Muscle force estimation method based on surface electromyogram signals and adjustable input characteristics

Abstract

The invention discloses a muscle force estimation method based on surface electromyogram signals and adjustable input characteristics. The method comprises the steps that the surface electromyogram signals and force signals in a muscle contraction state are synchronously collected; the EMD-DWT method is adopted to preprocess the surface electromyogram signals; segmenting the surface electromyogram signals and the force signals by adopting a sliding window, extracting features of the surface electromyogram signals, extracting an average value of the force signals, and making sample labels; a Pearson's correlation coefficient method is adopted to calculate correlation coefficients of the extracted features and the force signals; sequentially adding features with high correlation coefficients to form an input vector of the model, constructing a muscle force estimation regression model for training and testing, and adjusting the number of input features according to errors of a training set and a testing set. According to the method, effective features can be accurately and rapidly extracted, the number of the features can be determined, and it is guaranteed that high muscle force estimation precision is achieved under the condition that the input dimension is low.

Description

Technical Field

[0001] The invention relates to the fields of human biological signal processing and pattern recognition, and in particular to a muscle force estimation method based on surface electromyography signals and adjustable input characteristics. Background Art

[0002] In recent years, with the development and progress of medical technology, the treatment effects of more and more patients have been improved, but the number of disabled people has increased year by year. The main disability of disabled patients is the loss of limb function, which affects their daily self-care ability. However, disabled patients can restore their limb movement ability by wearing prostheses. At present, prostheses can be controlled by surface electromyography signals, electroencephalogram signals, sounds, manual commands, etc. Compared with several other control sources, surface electromyography signals are a signal containing a large amount of human intention information. It has the advantages of non-invasiveness and high real-time performance. The signal acquisition is simple and can collect signals from multiple muscles at the same time, which is conducive to identifying the human body's movement state. Analyzing surface electromyography signals can more accurately identify human intentions and reduce information delays. At the same time, wearing prostheses containing surface electromyography sensors is more comfortable, so surface electromyography signals are widely used in the field of intelligent prosthetic control.

[0003] Although surface electromyographic signals have obvious advantages in the field of intelligent prosthetic control, there are still challenges in practical applications. Surface electromyographic signals are random and non-stationary, and are easily affected by environmental interference. Especially in the state of motion, the motion artifacts generated by the movement of muscles, skin and electrode positions cause the baseline drift of the signal. The change of joint angle during movement causes the change of surface electromyographic signal amplitude and frequency. In addition, it will be interfered by various noise sources such as power system power frequency interference. These factors will reduce the accuracy of muscle force estimation. The processing steps for estimating muscle force based on surface electromyographic signals are divided into preprocessing, feature extraction and regression model prediction. The preprocessing operation includes filtering, window segmentation, normalization, etc. Features are extracted from the preprocessed surface electromyographic signals, generally including time domain features, frequency domain features and time-frequency domain combined features. Finally, the extracted features are input into the regression model to estimate the size of muscle force. The accuracy of muscle force estimation depends to a large extent on the quality of the model input features. Existing studies extract a large number of features from surface electromyographic signals and propose a series of feature screening methods to screen effective features from the extracted features to improve the accuracy of muscle force estimation. Too high a number of input features will cause input redundancy, and too low a number of input features will cause underfitting of the regression model. Therefore, accurately extracting effective features and the number of features is an issue worth studying.

[0004] The existing muscle force estimation methods based on surface electromyography have the following disadvantages:

[0005] 1. The interference of various noise sources such as motion artifacts generated in the motion state and power system power frequency interference makes the surface electromyographic signal mixed with a large amount of noise, reducing the accuracy of muscle force estimation;

[0006] 2. It is impossible to accurately extract effective features and it is difficult to determine the number of features, resulting in input redundancy and decreased accuracy of muscle force estimation. Summary of the invention

[0007] The object of the present invention is to provide a muscle force estimation method based on surface electromyography signals and adjustable input characteristics with good noise reduction effect and high accuracy.

