Compressive Sensing system for recording of ECoG signals in-vivo

7. Conclusions7.1. RemarksAs it is shown along the previous chapters, this project has been arranged as collaborationbetween LSM-EPFL and IMEC. During the three months in LSM, it has been carried out a deepstudy about Compressive Sensing oriented to the implementation of a system on-chip tocompress neural signals and send it by saving power consumption and heat dissipation in thetransmitter.Due to the fact that Compressive Sensing technique applied to real implementations is in anearly phase, an extensive state of the art has been accomplished in order to be able to specifythe known limitations of the designs based on Compressive Sensing. This study encompassesall the actual trends of design, both in digital and analog domain, and so it is considered a goodstarting point for future related projects. It has been scheduled the circuit implementation of ananalog-liked multi-path/multi-channel compression system, in order to deepen into the feasibilityof this configuration instead of the digital implementation, which is the most used one evenwhen the available power consumption comparison between digital and analog approacheshave revealed not to be enough conclusive to dismiss the latter one.For this purpose, foremost, a complete multichannel implementation has been developed bothin Matlab and Simulink. In order to carry through this task, a neural signal, (AP) has beenconsidered as sparse input. The reconstruction has been accomplished by applying BPDN andLASSO reconstruction methods, for which the mathematical bases of Compressive Sensinghave been studied.In order to implement a parallel and more compact random generator block to implement themeasurements matrix, a new design has been settled up in Cadence 5, by using the technologyUMC 0.18μm, for a variable number of multiple outputs.Under the scope of the regular operation of Compressive Sensing, a new approach to thismethod has been developed by realizing that sparsity conditions are fulfilled in an array ofmultielectrodes. That it, if all the sensors are simultaneously sampled, according to the usualspikes propagation among neurons, just few of these electrodes receive a peak, what results ina spatial sparse signal which can be compressed by Compressive Sensing. This approach hasbeen implemented in Matlab and Cadence, and results have been submitted in the paper [31].During the period in IMEC, in order to implement the mixer and the integrator blocks which areneed in each of the paths (rows) of a channel, a combined design has been implemented inCadence 6, by using the technology TSMC018.In order to achieve a feasible analog implementation, different integrators have been tested.Integrators in continues time domain have been proved not to be an optimal solution due to the71