AM62A7: AM62A7: EdgeAI to read ADC data

Hi Key,

Yes, this SoC and its software built to serve vision-tasks. Other tasks can work, but will not have a similarly built flow. Unfortunately, tools like gstreamer do not account for time-series data pipelines well (audio being an exception)

We do not currently support any specific standard for piping in ADC data to AI inference, nor is there an ADC on the device. Since there is no ADC in this processor, you will need to rely on an external IC to provide this data over SPI, I2C, or similar interface. Intefaces in Linux like IIO (industrial IO interface)[1] are a good option for acquiring data through sysfs, but will likely require a device-tree change to enable that sensor and corresponding driver . 

Processing piplines for time-series signals (non-audio) are generally more custom-built to my knowledge. The preprocessing stages are generally more complex than vision models, despite the model itself for time-series being simpler.

The neural network you use will need to fit the supported operators list [0]. Edgeai-tidl-tools is built for vision-tasks otherwise, so I expect you will need to develop some code for your task-type to handle preprocessing input/calibration data and postprocessing on the output. 

If there is any particular standard or methodology for acquiring ADC data and passing through preprocessing stages towards AI inference, I would be interested to know what softwares are worth mentioning. 

[0] https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/supported_ops_rts_versions.md

[1] https://docs.kernel.org/iio/iio_adc.html