Event-driven electromyographic signal processing system based on memristor

Abstract

The invention relates to a memristor-based event-driven electromyographic signal processing system, which comprises an electromyographic acquisition electrode, an FPGA (Field Programmable Gate Array) control panel and a memristor array, the FPGA control panel comprises a control module, a WIFI module, a pulse scheduling module, an electromyographic signal coding module, a read operation module, a write operation module, a neural network parameter storage module, an LIF neuron state updating module and an LIF neuron state storage module, and the electromyographic signal coding module comprises an event-driven coding module. Compared with the prior art, the method has the advantages of reducing the energy consumption of the electromyographic signal processing system and the like.

Description

Technical Field

[0001] The present invention relates to an electromyographic signal processing system, and in particular to an event-driven electromyographic signal processing system based on a memristor. Background Art

[0002] Electromyography (EMG), a bioelectric signal generated by muscles during movement, has a wide range of applications in medical rehabilitation, human-computer interaction, motion analysis, and other fields. For example, in this field, EMG is often used to control prosthetic limbs and exoskeletons. By identifying subtle muscle movements in the residual limb, intelligent prostheses can mimic the user's natural movements, provide precise motion control, and offer a more natural and intuitive user experience. The development of wearable devices and intelligent prosthetic technology has placed higher demands on the real-time processing and analysis of EMG signals.

[0003] In these application scenarios, EMG signal processing systems must offer high real-time performance and low power consumption to meet the demands of edge devices and long-term use. Furthermore, as neural networks are increasingly used in EMG signal processing, existing neural network-based processing solutions often require the storage of a large number of parameters to support complex network structures and a large number of computing units to complete neural network inference. This not only increases hardware costs but also increases energy consumption, reducing the system's ability to continue operating. Summary of the Invention

[0004] The purpose of the present invention is to provide an event-driven myoelectric signal processing system based on a memristor in order to reduce the energy consumption of the myoelectric signal processing system.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A memristor-based event-driven electromyographic signal processing system includes an electromyographic acquisition electrode, an FPGA control board, and a memristor array. The FPGA control board includes a control module, a Wi-Fi module, a pulse scheduling module, an electromyographic signal encoding module, a read operation module, a write operation module, a neural network parameter storage module, a LIF neuron state update module, and a LIF neuron state storage module. The electromyographic signal encoding module includes an event-driven encoding module.

[0007] The event-driven encoding module includes a comparator, a selector and an address event representation module. The input of the event-driven encoding module is the original electromyographic signal. The original electromyographic signal and the previous iterative signal are added to generate the current signal. The signal of the current signal after passing through the comparator is input into the address event representation module to obtain the pulse event signal output by the event-driven encoding module to the pulse scheduling module. The output of the pulse scheduling module serves as the input of the read operation module. The read operation module applies a read pulse to the memristor array and receives the current feedback from the memristor array. The feedback current Current_in and the state State_read stored in the LIF neuron state storage module at the previous moment are used as the input of the LIF neuron state update module. The LIF neuron state update module adds the current Current_in and the read state State_read to output a new membrane potential. If the new membrane potential exceeds the pulse threshold Vth, the pulse output Output is sent from the output end, and the new membrane potential is sent to the LIF neuron state storage module as the write state State_write to complete the update of the neuron membrane potential state.

[0008] Furthermore, the current signal is subtracted from the preset value, and the resulting signal is input to port 1 of the selector, the current signal is input to port 0 of the selector, the output of the comparator is input to the control port of the selector, and the current iteration signal is obtained by subtracting the subtrahend leak from the output of the selector.

[0009] Furthermore, the negative input of the comparator is a preset value, the positive input of the comparator is a current signal, and the output of the comparator serves as the input of the address event representation module.

[0010] Furthermore, the input of the WIFI module is connected to the output of the electromyography acquisition electrode, the output of the WIFI module serves as the input of the electromyography signal encoding module, and the pulse event signal output by the electromyography signal encoding module serves as the input of the pulse scheduling module.

[0011] Furthermore, the output of the neural network parameter storage module serves as the input of the write operation module, and the output of the write operation module is connected to the memristor array.

