The new generation of wireless technologies, fitness trackers, and deviceswith embedded sensors can have a big impact on healthcare systems and qualityof life. Among the most crucial aspects to consider in these devices are theaccuracy of the data produced and power consumption. Many of the events thatcan be monitored, while apparently simple, may not be easily detectable andrecognizable by devices equipped with embedded sensors, especially on deviceswith low computing capabilities. It is well known that deep learning reducesthe study of features that contribute to the recognition of the differenttarget classes. In this work, we present a portable and battery-poweredmicrocontroller-based device applicable to a wobble board. Wobble boards arelow-cost equipment that can be used for sensorimotor training to avoid ankleinjuries or as part of the rehabilitation process after an injury. The exerciserecognition process was implemented through the use of cognitive techniquesbased on deep learning. To reduce power consumption, we add an adaptivity layerthat dynamically manages the device's hardware and software configuration toadapt it to the required operating mode at runtime. Our experimental resultsshow that adjusting the node configuration to the workload at runtime can saveup to 60% of the power consumed. On a custom dataset, our optimized andquantized neural network achieves an accuracy value greater than 97% fordetecting some specific physical exercises on a wobble board.