Showing results for 
Search instead for 
Did you mean: 

Q&A from the webinar: Create an Edge AI solution for STM32 without any AI knowledge using NanoEdge AI Studio Part1

Bertrand PINEAU
ST Employee 

Part1 - NanoEdge AI Studio: 

What input formats does NanoEdge AI support for data sets for the classification use case? 

> NEAI Studio supports:   

  • File (.txt / .csv), 
  • Serial Port (USB VCP) of datalogger, 
  • SD card of datalogger (.dat) 

Please see: 

What if NanoEdge AI Studio cannot find an adequate algorithm? Is there a process for adding new algorithms? 

> No, for the user it is impossible It is covered by ST R&D 

What is the average RAM/flash size for an algorithm? 

> It is Highly depending on use case, size of input data, type of model selected, and so on. It is in the range of several KB of RAM/Flash usually, please see: 

How does the algorithm determine what is nominal and what is abnormal in the field?  

> User provides signal examples for both nominal and abnormal states, which helps select a library (gives context for the benchmark). Then, the (untrained) library learns only nominal states, and the rest (the patterns "not learned") is considered as abnormal. 

How about if the device is not working right (nominal) from the initial deployment? 

> Considering that it is Anomaly Detection use case the nominal condition must be provided during learning stage of dynamic model library. whatever state is learned on the microcontroller, will be considered as "the normal" even if it does not correspond to a 100% nominal machine state. 

How does anomaly detection account for model drift? Is the model looking for abrupt changes or absolute changes from a static baseline? 

> It depends on the use case, the type of model selected... the result of the inference is a comparison ("mathematical distance") between a given signal, and the knowledge learned by the model. so how this mathematical distance is calculated is different for each library / model.  

Note: if the behavior on a machine drift from nominal operation, it will be reflected in the inference results. Unless the user runs a few more learning iterations to take this drift into account (adjust the knowledge of the library). 

Is NanoEdge AI supporting regression problem solving? 

> Yes, it is. Please see

What is the minimal duration of the input signals for anomaly detection? 

> It depends on particular use case / physical phenomenon to be controlled but in practice e.g., for vibration sensor it is in range of several hundred [ms]. Please see: 

Is it possible to learn different conditions of the system and detect them as nominal? 

> Yes, it is possible e.g., concatenating different nominal conditions within one csv input file. 

Can different models be used at the same time? 

> Yes, it is possible, please see: 

In some applications anomalies are not known or can be simulated without destroying the device. How would you do this with NanoEdge AI? 

> 1-class classification use case can be considered, please see:

Can NanoEdge AI do continuous learning and improve over time? 

> Yes, it is possible, learning is incremental and new iterations can be run at any time to complement and enrich the knowledge / model. Please also see:

Is it needed to do the learning on every device? Or can the library be used on all devices of the same class?  

> It is possible to use the same library (both static and dynamic) on alle devices of the same class. The following scenario is possible: prepare representative HW/SW setup then learn the dynamic model (Anomaly Detection), backup the knowledge and clone the knowledge to every device (working under similar conditions), please see 

Can NanoEdge AI be used without external sensor input? Scenario: I have files of data and need to perform calculation on a model on my STM? 

> It is emulator mode, please see 


Accepted Solutions
Bertrand PINEAU
ST Employee
Bertrand PINEAU
ST Employee