Method for recognizing gesture by using transient-phase electromyographic signal on basis of electromyographic activation intensity transfer features and adaptive analysis-window length

Abstract

Provided in the present invention is a method for recognizing a gesture by using a transient-phase electromyographic signal on the basis of electromyographic activation intensity transfer features and an adaptive analysis-window length, which method aims to solve problems existing in conventional electromyographic signal processing technology when utilizing transient-phase signals, including the imbalance between spatial features and computational costs, the imbalance between recognition accuracy and response speed, and the low algorithm generalization across individuals and tasks. The method comprises: first, in an offline mode, performing signal acquisition and preprocessing, start point detection, signal frame formatting and feature extraction, and training an SVM classifier; and in an online mode, acquiring a preprocessed signal, detecting a start point, formatting the signal, extracting electromyographic activation intensity transfer features, inputting same into the offline SVM classifier, dynamically adjusting the length of an analysis window on the basis of confidence, and finally feeding back and adjusting a confidence threshold in real time. The practice proves that the method can effectively take into account both spatial features and computational costs, can achieve a balance between recognition accuracy and response speed across different individuals and tasks, and can significantly improve algorithm generalization, thus providing more efficient and accurate technical support in fields such as prosthesis control and exercise rehabilitation devices.

Description

A Transient Phase Electromyographic Signal Gesture Recognition Method Based on Electromyographic Activation Intensity Transfer Features and Adaptive Analysis Window Length Technical Field

[0001] This invention relates to the field of biosignal processing and prediction technology, particularly to a technique for predicting motor intentions using the transient phase of electromyographic signals, which is widely used in fields such as prosthetic control and sports rehabilitation equipment. Background Technology

[0002] Electromyography (EMG) signals are bioelectrical signals generated on the skin surface during muscle contraction. They contain rich information about muscle activity and can be used to predict motor intent. Traditional EMG signal processing methods typically rely on steady-state signals for feature extraction, but transient signals (i.e., the transition from rest to motion) contain important information about motor intent, which is particularly important for improving the speed of motor intent recognition. However, existing technologies utilizing transient signals are based on fixed-length analysis windows, which has the following problems:

[0003] The imbalance between spatial features and computational cost: Existing analysis methods either rely on analyzing the conventional time-domain features of discrete channel signals that have lost their spatial characteristics, or use computationally expensive deep learning methods such as CNN and LSTM models. Therefore, it is difficult to achieve both spatial features and low computational cost.

[0004] Imbalance between recognition accuracy and response speed: Increasing the analysis window length improves recognition accuracy but reduces response speed; conversely, shortening the analysis window length improves response speed but reduces recognition accuracy. Therefore, a fixed analysis window length cannot balance recognition accuracy and response speed.

[0005] Low algorithmic generalization between individuals and tasks: Existing electromyography signal processing methods are usually based on a fixed analysis window length, which cannot adapt to changes in different individuals and different tasks.

[0006] This invention proposes a gesture recognition method that can dynamically adjust the analysis window length, which can adaptively adjust the analysis window length for different individuals under different tasks, and achieve a balance between recognition accuracy and response speed by evaluating the confidence level. Summary of the Invention

[0007] This invention proposes a transient phase electromyographic signal gesture recognition method based on electromyographic activation intensity transfer characteristics and adaptive analysis window length. It can adaptively adjust the analysis window length for different individuals under different tasks, and achieve a balance between recognition accuracy and response speed by evaluating the confidence level.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows: a method for gesture recognition of transient phase electromyographic signals based on electromyographic activation intensity transfer characteristics and adaptive analysis window length, the method comprising the following steps:

[0009] S1: Signal acquisition and preprocessing in offline mode: Offline acquisition and preprocessing of electromyographic signals are performed using a high-density electromyography sensor.

[0010] S2: Starting point detection in offline mode: Using a starting point T0 detection algorithm based on detEMG, the starting moment of the motion intention is identified.

[0011] S3: Signal frame formatting in offline mode: Frame formatting is performed on the signal throughout the entire transient phase after the starting point T0.