[0008] To achieve the above object, the technical solution of the present invention is as follows:

[0009] A muscle force estimation method based on surface electromyography signals and adjustable input features comprises the following steps:

[0010] S1, synchronously collect surface electromyographic signals and force signals under muscle contraction;

[0011] S2, surface EMG signals were preprocessed using the EMD-DWT method;

[0012] S3, using sliding window to segment surface electromyographic signal and force signal, extracting the features of surface electromyographic signal, extracting the average value of force signal, and making sample set;

[0013] S4, the correlation coefficient between the extracted features and the force signal was calculated using the Pearson correlation coefficient method;

[0014] S5, add features with high correlation coefficients in turn to form the input vector of the model, build a muscle strength estimation regression model for training and testing, and adjust the number of input features according to the errors of the training set and the test set.

[0015] In step S1, it includes but is not limited to using a surface electromyography signal sensor to collect surface electromyography signals at the brachioradialis, deltoid and biceps respectively to obtain multi-channel surface electromyography signals; and using an isokinetic device to collect force signals of the elbow joint.

[0016] Wherein step S2 comprises the following steps:

[0017] S21, using the EMD method to decompose the surface electromyographic signal into a series of intrinsic modal components imf;

[0018] S22, calculate the energy ratio EP of each intrinsic modal component in the imf; EP refers to the ratio of the signal energy in the 0-10Hz frequency band to the energy in the 11-400Hz frequency band;

[0019] S23, setting an upper threshold t1 and a lower threshold t2, removing the connotation modal components with an EP greater than the threshold t1, taking the connotation modal components with an EP less than t2 as high-frequency components imfsH, and taking the remaining connotation modal components as low-frequency components imfsL;

[0020] S24, the signal reconstructed by the high-frequency component imfsH is denoted as signalHigh, and the signal reconstructed by the low-frequency component imfsL is denoted as signalLow;

[0021] S25, determining the wavelet basis and the number of decomposition layers, and performing DWT processing on the signalHigh signal and the signalLow signal;

[0022] S26, the first high-frequency approximate coefficient obtained by decomposing the signalHigh signal is reset to zero and then reconstructed, and the reconstructed signal is recorded as DEsignalLow, and the low-frequency approximate coefficient obtained by decomposing the signalLow signal is reset to zero and then reconstructed, and the reconstructed signal is recorded as DEsignalLow;

[0023] S27, further reconstructing DEsignalHigh and DEsignalLow, and the final reconstructed signal is the filtered surface electromyography signal.

[0024] In step S22, the steps of calculating EP are as follows:

[0025] S221, calculate the signal energy of each frequency band, the calculation formula is:

[0026]

[0027] Where x[n] is the nth data of the time domain signal, N is the number of data sampling points of the signal, and k is the signal frequency;

[0028] S222, calculate the value of EP, the calculation formula is as follows:

[0029]

[0030] Wherein step S3 comprises the following steps:

[0031] S31, using a sliding window to segment the surface electromyographic signal and the force signal, the sliding window moving step is half of the window length, and the window length ranges from 50ms to 200ms;

[0032] S32, extracting features of the surface electromyographic signal of each channel, including but not limited to root mean square, standard deviation, variance, mean value, integrated electromyographic value, waveform length, William amplitude, slope change rate, average frequency, median frequency, peak frequency, and power spectral density;

[0033] S33, extracting the average value of the force signal and using it as a sample label to create a sample set.

[0034] Wherein step S4 comprises the following steps:

[0035] S41, the Pearson correlation coefficient method is used to calculate the correlation coefficient between the features extracted from each channel and the label value;

[0036] S42, the result of adding the correlation coefficients of the corresponding features in all channels is used to evaluate the correlation between the feature and the force signal. The larger the coefficient, the higher the correlation.

