Electromyographic signal decoding method and device based on long short-term memory network improved by genetic algorithm

Abstract

The invention provides an electromyographic signal decoding method and device based on a genetic algorithm improved long short-term memory network, and the method comprises the steps: firstly, obtaining an electromyographic signal sample set, and carrying out the data preprocessing; thirdly, initializing individuals, including six hyper-parameter values including the number of first-layer neurons, the number of second-layer neurons, the first-layer dropout rate, the second-layer dropout rate, the learning rate and the batch size, of the LSTM model, so as to initialize a population; then, training and verifying LSTM models configured by different individuals by using the electromyographic signal sample set, calculating fitness based on verification indexes, and screening the individuals according to the fitness; thirdly, performing crossover operation and mutation operation on the hyper-parameter values of the screened individuals to obtain a new population, and repeating iteration until the individual with the highest fitness is converged or iteration is finished; and finally, training the LSTM model of the individual configuration with the highest fitness for electromyographic signal decoding.

Description

Technical Field

[0001] The present invention relates to the field of human-computer interaction, and in particular to a myoelectric signal decoding method. Background Art

[0002] Surface electromyography (sEMG) is a bioelectric signal collected from the skin surface when skeletal muscles contract. It can provide detailed information on muscle activity and is widely used in human-machine interaction systems such as robot-assisted rehabilitation, prosthetic control, industrial robot operation, smart supply chain, and sports injury prevention. EMG signal decoding refers to the process of converting sEMG into useful information (such as muscle activity level, muscle strength, movement intention, etc.). sEMG is the superposition of a large number of motor unit action potentials in time and space. It has complex characteristics such as non-invasiveness and real-time, which makes the EMG signal decoding process very complicated, especially in the recognition of hand movement intention.

[0003] With the development of artificial intelligence, gesture recognition methods based on deep learning have become a research hotspot, including the use of convolutional neural networks to achieve gesture classification. However, in gesture recognition methods based on deep learning, the accuracy of hand movement intention recognition depends largely on feature selection and model parameter optimization. In the surface electromyography signal classification task, the data contribution of different channels and time periods varies, and the feature performance of different gestures varies significantly. A single feature extraction method is difficult to meet the needs of all gesture classifications. Traditional manual adjustment of model hyperparameters requires multiple experiments, and it is difficult to find the global optimal solution, which increases the need for human intervention to a certain extent and reduces the generalization ability of the model. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a method and device for myoelectric signal decoding based on genetic algorithm and long short-term memory (LSTM) network.

[0005] According to a first aspect, the present invention provides a method for decoding an electromyographic signal, comprising the following steps: S0. Obtain an electromyographic signal sample set, and divide the electromyographic signal training sample set and the electromyographic signal verification sample set based on the electromyographic signal sample set, wherein the feature of each electromyographic signal sample is electromyographic signal time series data, and the label is a gesture category.

[0006] S1. Initialize multiple hyperparameter value combinations for the LSTM model, each hyperparameter value combination includes the number of neurons in the first layer, the number of neurons in the second layer, the dropout rate in the first layer, the dropout rate in the second layer, the learning rate, and the batch size of the LSTM model.

[0007] S2. For the first LSTM model configured by using any combination of hyperparameter values, first train it using the training sample set, then verify the trained first LSTM model using the verification sample set, calculate the fitness based on the verification index, and then select the hyperparameter value combination according to the fitness.

[0008] S3. By performing crossover and mutation operations on the hyperparameter value combinations screened out in step S2, multiple new hyperparameter value combinations are obtained.

[0009] S4. Repeat iterative steps S2-S3 until the hyperparameter value combination with the highest fitness converges.

[0010] S5. Use the hyperparameter value combination with the highest fitness finally selected to configure the LSTM model, and train it based on the electromyographic signal sample set to obtain the final LSTM model for electromyographic signal decoding.

[0011] According to a second aspect, the present invention provides an electromyographic signal decoding device, the device comprising: The sample processing module is used to obtain an electromyographic signal sample set, and divide the electromyographic signal sample set into an electromyographic signal training sample set and an electromyographic signal verification sample set based on the electromyographic signal sample set, wherein the feature of each electromyographic signal sample is the electromyographic signal time series data, and the label is the gesture category.

