.11, byte-level bit-shifting, SAE J1939 CAN frame encoder, Modbus TCP 16-bit register mapper.
  • Physics & ML Engine: NumPy, SciPy (Spatial Distance cdist, Cholesky Factorization cho_solve, Linear Least Squares), Randles Circuit & 1st-Order Thevenin ECM.
  • Backend Architecture: Asynchronous FastAPI framework, Pydantic v2 schemas, REST endpoints.
  • Frontend SCADA UI: Pure Vanilla CSS3 & Modern ES6 JavaScript (Zero external UI bloat), Canvas 2D renderers for real-time GPR uncertainty envelopes and Nyquist impedance arcs.
  • Testing & Verification: 22 automated Pytest unit and integration tests (100% pass rate in 0.6s) and sub-millisecond benchmark suite.

⚡ Technical Benchmarks

  • CAN-bus J1939 Packing: 6,701 frames / sec
  • Thermal Runaway Detection Latency: 142 µs / check (Sub-millisecond real-time edge speed)
  • Bayesian GPR Inference: 282 full forecast curves / sec
  • EIS Nyquist Synthesizer: 1,919 spectra / sec
  • Late-Data Reconciliation: 42 ms / full timeline re-evaluation

🧠 Challenges We Ran Into

  • Covariance Matrix Numerical Stability: Handling near-collinear historical battery cycles during Cholesky factorization. We solved this with an adaptive numerical jitter ladder ($10^{-10} \to 10^{-4}$) to guarantee positive-definiteness without distorting predictive uncertainty.
  • Out-of-Order IoT Telemetry Ingestion: Ensuring that delayed measurements do not corrupt the active timeline by implementing deterministic event-time indexing and parameter diff rollback.

🏆 Accomplishments That We're Proud Of

  • Sub-Millisecond Life-Safety Interlock: Achieving 54.6 µs micro-short detection to prevent thermal propagation before fire onset.
  • 1-Click Judge Tour: Building an automated 30-second live demonstration with synthetic speech announcements that walks judges through every subsystem.
  • 100% Passing Test Suite: 22/22 unit and integration tests verifying battery electrochemistry, CAN protocols, and API routes.

🎓 What We Learned

We deepened our expertise in fusing automotive CAN protocols (J1939) with non-parametric Bayesian regression (GPR) and electrochemical impedance modeling, demonstrating how edge AI can directly govern physical life-safety contactors.

🔮 What's Next for VoltPulse AI

  1. Hardware In-The-Loop (HIL) Testbench: Interfacing with physical STM32F4 / ESP32 CAN transceivers (MCP2515) and TI BQ76952 BMS evaluation boards.
  2. Embedded C/Rust Firmware Kernel: Porting the sub-millisecond micro-short detection routine to Bare-Metal Rust on ARM Cortex-M4.
  3. Grid Fleet Aggregation: Coordinating multi-gigawatt utility BESS installations for national grid frequency regulation.

Built With

Share this project:

Updates

Submission history