A training method and a recognition method of a motion intention recognition model, a device, an electronic device, a storage medium and a computer program product

Abstract

The present disclosure relates to a training method and a recognition method of a motion intention recognition model, a device, an electronic device, a storage medium and a computer program product, comprising: inputting MRCP features, alpha and beta band features and electromyographic features into a pre-processing module respectively to obtain pre-processing features; inputting the pre-processing features into a classifier to obtain a classification result; calculating an independent loss based on the classification result and a motion intention true label; calculating a fusion feature based on the pre-processing features under each branch and a fusion weight; inputting the fusion feature into the classifier to obtain a fusion classification result; calculating a fusion loss based on the fusion classification result and the motion intention true label; and adjusting parameters of the motion intention recognition model based on the independent loss and the fusion loss under each branch. In this way, the fusion weights corresponding to various types of electrophysiological signals can be dynamically adjusted, so that the advantages of various types of electrophysiological signals can be flexibly integrated.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and more specifically, to training methods, recognition methods, devices, electronic devices, storage media, and computer program products for motion intention recognition models. Background Technology

[0002] Neurological injuries often lead to severe motor dysfunction, such as muscle weakness and spasticity, impairing fine motor control and making it difficult for patients to perform basic daily activities like eating, dressing, and writing. Furthermore, the severity of these motor disorders varies considerably, ranging from partial weakness to near-complete paralysis. Therefore, developing adaptive decoding methods that can reliably interpret motor intentions across different degrees of injury to support personalized rehabilitation and assistance has become a current research focus.

[0003] Electrophysiological decoding methods offer hope for assisting and restoring motor function, their core being the direct inference of a patient's motor intention from electrophysiological signals. However, as mentioned earlier, the severity of motor dysfunction varies significantly, posing a considerable challenge to reliable decoding. For example, for individuals with partially preserved motor output capacity, electromyography (EMG) can provide clearly distinguishable hand motor features; while for individuals with severe impairment of motor output capacity, when muscle activation is absent, electroencephalography (EEG) becomes a more reliable source of motor intention. This cross-individual variability highlights the limitations of unimodal and non-adaptive decoding models, which often struggle to achieve effective generalization across different degrees of impairment.

[0004] In related technologies, unimodal and multimodal fusion strategies have emerged for recognizing patients' motor intentions. However, unimodal strategies often fail in severely ill patients, while multimodal fusion strategies generally lack adaptive mechanisms and cannot adjust the relative contributions of EEG and EMG according to the severity of the patient's motor impairment, thus making it difficult to cover patients with different degrees of motor impairment. Summary of the Invention

[0005] This disclosure provides training methods, recognition methods, devices, electronic devices, storage media, and computer program products for motion intention recognition models, to at least solve the problem in the aforementioned related technologies that multimodal fusion strategies cannot adjust the relative contributions of EEG and EMG according to the degree of motion impairment of patients, thus making it difficult to cover patients with different degrees of motion impairment.

[0006] According to a first aspect of the present disclosure, a training method for a motion intention recognition model is provided. The motion intention recognition model includes a preprocessing module and a classifier. The training method includes: acquiring motion-related feature samples, wherein the motion-related feature samples include motion-related cortical potential (MRCP) features, α, β band features, and electromyographic (EMG) features, and the motion-related feature samples correspond to real labels of motion intentions; inputting the MRCP features, the α, β band features, and the EMG features into the preprocessing module under the corresponding branch to obtain preprocessed features under the corresponding branch; and inputting the preprocessed features under the corresponding branch into the classifier to obtain classification results under the corresponding branch. Based on the classification results under the corresponding branch and the true motion intention label, calculate the independent loss under the corresponding branch; based on the preprocessing features under each branch and the fusion weights of each branch, calculate the fusion features, wherein the fusion weights of each branch are obtained based on the independent losses of each branch in the previous training process of the current training; input the fusion features into the classifier to obtain the fusion classification result; based on the fusion classification result and the true motion intention label, calculate the fusion loss; based on the independent losses under each branch and the fusion loss, calculate the total loss; adjust the parameters of the motion intention recognition model based on the total loss to train the motion intention recognition model.

[0007] Optionally, acquiring motion-related feature samples includes: acquiring EEG data samples and EMG data samples; predicting the motion initiation time based on the EMG data samples; and extracting the MRCP features from a first preset window of the EEG data samples, extracting the α and β frequency band features from a second preset window of the EEG data samples, and extracting the EMG features from a third preset window of the EMG data samples, based on the motion initiation time, wherein the first preset window and the second preset window have different time lengths.

[0008] Optionally, the fusion weights of the respective branches are calculated using the following formula:

[0009]

[0010] in, This represents the fusion weight of the i-th branch in the current training iteration. This represents the independent loss of the i-th branch in the previous training iteration. This represents the independent loss of the j-th branch in the previous training iteration. It is a constant.

[0011] Optionally, the fusion weights of each branch during the initial training process can be calculated using the following formula:

[0012]

[0013]

[0014] in, This represents the fusion weight of the branch corresponding to the electromyographic feature during the initial training process. This represents the fusion weight of the branch corresponding to the MRCP feature during the initial training process. This represents the fusion weights of the branches corresponding to the features of the α and β frequency bands during the initial training process. This represents the Fugl-Meyer hand assessment subscale score of the identified object.

[0015] Optionally, the training method further includes: calculating the regularization term loss; the calculation of the total loss based on the independent loss under each branch and the fusion loss includes: calculating the total loss based on the independent loss under each branch, the fusion loss, and the regularization term loss; wherein the regularization term loss is calculated using the following formula:

[0016]

[0017] in, This represents the loss of the regularization term. This indicates the number of motion-related feature samples contained in a single training batch. Indicates the first A number of motion-related feature samples in the branch The fusion weights below Indicates a constant.

[0018] According to a second aspect of the present disclosure, a motion intention recognition method based on a motion intention recognition model is provided. The motion intention recognition model is trained using a training method according to the present disclosure. The motion intention recognition method includes: acquiring target motion-related features of the target object, wherein the target motion-related features include target motion-related cortical potential (MRCP) features, target α and β frequency band features, and target electromyographic (EMG) features; inputting the target MRCP features, the target α and β frequency band features, and the target EMG features into the preprocessing module under the corresponding branch, respectively, to obtain target preprocessing features under the corresponding branch; and inputting the corresponding... The target preprocessing features under each branch are input into the classifier to obtain independent classification results for the corresponding branches. These independent classification results include the predicted motion category and its corresponding confidence level. Based on the target preprocessing features under each branch and the target fusion weights of each branch, target fusion features are calculated. The target fusion weights of each branch are the fusion weights of each branch obtained after the motion intent recognition model has been trained. The target fusion features are input into the classifier to obtain target fusion classification results. The motion category corresponding to the highest confidence level among the independent classification results under each branch and the target fusion classification results is taken as the motion intent of the identified object.