[0012] The initialization module is used to initialize multiple current hyperparameter value combinations of the LSTM model. Each hyperparameter value combination includes the values ​​of the number of neurons in the first layer, the number of neurons in the second layer, the dropout rate in the first layer, the dropout rate in the second layer, the learning rate, and the batch size of the LSTM model.

[0013] The model configuration module is used to obtain the first LSTM model by using any current hyperparameter value combination configuration.

[0014] A training and verification module is used to train and verify the first LSTM model and calculate the fitness based on the verification index; the training is performed based on the training sample set; and the verification is performed based on the verification sample set.

[0015] The screening module is used to screen the hyperparameter value combinations according to the fitness and input them into the update module, and to determine whether the hyperparameter value combination with the highest fitness has converged.

[0016] The update module is used to perform crossover and mutation operations on the selected hyperparameter value combinations, update the current multiple hyperparameter value combinations, and input them into the model configuration module for iteration.

[0017] The training module is used to train the hyperparameter value combination with the highest fitness to configure the LSTM model after the hyperparameter value combination with the highest fitness converges, so as to obtain the final LSTM model for electromyographic signal decoding; the training is performed based on the electromyographic signal sample set.

[0018] The electromyographic signal decoding method provided by the present invention uses a genetic algorithm to optimize specific key hyperparameters in the LSTM model, which simplifies the selection process of the LSTM model parameters and improves the generalization ability of the model. In addition, for gesture recognition classification tasks with significant differences in feature performance, this method can significantly improve the accuracy of gesture recognition without tedious manual feature extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0020] Figure 1 A flow chart of a myoelectric signal decoding method provided by the present invention; Figure 2 A schematic diagram of a LSTM model storage unit structure provided by the present invention; Figure 3 This is a comparison chart of the accuracy of the electromyographic signal decoding method provided by the present invention and the traditional classifier; Figure 4 This is the second comparison chart of the accuracy of the electromyographic signal decoding method provided by the present invention and the traditional classifier. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0022] The present invention provides an electromyographic signal decoding method and device based on a genetic algorithm and a long short-term memory network. The core idea of ​​the electromyographic signal decoding method is to optimize specific hyperparameters in LSTM through a genetic algorithm (GA), determine the optimal value of the specific hyperparameter of the LSTM model, and thereby construct a final LSTM model for electromyographic signal decoding. The method optimizes specific hyperparameters based on a genetic algorithm, regards a combination of hyperparameter values ​​as an individual (chromosome) of a genetic algorithm, regards the value of each hyperparameter as a gene, regards multiple hyperparameter value combinations as a population, and regards the verification effect of the LSTM model trained by the hyperparameter value combination as fitness. The process of optimizing hyperparameter values ​​is a process in which the population continuously iterates according to fitness and the individuals finally converge.

[0023] The electromyographic signal decoding method provided by the present invention is described in detail below in conjunction with specific embodiments.

[0024] The LSTM network model used in the embodiment includes two layers, each layer is composed of a number of storage units (i.e., neurons). LSTM is a recursive neural network specifically used to process time series data. It learns the long-term and short-term dependencies of sequence data through storage units c. Storage unit c can store the time state of the network. Its structural diagram is shown in FIG. Figure 2 As shown in Figure 1, the LSTM storage unit is controlled by three gates: forget gate, input gate, and output gate.

[0025] The forget gate uses the sigmoid function (symbol represents) determines what information to discard from the previous cell state. This function is based on the previous hidden state h t-1 and the current input x t Generate a weight value between 0 and 1, indicating how much information to discard, and multiply the output weight value by the unit state at the previous moment element by element, retaining important information to the current unit state, as shown in formula (1).

[0026] (1)

[0027] In the formula, f t To preserve the information of the current unit state, W f is the weight matrix of the forget gate, b f is the bias term of the forget gate.

[0028] The input gate is responsible for deciding what information needs to be updated from the current input to the cell state. First, a sigmoid function is used to decide how much information from h should be added. t-1 and x t information, as shown in formula (2).

[0029]

[0030] In the formula, i t For added information, W i is the weight matrix of the input gate, b i is the bias term of the input gate.

[0031] Subsequently, the temporary cell state generates a candidate cell state through the tanh function , contains information that can be added, as shown in formula (3).

[0032]

[0033] Where W c is the weight matrix of the candidate unit state, b c is the bias term of the candidate unit state.

[0034] Finally, the added information i output by the input gate t and candidate cell status Multiply them and add the result to the cell state to get the current cell state c t , as shown in formula (4).

[0035]

[0036] Where c t-1 is the unit state at the previous moment.

[0037] The output gate is responsible for selecting valid information in the cell state as output. First, the sigmoid function determines the output from h t-1 and x t Which values ​​are extracted as output, as shown in formula (5).

[0038]

[0039] In the formula, o t From h t-1 and x t The output extracted from o is the weight matrix of the output gate, b o is the bias term of the output gate.

[0040] Then, use the tanh function to convert the current cell state c t Scale to the range of -1 to 1, and then multiply it with the previously extracted output to get the current hidden state h t , and used as the input of the next moment, as shown in formula (6).

[0041]

[0042] The above introduces the basic structure of the LSTM model used in the embodiment. Next, a specific method for optimizing the LSTM model based on a genetic algorithm and applying it to electromyographic signal decoding is introduced.

[0043] The flowchart of the electromyographic signal decoding method provided by the present invention is as follows: Figure 1 As shown, the method mainly includes six steps, and the specific implementation methods of each step are introduced below.

[0044] S0. Obtain an electromyographic signal sample set, and divide the electromyographic signal sample set into an electromyographic signal training sample set and an electromyographic signal verification sample set based on the electromyographic signal sample set, which are subsequently used for training and verification of an LSTM model, respectively.

[0045] In the EMG signal sample set, the feature of each EMG signal sample is the EMG signal time series data, and the label is the gesture category.

[0046] The electromyographic signal time series data is obtained by normalizing the corresponding original sequence data and performing overlapping sliding window segmentation processing. The original sequence data is the surface electromyographic signal time series data recorded by the armband when the subject wears the armband and performs gestures.

[0047] Surface electromyographic signal is an electrophysiological signal generated during muscle activity and is a time series data. The present invention directly uses time series data as a feature input model, rather than first determining the corresponding feature image based on the time series data and then inputting the model. In this way, the original time series information of the electromyographic signal can be prevented from being lost during the conversion to the feature image, allowing the model to fully utilize the original characteristics of the signal and give full play to the advantages of LSTM in time domain feature classification, especially to capture long-term dependencies.

[0048] When preprocessing EMG time series data, using sliding windows can capture the time dependency and trend of the data. Based on this, the EMG signal can be divided into multiple overlapping subsequences, which helps the model to better understand and predict the behavior of time series data. At the same time, through the overlapping sliding window method, multiple samples can be generated from the original data, and the scale of the data set can be expanded through sliding windows even if the amount of data is small.

[0049] The above is the specific process of obtaining samples. The following introduces the initialization of hyperparameter value combinations.

[0050] S1. Initialize multiple hyperparameter value combinations for the LSTM model, each hyperparameter value combination includes the number of neurons in the first layer, the number of neurons in the second layer, the dropout rate in the first layer, the dropout rate in the second layer, the learning rate, and the batch size of the LSTM model.

[0051] In this step, multiple hyperparameter value combinations are established, and a random initial value is assigned to each hyperparameter in each hyperparameter value combination.

[0052] The present invention chooses to optimize the specific hyperparameters of the LSTM model, such as the number of neurons in the first layer, the number of neurons in the second layer, the dropout rate in the first layer, the dropout rate in the second layer, the learning rate, and the batch size, for the following reasons:

[0053] The number of neurons in LSTM determines the capacity and representation ability of the model. Too many neurons allow the model to learn more complex time series patterns, but it may also lead to overfitting. Too few neurons are not enough to capture the complex relationships in the data. Therefore, only the right number of units can find a balance between the model's training ability and overfitting.

[0054] The dropout rate determines the proportion of neurons randomly discarded during training, thereby preventing the model from overfitting the training data. Therefore, a moderate dropout rate is required to balance training effect and regularization.

[0055] The learning rate controls how much the model adjusts the weights at each update. Too high a learning rate may cause unstable or divergent training, while too low a learning rate may lead to slow training or even fall into a local optimal solution. Therefore, it is necessary to adaptively adjust the learning rate in the genetic algorithm.

[0056] The batch size determines the number of data samples used for each parameter update. Larger batches can speed up the training process and may lead to more stable gradient estimates, but also require more computing resources. Smaller batches can increase the generalization ability of the model, but may lead to longer training times.