[0012] S4: Feature extraction in offline mode: For each frame of signal, calculate the electromyographic activation intensity transfer (directional activation intensity transfer statistics) features sequentially.

[0013] S5: Training SVM classifiers: This stage is equivalent to training multiple classifiers corresponding to different window lengths.

[0014] S6: Signal acquisition and preprocessing in online mode: Online signal acquisition and preprocessing are performed using high-density electromyography sensors.

[0015] S7: Signal start point detection in online mode: using mEMG-based detection onset Online signal start point T0 online The detection algorithm identifies the start time of the motion intent.

[0016] S8: Signal frame formatting in online mode: from start point T0 online The signals during the entire transient phase thereafter undergo frame formatting.

[0017] S9: Feature extraction in online mode: For each frame of signal, calculate the electromyographic activation intensity transfer (directional activation intensity transfer statistics) features sequentially.

[0018] S10: Input offline SVM classifier: Calculate the prediction type and confidence level of the current frame and perform dynamic window length adjustment.

[0019] S11: Real-time feedback and adjustment: Adjust the confidence threshold based on real-time results.

[0020] Furthermore, the high-density electromyography sensor described in S1 needs to be able to provide 64 or more channels of electromyography signals in order to acquire electrical activity from different muscle regions.

[0021] Furthermore, the signal preprocessing method described in S1 employs the following steps:

[0022] S1-1: 50Hz notch filter: Removes power frequency interference.

[0023] S1-2: Bandpass filter (30Hz-500Hz): Simulates muscle characteristics, removes low-frequency and high-frequency noise, and retains effective electromyographic signals.

[0024] Furthermore, in S2, the start point detection in offline mode uses mEMG-based methods. onset Starting point T0 detection algorithm. This algorithm calculates detEMG to find the starting point of signal abrupt change, thus identifying the start of motion intent. The specific steps are as follows:

[0025] S2-1: First, sum the data from all channels, which is equivalent to compressing the high-density electromyography signal into one channel. Based on this, calculate the detEMG using the following formula:

[0026] Where N is the total number of data points, |x i | is the absolute value of the i-th data point, mEMG i T is the electromyographic activation intensity value of the i-th data point. det It detects the time span, detEMG i It is the detection signal of the i-th data point.

[0027] S2-2: Calculate the moment when the value reaches 0.5 times the maximum value; this moment is the starting point T0.

[0028] S2-3: Simultaneously obtain the mEMG at this time. i And denoted as mEMG onset This value will be used for online identification later.

[0029] Furthermore, the method for formatting the signal frames in the entire transient phase in S3 is as follows: starting from T0, at fixed time intervals ΔT, the portion of the signal from each channel between time t(i-1) and time t(i) is averaged, and the result is defined as the signal of the i-th frame (i = 1, 2, 3... f_sum). f_sum is the total number of frames in the entire transient phase. The specific calculation formula is as follows:

[0030] Among them, F i This is the signal of the i-th frame (multi-channel data; if the sensor size is 8*8, then it is in 8*8 matrix format), t i =T0+i*Δt is the end time of the i-th frame.

[0031] f_sum represents the total number of frames in the entire transient phase, and is calculated using the following formula:

[0032] Where Δt is the time interval, T transient This is the duration of the entire transient phase, consistent with offline mode.

[0033] Furthermore, in S4, the feature extraction method in offline mode involves calculating electromyographic activation intensity transfer (directional activation intensity transfer statistics) features for each frame of signal. Electromyographic activation intensity transfer features can extract texture information of the signal within local regions, making them suitable for describing the spatial characteristics of electromyographic signals. These features have a strong ability to capture changes in muscle activity patterns. The specific calculation steps are as follows:

[0034] S4-1: Calculation of electromyographic activation intensity transfer characteristics

[0035] First, for each electromyography sensor channel in the current frame, calculate its ST in the horizontal direction. x and vertical direction ST y Electromyographic intensity transfer value: ST x =mEMG(x+1,y)-mEMG(x-1,y)ST y =mEMG(x,y+1)-mEMG(x,y-1)

[0036] Where mEMG(x,y) is the electromyographic activation intensity value at the electromyographic sensor channel (x,y) in the current frame.

[0037] Then, the activation intensity transfer amplitude ST and activation intensity transfer direction θ for each electromyography sensor channel (x,y) are calculated:

[0038] S4-2: Constructing a statistical graph of electromyographic activation intensity transfer

[0039] The current frame is divided into several subframes, each with 2x2 channels. Within each subframe, an activation intensity transition graph is calculated based on the activation intensity transition direction of each channel. The specific steps are as follows:

[0040] S4-2-1 divides the activation intensity transfer direction θ (0-180°) into several intervals.

[0041] S4-2-2 assigns the activation intensity transfer amplitude ST of each electromyography (EMG) sensor channel to the corresponding statistical graph slot according to its activation intensity transfer direction θ. Specifically, if the activation intensity transfer direction θ of the EMG sensor channel falls in the k-th interval, then the activation intensity transfer amplitude ST of that EMG sensor channel is added to the statistical graph of the k-th interval.

[0042] S4-2-3 converts the statistical graph into a feature vector, resulting in the final electromyographic activation intensity transfer feature vector.

[0043] Furthermore, training an SVM classifier in S5 involves the following steps:

[0044] S5-1: Divide the training set and the test set

[0045] All gesture sample data are divided into training and test sets according to a certain ratio, which can be selected between 1:4 and 1:2.

[0046] S5-2: Signal Input

[0047] Starting from frame m (the initial input frame), the electromyographic activation intensity transfer features of the signals from frame m to frame m+1 are averaged across dimensions and then input into the SVM classifier.

[0048] S5-3: Prediction results and confidence calculation. For each gesture category, the SVM classifier outputs a prediction result and its confidence value. The confidence value represents the reliability of the prediction result.

[0049] S5-4: Adjust the signal input, continuously extend the signal window length range of the input SVM classifier, execute 5-3 and 5-4 until all frames in the transient phase are covered.

[0050] Furthermore, the signal preprocessing method described in S6 employs the following steps:

[0051] S6-1: 50Hz notch filter: Removes power frequency interference.

[0052] S6-2: Bandpass filter (30Hz-500Hz): Simulates muscle characteristics, removes low-frequency and high-frequency noise, and retains effective electromyographic signals.

[0053] Furthermore, the EMG-based method described in S7 onset Online signal start point T0 online The detection algorithm employs the following steps:

[0054] S7-1: First, sum the data from all channels, which is equivalent to compressing the high-density electromyography signal into one channel. Based on this, calculate the mEMG using the following formula. i :

[0055] Where N is the total number of data points, |x i | is the absolute value of the i-th data point, mEMG i It is the electromyographic activation intensity value of the i-th data point.

[0056] S7-2: Compare current mEMG i With mEMG onset (starting time mEMG)i (threshold), until the current mEMG i Greater than or equal to mEMG onset This point is now marked as the starting point T0. online .

[0057] Furthermore, the specific method for formatting the signal frame described in S8 is as follows: from T0 online Then, at fixed time intervals ΔT, the portion of the signal from each channel between time t(i-1) and time t(i) is averaged, and the result is defined as the signal of the i-th frame (i = 1, 2, 3... f_sum). f_sum is the total number of frames in the entire transient phase. Refer to S3 for the specific calculation formula.

[0058] Furthermore, in S9, the feature extraction method in online mode involves calculating electromyographic activation intensity transfer (directional activation intensity transfer statistics) features for each frame of signal. The specific calculation method and formula are detailed in S4.

[0059] Furthermore, the specific steps for inputting the offline SVM classifier in S10 are as follows:

[0060] S10-1: Calculate the prediction type and its confidence level for the current frame.

[0061] Starting from frame m (the initial input frame), the electromyographic activation intensity transfer features of the signals from frame m to frame m+1 are averaged dimensionally and input into the corresponding offline SVM classifier. The current prediction result (Result) and its confidence level (Confidence) are then calculated.

[0062] S10-2: Dynamic window length adjustment. Adjust the confidence level (Confidence) from the previous step to the set confidence threshold (Confidence). accept The comparison is performed. If the confidence level reaches the set acceptable confidence level, the current prediction result is output and further data acquisition stops. If not, the input signal analysis window length is increased, and step 10-1 is executed again until the confidence level reaches the set acceptable confidence level.

