PORRITTINC
( LAYTON, UT )
POSITION PAPER · NEUROMORPHIC R&D

Spiking anomaly
detection.

Conventional condition monitoring samples a tag on a scan cycle and compares it to a threshold. An event-driven sensor reports the change itself. This paper sets out why we believe that inversion matters for plant protection.

← Back to Spiking Anomaly Detection

POSITION PAPER · TECHNICAL OPINION

This is a technical position paper stating Porritt Inc.'s engineering view. It is not a peer-reviewed publication and does not report experimental results. Where it forecasts capability, it is stating design intent rather than achieved performance.

Position
PREMISE

Sampling discards the transient

A scan cycle is a low-pass filter. Fast mechanical events — cavitation onset, bearing spall, arc initiation — are attenuated or missed entirely between samples.

APPROACH

Encode change, not state

Event-driven sensors emit on threshold crossing at the transducer. Silence costs nothing; activity gets immediate attention. Duty cycle follows the process rather than the clock.

CLASSIFICATION

Temporal signature

A spiking network classifies the shape of an event over time rather than a scalar magnitude, which is what distinguishes a pump cavitating from a pump merely loaded.

HONEST LIMIT

Not yet demonstrated at plant scale

We are stating a research direction. We have not deployed this at scale in an operating facility, and this paper should not be read as reporting field results.

Objectives
μs
Latency objective
mW
Power objective
EDGE
On-device
R&D
Not deployed
Open questions

Training data for rare failure modes is scarce by definition — the events worth detecting are the ones that seldom happen. Transfer between machines and duty cycles is unproven. Certification of an event-driven protective function under existing standards is unresolved and would need to be worked with the certifying body.

Discuss the research.

We welcome technical challenge on these positions. Collaboration enquiries welcome.

CONTACT TIMOTHY → Back to Spiking Anomaly Detection →