SNN-based pulse muscle training neural signal monitoring method

Abstract

The invention discloses a pulse muscle training neural signal monitoring method based on SNN. The method specifically comprises the steps that S1, electromyographic signals generated by target muscles under the action of movement and electrical stimulation are collected; s2, converting the electromyographic signals into a time-ordered pulse sequence; s3, inputting the pulse sequence into a liquid state machine reservoir layer formed by threshold triggering pulse neurons, and outputting a pulse state; s4, forming a muscle state vector containing a fatigue component and an activation stability component according to the pulse state; s5, updating the time constant parameter of the spiking neuron according to the muscle state vector, and driving the reservoir layer to update the state; s6, constructing a model configuration vector, and completing grouping search updating through a grid adaptive direct search algorithm; and S7, generating electrical stimulation parameters according to the target model configuration vector, and applying electrical stimulation to form a closed-loop operation structure. According to the invention, SNN myoelectricity perception and stimulation closed loop are realized.

Description

Technical Field

[0001] This invention relates to the field of electromyography signal monitoring and closed-loop control technology based on spiking neural networks, and particularly to a method for monitoring nerve signals during spiking muscle training based on spiking neural networks. Background Technology

[0002] With the increasing demand for rehabilitation training, exercise training, and muscle function assessment, muscle state monitoring and electrical stimulation control technologies based on electromyography (EMG) signals have been widely applied. Current technologies typically involve collecting EMG signals generated by muscles during exercise or electrical stimulation, analyzing muscle activation, and then combining these with preset rules or models to output training assessment results or electrical stimulation parameters. However, EMG signals are non-stationary and exhibit significant individual variability. During prolonged training or repetitive stimulation, the signal distribution drifts with changes in muscle fatigue and activation state, leading to insufficient stability in monitoring results.

[0003] Some technical solutions incorporate neural network models to process electromyographic signals, but these often employ continuous numerical model structures, limiting their ability to characterize the continuously changing characteristics of muscle discharges. Furthermore, the computational process relies on continuous high-frequency updates, making it difficult to balance real-time performance with energy constraints. When combined with electrical muscle stimulation, existing methods typically employ fixed or semi-fixed stimulation parameter adjustment strategies, making it difficult to dynamically adjust the stimulation mode based on real-time muscle state changes. This can easily lead to a mismatch between the stimulation and the actual muscle state.

[0004] Furthermore, in terms of model parameter configuration, existing methods mostly rely on manual experience or single parameter search, which lacks sufficient support for the joint adjustment of model structural and kinetic parameters, making it difficult to achieve stable and consistent monitoring and control effects in complex training scenarios. Therefore, it is necessary to propose a technical solution that can adapt to the temporal characteristics of electromyographic signals, support adaptive parameter adjustment, and form a closed loop with electrical stimulation. Summary of the Invention

[0005] One objective of this invention is to propose a method for monitoring neural signals during pulsatile muscle training based on a pulsatile neural network (SNN). This invention introduces a pulsatile neural network liquid state machine and a time constant adaptive modulation mechanism to perform pulse-level analysis on electromyographic signals and dynamically generate electrical stimulation parameters, thereby constructing a continuously evolving closed-loop control process for muscle training. This method has the advantages of fine state perception, stable parameter updates, and consistent closed-loop regulation.

[0006] A method for monitoring neural signals during pulsating muscle training based on SNN according to an embodiment of the present invention includes the following steps: S1. Collect electromyographic signals generated by the target muscle under the action of exercise and electrical stimulation; S2. Convert the electromyographic signal into a time-ordered pulse sequence according to the pulse coding rules; S3. Input the pulse sequence into the reservoir layer of the liquid state machine. The reservoir layer is composed of threshold-triggered spiking neurons. Each spiking neuron is configured with a time constant parameter and outputs pulse states arranged by time index. S4. The pulse state output from the reservoir layer forms a muscle state vector, which includes a fatigue component and an activation stability component. S5. Perform numerical modulation on the time constant parameters using the muscle state vector to form a set of time constant parameters, and use the set of time constant parameters to drive the state update of the reservoir layer. S6. Construct the model configuration vector, which contains the liquid state machine structure identifier and time constant parameter set. Use the grid adaptive direct search algorithm to perform grouped search update on the model configuration vector. The grouping includes the structure identifier variable group and the continuous parameter variable group. Set the grid scale for different variable groups to complete the iterative update and output the target model configuration vector. S7. Generate a set of electrical stimulation parameters using the target model configuration vector and apply electrical stimulation, while continuing to collect electromyographic signals and return them to S1 to form a closed loop.

[0007] Optionally, S2 specifically includes: The acquired electromyographic signals are subjected to time axis expansion processing to form a continuous amplitude sequence arranged according to the sampling time. The continuous amplitude sequence is then subjected to amplitude normalization processing to obtain a normalized electromyographic amplitude sequence. A fixed-length encoding time window is set for the normalized electromyography amplitude sequence along the time axis. Within each encoding time window, a sampling point comparison operation is performed on the normalized electromyography amplitude sequence according to a preset pulse trigger threshold. When the amplitude of the sampling point reaches the pulse trigger threshold, a pulse event marker is generated at the corresponding sampling time position. The pulse event markers generated within each encoding time window are sorted according to their chronological order. The sorted pulse event markers are then concatenated with their corresponding time indices to form a pulse event sequence that is continuously arranged on the time axis. The output is a time-ordered pulse sequence.