[0019] According to a third aspect of the present disclosure, a training apparatus for a motion intention recognition model is provided. The motion intention recognition model includes a preprocessing module and a classifier. The training apparatus includes: a sample acquisition module configured to acquire motion-related feature samples, wherein the motion-related feature samples include motion-related cortical potential (MRCP) features, α, β band features, and electromyographic (EMG) features, and the motion-related feature samples correspond to real labels of motion intention; an independent loss acquisition module configured to input the MRCP features, the α, β band features, and the EMG features into the preprocessing module under the corresponding branch to obtain preprocessed features under the corresponding branch; input the preprocessed features under the corresponding branch into the classifier to obtain classification results under the corresponding branch; and based on the phase... The system calculates the independent loss for each branch based on the classification results under that branch and the true label of the motion intent. A fusion loss acquisition module is configured to calculate fusion features based on the preprocessing features of each branch and the fusion weights of each branch, wherein the fusion weights of each branch are obtained based on the independent losses of each branch during the previous training process. The fusion features are input into the classifier to obtain the fusion classification result. A fusion loss is calculated based on the fusion classification result and the true label of the motion intent. A total loss calculation module is configured to calculate the total loss based on the independent losses of each branch and the fusion loss. A parameter adjustment module is configured to adjust the parameters of the motion intent recognition model based on the total loss to train the motion intent recognition model.

[0020] Optionally, the sample acquisition module is configured to: acquire EEG data samples and EMG data samples; predict the start time of movement based on the EMG data samples; and extract the MRCP features from a first preset window of the EEG data samples, extract the α and β frequency band features from a second preset window of the EEG data samples, and extract the EMG features from a third preset window of the EMG data samples, using the start time of movement as a reference, wherein the first preset window and the second preset window have different time lengths.

[0021] Optionally, the fusion weights of the respective branches are calculated using the following formula:

[0022]

[0023] in, This represents the fusion weight of the i-th branch in the current training iteration. This represents the independent loss of the i-th branch in the previous training iteration. This represents the independent loss of the j-th branch in the previous training iteration. It is a constant.

[0024] Optionally, the fusion weights of each branch during the initial training process can be calculated using the following formula:

[0025]

[0026]

[0027] in, This represents the fusion weight of the branch corresponding to the electromyographic feature during the initial training process. This represents the fusion weight of the branch corresponding to the MRCP feature during the initial training process. This represents the fusion weights of the branches corresponding to the features of the α and β frequency bands during the initial training process. This represents the Fugl-Meyer hand assessment subscale score of the identified object.

[0028] Optionally, the training device further includes: a regularization term loss calculation module configured to calculate the regularization term loss; and a total loss calculation module configured to calculate the total loss based on the independent loss under each branch, the fusion loss, and the regularization term loss.

[0029] The regularization term loss is calculated using the following formula:

[0030]

[0031] in, This represents the loss of the regularization term. This indicates the number of motion-related feature samples contained in a single training batch. Indicates the first A number of motion-related feature samples in the branch The fusion weights below Indicates a constant.

[0032] According to a fourth aspect of the present disclosure, a motion intention recognition device based on a motion intention recognition model is provided. The motion intention recognition model is trained using a training method according to the present disclosure. The motion intention recognition device includes: a feature acquisition module configured to acquire target motion-related features of the object to be recognized, wherein the target motion-related features include target motion-related cortical potential (MRCP) features, target α and β frequency band features, and target electromyographic (EMG) features; and an independent classification result acquisition module configured to input the target MRCP features, the target α and β frequency band features, and the target EMG features into the preprocessing module under the corresponding branch, respectively, to obtain target preprocessing features under the corresponding branch; and to process the corresponding branch... The target preprocessing features are input into the classifier to obtain independent classification results for each branch, wherein the independent classification results include the predicted motion category and the corresponding confidence level; the target fusion classification result acquisition module is configured to calculate target fusion features based on the target preprocessing features under each branch and the target fusion weights of each branch, wherein the target fusion weights of each branch are the fusion weights of each branch obtained after the motion intent recognition model has been trained; the target fusion features are input into the classifier to obtain the target fusion classification result; the motion intent determination module is configured to take the motion category corresponding to the highest confidence level among the independent classification results under each branch and the target fusion classification results as the motion intent of the object to be identified.

[0033] According to a fifth aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement a training method for a motion intention recognition model according to the present disclosure or a motion intention recognition method based on a motion intention recognition model.

[0034] According to a sixth aspect of the present disclosure, a computer-readable storage medium is provided, wherein when instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform a training method for a motion intention recognition model according to the present disclosure or a motion intention recognition method based on a motion intention recognition model.

[0035] According to a seventh aspect of the present disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements a training method for a motion intention recognition model according to the present disclosure or a motion intention recognition method based on a motion intention recognition model.

[0036] The technical solutions provided by the embodiments of this disclosure have at least the following beneficial effects:

[0037] In this disclosure, during the training of the motion intention recognition model, the fusion weights corresponding to various types of electrophysiological signals can be dynamically adjusted, thereby flexibly integrating the advantages of various types of electrophysiological signals. In this way, the trained motion intention recognition model can control the contribution of various types of electrophysiological signals in the motion intention recognition task according to the trained fusion weights, fully leveraging the advantages of various types of electrophysiological signals, thereby improving the accuracy and generalization ability of motion intention recognition, and effectively covering recognition objects with different degrees of movement impairment.