[0057] Therefore, several specific hyperparameters of the LSTM model, including the number of neurons in the first layer, the number of neurons in the second layer, the dropout rate in the first layer, the dropout rate in the second layer, the learning rate, and the batch size, are selected for optimization.

[0058] S2. For the first LSTM model configured by using any combination of hyperparameter values, first train it using the training sample set, then verify the trained first LSTM model using the verification sample set, calculate the fitness based on the verification index, and then select the hyperparameter value combination according to the fitness.

[0059] The verification index may be the overall accuracy of gesture recognition, and the higher the accuracy, the higher the adaptability.

[0060] S3. By performing crossover and mutation operations on the hyperparameter value combinations screened out in step S2, multiple new hyperparameter value combinations are obtained.

[0061] In some embodiments, a crossover operation may be performed first, and then a mutation operation may be performed.

[0062] The crossover operations include:

[0063] For any two combinations of the screened hyperparameter value combinations, some hyperparameters in one of the hyperparameter value combinations are combined with another part of the hyperparameters in another hyperparameter value combination to obtain a crossover hyperparameter value combination.

[0064] The mutation operations include:

[0065] For the crossover hyperparameter value combinations, some of the hyperparameter values ​​are randomly replaced and then included in the new multiple hyperparameter value combinations.

[0066] In some other embodiments, a mutation operation may be performed first, and then a crossover operation may be performed based on the result of the mutation operation.

[0067] S4. Repeat iterative steps S2-S3 until the hyperparameter value combination with the highest fitness converges.

[0068] There are many ways to determine whether convergence has occurred. For example, the number of iterations reaches a predetermined threshold. For another example, if the rate of change of the fitness of the hyperparameter value combination with the highest fitness is less than a preset threshold in multiple consecutive iterations, it is considered to have converged.

[0069] S5. Use the hyperparameter value combination with the highest fitness finally selected to configure the LSTM model, and train it based on the electromyographic signal sample set to obtain the final LSTM model for electromyographic signal decoding.

[0070] The above is the process of the electromyographic signal decoding method provided by the present invention. In order to verify the effect of the electromyographic signal decoding method in the gesture recognition task, an electromyographic signal decoding experiment was carried out using the LSTM model obtained by the method. The following is the experimental process and results.

[0071] The experiment uses the BandMyo dataset as the electromyographic signal sample set. The dataset includes 15 gestures, including finger gestures, wrist gestures and other gestures. When the dataset was collected, the subjects completed all 15 gestures under the guidance of the video. The sEMG signals were synchronously collected through the Myo armband worn on the forearm of the subjects. The sampling frequency was 200Hz, the number of sampling points was about 628, and the corresponding gestures were considered to be completed within 3 seconds. When all 15 gestures were completed, the subjects took off the armband to rest, and then put on the armband again to repeat the experiment. Finally, a small number of sampling points were intercepted to obtain sEMG sequence data with an average length of 624 sampling points. The experiment selected the five most common gestures in daily life from the dataset: "Spread", "Fist", "Point", "Flex" and "Extend", and used the LSTM model with specific hyperparameters optimized by genetic algorithm to decode the electromyographic signals. The gesture classification accuracy is shown in Table 1.

[0072] As can be seen from Table 1, for most subjects, the motion intention of “Extend” can be fully predicted. For some subjects (such as subject 004), the motion intention of all gestures can be fully predicted. The classification accuracy of the five gestures of subject 001 fluctuates the most.

[0073] Table 1 Classification accuracy of five gestures of LSTM model optimized by genetic algorithm for specific hyperparameters

[0074]

[0075] In order to more intuitively demonstrate the advantages of the electromyographic signal decoding method provided by the present invention, the experiment selected three traditional classifiers, namely K-nearest neighbor (KNN), support vector machine (SVM) and decision tree, as control methods, and added two gestures: "Adduct" and "Abduct". To this end, it is necessary to first extract time domain features from the data set - mean absolute value (MAV), slope sign change (SSC) and zero crossing number (Zerocrossing, ZC) as the input of these three traditional classifiers. Using four different methods, 7 gestures were decoded for electromyographic signals, and the classification results obtained by 5-fold cross validation are shown as follows. Figure 3 shown.