[0008] Optionally, S4 specifically includes: The pulse states output by the reservoir layer are read in time index order, and a parsing time window is set in the time axis direction. The parsing time window and the encoding time window are aligned on the start time index and end time index. The number of pulse event markers is counted and the pulse density value of the window is calculated in each parsing time window. At the same time, the time index difference between adjacent pulse event markers is extracted to form a pulse interval sequence. Within each analytical time window, the interval mean and interval variance are calculated for the pulse interval sequence. The pulse density difference and the interval mean difference are calculated between adjacent analytical time windows. A linear combination of the pulse density difference and the interval mean difference is performed to obtain the fatigue component values. Normalization mapping is performed on the interval variance corresponding to each analytical time window to obtain the activation stability component value, and the fatigue component value and the activation stability component value are concatenated according to fixed vector positions to form a muscle state vector.

[0009] Optionally, the calculation of the window pulse density value in step S4 specifically includes: Read the start time index and end time index of the current parsing time window, and calculate the parsing time window length. The parsing time window length is equal to the end time index minus the start time index. Within the time index range covered by the current parsing time window, the pulse states output by the reservoir layer are traversed in time index order, the number of pulse event markers is counted, and the window pulse event count value is obtained. Divide the window pulse event count by the parsing time window length to obtain the pulse density value of the corresponding parsing time window, and write the pulse density value into a pulse density sequence arranged by time index.

[0010] Optionally, the numerical modulation performed in step S5 specifically includes: Read the muscle state vector in time index order, and read the fatigue component value and activation stability component value from the muscle state vector, and form state component pairs according to the fixed component position order; Multiply each pair of state components by the corresponding time constant mapping coefficient to obtain the fatigue mapping amount and the stability mapping amount, and then perform a summation operation on the fatigue mapping amount and the stability mapping amount to obtain the time constant offset; The time constant offset is summed with the initial time constant parameter of the spiking neuron to obtain the updated time constant parameter, and the updated time constant parameter is written into the parameter set of the spiking neuron.

[0011] Optionally, the reservoir layer state update in step S5 specifically includes: The updated time constant parameters of each spiking neuron in the reservoir layer of the liquid state machine are read sequentially according to the time index, and the time constant parameters of each spiking neuron are written into the corresponding neuron parameter storage location. At each time index position, the pulse event markers input to the reservoir layer are read, and the set of spiking neurons participating in the current time index state update is determined according to the internal connection relationship of the reservoir layer. For each spiking neuron participating in the state update, the current membrane potential value, the corresponding time constant parameter, and the pulse event markers from adjacent spiking neurons are read. A membrane potential decay operation is performed on each spiking neuron, using the current membrane potential value and the corresponding time constant parameter to calculate the decayed membrane potential value, and the decayed membrane potential value is written into the membrane potential buffer. After the membrane potential decay operation is completed, a pulse input accumulation operation is performed on each spiking neuron, accumulating the pulses from adjacent pulses. Neuron spiking event markers are mapped to membrane potential increments, and the membrane potential increments are summed with the decayed membrane potential values ​​to obtain updated membrane potential values. A comparison operation is performed between the updated membrane potential value and the spiking trigger threshold. When the updated membrane potential value reaches the spiking trigger threshold, a new spiking event marker is generated and written to the reservoir layer output buffer. After a spiking event marker is generated, the membrane potential value of the corresponding spiking neuron is reset to a preset reset value, and the reset membrane potential value is written to the membrane potential buffer. If no spiking event marker is generated at the current time index position, the updated membrane potential value is directly written to the membrane potential buffer and retained until the next time index. After updating the state of all spiking neurons in the reservoir layer according to the time index order, the spiking event markers in the reservoir layer output buffer are used as the input data source for the spiking state in the next parsing time window.

[0012] Optionally, the pulse input accumulation operation in step S5 specifically includes: At the current time index position, determine the set of adjacent spiking neurons connected to the target spiking neuron according to the internal connectivity of the reservoir layer, and read the pulse event markers generated by the adjacent spiking neurons at the current time index position; for each adjacent spiking neuron's pulse event marker, read the corresponding connection weight parameter, and perform a multiplication operation between the pulse event marker and the connection weight parameter to obtain the corresponding membrane potential increment component; at the current time index position, perform a term-by-term summation operation on all membrane potential increment components according to the corresponding order of the connection weight parameters to obtain the total membrane potential increment value; perform a summation operation between the total membrane potential increment value and the decayed membrane potential value to obtain the updated membrane potential value, and write the updated membrane potential value into the membrane potential buffer.

[0013] Optionally, S6 specifically includes: S61. Read the model configuration vector and split it into a structural identifier variable group and a continuous parameter variable group according to the preset variable type classification rules. The structural identifier variable group includes the reservoir layer scale identifier and connection structure identifier of the liquid state machine, and the continuous parameter variable group includes the time constant parameter set. S62. Set a first initial grid scale for the structural identifier variable group and a second initial grid scale for the continuous parameter variable group. The first initial grid scale and the second initial grid scale are numerically independent. S63. Using the current model configuration vector as the center point, generate a candidate structural configuration set for the structural identifier variable group within the range limited by the first initial grid scale, and generate a candidate parameter configuration set for the continuous parameter variable group within the range limited by the second initial grid scale. Form a candidate model configuration vector set according to the permutation and combination rules of the structural identifier variable group and the continuous parameter variable group. S64. For each candidate model configuration vector in the candidate model configuration vector set, write it sequentially into the liquid state machine configuration area and execute a complete closed-loop operation process. The closed-loop operation process includes pulse sequence input, The process involves updating the reservoir layer pulse state, resolving the muscle state vector, updating the time constant parameter, and generating the electrical stimulation parameter, while recording the evaluation value corresponding to each closed-loop operation. S65: Select the model configuration vector whose evaluation value satisfies the preset constraints from the candidate model configuration vector set as the current optimal model configuration vector, and replace the original model configuration vector with the current optimal model configuration vector. S66: If no model configuration vector whose evaluation value satisfies the preset constraints is selected at the current grid scale, perform a shrinkage operation on the first and second initial grid scales respectively to generate an updated grid scale, and return to S63 to regenerate the candidate model configuration vector set. S67: If a model configuration vector whose evaluation value satisfies the preset constraints is selected at the current grid scale, keep the first and second initial grid scales unchanged, and return to S63 to continue generating the candidate model configuration vector set. S68: Repeat steps S63 to S67 until the grid scale shrinks to the preset termination condition, output the final target model configuration vector, and write it to the model configuration buffer.

[0014] Optionally, S7 specifically includes: Read the target model configuration vector and extract the time constant parameter set from the target model configuration vector, and arrange the time constant parameter set according to the spiking neuron index order; The time constant parameter set is mapped into an electrical stimulation parameter vector that corresponds one-to-one with the stimulation channel identifier according to the preset parameter mapping order, and the numerical range pruning operation is performed on the electrical stimulation parameter vector to form an electrical stimulation parameter set. The electrical stimulation parameter set is read sequentially according to the time index and electrical stimulation is applied. At the same time, the electromyographic signals under the corresponding time index are collected and returned to step S1 to form a closed loop.

[0015] The beneficial effects of this invention are: (1) This invention converts electromyographic signals into time-ordered pulse sequences and introduces a liquid state machine reservoir layer composed of threshold-triggered pulse neurons, thereby realizing pulse-level modeling of the temporal characteristics of muscle discharge. This enables discrete neural discharge behaviors generated during muscle training to be processed in a unified pulse form, enhancing the adaptability to complex temporal changes.

[0016] (2) The present invention introduces a set of time constant parameters in the liquid state machine and performs numerical updates on the time constant parameters by the muscle state vector, so that the evolution process of the reservoir state is synchronously associated with the muscle fatigue component and the activation stability component, thereby maintaining the consistency between the reservoir dynamic parameters and the muscle state changes in continuous training or repetitive stimulation scenarios.

[0017] (3) The present invention further realizes the joint configuration adjustment of the structural parameters and dynamic parameters of the liquid state machine by constructing a model configuration vector containing a group of structural identifier variables and a group of continuous parameter variables, and by using a grid adaptive direct search algorithm to perform group search update, thereby avoiding the reliance on manual experience to set model parameters.

[0018] (4) After the model configuration vector is determined, the present invention maps the time constant parameter set to the electrical stimulation parameter set and applies electrical stimulation. At the same time, the electromyographic signal is continuously returned to form a closed-loop operation structure, so that electromyographic acquisition, pulse modeling, parameter updating and electrical stimulation execution are completed in a unified process. It is suitable for closed-loop application scenarios of muscle training and nerve signal monitoring. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0020] Figure 1 This is a flowchart of a method for monitoring neural signals in pulsating muscle training based on SNN proposed in this invention. Figure 2 This is a time constant numerical modulation diagram of a pulse muscle training neural signal monitoring method based on SNN proposed in this invention; Figure 3 This is a grouped grid adaptive direct search graph for a method of monitoring neural signals in pulsating muscle training based on SNN proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1-3 A method for monitoring neural signals during pulsating muscle training based on SNN, comprising the following steps: S1. Collect electromyographic signals generated by the target muscle under the action of exercise and electrical stimulation; S2. Convert the electromyographic signal into a time-ordered pulse sequence according to the pulse coding rules; S3. Input the pulse sequence into the reservoir layer of the liquid state machine. The reservoir layer is composed of threshold-triggered spiking neurons. Each spiking neuron is configured with a time constant parameter and outputs pulse states arranged by time index. S4. The pulse state output from the reservoir layer forms a muscle state vector, which includes a fatigue component and an activation stability component. S5. Perform numerical modulation on the time constant parameters using the muscle state vector to form a set of time constant parameters, and use the set of time constant parameters to drive the state update of the reservoir layer. S6. Construct the model configuration vector, which contains the liquid state machine structure identifier and time constant parameter set. Use the grid adaptive direct search algorithm to perform grouped search update on the model configuration vector. The grouping includes the structure identifier variable group and the continuous parameter variable group. Set the grid scale for different variable groups to complete the iterative update and output the target model configuration vector. S7. Generate a set of electrical stimulation parameters using the target model configuration vector and apply electrical stimulation, while continuing to collect electromyographic signals and return them to S1 to form a closed loop.

[0023] In this embodiment, S2 specifically includes: The acquired electromyographic signals are subjected to time axis expansion processing to form a continuous amplitude sequence arranged according to the sampling time. The continuous amplitude sequence is then subjected to amplitude normalization processing to obtain a normalized electromyographic amplitude sequence. A fixed-length encoding time window is set for the normalized electromyography amplitude sequence along the time axis. Within each encoding time window, a sampling point comparison operation is performed on the normalized electromyography amplitude sequence according to a preset pulse trigger threshold. When the amplitude of the sampling point reaches the pulse trigger threshold, a pulse event marker is generated at the corresponding sampling time position. The pulse event markers generated within each encoding time window are sorted according to their chronological order. The sorted pulse event markers are then concatenated with their corresponding time indices to form a pulse event sequence that is continuously arranged on the time axis. The output is a time-ordered pulse sequence.

[0024] In this embodiment, S3 specifically includes: In this embodiment, the pulse sequence is input to the reservoir layer of the liquid state machine in time index order. The reservoir layer consists of multiple threshold-triggered spiking neurons, which form an internal propagation structure according to a preset connection relationship. Each spiking neuron is configured with a corresponding time constant parameter to constrain the decay process of the membrane potential over time.

[0025] During the operation of the reservoir layer, pulse event markers in the input pulse sequence are read at each time index position, and the pulse event markers are distributed to the spiking neurons that have connections with them. Each spiking neuron reads its current membrane potential value, time constant parameter, and pulse event markers from neighboring spiking neurons at the corresponding time index position, and sequentially completes the membrane potential decay calculation and pulse input accumulation calculation.

[0026] After the membrane potential is updated, a comparison operation is performed between the updated membrane potential value and the pulse trigger threshold. When the updated membrane potential value reaches the pulse trigger threshold, a corresponding pulse event marker is generated and recorded at the current time index position; when the pulse trigger threshold is not reached, no pulse event marker is generated, and only the updated membrane potential value is retained for the state evolution at the next time index position.

[0027] Through the above process, the reservoir layer outputs a set of pulse event markers generated by multiple spiking neurons at each time index position. The pulse event markers at each time index position are arranged in chronological order, forming the pulse state output by the reservoir layer. The pulse state serves as the input data for muscle state vector analysis and time constant parameter update in subsequent steps.

[0028] In this embodiment, the method for monitoring pulsed muscle training neural signals based on SNN according to claim 1 is characterized in that S4 specifically includes: The pulse states output by the reservoir layer are read in time index order, and a parsing time window is set in the time axis direction. The parsing time window and the encoding time window are aligned on the start time index and end time index. The number of pulse event markers is counted and the pulse density value of the window is calculated in each parsing time window. At the same time, the time index difference between adjacent pulse event markers is extracted to form a pulse interval sequence. Within each analytical time window, the interval mean and interval variance are calculated for the pulse interval sequence. The pulse density difference and the interval mean difference are calculated between adjacent analytical time windows. A linear combination of the pulse density difference and the interval mean difference is performed to obtain the fatigue component values. Normalization mapping is performed on the interval variance corresponding to each analytical time window to obtain the activation stability component value, and the fatigue component value and the activation stability component value are concatenated according to fixed vector positions to form a muscle state vector.

[0029] In this embodiment, the calculation of the window pulse density value in step S4 specifically includes: Read the start time index and end time index of the current parsing time window, and calculate the parsing time window length. The parsing time window length is equal to the end time index minus the start time index. Within the time index range covered by the current parsing time window, the pulse states output by the reservoir layer are traversed in time index order, the number of pulse event markers is counted, and the window pulse event count value is obtained. Divide the window pulse event count by the parsing time window length to obtain the pulse density value of the corresponding parsing time window, and write the pulse density value into a pulse density sequence arranged by time index.

[0030] In this embodiment, the execution of numerical modulation in step S5 specifically includes: Read the muscle state vector in time index order, and read the fatigue component value and activation stability component value from the muscle state vector, and form state component pairs according to the fixed component position order; Multiply each pair of state components by the corresponding time constant mapping coefficient to obtain the fatigue mapping amount and the stability mapping amount, and then perform a summation operation on the fatigue mapping amount and the stability mapping amount to obtain the time constant offset; The time constant offset is summed with the initial time constant parameter of the spiking neuron to obtain the updated time constant parameter, and the updated time constant parameter is written into the parameter set of the spiking neuron.

[0031] In this embodiment, the reservoir layer state update in step S5 specifically includes: The updated time constant parameters of each spiking neuron in the reservoir layer of the liquid state machine are read sequentially according to the time index, and the time constant parameters of each spiking neuron are written into the corresponding neuron parameter storage location. At each time index position, the pulse event markers input to the reservoir layer are read, and the set of spiking neurons participating in the current time index state update is determined according to the internal connection relationship of the reservoir layer. For each spiking neuron participating in the state update, the current membrane potential value, the corresponding time constant parameter, and the pulse event markers from adjacent spiking neurons are read. A membrane potential decay operation is performed on each spiking neuron, using the current membrane potential value and the corresponding time constant parameter to calculate the decayed membrane potential value, and the decayed membrane potential value is written into the membrane potential buffer. After the membrane potential decay operation is completed, a pulse input accumulation operation is performed on each spiking neuron, accumulating the pulses from adjacent pulses. Neuron spiking event markers are mapped to membrane potential increments, and the membrane potential increments are summed with the decayed membrane potential values ​​to obtain updated membrane potential values. A comparison operation is performed between the updated membrane potential value and the spiking trigger threshold. When the updated membrane potential value reaches the spiking trigger threshold, a new spiking event marker is generated and written to the reservoir layer output buffer. After a spiking event marker is generated, the membrane potential value of the corresponding spiking neuron is reset to a preset reset value, and the reset membrane potential value is written to the membrane potential buffer. If no spiking event marker is generated at the current time index position, the updated membrane potential value is directly written to the membrane potential buffer and retained until the next time index. After updating the state of all spiking neurons in the reservoir layer according to the time index order, the spiking event markers in the reservoir layer output buffer are used as the input data source for the spiking state in the next parsing time window.

[0032] In this embodiment, the pulse input accumulation operation in step S5 specifically includes: At the current time index position, determine the set of adjacent spiking neurons connected to the target spiking neuron according to the internal connectivity of the reservoir layer, and read the pulse event markers generated by the adjacent spiking neurons at the current time index position; for each adjacent spiking neuron's pulse event marker, read the corresponding connection weight parameter, and perform a multiplication operation between the pulse event marker and the connection weight parameter to obtain the corresponding membrane potential increment component; at the current time index position, perform a term-by-term summation operation on all membrane potential increment components according to the corresponding order of the connection weight parameters to obtain the total membrane potential increment value; perform a summation operation between the total membrane potential increment value and the decayed membrane potential value to obtain the updated membrane potential value, and write the updated membrane potential value into the membrane potential buffer.

[0033] In this embodiment, S6 specifically includes: S61. Read the model configuration vector and split it into a structural identifier variable group and a continuous parameter variable group according to the preset variable type division rules. The structural identifier variable group includes the liquid state machine reservoir layer scale identifier and the connection structure identifier, and the continuous parameter variable group includes the time constant parameter set. S62. Set a first initial grid scale for the structural identifier variable group and a second initial grid scale for the continuous parameter variable group. The first initial grid scale and the second initial grid scale are numerically independent. The grid scale is used to limit the range of variable values ​​in each search iteration. S63. Using the current model configuration vector as the center point, generate a candidate structure configuration set for the structure identifier variable group within the first initial grid scale limit, and generate a candidate parameter configuration set for the continuous parameter variable group within the second initial grid scale limit. Form a candidate model configuration vector set according to the permutation and combination rules of the structure identifier variable group and the continuous parameter variable group. S64. For each candidate model configuration vector in the candidate model configuration vector set, sequentially write it into the liquid state machine configuration area and execute a complete closed-loop operation process. The closed-loop operation process includes pulse sequence input, reservoir layer pulse state update, muscle state vector resolution, time constant parameter update, and electrical stimulation parameter generation. Record the evaluation value corresponding to each closed-loop operation. S65. Select evaluation values ​​from the candidate model configuration vector set that satisfy preset constraints. The model configuration vector of the component is used as the current optimal model configuration vector, and the current optimal model configuration vector replaces the original model configuration vector; S66, when no model configuration vector with evaluation value that meets the preset constraint conditions is selected under the current grid scale, the first initial grid scale and the second initial grid scale are respectively shrunk to generate the updated grid scale, and return to S63 to regenerate the candidate model configuration vector set; S67, when a model configuration vector with evaluation value that meets the preset constraint conditions is selected under the current grid scale, the first initial grid scale and the second initial grid scale are kept unchanged, and return to S63 to continue to generate the candidate model configuration vector set; S68, repeat S63 to S67 until the grid scale is shrunk to the preset termination condition, output the final target model configuration vector and write it to the model configuration cache area.

[0034] In this embodiment, the division rules in step S61 specifically include: The variables in the model configuration vector are first distinguished according to whether they participate in the discrete structure definition of the state evolution structure. Variables used to describe the topological scale, connection relationship or structural form of the reservoir layer of the liquid state machine are uniformly classified as structural identifier variables; variables used to participate in the numerical calculation of the state evolution of spiking neurons and take values ​​in the continuous numerical domain are uniformly classified as continuous parameter variables.

[0035] Under this classification rule, the structural identifier variable of the reservoir layer in the liquid state machine is used to characterize the static structural properties of the reservoir layer. This type of variable maintains a constant value within one state evolution cycle, and its value changes are reflected through structural switching. The time constant parameter set is used to characterize the numerical characteristics of the membrane potential of spiking neurons changing over time. This type of variable changes in a continuous numerical domain, and its value changes are reflected through numerical update.

