Hunting the "Death Signal": From Occult Hypothesis to Non-Equilibrium Engineering

#python #dsp #javascript #opensource

What happens when you try to audit the structural collapse of a system using nothing but pure math?

1. The Occult Origin Story

This project didn't start in a boardroom. It started with a haunting biological hypothesis: "Do organisms or complex systems emit a specific physical signal at the exact moment of their structural collapse—the moment of death?".

I wanted to find the "scream" of a system as it moves irreversibly from an ordered state to chaos. While this sounds like sci-fi, translating it into engineering terms revealed a universal challenge: detecting short-lived non-equilibrium transitions buried in massive environmental noise.

2. The Concept: Structural Collapse

To catch these "ghosts," we need to understand the physics of a regime break.

Energy Landscape: Most systems sit in a stable basin. A collapse occurs when a perturbation pushes the system across a barrier into a new, often destructive, regime.
Transition Signatures: This shift isn't just an anomaly; it has a structure. We look for Precursors (subtle instabilities) and the Winner Event (the primary transition).

3. The Architecture: Dual-Branch Auditing

A single filter isn't enough to isolate a regime break. My pipeline separates analysis into two distinct branches.

Environment Map (LF Branch): Tracks slow behavior, baseline drift, and environmental interference. This branch ensures we don't mislabel a simple drift as a structural event.
Transition Scan (HF Branch): This is the hunter. It isolates rapid changes and audits them using physical metrics like Irreversibility (the "arrow of time") and Persistence.

4. The Muddy Reality: 3 Deadly DSP Traps

Theory is easy; reality is brutal. In my early prototypes, my own tools were assassinating the signal. Here is how I solved it in v15 of the auditor.

Trap 1: The Self-Destruction of CMR

I initially used Common Mode Rejection (CMR) to kill noise.

The Failure: My "Death Signal" was a collapse hitting all channels at once.
The Logic: Subtracting the average across channels meant subtracting the target from itself.
The Fix: Disabled CMR by default to preserve common-mode anomalies.

Trap 2: The IIR Filter Rampage

I fed drifting signals straight into sharp Notch filters for 50Hz hum.

The Failure: Sharp IIR filters "ring" violently when they hit massive drifts.
The Result: The filter creates fake transients that look exactly like structural collapses.
The Fix: Order matters. Apply a Bandpass filter first to flatten the drift, then run the Notch stack.

Trap 3: The DC Offset Illusion

SNR calculations were failing even with clear pulses.

The Failure: 500ms windows often have a local DC offset that dominates power calculations.
The Logic: The energy of this offset hides the AC power of the actual signal.
The Fix: Implemented local mean-centering before RMS calculation to isolate the true signal.

5. The Solution: non-equilibrium-auditor

The final result is a transparent, explainable toolkit. No black boxes. Just audited physics.

Python Engine: death_audit_v15.py processes multi-channel data, ranking events based on irreversibility and artifact ratios.
Interactive UI: A browser-based dashboard where you can visualize the "moment of decay" and run audits on raw CSV data.

Final Thoughts

Is there a "Death Signal"? I’m still auditing. But the journey has taught me that the most important signals are often the ones we accidentally filter out.