SPARSE SENTINEL

Plain English Technical Detail

What you're looking at

Your data centre sensors already collect temperature, humidity, and power readings every 30 seconds. Right now, you get an alert when a single reading crosses a fixed limit — for example, when exhaust temperature exceeds 38°C.

The problem: by the time a single reading breaches a threshold, the fault has been developing for hours or days. A failing fan doesn't suddenly hit 38°C — it drifts upward slowly, half a degree per hour, while every individual reading stays "within limits."

What Sentinel does: instead of watching individual readings, it learns the normal pattern across all six sensors together. It understands that when intake temperature rises, exhaust should rise proportionally. When one power phase creeps up while others don't, something is wrong — even though no single reading has breached any limit.

This dashboard shows four common failure modes on synthetic (simulated) data:

• Fan degradation — gradual thermal drift as airflow reduces
• PDU overload — one power phase creeping up (asymmetry)
• Hot spot — exhaust temperature rising while intake stays normal
• HVAC drift — humidity becoming unstable as cooling degrades

For each fault, the "Lead time" shows how many hours earlier Sentinel detects the problem compared to a traditional single-channel threshold alert.

Beyond individual fault detection, Sentinel also tracks system health over time. Some problems don't show up as a single unusual reading — they show up as patterns. An HVAC system that oscillates between 20°C and 24°C every ten minutes. A sensor that's been stuck on the same reading for hours. A power supply that trips and resets every 45 minutes. Sentinel catches these by watching the sequence of readings, not just individual values.

The "System Health Trajectory" card shows whether each fault type is healthy, slowly degrading, or actively pathological — and estimates how many hours until the problem becomes critical if the current trend continues.

How it works

Sentinel encodes multivariate sensor windows into sparse representations using a stack of signal processing primitives developed by Sparse Supernova.

1. Windowing & Encoding

Six sensor channels (temp_in, temp_out, humidity, current L1/L2/L3) are sampled every 30 seconds. A sliding window of 10 samples (5 minutes) is flattened to a 60-dimensional feature vector, normalised per-channel against the site's baseline statistics.

2. USL Layer Sizing

The Universal Saturation Law (FRAI = D / (D + 1/dim)) predicts the optimal hidden layer width from the data's intrinsic drift parameter D. For typical data centre sensors, D ≈ 0.19, yielding a hidden layer of ~128 units at 95% FRAI. This replaces trial-and-error architecture search with a single analytical formula.

3. RBM Feature Extraction

A Restricted Boltzmann Machine (60 → 128 hidden units) learns the joint distribution of normal sensor behaviour via contrastive divergence. Training is accelerated by USAD gating — a conformal anomaly detector that skips CD updates on already well-modelled samples, reducing training compute by 40–70%.

4. Anomaly Detection

At inference, each window's reconstruction error through the RBM is computed as a nonconformity score. Scores are smoothed with an EMA (α=0.1) and compared against a conformal threshold (75th percentile of calibration scores). Detection requires 30 consecutive anomalous windows (15 minutes sustained) to avoid false positives from noise.

5. Sequence Pathology (USPD v2)

While USAD scores individual windows, USPD (Universal Sequence Pathology Detector) analyses the sequence of classifications over hours or days. It detects chronic pathologies invisible to per-window scoring:

• Oscillation — system cycling between states (high ordering defect)
• Stuck sensor — no state variation (low coverage health)
• Intermittent faults — periodic state cycling across multiple scales
• Degradation trajectory — pathology score increasing over time (positive trend slope)

USPD v2 uses MAD-normalised slope scoring with a two-pass calibration: pass 1 estimates the slope distribution from calibration sequences, pass 2 rescores using the calibrated MAD as the denominator. The Dslope component is capped at 10.0 to prevent runaway scores on short sequences. Pathology is declared when the z-score exceeds 2.0 against the calibrated null distribution.

6. Classification

A softmax head (128 → 5 classes) classifies the fault type: normal, fan degradation, PDU overload, hot spot, or HVAC drift. Trained on synthetic data; re-trains on site-specific labelled incidents when available.

View the Dashboard