[0038] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0039] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0040] Figure 1 This is a flowchart illustrating a training method for a motion intent recognition model according to an exemplary embodiment of the present disclosure;

[0041] Figure 2 This is a schematic diagram illustrating a framework of a brain-muscle fusion motor intention decoding model according to exemplary embodiments of the present disclosure;

[0042] Figure 3 This is a flowchart illustrating a motion intent recognition method based on a motion intent recognition model according to an exemplary embodiment of the present disclosure;

[0043] Figure 4 This is a block diagram illustrating a training apparatus for a motion intent recognition model according to an exemplary embodiment of the present disclosure;

[0044] Figure 5 This is a block diagram illustrating a motion intention recognition device based on a motion intention recognition model according to an exemplary embodiment of the present disclosure;

[0045] Figure 6 This is a block diagram illustrating an electronic device according to exemplary embodiments of the present disclosure. Detailed Implementation

[0046] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0047] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following examples do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0048] It should be noted that the phrase "at least one of several items" in this disclosure refers to three parallel cases: "any one of the several items", "a combination of any number of the several items", and "all of the several items". For example, "including at least one of A and B" includes the following three parallel cases: (1) including A; (2) including B; (3) including A and B. Another example is "performing at least one of step one and step two", which means the following three parallel cases: (1) performing step one; (2) performing step two; (3) performing both step one and step two.

[0049] Figure 1 This is a flowchart illustrating a training method for a motion intent recognition model according to an exemplary embodiment of the present disclosure. The motion intent recognition model may include a preprocessing module and a classifier.

[0050] Reference Figure 1 In step 101, motion-related feature samples can be obtained, which may include motion-related cortical potential (MRCP) features, α and β band features and electromyographic features, and the motion-related feature samples may correspond to a true label of motion intention.

[0051] According to an exemplary embodiment of this disclosure, electroencephalogram (EEG) data samples and electromyogram (EMG) data samples can be acquired. Then, the start time of movement can be predicted based on the EMG data samples. Next, based on the start time of movement, MRCP features can be extracted from a first preset window (e.g., a [-1, 2] s window) of the EEG data samples, α and β frequency band features can be extracted from a second preset window (e.g., a [0, 3] s window) of the EEG data samples, and EMG features can be extracted from a third preset window (e.g., a [0, 3] s window) of the EMG data samples.

[0052] In this disclosure, MRCP components can be extracted from the [-1, 2] s window of the EEG data sample based on the detected onset position to characterize the motor planning phase; and electromyographic features can be extracted from the [0, 3] s window of the EMG data sample, as well as α and β frequency band components can be extracted from the [0, 3] s window of the EEG data sample to capture the characteristics of the motor execution phase.

[0053] It should be noted that, in this disclosure, for EEG data, firstly, a 50 Hz notch filter can be used to remove power frequency interference. Then, a fourth-order Butterworth high-pass filter with a cutoff frequency of 0.1 Hz can be applied to eliminate low-frequency noise. Next, a second-order Butterworth bandpass filter can be used to extract MRCP components in the range of [0.5, 5] Hz and motion-related α and β frequency band components in the range of [8, 30] Hz. For EMG data, a fourth-order Butterworth bandpass filter in the range of [20, 450] Hz can be used to capture relevant muscle activity, and it can be further downsampled to 256 Hz to maintain consistency with the EEG sampling rate.

[0054] According to exemplary embodiments of this disclosure, high-precision motion initiation point identification can be achieved using multi-channel EMG signals. For example, the input can be a segmented EMG signal preprocessed according to an experimental paradigm. Where C is the number of channels and T is the number of sampling points for each trial segment.

[0055] First, the instantaneous energy characteristics of EMG signals can be enhanced using the Teager-Kaiser energy operator (TKEO):

[0056]

[0057] in, For time point n of channel c, This represents the TKEO energy value corresponding to that point.

[0058] Then, the Constant False-Alarm Rate (CFAR) method can be introduced to achieve adaptive threshold detection. Unlike traditional methods that rely on empirically setting the sensitivity factor α, this disclosure can preset the false alarm probability (PFA=0.05) and combine it with CFAR window parameters. Calculate the value of α:

[0059]

[0060] in, For hyperparameters, This indicates that for a preset false alarm probability (PFA), take ( ) times.

[0061] Next, we can proceed by monitoring the protected areas. Training windows on both sides The data is averaged to estimate the noise. :

[0062]

[0063] in, For hyperparameters, For time points The corresponding TKEO energy value.

[0064] Then, the adaptive threshold can be calculated. :

[0065]

[0066] Next, for each channel, the energy point that first exceeds the adaptive threshold can be... It was determined to be a candidate starting point.

[0067] Then, the median of multiple candidate starting points corresponding to all channels can be used as the final motion starting point. This strategy effectively improves the accuracy and robustness of motion starting point detection, thus providing reliable assurance for cross-individual motion intent decoding.

[0068] Figure 2 This is a schematic diagram illustrating a framework of a brain-muscle electrofusion motor intention decoding model according to an exemplary embodiment of the present disclosure. (Refer to...) Figure 2 The motion intention recognition model provided in this disclosure can include four branches: three independent branches and one fusion branch. The three independent branches can be: MRCP branch, α and β band branch and electromyography branch.

[0069] For the MRCP branch, the input can be EEG-MRCP components in the range of [0.5, 5] Hz (EEG-MRCP[0.5, 5]Hz) as described above. This branch can contain sequentially arranged temporal convolutional modules, spatial convolutional modules, separateable pointwise convolutional modules, stacked temporal convolution networks (TCN), and an MRCP classifier. Furthermore, all modules in this MRCP branch except the MRCP classifier can be considered as "preprocessing modules" within this branch. Additionally, the TCN consists of dilated causal convolution, batch normalization, ReLU activation, dropout, and residual connections. Moreover, by employing exponentially increasing dilation rates (1, 2, and 4), long-range temporal dependencies can be effectively captured, making it suitable for decoding dynamic EEG activity.

[0070] For the α and β band branches, the input can be motion-related α and β band components (EEG-αβ[8, 30]Hz) within the range of [8, 30] Hz as described above. Similar to the MRCP branch, this α and β band branch can also contain sequentially arranged temporal, spatial, separateable, and stacked TCNs, as well as α and β classifiers (αβ_classifier). Furthermore, all modules in this α and β band branch except for the α and β classifiers can be regarded as "preprocessing modules" under this branch.

[0071] For the electromyography (EMG) branch, the input can be the EMG features (EMG[20, 450] Hz) in the range of [20, 450] Hz as described above. This EMG branch can contain sequentially arranged temporal features, stacked TCNs, and an EMG classifier. Furthermore, all modules in this EMG branch except for the EMG classifier can be regarded as "preprocessing modules" under this branch.

[0072] The fusion branch mainly includes a fusion module and a fusion classifier. The fusion module is mainly used to fuse the feature processing results of the three independent branches to obtain fused features.

