Conception of a System-on-Chip (SoC) Platform to Enable EMG-Guided Robotic Neurorehabilitation

Abstract

Electromyography (EMG) signals are fundamental in neurorehabilitation as they provide a non-invasive means of capturing the electrical activity of muscles, enabling precise detection of motor intentions. This capability is essential for controlling assistive devices, such as therapeutic exoskeletons, that aim to restore mobility and improve motor function in patients with neuromuscular impairments. The integration of EMG into neurorehabilitation systems allows for adaptive and patient-specific interventions, addressing the variability in motor recovery needs. However, achieving the high fidelity, low latency, and robustness required for real-time control of these devices remains a significant challenge. This paper introduces a novel multi-channel electromyography (EMG) acquisition system implemented on a System-on-Chip (SoC) architecture for robotic neurorehabilitation. The system employs the Zynq-7000 SoC, which integrates an Advanced RISC Machine (ARM) processor, for high-level control and an FPGA for real-time signal processing. The architecture enables precise synchronization of up to eight EMG channels, leveraging high-speed analog-to-digital conversion and advanced filtering techniques implemented directly at the measurement site. By performing filtering and initial signal processing locally, prior to transmission to other subsystems, the system minimizes noise both through optimized processing and by reducing the distance to the muscle, thereby significantly enhancing the signal-to-noise ratio (SNR). A dedicated communication interface ensures low-latency data transfer to external controllers, crucial for adaptive control loops in exoskeletal applications. Experimental results validate the system’s capability to deliver high signal fidelity and low processing delays, outperforming commercial alternatives in terms of flexibility and scalability. This implementation provides a robust foundation for real-time bio-signal processing, advancing the integration of EMG-based control in neurorehabilitation devices.

1. Introduction

Mobility is an integral part of human life that we often take for granted until it is compromised by an injury or disability. It is estimated that approximately 15% of the global population suffers from some form of disability. Considering Spain’s current population of approximately 48.95 million, it is notable that in 2020, 4.38 million individuals reported having some form of disability. Among these, mobility-related disabilities—such as a difficulty or inability to move or handle objects—were the most prevalent, affecting $38.9$% of men and $68.5$% of women. These mobility difficulties, due to various illnesses, significantly affect the quality of life and the ability to perform daily activities for those affected. Among these individuals, a not very numerous group (500,000 people per year worldwide in 2013 and 915 individuals in total in 2021 in Spain) but particularly vulnerable are those with spinal cord injuries.

These injuries, typically caused by traffic accidents, falls, or sports injuries, can result in a total or partial loss of motor function, often leading to a significant loss of independence and autonomy, and it is common for them to result in depression and a higher risk of premature death. Among existing spinal cord injuries, the only ones that have a potential to regaining mobility to some degree are those where there has not been total damage to the spinal cord; in other words, the so-called “incomplete spinal cord injuries” [1]. Incomplete spinal cord injury encompasses a wide range of symptoms and levels of disability, making treatment and rehabilitation a complex challenge. Traditionally, treatments to improve mobility in patients with incomplete spinal cord injuries have focused on traditional physiotherapy, which includes specific physical exercises and rehabilitation techniques to strengthen weakened muscles and improve motor function. Although these treatments have proven effective in many cases, their success largely depends on the severity of the injury and the individual patient’s responsiveness [1,2].

In recent decades, there have been significant advances in the field of motor rehabilitation, particularly with therapies based on exoskeletons [1,2,3,4,5,6]. These mobility robotic aids have revolutionized the way we approach rehabilitation for patients with incomplete spinal cord injuries and other diseases that affect movement. Exoskeletons are designed to provide patients with support and assistance for controlled movement, enabling them to perform movements that would otherwise be impossible [1,2,3,4,5,6].

When comparing conventional methods with exoskeleton-based approaches, it is difficult to establish which method is more effective for patient rehabilitation [2]. The use of exoskeletons in spinal cord injury rehabilitation holds promise in improving motor function through activation of neuroplasticity processes. While they do not directly induce axon or tissue regeneration, their ability to facilitate movement repetition and improve physical independence contributes to neural reorganization and enhances residual mobility. Integrating them with other therapies, such as electrical stimulation or stem cell treatments, could further enhance these effects, helping to restore motor functions in patients with spinal cord injuries [1,2].

