Physics-Informed Anomaly Detection vs. The 'Black Box' AI Trap
In critical energy infrastructure, "close enough" is not comforting. It is how teams end up explaining a problem after it has already become expensive.
Traditional AI and machine learning models, often trained on generic statistical variance, can miss the subtle signatures of physical degradation until the site is already drifting in the wrong direction.
At Oxaide, we believe that auditing industrial assets requires more than statistical guessing; it requires Deterministic Physics.
The Problem with Statistical "Black Boxes"
Standard machine learning models (and even Neural Networks) treat telemetry (voltage, temperature, current) as abstract data points. They look for statistical outliers. But in a Battery Energy Storage System (BESS), an outlier isn't always an anomaly, and a critical anomaly isn't always an outlier.
A sub-second voltage curve shift might be statistically insignificant across a 24-hour window, but when cross-referenced against electrochemistry via Incremental Capacity Analysis (ICA - dQ/dV), it becomes a definitive signal of cell degradation or lithium plating.
A Neural Network will tell you: "Anomaly Detected." It cannot tell you: "Loss of Lithium Inventory at SEI growth rate X, leading to Knee Point in 14 months."
The physics does.
The Oxaide Horizon Methodology: Physics-First, AI-Second
We do not sell another black box that guesses when a battery might fail. We start with the physics and build from there.
Oxaide Horizon relies on a 5-Pillar Architecture, with the foundational Pillar 1 built entirely on deterministic equations. We derive chemistry first, then Machine Learning builds on top.
1. Deterministic Chemistry Extraction (ICA)
Most enterprise analytics aggregate logs into 1-minute or 5-minute averages and feed them into an algorithm. Oxaide Horizon uses a signal-separation kernel to reconstruct clean incremental-capacity curves from noisy field data. We calculate the derivative because that is where the chemistry becomes easier to read.
2. Physical Ground Truth Mapping
We map telemetry against the asset's established phase transitions (validated against models from Bloom and Dubarry). If the voltage curve plateaus improperly, our deterministic engine catches the fundamental electrochemical change, firing an alert weeks before temperature sensors detect a spike.
3. Machine Learning as a Scaling Layer (Phase 2)
We do not use Neural Networks to identify mechanisms. ML enters in Phase 2 in specific, mathematically constrained roles: Extended Kalman Filters (EKF) for ±2% State of Charge (SOC) estimation, and adaptive modeling for Knee Point prediction. Our physics engine generates the ground-truth labels that supervise the ML layer. Physics trains the model. The model scales the physics.
Case Study: BESS Forensic Scrubbing
Consider a utility-scale BESS where the existing black-box AI stack keeps flagging "ghost faults" without a usable root cause.
- The situation: Site teams keep getting alerts, but the alert trail is too vague to support action.
- The analysis: Offline Incremental Capacity Analysis shows a shifting dQ/dV shoulder peak in one block.
- The takeaway: That signature is consistent with active lithium plating. Standard SCADA and generic anomaly models can miss it because the signal is subtle until the damage is harder to ignore.
The Verdict: Audit-Grade Rigor
For asset owners and engineering teams, AI is a liability if it cannot be audited. The Oxaide Phase 1 pilot generates a reproducible method. Your engineering team can check each step against the literature and against the telemetry itself. That is what audit-grade transparency looks like.
We don't guess. We derive.
Request a Sample Triage (Free CSV Scan) | Talk to the team
Related reading: