Wearable motion intention recognition method based on bidirectional cross attention
By constructing a bidirectional cross-attention feature fusion module and synchronizing multi-source data, the problem of insufficient accuracy in single signal source recognition in existing technologies is solved, achieving high-precision and robust recognition of human movement intentions, which is suitable for assisted rehabilitation and daily activity assistance in wearable exoskeleton devices.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CAPITAL UNIV OF PHYSICAL EDUCATION & SPORTS
- Filing Date
- 2026-01-31
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies rely on a single signal source for human motion intention recognition, failing to fully integrate features from multiple signal sources. This limits the accuracy and comprehensiveness of recognition, especially making it difficult to accurately capture sudden changes in intention during the transition phase of a movement.
A wearable motion intention recognition method based on bidirectional cross-attention is adopted. Through multi-source data synchronization, feature fusion and model training, a cross-modal bidirectional cross-attention feature fusion module is constructed to fully explore the deep semantic association between bioelectric signals and kinematic signals. By utilizing the bidirectional cross-attention mechanism and explicit temporal position coding, the efficient fusion and accurate recognition of heterogeneous temporal signals are achieved.
It significantly improves the accuracy and robustness of motion intent recognition, enabling more comprehensive extraction of motion intent features, especially in the accuracy of intent mutation recognition during motion transitions, and enhances the model's adaptability to complex motion scenarios.
Smart Images
Figure CN122020547A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of human motion intention recognition and robotic exoskeleton technology, specifically relating to a wearable motion intention recognition method based on bidirectional cross-attention. Background Technology
[0002] In the fields of human motion intent recognition and robotic exoskeleton technology, wearable exoskeleton devices need to accurately identify various human motion states to provide data support for assisting rehabilitation and daily activities. One of the core technologies is the effectiveness of the motion intent recognition method. Currently, similar technical solutions have been applied in this field, such as the solution proposed in patent CN113011458A, "Load-Moving Exoskeleton Human Motion Intent Recognition Method and Exoskeleton System".
[0003] The existing solution acquires foot GCF and IMU signals through the exoskeleton's sensing system. It first classifies human movement states into non-periodic activities (sitting, standing) and periodic activities (running, walking, climbing stairs) based on the periodicity of the signal characteristics. For non-periodic activities, the knee joint angles of both legs are calculated using IMU signals, and the difference in knee joint angles between sitting and standing is used for differentiation. For periodic activities, the foot GCF signals are first used to determine whether both feet are supported to distinguish running from other activities. Then, a fuzzy inference system is used to identify walking, climbing stairs, or descending stairs based on the knee joint angles at the time of heel or toe impact in the foot GCF signals.
[0004] However, existing technologies have significant shortcomings. Their classification of human movement states primarily relies on the periodicity of signals from the exoskeleton sensor system. After periodic classification, further judgment is based solely on a single signal source, failing to fully integrate the characteristics of non-periodic and periodic signals within the same state mode, nor comprehensively consider the characteristics of the two types of signals under different human movement modes. The core reason for this shortcoming lies in the fact that the guiding principle of existing technologies is limited to determining movement states based on whether different human movement states are periodic and on a single signal source after periodic judgment. This lack of sufficient exploration and utilization of multi-source signal information under various movement states limits the accuracy and comprehensiveness of motion intention recognition, especially in accurately capturing sudden changes in intention during movement transitions. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a wearable motion intention recognition method based on bidirectional cross-attention to solve the problems mentioned in the background technology.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a wearable motion intention recognition method based on bidirectional cross-attention, comprising the following steps: S1. Multi-source data synchronization: The control module sends a synchronization trigger signal to trigger at least two types of signal acquisition modules to start collecting human-related signals synchronously. S2. Data cleaning: Preprocessing the synchronized multi-source data, including at least one of filtering, noise reduction, normalization, motion intention category labeling, and window segmentation. S3, Feature Fusion: S31. Perform feature extraction on the preprocessed bioelectrical signal and kinematic signal respectively to obtain the bioelectrical signal feature sequence and the kinematic signal feature sequence; S32. Time-position coding is introduced into the bioelectric signal feature sequence and the kinematic signal feature sequence respectively to obtain two types of feature sequences after position coding; the time-position coding is constructed by sine function and cosine function to characterize the relative positional relationship between different time windows; S33. Construct a bidirectional cross-attention mechanism, including a positive cross-attention mechanism and a negative cross-attention mechanism; S331, Positive Cross-Attention Mechanism: Using the bioelectrical signal features after position encoding as the query and the kinematic signal features as the key and value, positive fusion features are obtained through projection, attention weight calculation and weighted summation; S332, Reverse Cross-Attention Mechanism: Using the kinematic signal features after position encoding as the query and the bioelectric signal features as the key and value, the reverse fusion features are obtained through projection, attention weight calculation and weighted summation; S34. The positive fusion feature and the negative fusion feature are fused to form a cross-modal joint feature; S4. Motion Intent Recognition: Input the cross-modal joint features into the intent recognition module and output the motion intent recognition result; S5. Model Training: The model is trained using a total loss function that includes classification loss and modal collaboration loss. The total loss function is used to optimize the accuracy of cross-modal feature representation and motion intent recognition. The training process adopts a phased training strategy.
