Upper limb robot intention recognition method and system based on electromyographic signals

Abstract

The invention relates to the technical field of robot control, and discloses an upper limb robot intention recognition method and system based on electromyographic signals. The method comprises the following steps: acquiring an upper limb electromyographic signal data stream and carrying out noise suppression processing to obtain an electromyographic signal data stream; extracting frequency component characteristics representing the motion intention of the user through time-frequency analysis, and determining a signal component set; time-frequency features and interference suppression degree information are extracted, and a stable component feature combination reflecting a user motion intention mode is generated; on the basis, a multi-frequency superposition characteristic matrix is constructed, and system response delay fluctuation is monitored in real time; dynamically adjusting the logic and time sequence of the instruction queue according to the delay fluctuation analysis result, and generating a corrected control instruction sequence; and carrying out smooth processing, dynamic adjustment, time sequence fine adjustment and synchronous verification on the sequence to generate a final robot control signal matched with the action intention of the user. The control reliability, accuracy and real-time performance of the electromyographic signals in a complex environment are effectively improved.

Description

Technical Field

[0001] This application relates to the field of robot control technology, and in particular to a method and system for upper limb robot intention recognition based on electromyographic signals. Background Technology

[0002] In the field of modern human-computer interaction and rehabilitation engineering, upper limb robots (such as rehabilitation exoskeletons and intelligent prostheses) controlled by surface electromyography (EMG) signals have become a key technology for achieving natural and active rehabilitation training and functional compensation. This technology decodes the user's upper limb movement intentions and drives the robot to complete the corresponding actions, thereby establishing a direct human-machine control pathway.

[0003] However, existing control methods based on electromyography (EMG) signals still face a series of severe challenges in practical applications, resulting in poor control performance. First, the acquired raw EMG signals are extremely weak and easily contaminated by power frequency interference, electrode movement noise, and other physiological electrical signals (such as ECG). These noises overlap with the target signal frequency band, and traditional filtering methods often distort useful EMG features or leave a large amount of residual noise while denoising, making subsequent intention feature extraction unreliable. Second, existing methods do not adequately consider the non-stationary characteristics of EMG signals and individual user differences, resulting in limited stability and discriminative power in feature extraction. This makes it difficult to continuously and reliably represent the user's true movement intention in complex dynamic environments (such as when muscle fatigue changes). This directly leads to control command drift, false triggering, and even erroneous execution. Furthermore, throughout the entire process from signal acquisition to command generation, the system response delay fluctuates significantly and lacks an effective dynamic compensation mechanism, making it impossible to precisely synchronize the robot's movement with the user's intention in time, resulting in poor motion smoothness and a poor user experience.

[0004] Therefore, the core problem that existing technologies urgently need to solve is: how to achieve stable, accurate, and low-latency recognition and translation of the user's upper limb movement intentions in complex environments with strong noise and dynamic changes, and generate synchronous and smooth robot control signals. Summary of the Invention

[0005] This application provides a method and system for upper limb robot intention recognition based on electromyography (EMG) signals, aiming to solve the following technical problems: how to overcome the problems of inaccurate control commands, poor real-time performance, and low robustness caused by noise interference, unstable feature extraction, and system delay fluctuations in existing EMG signal control methods, thereby providing a method that can stably, accurately, and with low latency recognize user movement intentions and generate synchronous and smooth control signals.

[0006] In a first aspect, this application provides a method for upper limb robot intention recognition based on electromyographic signals, the method comprising: Step 1: Acquire upper limb electromyography (EMG) signal data stream and perform preliminary noise suppression processing to obtain a preliminary clean EMG signal data stream; Step 2: Perform time-frequency analysis on the electromyography signal data stream, extract frequency component features that characterize the user's movement intention, and determine the signal component set; Step 3: Extract time-frequency features and interference suppression information based on the signal component set to obtain the component feature combination that reflects the user's motion intention pattern; Step 4: Construct a multi-frequency superimposed feature matrix describing the user's upper limb movement intention based on the component feature combination, and monitor the system's response delay fluctuation in real time to obtain a delay fluctuation analysis report; Step 5: Analyze the delay fluctuation analysis report, dynamically adjust the logic and timing of the instruction queue based on the analysis results, and generate a corrected control instruction sequence; Step 6: Smooth and dynamically adjust the control instruction sequence to generate an optimized instruction execution signal. Step 7: Fine-tune the timing and verify the synchronization of the instruction execution signal to generate upper limb robot control signals that match the user's action intentions.

[0007] Secondly, this application provides an upper limb robot intention recognition system based on electromyographic signals, the system comprising: The data acquisition module is used to acquire upper limb electromyographic signal data streams and perform preliminary noise suppression processing to obtain a preliminary clean electromyographic signal data stream. The set determination module is used to perform time-frequency analysis on the electromyographic signal data stream, extract frequency component features that characterize the user's movement intention, and determine the set of signal components; The feature extraction module is used to extract time-frequency features and interference suppression information based on the signal component set, and obtain the component feature combination that reflects the user's motion intention pattern; The matrix construction module is used to construct a multi-frequency superimposed feature matrix describing the user's upper limb movement intention based on the combination of component features, and to monitor the system's response delay fluctuations in real time and obtain a delay fluctuation analysis report. The instruction correction module is used to analyze the delay fluctuation analysis report, dynamically adjust the logic and timing of the instruction queue based on the analysis results, and generate a corrected control instruction sequence. The smooth transition module is used to smooth and dynamically adjust the control instruction sequence to generate transition-optimized instruction execution signals. The signal generation module is used to fine-tune the timing and verify the synchronization of the instruction execution signal, and generate upper limb robot control signals that match the user's action intentions.

[0008] Compared with the prior art, the beneficial effects of the technical solution of this application are at least as follows: 1. By combining bandpass filtering, frequency spacing uniformity analysis, and notch filter noise suppression technology, harmonic interference and noise in electromyography signals can be accurately separated and filtered out. Combined with signal integrity verification and interpolation compensation, the purity and reliability of signal processing are improved, ensuring high quality and high integrity of the source signal used for intent recognition.

[0009] 2. By constructing multiple frequency superposition patterns through time-frequency decomposition and hierarchical clustering, and combining interference suppression analysis with secondary filtering, this method can stably extract the essential feature combination that best represents the user's intent from complex signals. This method is highly adaptable to signal non-stationarity and individual differences, effectively avoiding feature drift and misidentification, and enhancing the accuracy and stability of motion intent recognition.

[0010] 3. By introducing a delay detection time window to monitor and process delay fluctuations in real time, and dynamically adjusting the timing of the instruction queue based on the fluctuation characteristics, proactive compensation for system delay is achieved. This ensures that the final generated control commands are highly synchronized with the user's original motion intentions in time, guaranteeing the real-time and synchronous nature of the system response and significantly reducing the lag in human-computer interaction.

