Data analysis system for upper limb function assessment

Abstract

The present invention relates to the technical field of data analysis, and discloses a data analysis system for upper limb function evaluation. The system includes image data acquisition, data preprocessing, bone node acquisition, bone node processing and function evaluation modules. Color image data collected from multiple angles and multiple scenes are used, and the image data is processed by Gaussian filtering for noise reduction, histogram equalization for contrast enhancement and normalization. The improved OpenPose algorithm is combined with a transfer learning strategy and an adversarial training mechanism to improve the accuracy of key point detection. The bone node processing adopts Kalman filter smoothing and a density-based spatial clustering algorithm to eliminate outliers. The function evaluation module adopts a deep learning method based on graph convolution to mine the spatial relationship between bone nodes and realize the quantitative evaluation of upper limb function. The present invention improves the accuracy and efficiency of the evaluation and reduces the manpower and time costs.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and in particular to a data analysis system for upper limb function assessment. Background Art

[0002] In traditional upper limb function assessment, doctors or rehabilitation therapists usually rely on subjective observation and manual measurement. This method is not only time-consuming and labor-intensive, but the evaluation results are easily affected by personal experience and subjective judgment. With the rapid development of computer vision and deep learning technology, it has become possible to use image processing and data analysis technology to objectively and automatically evaluate the upper limb function of the human body. However, existing image processing and data analysis methods still have many challenges in upper limb key point detection, data preprocessing, and functional evaluation, such as insufficient key point detection accuracy, data noise interference, and poor generalization ability of evaluation models. Therefore, it is particularly important to develop a data analysis system that can efficiently and accurately evaluate upper limb function. Summary of the invention

[0003] In view of the above-mentioned technical deficiencies, the purpose of the present invention is to provide a data analysis system for upper limb function assessment to solve the problems of insufficient upper limb key point detection accuracy and data noise interference in the prior art.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a data analysis system for upper limb function assessment, the system comprising: An image data acquisition module is used to collect real-time human upper limb activity image data; A data preprocessing module, used for performing data preprocessing on the collected image data; The skeleton node acquisition module uses the human upper limb key point detection algorithm to analyze the preprocessed image, identify and extract the coordinate position information of the human upper limb skeleton nodes; The skeleton node processing module performs smoothing processing on the acquired skeleton node coordinate data and removes abnormal values; The function assessment module adopts a deep learning method based on graph convolution, takes the processed bone node data as input, performs quantitative assessment of the upper limb function, and outputs the upper limb function assessment results.

[0005] Preferably, in a possible implementation manner of the first aspect, the human upper limb activity image data is color image data collected from multiple angles and multiple scenes, including images during stretching and rotation movements.

[0006] Preferably, in a possible implementation manner of the first aspect, the data preprocessing includes noise reduction, contrast enhancement and normalization; Use Gaussian filtering algorithm to reduce noise and remove noise interference in the image; Enhance contrast through histogram equalization method; Normalization maps the pixel value range of image data to the interval [0,1].

[0007] Preferably, in a possible implementation manner of the first aspect, the human upper limb key point detection algorithm adopts an improved OpenPose algorithm, which is trained using a human upper limb key point detection dataset, where the human upper limb key point detection dataset consists of historical human upper limb activity images and annotations; A transfer learning strategy is introduced during the training process, and the network parameters are initialized with the help of a pre-trained model of a large-scale general human posture dataset. An adversarial training mechanism is adopted, and a generative adversarial network structure is introduced to generate adversarial samples.

[0008] Preferably, in a possible implementation of the first aspect, the improved OpenPose algorithm specifically includes: Embedded attention mechanism module automatically learns the importance of different upper limb bone node features and pays more attention to key parts; The loss function is optimized, combining focal loss and L1 smoothness loss to give higher weights to difficult samples.

[0009] Preferably, in a possible implementation manner of the first aspect, the skeleton node processing includes: The Kalman filter algorithm is used to smooth the acquired bone node coordinate data; A density-based spatial clustering algorithm is used to remove outliers. The density of data points is calculated to remove outliers.

[0010] Preferably, in a possible implementation of the first aspect, the graph convolution-based deep learning method includes: Construct a multi-layer graph convolutional neural network, which consists of an input layer, a graph convolution layer, a pooling layer, a fully connected layer, and an output layer; The input layer receives the processed skeleton node data and converts it into graph structure data suitable for network processing; The graph convolution layer extracts features from graph structure data through spatiotemporal graph convolution operations to capture dynamic changes during upper limb movement. The pooling layer reduces the dimension of the features output by the graph convolution layer to reduce the amount of data; The fully connected layer maps the features output by the pooling layer to the final feature space; The output layer outputs the upper limb function assessment score and upper limb function assessment results based on the results of the fully connected layer.

[0011] Preferably, in a possible implementation of the first aspect, a graph convolution kernel approximated by a Chebyshev polynomial is used in the graph convolution layer, and the convolution kernel can perform a convolution operation on an irregular graph structure to extract local features between bone nodes; A batch normalization layer and ReLU activation function are added after each graph convolution layer. The batch normalization layer is used to normalize the input data to accelerate network convergence, and the ReLU activation function is used to increase the nonlinear expression ability of the network.