[0007] Preferably, the signal acquisition module includes a bioelectric signal acquisition unit and a kinematic signal acquisition unit; each signal acquisition unit continuously acquires data at its own fixed sampling frequency, and during the data synchronization processing stage, all signal source data are unified to the same sampling rate; The synchronization trigger signal is a fixed-voltage TTL signal, and the fixed voltage is 3V-5V; The bioelectric signal acquisition unit includes a surface muscle electrophysiological signal measuring instrument, and the kinematic signal acquisition unit includes at least one of an inertial measurement unit and a motor encoder; The same sampling rate is the highest sampling rate among the sampling rates of each signal acquisition module. Signal source data with a sampling rate lower than this highest sampling rate are unified through upsampling.
[0008] Preferably, the preprocessing in step S2 includes: applying Kalman filtering and standardization to the data collected by the inertial measurement unit; converting the format of the data collected by the motor encoder to obtain angle signals and / or angular velocity signals; filtering and denoising the bioelectric signals using at least one of notch filter, low-pass filter, and band-pass filter, and then normalizing them; and completing the motion intention category labeling based on human motion-related features.
[0009] Preferably, in step S31, features are extracted using a sliding window segmentation method, and continuous features are selected. T Each time window is used as a time series input sample; Let the dimension of the bioelectrical signal features extracted within each time window be . The kinematic signal feature dimension is Select continuous T If each time window is used as a time-series input sample, then the corresponding two types of modal features are represented as follows:
[0010] in, Represents the characteristic sequence of bioelectrical signals. This represents the characteristic sequence of kinematic signals.
[0011] Preferably, the time position encoding in step S32 is defined as:
[0012] in, Indicates the time window index. , For attention, the latent space dimension, Number of time windows; The location-encoded bioelectrical signal feature sequence is The kinematic signal feature sequence after position encoding is: .
[0013] Preferably, the specific calculation process of the positive cross-attention mechanism in step S331 is as follows:
[0014] in, , , ; The correlation between bioelectrical signals and kinematic signals over time is calculated to obtain the cross-attention weight matrix:
[0015]
[0016] in, This is a query vector for bioelectrical signals. The kinematic signal key vector, For kinematic signal value vectors, This is the weight matrix. For bias vectors, For attention weights, This is a positive fusion feature. For attention, the hidden space dimension.
[0017] Preferably, the specific calculation process of the reverse cross-attention mechanism in step S332 is as follows:
[0018] in, , ; Calculate the reverse cross-attention and obtain the fused output:
[0019] in, For kinematic signal query vectors, For bioelectrical signal bond vectors, This is a vector of bioelectrical signal values. This is the weight matrix. For bias vectors, This is a reverse fusion feature. For attention, the hidden space dimension.
[0020] Preferably, the fusion process in step S34 includes feature splicing or weighted fusion; The intent recognition module includes at least one of a fully connected neural network and a long short-term memory network. The exercise intention includes one or more of the following: sitting, standing, walking, climbing stairs, descending stairs, going uphill, going downhill, and switching between sitting and standing.