[0012] Furthermore, the current fed back by the memristor array is the current calculated by the synapses of the neurons of the spiking neural network;

[0013] After the memristor array receives a read pulse, the read pulse is applied to the forward neuron of the current layer as the current neuron input. The output of the forward neuron is applied to the bit line BL of the memristor array. The memristor array source line SL outputs the aggregate current of each column of the memristor array. The difference between the aggregate currents of two adjacent columns is used as the input of the backward neuron of the next layer. The above steps are repeated to finally obtain the current fed back by the memristor array.

[0014] Furthermore, the control module is connected to communicate with the WIFI module, the electromyographic signal encoding module, the neural network parameter storage module, the LIF neuron state updating module and the LIF neuron state storage module.

[0015] Furthermore, the pulse event signal is a pulse event sequence containing only 0s and 1s.

[0016] Furthermore, the memristor array is a non-volatile device.

[0017] Furthermore, the modules of the FPGA control board communicate with each other through handshake.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] The present invention can reduce the energy consumption of myoelectric signal processing systems: it uses a memristor array to perform synaptic weight calculations between neural network layers. Because memristors can directly perform "vector-matrix multiplication and accumulation" operations (BL input voltage V, the memristor's own conductance value G stores information, I = V * G, and the resulting current value is the multiplication and accumulation result), compared to operations in digital circuits, it can reduce energy consumption. The present invention also exploits the temporal and spatial characteristics of myoelectric signals. Myoelectric signals are encoded into pulse events by LIF neurons. During the encoding process, LIF neurons can remember signals from past moments. This exploits the temporal correlation between myoelectric signals and the past. The multi-bit myoelectric signals are then converted into pulse sequences. The pulses in the pulse sequence are only "0" and "1". A pulse event is only generated when the membrane potential of the LIF neuron reaches the pulse emission threshold. This consumes fewer resources than multi-bit values of unprocessed data. While reducing data density, it is better adapted to the reasoning of spiking neural networks and the use of memristor chips, thereby reducing the energy consumption of the myoelectric signal processing system. The present invention also adopts multiple myoelectric acquisition electrodes corresponding to respective encoding channels, which are processed separately, thereby exploring the spatial regularity of the myoelectric signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a structural schematic diagram of the present invention;

[0021] Figure 2 This is a circuit diagram of an event-driven encoding module of the present invention;

[0022] Figure 3 This is a diagram of the memristor array deployment and forward reasoning structure of the spiking neural network;

[0023] Figure 4 It is the structure diagram of the system control scheme;

[0024] Figure 5 Encoding results for raw EMG signals and event-driven coding;

[0025] Figure 6 Confusion matrix of system inference results for the test set;

[0026] Figure 7 Provide real-time inference results for the system. DETAILED DESCRIPTION

[0027] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0028] This paper proposes an EMG signal processing system based on event triggering and memristors. This system exploits the temporal and spatial characteristics of EMG signals, addressing the high power consumption and latency issues of existing solutions. Electrodes worn on the arm collect EMG signals, and an FPGA controls the system data flow, pre-processing the signals. The memristor array then processes and identifies the EMG signals using a spiking neural network, outputting corresponding control instructions to control external devices such as prosthetic limbs and robotic arms.

[0029] The system includes electromyographic acquisition electrodes, an FPGA control board and a memristor array. The FPGA control board includes a control module, a WiFi module, a pulse scheduling module, an electromyographic signal encoding module, a read operation module, a write operation module, a neural network parameter storage module, a LIF neuron state update module and a LIF neuron state storage module. The electromyographic signal encoding module includes an event-driven encoding module.

[0030] The event-driven encoding module includes a comparator, a selector and an address event representation module. The input of the event-driven encoding module is the original electromyographic signal. The original electromyographic signal and the previous iterative signal are added to generate the current signal. The signal of the current signal after passing through the comparator is input into the address event representation module to obtain the pulse event signal output by the event-driven encoding module to the pulse scheduling module. The output of the pulse scheduling module serves as the input of the read operation module. The read operation module applies a read pulse to the memristor array and receives the current feedback from the memristor array. The feedback current Current_in and the state State_read of the previous moment stored in the LIF neuron state storage module serve as the input of the LIF neuron state update module. The LIF neuron state update module outputs a new membrane potential. If the new membrane potential exceeds the pulse threshold Vth, the pulse output Output is sent from the output end, and the new membrane potential is sent to the LIF neuron state storage module as the write state State_write to complete the update of the neuron membrane potential state.

