Edge AI for machine health
Failure detection engineered to your latency target and deployed on-edge: a dedicated, vendor-independent detection layer for machines and components, running inside the control loop on microcontrollers and FPGAs.
What we do
Vendor-independent, built for a defined set of failure modes, and validated on simulated and real fault data. You bring the domain knowledge of your machine; we bring sensing, detector design and hardware implementation.
Where to place sensors in a machine so a given set of failure modes actually becomes observable.
Observability analysis and experimental design on simulation models of your machine, so instrumentation is decided on evidence rather than habit.
A combination of detectors, complementary in what they catch and how fast, engineered against an explicit target detection latency.
Physics- and simulation-informed design with stated uncertainty bounds and worst-case scenarios before anything reaches your machine.
Detectors run on-edge on microcontrollers and FPGAs, inside the control loop, fast enough to act on a fault while it is still forming.
Delivered as firmware or as an IP core for an FPGA already in your bill of materials. No cloud connection required.
Why Dynetic
For the failure modes where microseconds decide whether a part is scrapped or a spindle is damaged, we run reservoir-computing detectors directly in FPGA fabric. Sensor streams are processed at up to 667 MHz with fixed, predictable latency.
Where it pays off
Examples from precision machining, the first domain we work in. Each one is a fault that forms faster than periodic analysis can react.
Problem
A chipped edge drifts parts out of tolerance; a breakage can wreck the workpiece and the spindle in one revolution.
How we catch it
Spindle-current and vibration detectors, placed for observability of the cutting zone, separate normal wear from chipping and flag a breakage transient in time to retract.
Problem
Self-excited tool-workpiece vibration scars the surface, shortens tool life, hammers the spindle and can escalate from stable to unstable during one pass of the tool.
How we catch it
A fast detector tracks the onset signature of regenerative instability as it grows, giving the control system a usable window to act before the surface is marked.
Problem
Over a long run, process and drive heat move the machine geometry and walk parts out of tolerance, which often remains unseen until inspection.
How we catch it
Detectors fuse temperatures across the machine structure with axis-position feedback to separate genuine drift from noise, so the trend is caught and can be compensated before parts fall out of tolerance.
The same approach applies to
Let's talk
Come with one machine, one failure mode and the time you would need to react. We will sketch a detector for it: which signals, where to sense them, and what hardware it takes.
Who we are
Co-Founder
Doctoral researcher at TU Munich, Chair of Integrated Systems, on adaptive AI for real-time analysis of high-frequency sensor data. He is lead author of ReLiCADA, a published reservoir-computing design algorithm, and works in FPGA design and embedded neural networks.
LinkedIn profile of Jonas Kantic (opens in a new tab)
Co-Founder
Aerospace and robotics engineer; industrial PhD at Airbus Defence and Space on onboard anomaly detection for spacecraft telemetry. He built machine-learning fault detectors for multivariate sensor streams and ran them on space-grade FPGA and real-time hardware.
LinkedIn profile of Andreas Koch (opens in a new tab)