[0037] Wherein step S5 specifically comprises the following steps:

[0038] S51, divide the sample set into a training set and a test set in a ratio of 8:2;

[0039] S52, select the BiLSTM model as the muscle strength estimation model, determine the number of network layers, the size of the hidden layer, the optimizer, and the dimension of the input;

[0040] S53, based on the previous input, sequentially adding features with high correlation coefficients to form the input vector of the model;

[0041] S54, train and test the BiLSTM model, and record the root mean square error of the training set and the test set;

[0042] S55, repeat steps S53 and S54, and stop training and testing the model when the root mean square error of the training set and the test set tends to be stable.

[0043] Compared with the prior art, the present invention has the following obvious advantages:

[0044] 1. A filtering processing method combining empirical mode decomposition (EMD) and discrete wavelet transform (DWT) is proposed. This method innovatively uses the energy ratio of high-frequency and low-frequency signals as a threshold to filter out modal components containing a large amount of low-frequency noise. At the same time, it solves the common modal aliasing problem in EMD decomposition, and the problem of baseline drift and power frequency interference in surface electromyographic signals under motion, thus improving the noise reduction effect.

[0045] 2. In order to solve the problems of accurate and fast extraction of effective features and the number of features as well as the low accuracy of muscle force estimation in motion, the Pearson correlation coefficient method is used to calculate the correlation coefficient between the extracted features and the force signal, and the features with higher correlation are added as the input of the regression model in sequence. The number of input features is adjusted according to the root mean square error of the training set and the test set. At the same time, the EMD-DWT filtering method is used to solve the problem of mixed noise of surface electromyographic signals in motion. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0047] Figure 1 is a flow chart of the method in the present invention;

[0048] Figure 2 It is a structural schematic diagram of the EMD-DWT filtering method in the present invention;

[0049] Figure 3 It is a sequence diagram of the original surface electromyography force signal in the present invention;

[0050] Figure 4 It is a sequence diagram of surface electromyography force signal after EMD-DWT filtering in the present invention;

[0051] Figure 5 This is a structural diagram of the BiLSTM model in the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose and technical solution of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are not limited to the present invention.

[0053] like Figure 1 As shown, a muscle force estimation method based on surface electromyography signals and adjustable input features includes the following steps:

[0054] Step 1, synchronously collecting surface electromyographic signals and force signals under muscle contraction state;

[0055] Step 2: Use EMD-DWT method to preprocess surface electromyographic signals;

[0056] Step 3: Use a sliding window to segment the surface electromyographic signal and the force signal, extract the features of the surface electromyographic signal, extract the average value of the force signal, and make a sample set;

[0057] Step 4, using the Pearson correlation coefficient method to calculate the correlation coefficient between the extracted features and the force signal;

[0058] Step 5: Add features with high correlation coefficients to form the input vector of the model in sequence, build a muscle strength estimation regression model for training and testing, and adjust the number of input features according to the errors of the training set and the test set.

[0059] In step 1, it includes but is not limited to using a surface electromyography signal sensor to collect surface electromyography signals from the brachioradialis, deltoid and biceps respectively to obtain multi-channel surface electromyography signals; and using an isokinetic device to collect force signals of the elbow joint.

[0060] like Figure 2 As shown, step 2 includes the following steps:

[0061] Step 2.1, using the EMD method to decompose the surface electromyographic signal into a series of intrinsic modal components imf;

[0062] Step 2.2, calculate the energy ratio EP of each intrinsic modal component in the imf; EP refers to the ratio of the signal energy in the 0-10 Hz frequency band to the energy in the 11-400 Hz frequency band;

[0063] Step 2.3, set the threshold t=5, remove the corresponding connotation modal components whose EP is greater than the threshold t, and record the remaining connotation modal components as imfs;

[0064] Step 2.4, according to the EP value corresponding to the intrinsic modal component in imfs, the imfs is divided into a high-frequency component imfsH and a low-frequency component imfsL, the intrinsic modal component corresponding to the EP less than 0.02 is determined as the high-frequency component, and the remaining intrinsic modal components are regarded as low-frequency components;

[0065] Step 2.5, reconstruct the high-frequency component imfsH and the low-frequency component imfsL respectively. The reconstructed signal of the high-frequency component imfsH is denoted as signalHigh, and the reconstructed signal of the low-frequency component imfsL is denoted as signalLow.

