BMS Intelligent Battery Multi-Fleet Hub

Select Active Fleet:

Loading fleet details...

Base: NASA Benchmark

Fleet Size

-- Packs

Operational Units

Fleet Average SOC

-- %

Remaining Stored Charge

Fleet Average SOH

Capacity Retention Factor

Health Status Counts

-- H -- C -- EOL

NASA Ames Aging Metric

Filter Packs:

💡 Click on any battery row to view its complete 32-step telemetry waveforms & cycle history.

Battery ID	Vehicle / Unit	Cycle Count	Current Capacity	Avg Voltage	Avg Current	Peak Temp	Predicted SOC	Predicted SOH	Status	Action
Loading fleet...

Select Battery Across All 5 Fleets

Pick any specific pack from the 200 fleet units to run live cloud inference.

Choose Battery Pack:

State of Charge (SOC) NORMAL

-- %

Remaining Usable Stored Energy

0% (Empty) 50% 100% (Full)

State of Health (SOH) HEALTHY

Capacity Retention Factor

0.0 (End of Life) 0.70 (Retire) 1.0+ (Nominal)

Cloud Inference Execution

EC2 Inference Latency: -- ms

Input Sequence Dimension: 32 steps × 3 features

Architecture Type: Dual-Head Transformer

Parameter Count: 41,986 params (~216 KB)

Server: AWS EC2 t2.micro (13.203.220.49:8000)

32-Timestep Sensor Waveforms

Temporal sensor readings fed into the Transformer

Voltage (V) Current (A) Temp (°C)

Live JSON Output

Payload returned from AWS /predict

Click 'Run Neural Prediction'

POST http://13.203.220.49:8000/predict

Select NASA PCoE Benchmark Cell

Switch between benchmark datasets to analyze capacity fade curves and discharge profiles.

NASA Cell:

Capacity Fade & SOH vs. Cycles

B0005

Capacity degradation across 168 charging cycles (nominal 2.0 Ah down to 1.4 Ah EOL).

Discharge Voltage Profiles

Aging Knees

Terminal voltage curves comparing Cycle 1 vs Cycle 80 vs Cycle 160 under 2.0A discharge.

Surface Thermal Elevation

Joule Heating

Cell thermocouple rise under load. Increased impedance creates +14°C elevation.

Slide 8 Experimental Verification

Empirical Benchmark: Transformer vs. State-Space Model (SSM / Mamba)

Comparison of accuracy, model size, and deployment latency for serverless AWS BMS architectures.

Figure 1: Slide 8 Experimental Results (SOC Estimation, Parameters & Latency)

Dual-Head Transformer

0.68 ms

Inference Latency (Fast)

~25.6k Parameters

State-Space Model (SSM)

5.60 ms

Inference Latency (~9x Slower)

~7.8k Parameters (3x Smaller)

✓

Why Transformer Wins in Deployment Speed: The Transformer uses dense matrix multiplications (BLAS), running fully parallelized on CPU/ONNX Runtime in <1 ms.

✓

Why SSM Has Fewer Parameters: State-Space Models compress state representations into a recurrence matrix, requiring 3x fewer weights (~7.8k vs ~25.6k).

✓

Key Takeaway for Evaluators: Parameter count alone does not dictate cloud execution cost. Without fused parallel-scan CUDA/Triton kernels, recurrent SSM loops run slower sequentially on commodity cloud hardware.

Compiled Benchmark Results Across NASA Datasets

Dataset Cell	Transformer SOC MAE	SSM SOC MAE	Transformer Latency	SSM Latency	Best Cloud Choice
NASA B0005 (Standard)	2.84%	1.42%	0.65 ms	5.58 ms	Transformer (8.5x Faster)
NASA B0006 (Deep Cutoff)	3.12%	1.65%	0.66 ms	5.62 ms	Transformer (8.5x Faster)
NASA B0007 (High Cycle)	2.45%	1.28%	0.65 ms	5.55 ms	Transformer (8.5x Faster)
NASA B0018 (Thermal Stress)	3.48%	1.88%	0.68 ms	5.65 ms	Transformer (8.3x Faster)