[0036] When reading the model configuration vector, the variables are parsed according to their predefined position order in the configuration vector, and the parsed variables are written into the structure identifier variable group and the continuous parameter variable group based on their function attributes. Variables occupying fixed positions and used to determine the reservoir layer structure are written into the structure identifier variable group, while variables occupying positions in the continuous parameter interval and used to participate in membrane potential decay calculations and state update calculations are written into the continuous parameter variable group.

[0037] In this embodiment, S7 specifically includes: Read the target model configuration vector and extract the time constant parameter set from the target model configuration vector, and arrange the time constant parameter set according to the spiking neuron index order; The time constant parameter set is mapped into an electrical stimulation parameter vector that corresponds one-to-one with the stimulation channel identifier according to the preset parameter mapping order, and the numerical range pruning operation is performed on the electrical stimulation parameter vector to form an electrical stimulation parameter set. The electrical stimulation parameter set is read sequentially according to the time index and electrical stimulation is applied. At the same time, the electromyographic signals under the corresponding time index are collected and returned to step S1 to form a closed loop.

[0038] Example 1: To verify the feasibility and stability of this invention in practical applications, it was applied to a scenario targeting muscle training and state monitoring. In this scenario, subjects continuously acquired electromyographic (EMG) signals generated by the target muscles under exercise and electrical stimulation while performing repetitive muscle contractions. These signals exhibit significant temporal variations during training, and the muscle activation state and fatigue level dynamically change as training progresses. Traditional methods based on fixed parameters or single statistical features struggle to maintain stable monitoring and regulation effects. In this application scenario, this invention first performs pulse coding processing on the continuously acquired EMG signals, converting the original continuous signals into a time-ordered pulse sequence. The pulse sequence is input into a reservoir layer of a liquid state machine composed of threshold-triggered spiking neurons. Within the reservoir layer, the input pulses propagate and evolve through the connections between spiking neurons, outputting a pulse state reflecting the current muscle discharge state. Further analysis of the pulse state forms a muscle state vector containing fatigue and activation stability components, used to characterize the changing trends of muscle state during training. During continuous training, the time constant parameters of spiking neurons in the reservoir layer are updated based on the muscle state vector to ensure that the state evolution of the reservoir layer is consistent with the current muscle state. Simultaneously, a state evolution structure configuration vector containing the liquid state machine structure identifier and time constant parameter set is constructed. A grid-adaptive direct search method is used to group and update the structure identifier variable group and the continuous parameter variable group, thereby obtaining stable configuration results without relying on human experience. Based on the final target configuration vector, a corresponding set of electrical stimulation parameters is generated and electrical stimulation is applied, while new electromyographic signals are continuously collected to form a closed-loop operation. During this implementation, several key indicators are continuously recorded, including changes in pulse density, mean pulse interval, fatigue component values, activation stability component values, and changes in electrical stimulation parameters. Comparative analysis shows that in the early stages of training, pulse density and activation stability are relatively concentrated; as training progresses, the fatigue component gradually increases, and the time constant parameters and electrical stimulation parameters are adjusted accordingly to maintain continuous changes in the pulse state distribution without significant abrupt changes. Compared to a control method that does not employ the closed-loop structure of this invention, this invention exhibits a more stable parameter change trajectory and more consistent electromyographic signal response characteristics during long-term training. The table below presents the statistical results of key indicators at different stages during continuous training, visually demonstrating the operational status and performance of this invention during implementation.

[0039] Table 1: Statistical Table of Changes in Key Indicators During Muscle Training

[0040] The changes in the key indicators listed in the table show that, during continuous training, the values ​​exhibit a stable and interpretable evolutionary trend as the training phase progresses. The pulse density values ​​change gradually between different stages without sudden jumps, indicating that the pulse-encoded electromyographic signals maintain continuous temporal expression characteristics within the reservoir layer, and the distribution of pulse events along the time axis is well-traceable. This process is consistent with the actual situation of the gradual adjustment of muscle firing rhythm during training.

[0041] The mean pulse interval shows a gradual changing trend in each stage, reflecting the adjustment of the distribution of adjacent pulse events in the time dimension as training progresses. This change is consistent with the change direction of the pulse density value, indicating that the pulse state output by the reservoir layer has good inherent consistency between different resolution stages, and there is no contradiction between the density change and the interval change.

[0042] The fatigue component values ​​showed a continuous upward trend across all stages, with stable amplitudes between adjacent stages and no abnormal spikes or drops. This indicates that the fatigue component obtained from impulse state analysis can consistently reflect the cumulative changes in muscle state during training. Meanwhile, the activation stability component remained within a relatively concentrated range, without drastic fluctuations due to changes in the fatigue component. This demonstrates that even with gradual fatigue accumulation, the activation rhythm analysis results remain relatively stable.

[0043] The mean values ​​of the time constant and electrical stimulation parameters are adjusted synchronously with changes in muscle state at each stage, and the changes are smooth and continuous without obvious discontinuities. This indicates that after updating the time constant parameters numerically using the muscle state vector, the evolution of the reservoir state can remain coordinated with the training state, while the generated electrical stimulation parameters maintain a controllable range of change at the numerical level, which is beneficial for the continuous operation of the closed loop.