[0011] 4. By performing smooth curve fitting, boundary processing, and dynamic adjustment of gain parameters based on frequency feedback on the control command sequence, abrupt changes and jitters between commands are effectively eliminated, generating a smooth and continuous command execution signal with natural transitions. This optimizes the smoothness and continuity of robot motion execution, thereby driving the robot to complete smooth and human-like movements. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of the upper limb robot intention recognition method based on electromyography signals according to this application; Figure 2 This is a schematic diagram of the upper limb robot intention recognition system based on electromyography signals according to this application. Detailed Implementation

[0014] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0015] For ease of understanding, the specific process of the embodiments of this application is described below. Figure 1 The diagram shows a flowchart of the upper limb robot intention recognition method based on electromyography signals provided by the present invention. The flowchart specifically includes the following steps: Step 1: Acquire upper limb electromyography (EMG) signal data stream and perform preliminary noise suppression processing to obtain a preliminary clean EMG signal data stream.

[0016] In one specific embodiment, the process of performing step 1 may specifically include the following steps: Acquire upper limb electromyography signal data streams through real-time acquisition devices; The upper limb electromyography signal data stream was noise suppressed by bandpass filtering technology, and harmonic interference frequencies were separated by frequency interval uniformity analysis. Then, the frequencies were filtered out by notch filter to generate an intermediate filtered data stream. The signal integrity of the intermediate filtered data stream is checked. If data loss or abnormal fluctuations are detected, interpolation compensation technology is used for repair. Based on the repaired data stream, the continuity and stability of the signal are evaluated, and a preliminary clean electromyographic signal data stream is generated based on the evaluation results.

[0017] Specifically, in practical applications of upper limb robot intention recognition based on electromyography (EMG) signals, such as rehabilitation robots assisting stroke patients in upper limb function recovery and industrial robotic arms achieving human-machine collaborative operation through EMG signals, upper limb EMG signals are susceptible to interference from external electromagnetic waves and signal fluctuations caused by muscle fatigue. These interferences can lead to signal distortion, making it impossible for the robot to accurately recognize the user's intentions or even causing misoperations. Therefore, it is necessary to process the acquired raw signals to eliminate external noise and interference and obtain high-quality EMG signal data streams, enabling the upper limb robot to carry out subsequent intention recognition work based on more reliable signal data.

[0018] An electrode array attached to the surface of target muscle groups in the upper limb (such as the biceps and triceps brachii) continuously acquires time-domain voltage signals at a preset sampling rate, forming a raw electromyographic (EMG) signal data stream. This data stream typically contains effective EMG components with amplitudes ranging from 50 microvolts to 5 millivolts, but it is also mixed with various interferences: baseline drift manifests as a slow-varying component with frequencies below 20 Hz, while power frequency interference and its harmonics in the environment are concentrated at discrete frequency points such as 50 Hz and 100 Hz. In addition, there is high-frequency transient noise caused by poor electrode contact or limb movement.

[0019] To address this complex noise characteristic, a bandpass filter was used to filter the original data stream in the frequency domain. The passband range was set from 20 Hz to 500 Hz, covering the main energy distribution range of the surface electromyography (EMG) signal. This filter effectively suppressed baseline drift and most high-frequency noise, resulting in a preprocessed data stream. However, power frequency interference and its integer harmonics, overlapping with the EMG signal frequency band, still remained in the preprocessed data stream. To eliminate these specific frequency interferences, a frequency spacing uniformity analysis method was used to detect the spectrum of the preprocessed data stream: a fast Fourier transform was performed on the preprocessed data stream to obtain its frequency domain representation, identifying equally spaced peaks with a basic interval of 50 Hz. These peaks corresponded to the harmonic components of the power frequency interference. For each identified interference frequency, a notch filter with a suitable quality factor was designed and applied to form an extremely narrow stopband at that frequency, thereby attenuating these discrete interference frequency components from the signal and generating an intermediate filtered data stream.

[0020] The intermediate filtered data stream may still contain data anomalies in the time domain due to momentary failures in the acquisition system or violent limb movements. To detect these anomalies, signal integrity verification is performed on the intermediate filtered data stream: the numerical difference between adjacent sampling points is calculated, and when the difference exceeds a threshold set based on the overall root mean square value of the signal, it is determined to be an abnormal fluctuation; simultaneously, the continuity of the sampling time series is checked, and when the time interval between adjacent sampling points exceeds 1.5 times the theoretical sampling period, it is determined to be data loss. For the identified data loss segments, a linear interpolation algorithm is used for repair, and the estimated value of the intermediate point is calculated according to the time ratio based on the values ​​of the valid data points before and after the lost segment. For abnormal fluctuation regions, a cubic spline interpolation algorithm is used, and a smooth curve is fitted based on the normal data segments before and after the fluctuation region to replace the abnormal data points, forming a repaired data stream.

[0021] The repaired data stream needs to undergo quality assessment to be confirmed as a preliminary clean electromyographic signal data stream suitable for subsequent intent recognition. The assessment metrics include two aspects: continuity and stability. Continuity is evaluated by calculating the autocorrelation function of the repaired data stream. When the area of ​​the integral of the autocorrelation function before the first zero-crossing point after zero delay exceeds a set threshold (e.g., 0.9), the signal continuity is considered satisfactory. Stability is evaluated by calculating the variance of the repaired data stream within a time window. A variance below a set threshold (e.g., 0.1 mV) is considered acceptable. 2 This indicates that the signal amplitude fluctuation is within an acceptable range. Only data streams that simultaneously meet the requirements of continuity and stability will be recognized as initially pure electromyographic signal data streams and enter the subsequent processing flow. If the quality assessment fails, iterative repairs will be performed until the quality assessment passes.

[0022] This comprehensive noise suppression and signal restoration mechanism effectively solves the inherent noise interference and data incompleteness problems in the electromyography signal acquisition process, laying a reliable foundation for subsequent intention feature extraction.

[0023] Step 2: Perform time-frequency analysis on the electromyographic signal data stream, extract the frequency component features that characterize the user's movement intention, and determine the set of signal components.

[0024] In one specific embodiment, the process of performing step 2 may specifically include the following steps: Time-frequency decomposition technology was used to decompose the electromyographic signal data stream and extract frequency component features; The peak energy distribution range of the frequency components is analyzed, and based on the peak energy distribution range, a hierarchical clustering algorithm is used to determine the signal superposition level. Based on the signal superposition hierarchy, each signal superposition hierarchy is mapped to a frequency subset and assigned a weight based on the energy ratio to construct a multi-frequency superposition mode. The multi-frequency superposition pattern is convolved with the electromyographic signal data stream, and a threshold is applied to separate the initial signal component set. The initial signal component set is subjected to frequency domain feature verification to determine whether there is frequency component overlap. If so, the initial signal component set is optimized and adjusted using spectrum separation technology to obtain the final signal component set.

[0025] Specifically, in upper limb robot intention recognition applications based on electromyography (EMG) signals, such as rehabilitation robot-assisted upper limb function training and human-machine collaboration in industrial robotic arms, there is a technical problem that the frequency components of EMG signals are mixed due to non-stationarity and multi-muscle coordination, making it impossible to accurately extract features associated with movement intentions. This leads to deviations in the robot's recognition of the user's action intentions. Therefore, it is necessary to extract effective features from the initially pure EMG signal data stream through time-frequency analysis and multi-stage data processing.