[0031] This paper proposes event-driven coding and computational solutions for signal preprocessing and subsequent processing modules, reducing system energy consumption and enabling real-time, low-power signal processing. Furthermore, in conjunction with peripherals, it can implement gesture-based robotic control tasks.

[0032] System FPGA control core: such as Figure 1 As shown in the figure, the system consists of three components: EMG acquisition electrodes, an FPGA control board, and a memristor array. After the electrodes collect the EMG signals, signal preprocessing and calculation are performed by the FPGA, utilizing its internal resources and the memristor chip. The FPGA includes the following modules: a control module, a Wi-Fi module, a pulse scheduling module, an EMG signal encoding module, a read operation module, a write operation module, a neural network parameter storage module, a LIF neuron state update module, and a LIF neuron state storage module. The control module, as the top-level module, communicates and controls the remaining modules. After the EMG signals are received by the Wi-Fi module, the built-in EMG signal encoding module performs LIF neuron-based encoding operations, encoding the multi-bit signals into pulse trains, which are then fed into the pulse scheduling module for further processing.

[0033] Event-driven encoding: This encoding process converts electromyographic signals into pulse events and embeds them in the "electromyographic signal encoding module". While reducing data density, it better adapts to the reasoning of pulse neural networks and the use of memristor chips. Figure 2 The LIF neuron circuit structure shown converts the input multi-bit raw data into a pulse event sequence containing only "0" and "1", and transmits it to the subsequent pulse scheduling module for use.

[0034] Memristor array: The memristor array can be operated by applying signals through FPGA. The write operation module can write the pre-trained pulse neural network weight parameters to the memristor device in the memristor chip; the read operation module can complete the neuron synaptic calculation of the pulse neural network in the memristor chip, and return the calculation results to the next layer of neurons in the neural network for further reasoning, thus achieving the realization of "storage and computing in one". Specific synaptic weight deployment plan Figure 3 As shown, the synaptic weights of the spiking neural network will use a differential method to deploy positive and negative weights, so each weight is represented by two memristor devices (the difference in conductance between two adjacent columns is the weight value). The output pulse corresponding to the forward neuron (Pre-Neuron) will be applied to the BL of the memristor array, and the current accumulated in each column will be read on the corresponding SL. The currents of two adjacent columns are subtracted to obtain the input of the corresponding backward neuron (Post-Neuron).

[0035] Peripheral control: After the network completes the inference of the EMG signal, it will send the corresponding control instructions to the peripheral device to realize the EMG signal processing system based on the memristor.

[0036] like Figure 4 As shown, the system consists of four components: EMG acquisition electrodes, an FPGA control board, a memristor array, and peripherals. The core of the system is the FPGA, which communicates and controls the external EMG acquisition electrodes, memristor chip, and peripherals. After the EMG acquisition electrodes collect EMG signals, they send EMG_data to the FPGA control board's Wi-Fi module, which has an internal FIFO module for data buffering. The data is then input to the EMG signal encoding module, where internal encoding logic processes the input data and outputs a pulse sequence, Spike_data, to the pulse scheduling module. The read operation module identifies the address of the Spike signal from the pulse scheduling module, applies a Read_pulse to the memristor chip, and receives the read Current, completing the calculation of the synaptic weights between the neural networks based on the memristor chip. This read current value serves as the input Current_in to the LIF neuron state update module, which stimulates the neuron's read state State_read and determines whether to fire a pulse. The neuron's membrane potential state State_write is then stored in the LIF neuron state storage module, completing the neuron's membrane potential state update.

[0037] The neural network parameters are written to the memristor chip via a write module. The memristor device used in this invention is non-volatile, so the write module is only needed to adjust the resistance of the memristor device in the memristor chip when retraining the neural network parameters or recalibrating the memristor chip. Each module in the system communicates through Valid and Ready handshakes to ensure efficient data transmission and correct processing.

[0038] EMG signal encoding results: By using neurons to pulse encode the original EMG signal, the results are as follows: Figure 5 shown. Figure 5 The left side shows the normalized EMG waveform corresponding to the eight electrode channels. It can be seen that the unprocessed data is complex and dense. The pulse sequence after pulse encoding is shown on the right side, which shows that the data complexity is greatly reduced, making it more convenient for transmission.

