Lower limb function rehabilitation evaluation method based on multi-dimensional feature fusion network

Abstract

The invention relates to a lower limb function rehabilitation evaluation method based on a multi-dimensional feature fusion network, and belongs to the technical field of intelligent evaluation of lower limb function rehabilitation conditions. The method comprises the following steps: acquiring multichannel surface electromyogram signals of a stroke patient and preprocessing the multichannel surface electromyogram signals to construct a data set; constructing a lower limb function rehabilitation evaluation model which comprises a time domain feature extraction branch, a frequency domain feature extraction branch, a first cross attention module, a second cross attention module and a classifier; inputting samples in the data set into a lower limb function rehabilitation evaluation model, and training the model; optimizing the model by adopting a loss function, updating parameters by using an AdamW optimizer, and minimizing loss through a back propagation algorithm until the model converges to obtain a trained model; and preprocessing the to-be-evaluated surface electromyogram signal, and inputting the preprocessed to-be-evaluated surface electromyogram signal into the trained model to obtain an evaluation result The accuracy of lower limb function rehabilitation evaluation can be improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent assessment technology for lower limb functional rehabilitation, specifically relating to a method for assessing lower limb functional rehabilitation based on a multidimensional feature fusion network. Background Technology

[0002] The aging global population and changes in lifestyle have led to a significant increase in the number of patients with lower limb dysfunction. These conditions are often caused by stroke, trauma, and other injuries. Effective rehabilitation of these patients requires precise and objective functional assessments. However, traditional rehabilitation assessment methods rely heavily on the subjective judgment of physicians, resulting in significant differences among assessors and low repeatability. Furthermore, these methods are mostly static and superficial assessments, which cannot accurately reflect the actual functional recovery status of patients during rehabilitation training. Therefore, the field of rehabilitation medicine urgently needs more objective and quantifiable assessment methods.

[0003] The non-invasive and real-time acquisition characteristics of surface electromyography (sEMG) signals can directly reflect the activity state of the neuromuscular system, providing a new path to solve the above problems. Furthermore, research on intelligent rehabilitation assessment based on sEMG combined with artificial intelligence has made some progress. Its related indicators can effectively quantify muscle activation level and coordination, and are also significantly correlated with clinical scale scores. However, existing sEMG-based assessment methods mostly consider only single-domain information and rely on manually designed features, making it difficult to fully explore the deep high-dimensional pattern information in the signal, and also unable to capture the dynamic evolution characteristics of limb movement. Therefore, it is urgent to develop new technical solutions to achieve a more refined, reliable, and objective intelligent quantitative assessment of lower limb dysfunction by deeply exploring the representation capabilities of sEMG signals and designing multi-dimensional cross-domain collaborative fusion mechanisms. Summary of the Invention

[0004] To achieve the above objectives, the present invention employs the following technical solution: This invention provides a lower limb functional rehabilitation assessment method based on a multidimensional feature fusion network, comprising the following steps: S1. Acquire multi-channel surface electromyography signals from stroke patients performing standard rehabilitation movements and preprocess them to construct a dataset; S2. Construct a lower limb functional rehabilitation assessment model, including a time-domain feature extraction branch, a frequency-domain feature extraction branch, a first cross-attention module, a second cross-attention module, and a classifier; input samples from the dataset into the lower limb functional rehabilitation assessment model to train the model; S3. Optimize the model using a loss function, update the parameters using the AdamW optimizer, minimize the loss using the backpropagation algorithm until the model converges, and obtain the trained lower limb functional rehabilitation assessment model. S4. After preprocessing, the electromyographic signals of the surface to be evaluated are input into the trained lower limb functional rehabilitation assessment model to obtain the assessment results.

[0005] Furthermore, the raw multi-channel electromyography data were collected. A standardized preprocessing procedure is performed: the signal of each channel is cropped using a thresholding method based on Z-scores, and the cropped EMG signal is processed using a Butterworth bandpass filter to retain the main components of the EMG signal from 20Hz to 150Hz, resulting in a clean EMG signal; the clean EMG signal is processed using a downsampling method based on anti-aliasing filtering to obtain a downsampled EMG signal; the DC component of the signal in each channel of the downsampled EMG data is removed to obtain the preprocessed EMG signal.