Robotic exoskeletons are mechanical devices designed to attach to the human body and assist, enhance, or support movement, typically powered by electric actuators. These devices are widely used in rehabilitation to help patients regain mobility through controlled and repetitive movements [7,8].

An advantage of exoskeleton-assisted methods (EAM) is that rehabilitation sessions can be significantly extended, they eliminate the need to account for physiotherapist fatigue during exercises. However, this does not appear to result in improved outcomes. This lack of improvement has been attributed to the passive involvement of the patient, who may perform the exercises without active engagement, potentially leading to mental disengagement and, consequently, a diminished capacity to enhance neuroplasticity. To overcome this limitation, the integration of electromyography (EMG) signals into exoskeleton control systems has been proposed as one of several techniques to facilitate active rehabilitation, thereby improving the overall effectiveness of EAM for rehabilitation [7,8].

EMG is a reliable and well-established method for recording the electrical activity of muscles, enabling precise tracking of muscular activity during movement. The response range extends from 0.01 Hz to 10 kHz in bandwidth, encompassing low-frequency components often associated with physiological or mechanical artifacts, as well as high-frequency details relevant for advanced intramuscular EMG studies. This broad range provides flexibility for experimental setups and specialized applications, although the typical useful bandwidth for surface EMG lies between 20 Hz and 500 Hz. By quantifying the intensity of muscle-generated electrical signals, EMG provides valuable insights into muscle function, treatment effectiveness, and the degree of force exerted, facilitating the personalization of rehabilitation sessions [7]. This makes EMG a powerful tool for integration into therapeutic exoskeletons, as it allows the evaluation of both the patient’s intent and movement force, even when the generated force is insufficient or gradually improving [9,10,11].

EMG-driven systems could enable exoskeletons to dynamically adjust their assistance to match the user’s physical capabilities, potentially supporting real-time adaptive control that fosters personalized human–machine interaction [12,13]. This approach may improve the effectiveness of assisted movements while encouraging active user participation, which could play a crucial role in promoting neuroplasticity and enhancing rehabilitation outcomes.

Despite its advantages, using EMG in exoskeleton control systems presents challenges, primarily due to the low amplitude and noise susceptibility of signals sampled, compounded by external interference, movement artifacts, and electrode placement variations (low SNR) [7,14]. These issues require advanced filtering and conditioning techniques. Additionally, integrating EMG-based control demands real-time signal processing with high-performance hardware and software to ensure seamless interaction and minimize latency, which is critical for user confidence and safety. Nevertheless, EMG remains invaluable for assessing force and movement intent, enabling adaptive assistance tailored to the user’s needs. Advancements in sensor technology and machine learning algorithms continue to enhance the accuracy and reliability of EMG-based systems, making them fundamental to personalized rehabilitation strategies.

While commercial EMG systems are well-suited for standard conditions, their sampling rates and reliance on external processing devices often limit their adaptability to advanced applications, such as real-time motion control in therapeutic exoskeletons. To address these limitations, this study proposes the development of a high-performance multichannel EMG acquisition system with a sampling rate of 8 kSPS per differential channel and real-time in situ signal processing capabilities [15].

The proposed system incorporates edge computing to locally process EMG signals, reducing latency and eliminating the need to transmit noisy raw signals. Additionally, by minimizing cable length and optimizing signal acquisition, the system mitigates electromagnetic interference and enhances portability. Designed for seamless integration into ROS2-based control architectures, the system transmits only processed data to robotic nodes, ensuring low-latency operation and enabling modular and scalable designs.

This approach offers a robust and flexible solution for improving the control of therapeutic exoskeletons, particularly in applications that require adaptive and real-time responses. The proposed system is intended to facilitate iterative development, improve signal reliability, and support diverse experimental setups, making it a valuable contribution to the field of rehabilitation robotics. To address the substantial computational demands associated with processing eight differential channels at a sampling rate of 8 kSPS, this work proposes the use of System-on-Chip (SoC) devices based on FPGAs. These platforms allow for the parallel processing of channels and the execution of necessary signal conditioning procedures, while also facilitating the integration of ROS2 into the embedded microprocessor within the SoC architecture. For implementation, a development board featuring the Zynq-7000 platform is employed, leveraging its tightly coupled ARM processor and FPGA fabric to balance real-time processing requirements with system flexibility.

