AI Service Build Signal & Sensor AI Industrial ML Systems

From field inspection to
continuous intelligence:
monitoring high voltage circuit breakers.

Critical infrastructure doesn't get second chances. We built a full-stack monitoring system — from ADC signal capture to AI-powered failure prediction — that went from lab prototype to mass production.

Lab to mass production
Full stack ownership: ADC → AI diagnostic
Years of inspection data benchmarked
24/7 continuous health monitoring
Engagement Snapshot
SectorPower Infrastructure / Industrial AI
ServiceFull-Stack System Design & Delivery
PlatformIndustrial PC (IPC) with ADC signal capture
AI ApproachRule-based system on wavelet & FFT signal features
ValidationLab → field deployment → historical benchmarking
OutcomePrototype → mass production product differentiator

Key Highlights

Full stack ownership — hardware signal capture through to AI diagnostic layer and dispatch reporting
Novel signal engineering — wavelet decomposition and FFT for interpretable features from switching transients
Rules derived from decades of field inspection data — expert knowledge encoded, not approximated
Lab tested, field validated against years of historical inspection records
Went to mass production — delivered as a product differentiator, not a research prototype
Context & Challenge

High voltage circuit breakers are among the most critical components in power infrastructure. When they fail, the consequences are severe — outages, safety incidents, costly emergency repairs. And yet, before this system existed, the only way to assess their health was to send engineers into the field for periodic manual inspections. Reactive, infrequent, and entirely dependent on the knowledge of the engineer standing in front of the equipment.

The manufacturer behind this engagement had no monitoring component at all. They wanted to change that — not just to reduce maintenance costs for their customers, but to add something their competitors couldn't match. A circuit breaker that could report on its own health, predict failures before they happened, and tell operators when maintenance was actually needed, rather than on a fixed schedule.

The challenge was that the signals coming off a high voltage circuit breaker are not clean data. They are complex, physically meaningful waveforms — and interpreting them correctly requires understanding both the electronics and the physics of what's happening inside the breaker at the moment of switching. That's the domain boundary where most AI projects fail. We didn't have that boundary.

What the System Monitors

From three-phase voltage
to switching transients.

Three-Phase Voltage & Current

Continuous measurement across all three phases — the baseline of breaker health under normal operating conditions.

Switching Transients

Voltage and current at the moment of switching — the most diagnostically rich signal. Contact erosion, arc behaviour, and mechanical degradation all leave traces here.

Temperature

Thermal monitoring across key components — an early indicator of insulation degradation and abnormal resistance.

Low Voltage Electronic Signals

Control and auxiliary circuit monitoring — capturing the behaviour of the electronic systems that govern breaker operation.

High Voltage Signals (Converted)

High voltage measurements digitised via ADC into safe, processable values — without losing the physical information encoded in the waveform.

Alarms & Events

Timestamped event logging, alarm thresholds, and dispatch centre reporting — the operational layer that turns diagnostic intelligence into actionable alerts.

Approach & Solution

Capture. Interpret. Diagnose. Ship.

Stage 01 — Capture

Building the signal capture layer from scratch.

The system starts at the physical layer — ADC hardware digitising both low voltage electronic signals and high voltage measurements (converted to safe processable values) at the sampling rates needed to capture switching transients accurately. The IPC platform handles real-time ingestion, timestamping, and storage — the foundation everything else runs on.

Multi-channel ADC capture across voltage, current, and temperature signals
Real-time digitisation at switching-appropriate sampling rates
Industrial PC deployment — designed for the environmental demands of substation operation
Line protocol communication to dispatch centre and reporting systems
Stage 02 — Interpret

Making sense of switching transients — the hard problem.

A raw switching transient is not directly interpretable. At the moment a high voltage breaker opens or closes, the voltage and current waveforms contain rich information about contact condition, arc behaviour, and mechanical state — but only if you know how to decompose them.

We used wavelet decomposition to break switching signals into frequency bands, isolating the components that carry diagnostic meaning at each timescale. FFT analysis was applied to other signal types where frequency-domain features were more informative than time-domain. The result: a set of engineered signal features that are physically meaningful — not just numbers, but representations of what is actually happening inside the breaker.

Most monitoring systems apply rules directly to raw signals. That works for simple thresholds. It doesn't work for contact erosion, which expresses itself as subtle changes in transient waveform shape that are invisible without the right decomposition.

Stage 03 — Diagnose

Rules extracted from decades of field knowledge.

The AI diagnostic layer is a hierarchical rule-based system — but the rules are not simple thresholds. They operate on the engineered signal features from Stage 02, and they compose: simple rules combine into deeper rules that capture more complex failure modes and degradation patterns.

The ruleset was derived from years of historical inspection data — records collected manually by field engineers over the operational lifetime of installed breakers. That accumulated expert knowledge, previously locked in documents and human memory, was systematically extracted, formalised, and encoded into the diagnostic engine.

Hierarchical rule composition — simple observations combine into complex diagnostic conclusions
Health status reporting — continuous assessment, not periodic snapshots
Failure prediction — identifying degradation trajectories before they become failures
Maintenance scheduling — telling operators when intervention is actually needed
Stage 04 — Ship

Lab to field to mass production.

The system was built and validated in a controlled lab environment first — verifying signal capture fidelity, ADC calibration, and diagnostic accuracy against known conditions. It was then deployed in the field for real-world validation, where performance was benchmarked against the historical inspection records that had been collected manually over years of field operation. The prototype was delivered to the manufacturer. Mass production followed.

Lab validation — controlled environment testing of hardware and diagnostic accuracy
Field deployment — real-world validation against live equipment
Historical benchmarking — diagnostic output evaluated against years of manual inspection records
Prototype → mass production — delivered as an integrated product component
Results & Impact

What changed.

Before Ferrous Labs After Ferrous Labs
No monitoring component existed — health assessment was entirely manual Continuous, automated health monitoring — no engineer required in the field
Field engineers required for periodic inspection — slow, expensive, and reactive Failure prediction and maintenance scheduling based on actual equipment condition
Failure detected after the fact, not predicted in advance Expert field knowledge encoded into a system that runs without the expert present
Decades of inspection knowledge locked in documents and human memory Validated against years of historical inspection data — not just lab conditions
No product differentiator against competing manufacturers Delivered as a mass production product — a genuine commercial differentiator
Most AI projects fail at the domain boundary — where the signal meets the model. We didn't have that boundary. The same person who understood the physics of a switching transient built the diagnostic system on top of it. That's not a process advantage. It's the only way this gets built correctly.
Ferrous Labs engineering note
Technology

Stack

C++ Python Wavelet Transform FFT ADC Signal Processing Industrial PC (IPC) Line Protocols (IEC 61850) Real-Time Data Acquisition
Got an industrial signal you need to interpret?

Talk to engineering.

Time-series, sensor fusion, edge deployment, signal-to-AI pipelines — we've delivered this end to end. Book a discovery call.