[0021] Preferably, in step 5, the total loss function is:
[0022] in, These are the weighting coefficients. This is the main classification loss function, used to directly optimize classification accuracy, penalizing the deviation between the predicted probability and the true label. Its specific form is:
[0023] in, The number of categories of movement intentions. , The total number of samples, This represents the sample size during the transition phase. These are class weights, used to address the problem of imbalanced samples. It's a real label. It is the class probability predicted by the model; The modal collaborative consistency loss function is:
[0024] in, and These are the time steps after attention fusion. The bioelectrical and kinematic signal feature vectors, This indicates an element-wise multiplication operation. The length of the time window; The phased training strategy is as follows: first, fix the kinematic signal-related network branches, train the bioelectric signal cross-attention related parameters, then unfreeze the kinematic signal-related network branches, and jointly fine-tune all model parameters.
[0025] A wearable motion intention recognition system based on bidirectional cross-attention includes a wearable exoskeleton, at least two types of signal acquisition modules, a control module, and a processing module. The signal acquisition module includes a bioelectric signal acquisition module and a kinematic signal acquisition module, which are used to acquire human-related signals. The control module is used to send a synchronization trigger signal to control each signal acquisition module to acquire data synchronously. The processing module is used to execute the steps of the wearable motion intent recognition method with bidirectional cross-attention to achieve motion intent recognition.
[0026] Compared with existing technologies, this invention provides a wearable motion intent recognition method based on bidirectional cross-attention, which has the following beneficial effects: Significantly improved recognition accuracy: This invention constructs a cross-modal bidirectional cross-attention feature fusion module to fully explore the deep semantic relationship between bioelectrical signals (such as electromyographic signals) and kinematic signals (angle acquired by motor encoders, IMU signals, etc.). It not only models the positive influence of bioelectrical signals on kinematic motion states, but also considers the inverse modulation relationship of kinematic feedback on changes in bioelectrical signals. This overcomes the shortcomings of existing technologies that rely on only a single signal source or do not fully fuse the two types of signals. It can extract action intention features more comprehensively and greatly improve the recognition accuracy of various action states such as sitting, standing, walking, and climbing stairs, as well as sudden changes in intention during transition phases.
[0027] Enhanced temporal sensitivity and robustness: By introducing explicit temporal position coding (TPE), the relative positional relationship between different time windows is effectively represented, which strengthens the model's ability to perceive temporal order and causal relationship. This solves the problem of insufficient capture of intentional abrupt changes during the transition phase of the action in existing technologies, making the model more robust when facing complex and ever-changing human motion scenarios and able to stably output accurate motion intention recognition results.
[0028] Multi-source data utilization efficiency optimization: This invention designs a comprehensive multi-source data synchronization scheme. It achieves time synchronization of signals acquired by different hardware through TTL signal triggering, and then unifies the sampling rate through upsampling. Combined with targeted data cleaning and preprocessing (such as Kalman filtering, notch filter denoising, normalization, etc.), it ensures the effectiveness and consistency of multi-source data. At the same time, the bidirectional cross-attention mechanism realizes the efficient fusion of heterogeneous time-series signals, maximizing the value of each signal source and making fuller use of data information compared with existing technologies.