[0026] Time-frequency decomposition employs the short-time Fourier transform method, dividing the electromyography (EMG) signal data stream into time windows of 256 sampling points, with adjacent windows overlapping by 128 sampling points, resulting in a window overlap rate of 50%. By multiplying the voltage data within each time window by a Hanning window function and then performing a Fourier transform, the time-domain voltage signal is converted into a two-dimensional time-frequency data matrix. Rows of the matrix correspond to time windows, columns to frequency points, and element values ​​represent the signal amplitude at the corresponding time and frequency. The amplitude values ​​of each frequency point in different time windows are extracted from this matrix, forming frequency component feature data. This data covers the energy distribution of each frequency within the typical frequency range of 20–500 Hz EMG signals, establishing a one-to-one time-frequency mapping relationship with the initially purified EMG signal data stream.

[0027] Peak energy analysis is performed based on the frequency component feature set. The average energy E(f) = ∑|TFR(t,f)|² at each frequency point in the time dimension is calculated, where TFR(t,f) represents the signal amplitude at time t and frequency dimension f. Continuous frequency regions with energies exceeding a preset multiple (e.g., 1.5 times) of the global energy mean are identified; these regions constitute the peak energy distribution range. In typical electromyographic signals, these ranges may appear in multiple frequency bands such as 20–50 Hz and 80–150 Hz. Hierarchical clustering algorithms are used to analyze these energy peak regions, calculating the Euclidean distance of each region on the frequency axis. Regions with close distances are merged into clusters using a bottom-up aggregation method, with each cluster corresponding to a signal superposition level. For example, through peak energy analysis, three significant peak energy distribution ranges were identified from the frequency domain characteristics of the electromyographic signal. The corresponding peak energy distribution range data are P1: [22, 28, 25.0, 105], P2: [45, 52, 48.5, 180], and P3: [95, 108, 101.5, 195]. The meaning of each parameter in the data sequence is based on the starting frequency, ending frequency, range center frequency, and range maximum energy. Feature vectors are constructed based on the range center frequency and range maximum energy of the sequence, namely V_1=[25.0, 105], V_2=[48.5, 180], and V_3=[101.5, 195]. Agglomerative hierarchical clustering begins, with each peak range initially forming its own cluster: {C1: V1}, {C2: V2}, {C3: V3}. Subsequently, Euclidean distance is used to calculate the distance between every two clusters, forming an initial distance matrix. A merging threshold is set (e.g., 75.0). In the distance matrix, the distance between C2 and C3, D23 = 71.1, is the minimum and less than the threshold. Therefore, C2 and C3 are merged into a new cluster C4. The eigenvector of the new cluster C4 is obtained by calculating the mean of all its members (V2, V3): V_4 = [75.0, 187.5]. At this point, the cluster state is updated to: {C1: V1}, {C4: V4}. The distance between the remaining clusters C1 and C4 is recalculated: D14 ≈ 96.5, which is greater than the merging threshold of 75.0. Therefore, merging is no longer performed, and the clustering process terminates. Finally, two independent clusters are output, thus determining the signal superposition level to be 2 levels.

[0028] The determined signal superposition levels are mapped to specific frequency subsets, with each level containing a set of frequency units. A weighting coefficient w(f) = E(f) / ∑E(f) is calculated based on the energy proportion of each frequency unit within its level, where the summation iterates through all frequency units in that level. All weighting coefficients constitute a weight vector, which, together with the corresponding frequency subsets, forms a multi-frequency superposition pattern. For example, two signal superposition levels have been determined, and the original frequency points and their energy values ​​contained in each level are known, as shown in Tables 1 and 2.

[0029] Table 1, Level A (composed of cluster C1)

[0030] Table 2, Level B (consisting of cluster C1, including the ranges of P2 and P3)

[0031] For level A, the frequency subset is [22, 23, 24, 25, 26, 27, 28] Hz, and the total energy of the level is 705 μV. 2 The weighting coefficients were calculated as [0.139, 0.145, 0.149, 0.149, 0.143, 0.140, 0.135], by calculating the proportion of energy at each frequency point in the total energy of the hierarchy. For hierarchy B, the frequency subset is [45, 46, 47, 48, 49, 95, 96, 97, 98, 99] Hz. Frequencies 50 and 100 were excluded from the subset because their energy was significantly lower than the surrounding peaks. The total energy of the hierarchy is 1810 μV. 2 The weighting coefficients were calculated as [0.091, 0.097, 0.099, 0.098, 0.094, 0.099, 0.105, 0.108, 0.106, 0.102] by calculating the proportion of energy at each frequency point in the total energy of the hierarchy. The constructed multi-frequency superposition mode is shown in Table 3.

[0032] Table 3

[0033] The multi-frequency superposition pattern matrix is ​​convolved with the initially purified electromyographic signal data stream. During the convolution process, the pattern matrix is ​​used as a filter template. The dot product between the template and the signal data is calculated segment by segment through a sliding window to amplify the frequency components that match the pattern. Then, a threshold is applied to the signal amplitude data after the convolution operation. The threshold is set to 1.2 times the average signal amplitude, and signal components with amplitudes exceeding the threshold are retained to form an initial set of signal components. Each component in this set corresponds to a specific frequency subset and weight, and has a data dimension correspondence with the multi-frequency superposition pattern and the electromyographic signal data stream before the convolution operation.

[0034] The initial signal component set is verified for frequency domain features. Each component is converted into a frequency domain spectrum using a Fast Fourier Transform (FFT). The overlap ratio between different component spectra is calculated. If the ratio exceeds 10%, frequency component overlap is identified. Independent Component Analysis (ICA), a spectral separation technique, is then used to optimize the initial signal component set. A separation matrix is ​​constructed by maximizing the non-Gaussianity of the components. Overlapping frequency components are demixed. During demixing, the separation matrix is ​​iteratively updated until the mutual information between components is less than 0.1. After optimization, the final signal component set is obtained. The signal component set is a collection of sub-signals representing different potential source signals or feature patterns, separated from the initially clean electromyographic signal data stream through a series of signal processing techniques. The components in this set are independent of each other, with clear frequency characteristics, effectively solving the problem of intention recognition deviation caused by frequency component mixing, and can more accurately represent different muscle activity patterns.

[0035] Step 3: Extract time-frequency features and interference suppression information based on the signal component set to obtain the component feature combination that reflects the user's motion intention pattern.

[0036] In one specific embodiment, the process of performing step 3 may specifically include the following steps: The dynamic frequency change tracking technique is used to perform in-depth extraction of time-frequency features for each component in the signal component set; Based on the extracted time-frequency feature depth, analyze the frequency variation trend of each component; Based on the frequency variation trend, calculate the interference suppression information between components; Determine whether the interference suppression information is lower than the preset suppression threshold. If so, enhance the independence of components through secondary filtering. Based on the result of secondary filtering, generate an initial component feature combination. The initial component feature combination is subjected to feature consistency verification to determine whether there is feature deviation. If not, the initial component feature combination is taken as the final component feature combination. If so, the initial component feature combination is adjusted by feature correction algorithm, and the feature combination with the deviation corrected is determined as the component feature combination.