2. State of the Art

To acquire an EMG signal, a relatively simplex system is required that functions both as a signal amplifier and as a processing unit. The signal amplifier boosts the weak electrical signals generated by the muscles, as these signals typically have low amplitude and are highly affected by noise [7,8].

Needle electrodes or intramuscular electrodes (iEMG), due to their direct contact with muscle fibers (invasive), offer more precise electrical measurements with reduced noise and interference compared to surface electrodes. While iEMG are less affected by external noise and movement artifacts, surface electrodes (sEMG), being non-invasive, are more prone to these issues and capture signals from broader areas, which lowers the signal-to-noise ratio [14].

Once amplified, the signals must be processed to filter out this unwanted noise, enhance relevant frequencies, and convert the analog signals into digital formats for further analysis. The processing unit then interprets these cleaned signals to assess muscle activity accurately, enabling real-time monitoring and control of muscle movement.

Processing EMG signals involves multiple steps, from preprocessing and feature extraction to real-time signal classification and integration into control systems. Each step contributes to enhancing the accuracy, efficiency, and usability of EMG-based applications, particularly in prosthetics and rehabilitation technologies [7,8,15].

1.

Preprocessing (Signal Conditioning): preprocessing ensures that EMG signals are clean and ready for analysis. Key processes include the following:

-

Amplification and Filtering: weak signals are amplified and processed using high-pass and low-pass filters to limit the signal bandwidth and remove noise or aliasing effects. A notch filter is employed to eliminate power-line interference.

-

Full-wave rectification and smoothing: an overall trend of muscle activity is obtained using an envelope detector.

2.

Feature Extraction Common methods include the following:

-

Time-domain features: Mean Absolute Value (MAV), Root Mean Square (RMS), Zero-Crossing Rate (ZCR), and Slope Sign Changes (SSC), which describe signal magnitude and dynamics.

-

Frequency-domain features: metrics such as Median Frequency (MDF), Mean Power Frequency (MPF), and Power Spectral Density (PSD) are used to detect muscle fatigue and activity.

-

Wavelet Transform: Provides detailed time-frequency analysis, offering a multi-scale view of muscle activity.

3.

Signal Classification: Processed signals are interpreted for real-time control, often using Supervised or Unsupervised Learning Techniques.

4.

Real-Time Processing: Critical for applications such as prosthetics or exoskeletons.

5.

Signal Integration (Control Systems): Once processed, EMG signals are integrated into control systems to operate external devices:

-

Command Generation: Signals are translated into specific actions, such as controlling prosthetic or robotic actuators.

-

Feedback Systems: Device feedback adjusts user movements and enhances system responsiveness and accuracy.

Nowadays, EMG systems can be classified on the following:

• Single Amplification and Filtering Systems: The simplest devices available. Consist of sEMG or iEMG electrodes connected to amplifiers and filters. They can support multiple wired channels to acquire signals from different locations using at least three electrodes, where one is used as a reference and the other two as a differential channel.
• High-Density EMG (HDEMG) Systems: Capture high-density EMG signals using more electrodes, enabling higher resolution by recording from multiple muscle sites simultaneously. While well-suited for research and diagnostics, they generate large volumes of data, which increase complexity, computational demands, and cost, making them ideal primarily for specific studies in controlled environments.

Without losing the generality of the previous classification, some emerging technologies can be highlighted that add value to these systems and show promise in the scientific field studying these topics:

• Wearable and portable sensors: Compact and flexible sensors, such as those based on conductive polymers, enable continuous and comfortable monitoring [16,17,18].
• Advanced digital processing: Machine learning algorithms and adaptive filters could improve pattern recognition in EMG signals, reducing noise and identifying patients’ movement intent.
• Integration with smart controllers: Modern exoskeletons utilize control architectures based on EMG signals to provide personalized assistance, adjusting parameters such as force and speed in real-time.

Typical EMG equipment features high-channel acquisition capabilities, with some systems supporting up to hundreds of channels for biomedical signals (including EMG and EEG). These systems commonly include wireless transmission of data to external computers, with battery life ranging from 4 to 8 h depending on the model.