[0029] The model training performance is superior: the total loss function combines weighted cross-entropy loss and modal coherence loss, which not only solves the sample imbalance problem through class weights and ensures classification accuracy, but also ensures the semantic relevance of different signal features at the same time step through modal coherence constraints, helping the model learn better cross-modal representations; the progressive training method further optimizes the training process, improves the model's convergence speed and performance stability, and enables the model to quickly adapt to the needs of wearable exoskeleton devices in practical applications, providing reliable data support for assisting rehabilitation and daily activities. Attached Figure Description
[0030] Figure 1 This is a schematic diagram illustrating the steps of the wearable motion intention recognition method based on bidirectional cross-attention of the present invention; Figure 2 This is a schematic diagram of the structure of the flowchart for the bidirectional cross-attention fusion and intent recognition of bioelectric-kinematic signals of the present invention. Detailed Implementation
[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0032] Example 1: See attached document Figures 1 to 2 A wearable motion intent recognition method based on bidirectional cross-attention includes the following steps: S1. Multi-source data synchronization: The control module sends a synchronization trigger signal to trigger at least two types of signal acquisition modules to start collecting human-related signals synchronously. The signal acquisition module includes a bioelectric signal acquisition unit and a kinematic signal acquisition unit; each signal acquisition unit continuously acquires data at its own fixed sampling frequency, and during the data synchronization processing stage, all signal source data are unified to the same sampling rate; The synchronous trigger signal is a fixed-voltage TTL signal, with a fixed voltage of 3V-5V; the bioelectric signal acquisition unit includes a surface muscle electrophysiology signal measuring instrument, and the kinematic signal acquisition unit includes at least one of an inertial measurement unit (IMU) and a motor encoder; the same sampling rate is the highest sampling rate among the sampling rates of each signal acquisition module, and signal source data with a sampling rate lower than this highest sampling rate are unified through upsampling; Specifically, addressing the issue of inconsistent signal timing sources among different hardware during human motion acquisition, both bioelectrical signals and exoskeleton kinematic signals support receiving a fixed 3.3V TTL signal as the start of hardware acquisition and recording. During actual synchronization, the exoskeleton device's controller sends a TTL signal to trigger all signal acquisition sensors to begin data recording. The main signal sources include IMUs (Inertial Measurement Units) worn on various parts of the body, lower limb hip and knee joint angles and angular velocities measured via motor encoders, and muscle electrical signals measured by a surface electromyography (EMG) sensor.
[0033] To address the issue of inconsistent sampling frequencies from different signal sources, a TTL source signal is used to trigger the recording of the start of the record for different signal sources. Subsequently, data is continuously collected from different signal sources at their respective fixed sampling frequencies. During the data synchronization processing phase, data from other signal sources with sampling rates lower than the highest sampling rate are upsampled to the signal source with the highest sampling rate, and this highest sampling rate is used as the consistent sampling rate for the fused data.
[0034] S2. Data cleaning: Preprocess the synchronized multi-source data. Preprocessing includes at least one of the following: filtering, noise reduction, normalization, motion intent category labeling, and window segmentation. The preprocessing in this step includes: applying Kalman filtering and normalization to the data acquired by the inertial measurement unit (IMU); converting the format of the data acquired by the motor encoder to obtain angle and / or angular velocity signals; filtering and denoising the bioelectrical signals using at least one of a notch filter, low-pass filter, and band-pass filter, followed by normalization; labeling the motion intent category based on human motion-related features; the center frequency of the notch filter is 50Hz or 60Hz, the cutoff frequency of the low-pass filter is 10Hz-50Hz, and the passband frequency of the band-pass filter is 20Hz-1000Hz; the normalization process includes Z-Score normalization; the human motion-related features include at least one of hip and knee joint angles and hip and knee joint angular velocities. This stage primarily involves preprocessing data from different signal sources, including filtering, denoising, normalization, motion intent category labeling, and window segmentation, to facilitate subsequent feature extraction and model training. Kalman filtering is applied to IMU data sources worn on various body parts, followed by Z-score normalization. Data obtained from motor encoder measurements is uniformly converted into corresponding angle and angular velocity formats. Surface electromyography (EMG) signals are filtered and denoised using notch filters, low-pass filters, and band-pass filters, and then normalized using the largest muscle sample obtained in the experiment. Motion intent category labeling is performed on the collected data based on the hip and knee joint angle and angular velocity characteristics under different movement states.
[0035] S3, Feature Fusion: S31. Perform feature extraction on the preprocessed bioelectrical signal and kinematic signal respectively to obtain the bioelectrical signal feature sequence and the kinematic signal feature sequence; In this step, features are extracted using a sliding window segmentation method, and continuous features are selected. T Each time window is used as a time series input sample; Let the dimension of the bioelectrical signal features extracted within each time window be . The kinematic signal feature dimension is Select continuous T If each time window is used as a time-series input sample, then the corresponding two types of modal features are represented as follows:
[0036] in, Represents the characteristic sequence of bioelectrical signals. This represents the characteristic sequence of kinematic signals.
[0037] S32. Time-position coding is introduced into the bioelectric signal feature sequence and the kinematic signal feature sequence respectively to obtain two types of feature sequences after position coding; the time-position coding is constructed by sine function and cosine function to characterize the relative positional relationship between different time windows; In this step, the Time Position Encoding (TPE) is defined as follows:
[0038] in, Indicates the time window index. , For attention, the latent space dimension, Number of time windows; The location-encoded bioelectrical signal feature sequence is The kinematic signal feature sequence after position encoding is: .
[0039] S33. Construct a bidirectional cross-attention mechanism, including a positive cross-attention mechanism and a negative cross-attention mechanism; S331, Positive Cross-Attention Mechanism: Using the bioelectrical signal features after position encoding as the query and the kinematic signal features as the key and value, positive fusion features are obtained through projection, attention weight calculation and weighted summation; The specific calculation process of the positive cross-attention mechanism in this step is as follows:
[0040] in, , , ; The correlation between bioelectrical signals and kinematic signals over time is calculated to obtain the cross-attention weight matrix:
[0041]
[0042] in, This is a query vector for bioelectrical signals. The kinematic signal key vector, For kinematic signal value vectors, This is the weight matrix. For bias vectors, Attention weights (specifically representing the first) The bioelectrical signal characteristics of the first time window for the second (the degree of dependence of kinematic signal features within a time window) This is a positive fusion feature. For attention, the hidden space dimension.
[0043] S332, Reverse Cross-Attention Mechanism: Using the kinematic signal features after position encoding as the query and the bioelectric signal features as the key and value, the reverse fusion features are obtained through projection, attention weight calculation and weighted summation; The specific calculation process of the reverse cross-attention mechanism in this step is as follows:
[0044] in, , ; Calculate the reverse cross-attention and obtain the fused output:
[0045] in, For kinematic signal query vectors, For bioelectrical signal bond vectors, This is a vector of bioelectrical signal values. This is the weight matrix. For bias vectors, This is a reverse fusion feature. For the attention latent space dimension; the positive cross-attention output will be... With reverse cross-attention output Feature concatenation or weighted fusion is performed to form a cross-modal joint feature representation. The fused features are then input into the intent recognition module for classification decision-making. The intent recognition module uses a fully connected neural network and a Long Short-Term Memory (LSTM) classification model for recognition.
[0046] By classifying different signal acquisition sources into bioelectrical signals (such as electromyographic signals) and kinematic signals (such as angles acquired by motor encoders and signals from IMU), and considering the lead-lag relationship between bioelectrical signals and kinematic signals in the time dimension, a cross-modal bidirectional cross-attention feature fusion module is constructed to achieve deep semantic alignment and collaborative modeling between the two types of heterogeneous time-series signals.
[0047] S34. Perform fusion processing on the forward fusion features and the reverse fusion features to form cross-modal joint features; the fusion processing in this step includes feature splicing or weighted fusion. The intent recognition module includes at least one of the following: fully connected neural network, long short-term memory network (LSTM), recurrent neural network (RNN), and Transformer; Movement intentions include one or more of the following: sitting, standing, walking, climbing stairs, descending stairs, going uphill, going downhill, and switching between sitting and standing. S4. Motion Intent Recognition: Input the cross-modal joint features into the intent recognition module and output the motion intent recognition result; S5. Model Training: The model is trained using a total loss function that includes classification loss and modal collaboration loss. The total loss function is used to optimize cross-modal feature representation and motion intent recognition accuracy. A phased training strategy is adopted in the training process. In this step, the total loss function is
[0048] in, These are the weighting coefficients. This is the main classification loss function, used to directly optimize classification accuracy, penalizing the deviation between the predicted probability and the true label. Its specific form is:
[0049] in, The number of categories of movement intentions. , The total number of samples, This represents the sample size during the transition phase. These are class weights, used to address the problem of imbalanced samples. It's a real label. It is the class probability predicted by the model; The modal cooperative consistency loss function is:
[0050] in, and These are the time steps after attention fusion. The bioelectrical and kinematic signal feature vectors, This indicates an element-wise multiplication operation. As above, this is the length of the time window; the modal coherence loss function is to ensure the semantic correlation of bioelectrical signal and kinematic signal features at the same time step. The phased training strategy is as follows: first, fix the kinematic signal-related network branches and train the cross-attention parameters of bioelectric signals; then, unfreeze the kinematic signal-related network branches and jointly fine-tune all model parameters.