[0006] Furthermore, in the temporal feature extraction branch, for each input sample The features at each time step are transformed from the original channel dimension through a linear projection layer. Mapping to the latent space dimension yields the projected sequence. For the projected sequence By adding positional encoding, we obtain the positional encoding sequence. ; in the positional encoding sequence A learnable special marker is concatenated at the starting position. The spliced ​​sequence is obtained. The formula is expressed as follows: , in, Indicates feature concatenation operation; Represents a trainable parameter vector; concatenates the sequence The input is a module consisting of four stacked Transformer encoder layers, which is processed to obtain the encoded sequence. Take the encoded sequence The features at the starting position are used as the output of the temporal feature extraction branch to obtain the temporal features. The formula is expressed as follows: , in, Indicates taking the encoded sequence Operations on the starting position feature.

[0007] Furthermore, in the frequency domain feature extraction branch, for each input sample Applying a Fast Fourier Transform along the time dimension to each channel calculates the amplitude spectrum, yielding the frequency domain representation. Representing the frequency domain Input a multi-scale one-dimensional convolutional module to obtain frequency domain features. .

[0008] Furthermore, the multi-scale one-dimensional convolution module includes three parallel one-dimensional convolution paths with kernel sizes of 3, 5, and 7 from top to bottom. Each path includes convolution, batch normalization, and the GELU activation function; the frequency domain features It is obtained by concatenating the outputs of each one-dimensional convolutional path, then processing it through a hybrid convolutional layer and global average pooling; the formula is expressed as follows: , in, This represents parallel one-dimensional convolutions with kernels of 3, 5, and 7. This represents a 1x1 hybrid convolutional layer; This indicates a global average pooling operation.

[0009] Furthermore, the first and second cross-attention modules have identical structures and independent parameters, both employing a multi-head attention mechanism; [the text abruptly shifts to a different topic] ...temporal features As a query Frequency domain characteristics As a key Sum Input the first cross-attention module Temporal enhancement features were obtained. The formula is expressed as follows: , in, Presentation layer normalization operation; This indicates the multi-head attention mechanism operation of the first cross-attention module; frequency domain features As a query Temporal characteristics As a key Sum Input the second cross-attention module Output frequency domain enhancement features The formula is expressed as follows: , in, This indicates the multi-head attention mechanism operation of the second cross-attention module.

[0010] Furthermore, the temporal domain enhancement features Frequency domain enhancement features The features are concatenated along the feature dimension to obtain the fused features. ; to integrate features Input the classifier to obtain the predicted probability of each category, and take the category with the highest probability as the final evaluation and staging result.

[0011] Furthermore, the loss function adopted is the cross-entropy loss function.

[0012] The advantages of this invention are: This invention, through a dual-stream deep feature extraction framework oriented towards rehabilitation staging, is the first to systematically combine deep Transformer with time-frequency domain analysis for sEMG signal analysis. The time-domain branch accurately captures the temporal coordination patterns of muscle activation through a self-attention mechanism, while the frequency-domain branch analyzes the spectral energy distribution of muscle contraction through multi-scale convolution. This dual-perspective deep feature learning overcomes the shortcomings of traditional methods that rely on manual features, resulting in incomplete information and limited representation capabilities. The bidirectional attention fusion mechanism proposed in this invention does not simply concatenate time-domain and frequency-domain features, but rather allows the features of the two modalities to interact as queries and contexts, engaging in a deep "information exchange." For example, when the model analyzes a signal suspected to be in stage III, the time-domain branch may capture an abnormal coordinated contraction temporal pattern. At this time, through cross-attention, it can extract relevant spectral evidence from the frequency-domain features that may manifest as abnormally wide bandwidth or energy dispersion, and vice versa. This dynamic, adaptive feature recalibration process can effectively mitigate the risk of misjudgment caused by individual differences, slight changes in electrode position, or measurement noise in single-modal signals, and improve the model's generalization ability and robustness when facing real, complex, and noisy clinical data. Attached Figure Description