Maximum sampling rates typically range from 2 kSPS to 10.24 kSPS, though most systems are limited to around 3 kSPS, which may not fully capture the broad EMG bandwidth (0.01 Hz to 10 kHz). Some devices include inertial measurement units (IMUs) for additional data, and their sensors may use patch electrodes or mesh arrangements to capture signals. Data storage options, such as flash memory cards, are also available for longer-range recording (over 100 m).

Despite these capabilities, all devices generally require connection to an external computer for signal processing, making them less portable and limiting their use in mobile applications, such as integration with exoskeletons. Moreover, aside from the need to be connected to a computer, if we combine the total size of all the devices necessary for capturing and processing the EMG signal in each individual system, they result in large systems that are hardly portable and, therefore, difficult to apply in exoskeletons.

3. System Requirements and Architecture Implementation

When designing our system of acquisition and processing, several key requirements must be met to ensure its functionality, precision, and adaptability. The following outlines the primary characteristics necessary for achieving optimal performance in EMG-based systems. The primary requirements are as follows:

• Four channels per muscle group: For accurate measurement of a muscle group, a minimum of four differential channels per group is required, with an additional reference channel. This setup allows for the reliable characterization of both contractions and relaxations of antagonistic muscles within the same group, ensuring precise movement analysis.
• High acquisition frequency and resolution: To capture the majority of the EMG signal, the acquisition rate should be as close as possible to twice the upper limit of the signal’s bandwidth, in accordance with the Nyquist–Shannon sampling theorem. Given that the upper bandwidth limit for an EMG signal is 10 kHz, the minimum acquisition frequency should be approximately 20 kHz or higher. Additionally, due to the low amplitude of the EMG signal, the acquisition system must include an amplification phase and a resolution of at least 16 bits to accurately capture signal details.
• High processing capacity: To significantly reduce the bandwidth required for transmitting information outside the device, an integrated system with sufficient computational capacity to process the acquired signals is essential. This approach minimizes data transmission by sending only specific commands to the exoskeleton actuators, thereby enhancing efficiency.
• Fast processing and flexibility: The device must be sufficiently flexible to allow for the programming and construction of the entire EMG signal processing architecture. It should be capable of handling near real-time processing to ensure that the resulting exoskeleton movements are reliable and fluid.
• Portability: The acquisition system should be compact enough to be transported with one hand, facilitating its use in various contexts beyond the exoskeleton application. This portability feature enables convenient testing and measurement in diverse environments.

The analog signal captured by the electrodes must be conditioned and digitized to facilitate its analysis and use. While final products often employ custom-designed circuits tailored to the specific application, the complexity of designing, calibrating, and managing communication in these circuits makes commercial acquisition AFEs (Analog Front-Ends) a practical choice for early development stages.

Commercial AFEs, although more expensive and less application-specific, provide a reliable and rapid solution for prototyping. For this project, the ADS1298 [19], part of Texas Instruments^® (headquartered in Dallas, TX, USA) ADS129x family, was selected due to its configurable multichannel architecture, low power consumption, and high performance in biomedical signal acquisition, such as EMG, ECG (Electrocardiography), and EEG (Electroencephalography).

The ADS1298 integrated circuit offers eight channels, providing greater flexibility compared to other devices in the series. It also includes a respiration module which, although not used in this work, could be valuable for future developments. For the initial development phase, we utilized the Texas Instruments ADS1298RECGFE development board as analog front end (AFE), which integrates the ADS1298R chip. This enabled us to refine and mature our design concept. Subsequently, using the same chip, we developed custom hardware, ensuring reliability and functionality after validating and testing the work conducted with the development board.

The ADS1298 supports sampling rates of up to 32 kSPS, but at higher frequencies, the bit resolution decreases, as specified in the datasheet: 24 bits for up to 8 kSPS, 19 bits at 16 kSPS, and 17 bits at 32 kSPS. For this project, a sampling frequency of 8 kSPS was selected to ensure maximum resolution. This flexibility enables potential adjustments for future applications or specific signal acquisition needs. Conversely, the use of 24-bit resolution provides a voltage resolution as low as 300 μV, thereby enabling the detection of even minimal changes without the necessity of significantly increasing the gain of the acquisition system.