[0073] In step 102, MRCP features, α and β band features, and electromyography features can be input into the preprocessing module of the corresponding branch to obtain the preprocessed features for that branch. Then, the preprocessed features of the corresponding branch can be input into the classifier to obtain the classification result for that branch. Next, based on the classification result and the true motion intent label of the corresponding branch, the independent loss for that branch can be calculated.

[0074] It should be noted that EEG data can be divided into MRCP branches and alpha / beta frequency branches. These two branches are chosen due to their complementary advantages in motor decoding. MRCP, as motor cortical potentials closely related to motor preparation and initiation, can reveal motor intentions in advance, thereby facilitating a faster response. In contrast, alpha / beta rhythms are the most widely used oscillatory features in motor-related research, reflecting sensorimotor regulation encompassing motor planning and execution. By combining the two, a balanced representation from motor planning to motor execution can be provided.

[0075] For example, for either the MRCP branch or the α / β band branch, to transform the neurophysiological features of the corresponding branch into discriminative features, temporal convolution can be used first to capture short-term temporal dynamics. Then, spatial convolution can be used to highlight the temporal features within a channel while maintaining the spatial independence between channels. Next, separate convolution can be used to integrate the information extracted by spatial convolution across channels, achieving cross-channel feature interaction with low computational overhead. The extracted features can then be input into a stacked TCN. As mentioned earlier, this TCN consists of dilated causal convolution, batch normalization, ReLU activation, dropout, and residual connections. Furthermore, by employing exponentially increasing dilation rates (1, 2, and 4), long-term temporal dependencies can be effectively captured, making it suitable for decoding dynamic EEG activity.

[0076] The electromyography (EMG) branch primarily captures muscle activation signals that directly reflect motor execution. Unlike electroencephalography (EEG), which mainly encodes cortical motor planning and preparation, EMG data provides peripheral measurements of neuromuscular output. It boasts high temporal resolution and strong discriminative power for executed movements. Notably, for subjects with relatively preserved motor function, EMG typically exhibits a higher signal-to-noise ratio than EEG, making it a reliable modality for decoding fine motor intentions. Furthermore, to effectively extract these features, this EMG branch primarily employs a temporal coding structure to model high-resolution temporal patterns. Moreover, due to the relatively limited number of EMG channels and restricted spatial interactions, temporal dynamics become the primary source of discriminative information.

[0077] For example, firstly, the electromyographic (EMG) signal can be processed through a convolutional layer. Then, batch normalization and ReLU activation can be performed. Next, temporal features can be captured across scales by stacking three dilated TCNs with dilation rates of 1, 2, and 4. Similar to the EEG branch, residual connections are introduced into each module in the EMG branch to ensure training stability.

[0078] In step 103, fusion features can be calculated based on the preprocessing features of each branch and the fusion weights of each branch. The fusion weights of each branch can be calculated based on the independent losses of each branch in the previous training process. Then, the fusion features can be input into the classifier to obtain the fusion classification result. Next, the fusion loss can be calculated based on the fusion classification result and the true motion intent label.

[0079] It should be noted that, in order to ensure that the fusion strategy can dynamically reflect the relative reliability of each modality during training, this disclosure introduces an adaptive modality fusion weighting mechanism based on training loss. Unlike methods that learn gating weights through additional trainable parameters, this disclosure adopts a lightweight and interpretable strategy: updating the fusion weights at the epoch level based on the inverse of the training loss.

[0080] For example, suppose , and Let each represent an independent training loss for one of the three modalities. Then, based on these three independent training losses obtained in the current training, we can calculate the fusion weights for each of the three modalities in the next epoch.

[0081] According to an exemplary embodiment of this disclosure, the fusion weight of each branch can be calculated using the following formula:

[0082]

[0083] in, This represents the fusion weight of the i-th branch in the current training iteration. This represents the independent loss of the i-th branch in the previous training iteration. This represents the independent loss of the j-th branch in the previous training iteration. For example, a constant. It is a small constant used to ensure numerical stability.

[0084] In this way, the fusion weights are normalized so that their sum is 1, and can be applied to both feature-level and decision-level fusion operations in the next training epoch. Furthermore, this update mechanism emphasizes more reliable modalities, i.e., those with lower training loss, while suppressing poorly performing modalities.

[0085] It should be noted that, in order to introduce individualized prior knowledge during the fusion process, this disclosure also designs a fusion weight initialization scheme guided by a clinical motor assessment scale. For example, the fusion weights for the three modalities (EEG-MRCP, EEG-αβ, and EMG) can be initialized using the Fugl-Meyer Motor Assessment (FMA) hand component score (range 0 to 14) of the identified subject. For instance, the fusion weight for EMG can be linearly scaled by dividing the FMA score by 14, while the fusion weights for the other branches can be evenly distributed between the two EEG components.

[0086] According to an exemplary embodiment of this disclosure, the fusion weights of each branch during the initial training process can be calculated using the following formula:

[0087]

[0088]

[0089] in, This represents the fusion weights of the branches corresponding to electromyographic features (EMG branches) during the initial training process. This represents the fusion weights of the branch corresponding to the MRCP feature (EEG-MRCP branch) during the initial training process. This represents the fusion weights of the branches (EEG-αβ branches) corresponding to the features of the α and β frequency bands during the initial training process. This indicates the Fugl-Meyer Motor Assessment (FMA) hand component score of the identified object.

[0090] In step 104, the total loss can be calculated based on the independent loss and fusion loss under each branch. It should be noted that, in order to regulate the optimization process and avoid the model from over-relying on a single mode, this disclosure further introduces an entropy-based regularization term.

[0091] According to exemplary embodiments of this disclosure, regularization term loss can also be calculated. Then, the total loss can be calculated based on the independent loss, fusion loss, and regularization term loss under each branch.

[0092] The regularization loss can be calculated using the following formula:

[0093]

[0094] in, This represents the loss due to regularization. This indicates the number of motion-related feature samples contained in a single training batch. Indicates the first A number of motion-related feature samples in the branch The fusion weights below Indicates a constant.

[0095] It's important to note that this entropy regularization term primarily encourages balanced contributions between modalities by penalizing low-entropy distributions. For example, when a modality becomes dominant, i.e., its fusion weights approach 1, its entropy will approach zero, thus increasing the regularization penalty. This mechanism effectively prevents modality collapse in the early stages of training and promotes dynamic adaptation based on the performance of each modality during training. Furthermore, this mechanism complements FMA-based fusion weight initialization: although subjects with high FMA scores receive higher EMG weights during fusion weight initialization, entropy regularization prevents the model from prematurely converging to rely solely on EMG and supports adaptive rebalancing of fusion weights during training.