[0066] Step 2.6, determine the wavelet basis as db4, the decomposition level as 4, and perform DWT processing on the signalHigh signal and the signalLow signal;

[0067] Step 2.7, the low-frequency approximate coefficient obtained by decomposing the signalHigh signal is recorded as cA4, and the high-frequency detail coefficients are recorded as cD1, cD2, cD3, cD4; the low-frequency approximate coefficient obtained by decomposing the signalLow signal is recorded as ca4, and the high-frequency detail coefficients are recorded as cd1, cd2, cd3, cd4; among which, the signal frequencies corresponding to the coefficients of cD1, cD2, cD3, cD4, cA4 and cd1, cd2, cd3, cd4, ca4 decrease in sequence;

[0068] Step 2.8, set the cD1 coefficient to zero, and record the signal reconstructed by the cD1, cD2, cD3, cD4, and cA4 coefficients as DEsignalHigh; set the ca4 coefficient to zero, and record the signal reconstructed by the cd1, cd2, cd3, cd4, and ca4 coefficients as DEsignalLow, and further reconstruct DEsignalHigh and DEsignalLow, as follows Figure 4 As shown, the final reconstructed signal is the filtered surface electromyography signal.

[0069] In step 2.2, the steps to calculate EP are as follows:

[0070] Step 2.2.1, calculate the signal energy of each frequency band, the calculation formula is:

[0071]

[0072] Where x[n] is the nth data of the time domain signal, N is the amount of signal data, and k is the signal frequency;

[0073] Step 2.2.2, calculate the value of EP, the calculation formula is as follows:

[0074] Step 3 includes the following steps:

[0075] Step 3.1: Use a sliding window to segment the surface electromyographic signal and the force signal. The sliding window moving step is half of the window length, and the window length is set to 100ms.

[0076] Step 3.2, extracting the characteristics of the surface electromyographic signal of each channel, including but not limited to root mean square, standard deviation, variance, mean value, integrated electromyographic value, waveform length, William amplitude, slope change rate, average frequency, median frequency, peak frequency, and power spectral density;

[0077] Step 3.3: Extract the average value of the force signal and use it as the sample label to create a sample set.

[0078] Wherein step 4 specifically comprises the following steps:

[0079] Step 4.1, use the Pearson correlation coefficient method to calculate the correlation coefficient between the features extracted from each channel and the label value;

[0080] Step 4.2: The result of adding the correlation coefficients of the corresponding features in all channels is used to evaluate the correlation between the feature and the force signal. The larger the coefficient, the higher the correlation.

[0081] Wherein step 5 specifically comprises the following steps:

[0082] Step 5.1, divide the sample set into training set and test set in a ratio of 8:2;

[0083] Step 5.2: Select the BiLSTM model as the muscle strength estimation model, determine the number of network layers, the size of the hidden layer, the optimizer, and the input dimension;

[0084] Step 5.3, based on the previous input, add features with high correlation coefficients to form the input vector of the model;

[0085] Step 5.4: Train and test the BiLSTM model, and record the root mean square error of the training set and the test set;

[0086] Step 5.5, repeat steps S53 and S54, and stop training and testing the model when the root mean square error of the training set and the test set tends to be stable.

[0087] Figure 3 For the original surface electromyography force signal sequence diagram in the present invention, a surface electromyography signal sequence of one channel is selected;

[0088] Figure 4 is the surface electromyography signal sequence diagram after EMD-DWT filtering in the present invention, and Figure 3 The surface electromyography force signal sequence was compared to eliminate the baseline drift of the original signal, proving that the EMD-DWT filtering effect is good;

[0089] Figure 5 This is a structural diagram of the BiLSTM model in the present invention. BiLSTM is a recurrent neural network suitable for processing sequence data. It can capture the dependency relationship between the previous and next moments in the time series and achieve more accurate muscle strength estimation.

[0090] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Without departing from the original invention technology, several improvements and variations may be made, which shall all be included in the protection scope of the present invention.