A microcontroller-based system cannot handle high acquisition rates, restricting the bandwidth for capturing signals. While sufficient for low-frequency EMG signals, the system fails to capture signals with higher frequencies, as it operates at only 1 kSPS. Given that EMG signals span a bandwidth between 0.01 Hz and 10 kHz, most of this range remains unsupported. Moreover, its limited processing capacity prevents the implementation of real-time processing necessary for the exoskeleton’s control loop.

To address these challenges, a SoC FPGA-based system was developed. The selected device, the ZYNQ XC7Z010-1CLG400C is manufactured by AMD-Xilinx (headquartered in San Jose, CA, USA) and integrates a high-performance dual-core ARM^® Cortex™-A9 MPCore^TM processor with 28 nm programmable FPGA logic. This architecture combines on-chip memory, external memory interfaces, and a broad range of peripheral connectivity interfaces, offering the processing power, flexibility, and real-time capabilities required for the system. This flexibility reduces latency and enhances the integration of ROS2 nodes, ensuring modularity and scalability for the robotic subsystems.

High-speed data acquisition on Zynq-7000 boards is supported by its capability to handle computationally intensive tasks without external latency. With its multi-channel processing and integrated ARM cores, the Zynq-7000 enables systems to preprocess and analyse EMG data directly on the hardware, minimizing the data bandwidth required for external transmission.

Zynq-7000 platforms support ROS2 integration through tools like meta-FOrEST (https://meta-forest.readthedocs.io/en/latest/introduction.html, accessed on 5 November 2024), which is an automatic creation tool for nodes of ROS2 systems integrating several FPGA logics into a ROS2 system (ROS2-FPGA nodes) and enables seamless communication between the FPGA and ROS2 nodes. This facilitates the deployment of ROS2-based robotic frameworks by automating the generation of FPGA nodes, reducing programming complexity. For example, meta-FOrEST can link the FPGA processing cores directly with ROS2 middleware, enhancing performance while maintaining a flexible modular architecture for robotic subsystems [20,21].

The integration of the ADS1298 analog front-end and its connection to a computer for efficient signal processing are detailed below, as shown in Figure 1.

The architecture comprises an IP block that oversees the AFE Controller (ADS SPI Controller), which is linked through a multiplexer to a buffer where data is temporarily stored before being transmitted to DDR memory, via DMA. The blocks used are as follows:

• ADS SPI Ctrl: This custom IP block acts as the controller for SPI communication when receiving data from the AFE. It temporarily stores and manages the captured data until the DMA is ready for transmission. It includes an AXI4-Lite slave interface for interaction with the processing block, an AXI-Stream master interface for data transfer to the DMA, and an SPI communication module with a 128 × 32 FIFO memory to ensure no data loss. Additionally, it features a clock divider that converts the system clock (50 MHz) to a 4 MHz clock for SPI communication.
• ADS SPI Mux: This custom IP block functions as a multiplexer that allows switching between two SPI data streams connected to the AFE. The first stream, linked to the AXI Quad SPI block, configures the AFE, while the second stream, connected to the output of ADS Ctrl SPI, manages the captured data received from the AFE. The multiplexer selection is controlled via an AXI4-Lite slave interface.
• AXI Quad SPI: This is a standard Vivado IP block responsible for managing and establishing SPI communication between the AFE and the processor. It reduces the input clock frequency to achieve a 4 MHz SPI output clock necessary for AFE communication. The block provides SPI outputs for data transmission (MOSI), clock (SCK), and chip select (CS), along with an input for data reception (MISO).
• Processing Algorithm: This IP has been generated from Vitis HLS, and is responsible for implementing the low-level processing algorithms required for the data streams received from the ADS1298. These include signal processing to obtain the on-set detection [22], and buffering for each channel (e.g., ch0 buffer, ch1 buffer, etc.). It is connected to the AXI4-Lite bus for configuration and uses the AXI Stream interface to interact with the DMA, enabling efficient data transfer to the DDR memory for further processing. Different on-set detection algorithms have been implemented with Vitis HLS, such as moving average, moving RMS, low-pass filter with FIR (Finite Impulse Response), and TKEO (Teager–Kaiser Energy Operator). Due to its simplicity and efficiency, the moving average implementation with a fixed threshold has been used for this work. This consists of rectifying the scaled data and calculating the moving RMS for 64 samples, as shown in Equation (1), where $x_{rms}$ is the result of the moving RMS; N is the number of samples, in this case 64; and $|x_i|$ is the absolute value of each scaled sample.

