What local analysis is
Local analysis is the compute layer that lives inside the train. Instead of sending millions of raw samples per second to the cloud, IN-SIGHT processes the signal on board on a Raspberry Compute Module 4 (CM4): it turns vibration into interpretable features and decides what is worth transmitting.
This strategy —edge computing— reduces bandwidth by several orders of magnitude and makes it possible to react immediately to a critical event, even in tunnels or areas with no coverage.
Zero cloud latency: primary detection happens on board in milliseconds. Connectivity is only needed to aggregate and report, not to detect.
How IN-SIGHT does it
The on-board pipeline transforms the raw signal into structured telemetry through an optimised DSP chain:
- Preprocessing and FFT: Hann window and Fast Fourier Transform to move from the time domain to the spectral domain.
- Per-subsystem filtering: band-pass IIR filters separate bearings, wheel/rail and low frequency.
- Feature extraction: RMS, kurtosis, crest factor and spectral peaks associated with each failure mode.
- Local EKF + Golden Run: an Extended Kalman Filter compares against the baseline to classify between normal state and anomaly.
In railway practice
For the operator, local analysis means the system works even when the train runs through a section with no network. A flat spot or a bearing overheating is detected on board the instant it appears, without depending on the latency of a remote server.
In normal state, the CM4 aggregates and sends a summary every 30 seconds; on an anomaly, it triggers an immediate event with its classification. This separates "noise" from valuable data before it leaves the vehicle.