Claims

1. A muscle force estimation method based on surface electromyography signals and adjustable input features, characterized in that: The following steps are involved: S1, synchronously collect surface electromyographic signals and force signals under muscle contraction; S2, surface EMG signals were preprocessed using the EMD-DWT method; S3, using sliding window to segment surface electromyographic signal and force signal, extracting the features of surface electromyographic signal, extracting the average value of force signal, and making sample set; S4, the correlation coefficient between the extracted features and the force signal was calculated using the Pearson correlation coefficient method; S5, add features with high correlation coefficients in turn to form the input vector of the model, build a muscle strength estimation regression model for training and testing, and adjust the number of input features according to the errors of the training set and the test set.

2. A muscle force estimation method based on surface electromyography signals and adjustable input features according to claim 1, characterized in that: Step S1 includes but is not limited to using a surface electromyography signal sensor to respectively collect surface electromyography signals of the brachioradialis, deltoid and biceps brachii to obtain multi-channel surface electromyography signals; The force signals of the elbow joint were collected using an isokinematic device.

3. The muscle force estimation method based on surface electromyography signal and adjustable input characteristics according to claim 1, characterized in that: Step S2 includes the following steps: S21, using the EMD method to decompose the surface electromyographic signal into a series of intrinsic modal components imf; S22, calculate the energy ratio EP of each intrinsic modal component in the imf; EP refers to the ratio of the signal energy in the 0-10Hz frequency band to the energy in the 11-400Hz frequency band; S23, setting an upper threshold t1 and a lower threshold t2, removing the connotation modal components with an EP greater than the threshold t1, taking the connotation modal components with an EP less than t2 as high-frequency components imfsH, and taking the remaining connotation modal components as low-frequency components imfsL; S24, the signal reconstructed by the high-frequency component imfsH is denoted as signalHigh, and the signal reconstructed by the low-frequency component imfsL is denoted as signalLow; S25, determining the wavelet basis and the number of decomposition layers, and performing DWT processing on the signalHigh signal and the signalLow signal; S26, the first high-frequency approximate coefficient obtained by decomposing the signalHigh signal is reset to zero and then reconstructed, and the reconstructed signal is recorded as DEsignalLow, and the low-frequency approximate coefficient obtained by decomposing the signalLow signal is reset to zero and then reconstructed, and the reconstructed signal is recorded as DEsignalLow; S27, further reconstructing DEsignalHigh and DEsignalLow, and the final reconstructed signal is the filtered surface electromyography signal.

4. The muscle force estimation method based on surface electromyography signal and adjustable input characteristics according to claim 1, characterized in that: In step S3, a sliding window is used to segment the surface electromyographic signal and the force signal. The sliding window moves in a step size of half the window length. The features of the surface electromyographic signal of each channel are extracted, including but not limited to root mean square, standard deviation, variance, mean value, integrated electromyographic value, waveform length, William amplitude, slope change rate, average frequency, median frequency, peak frequency, and power spectral density. The average value of the force signal is extracted and used as the sample label.

5. The muscle force estimation method based on surface electromyography signal and adjustable input characteristics according to claim 1, characterized in that: In step S4, the Pearson correlation coefficient method is used to calculate the correlation coefficient between the feature extracted from each channel and the force signal. The result of adding the correlation coefficients of the corresponding features in all channels is used to evaluate the correlation between the feature and the force signal. The larger the coefficient, the higher the correlation.

6. The muscle force estimation method based on surface electromyography signal and adjustable input characteristics according to claim 1, characterized in that: Step S5 specifically includes the following steps: S51, divide the sample set into a training set and a test set in a ratio of 8:2; S52, select the BiLSTM model as the muscle strength estimation model, determine the number of network layers, the size of the hidden layer, the optimizer, and the dimension of the input; S53, based on the previous input, sequentially adding features with high correlation coefficients to form the input vector of the model; S54, train and test the BiLSTM model, and record the root mean square error of the training set and the test set; S55, repeat steps S53 and S54, and stop training and testing the model when the root mean square error of the training set and the test set tends to be stable.

Images