  $$\begin{array}{c}\hfill x_{rms}={\displaystyle \frac{1}{N}}\sum _{i=0}^{N-1}\left|x_i\right|\end{array}$$

  (1)
  Then, a threshold is applied to the moving RMS signal to determine whether muscle activation has occurred. This threshold is configurable and is used to obtain a digital signal, as shown in Equation (3), where $x_{th}$ is the result of the thresholding; $x_{rms}$ is the moving RMS signal; and $th$ is the threshold.

  $$x_{th}=\left\{\begin{array}{cc}\mathrm{if}{x}_{rms}>th\hfill & 1\hfill \\ \mathrm{if}{x}_{rms}\le th\hfill & 0\hfill \end{array}\right.$$

  (2)

4. Experimental Results

Experimental results validate the proposed SoC architecture for EMG acquisition with a sampling frequency of 8 kSPS. The effectiveness of the developed prototype is compared with commercial systems, showing significant improvements in signal-to-noise ratio and overall performance. A discussion on latency and effectiveness in the context of active rehabilitation therapies is included. The final setup for the system validation is shown in Figure 2.

Four key experimental tests were conducted to evaluate the system’s performance. These were the following: functional validation, latency assessment, time and frequency analysis of captured EMG signals, and resource assessment on SoC-FPGA. Prior to the description of the experimental procedure, the EMG signal acquisition protocol performed in this study is described below.

4.1. Signal Acquisition Protocol

In order to validate the architecture, the Rectus Femoris and Vastus Lateralis muscle were measured and the subject’s skin was prepared for electrode placement. This procedure minimizes the attenuation and possible distortion of EMG signals. Following the recommendations of the SENIAM project guidelines [23], the electrodes are placed at the midpoint of the imaginary line connecting the superior iliac spine and the top of the patella.

The subject is initially positioned seated on a table, with the knees slightly flexed and the upper body slightly reclined. This position facilitates knee extension without thigh rotation during measurements. For data capture, the subject extends the knee while an assistant applies pressure against the leg above the ankle in the direction of flexion. This induces tension in the target muscle. An initial capture is performed without muscle activation, followed by ten normal measurements for each test. Each measurement lasts 30 s, with the first 5 s without activation, followed by intervals of 5 s of muscle activation (maximum possible force) and 5 s of relaxation. This cycle is repeated three times per capture. A 1-min rest is allowed between captures.

A total of ten recordings are performed, as described above. The prototype is configured to acquire data at 8 kSPS with a gain of 12. This gain was selected by recording the maximum voluntary contraction. The approximate total duration of the test is three hours. This time includes explaining the protocol to the subject, preparing the area where the electrodes will be placed, and conducting ten signal recordings, each with an average duration of twenty minutes.

It is important to note that irregular contact between the electrodes and the skin, the interface between the conductor cables and the electrodes, and the connection to the ADS1298 module may introduce potential artifacts. This protocol was followed for the experimental validation of the proposed architecture, ensuring the accuracy and reliability of the obtained measurements.

Finally, the functional validation was performed by measuring a single muscle group, the Rectus Femoris and Vastus Lateralis. However, the remaining channels were configured to acquire the internal test signal generated by the ADS1298. This means that while data are acquired and processed on these channels, it does not correspond to meaningful EMG information.

4.2. Functional Validation

The system’s functionality was confirmed by programming specific channels of the AFE to acquire internal test signals and noise. Channels 1, and 4 to 8 were configured to acquire a 2 Hz square wave test signal, while channel 2 captures signals from the Vastus Lateralis muscle, and channel 3 captures signals from the Rectus Femoris muscle. Results demonstrated accurate transmission of the expected square wave signal, correct differentiation between channels, adherence to the desired data acquisition rate of 8 kSPS, and no sample loss during UART transmission.

Figure 3 shows the reception on channels 1 to 8 for functional validation. Channel 2 captures signals from the Vastus Lateralis muscle, while channel 3 captures signals from the Rectus Femoris muscle. The remaining channels represent the internal test signal.