[0063] S10-3: Output the recognition result and stop data acquisition. If the length of the input signal has covered the entire transient phase, also output the recognition result.

[0064] Furthermore, the specific methods for real-time feedback and adjustment in S11 are as follows:

[0065] S11-1: Preset initial confidence threshold accept_0

[0066] S11-2: Based on real-time results, adjust the confidence threshold Confidenceaccept_new to adapt to the needs of different individuals and tasks, thereby dynamically adjusting the signal acquisition and analysis strategy to achieve a balance between prediction accuracy and response speed.

[0067] Compared with the prior art, the advantages of the present invention are as follows:

[0068] 1. Reduced computational cost: Existing analysis methods, while preserving spatial features, can only utilize costly deep learning methods such as CNN and LSTM models. This solution, using electromyography activation intensity transfer features, can significantly reduce computational costs.

[0069] 2. Improved response speed: Compared with other gesture recognition solutions based on electromyography signals, this solution directly processes the transitional signal before the signal stabilizes, which can advance the response time by about 400ms.

[0070] 3. Achieving a balance between recognition accuracy and response speed: This invention proposes a gesture recognition method that can dynamically adjust the analysis window length, which can adaptively adjust the analysis window length for different individuals under different tasks, and achieve a balance between recognition accuracy and response speed by evaluating the confidence level. Attached Figure Description

[0071] Figure 1 is a schematic diagram of the method flow of the present invention.

[0072] Figure 2 is a schematic diagram of the starting point detection results in an embodiment of this application.

[0073] Figure 3 is a schematic diagram showing the changes in analysis window length and average accuracy with confidence threshold settings in the embodiments of this application. Detailed Implementation

[0074] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0075] Example: Referring to Figures 1-3, a method for gesture recognition of transient phase electromyographic signals based on electromyographic activation intensity transfer features and adaptive analysis window length is described. The method includes the following steps:

[0076] S1: Signal acquisition and preprocessing in offline mode. Gesture classification data from an open-source dataset was used, which employed a 128-channel high-density electromyography (EMG) sensor to acquire electrical activity from different muscle regions.

[0077] S1-1: 50Hz notch filter: Removes power frequency interference.

[0078] S1-2: Bandpass filter (30Hz-500Hz): Simulates muscle characteristics, removes low-frequency and high-frequency noise, and retains effective electromyographic signals.

[0079] S2: Starting point detection in offline mode: Using a starting point T0 detection algorithm based on detEMG, the starting moment of the motion intention is identified.

[0080] In S2's offline mode, the start point detection uses a detEMG-based start point T0 detection algorithm. This algorithm calculates the detEMG signal to find the start point of signal abrupt change, identifying the beginning of motion intent. The specific steps are as follows:

[0081] S2-1: First, sum the data from all channels, which is equivalent to compressing the high-density electromyography signal into one channel. Based on this, calculate the detEMG using the following formula:

[0082] Where N is the total number of data points, |x i | is the absolute value of the i-th data point, mEMG i T is the electromyographic activation intensity value of the i-th data point. det It detects the time span, detEMG i S2-2: Calculate the moment when the signal reaches 0.5 times the maximum value; this moment is the starting point T0.

[0083] S2-3: Simultaneously obtain the mEMG at this time. i And denoted as mEMG onset This value will be used for online identification later.

[0084] S3: Signal Frame Formatting in Offline Mode: The signal is formatted into frames throughout the entire transient phase starting from point T0. After T0, at fixed time intervals ΔT, the portion of the signal from each channel between time t(i-1) and time t(i) is averaged. The result is defined as the i-th frame signal (i = 1, 2, 3... f_sum), where f_sum is the total number of frames in the entire transient phase. The specific calculation formula is as follows:

[0085] Among them, F i This is the signal of the i-th frame (multi-channel data; if the sensor size is 8*8, then it is in 8*8 matrix format), t i =T0+i*Δt is the end time of the i-th frame.