[0037] Specifically, the signal component set contains multiple signal components separated in the time-frequency domain, each representing a potential pattern of muscle activity. A dynamic frequency change tracking technique combining wavelet transform and Kalman filtering is used to process each component in the signal component set. The signal component set contains multiple independent sub-signals, each presenting voltage-time time-domain data at a preset sampling rate. During processing, each sub-signal is first decomposed into 5 levels of wavelet coefficients using a db4 wavelet basis, extracting high-frequency and low-frequency coefficients at each scale. Then, Kalman filtering is used to track and predict the high-frequency coefficients. During the filtering process, the preceding sampled values ​​of the wavelet coefficients are used as state variables, and the current sampled value is used as the observation variable. The state equation is set as z(k) = a1 × z(k-1) + a2 × v(k), where a1 is the state transition coefficient (taken as 0.98), a2 is the noise figure (taken as 0.02), and v... (k) represents Gaussian white noise. The filtered wavelet coefficients are obtained through iterative calculation. The instantaneous frequency, peak energy, frequency bandwidth, and other time-frequency characteristic depth data of each component are extracted from these coefficients. These data form a one-to-one mapping relationship with each sub-signal in the signal component set. The instantaneous frequency reflects the frequency change of the component at different times, the peak energy reflects the intensity characteristics of the component, and the frequency bandwidth characterizes the frequency distribution range of the component.

[0038] Based on the extracted time-frequency feature depth data, a linear fitting method is used to fit the trajectory of the instantaneous frequency change of each component over time, obtaining the slope and intercept of the frequency change trend. A positive slope indicates that the frequency increases over time, while a negative slope indicates that the frequency decreases over time. For example, the slope of the instantaneous frequency fitting for the component corresponding to the upper limb grasping action is 2 Hz / ms, and the slope of the fitting for the component corresponding to the extension action is -1.5 Hz / ms. By analyzing the frequency change trend of each component through these fitting parameters, this trend data is directly related to the instantaneous frequency data in the time-frequency feature depth data, which is a further quantification of the dynamic characteristics of the component frequency.

[0039] Based on the fitting parameters of the frequency variation trend, a frequency variation trend vector is constructed. This vector is then used to calculate the interference suppression degree between components. A cosine similarity algorithm is employed to evaluate the directional consistency of the two component trend vectors. Given the trend vectors of two components, V_m and V_n, their interference suppression degree is calculated as follows: , where the symbol · represents the vector dot product, and ||·|| represents the vector magnitude. When the calculated interference suppression degree is lower than a preset threshold (e.g., 0.7), it is determined that there is significant interference between components, and a secondary filtering process is initiated to enhance component independence.

[0040] The signal component set is optimized by a secondary filtering process using adaptive notch filtering. During the filtering process, the interference frequency between components is used as the notch center frequency, and the notch depth is dynamically adjusted according to the interference suppression degree. The lower the interference suppression degree, the greater the notch depth. For example, when the interference suppression degree is 0.5, the notch depth is set to 30dB. After the secondary filtering process, the frequency interference between components is further suppressed, generating an initial component feature combination composed of the time-frequency feature depth data of each component after filtering. This combination forms a data dimension correspondence with the signal component set before the secondary filtering.

[0041] The initial component feature combinations are subjected to feature consistency verification. Feature deviation is determined by calculating the statistical distribution similarity of each component within the combination. Jensen-Shannon divergence is used to assess the probability distribution difference of each pair of components. When the divergence value of any two components exceeds a threshold (e.g., 0.1), feature deviation is identified. For feature combinations with deviation, principal component analysis (PCA) is applied for feature correction. The feature vectors are projected onto a subspace spanned by the principal feature vectors, retaining principal components with a cumulative contribution rate exceeding 85%. Redundancy and deviation between features are eliminated through reconstruction. The corrected feature combinations form the final component feature combinations. Each feature in this combination exhibits good consistency and discriminative power, stably representing different upper limb movement intention patterns and providing reliable feature input for robot control.

[0042] Step 4: Construct a multi-frequency superimposed feature matrix describing the user's upper limb movement intention based on the combination of component features, and monitor the system's response delay fluctuations in real time to obtain a delay fluctuation analysis report.

[0043] In one specific embodiment, the process of performing step 4 may specifically include the following steps: Based on the combination of component features, the classification criteria for signal components are obtained by analyzing the time-frequency feature depth of each component, and the components are classified into fundamental frequency components and harmonic components accordingly. Based on the classification results, the energy superposition ratio of the fundamental frequency component and the harmonic components is calculated to determine the superposition level characteristics of the signal. Based on the superposition hierarchical characteristics, the weights of frequency components in each level are calculated, and the weight values ​​are filled into a preset matrix structure to construct a multi-frequency superposition feature matrix. A delayed detection time window is introduced to monitor the processing from the stabilization of the signal component set to the completion of the construction of the multi-frequency superposition feature matrix in real time; Based on the processing timing data collected within the delay detection time window, it is determined whether there are response delay fluctuations. If so, the time point and amplitude data of the fluctuations are recorded, and a delay fluctuation analysis report is generated.

[0044] Specifically, based on the time-frequency feature depth data in the component feature combination, the fundamental frequency, harmonic frequencies, and corresponding energy proportions of each component are extracted as the basis for signal component classification. The component feature combination contains multiple corrected feature data groups, each associated with a specific electromyographic signal component. During processing, the time-domain voltage data of the component is converted into a frequency-domain spectrum using a Fast Fourier Transform. The frequency with the highest energy proportion in the spectrum is identified as the fundamental frequency, and frequencies that are integer multiples of the fundamental frequency are determined as harmonics. For example, if the 30Hz frequency has an energy proportion of 65% in a component's spectrum, it is identified as the fundamental frequency. The 60Hz and 90Hz frequencies are the second and third harmonics, respectively, with energy proportions of 20% and 10%. According to this rule, the components in the component feature combination are divided into fundamental frequency components and harmonic components. The classification results form a one-to-one mapping relationship with each feature data group in the component feature combination. The fundamental frequency component corresponds to the signal feature reflecting the user's core action intention, while the harmonic component is a signal feature that assists in representing the intention.

[0045] Based on the classification results, the ratio of the total energy of the fundamental frequency component to the total energy of the harmonic components is calculated. This ratio is the energy superposition ratio. The total energy of the fundamental frequency component is obtained by summing the energy values ​​of all fundamental frequency components, while the total energy of the harmonic components is the sum of the energy values ​​of all harmonic components. For example, if the total energy of the fundamental frequency component is 850 μV² and the total energy of the harmonic components is 340 μV², the energy superposition ratio is 2.5:1. Based on this ratio and the frequency distribution range of the components, the superposition level characteristics of the signal are determined. If the energy superposition ratio is greater than 2:1, it is determined to be a two-level superposition level, with the fundamental frequency component as the first level and the harmonic components as the second level. If the ratio is less than 2:1, it is merged into a single superposition level. The superposition level characteristic data is directly related to the fundamental frequency and harmonic component energy data in the classification results and is the structural basis for constructing the feature matrix.