4.3. Latency Assessment

This test involved capturing time measurements during data transmissions in MATLAB^® R2024b. Note: This is necessary to be done, otherwise it will affect publication. The setup included sending a continuous data stream from the prototype to a client and comparing the elapsed time for transmission with the effective signal duration captured. The effective duration was calculated based on the known data acquisition rate of 8 kSPS, verified during the functional validation test. The measurements demonstrated that capturing 58 s of effective signal required an average total time of 101 s (1 min and 41 s). This latency highlights the enhanced efficiency of the UART-based transmission system combining DMA concept.

The prototype architecture uses a regular DMA-UART-based system to transfer data between the proposed SoC architecture and the computer, minimizing latency between frames. Experimental tests were conducted to measure statistical latency metrics and evaluate the system’s ability to meet timing requirements for critical applications. Two latency perspectives were analyzed, byte-level and burst-level (3 bytes) metrics, to understand the effects of system jitter and bursts, respectively. The results highlight the importance of transmission speed and stability in ensuring data integrity and reliable system performance.

The UART system was tested with 9 data channels (8 channels and a status word channel), each transmitting 24-bit samples. Due to UART protocol overhead, each sample expands to 30 bits (including start and stop bits). Data were continuously transmitted for one minute at two baud rates: 115,200 bps and 230,400 bps. The maximum number of data points transmitted per channel in one minute was calculated for both baud rates.

The operating system (OS) and receiver hardware affect the accuracy of the measurement. Non real-time OSs, like Windows or Linux, introduce jitter due to multitasking, causing latency variability. Interrupt-driven processes can delay serial data handling, especially under heavy system load. UART buffer management is also crucial if the buffer fills faster than it can be processed, data loss or delays occur, particularly at high baud rates. Additionally, hardware receiver performance (CPU, memory, UART controller) affects data processing speed, with older or limited systems struggling to handle high-speed transmissions, leading to inaccuracies or dropped data.

Table 1. Typical statistical results of latency measurements using two different baud rates: 115,200 bps and 230,400 bps.

Metric	Baud Rate: 115,200bps		Baud Rate: 230,400bps
	Byte-Level	Burst-Level	Byte-Level	Burst-Level
Max. Latency (ms)	6.22	6.20	7.18	7.2
Min. Latency (μs)	0.4	17.9	0.3	14.8
Mean Latency (ms)	0.25	0.85	0.13	0.43
Median Latency (ms)	0.002	0.98	0.001	0.090
Standard Deviation (ms)	0.43	0.34	0.31	0.44

shows statistical metrics describing latency behaviour at two baud-rates: 115,200 bps and 230,400 bps. The values include typical latency statistics measured both by byte-level and burst-level during the continuous transmission of 24-bit data. The observed latencies were generally low for both cases, with medians below 1 ms, indicating consistent performance. However, the higher speed (223,400 bps) showed slightly greater variations in maximum latency, possibly due to buffer management or system interruptions. These findings emphasize the importance of considering both byte-level and burst-level metrics to evaluate typical performance and detect rare delays in time-sensitive applications.

The latency analysis demonstrates that the system operates well within the required response time needed for real-time operation, even under the highest measured latency conditions. Both baud rates (115,200 bps and 230,400 bps) deliver consistent performance, with median and mean latencies significantly below 1 ms, highlighting the system’s suitability for time-sensitive applications. While higher transmission speeds slightly increase maximum latency, the variations remain minimal and manageable, ensuring reliable data transfer and real-time response capabilities.

To address issues such as jitter and buffer overflow, practical solutions can be implemented. These include using real-time operating systems (RTOS) to ensure deterministic task management and optimizing UART configurations to enhance buffer handling and transmission efficiency.

4.4. Time-Frequency Analysis of EMG Signals

EMG signals were captured from the Rectus Femoris and Vastus Lateralis. This was achieved by varying the AFE gains, highlighting a clear activation signal, particularly at gains of 8 and 12, with a high signal-to-noise ratio. The frequency spectrum analysis revealed that the signal was predominantly within 3 Hz to 250 Hz, attenuating to 2.8 kHz. High CMRR of the differential amplifier rejected mains frequency interference, negating the need for additional filtering. Figure 4 shows the time (top) and frequency (bottom) representations of the EMG signals received by the AFE at gain 12. It is worth noting that captures 1 and 2 were taken at different times, but they represent the acquisition of the EMG for the Rectus Femoris, which is the muscle most activated by the movement of the protocol.