[0013] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0014] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 This is a graph showing the accuracy results for each round of training using the method of this invention. Figure 3 This is a graph showing the macro F1 score results for each round of training using the method of this invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Example 1 In this embodiment, as Figure 1As shown, this invention provides a lower limb functional rehabilitation assessment method based on a multidimensional feature fusion network, the specific steps of which include: S1. Acquire multi-channel surface electromyography signals from stroke patients performing standard rehabilitation movements and preprocess them to construct a dataset; Specifically, 23 sets of lower limb surface electromyographic signals were collected from 13 stroke patients and 1 healthy individual while performing five standard movements: knee extension, knee flexion, plantar flexion, dorsiflexion, and standing. Each acquisition covered multiple target muscle channels, obtaining raw multichannel electromyographic data. ,in Represents the original signal length. Represents the number of channels; for raw multichannel electromyography data Perform a standardized preprocessing procedure: The signal for each channel is cropped using a Z-score-based thresholding method to suppress transient spikes caused by hardware noise. For the first channel... Signal of each channel The clipped signal is The formula is expressed as follows: , in, They represent The mean and standard deviation; This is the threshold factor (usually set to 5); The method involves using a thresholding approach based on Z-scores; processing the cropped EMG signal using a Butterworth bandpass filter to retain the main components of the EMG signal from 20Hz to 150Hz, filtering out low-frequency motion artifacts and high-frequency noise to obtain a clean EMG signal; processing the clean EMG signal using a downsampling method based on anti-aliasing filtering to obtain a downsampled EMG signal; removing the DC component from the signal of each channel in the downsampled EMG data to eliminate signal baseline drift and normalizing it to obtain a preprocessed EMG signal.

[0017] The downsampling method based on anti-aliasing filtering works as follows: A finite impulse response (FIR) low-pass filter is used to preprocess the signal to prevent spectral aliasing of high-frequency components after downsampling. Subsequently, the filtered signal is decimated by an integer multiple, reducing the sampling rate from the original 2048 Hz to a lower frequency (256 Hz) more suitable for model processing. Downsampling reduces data redundancy while preserving important information, thus improving computational efficiency.

[0018] A sliding window is used to segment the continuous preprocessed electromyographic signals into samples of fixed length. Let the window length be 1. Step size is Then the first Sample Represented as: , in, This represents the preprocessed electromyographic signal; each sample This corresponds to a Brunnstrom staging label that is jointly diagnosed by multiple professional doctors. , These correspond to stages II through VI, respectively; all samples and their corresponding Brunnstrom stage labels together constitute the sample set. The training set was randomly divided into two parts in a 7:2:1 ratio. Validation set and test set .

[0019] S2. Construct a lower limb functional rehabilitation assessment model, including a time-domain feature extraction branch, a frequency-domain feature extraction branch, a first cross-attention module, a second cross-attention module, and a classifier; Specifically, in the temporal feature extraction branch, for each input sample The features at each time step are transformed from the original channel dimension through a linear projection layer (fully connected layer). Mapped to a higher-dimensional latent space dimension =128, the 128-dimensional feature space is sufficient to encode the complex temporal features in sEMG signals, providing a rich information basis for subsequent attention fusion to obtain the projected sequence. The formula is expressed as follows: , in, Represents the projection weight matrix; This represents the bias vector; since the Transformer itself does not have the ability to perceive the sequence order, positional encoding must be explicitly added. For the projected sequence... By adding positional encoding, we obtain the positional encoding sequence. The formula is expressed as follows: , in, Represents the position encoding matrix; This represents an element-wise addition operation; to aggregate information from the entire sequence and generate a global representation, the sequence is encoded at position. A learnable special marker is concatenated at the starting position. The spliced ​​sequence is obtained. The formula is expressed as follows: , in, Indicates feature concatenation operation; Represents a trainable parameter vector; concatenates the sequence The input is a module consisting of four stacked Transformer encoder layers, which is processed to obtain the encoded sequence. Each Transformer encoder layer contains an 8-head multi-head self-attention mechanism, a feedforward neural network, residual connections, and layer normalization to model long-range dependencies between signal points; the encoded sequence is then extracted. The features at the starting position are used as the output of the temporal feature extraction branch to obtain the temporal features. The formula is expressed as follows: , in, Indicates taking the encoded sequence Operations on the starting position feature.