[0046] Based on the defined superposition layer characteristics, weights are assigned to the frequency components within each layer. For a layer L_k containing I frequency components, the weight value for each frequency component is calculated. Where wj is the weight value of the j-th frequency component, Ej is the energy value of the j-th frequency component, and Ei is the energy value of the i-th frequency component. The calculated weight values ​​are filled into a preset r×c matrix structure according to the level number and frequency index, where the row dimension r corresponds to the number of superposition levels, the column dimension c corresponds to the total number of frequency components supported by the system, and the matrix elements store the weight values ​​of a specific level on a specific frequency component, thereby constructing a multi-frequency superposition feature matrix.

[0047] While constructing the feature matrix, a delay detection time window is introduced to monitor the processing in real time. The starting point of the time window is the moment t_s when the independence index of all components in the signal component set reaches a preset threshold, and the ending point is set to the moment t_e when the construction of the multi-frequency superposition feature matrix is ​​completed. Within the time window, processing time-series data is collected at a preset sampling rate, including component classification time, energy calculation time, matrix filling time, etc. These processing time-series data are correlated with the time nodes of the delay detection time window, and each sampling point corresponds to the time consumption value of a processing step.

[0048] Based on the processing timing data collected within the delay detection time window, the difference between the processing time of each stage and the preset baseline time is calculated. The preset baseline time is determined based on multiple experiments; for example, the baseline time for component classification is 50ms, the baseline time for energy calculation is 30ms, and the baseline time for matrix filling is 20ms. If the absolute value of the difference between the processing time of a certain stage and the baseline time exceeds 5ms, it is determined that there is a response delay fluctuation. For any stage identified as fluctuating, the time point and magnitude of the difference are recorded. For example, if the component classification stage fluctuates at 30ms in the time window, with a processing time of 60ms and an amplitude of 10ms, the time point and amplitude of all fluctuations are integrated with the corresponding processing stage information to generate a delay fluctuation analysis report. The report also includes frequency statistics of delay fluctuations and the distribution of fluctuation proportions in each stage. This report is directly related to the processing timing data within the delay detection time window and can intuitively reflect the delay status of the system signal processing process. This effectively solves the problem of not being able to monitor response delay fluctuations in real time. At the same time, the multi-frequency superimposed feature matrix accurately represents the user's upper limb movement intention in a structured form, avoiding robot movement deviations caused by ambiguous intention representations. This provides quantified delay data and accurate intention feature basis for dynamically adjusting the instruction queue.

[0049] Step 5: Analyze the delay fluctuation analysis report, dynamically adjust the logic and timing of the instruction queue based on the analysis results, and generate a corrected control instruction sequence.

[0050] In one specific embodiment, the process of performing step 5 may specifically include the following steps: Based on the delay fluctuation analysis report, analyze whether the fluctuation amplitude exceeds the preset response delay threshold. If so, extract the fluctuation cycle characteristics through delay fluctuation cycle detection technology. Based on the characteristics of the fluctuation cycle and the command triggering conditions, the timing adjustment step size is determined; Based on the timing adjustment step size, the instructions in the instruction queue are sorted by execution priority. Based on the priority sorting results, adjust the instruction queue reconstruction logic to reorganize the queue structure; Based on the adjusted instruction queue, an initial control instruction sequence is generated; The initial control command sequence is logically consistent to determine whether there is a conflict. If not, the initial control command sequence is used as the corrected control command sequence. If so, the control command sequence is optimized through a conflict resolution algorithm to obtain the corrected control command sequence.

[0051] Specifically, the delay fluctuation analysis report includes information on the fluctuation time point, amplitude, and corresponding processing steps. First, the fluctuation amplitude data of each step recorded in the report is analyzed and compared with a preset response delay threshold, which is set according to the real-time requirements of the robot control system (e.g., 10ms). When the fluctuation amplitude of any step exceeds this threshold, delay fluctuation period detection technology is activated. A fast Fourier transform is performed on the fluctuation time series recorded in the report to extract the dominant frequency component of the fluctuation. Based on the separated dominant frequency component, the average and variance of the fluctuation period are calculated as fluctuation period characteristics. The average represents the typical period length, and the variance reflects the fluctuation stability.

[0052] Based on the extracted fluctuation cycle features, the timing adjustment step size is determined by combining the instruction triggering conditions from the electromyography (EMG) signal processing pipeline. Instruction triggering conditions include a signal feature intensity threshold and an intent feature matching degree threshold (e.g., 0.8). A valid trigger is considered when the feature intensity is higher than the intensity threshold and the feature matching degree is greater than the matching degree threshold. The signal feature intensity can be obtained by calculating the root mean square (RMS) value of the EMG signal components, or by extracting the ratio of peak signal energy to background noise energy as an intensity index. The feature matching degree of the user's upper limb movement intent is the quantified similarity value between the extracted EMG signal features (e.g., component feature combinations, multi-frequency superposition feature matrices) and the preset upper limb movement intent feature templates (e.g., grasping, extending, and rotating feature templates). The timing adjustment step size T_s is calculated using the formula T_s = T_base + K_p × P_avg, where T_base is the base step size, K_p is the proportional coefficient adjusted according to the feature matching degree, and P_avg is the average fluctuation cycle. This calculation ensures that a larger adjustment step size is used when the fluctuation period is long and the intention is clear, and a smaller adjustment step size is used when the fluctuation is frequent and the intention is ambiguous.

[0053] Based on the calculated timing adjustment step size, the instructions to be executed in the instruction queue are dynamically prioritized. The sorting algorithm calculates the comprehensive priority score of each instruction using a priority scoring formula: P_score = α×I_imp+β×U_urg+γ×T_s, where I_imp represents the inherent importance parameter preset based on the criticality of the robot's action in the task, U_urg represents the urgency parameter determined in real time based on the intensity of electromyographic signal characteristics, and α, β, and γ are preset weighting coefficients. The timing adjustment step size T_s serves as a dynamic correction factor, and its value directly and positively affects the comprehensive priority score, allowing corresponding instructions to receive higher scheduling priority when system latency fluctuations are significant. A max-heap data structure is used to implement the priority queue, moving high-priority instructions to the front of the queue to ensure that critical instructions are processed first.

[0054] Based on the priority ranking results, the instruction queue reconstruction logic is adjusted. The reconstruction process uses a doubly linked list structure to store the instruction sequence, with each node containing the instruction content, the adjusted execution timestamp, and a comprehensive priority score. The execution timestamp of each instruction is recalculated based on the timing adjustment step size. Instructions originally scheduled to execute at time t1 are rescheduled to a new execution time t1' = t1 + Δt, where Δt is the offset calculated based on instruction priority and fluctuation cycle characteristics. Through this mechanism, the instruction queue is reorganized on the timeline, forming a new execution sequence that matches the system's processing capacity.