In terms of SNR, Capture 1 achieved an SNR of 34.66 dB, while Capture 2 recorded an SNR of 26.75 dB. The SNR in decibels (dB) is calculated as shown in (3), where ($P_{\mathrm{signal}}$) represents the power of the signal, and ($P_{\mathrm{noise}}$) represents the power of the noise.

$$\begin{array}{c}\hfill \mathrm{SNR}\left(\mathrm{dB}\right)=10·{log}_{10}\left({\displaystyle \frac{P_{\mathrm{signal}}}{P_{\mathrm{noise}}}}\right)\end{array}$$

(3)

The results demonstrate that the SNR values for both captures are high, confirming that the EMG signals were recorded with minimal noise interference. This is crucial for the accurate analysis and interpretation of muscle activity, as it ensures that the data reflect true physiological signals rather than artifacts or noise.

4.5. Resource Utilization

Resource consumption on the TE0727 (Z7-10) SoC-FPGA was minimal, using 15.9% of LUTs, 11.9% of registers, and 5% of available RAM blocks, leaving capacity for additional processing logic as can be summarized in

Table 2. Hardware resource consumption for TE0727 (Z7-10) SoC-FPGA.

Component	Slice LUTs	Slice Registers	Block RAM Tile
Global System	2798 (15.9%)	4189 (11.9%)	3 (5%)
ADS Ctrl SPI	130 (0.74%)	273 (0.78%)	0.5 (0.83%)
ADS Mux SPI	54 (0.31%)	170 (0.48%)	0 (0%)
DMA	1013 (5.76%)	1520 (4.32%)	2.5 (4.17%)
AXI Quad SPI	376 (2.14%)	575 (1.63%)	0 (0%)
Additional blocks	778 (4.42%)	921 (2.62%)	0 (0%)

. However, it is important to note that using a higher sampling frequency, such as the indicated 16 kSPS and 32 kSPS, results in increased FPGA resource consumption. This is because, to achieve the same capture time, the amount of generated data is two or four times greater, respectively.

One identified limitation was that the higher transfer speed of DMA compared to UART transmission caused data overwrites in DDR memory. Increasing allocated DDR mitigated this issue, enabling continuous signal acquisition for up to 7 min and 38 s, deemed sufficient for volunteer evaluations. A permanent solution is proposed for future work.

5. Conclusions

The proposed architecture based on a System-on-Chip (SoC), embodies key edge computing principles for EMG signal processing. By performing local data processing at the acquisition site, such as high-speed analog-to-digital conversion, advanced filtering, and initial signal processing, it significantly reduces latency and enables responsiveness in real-time. This approach minimizes raw data transmission, optimizing bandwidth while enhancing signal-to-noise ratio (SNR) and overall data integrity, which are critical for reliable robotic neurorehabilitation applications.

The integration of the ADS1298R chip with an FPGA-based architecture, specifically the Zynq-7000 platform, addresses the limitations of microcontroller-based systems, such as their inability to handle high acquisition rates and to support the full bandwidth of EMG signals (0.01 Hz to 10 kHz). This approach not only reduces dependency on external processing but also improves modularity and scalability, making it suitable for advanced applications like integration with exoskeletons and other rehabilitation technologies.

While the developed system demonstrates significant advancements, it still faces challenges, such as the computational load associated with higher channel counts and sampling frequencies. Nevertheless, the FPGA architecture offers the flexibility needed to overcome these limitations in future iterations. The next development stages will include integrating real-time processing capabilities directly into the SoC and continuously optimizing system resources.

Finally, the proposed system represents not only a technical breakthrough but also establishes a solid foundation for applications in robotic rehabilitation and adaptive control of medical devices, reaffirming its relevance in the field of neurorehabilitation. This architecture is highly adaptable and can be utilized in various neurorehabilitation applications, including the control of assistive robots for rehabilitation. Its modularity and scalability make it an ideal candidate for integration with exoskeletons and other advanced rehabilitation technologies, ensuring its continued relevance and applicability in the evolving field of medical device control and neurorehabilitation.