[0039] System dataset and real-time reasoning result analysis: We use the test set of the dataset to perform reasoning tests on the system. The confusion matrix of the reasoning results is as follows: Figure 6 As shown in Figure 2, in the five gesture classification tasks, the classification accuracy of the memristor-based gesture classification can reach 95.76%. The continuous memristor-based gesture classification results and the correct results are shown in Figure 2. Figure 7As shown in the figure, the EMG signal processing and classification based on this system achieves excellent results in real-time applications. By comparing the actual values and the predicted values, we can see that the classification is correct in most cases, proving the reliability of the system.

[0040] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. An event-driven electromyographic signal processing system based on memristor, characterized in that: It includes electromyographic acquisition electrodes, an FPGA control board and a memristor array. The FPGA control board includes a control module, a WIFI module, a pulse scheduling module, an electromyographic signal encoding module, a read operation module, a write operation module, a neural network parameter storage module, a LIF neuron state update module and a LIF neuron state storage module. The electromyographic signal encoding module includes an event-driven encoding module. The event-driven encoding module includes a comparator, a selector and an address event representation module. The input of the event-driven encoding module is the original electromyographic signal. The original electromyographic signal and the previous iterative signal are added to generate the current signal. The signal of the current signal after passing through the comparator is input into the address event representation module to obtain the pulse event signal output by the event-driven encoding module to the pulse scheduling module. The output of the pulse scheduling module serves as the input of the read operation module. The read operation module applies a read pulse to the memristor array and receives the current feedback from the memristor array. The feedback current Current_in and the state State_read stored in the LIF neuron state storage module at the previous moment are used as the input of the LIF neuron state update module. The LIF neuron state update module adds the current Current_in and the read state State_read to output a new membrane potential. If the new membrane potential exceeds the pulse threshold Vth, the pulse output Output is sent from the output end, and the new membrane potential is sent to the LIF neuron state storage module as the write state State_write to complete the update of the neuron membrane potential state.

2. The event-driven electromyographic signal processing system based on memristor according to claim 1, characterized in that: The current signal is subtracted from the preset value, and the resulting signal is input to port 1 of the selector. The current signal is input to port 0 of the selector. The output of the comparator is input to the control port of the selector. The current iteration signal is obtained by subtracting the subtrahend leak from the output of the selector.

3. The event-driven electromyographic signal processing system based on memristor according to claim 2, characterized in that: The negative input of the comparator is a preset value, the positive input of the comparator is a current signal, and the output of the comparator serves as the input of the address event representation module.

4. The event-driven electromyographic signal processing system based on memristor according to claim 1, characterized in that: The input of the WIFI module is connected to the output of the electromyography acquisition electrode, the output of the WIFI module serves as the input of the electromyography signal encoding module, and the pulse event signal output by the electromyography signal encoding module serves as the input of the pulse scheduling module.

5. The event-driven electromyographic signal processing system based on memristor according to claim 4, characterized in that: The output of the neural network parameter storage module serves as the input of the write operation module, and the output of the write operation module is connected to the memristor array.

6. The event-driven electromyographic signal processing system based on memristor according to claim 5, characterized in that: The current fed back by the memristor array is the current calculated by the neuron synapses of the spiking neural network; After the memristor array receives a read pulse, the read pulse is applied to the forward neuron of the current layer as the current neuron input. The output of the forward neuron is applied to the bit line BL of the memristor array. The memristor array source line SL outputs the aggregate current of each column of the memristor array. The difference between the aggregate currents of two adjacent columns is used as the input of the backward neuron of the next layer. The above steps are repeated to finally obtain the current fed back by the memristor array.

7. The event-driven electromyographic signal processing system based on memristor according to claim 1, characterized in that: The control module is connected to communicate with the WIFI module, the electromyographic signal encoding module, the neural network parameter storage module, the LIF neuron state updating module and the LIF neuron state storage module.

8. The event-driven electromyographic signal processing system based on memristor according to claim 1, characterized in that: The pulse event signal is a pulse event sequence consisting of only 0s and 1s.

9. The event-driven electromyographic signal processing system based on memristor according to claim 1, characterized in that: The memristor array is a non-volatile device.

10. The event-driven electromyographic signal processing system based on memristor according to claim 1, characterized in that: Handshake communication between various modules of the FPGA control board.

Images