[0055] An initial control command sequence is generated based on the reconstructed command queue. This sequence contains robot joint control commands arranged according to the new timestamps. Logical consistency is checked against this sequence to identify kinematic conflicts between adjacent commands, such as joint angle exceeding limits or end effector path interference. The verification process simulates command execution using the robot's kinematic model to detect violations of physical constraints. When a command conflict is detected, a conflict resolution mechanism based on a greedy algorithm is activated. While keeping high-priority commands unchanged, conflicting low-priority commands are adjusted or deleted. For example, a level 2 extension command is deleted, or a delayed command is inserted between grasping and extension commands, resulting in a conflict-free corrected control command sequence. This sequence compensates for timing errors caused by system processing delays and ensures the executability of commands at the robot's kinematic level, providing stable and reliable control input for the upper limb robot.

[0056] Step 6: Smooth and dynamically adjust the control instruction sequence to generate an optimized instruction execution signal.

[0057] In one specific embodiment, the process of performing step 6 may specifically include the following steps: The control command sequence is processed using a smooth transition curve fitting technique, and the smoothed command sequence is obtained based on the fitting results. For the smoothed instruction sequence, signal boundary smoothing processing technology is applied to optimize the boundary continuity of the instruction sequence and obtain an optimized instruction sequence; By combining the feedback frequency dynamic adjustment mechanism, the frequency response characteristics of the optimized instruction sequence are extracted; Adjust the real-time correction gain parameter used to maintain the continuity of the command sequence based on the frequency response characteristics; Based on the adjusted real-time correction gain parameters, the optimized instruction sequence is continuously controlled to suppress abrupt changes between instructions; The instruction execution signal is generated based on the adjusted instruction sequence to achieve transition optimization.

[0058] Specifically, the corrected control command sequence contains robot joint angle or velocity commands arranged by timestamps. The sequence is processed using a smooth transition curve fitting technique, and a continuous and differentiable fitting curve is constructed between adjacent command points using a cubic spline interpolation algorithm. This algorithm solves for the interpolation polynomial coefficients with the constraint of ensuring the continuity of the second derivative of the curve, generating a command sequence that is mathematically guaranteed to be smooth.

[0059] For the smoothed instruction sequence, a signal boundary smoothing technique based on moving weighted average is applied. The boundary points of the beginning and end of the instruction sequence and adjacent instruction segments are selected as the processing objects. Ten adjacent data points, including the boundary points, are set as the calculation window. The weight of the data points within the window decreases linearly with the distance from the boundary points, and the closer the distance, the greater the weight. For example, the weight of the boundary point is 0.2, the weight of the first adjacent data point is 0.18, and so on, decreasing until the weight of the data point at the edge of the window is 0.02. The instruction value of the original boundary point is replaced by the weighted average of the data points within the window, which optimizes the boundary continuity of the instruction sequence and generates an optimized instruction sequence with continuous boundaries in the time domain. The difference between the instruction value of the boundary point and the surrounding data points in this sequence is less than 0.5°, and the derivative of the curve at the boundary is continuous, forming a data dimension correspondence with the smoothed instruction sequence, thus solving the problem of abrupt boundary transitions in the instruction sequence.

[0060] By combining the dynamic adjustment mechanism of feedback frequency, frequency domain characteristic analysis is performed on the optimized instruction sequence. The instruction sequence is transformed from the time domain to the frequency domain through fast Fourier transform to obtain its frequency response characteristics, including the main frequency value, amplitude spectrum peak value and phase spectrum characteristics. The main frequency value reflects the change rhythm of the instruction sequence, the amplitude spectrum peak value reflects the intensity of the instruction change, and the phase spectrum characterizes the time delay characteristics of the instruction sequence.

[0061] Based on the extracted frequency response characteristics, the real-time correction gain parameter used to maintain the continuity of the command sequence is adjusted. This parameter is a dynamic adjustment coefficient for the signal continuity maintenance strategy, with a value range of 0.8 to 1.5. For example, the adjustment rule is as follows: when the main frequency value is higher than 10Hz (corresponding to a fast action intention), the gain parameter is increased by 20% to enhance the intensity of continuity control; when the main frequency value is lower than 5Hz (corresponding to a slow action intention), the gain parameter is decreased by 15% to avoid over-control leading to loss of intention; if the amplitude spectrum peak value exceeds 50°, an additional 5% gain compensation is added. For example, if the main frequency of the command sequence is 8Hz and the amplitude spectrum peak value is 60°, the gain parameter is adjusted from the default value of 1.0 to 1.05. The adjusted gain parameter is dynamically correlated with the main frequency and amplitude spectrum peak data in the frequency response characteristics, directly determining the strength of subsequent continuity control. Preferably, a real-time correction gain adjustment function based on the main frequency value, amplitude spectrum peak value, and phase spectrum characteristics can also be constructed.

[0062] Based on the extracted frequency response characteristics, the real-time correction gain parameter used to maintain the continuity of the command sequence is adjusted. This adjustment is based on the energy ratio of the high-frequency components to the low-frequency components in the frequency response, calculated using the formula G_k = G_0 × (1 + λ1 × (E_h / E_l)), where G_0 is the reference gain, E_h and E_l represent the energy values ​​of the high-frequency and low-frequency bands, respectively, and λ1 is the adjustment coefficient. When a relative increase in high-frequency energy is detected, the gain parameter is increased accordingly to enhance the suppression capability of abrupt command changes.

[0063] The adjusted real-time correction gain parameters are applied to continuously regulate the optimized command sequence. This process is achieved by calculating the difference between adjacent command points and applying gain control. For adjacent command points, the adjusted command value A'_i = A_i + G_k × (A_{i+1} - A_i) × W_i, where A_i represents a specific control quantity sent to a joint of the robot (such as the elbow joint) at time point t_i. This control quantity can be a target position (unit: degrees), a target velocity (unit: degrees / second), or a target torque (unit: Newton-meter). A_{i+1} represents the next control quantity sent to the same joint at the next time point t_{i+1}, and W_i is a weighting coefficient based on the command point position. Preferably, if the difference between adjacent command points (A_{i+1} - A_i) exceeds the preset maximum allowable difference, the original difference is replaced by the maximum allowable difference. This regulation mechanism effectively suppresses instantaneous changes between commands while maintaining the overall trend of the command sequence. The instruction sequence, after continuous adjustment and processing, is finally generated into a transitionally optimized instruction execution signal. This signal maintains the accuracy of the original control intention while possessing the smoothness and continuity required for execution, providing a shock-free and vibration-free motion control basis for upper limb robots.

[0064] Step 7: Fine-tune the timing and verify the synchronization of the instruction execution signal to generate upper limb robot control signals that match the user's action intentions.

[0065] In one specific embodiment, the process of performing step 7 may specifically include the following steps: Based on the instruction execution signal, the frequency correction feedback period is extracted, and the dynamic step size correction weight is determined accordingly. Adjust the timing parameters of instruction execution based on the dynamic step size correction weight; The adjusted timing parameters are calibrated by combining a real-time timing calibration mechanism. The timing parameters after calibration are synchronously verified for instruction execution, and the timing deviation is determined based on the verification results. If timing deviations exist, the instruction execution signals are adjusted using a timing fine-tuning algorithm; The final upper limb robot control signal is generated based on the adjusted instruction execution signal.