[0096] For example, the total training loss of the model can be defined as:

[0097]

[0098] in, This represents the total loss of the model. Indicates fusion loss, Indicates the independent loss of the EEG-MRCP branch. This represents the independent loss of the EEG-αβ branch. This represents the independent loss of the EMG branch. This represents the loss due to regularization. and As preset weights, and, It can be set to 0.2. It can be set to 0.1.

[0099] In this disclosure, in addition to the fusion representation, modality-specific classification branches such as the EEG-MRCP branch, EEG-αβ branch, and EMG branch are retained. Inspired by the residual connectivity paradigm, these branches allow each modality to output predictions independently, thus preserving discriminative features that might be diluted during the fusion process. This not only enables the model to fully leverage the strengths of each modality but also enhances overall robustness when a particular modality is more informative for a specific subject or trial.

[0100] In step 105, the parameters of the motion intent recognition model can be adjusted based on the total loss to train the motion intent recognition model.

[0101] Figure 3 This is a flowchart illustrating a motion intent recognition method based on a motion intent recognition model according to an exemplary embodiment of the present disclosure, the motion intent recognition model being trained by a training method for a motion intent recognition model according to the present disclosure.

[0102] Reference Figure 3 In step 301, target motion-related features of the identified object can be acquired. These target motion-related features may include target motion-related cortical potential (MRCP) features, target α and β frequency band features, and target electromyographic features. Furthermore, the methods for acquiring these three modal features have been described in detail in the previous embodiment and will not be repeated here.

[0103] In step 302, the target MRCP features, target α and β band features, and target electromyography features can be input into the preprocessing module of the corresponding branch to obtain the target preprocessing features of the corresponding branch. Then, the target preprocessing features of the corresponding branch can be input into the classifier to obtain the independent classification result of the corresponding branch, wherein the independent classification result can include the predicted motion category and the corresponding confidence level.

[0104] In step 303, target fusion features can be calculated based on the target preprocessing features under each branch and the target fusion weights of each branch. The target fusion weights of each branch can be the fusion weights of each branch obtained after the motion intent recognition model has been trained. Then, the target fusion features can be input into a classifier to obtain the target fusion classification result.

[0105] In step 304, the motion category corresponding to the highest confidence in the independent classification results under each branch and the target fusion classification results can be used as the motion intent of the identified object.

[0106] For example, suppose , , , These represent the categories (logits) predicted by the EEG-MRCP branch, EEG-αβ branch, EMG branch, and fusion branch, respectively, where C represents the number of action categories. For example, C can represent the number of gesture categories of the identified object. For instance, the identified object's gesture categories could include 5 types: "gesturing to the number 1 with a finger," "gesturing to the number 2 with a finger," "gesturing to the number 3 with a finger," "gesturing to the number 4 with a finger," and "gesturing to the number 5 with a finger."

[0107] Furthermore, the confidence score for the predicted motion category for each branch can be a softmax-based confidence score, and its formula can be as follows:

[0108] .

[0109] Then, the motion category corresponding to the highest confidence level among the three categories of the three independent branches and the category confidence level under the fusion branch can be used as the motion intent of the object to be identified:

[0110]

[0111] in, This represents the prediction confidence, or prediction probability, of branch i in motion category c.

[0112] This confidence-guided decision-making strategy allows the model to fall back to a more reliable single-modal prediction when there is uncertainty in the fusion output, thus achieving an effective combination of fusion and competition between modalities.

[0113] Figure 4 This is a block diagram illustrating a training apparatus 400 for a motion intent recognition model according to an exemplary embodiment of the present disclosure. The motion intent recognition model may include a preprocessing module and a classifier.

[0114] Reference Figure 4 The training device 400 for the motion intention recognition model may include a sample acquisition module 401, an independent loss acquisition module 402, a fusion loss acquisition module 403, a total loss calculation module 404, and a parameter adjustment module 405.

[0115] The sample acquisition module 401 can acquire motion-related feature samples, which may include MRCP features, α and β frequency band features and electromyographic features, and may correspond to a true label of motion intention.

[0116] According to an exemplary embodiment of this disclosure, the sample acquisition module 401 can acquire EEG data samples and EMG data samples. Then, the sample acquisition module 401 can predict the start time of movement based on the EMG data samples. Next, the sample acquisition module 401 can extract MRCP features from a first preset window (e.g., a [-1, 2] s window) of the EEG data samples, extract α and β frequency band features from a second preset window (e.g., a [0, 3] s window) of the EEG data samples, and extract EMG features from a third preset window (e.g., a [0, 3] s window) of the EMG data samples.

[0117] The independent loss acquisition module 402 can input MRCP features, α and β band features, and electromyography features into the preprocessing module of the corresponding branch to obtain the preprocessed features for that branch. Then, the preprocessed features of the corresponding branch can be input into the classifier to obtain the classification result for that branch. Next, based on the classification result and the true label of the motion intent, the independent loss for that branch can be calculated.

[0118] The fusion loss acquisition module 403 can calculate the fusion features based on the preprocessed features of each branch and the fusion weights of each branch. The fusion weights of each branch can be obtained by calculating the independent losses of each branch in the previous training process. Then, the fusion features can be input into the classifier to obtain the fusion classification result. Next, the fusion loss can be calculated based on the fusion classification result and the true motion intent label.

[0119] According to an exemplary embodiment of this disclosure, the fusion weight of each branch can be calculated using the following formula:

[0120]

[0121] in, This represents the fusion weight of the i-th branch in the current training iteration. This represents the independent loss of the i-th branch in the previous training iteration. This represents the independent loss of the j-th branch in the previous training iteration. For example, a constant. It is a small constant used to ensure numerical stability.

[0122] According to an exemplary embodiment of this disclosure, the fusion weights of each branch during the initial training process can be calculated using the following formula:

[0123]

[0124]

[0125] in, This represents the fusion weights of the branches corresponding to electromyographic features (EMG branches) during the initial training process. This represents the fusion weights of the branch corresponding to the MRCP feature (EEG-MRCP branch) during the initial training process. This represents the fusion weights of the branches (EEG-αβ branches) corresponding to the features of the α and β frequency bands during the initial training process. This indicates the Fugl-Meyer Motor Assessment (FMA) hand component score of the identified object.