[0044] As can be seen from the data in the table, during the implementation of this invention, the evolutionary relationship between the key indicators remains consistent, the parameter update process is continuous, and the state analysis results are stable. This enables the monitoring of electromyographic signals and the closed-loop operation of electrical stimulation under long-term training conditions, demonstrating good feasibility and operational consistency.

[0045] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring neural signals during pulsating muscle training based on SNN, characterized in that, Includes the following steps: S1. Collect electromyographic signals generated by the target muscle under the action of exercise and electrical stimulation; S2. Convert the electromyographic signal into a time-ordered pulse sequence according to the pulse coding rules; S3. Input the pulse sequence into the reservoir layer of the liquid state machine. The reservoir layer is composed of threshold-triggered spiking neurons. Each spiking neuron is configured with a time constant parameter and outputs pulse states arranged by time index. S4. The pulse state output from the reservoir layer forms a muscle state vector, which includes a fatigue component and an activation stability component. S5. Perform numerical modulation on the time constant parameters using the muscle state vector to form a set of time constant parameters, and use the set of time constant parameters to drive the state update of the reservoir layer. S6. Construct the model configuration vector, which contains the liquid state machine structure identifier and time constant parameter set. Use the grid adaptive direct search algorithm to perform grouped search update on the model configuration vector. The grouping includes the structure identifier variable group and the continuous parameter variable group. Set the grid scale for different variable groups to complete the iterative update and output the target model configuration vector. S7. Generate a set of electrical stimulation parameters using the target model configuration vector and apply electrical stimulation, while continuing to collect electromyographic signals and return them to S1 to form a closed loop.

2. The method for monitoring nerve signals in pulsed muscle training based on SNN according to claim 1, characterized in that, S2 specifically includes: The acquired electromyographic signals are subjected to time axis expansion processing to form a continuous amplitude sequence arranged according to the sampling time. The continuous amplitude sequence is then subjected to amplitude normalization processing to obtain a normalized electromyographic amplitude sequence. A fixed-length encoding time window is set for the normalized electromyography amplitude sequence along the time axis. Within each encoding time window, a sampling point comparison operation is performed on the normalized electromyography amplitude sequence according to a preset pulse trigger threshold. When the amplitude of the sampling point reaches the pulse trigger threshold, a pulse event marker is generated at the corresponding sampling time position. The pulse event markers generated within each encoding time window are sorted according to their chronological order. The sorted pulse event markers are then concatenated with their corresponding time indices to form a pulse event sequence that is continuously arranged on the time axis. The output is a time-ordered pulse sequence.

3. The method for monitoring nerve signals in pulsed muscle training based on SNN according to claim 2, characterized in that, S4 specifically includes: The pulse states output by the reservoir layer are read in time index order, and a parsing time window is set in the time axis direction. The parsing time window and the encoding time window are aligned on the start time index and end time index. The number of pulse event markers is counted and the pulse density value of the window is calculated in each parsing time window. At the same time, the time index difference between adjacent pulse event markers is extracted to form a pulse interval sequence. Within each analytical time window, the interval mean and interval variance are calculated for the pulse interval sequence. The pulse density difference and the interval mean difference are calculated between adjacent analytical time windows. A linear combination of the pulse density difference and the interval mean difference is performed to obtain the fatigue component values. Normalization mapping is performed on the interval variance corresponding to each analytical time window to obtain the activation stability component value, and the fatigue component value and the activation stability component value are concatenated according to fixed vector positions to form a muscle state vector.

4. The method for monitoring neural signals during pulsed muscle training based on SNN according to claim 3, characterized in that, The calculation of the window pulse density value in step S4 specifically includes: Read the start time index and end time index of the current parsing time window, and calculate the parsing time window length. The parsing time window length is equal to the end time index minus the start time index. Within the time index range covered by the current parsing time window, the pulse states output by the reservoir layer are traversed in time index order, the number of pulse event markers is counted, and the window pulse event count value is obtained. Divide the window pulse event count by the parsing time window length to obtain the pulse density value of the corresponding parsing time window, and write the pulse density value into a pulse density sequence arranged by time index.

5. The method for monitoring neural signals during pulsating muscle training based on SNN according to claim 4, characterized in that, The execution of numerical modulation in step S5 specifically includes: Read the muscle state vector in time index order, and read the fatigue component value and activation stability component value from the muscle state vector, and form state component pairs according to the fixed component position order; Multiply each pair of state components by the corresponding time constant mapping coefficient to obtain the fatigue mapping amount and the stability mapping amount, and then perform a summation operation on the fatigue mapping amount and the stability mapping amount to obtain the time constant offset; The time constant offset is summed with the initial time constant parameter of the spiking neuron to obtain the updated time constant parameter, and the updated time constant parameter is written into the parameter set of the spiking neuron.