[0020] Specifically, in the frequency domain feature extraction branch, for each input sample Applying a Fast Fourier Transform along the time dimension to each channel calculates the amplitude spectrum to obtain the frequency distribution information of the signal, thus obtaining the frequency domain representation. , The number of frequency points; representing the frequency domain. Input a multi-scale one-dimensional convolutional module to obtain frequency domain features. The multi-scale one-dimensional convolution module includes three parallel one-dimensional convolution paths with kernel sizes of 3, 5, and 7 from top to bottom. Each path includes convolution, batch normalization, and the GELU activation function; the frequency domain features It is obtained by concatenating the outputs of each one-dimensional convolutional path, then processing it through a hybrid convolutional layer and global average pooling; the formula is expressed as follows: , in, This represents parallel one-dimensional convolutions with kernels of 3, 5, and 7. This represents a 1x1 hybrid convolutional layer used for fusing and reducing the dimensionality of the concatenated high-dimensional features; This indicates a global average pooling operation.

[0021] Specifically, the first and second cross-attention modules have the same structure and independent parameters; both employ a multi-head attention mechanism, as expressed in the following formula: , in, These represent the query, key, and value, respectively. Indicator key The dimension; Time domain features As a query Frequency domain characteristics As a key Sum Input the first cross-attention module This allows time-domain features to focus on the frequency-domain components most relevant to the current time-domain context, resulting in enhanced time-domain features. , The formula is expressed as follows: , in, Presentation layer normalization operation; This indicates the multi-head attention mechanism operation of the first cross-attention module; frequency domain features As a query Temporal characteristics As a key Sum Input the second cross-attention module This enables frequency domain features to be associated with their corresponding time domain activity stages, outputting frequency domain enhanced features. , The formula is expressed as follows: , in, This indicates the multi-head attention mechanism operation of the second cross-attention module.

[0022] Temporal enhancement features Frequency domain enhancement features The features are concatenated along the feature dimension to obtain the fused features. , The formula is expressed as follows: ; Fusion features The input is processed by a classifier to obtain the predicted probability of each category. The category with the highest probability is taken as the final evaluation and classification result. The classifier consists of a two-layer feedforward network. The first layer maps the dimension to 256 and is followed by a GELU activation function. The second layer maps the dimension from 256 to 5 and applies a Softmax function to the 5-dimensional vector to obtain the predicted probability of each category. .

[0023] S3. Optimize the model using a loss function, update the parameters using the AdamW optimizer, minimize the loss using the backpropagation algorithm until the model converges, and obtain the trained lower limb functional rehabilitation assessment model. Specifically, the loss function adopted is the cross-entropy loss function, which is expressed by the following formula: , in, Represents the cross-entropy loss function; Indicates batch size; Indicates the number of categories; Indicates sample The true label; Indicates sample The predicted probability; S4. After preprocessing, the electromyographic signals of the surface to be evaluated are input into the trained lower limb functional rehabilitation assessment model to obtain the assessment results.

[0024] Example 2 This embodiment provides an experimental comparison of the method of the present invention. The experiment was conducted on a workstation equipped with an NVIDIA RTX 3090 GPU (24GB VRAM), using the PyTorch 1.13.0 deep learning framework and CUDA version 11.8. The operating system was Windows 11, and the memory capacity was 64GB, ensuring efficient processing of large-scale electromyography signal data and model training. All code was implemented using Python 3.8, relying on scientific computing libraries including NumPy, Pandas, and Scikit-learn.

[0025] The experimental data were derived from clinically collected surface electromyography (EMG) signals. Specifically, 23 sets of lower limb EMG signals were collected from 13 stroke patients and 1 healthy individual while performing five standard movements: knee extension, knee flexion, plantar flexion, dorsiflexion, and standing. Raw signals from 10 major lower limb muscles were included, with a sampling frequency of 2048 Hz. The AdamW optimizer was used during training, with an initial learning rate set to 3 × 10⁻⁶. -5 The weight decay coefficient is 1×10 -4 To control overfitting, a batch size of 512 was set to fully utilize the parallel computing power of the GPU. Cross-entropy loss was chosen as the loss function, and the training cycle was 30 epochs. To augment the training set, data augmentation techniques were applied, including random scaling and the addition of Gaussian noise, enhancing the model's generalization ability.

[0026] After systematic training and evaluation, the MFFNet model proposed in this invention demonstrated outstanding performance in the Brunnstrom motor function grading task. As shown in Table 1, the complete MFFNet model achieved an accuracy of 93.10%, a macro-average F1 score of 0.9087, and a top-3 accuracy of 0.9904. This result indicates that the model can not only accurately predict the patient's motor function level, but also, even when an absolutely correct judgment cannot be made, it can still include the true class in the top three candidate predictions with a probability exceeding 99%, which provides extremely high practical value for clinical auxiliary diagnosis.

