Training a sensor using generated data can be tricky because your generation program needs to mimic as much as possible the sensor behavior (in terms of output data rate, full scale, sensitivity, noise, etc...). If, ideally, you have a good sensor model that can emulate the physical sensor, using simulated training data for the MLC is possible.