[0051] Bioelectric signals include at least one of electromyography (EMG), electrocardiography (ECG), and electroencephalography (EEG); kinematic signals include at least one of angle signals, angular velocity signals, acceleration signals, and posture signals.
[0052] A wearable motion intention recognition system based on bidirectional cross-attention includes a wearable exoskeleton, at least two types of signal acquisition modules, a control module, and a processing module. The signal acquisition module includes a bioelectric signal acquisition module and a kinematic signal acquisition module, which are used to acquire human-related signals; The control module is used to send a synchronization trigger signal to control each signal acquisition module to acquire data synchronously. The processing module is used to execute the steps of the wearable motion intent recognition method with bidirectional cross-attention to achieve motion intent recognition.
[0053] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A wearable motion intent recognition method based on bidirectional cross-attention, characterized in that, Includes the following steps: S1. Multi-source data synchronization: The control module sends a synchronization trigger signal to trigger at least two types of signal acquisition modules to start collecting human-related signals synchronously. S2. Data cleaning: Preprocessing the synchronized multi-source data, including at least one of filtering, noise reduction, normalization, motion intention category labeling, and window segmentation. S3, Feature Fusion: S31. Perform feature extraction on the preprocessed bioelectrical signal and kinematic signal respectively to obtain the bioelectrical signal feature sequence and the kinematic signal feature sequence; S32. Time-position coding is introduced into the bioelectric signal feature sequence and the kinematic signal feature sequence respectively to obtain two types of feature sequences after position coding; the time-position coding is constructed by sine function and cosine function to characterize the relative positional relationship between different time windows; S33. Construct a bidirectional cross-attention mechanism, including a positive cross-attention mechanism and a negative cross-attention mechanism; S331, Positive Cross-Attention Mechanism: Using the bioelectrical signal features after position encoding as the query and the kinematic signal features as the key and value, positive fusion features are obtained through projection, attention weight calculation and weighted summation; S332, Reverse Cross-Attention Mechanism: Using the kinematic signal features after position encoding as the query and the bioelectric signal features as the key and value, the reverse fusion features are obtained through projection, attention weight calculation and weighted summation; S34. The positive fusion feature and the negative fusion feature are fused to form a cross-modal joint feature; S4. Motion Intent Recognition: Input the cross-modal joint features into the intent recognition module and output the motion intent recognition result; S5. Model Training: The model is trained using a total loss function that includes classification loss and modal collaboration loss. The total loss function is used to optimize the accuracy of cross-modal feature representation and motion intent recognition. The training process adopts a phased training strategy.
2. The wearable motion intent recognition method based on bidirectional cross-attention according to claim 1, characterized in that, The signal acquisition module includes a bioelectric signal acquisition unit and a kinematic signal acquisition unit; each signal acquisition unit continuously acquires data at its own fixed sampling frequency, and during the data synchronization processing stage, all signal source data are unified to the same sampling rate. The synchronization trigger signal is a fixed-voltage TTL signal, and the fixed voltage is 3V-5V; The bioelectric signal acquisition unit includes a surface muscle electrophysiological signal measuring instrument, and the kinematic signal acquisition unit includes at least one of an inertial measurement unit and a motor encoder; The same sampling rate is the highest sampling rate among the sampling rates of each signal acquisition module. Signal source data with a sampling rate lower than this highest sampling rate are unified through upsampling.