[0066] Specifically, the over-optimized command execution signal is a continuous command stream that has been smoothed and dynamically adjusted. The time interval between a complete command transmission-execution-feedback operation is extracted from this signal as the frequency correction feedback cycle. The command execution signal is transmitted to the robot actuator at a frequency of 100Hz. Each signal frame contains command information such as joint angle and execution duration. By recording the time difference between the signal transmission time and the completion time of the robot actuator's feedback action, the average value of multiple feedback cycles is calculated. For example, if the time differences of 10 consecutive feedback cycles are 10ms, 11ms, 9ms, 10ms, 12ms, 9ms, 10ms, 11ms, 9ms, and 10ms respectively, the average value is 10ms, which is the frequency correction feedback cycle. The dynamic step size correction weight is determined based on the extracted frequency correction feedback cycle. For example, the weight is calculated using the formula W_s = 1 + (T_f - 10) / 50, where T_f is the actual extracted frequency correction feedback cycle. The dynamic step size correction weight is positively correlated with the frequency correction feedback cycle, and the weight data directly relates to the adjustment magnitude of subsequent timing parameters.

[0067] The timing parameters of instruction execution are adjusted using dynamic step-size correction weights. These timing parameters include the instruction start timestamp, execution interval, and execution duration. For example, the adjustment formula is tp' = tp × (1 + W_s × ε), where tp is the original timing parameter and ε is a correction factor based on historical timing deviations. This adjustment mechanism allows the timing parameters to adapt to changes in the system's dynamic characteristics, maintaining synchronization with the user's intended execution rhythm.

[0068] The adjusted timing parameters are calibrated using a real-time timing calibration mechanism. This mechanism uses the robot system's high-precision system clock as a reference. By comparing the difference between the adjusted timing parameter timestamp and the system clock, a linear interpolation method is used to correct the deviation. For example, if the adjusted start timestamp is 208ms and the system clock shows the current time as 207.5ms, the difference is 0.5ms. Therefore, the start timestamp is corrected to 207.5ms, and the execution interval and duration are adjusted proportionally. The calibrated timing parameters are synchronized with the system clock, eliminating timing errors caused by hardware clock drift. The calibration data and the adjusted timing parameters form a one-to-one correction relationship, with the parameter deviation controlled within 0.1ms. Preferably, the timing parameters can also be compensated and calibrated by comparing the deviation δ_t = t_d - t_p between the planned instruction execution time t_p and the actual issuance time t_d. The calibration employs a first-order recursive filtering algorithm: t_c = α1 × t_p + (1 - α1) × (t_d - δ_avg), where α1 is the filtering coefficient, δ_avg is the average historical deviation, and t_c is the calibrated timing parameter. This calibration method effectively eliminates the system's accumulated timing error and improves the absolute accuracy of the timing parameters.

[0069] The calibrated timing parameters are synchronously verified during instruction execution. A timing deviation detection model is constructed, and the calibrated timing parameters are input into the model and compared with the timing of characteristic changes in the user's upper limb electromyography (EMG) signal. The characteristic changes in the EMG signal are characterized by extracting data such as the peak time and frequency change trend of the signal. For example, the peak of the EMG signal of the user's grasping action occurs at 200ms, while the calibrated grasping instruction start timestamp is 207.5ms, with a time difference of 7.5ms. If this difference exceeds the preset timing deviation threshold of 5ms, a timing deviation is determined to exist. The result of the synchronous verification is correlated with the calibrated timing parameters and the timing data of the EMG signal characteristics, directly determining whether to start the timing fine-tuning algorithm.

[0070] If timing deviations exist in the synchronization check, the command execution signal is adjusted using a Kalman filter timing fine-tuning algorithm. The algorithm treats the timing deviation as a state variable and the calibrated timing parameters as observation variables to construct a Kalman filter state-space model. In this model, the state transition matrix, control matrix, and observation matrix are configured according to the system model. The optimal timing correction value is obtained through iterative calculation of this model. The adjusted command execution signal eliminates timing deviations, and the signal transmission rhythm is consistent with the timing changes in the user's electromyography (EMG) signal characteristics.

[0071] Based on the adjusted instruction execution signal, it is encoded according to the communication protocol format of the robot actuator. The encoding includes an instruction header, a data segment, and a check bit. The data segment integrates the adjusted timing parameters and action instruction information. The check bit is calculated using the CRC32 algorithm to ensure the integrity of signal transmission. The encoded signal is output in the form of digital pulses, and the pulse frequency matches the receiving frequency of the robot actuator. Finally, an upper limb robot control signal is generated. This signal can directly drive the robot actuator to complete the action corresponding to the user's upper limb action intention, effectively solving the problem of time difference between robot action and user intention.

[0072] The foregoing described the upper limb robot intention recognition method based on electromyography (EMG) signals in the embodiments of this application. The following describes the upper limb robot intention recognition system based on EMG signals in the embodiments of this application. Please refer to [link to relevant documentation]. Figure 2 The present application provides a schematic diagram of the structure of an upper limb robot intention recognition system based on electromyography signals. The system includes: The data acquisition module 10 is used to acquire the upper limb electromyographic signal data stream and perform preliminary noise suppression processing to obtain a preliminary clean electromyographic signal data stream.

[0073] The set determination module 20 is used to perform time-frequency analysis on the electromyographic signal data stream, extract frequency component features that characterize the user's movement intention, and determine the set of signal components.

[0074] The feature extraction module 30 is used to extract time-frequency features and interference suppression information based on the signal component set, and obtain a combination of component features that reflect the user's motion intention pattern.

[0075] The matrix construction module 40 is used to construct a multi-frequency superimposed feature matrix describing the user's upper limb movement intention based on the combination of component features, and to monitor the system's response delay fluctuations in real time and obtain a delay fluctuation analysis report.

[0076] The instruction correction module 50 is used to analyze the delay fluctuation analysis report, dynamically adjust the logic and timing of the instruction queue based on the analysis results, and generate a corrected control instruction sequence.

[0077] The smooth transition module 60 is used to smooth and dynamically adjust the control instruction sequence to generate an instruction execution signal with optimized transition.

[0078] The signal generation module 70 is used to fine-tune the timing and verify the synchronization of the instruction execution signal, and generate upper limb robot control signals that match the user's action intentions.

[0079] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for upper limb robot intention recognition based on electromyographic signals, characterized in that, The method includes: Step 1: Acquire upper limb electromyography (EMG) signal data stream and perform preliminary noise suppression processing to obtain a preliminary clean EMG signal data stream; Step 2: Perform time-frequency analysis on the electromyographic signal data stream, extract frequency component features that characterize the user's movement intention, and determine the signal component set; Step 3: Extract time-frequency features and interference suppression information based on the signal component set to obtain a combination of component features that reflect the user's motion intention pattern; Step 4: Construct a multi-frequency superimposed feature matrix describing the user's upper limb movement intention based on the component feature combination, and monitor the system's response delay fluctuation in real time to obtain a delay fluctuation analysis report; Step 5: Analyze the delay fluctuation analysis report, dynamically adjust the logic and timing of the instruction queue based on the analysis results, and generate a corrected control instruction sequence; Step 6: Smooth and dynamically adjust the control command sequence to generate a transition-optimized command execution signal; Step 7: Perform timing fine-tuning and synchronization verification on the instruction execution signal to generate upper limb robot control signals that match the user's action intentions.