[0126] The total loss calculation module 404 can calculate the total loss based on the independent loss and fusion loss under each branch.

[0127] According to an exemplary embodiment of this disclosure, the training apparatus 400 may further include a regularization term loss calculation module.

[0128] The regularization term loss calculation module can calculate the regularization term loss. Then, the total loss calculation module 404 can calculate the total loss based on the independent loss, fusion loss, and regularization term loss under each branch.

[0129] The regularization loss can be calculated using the following formula:

[0130]

[0131] in, This represents the loss due to regularization. This indicates the number of motion-related feature samples contained in a single training batch. Indicates the first A number of motion-related feature samples in the branch The fusion weights below Indicates a constant.

[0132] The parameter adjustment module 405 can adjust the parameters of the motion intention recognition model based on the total loss in order to train the motion intention recognition model.

[0133] Figure 5 This is a block diagram illustrating a motion intention recognition device 500 based on a motion intention recognition model according to an exemplary embodiment of the present disclosure. The motion intention recognition model can be trained using a training method for a motion intention recognition model according to the present disclosure.

[0134] Reference Figure 5 The motion intent recognition device 500 may include a feature acquisition module 501, an independent classification result acquisition module 502, a target fusion classification result acquisition module 503, and a motion intent determination module 504.

[0135] The feature acquisition module 501 can acquire the target motion-related features of the identified object, wherein the target motion-related features may include target motion-related cortical potential (MRCP) features, target α and β frequency band features, and target electromyographic features.

[0136] The independent classification result acquisition module 502 can input the target MRCP features, target α and β frequency band features, and target electromyographic features into the preprocessing module under the corresponding branch to obtain the target preprocessing features under the corresponding branch. Then, the target preprocessing features under the corresponding branch can be input into the classifier to obtain the independent classification result under the corresponding branch, wherein the independent classification result can include the predicted motion category and the corresponding confidence level.

[0137] The target fusion classification result acquisition module 503 can calculate target fusion features based on the target preprocessing features under each branch and the target fusion weights of each branch. The target fusion weights of each branch can be the fusion weights of each branch obtained after the motion intent recognition model has been trained. Then, the target fusion features can be input into the classifier to obtain the target fusion classification result.

[0138] The motion intent determination module 504 can use the motion category with the highest confidence among the independent classification results under each branch and the target fusion classification results as the motion intent of the object to be identified.

[0139] Figure 6 This is a block diagram illustrating an electronic device 600 according to an exemplary embodiment of the present disclosure.

[0140] Reference Figure 6 The electronic device 600 includes at least one memory 601 and at least one processor 602. The at least one memory 601 stores a set of computer-executable instructions. When the set of computer-executable instructions is executed by the at least one processor 602, a training method for a motion intention recognition model according to an exemplary embodiment of the present disclosure or a motion intention recognition method based on the motion intention recognition model is executed.

[0141] As an example, electronic device 600 may be a PC, tablet, personal digital assistant, smartphone, or other device capable of executing the aforementioned set of instructions. Here, electronic device 600 is not necessarily a single electronic device, but may be a collection of any devices or circuits capable of executing the aforementioned instructions (or instruction sets) individually or in combination. Electronic device 600 may also be part of an integrated control system or system manager, or may be configured to interconnect with a portable electronic device locally or remotely (e.g., via wireless transmission) through an interface.

[0142] In electronic device 600, processor 602 may include a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller, or a microprocessor. By way of example and not limitation, processor may also include analog processors, digital processors, microprocessors, multi-core processors, processor arrays, network processors, etc.

[0143] The processor 602 can execute instructions or code stored in the memory 601, which can also store data. Instructions and data can also be sent and received over a network via a network interface device, which can employ any known transmission protocol.

[0144] The memory 601 may be integrated with the processor 602, for example, by placing RAM or flash memory within an integrated circuit microprocessor. Alternatively, the memory 601 may include a separate device, such as an external disk drive, a storage array, or other storage device that can be used by any database system. The memory 601 and the processor 602 may be operatively coupled, or may communicate with each other, for example, via I / O ports, network connections, etc., enabling the processor 602 to read files stored in the memory.

[0145] In addition, the electronic device 600 may also include a video display (such as a liquid crystal display) and a user interaction interface (such as a keyboard, mouse, touch input device, etc.). All components of the electronic device 600 can be interconnected via a bus and / or network.

[0146] According to exemplary embodiments of the present disclosure, a computer-readable storage medium storing instructions may also be provided, wherein when the instructions are executed by at least one processor, they cause at least one processor to perform a training method for a motion intention recognition model according to the present disclosure or a motion intention recognition method based on the motion intention recognition model. Examples of computer-readable storage media herein include: read-only memory (ROM), random access programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), flash memory, non-volatile memory, CD-ROM, CD-R, CD+R, CD-RW, CD+RW, DVD-ROM, DVD-R, DVD+R, DVD-RW, DVD+RW, DVD-RAM, BD-ROM, BD-R, BD-R LTH, BD-RE, Blu-ray or optical disc storage, hard disk drive (HDD), solid-state drive (SSD), card storage (such as multimedia cards, secure digital (SD) cards, or ultra-fast digital (XD) cards), magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, and any other device configured to store a computer program and any associated data, data files, and data structures in a non-transitory manner and to provide the computer program and any associated data, data files, and data structures to a processor or computer so that the processor or computer can execute the computer program. The computer program in the aforementioned computer-readable storage medium can run in an environment deployed in computer devices such as clients, hosts, agent devices, servers, etc. Furthermore, in one example, the computer program and any associated data, data files, and data structures are distributed across a networked computer system, such that the computer program and any associated data, data files, and data structures are stored, accessed, and executed in a distributed manner through one or more processors or computers.

[0147] According to exemplary embodiments of the present disclosure, a computer program product may also be provided, wherein the instructions in the computer program product can be executed by a processor of a computer device to complete a method for training a motion intention recognition model according to exemplary embodiments of the present disclosure or a method for motion intention recognition based on a motion intention recognition model.

[0148] According to the training method, recognition method, device, electronic device, storage medium, and computer program product of the motion intention recognition model disclosed herein, during the training process of the motion intention recognition model, the fusion weights corresponding to various types of electrophysiological signals can be dynamically adjusted, thereby flexibly integrating the advantages of various types of electrophysiological signals. In this way, the trained motion intention recognition model can control the contribution of various types of electrophysiological signals in the motion intention recognition task according to the trained fusion weights, fully leveraging the advantages of various types of electrophysiological signals, thereby improving the accuracy and generalization ability of motion intention recognition, and effectively covering recognition objects with different degrees of movement impairment.