3. The wearable motion intent recognition method based on bidirectional cross-attention according to claim 2, characterized in that, The preprocessing in step S2 includes: applying Kalman filtering and standardization to the data collected by the inertial measurement unit; converting the format of the data collected by the motor encoder to obtain angle signals and / or angular velocity signals; filtering and denoising the bioelectric signals using at least one of notch filters, low-pass filters, and band-pass filters, and then normalizing them; and labeling the motion intention category based on human motion-related features.
4. The wearable motion intent recognition method based on bidirectional cross-attention according to claim 1, characterized in that, In step S31, features are extracted using a sliding window segmentation method, and continuous features are selected. T Each time window is used as a time series input sample; Let the dimension of the bioelectrical signal features extracted within each time window be . The kinematic signal feature dimension is Select continuous T If each time window is used as a time-series input sample, then the corresponding two types of modal features are represented as follows: in, Represents the characteristic sequence of bioelectrical signals. This represents the characteristic sequence of kinematic signals.
5. The wearable motion intent recognition method based on bidirectional cross-attention according to claim 4, characterized in that, The time position encoding in step S32 is defined as follows: in, Indicates the time window index. , For attention, the latent space dimension, Number of time windows; The location-encoded bioelectrical signal feature sequence is The kinematic signal feature sequence after position encoding is: .
6. The wearable motion intent recognition method based on bidirectional cross-attention according to claim 5, characterized in that, The specific calculation process of the positive cross-attention mechanism in step S331 is as follows: in, , , ; The correlation between bioelectrical signals and kinematic signals over time is calculated to obtain the cross-attention weight matrix: in, This is a query vector for bioelectrical signals. The kinematic signal key vector, For kinematic signal value vectors, This is the weight matrix. For bias vectors, For attention weights, This is a positive fusion feature. For attention, the hidden space dimension.
7. The wearable motion intent recognition method based on bidirectional cross-attention according to claim 6, characterized in that, The specific calculation process of the reverse cross-attention mechanism in step S332 is as follows: in, , ; Calculate the reverse cross-attention and obtain the fused output: in, For kinematic signal query vectors, For bioelectrical signal bond vectors, This is a vector of bioelectrical signal values. This is the weight matrix. For bias vectors, This is a reverse fusion feature. For attention, the hidden space dimension.
8. The wearable motion intent recognition method based on bidirectional cross-attention according to claim 1, characterized in that, The fusion process in step S34 includes feature splicing or weighted fusion. The intent recognition module includes at least one of a fully connected neural network and a long short-term memory network. The exercise intention includes one or more of the following: sitting, standing, walking, climbing stairs, descending stairs, going uphill, going downhill, and switching between sitting and standing.
9. The wearable motion intent recognition method based on bidirectional cross-attention according to claim 1, characterized in that, In step 5, the total loss function is: in, These are the weighting coefficients. This is the main classification loss function, used to directly optimize classification accuracy, penalizing the deviation between the predicted probability and the true label. Its specific form is: in, The number of categories of movement intentions. , The total number of samples, This represents the sample size during the transition phase. These are class weights, used to address the problem of imbalanced samples. It's a real label. It is the class probability predicted by the model; The modal collaborative consistency loss function is: in, and These are the time steps after attention fusion. The bioelectrical and kinematic signal feature vectors, This indicates an element-wise multiplication operation. The length of the time window; The phased training strategy is as follows: first, fix the kinematic signal-related network branches, train the bioelectric signal cross-attention related parameters, then unfreeze the kinematic signal-related network branches, and jointly fine-tune all model parameters.
10. A wearable motion intention recognition system based on bidirectional cross-attention, used to implement the wearable motion intention recognition method according to any one of claims 1-9, characterized in that, It includes a wearable exoskeleton body, at least two types of signal acquisition modules, a control module, and a processing module; The signal acquisition module includes a bioelectric signal acquisition module and a kinematic signal acquisition module, which are used to acquire human-related signals. The control module is used to send a synchronization trigger signal to control each signal acquisition module to acquire data synchronously. The processing module is used to execute the steps of the wearable motion intent recognition method with bidirectional cross-attention to achieve motion intent recognition.