Executing Condition Monitoring Algorithms on ARM Cortex-M4 Using Tensorflow Lite for Microcontrollers

Condition monitoring systems (CMS) require execution of inverse models of physical systems close to the product to estimate the system’s status from sensor readings. This requires mapping and executing elaborate numerical models—potentially including machine learning models—onto computing devices with limited power budget, limited memory and limited support for floating-point arithmetic. Most existing ML frameworks provide only limited support for this task. We document in detail a workflow using Tensorflow Lite for Microcontrollers (TFLM) to map an LSTM, CNN and transformer model onto an Arduino Nano, featuring an ARM Cortex-M4 CPU. We discuss identified shortcomings and their implications on CMS design.