With the development of mobile applications (telephony, IoT, home automation, etc.), embedded systems have grown exponentially in recent years. The main characteristic of these new systems is the combination of high computing power and long operating autonomy. Unfortunately, as these characteristics are diametrically opposed, system designers are faced with the dilemma of limiting on-board power. To overcome this autonomy problem, more and more architectures are turning to the implementation of energy recovery stages from one or more sources. As part of a CIFRE thesis, the study focused on the design of an optimized energy recovery module interfacing with several sources. Two directions were identified to increase the amount of energy recovered by such a device:- the development of a multisource approach in order to propose an industrializable architecture;- the optimization of the amount of energy extracted from each source using new-generation algorithms to find the maximum operating point. These are optimized thanks to recent advances in Deep Learning and the commercialization of increasingly powerful low-power microcontrollers. The ultimate goal is to deploy large-scale, low-cost sensor networks, with improved or even infinite autonomy if the system manages to produce more energy than it consumes.