[0086] f_sum represents the total number of frames in the entire transient phase, and is calculated using the following formula:

[0087] Where Δt is the time interval, which can be set to 50-100ms depending on the task, and T transient It is the duration of the entire transient phase, which can be set to an integer multiple of Δt between 300ms and 500ms, depending on the task.

[0088] S4: Feature extraction in offline mode: For each frame of signal, calculate the electromyographic activation intensity transfer (directional activation intensity transfer statistics) features sequentially. The specific calculation steps are as follows:

[0089] S4-1: Calculation of electromyographic activation intensity transfer characteristics

[0090] First, for each electromyography sensor channel in the current frame, calculate its ST in the horizontal direction. x and vertical direction ST y Electromyographic intensity transfer value: ST x =mEMG(x+1,y)-mEMG(x-1,y)ST y =mEMG(x,y+1)-mEMG(x,y-1)

[0091] Where mEMG(x,y) is the electromyographic activation intensity value at the electromyographic sensor channel (x,y) in the current frame.

[0092] Then, the activation intensity transfer amplitude ST and activation intensity transfer direction θ for each electromyography sensor channel (x,y) are calculated:

[0093] S4-2: Constructing a statistical graph of electromyographic activation intensity transfer

[0094] The current frame is divided into several subframes, each with 2x2 channels. Within each subframe, an activation intensity transition graph is calculated based on the activation intensity transition direction of each channel. The specific steps are as follows:

[0095] S4-2-1 divides the activation intensity transfer direction θ (0-180°) into several intervals.

[0096] S4-2-2 assigns the activation intensity transfer amplitude ST of each electromyography (EMG) sensor channel to the corresponding statistical graph slot according to its activation intensity transfer direction θ. Specifically, if the activation intensity transfer direction θ of the EMG sensor channel falls in the k-th interval, then the activation intensity transfer amplitude ST of that EMG sensor channel is added to the statistical graph of the k-th interval.

[0097] S4-2-3 converts the statistical graph into a feature vector, resulting in the final electromyographic activation intensity transfer feature vector.

[0098] S5: Training SVM classifiers: This stage trains multiple classifiers corresponding to different window lengths.

[0099] S5-1: Divide the training and test sets. Divide all gesture sample data into training and test sets according to a certain ratio, which can be selected between 1:4 and 1:2.

[0100] S5-2: Signal Input

[0101] Starting from frame m (the initial input frame), the electromyographic activation intensity transfer features of the signals from frame m to frame m+1 are averaged across dimensions and then input into the SVM classifier.

[0102] S5-3: Prediction Results and Confidence Calculation

[0103] [Corrected according to Rule 91, 09.07.2025] For each gesture category, the SVM classifier outputs a prediction result and its confidence value, whereby the confidence value represents the reliability of the prediction result. Figure 3 shows how the accuracy of the 16 gesture recognition methods changes with the increase of the analysis window in this embodiment.

[0104] S5-4: Adjust signal input

[0105] The signal window length range input to the SVM classifier is continuously extended, and steps 5-3 and 5-4 are performed until all frames within the transient phase are covered.

[0106] S6: Signal acquisition and preprocessing in online mode: Online signal acquisition and preprocessing are performed using high-density electromyography sensors.

[0107] S6-1: 50Hz notch filter: Removes power frequency interference.

[0108] S6-2: Bandpass filter (30Hz-500Hz): Simulates muscle characteristics, removes low-frequency and high-frequency noise, and retains effective electromyographic signals.

[0109] S7: Signal start point detection in online mode: using mEMG-based detection onset Online signal start point T0 online The detection algorithm identifies the start time of the motion intent.

[0110] S7-1: First, sum the data from all channels, which is equivalent to compressing the high-density electromyography signal into one channel. Based on this, calculate the mEMG using the following formula. i :

[0111] Where N is the total number of data points, |x i | is the absolute value of the i-th data point, mEMG i It is the electromyographic activation intensity value of the i-th data point.

[0112] S7-2: Compare current mEMG i With mEMG onset (starting time mEMG) i (threshold), until the current mEMG i Greater than or equal to mEMG onset This point is now marked as the starting point T0. online .