2. The method according to claim 1, characterized in that, Step 1 includes: Acquire upper limb electromyography signal data streams through real-time acquisition devices; The upper limb electromyography signal data stream is noise suppressed by bandpass filtering technology, and harmonic interference frequencies are separated by frequency interval uniformity analysis method. Then, they are filtered out by notch filter to generate intermediate filtered data stream. The intermediate filtered data stream is subjected to signal integrity verification. If data loss or abnormal fluctuations are detected, interpolation compensation technology is used for repair. Based on the repaired data stream, the continuity and stability of the signal are evaluated, and a preliminary clean electromyographic signal data stream is generated based on the evaluation results.

3. The method according to claim 1, characterized in that, Step 2 includes: The electromyographic signal data stream is decomposed using time-frequency decomposition technology to extract frequency component features; Analyze the peak energy distribution range of the frequency component characteristics, and based on the peak energy distribution range, use a hierarchical clustering algorithm to determine the signal superposition level; Based on the aforementioned signal superposition levels, each signal superposition level is mapped to a frequency subset and assigned a weight based on energy ratio to construct a multi-frequency superposition mode. The multi-frequency superposition mode is convolved with the electromyographic signal data stream, and a threshold filter is applied to separate the initial signal component set. The initial signal component set is subjected to frequency domain feature verification to determine whether there is frequency component overlap. If so, the initial signal component set is optimized and adjusted using spectrum separation technology to obtain the final signal component set.

4. The method according to claim 1, characterized in that, Step 3 includes: The time-frequency features of each component in the signal component set are deeply extracted using dynamic frequency change tracking technology. Based on the extracted time-frequency feature depth, analyze the frequency variation trend of each component; Based on the frequency change trend, calculate the interference suppression information between components; Determine whether the interference suppression information is lower than a preset suppression threshold. If so, enhance the independence of components through secondary filtering and generate an initial component feature combination based on the result of the secondary filtering. The initial component feature combination is subjected to feature consistency verification to determine whether there is feature deviation. If not, the initial component feature combination is taken as the final component feature combination. If so, the initial component feature combination is adjusted by feature correction algorithm, and the feature combination with the deviation corrected is determined as the component feature combination.

5. The method according to claim 1, characterized in that, Step 4 includes: Based on the combination of component features, the classification criteria for signal components are obtained by analyzing the time-frequency feature depth of each component, and the components are classified into fundamental frequency components and harmonic components accordingly. Based on the classification results, the energy superposition ratio of the fundamental frequency component and the harmonic components is calculated to determine the superposition level characteristics of the signal. Based on the superposition layer characteristics, the weights of frequency components in each layer are calculated, and the weight values ​​are filled into a preset matrix structure to construct the multi-frequency superposition feature matrix. A delayed detection time window is introduced to monitor the processing from the stabilization of the signal component set to the completion of the construction of the multi-frequency superposition feature matrix in real time. Based on the processing timing data collected within the delay detection time window, it is determined whether there is a response delay fluctuation. If so, the time point and amplitude data of the fluctuation are recorded, and a delay fluctuation analysis report is generated.

6. The method according to claim 1, characterized in that, Step 5 includes: Based on the aforementioned delay fluctuation analysis report, it is analyzed whether the fluctuation amplitude exceeds the preset response delay threshold. If so, the fluctuation cycle characteristics are extracted using delay fluctuation cycle detection technology. Based on the fluctuation cycle characteristics and the instruction triggering conditions, the timing adjustment step size is determined; Based on the timing adjustment step size, priority sorting is performed on the instructions in the instruction queue; Based on the priority sorting results, adjust the instruction queue reconstruction logic to reorganize the queue structure; Based on the adjusted instruction queue, generate the initial control instruction sequence; The initial control command sequence is subjected to a logical consistency check to determine whether there is a conflict. If not, the initial control command sequence is used as the corrected control command sequence. If so, the control command sequence is optimized by a conflict resolution algorithm to obtain the corrected control command sequence.

7. The method according to claim 1, characterized in that, Step 6 includes: The control command sequence is processed using a smooth transition curve fitting technique, and a smoothed command sequence is obtained based on the fitting result. For the smoothed instruction sequence, signal boundary smoothing processing technology is applied to optimize the boundary continuity of the instruction sequence to obtain an optimized instruction sequence; By combining the feedback frequency dynamic adjustment mechanism, the frequency response characteristics of the optimized instruction sequence are extracted; Based on the frequency response characteristics, adjust the real-time correction gain parameter used to maintain the continuity of the command sequence; Based on the adjusted real-time correction gain parameters, the optimized instruction sequence is continuously regulated to suppress abrupt changes between instructions; The instruction execution signal is generated based on the regulated instruction sequence, and then optimized for transition.

8. The method according to claim 1, characterized in that, Step 7 includes: Based on the instruction execution signal, the frequency correction feedback period is extracted, and the dynamic step size correction weight is determined accordingly. The timing parameters for instruction execution are adjusted based on the dynamic step size correction weight. The adjusted timing parameters are calibrated by combining a real-time timing calibration mechanism. The timing parameters after calibration are synchronously verified for instruction execution, and the timing deviation is determined based on the verification results. If timing deviations exist, the instruction execution signal is adjusted using a timing fine-tuning algorithm; The final upper limb robot control signal is generated based on the adjusted instruction execution signal.

9. An upper limb robot intention recognition system based on electromyographic signals, used to implement the method as described in any one of claims 1 to 8, characterized in that, The system includes: The data acquisition module is used to acquire upper limb electromyographic signal data streams and perform preliminary noise suppression processing to obtain a preliminary clean electromyographic signal data stream. The set determination module is used to perform time-frequency analysis on the electromyographic signal data stream, extract frequency component features that characterize the user's movement intention, and determine the set of signal components; The feature extraction module is used to extract time-frequency features and interference suppression information based on the signal component set, and obtain a combination of component features that reflect the user's motion intention pattern. The matrix construction module is used to construct a multi-frequency superimposed feature matrix describing the user's upper limb movement intention based on the component feature combination, and to monitor the system's response delay fluctuation in real time and obtain a delay fluctuation analysis report. The instruction correction module is used to analyze the delay fluctuation analysis report, dynamically adjust the logic and timing of the instruction queue based on the analysis results, and generate a corrected control instruction sequence. A smooth transition module is used to smooth and dynamically adjust the control command sequence to generate a transition-optimized command execution signal; The signal generation module is used to perform timing fine-tuning and synchronization verification on the instruction execution signal to generate upper limb robot control signals that match the user's action intentions.

Images