During the years, neuroprosthetic applications have obtained a great deal of attention byudthe international research, especially in the bioengineering field, thanks to the huge investmentsudon several proposed projects funded by the political institutions which consider the treatment of this particular disease of fundamental importance for the global community.udThe aim of these projects is to find a possible solution to restore the functionalities lost by a patient subjected to an upper limb amputation trying to develop, according to physiological considerations, a communication link between the brain in which the significant signals areudgenerated and a motor prosthesis device able to perform the desired action. Moreover, the designed system must be able to give back to the brain a sensory feedback about the surrounding world in terms of pressure or temperature acquired by tactile biosensors placed at the surface of the cybernetic hand. It in fact allows to execute involuntarymovements when for example the armcomes in contact with hot objects.udThe development of such a closed-loop architecture involves the need to address some critical issues which depend on the chosen approach. Several solutions have been proposedudby the researches of the field, each one differing with respect to where the neural signals are acquired, either at the central nervous systemor at the peripheral one,most of themfollowing the former even that the latter is always considered by the amputees amore natural wayudto handle the artificial limb. This research work is based on the use of intrafascicular electrodes directly implanted in the residual peripheral nerves of the stump which represents a good compromise choice in terms of invasiveness and selectivity extracting electroneurographicud(ENG) signals from which it is possible to identify the significant activity of a quite limited number of neuronal cells. In the perspective of the hardware implementation ofudthe resulting solution which can work autonomously without any intervention by the amputee in an adaptive way according to the current characteristics of the processed signal and by using batteries as power source allowing portability, it is necessary to fulfill the tightudconstraints imposed by the application under consideration involved in each of the various phases which compose the considered closed-loop system.udRegarding to the recording phase, the implementation must be able to remove the unwanted interferences mainly due to the electro-stimulations of themuscles placed near theudelectrodes featured by an order of magnitude much greater in comparison to that of the signals of interest amplifying the frequency components belonging to the significant bandwidth, and to convert them with a high resolution in order to obtain good performance at the next processing phases. To this aim, a recording module for peripheral neural signals will be presented, based on the use of a sigma-delta architecture which is composed byudtwo main parts: an analog front-end stage for neural signal acquisition, pre-filtering and sigma-delta modulation and a digital unit for sigma-delta decimation and system configuration.udHardware/software cosimulations exploiting the Xilinx System Generator tool in Matlab Simulink environment and then transistor-level simulations confirmed that the systemudis capable of recording neural signals in the order of magnitude of tens of μV rejecting the huge low-frequency noise due to electromyographic interferences.udThe same architecture has been then exploited to implement a prototype of an 8-channel implantable electronic bi-directional interface between the peripheral nervous system and the neuro-controlled hand prosthesis. The solution includes a custom designed Integrated Circuit (0.35μm CMOS technology), responsible of the signal pre-filtering and sigma-delta modulation for each channel and the neural stimuli generation (in the opposite path) basedudon the directives sent by a digital control systemmapped on a low-cost Xilinx FPGA Spartan-3E 1600 development board which also involves the multi-channel sigma-delta decimationudwith a high-order band-pass filter as first stage in order to totally remove the unwanted interferences.udIn this way, the analog chip can be implanted near the electrodes thanks to its limited size avoiding to add a huge noise to theweak neural signals due to longwires connections and to cause heat-related infections, shifting the complexity to the digital part which can be hosted on a separated device in the stump of the amputeewithout using complex laboratory instrumentations. The system has been successfully tested from the electrical pointudof view and with in-vivo experiments exposing good results in terms of output resolution and noise rejection even in case of critical conditions.udThe various output channels at the Nyquist sampling frequency coming from the acquisition system must be processed in order to decode the intentions of movements of the amputee, applying the correspondent electro-mechanical stimulation in input to the cybernetic hand in order to perform the desired motor action. Different decoding approaches have been presented in the past, the majority of them were conceived starting from the relativeudimplementation and performance evaluation