[0113] S8: Signal frame formatting in online mode: from start point T0 online The signals during the entire transient phase thereafter undergo frame formatting. From T0 online Then, at fixed time intervals ΔT, the portion of the signal from each channel between time t(i-1) and time t(i) is averaged, and the result is defined as the signal of the i-th frame (i = 1, 2, 3... f_sum). f_sum is the total number of frames in the entire transient phase. Refer to S3 for the specific calculation formula.

[0114] S9: Feature extraction in online mode: For each frame of signal, calculate the electromyographic activation intensity transfer (directional activation intensity transfer statistics) features sequentially. Refer to S4 for specific calculation methods and formulas.

[0115] S10: Input offline SVM classifier: Calculate the prediction type and confidence level of the current frame and perform dynamic window length adjustment.

[0116] S10-1: Calculate the prediction type and its confidence level for the current frame.

[0117] Starting from frame m (the initial input frame), the electromyographic activation intensity transfer features of the signals from frame m to frame m+1 are averaged dimensionally and input into the corresponding offline SVM classifier. The current prediction result (Result) and its confidence level (Confidence) are then calculated.

[0118] S10-2: Dynamic window length adjustment.

[0119] Compare the confidence level (Confidence) from the previous step with the set confidence threshold (Confidence). accept The comparison is performed. If the confidence level reaches the set acceptable confidence level, the current prediction result is output and further data acquisition stops. If not, the input signal analysis window length is increased, and step 10-1 is executed again until the confidence level reaches the set acceptable confidence level.

[0120] S10-3: Output Results

[0121] Output the recognition result and stop collecting data. If the length of the input signal has covered the entire transient phase, also output the recognition result.

[0122] S11: Real-time feedback and adjustment: Adjust the confidence threshold based on real-time results.

[0123] S11-1: Preset initial confidence threshold accept_0

[0124] S11-2: Based on real-time results, adjust the confidence threshold Confidenceaccept_new to adapt to the needs of different individuals and tasks, thereby dynamically adjusting the signal acquisition and analysis strategy to achieve a balance between prediction accuracy and response speed.

[0125] Figure 3 shows the changes in analysis window length and average accuracy with confidence threshold settings in the embodiments of this application.

[0126] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention. Equivalent transformations or substitutions made based on the above technical solutions all fall within the scope of protection of the claims of the present invention.

Claims

1. A method for gesture recognition of transient phase electromyographic signals based on electromyographic activation intensity transfer features and adaptive analysis window length, characterized in that, The method includes the following steps: S1: Offline mode signal acquisition and preprocessing, using a high-density electromyography (EMG) sensor for offline acquisition and preprocessing of EMG signals. S2: Offline mode start point detection, using a detEMG-based start point T0 detection algorithm to identify the start time of motion intent. S3: Signal frame formatting in offline mode. The signal is formatted in frames throughout the entire transient phase after the starting point T0. S4: Feature extraction in offline mode. For each frame of signal, the electromyographic activation intensity transfer features are calculated sequentially. S5: Train the SVM classifier. This stage is equivalent to training multiple classifiers corresponding to different window lengths. S6: Online signal acquisition and preprocessing, using high-density electromyography sensors for online signal acquisition and preprocessing. S7: Signal start point detection in online mode, using the online signal start point T0 based on detEMG. online The detection algorithm identifies the start time of the motion intent. S8: Signal frame formatting in online mode, starting from point T0. online The signal during the entire transient phase thereafter undergoes frame formatting processing. S9: Feature extraction in online mode. For each frame of signal, the electromyographic activation intensity transfer features are calculated sequentially. S10: Input the offline SVM classifier, calculate the predicted type and confidence level of the current frame, and dynamically adjust the window length. S11: Real-time feedback and adjustment, adjusting the confidence threshold based on real-time results.

