In the world of megaprojects, "I don’t know how it works" is a liability.

As we integrate AI into safety-critical systems, we are often sold the "Black Box" : a model that predicts a failure without an explanation. In infrastructure, blind trust is dangerous. If we are going to use Machine Learning to prevent a multi-million dollar thermal event, we need interpretability. We need to know exactly which variable triggered the alarm before the smoke starts.

At Marsham Edge, we don't believe in a "one-size-fits-all" algorithm. We believe in finding the right model for the right physics.

The 3-Tier Defense: Predictive Engineering for Li-ion Safety

We are currently developing a three-tiered model architecture to take the mystery out of battery health. Here is how we are mapping specific algorithms to specific physical risks:

1. The Baseline Model (Forecasting the "Normal")

Algorithm: Prophet / SARIMA Safety starts with understanding the heartbeat of the asset. We use Prophet to disentangle daily and weekly operational "noise" from the true baseline. By accounting for seasonality and scheduled cycles, we can spot a small deviation in voltage or temperature that traditional SCADA systems would ignore as "within range."

2. The Long-Term Trend (The Health Index)

Algorithm: Random Forest (RF) / XGBoost Predicting the "End of Reliable Life" is a multi-dimensional feature problem. We look at depth of discharge, ambient temperature cycles and internal resistance (and other features).

Using Random Forestallows us to rank "feature importance." We can tell a Project Director exactly which operational habit is accelerating degradation. It turns data into a maintenance strategy.

3. The Spike Model (The High-Frequency Sentinel)

Algorithm: CNN-LSTM Hybrid This is our "needle in a haystack" detector. Thermal runaway is preceded by micro-anomalies that happen in milliseconds. We use a CNN to identify the "shape" of these voltage spikes and an LSTM to understand the temporal context of what happened just before. This hybrid approach provides the head start needed to trigger suppression systems or evacuate a site.

We Need Your Data (And You Need Our Eyes)

To make these models the industry standard for infrastructure safety, we need to stress-test them against real-world megaproject conditions.

The Ask: If you are sitting on historical battery performance data, thermal cycle logs or incident data from failures, I want to hear from you.

I’ll look under the hood of your current predictive tech, check your feature engineering and give you a candid, engineer-to-engineer assessment of where your "Black Boxes" are hiding and where your biggest risks remain.

The Bottom Line

Safety in the age of AI isn't about blind faith in an algorithm. It’s about rigorous, transparent engineering. Let’s stop guessing when the next spike is coming and start calculating it.

Reply to this email if you're ready to talk data.

Stay sharp,

Muriel Demarcus CEO, Marsham Edge