6. The method for monitoring neural signals during pulsating muscle training based on SNN according to claim 5, characterized in that, The reservoir layer state update in step S5 specifically includes: The updated time constant parameters of each spiking neuron in the reservoir layer of the liquid state machine are read sequentially according to the time index, and the time constant parameters of each spiking neuron are written into the corresponding neuron parameter storage location. At each time index position, the pulse event markers input to the reservoir layer are read, and the set of spiking neurons participating in the current time index state update is determined according to the internal connection relationship of the reservoir layer. For each spiking neuron participating in the state update, the current membrane potential value, the corresponding time constant parameter, and the pulse event markers from adjacent spiking neurons are read. A membrane potential decay operation is performed on each spiking neuron, using the current membrane potential value and the corresponding time constant parameter to calculate the decayed membrane potential value, and the decayed membrane potential value is written into the membrane potential buffer. After the membrane potential decay operation is completed, a pulse input accumulation operation is performed on each spiking neuron, accumulating the pulses from adjacent pulses. Neuron spiking event markers are mapped to membrane potential increments, and the membrane potential increments are summed with the decayed membrane potential values ​​to obtain updated membrane potential values. A comparison operation is performed between the updated membrane potential value and the spiking trigger threshold. When the updated membrane potential value reaches the spiking trigger threshold, a new spiking event marker is generated and written to the reservoir layer output buffer. After a spiking event marker is generated, the membrane potential value of the corresponding spiking neuron is reset to a preset reset value, and the reset membrane potential value is written to the membrane potential buffer. If no spiking event marker is generated at the current time index position, the updated membrane potential value is directly written to the membrane potential buffer and retained until the next time index. After updating the state of all spiking neurons in the reservoir layer according to the time index order, the spiking event markers in the reservoir layer output buffer are used as the input data source for the spiking state in the next parsing time window.

7. The method for monitoring neural signals during pulsating muscle training based on SNN according to claim 6, characterized in that, The pulse input accumulation operation specifically includes: At the current time index position, determine the set of adjacent spiking neurons connected to the target spiking neuron according to the internal connectivity of the reservoir layer, and read the pulse event markers generated by the adjacent spiking neurons at the current time index position; for each adjacent spiking neuron's pulse event marker, read the corresponding connection weight parameter, and perform a multiplication operation between the pulse event marker and the connection weight parameter to obtain the corresponding membrane potential increment component; at the current time index position, perform a term-by-term summation operation on all membrane potential increment components according to the corresponding order of the connection weight parameters to obtain the total membrane potential increment value; perform a summation operation between the total membrane potential increment value and the decayed membrane potential value to obtain the updated membrane potential value, and write the updated membrane potential value into the membrane potential buffer.

8. The method for monitoring neural signals during pulsating muscle training based on SNN according to claim 7, characterized in that, S6 specifically includes: S61. Read the model configuration vector and split it into a structural identifier variable group and a continuous parameter variable group according to the preset variable type classification rules. The structural identifier variable group includes the reservoir layer scale identifier and connection structure identifier of the liquid state machine, and the continuous parameter variable group includes the time constant parameter set. S62. Set a first initial grid scale for the structural identifier variable group and a second initial grid scale for the continuous parameter variable group. The first initial grid scale and the second initial grid scale are numerically independent. S63. Using the current model configuration vector as the center point, generate a candidate structural configuration set for the structural identifier variable group within the range limited by the first initial grid scale, and generate a candidate parameter configuration set for the continuous parameter variable group within the range limited by the second initial grid scale. Form a candidate model configuration vector set according to the permutation and combination rules of the structural identifier variable group and the continuous parameter variable group. S64. For each candidate model configuration vector in the candidate model configuration vector set, write it sequentially into the liquid state machine configuration area and execute a complete closed-loop operation process. The closed-loop operation process includes pulse sequence input, The process involves updating the reservoir layer pulse state, resolving the muscle state vector, updating the time constant parameter, and generating the electrical stimulation parameter, while recording the evaluation value corresponding to each closed-loop operation. S65: Select the model configuration vector whose evaluation value satisfies the preset constraints from the candidate model configuration vector set as the current optimal model configuration vector, and replace the original model configuration vector with the current optimal model configuration vector. S66: If no model configuration vector whose evaluation value satisfies the preset constraints is selected at the current grid scale, perform a shrinkage operation on the first and second initial grid scales respectively to generate an updated grid scale, and return to S63 to regenerate the candidate model configuration vector set. S67: If a model configuration vector whose evaluation value satisfies the preset constraints is selected at the current grid scale, keep the first and second initial grid scales unchanged, and return to S63 to continue generating the candidate model configuration vector set. S68: Repeat steps S63 to S67 until the grid scale shrinks to the preset termination condition, output the final target model configuration vector, and write it to the model configuration buffer.

9. The method for monitoring neural signals during pulsating muscle training based on SNN according to claim 8, characterized in that, Specifically, S7 includes: Read the target model configuration vector and extract the time constant parameter set from the target model configuration vector, and arrange the time constant parameter set according to the spiking neuron index order; The time constant parameter set is mapped to an electrical stimulation parameter vector corresponding one-to-one with the stimulation channel identifier according to the preset parameter mapping order, and the numerical range pruning operation is performed on the electrical stimulation parameter vector to form an electrical stimulation parameter set. The electrical stimulation parameter set is read sequentially according to the time index and electrical stimulation is applied. At the same time, the electromyographic signals under the corresponding time index are collected and returned to step S1 to form a closed loop.

Images