2. The method for gesture recognition of transient phase electromyographic signals based on electromyographic activation intensity transfer features and adaptive analysis window length according to claim 1, characterized in that, In S1, the high-density electromyography sensor needs to be able to provide 64 or more channels of electromyography signals in order to acquire electrical activity from different muscle regions; The signal preprocessing method in S1 adopts the following steps: S1-1: 50Hz notch filter: removes power frequency interference. S1-2: Bandpass filter (30Hz-500Hz): Simulates muscle characteristics, removes low-frequency and high-frequency noise, and retains effective electromyographic signals; In step S2, the offline mode start point detection uses a start point T0 detection algorithm based on detEMG. This algorithm finds the start point of signal abrupt change by calculating the detEMG of the signal, thus identifying the start of the motion intention. The specific steps are as follows: S2-1: First, sum the data from all channels, which is equivalent to compressing the high-density electromyography signal into one channel. Based on this, calculate the detEMG using the following formula: Where N is the total number of data points, |x i | is the absolute value of the i-th data point, mEMG i T is the electromyographic activation intensity value of the i-th data point. det It detects the time span, detEMG i It is the detection signal of the i-th data point. S2-2: Calculate the moment when the value reaches 0.5 times the maximum value; this moment is the starting point T0. S2-3: Simultaneously obtain the mEMG at this time. i And denoted as mEMG onset This value will be used for online identification later.

3. The method for gesture recognition of transient phase electromyographic signals based on electromyographic activation intensity transfer features and adaptive analysis window length according to claim 1, characterized in that, The method for formatting the signal frames in the entire transient phase in S3 is as follows: After T0, at fixed time intervals ΔT, the portion of the signal from each channel between time t(i-1) and time t(i) is averaged, and the result is defined as the signal of the i-th frame (i = 1, 2, 3... f_sum). f_sum is the total number of frames in the entire transient phase, and the specific calculation formula is as follows: Among them, F i This is the signal of the i-th frame (multi-channel data; if the sensor size is 8*8, then it is in 8*8 matrix format), t i =T0+i*Δt is the end time of the i-th frame. f_sum represents the total number of frames in the entire transient phase, and is calculated using the following formula: Where Δt is the time interval, which can be set to 50-100ms depending on the task, and T transient It is the duration of the entire transient phase, which can be set to an integer multiple of Δt between 300ms and 500ms, depending on the task.

4. The method for gesture recognition of transient phase electromyographic signals based on electromyographic activation intensity transfer features and adaptive analysis window length according to claim 1, characterized in that, In S4, the offline feature extraction method calculates the electromyographic activation intensity transfer feature for each frame of signal. The electromyographic activation intensity transfer feature can extract the texture information of the signal in the local region, which is suitable for describing the spatial features of electromyographic signals. The electromyographic activation intensity transfer feature has a strong ability to capture the pattern changes of muscle activity. The specific calculation steps are as follows: S4-1: Calculate the characteristics of electromyographic activation intensity transfer. First, for each electromyography sensor channel in the current frame, calculate its ST in the horizontal direction. x and vertical direction ST y Electromyographic intensity transfer values: ST x =mEMG(x+1,y)-mEMG(x-1,y) ST y =mEMG(x,y+1)-mEMG(x,y-1) Where mEMG(x,y) is the electromyographic activation intensity value at the electromyographic sensor channel (x,y) in the current frame. Then, the activation intensity transfer amplitude ST and activation intensity transfer direction θ for each electromyography sensor channel (x,y) are calculated: S4-2: Construct an electromyographic activation intensity transfer statistical map. Divide the current frame into several subframes, each with 2*2 channels. Within each subframe, calculate the activation intensity transfer statistical map based on the activation intensity transfer direction of each channel. The specific steps are as follows: S4-2-1 divides the activation intensity transfer direction θ (0-180°) into several intervals. S4-2-2 assigns the activation intensity transfer amplitude ST of each electromyography (EMG) sensor channel to the corresponding statistical graph slot according to its activation intensity transfer direction θ. Specifically, if the activation intensity transfer direction θ of the EMG sensor channel falls in the k-th interval, then the activation intensity transfer amplitude ST of that EMG sensor channel is added to the statistical graph of the k-th interval. S4-2-3 converts the statistical graph into a feature vector, resulting in the final electromyographic activation intensity transfer feature vector.