[0076] For most subjects, the effect of gesture recognition using the electromyographic signal decoding method provided by the present invention is better than the recognition effect of the traditional classifier on manually extracted features. In most cases, the electromyographic signal decoding method provided by the present invention can improve the accuracy of gesture recognition, especially in the complex seven-category task, the LSTM that optimizes specific parameters through genetic algorithm shows strong adaptability and performance improvement potential. In addition, the gesture recognition rate of the electromyographic signal decoding method provided by the present invention on subject No. 004 reached 100.0%, which is much higher than KNN (77.0%), SVM (86.7%), and Decision Tree (70.7%), indicating that the electromyographic signal decoding method provided by the present invention can significantly improve the accuracy of gesture recognition on subjects with lower accuracy.

[0077] The average accuracy of recognizing seven gestures using the traditional classifier, the traditional LSTM model and the method proposed in this invention is as follows: Figure 4 As shown. Figure 4 It can be seen that the average accuracy of gesture recognition of the electromyographic signal decoding method provided by the present invention is 11.2%, 3.5%, and 8.2% higher than the average accuracy of the traditional classifiers KNN, SVM, and Decision Tree, respectively, reaching 89.7%. And compared with the traditional LSTM model, the accuracy of gesture recognition is improved by 2.4%. It can be seen that the electromyographic signal decoding method provided by the present invention has better performance than the traditional method. This is because the final LSTM model obtained by the present invention has a faster training speed and can achieve a better recognition effect after parameter optimization of the genetic algorithm. In addition, the genetic algorithm can help automatically adjust the configuration of the model to adapt to the signal collection conditions of different users and different devices, making it more versatile and able to adapt well to different devices, environments, and usage scenarios, especially suitable for scenarios with limited computing resources.

[0078] It can be seen from the above that the electromyographic signal decoding method provided by the present invention can simplify the selection process of LSTM model hyperparameter values, optimize the LSTM model hyperparameters, thereby significantly improving the gesture recognition accuracy, and at the same time improve the generalization ability of the model, making it more universal.

[0079] The present invention also provides an electromyographic signal decoding device for implementing the above-mentioned electromyographic signal decoding method, the device comprising:

[0080] The sample processing module is used to obtain an electromyographic signal sample set, and divide the electromyographic signal sample set into an electromyographic signal training sample set and an electromyographic signal verification sample set based on the electromyographic signal sample set, wherein the feature of each electromyographic signal sample is the electromyographic signal time series data, and the label is the gesture category.

[0081] The initialization module is used to initialize multiple current hyperparameter value combinations of the LSTM model. Each hyperparameter value combination includes the values ​​of the number of neurons in the first layer, the number of neurons in the second layer, the dropout rate in the first layer, the dropout rate in the second layer, the learning rate, and the batch size of the LSTM model.

[0082] The model configuration module is used to obtain the first LSTM model by using any current hyperparameter value combination configuration.

[0083] A training and verification module is used to train and verify the first LSTM model and calculate the fitness based on the verification index; the training is performed based on the training sample set; and the verification is performed based on the verification sample set.

[0084] The screening module is used to screen the hyperparameter value combinations according to the fitness and input them into the update module, and to determine whether the hyperparameter value combination with the highest fitness has converged.

[0085] The update module is used to perform crossover and mutation operations on the selected hyperparameter value combinations, update the current multiple hyperparameter value combinations, and input them into the model configuration module for iteration.

[0086] The training module is used to train the hyperparameter value combination with the highest fitness to configure the LSTM model after the hyperparameter value combination with the highest fitness converges, so as to obtain the final LSTM model for electromyographic signal decoding; the training is performed based on the electromyographic signal sample set.

[0087] It should be noted that the above-mentioned device can execute the above-mentioned electromyographic signal decoding method. For the functions of each module, please refer to the above-mentioned introduction to the method and will not be described in detail.

[0088] In the description of the embodiments of the present application, words such as "exemplary", "for example" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary", "for example" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary", "for example" or "for example" is intended to present related concepts in a concrete way.

[0089] In the description of the embodiments of the present application, the term "and / or" is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, B exists alone, and A and B exist at the same time. In addition, unless otherwise specified, the term "plurality" means two or more.

