ERR:-1002 and the Arduino lib usage

A nano33 ble sense rev2 board is used for detecting the vibration style. I used accelerometer x,y,z, sample rate=100, window= stride=1000ms. A neural network is trained for classifying normal vibration and abnormal vibration. After I sampled and trained, then the ei binary FW is downloaded and burned into nan0 33 rev2. After running impulse in PowerShell for 2 minutes, error happens:
Starting inferencing in 2 seconds…
Sampling…
Timing: DSP 9 ms, inference 626 us, anomaly 0 ms, postprocessing 49 us
#Classification predictions:
abnormal: 0.003906
none: 0.996094
Starting inferencing in 2 seconds…
Sampling…
Can’t allocate new thread
Timing: DSP 5 ms, inference 602 us, anomaly 0 ms, postprocessing 49 us
#Classification predictions:
abnormal: 0.000000
none: 0.996094
Starting inferencing in 2 seconds…
Sampling…
Can’t allocate new thread
ERR: Failed to run DSP process (-1002)
Failed to run impulse (-5)

I am not sure, the bin FW is generated by EdgeImulse online studio, why errors?
Another question: I tried the Arduino library and the …\examples\nano_ble33_sense_rev2\nano_ble33_sense_rev2_accelerometer\nano_ble33_sense_rev2_accelerometer.ino, but the abnormal status could not be checked out,only normal checked will possibilities around >=0.98;meanwhile the same abnormal status is sampled and detected by “live classification” in Edge Impulse, abnormal is easily checked!!!
I wonder one thing: Do I need to modify MANUALLY the low level accelerometer driver(the BMI270 .c /.h) code to adapt to nano_ble33_sense_rev2_accelerometer.ino project? OR, is everything ready in the example nano_ble33_sense_rev2_accelerometer.ino?

Any response? Or do I miss something to get reply?

I did some tests to find the truth: the online studio’s sampling is different with that when use Arduino Library examples!!! The BMI270’s driver is not the same with that used in online studio.
With the same vibration operation(nano3 binded to a swinging fan), I sampled online and in the Arduino library example, and compare the waves of axes X,Y, Z data。 They are obviously different.
EdgeImpulse fellows, why don’t you tell me that point? And what configuration did you use when sampling online, the BMI270 sample rate, filter, continuous mode, opt mode? I tried many combinations but can not get the right wave style as the online sampling data.