of their off-line version. At the end of the research, it is necessary to develop these solutions on embedded systems performing an online processing of the peripheral neural signals. However, it is often possible only by usingudcomplex hardware platforms clocked at very high operating frequencies which are not be compliant with the low-power requirements needed to allow portability for the prostheticuddevice.udAt present, in fact, the important aspect of the real-time implementation of sophisticated signal processing algorithms on embedded systems has been often overlooked, notwithstanding the impact that limited resources of the former may have on the efficiency/effectivenessudof any given algorithm. In this research work it has been addressed the optimization of a state-of-the-art algorithmfor PNS signals decoding that is a step forward for its real-time, full implementation onto a floating-point Digital Signal Processor (DSP). Beyond low-leveludoptimizations, different solutions have been proposed at an high level in order to find the best trade-off in terms of effectiveness/efficiency. A latency model, obtained through cycle accurate profiling of the different code sections, has been drawn in order to perform a fairudperformance assessment. The proposed optimized real-time algorithmachieves up to 96% of correct classification on real PNS signals acquired through tf-LIFE electrodes on animals, and performs as the best off-line algorithmfor spike clustering on a synthetic cortical datasetudcharacterized by a reasonable dissimilarity between the spikemorphologies of different neurons.udWhen the real-time requirements are joined to the fulfilment of area and power minimizationudfor implantable/portable applications, such as for the target neuroprosthetic devices, only custom VLSI implementations can be adopted. In this case, every part of the algorithmshould be carefully tuned. To this aim, the first preprocessing stage of the decoding algorithmbased on the use of aWavelet Denoising solution able to remove also the in-band noise sources has been deeply analysed in order to obtain an optimal hardware implementation.udIn particular, the usually overlooked part related to threshold estimation has been evaluated in terms of required hardware resources and functionality, exploiting the commercial Xilinx System Generator tool for the design of the architecture and the co-simulation. Theudanalysis has revealed how the widely used Median Absolute Deviation (MAD) could lead o hardware implementations highly inefficient compared to other dispersion estimatorsuddemonstrating better scalability, relatively to the specific application.udFinally, two different hardware implementations of the reference decoding algorithm have been presented highlighting pros and cons of each one of them. Firstly, a novel approach based on high-level dataflow description and automatic hardware generation is presentedudand evaluated on the on-line template-matching spike sorting algorithmwhich represents the most complex processing stage. It starts from the identification of the single kernels with the greater computational complexity and using their dataflow description to generate the HDL implementation of a coarse-grained reconfigurable global kernel characterized by theminimumresources in order to reduce the area and the energy dissipation forudthe fulfilment of the low-power requirements imposed by the application. Results in the best case have revealed a 71%of area saving compared tomore traditional solutions,without any accuracy penalty. With respect to single kernels execution, better latency performance are achievable stillminimizing the number of adopted resources.udThe performance in terms of latency can also be improved by tuning the implemented parallelismin the light of a defined number of channels and real-time constraints, by usingudmore than one reconfigurable global kernel in order that they can be exploited to perform the same or different kernels at the same time in a parallel way, due to the fact that each one can execute the relative processing only in a sequential way. For this reason, a second FPGA-based prototype has been proposed based on the use of aMulti-Processor System-on-Chip (MPSoC) embedded architecture. This prototype is capable of respecting the real-timeudconstraints posed by the application when clocked at less than 50 MHz, in comparison to 300 MHz of the previous DSP implementation. Considering that the application workloadudis extremely data dependent and unpredictable due to the sparsity of the neural signals, the architecture has to be dimensioned taking into account critical worst-case operating conditions in order to always ensure the correct functionality. To compensate the resulting overprovisioningudof the system architecture, a software-controllable power management based on the use of clock gating techniques has been integrated in order tominimize the dynamicudpower consumption of the resulting solution.udSummarizing, this research work can be considered a sort of proof-of-concept for the proposed techniques considering all the design issues which characterize each stage of theudclosed-loop system in the perspective of a portable low-power real-time hardware implementation of the neuro-controlled prosthetic device.ud
展开▼