[0027] Table 1 Comparative Experiment Results As shown in Table 1, to gain a deeper understanding of the operational mechanisms of each component of MFFNet, this invention conducted systematic ablation and comparative experiments, the results of which are shown in Table 1. The four different networks involved in the comparison include: MFFNet (Time-only): It uses only the time-domain branch, processes the raw electromyographic temporal signals based on the Transformer encoder, captures long-term dependencies through a self-attention mechanism, and finally uses CLS tags for classification.

[0028] MFFNet (Freq-only): It uses only the frequency domain branch. First, it transforms the signal to the frequency domain, then extracts the spectral features through a multi-scale one-dimensional convolutional network (kernel size 3, 5, 7), and then feeds them into the classifier after global average pooling.

[0029] Transformer: The standard Transformer encoder architecture directly processes time-domain signals, including position encoding and multi-head self-attention layers, and relies entirely on temporal modeling capabilities for classification.

[0030] CNN: A traditional convolutional neural network that uses multiple layers of one-dimensional convolutions to extract local temporal features, gradually compresses the sequence dimension through pooling layers, and finally outputs the classification result through a fully connected layer.

[0031] The MFFNet variant using only the time-domain branch achieved an accuracy of 84.87%, an F1 score of 0.8099, and a top-3 accuracy of 0.9540, showing a significant performance decrease compared to the full model. This result indicates that although the Transformer architecture can effectively capture the temporal dependencies of electromyographic signals, the lack of frequency domain features limits the model's discriminative ability.

[0032] The model using only the frequency domain branch (Freq-only) slightly outperformed the time-domain variant, achieving an accuracy of 86.48%, an F1 score of 0.8264, and a top-3 accuracy of 0.9613. Multi-scale convolutional neural networks can extract features from different frequency bands from a frequency domain perspective, which is crucial for the classification of surface electromyography (EMG) signals, as the spectral characteristics of EMG signals are closely related to physiological parameters such as muscle activation state and fatigue level. However, the pure frequency domain model still scored 6.62 percentage points lower than the complete model, indicating that the lack of time-domain information also weakens the model's performance.

[0033] To verify the advantages of the two-stream architecture over the traditional single-modal model, this invention further compares the standalone Transformer model with a traditional Convolutional Neural Network (CNN). The pure Transformer model achieved an accuracy of 88.91%, an F1 score of 0.8565, and a top-3 accuracy of 0.9734. This performance outperforms the variant model of the single-modal architecture, demonstrating the powerful capabilities of the Transformer in temporal modeling. The traditional CNN model performed relatively weaker, with an accuracy of 82.55%, an F1 score of 0.7913, and a top-3 accuracy of 0.9386. This result is in line with expectations, as traditional CNNs lack the ability to model long-term temporal dependencies and do not fully utilize the frequency domain characteristics of electromyographic signals.

[0034] From the perspective of training dynamics, such as Figure 2 , Figure 3 As shown, the model learns rapidly within the first 15 epochs, with accuracy quickly increasing from an initial 26% to 72%. The performance on the validation set grows in sync with the training set, showing no significant deviation. Between epochs 15 and 25, the model enters a refined learning phase, with validation accuracy steadily improving from 72% to 90%. At this point, the performance curves of the training and validation sets almost overlap, indicating good generalization ability and no overfitting. After the 26th epoch, the model reaches a stable convergence state, with both training and validation accuracy stabilizing at around 92%, and the F1 score reaching 0.92, demonstrating robustness and stability. From a training efficiency perspective, the model converges within 30 epochs with stable training without drastic performance fluctuations, which is beneficial for fine-tuning and updating the model in practical deployments. The high consistency between training and validation set performance indicates that, through appropriate data augmentation and regularization strategies, deep learning models can learn well-generalized feature representations even on limited clinical data.