[0149] According to exemplary embodiments of this disclosure, the accuracy and robustness of motion origin detection can be effectively improved, thereby providing reliable assurance for motion intent decoding across individuals.

[0150] According to exemplary embodiments of this disclosure, by employing exponentially increasing inflation rates (1, 2, and 4), long-term time dependence can be effectively captured, making it suitable for decoding dynamic EEG activity.

[0151] According to exemplary embodiments of this disclosure, MRCP, as a motor cortical potential closely related to motor preparation and initiation, can reveal motor intentions in advance, thereby facilitating a faster response. In contrast, αβ rhythms are the most widely used oscillatory features in motor-related research, reflecting sensorimotor regulation encompassing both motor planning and execution. By combining the two, a balanced representation from motor planning to motor execution can be provided.

[0152] According to exemplary embodiments of this disclosure, the stability of training can be guaranteed by introducing residual connections in the electroencephalogram (EEG) branches and electromyogram (EMG) branches.

[0153] According to exemplary embodiments of this disclosure, the fusion weight update mechanism emphasizes more reliable modalities, i.e., modalities with lower training loss, while suppressing poorly performing modalities.

[0154] According to exemplary embodiments of this disclosure, by setting a regularization loss term, modality collapse in the early stages of training can be effectively prevented, and dynamic adaptation based on the performance of each modality during training can be promoted. Furthermore, this mechanism can complement FMA-based fusion weight initialization: although subjects with high FMA scores receive higher EMG weights during the initialization phase of fusion weights, entropy regularization can prevent the model from prematurely converging to rely solely on EMG, and can support adaptive rebalancing of fusion weights during training.

[0155] According to the disclosed exemplary embodiments, in addition to the fused representation, modality-specific classification branches such as the EEG-MRCP branch, EEG-αβ branch, and EMG branch are also preserved. Inspired by the residual connectivity paradigm, these branches allow each modality to output predictions independently, thereby preserving discriminative features that might be diluted during the fusion process. This not only enables the model to fully leverage the strengths of each modality but also enhances overall robustness when a particular modality is more informative for a specific subject or trial.

[0156] According to the disclosed exemplary embodiments, the confidence-guided decision-making strategy enables the model to fall back to a more reliable single-modal prediction when there is uncertainty in the fusion output, thereby achieving an effective combination of fusion and competition between modalities.

[0157] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0158] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A training method for a motion intent recognition model, characterized in that, The motion intent recognition model includes a preprocessing module and a classifier, and the training method includes: Acquire motion-related feature samples, wherein the motion-related feature samples include motion-related cortical potential (MRCP) features, α and β band features and electromyographic features, and the motion-related feature samples correspond to real labels of motion intention; The MRCP features, the α and β frequency band features, and the electromyographic features are respectively input into the preprocessing module under the corresponding branch to obtain the preprocessed features under the corresponding branch; the preprocessed features under the corresponding branch are input into the classifier to obtain the classification result under the corresponding branch; based on the classification result under the corresponding branch and the true label of the motion intention, the independent loss under the corresponding branch is calculated. Based on the preprocessing features of each branch and the fusion weights of each branch, the fusion features are calculated, wherein the fusion weights of each branch are obtained based on the independent losses of each branch in the previous training process of the current training; the fusion features are input into the classifier to obtain the fusion classification result; based on the fusion classification result and the true label of the motion intention, the fusion loss is calculated. The total loss is calculated based on the independent loss of each branch and the fusion loss. The parameters of the motion intent recognition model are adjusted based on the total loss to train the motion intent recognition model; The acquisition of motion-related feature samples includes: Acquire electroencephalogram (EEG) data samples and electromyogram (EMG) data samples; Based on the electromyography (EMG) data samples, the instantaneous energy characteristics of the EMG data samples in each channel and at each time point are enhanced using the Teager-Kaiser energy operator, and the noise at each time point is determined based on the enhanced instantaneous energy characteristics. The sensitivity factor was calculated using the constant false alarm rate method. For each time point, an adaptive threshold is determined based on the noise at that time point and the calculated sensitivity factor. For each channel, the candidate motion start time of that channel is determined based on the adaptive threshold at each time point and the enhanced instantaneous energy characteristics of that channel. The motion start time is determined based on the candidate motion start time of each channel; Based on the start time of the movement, the MRCP features are extracted from the first preset window of the EEG data sample, the α and β frequency band features are extracted from the second preset window of the EEG data sample, and the EMG features are extracted from the third preset window of the EMG data sample.

2. The training method as described in claim 1, characterized in that, The first preset window and the second preset window have different time lengths.

3. The training method as described in claim 1, characterized in that, The fusion weights of each branch are calculated using the following formula: in, This represents the fusion weight of the i-th branch in the current training iteration. This represents the independent loss of the i-th branch in the previous training iteration. This represents the independent loss of the j-th branch in the previous training iteration. It is a constant.

4. The training method as described in claim 1, characterized in that, The fusion weights of each branch during the initial training process are calculated using the following formula: in, This represents the fusion weight of the branch corresponding to the electromyographic feature during the initial training process. This represents the fusion weight of the branch corresponding to the MRCP feature during the initial training process. This represents the fusion weights of the branches corresponding to the features of the α and β frequency bands during the initial training process. This represents the Fugl-Meyer hand assessment subscale score of the identified object.

5. The training method as described in claim 1, characterized in that, The training method also includes: Calculate the loss of the regularization term; The calculation of the total loss based on the independent losses of each branch and the fusion loss includes: The total loss is calculated based on the independent loss of each branch, the fusion loss, and the regularization term loss. The regularization term loss is calculated using the following formula: in, This represents the loss of the regularization term. This indicates the number of motion-related feature samples contained in a single training batch. Indicates the first A number of motion-related feature samples in the branch The fusion weights below Indicates a constant.