5. The method for gesture recognition of transient phase electromyographic signals based on electromyographic activation intensity transfer features and adaptive analysis window length according to claim 1, characterized in that, Training an SVM classifier in S5 involves the following steps: S5-1: Divide the data into training and testing sets. All gesture sample data are divided into training and testing sets according to a certain ratio, which can be chosen between 1:4 and 1:

2. S5-2: Signal input. Starting from frame m (initial input frame), the electromyographic activation intensity transfer features of the signals from frame m to frame (m+1) are averaged across dimensions and input into the SVM classifier. S5-3: Prediction Results and Confidence Calculation. For each gesture category, the SVM classifier outputs a prediction result and its confidence value. The confidence value represents the reliability of the prediction result. S5-4: Adjust the signal input, continuously extend the signal window length range of the input SVM classifier, execute 5-3 and 5-4 until all frames in the transient phase are covered.

6. The method for gesture recognition of transient phase electromyographic signals based on electromyographic activation intensity transfer features and adaptive analysis window length according to claim 1, characterized in that, The signal preprocessing method in S6 adopts the following steps: S6-1: 50Hz notch filter to remove power frequency interference. S6-2: Bandpass filter (30Hz-500Hz), simulates muscle characteristics, removes low-frequency and high-frequency noise, and retains effective electromyographic signals; Based on mEMG in S7 onset Online signal start point T0 online The detection algorithm employs the following steps: S7-1: First, sum the data from all channels, which is equivalent to compressing the high-density electromyography signal into one channel. Based on this, calculate the mEMG using the following formula. i : Where N is the total number of data points, |x i | is the absolute value of the i-th data point, mEMG i It is the electromyographic activation intensity value of the i-th data point. S7-2: Compare current mEMG i With mEMG onset (starting time mEMG) i (threshold), until the current mEMG i Greater than or equal to mEMG onset This point is now marked as the starting point T0. online .

7. The method for transient phase electromyographic signal gesture recognition based on electromyographic activation intensity transfer features and adaptive analysis window length according to claim 1, characterized in that, The specific method for signal frame formatting in S8 is as follows: from T0 online Then, at fixed time intervals ΔT, the portion of the signal from each channel between time t(i-1) and time t(i) is averaged, and the result is defined as the signal of the i-th frame (i = 1, 2, 3... f_sum). f_sum is the total number of frames in the entire transient phase. The specific calculation formula is given in claim 5, where Δt is the time interval, and T is the sum of the time intervals. transient This is the duration of the entire transient phase, consistent with offline mode; In S9, the feature extraction method in online mode is to calculate the electromyographic activation intensity transfer feature for each frame of signal.

8. The method for gesture recognition of transient phase electromyographic signals based on electromyographic activation intensity transfer features and adaptive analysis window length according to claim 1, characterized in that, The specific steps for inputting the offline SVM classifier in S10 are as follows: S10-1: Calculate the prediction type and its confidence level for the current frame. Starting from frame m (the initial input frame), the electromyographic activation intensity transfer features of the signals from frame m to frame (m+1) are averaged dimensionally and input into the corresponding offline SVM classifier. The current prediction result Result and its confidence level Confidence are then calculated. S10-2: Dynamic window length adjustment Compare the confidence level (Confidence) from the previous step with the set confidence threshold (Confidence). accept The comparison is performed. If the confidence level reaches the set acceptable level, the current prediction result is output and further data collection stops. If not, the input signal analysis window length is increased, and step 10-1 is repeated until the confidence level reaches the set acceptable level. S10-3: Output the recognition result and stop collecting data. If the length of the input signal has covered the entire transient phase, the recognition result will also be output.

9. The method for gesture recognition of transient phase electromyographic signals based on electromyographic activation intensity transfer features and adaptive analysis window length according to claim 1, characterized in that, The specific methods for real-time feedback and adjustment in S11 are as follows: S11-1: Preset initial confidence threshold accept_0 S11-2: Based on real-time results, adjust the confidence threshold Confidenceaccept_new to adapt to the needs of different individuals and tasks, thereby dynamically adjusting the signal acquisition and analysis strategy to achieve a balance between prediction accuracy and response speed.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the transient phase electromyographic signal gesture recognition method based on electromyographic activation intensity transfer features and adaptive analysis window length as described in any one of claims 1 to 9.

Images