[0035] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for lower limb functional rehabilitation assessment based on a multidimensional feature fusion network, characterized in that, Includes the following steps: S1. Acquire multi-channel surface electromyography signals from stroke patients performing standard rehabilitation movements and preprocess them to construct a dataset; S2. Construct a lower limb functional rehabilitation assessment model, including a time-domain feature extraction branch, a frequency-domain feature extraction branch, a first cross-attention module, a second cross-attention module, and a classifier; input samples from the dataset into the lower limb functional rehabilitation assessment model to train the model; S3. Optimize the model using a loss function, update the parameters using the AdamW optimizer, minimize the loss using the backpropagation algorithm until the model converges, and obtain the trained lower limb functional rehabilitation assessment model. S4. After preprocessing, the electromyographic signals of the surface to be evaluated are input into the trained lower limb functional rehabilitation assessment model to obtain the assessment results.

2. The lower limb functional rehabilitation assessment method based on a multidimensional feature fusion network according to claim 1, characterized in that, Acquire raw multichannel electromyography data A standardized preprocessing procedure is performed: the signal of each channel is cropped using a thresholding method based on Z-scores, and the cropped EMG signal is processed using a Butterworth bandpass filter to retain the main components of the EMG signal from 20Hz to 150Hz, resulting in a clean EMG signal; the clean EMG signal is then processed using a downsampling method based on anti-aliasing filtering to obtain a downsampled EMG signal. The DC component of each channel in the downsampled electromyography (EMG) data is removed to obtain the preprocessed EMG signal.

3. The lower limb functional rehabilitation assessment method based on a multidimensional feature fusion network according to claim 2, characterized in that, In the temporal feature extraction branch, for each input sample The features at each time step are transformed from the original channel dimension through a linear projection layer. Mapping to the latent space dimension yields the projected sequence. For the projected sequence By adding positional encoding, we obtain the positional encoding sequence. ; in the positional encoding sequence A learnable special marker is concatenated at the starting position. The spliced ​​sequence is obtained. The formula is expressed as follows: ， in, Indicates feature concatenation operation; Represents a trainable parameter vector; concatenates the sequence The input is a module consisting of four stacked Transformer encoder layers, which is processed to obtain the encoded sequence. Take the encoded sequence The features at the starting position are used as the output of the temporal feature extraction branch to obtain the temporal features. The formula is expressed as follows: ， in, Indicates taking the encoded sequence Operations on the starting position feature.

4. The lower limb functional rehabilitation assessment method based on a multidimensional feature fusion network according to claim 3, characterized in that, In the frequency domain feature extraction branch, for each input sample Applying a Fast Fourier Transform along the time dimension to each channel calculates the amplitude spectrum, yielding the frequency domain representation. Representing the frequency domain Input a multi-scale one-dimensional convolutional module to obtain frequency domain features. .

5. The lower limb functional rehabilitation assessment method based on a multidimensional feature fusion network according to claim 4, characterized in that, The multi-scale one-dimensional convolution module includes three parallel one-dimensional convolution paths with kernel sizes of 3, 5, and 7 from top to bottom. Each path includes convolution, batch normalization, and a GELU activation function; the frequency domain features It is obtained by concatenating the outputs of each one-dimensional convolutional path, then processing it through a hybrid convolutional layer and global average pooling; the formula is expressed as follows: ， in, This represents parallel one-dimensional convolutions with kernels of 3, 5, and 7. This represents a 1x1 hybrid convolutional layer; This indicates a global average pooling operation.

6. The lower limb functional rehabilitation assessment method based on a multidimensional feature fusion network according to claim 5, characterized in that, The first and second cross-attention modules have the same structure and independent parameters, both employing a multi-head attention mechanism; they incorporate temporal features. As a query Frequency domain characteristics As a key Sum Input the first cross-attention module Temporal enhancement features were obtained. The formula is expressed as follows: ， in, Presentation layer normalization operation; This indicates the multi-head attention mechanism operation of the first cross-attention module; frequency domain features As a query Temporal characteristics As a key Sum Input the second cross-attention module Output frequency domain enhancement features The formula is expressed as follows: ， in, This indicates the multi-head attention mechanism operation of the second cross-attention module.

7. The lower limb functional rehabilitation assessment method based on a multidimensional feature fusion network according to claim 6, characterized in that, Temporal enhancement features Frequency domain enhancement features The features are concatenated along the feature dimension to obtain the fused features. ; to integrate features Input the classifier to obtain the predicted probability of each category, and take the category with the highest probability as the final evaluation and staging result.

8. The lower limb functional rehabilitation assessment method based on a multidimensional feature fusion network according to claim 7, characterized in that, The loss function used is the cross-entropy loss function.

Images