6. A motion intent recognition method based on a motion intent recognition model, characterized in that, The motion intent recognition model is trained using the motion intent recognition model training method as described in any one of claims 1 to 5, wherein the motion intent recognition method includes: Acquire target motion-related features of the identified object, wherein the target motion-related features include target motion-related cortical potential (MRCP) features, target α and β frequency band features, and target electromyographic features; The target MRCP features, the target α and β frequency band features, and the target electromyography features are respectively input into the preprocessing module under the corresponding branch to obtain the target preprocessing features under the corresponding branch; the target preprocessing features under the corresponding branch are input into the classifier to obtain the independent classification results under the corresponding branch, wherein the independent classification results include the predicted motion category and the corresponding confidence level; Based on the target preprocessing features under each branch and the target fusion weights of each branch, the target fusion features are calculated, wherein the target fusion weights of each branch are the fusion weights of each branch obtained after the motion intention recognition model is trained; the target fusion features are input into the classifier to obtain the target fusion classification result; The motion category with the highest confidence among the independent classification results under each branch and the target fusion classification results is taken as the motion intent of the identified object. The acquisition of target motion-related features of the identified object includes: Acquire the target EEG data and target EMG data of the identified object; Based on the target electromyography data, the instantaneous energy characteristics of the target electromyography data in each channel and at each time point are enhanced using the Teager-Kaiser energy operator, and the noise at each time point is determined based on the enhanced instantaneous energy characteristics. The sensitivity factor was calculated using the constant false alarm rate method. For each time point, an adaptive threshold is determined based on the noise at that time point and the calculated sensitivity factor. For each channel, the candidate motion start time of that channel is determined based on the adaptive threshold at each time point and the enhanced instantaneous energy characteristics of that channel. The motion start time is determined based on the candidate motion start time of each channel; Based on the start time of the movement, the target MRCP features are extracted from the first preset window of the target EEG data, the target α and β frequency band features are extracted from the second preset window of the target EEG data, and the target EMG features are extracted from the third preset window of the target EMG data.

7. A training device for a motion intention recognition model, characterized in that, The motion intent recognition model includes a preprocessing module and a classifier, and the training device includes: The sample acquisition module is configured to acquire motion-related feature samples, wherein the motion-related feature samples include motion-related cortical potential (MRCP) features, α and β band features and electromyographic features, and the motion-related feature samples correspond to real labels of motion intentions; The independent loss acquisition module is configured to input the MRCP features, the α and β frequency band features, and the electromyographic features into the preprocessing module of the corresponding branch to obtain the preprocessed features of the corresponding branch; input the preprocessed features of the corresponding branch into the classifier to obtain the classification result of the corresponding branch; and calculate the independent loss of the corresponding branch based on the classification result of the corresponding branch and the true label of the motion intention. The fusion loss acquisition module is configured to calculate fusion features based on the preprocessed features of each branch and the fusion weights of each branch, wherein the fusion weights of each branch are obtained based on the independent losses of each branch in the previous training process of the current training; the fusion features are input into the classifier to obtain the fusion classification result; and the fusion loss is calculated based on the fusion classification result and the true label of the motion intention. The total loss calculation module is configured to calculate the total loss based on the independent loss under each branch and the fusion loss. The parameter adjustment module is configured to adjust the parameters of the motion intention recognition model based on the total loss in order to train the motion intention recognition model; The sample acquisition module is further configured as follows: Acquire electroencephalogram (EEG) data samples and electromyogram (EMG) data samples; Based on the electromyography (EMG) data samples, the instantaneous energy characteristics of the EMG data samples in each channel and at each time point are enhanced using the Teager-Kaiser energy operator, and the noise at each time point is determined based on the enhanced instantaneous energy characteristics. The sensitivity factor was calculated using the constant false alarm rate method. For each time point, an adaptive threshold is determined based on the noise at that time point and the calculated sensitivity factor. For each channel, the candidate motion start time of that channel is determined based on the adaptive threshold at each time point and the enhanced instantaneous energy characteristics of that channel. The motion start time is determined based on the candidate motion start time of each channel; Based on the start time of the movement, the MRCP features are extracted from the first preset window of the EEG data sample, the α and β frequency band features are extracted from the second preset window of the EEG data sample, and the EMG features are extracted from the third preset window of the EMG data sample.

8. A motion intent recognition device based on a motion intent recognition model, characterized in that, The motion intention recognition model is trained using the training method for the motion intention recognition model as described in any one of claims 1 to 5, and the motion intention recognition device comprises: The feature acquisition module is configured to acquire target motion-related features of the identified object, wherein the target motion-related features include target motion-related cortical potential (MRCP) features, target α and β frequency band features, and target electromyographic features; The independent classification result acquisition module is configured to input the target MRCP features, the target α and β frequency band features, and the target electromyography features into the preprocessing module under the corresponding branch to obtain the target preprocessing features under the corresponding branch; and input the target preprocessing features under the corresponding branch into the classifier to obtain the independent classification result under the corresponding branch, wherein the independent classification result includes the predicted motion category and the corresponding confidence level; The target fusion classification result acquisition module is configured to calculate target fusion features based on the target preprocessing features under each branch and the target fusion weights of each branch, wherein the target fusion weights of each branch are the fusion weights of each branch obtained after the motion intention recognition model is trained; and input the target fusion features into the classifier to obtain the target fusion classification result. The motion intent determination module is configured to take the motion category corresponding to the highest confidence among the independent classification results under each branch and the target fusion classification results as the motion intent of the identified object. The feature acquisition module is further configured as follows: Acquire the target EEG data and target EMG data of the identified object; Based on the target electromyography data, the instantaneous energy characteristics of the target electromyography data in each channel and at each time point are enhanced using the Teager-Kaiser energy operator, and the noise at each time point is determined based on the enhanced instantaneous energy characteristics. The sensitivity factor was calculated using the constant false alarm rate method. For each time point, an adaptive threshold is determined based on the noise at that time point and the calculated sensitivity factor. For each channel, the candidate motion start time of that channel is determined based on the adaptive threshold at each time point and the enhanced instantaneous energy characteristics of that channel. The motion start time is determined based on the candidate motion start time of each channel; Based on the start time of the movement, the target MRCP features are extracted from the first preset window of the target EEG data, the target α and β frequency band features are extracted from the second preset window of the target EEG data, and the target EMG features are extracted from the third preset window of the target EMG data.

9. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the training method of the motion intention recognition model as described in any one of claims 1 to 5, or to implement the motion intention recognition method based on the motion intention recognition model as described in claim 6.

10. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is able to perform the training method of the motion intention recognition model as described in any one of claims 1 to 5, or to perform the motion intention recognition method based on the motion intention recognition model as described in claim 6.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the training method of the motion intention recognition model as described in any one of claims 1 to 5, or implements the motion intention recognition method based on the motion intention recognition model as described in claim 6.

Images