[0012] Preferably, in a possible implementation manner of the first aspect, the fully connected layer adopts Dropout technology to randomly discard some neuron connections to prevent network overfitting; During the network training process, the Adam optimizer is used to adjust the network parameters, with the goal of minimizing the loss function, and the network parameters are updated through the back propagation algorithm; The loss function adopts a mean square error loss function, which is used to measure the difference between the predicted upper limb function assessment result and the true value.

[0013] Preferably, in a possible implementation of the first aspect, the upper limb function assessment score interval is [0, 100], and the upper limb function assessment results are divided into five levels: excellent, good, medium, passing, and failing, and each level corresponds to a different preset assessment score interval.

[0014] The beneficial effects of the present invention are: by performing upper limb key point detection through an improved OpenPose algorithm, embedding an attention mechanism module and optimizing the loss function, the key point detection accuracy is improved, thereby providing a more accurate data basis for upper limb function assessment. At the same time, the deep learning method based on graph convolution can effectively mine the spatial relationship between bone nodes, making the assessment results more accurate and reliable.

[0015] The data preprocessing module performs noise reduction, contrast enhancement and normalization on the image data. The skeleton node processing module uses the Kalman filter algorithm for smoothing and the density-based spatial clustering algorithm to eliminate outliers, which effectively reduces data noise interference and improves data stability and reliability.

[0016] In the training process of the improved OpenPose algorithm, a transfer learning strategy and adversarial training mechanism are introduced. The pre-trained model of a large-scale general human posture dataset is used to initialize network parameters and generate adversarial samples, which enhances the generalization ability of the model and enables it to better adapt to upper limb function assessment in different scenarios and individuals.

[0017] This system can automatically collect, process and analyze upper limb activity image data to achieve quantitative evaluation of upper limb function. Compared with traditional subjective observation and manual measurement methods, it greatly improves evaluation efficiency and saves manpower and time costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 A structural diagram of a data analysis system for upper limb function assessment is provided for this application.

[0020] Explanation of the accompanying drawings: 1-image data acquisition module, 2-data preprocessing module, 3-skeletal node acquisition module, 4-skeletal node processing module, 5-function evaluation module. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] Embodiment 1: Figure 1 As shown, the present invention provides a data analysis system for upper limb function assessment, the system comprising: The image data acquisition module 1 is used to collect real-time human upper limb activity image data.

[0023] Specifically, a monocular camera is used as the image acquisition device, and the monocular camera is installed in a suitable position to completely and clearly capture the range of movement of the human upper limbs, ensure that the center of the camera's field of view is roughly aligned with the main area of ​​human upper limb movement, avoid occlusion, and obtain real-time human upper limb activity video. According to the video, the frame is cut at 5 times per second to obtain the human upper limb activity image data with timestamp.

[0024] The data preprocessing module 2 is used to perform data preprocessing on the collected image data.

[0025] Specifically, data preprocessing includes noise reduction, contrast enhancement and normalization. In this embodiment, the input image is first subjected to noise reduction by using a Gaussian filtering algorithm, using a 3x3 Gaussian kernel, whose standard deviation is automatically calculated by a function according to the kernel size, so as to eliminate Gaussian noise in the image.

[0026] Next, the contrast enhancement operation is performed by first converting the image to the YUV color space, performing histogram equalization on the Y channel, and then converting it back to the BGR color space.

[0027] Finally, normalization is performed to convert the contrast-enhanced image data type to float32 and divide it by 255.0 to map the pixel value range to the [0,1] interval, and finally return the preprocessed image.

[0028] The skeleton node acquisition module 3 uses the human upper limb key point detection algorithm to analyze the preprocessed image, identify and extract the coordinate position information of the human upper limb skeleton nodes.

[0029] Specifically, the human upper limb key point detection algorithm adopts the improved OpenPose algorithm, which is trained with the human upper limb key point detection dataset, which consists of historical human upper limb activity images and annotations.

[0030] In the dataset annotation process, we first collect historical images of human upper limb activities. These images cover scenes of different individuals and upper limbs in different postures and movements to ensure that the dataset is diverse and representative.

[0031] Subsequently, LabelMe was used for labeling. The labeled human upper limb bone nodes included the shoulder joint, elbow joint, wrist joint, as well as the bone connection points between the upper arm and forearm, and the key connection points between the palm and wrist. After the labeling is completed, a quality inspection link is set up, and a sampling ratio of no less than 20% of the total number of labeled images is used to check whether the coordinate positions of the labeled bone nodes such as the shoulder joint, elbow joint, and wrist joint meet the rules and whether there are any deviations or omissions.

[0032] Finally, the quality-checked annotated images and their corresponding skeletal node (shoulder joint, elbow joint, wrist joint, etc.) coordinate annotation information and the timestamp of the image are sorted and stored in JSON format to form a human upper limb key point detection dataset, which is divided into a training set and a validation set in the ratio of 8:2.

