Knowledge persistence on ISPU across power cycles — is this a known limit on IIS3DWB10IS too, and is anyone working around it?

[Meltalice]

Hi all,

I came across an older thread here ("Anomaly Detection on ISPU", ISM330ISN) where the conclusion was that NanoEdge AI's learned knowledge cannot be preloaded onto the ISPU — the "Include knowledge from benchmark" option isn't available when targeting the ISPU, because the model has to relearn after every power cycle. One of you confirmed this is expected behavior given the ISPU sits directly on the sensor with no on-board non-volatile memory for it.

I'm doing research on vibration-based bearing fault detection (a manuscript is currently under review at Mechanical Systems and Signal Processing, with validation across four public datasets — CWRU, Huang–Baddour, XJTU-SY, Paderborn), and this relearning constraint is directly relevant to what I'm working on: detectors whose healthy-state reference is small enough to be committed to flash and restored instantly, with no relearning window after a restart.

Two questions for the community / ST team:
1. Does this same constraint apply to the IIS3DWB10IS, or has anything changed there given it's a newer ISPU generation?
2. Has anyone found a practical workaround in the field — e.g., running a lightweight non-learning baseline alongside NanoEdge to bridge the gap right after a power-up, before the model has relearned?

I'd be glad to compare notes if others have hit the same wall, and happy to share what I've found from the detection side if useful.

Thanks!

[Federica Bossi]

Hi ​@Meltalice ,

The ISPU on IIS3DWB10IS, like previous generations, only provides volatile RAM, so the application code must be reloaded after every power cycle. That said, this does not necessarily imply that the model must be retrained each time. A practical approach is to store the trained knowledge in external flash and reload it at startup together with the ISPU firmware. In this way, the device can restore the learned state immediately after power-up without requiring a full relearning window.

[Meltalice]

Thank you Federica, that's very helpful — good to have it confirmed for the IIS3DWB10IS specifically.

One follow-up, if you don't mind: roughly how large is the knowledge block that needs to be reloaded from external flash (just the trained parameters, not the ISPU firmware itself)? And is there published guidance on the energy/time cost of that reload over SPI/I2C at startup?

Asking because I'm comparing this against a non-learning baseline approach in my own work, where the reference state is a handful of scalar values rather than a trained model, and I'd like to make a fair comparison rather than assume numbers.

Thanks again for the clear answer.