[0090] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "include", "comprises", "has" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0091] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. A myoelectric signal decoding method based on a long short-term memory network improved by a genetic algorithm, characterized in that: The following steps are involved: S0, obtaining an electromyographic signal sample set, and dividing the electromyographic signal training sample set and the electromyographic signal verification sample set based on the electromyographic signal sample set, wherein the feature of each electromyographic signal sample is the electromyographic signal time series data, and the label is the gesture category; S1. Initialize multiple hyperparameter value combinations for the LSTM model, each hyperparameter value combination includes the number of neurons in the first layer, the number of neurons in the second layer, the dropout rate in the first layer, the dropout rate in the second layer, the learning rate, and the batch size of the LSTM model; S2. For the first LSTM model configured by any hyperparameter value combination, first train it by using the electromyographic signal training sample set, then verify the trained first LSTM model by using the electromyographic signal verification sample set, calculate the fitness based on the verification index, and then select the hyperparameter value combination according to the fitness; S3, obtaining multiple new hyperparameter value combinations by performing crossover and mutation operations on the hyperparameter value combinations screened out in step S2; S4, repeat iterating S2-S3 until the hyperparameter value combination with the highest fitness converges; S5. Use the hyperparameter value combination with the highest fitness finally selected to configure the LSTM model, and train it based on the electromyographic signal sample set to obtain the final LSTM model for electromyographic signal decoding.

2. The electromyographic signal decoding method according to claim 1, characterized in that: The electromyographic signal time series data is obtained by processing the corresponding original sequence data through data normalization and overlapping sliding window segmentation.

3. The electromyographic signal decoding method according to claim 1, characterized in that: The crossover operation specifically includes: For any two combinations of the screened hyperparameter value combinations, some hyperparameters in one of the hyperparameter value combinations are combined with another part of the hyperparameters in another hyperparameter value combination to obtain a crossover hyperparameter value combination.

4. The method according to claim 3, characterized in that The mutation operation specifically includes: For the crossover hyperparameter value combinations, some of the hyperparameter values ​​are randomly replaced and then included in the new multiple hyperparameter value combinations.

5. A myoelectric signal decoding device, characterized in that: The device comprises: A sample processing module, used for acquiring an electromyographic signal sample set, and dividing the electromyographic signal training sample set and the electromyographic signal verification sample set based on the electromyographic signal sample set, wherein the feature of each electromyographic signal sample is the electromyographic signal time series data, and the label is the gesture category; The initialization module is used to initialize multiple current hyperparameter value combinations of the LSTM model. Each hyperparameter value combination includes the number of neurons in the first layer, the number of neurons in the second layer, the dropout rate in the first layer, the dropout rate in the second layer, the learning rate, and the batch size of the LSTM model. A model configuration module, used to obtain a first LSTM model by using any current hyperparameter value combination configuration; A training and verification module, used for training and verifying the first LSTM model, and calculating fitness based on a verification indicator; the training is performed based on the training sample set; the verification is performed based on the verification sample set; A screening module is used to screen the hyperparameter value combinations according to fitness and input them into the update module, and to determine whether the hyperparameter value combination with the highest fitness has converged; The update module is used to perform crossover and mutation operations on the selected hyperparameter value combinations, update the current multiple hyperparameter value combinations, and input them into the model configuration module for iteration; The training module is used to train the hyperparameter value combination with the highest fitness to configure the LSTM model after the hyperparameter value combination with the highest fitness converges, so as to obtain the final LSTM model for electromyographic signal decoding; the training is performed based on the electromyographic signal sample set.

6. The electromyographic signal decoding device according to claim 5, characterized in that: The sample processing module includes a preprocessing unit, which is used to perform data normalization and overlapping sliding window segmentation processing on the original data of the electromyographic signal time series to obtain the electromyographic signal time series data.

7. The electromyographic signal decoding device according to claim 5, characterized in that: The updating module includes a crossover unit, which is used to combine part of the hyperparameters in one of the hyperparameter value combinations with another part of the hyperparameters in another hyperparameter value combination for any two combinations of the screened hyperparameter value combinations to obtain a crossed hyperparameter value combination, and include the crossed hyperparameter value combination into the updated multiple hyperparameter value combinations.

8. The electromyographic signal decoding device according to claim 7, characterized in that: The updating module includes a mutation unit, which is used to randomly replace the values ​​of some hyperparameters in the hyperparameter value combinations after the crossover, and then classify them into the updated multiple hyperparameter value combinations.

Images