[0033] In this embodiment, the attention mechanism module AttentionModule is first defined. The module receives the number of channels of the input feature map as a parameter, and contains an adaptive average pooling layer nn.AdaptiveAvgPool2d(1) inside, which pools the input feature map into a feature vector of size 1x1. Next is a sequence nn.Sequential consisting of a fully connected layer, a ReLU activation function, another fully connected layer and a Sigmoid activation function. In the forward method, the batch size and number of channels of the input feature map are first obtained, and the feature vector is obtained by adaptive average pooling and reshaping. After being processed by the fully connected layer, the shape is adjusted to a four-dimensional tensor with the same number of channels as the input feature map, and finally it is multiplied element by element with the input feature map to achieve weighted attention to different feature channels.

[0034] The basic convolutional block ConvBlock is composed of the convolutional layer nn.Conv2d, the batch normalization layer nn.BatchNorm2d and the ReLU activation function nn.ReLU in sequence. In the forward method, the input data is processed by these three layers in sequence and then output.

[0035] The feature extraction module FeatureExtractor is built on the basis of some layers of VGG19. It contains multiple convolution blocks and the maximum pooling layer nn.MaxPool2d, which gradually extracts the features of the input image through multiple convolution and pooling operations. In the forward method, the input image is processed by these layers in turn, and the extracted feature map is finally output.

[0036] The confidence map and PAF (Part Affinity Fields) prediction module Stage1 of the first stage predicts PAF and confidence map respectively. For PAF prediction, the input feature map is processed by multiple convolution blocks and then passed through a convolution layer to obtain the PAF prediction result; for confidence map prediction, a similar structure is adopted, and the confidence map prediction result is obtained through a series of convolution blocks and a convolution layer.

[0037] The subsequent iterative refinement module RefinementStage first performs feature enhancement through the attention mechanism module AttentionModule before input processing, and then is processed through multiple convolution blocks, and finally obtains the prediction results of PAF and confidence map through the convolution layer respectively.

[0038] The improved OpenPose model ImprovedOpenPose includes a feature extraction module, a first-stage prediction module, and multiple iterative refinement modules. In the forward method, the input image first passes through the feature extraction module to obtain a feature map, and then the first-stage prediction module obtains the initial PAF and confidence map. The feature map, PAF, and confidence map are then concatenated as the input of the subsequent iterative refinement module, and are processed by multiple iterative refinement modules in turn to finally obtain a list of PAFs and confidence maps for each stage.

[0039] In terms of loss function, the focal loss function focal_loss is defined to solve the problem of class imbalance and give higher weights to difficult samples. The L1 smooth loss function smooth_l1_loss is also defined, and the two are combined to form combined_loss, which is used to calculate the total loss between the model prediction result and the target result.

[0040] According to the PAF and confidence map list in the final stage, the coordinate data of the bone nodes are calculated. First, for each confidence map corresponding to a specific bone node of the upper limb of the human body, the peak point is found through the non-maximum suppression algorithm to preliminarily determine the possible position of the bone node. Specifically, a non-maximum suppression function is defined first, which receives the confidence map and a parameter representing the neighborhood size (kernel size). Inside the function, half of the kernel size is first calculated as the filling amount. Then the input confidence map is filled by filling constants around to obtain a filled confidence map. Then create a suppression map with the same size as the original confidence map and all initial values ​​are 0. Then traverse each pixel of the original confidence map. For each pixel, extract a neighborhood window with the pixel as the center and the size of the kernel from the filled confidence map. If the confidence value of the current pixel in the original confidence map is the maximum value in this neighborhood window, the value of the corresponding position in the suppression map is set to the confidence value of the pixel in the original confidence map, otherwise keep the position in the suppression map as 0. Finally, the suppression graph is returned, and the retained peak points are the candidate positions of the corresponding bone nodes.

[0041] Next, PAF is used to verify and optimize these candidate positions. PAF describes the connection relationship between different bone nodes. For each pair of adjacent bone nodes (such as the shoulder joint and the elbow joint), starting from the peak point of each confidence map, the integration is performed along the direction of the PAF vector. It is expressed as starting from node A (coordinate ) to node B (coordinates are ), the PAF vector is , integrate along the path s connecting the two candidate points, the formula is ,in is the unit tangent vector on path s. If the cumulative sum is greater than the set threshold, the threshold used in this embodiment is 0.25, then the connection between the two candidate points is considered reasonable, thereby confirming the positions of the two skeletal nodes.

[0042] This embodiment introduces a transfer learning strategy during the training process, and initializes the network parameters with the help of a pre-trained model of a large-scale general human posture data set. When the improved OpenPose model is started to be trained, the parameters of the pre-trained model are loaded into the model, so that the model can use the feature representation learned by the pre-trained model on large-scale data to accelerate the convergence speed and improve performance. The adversarial training mechanism is adopted, and the generative adversarial network structure is introduced to generate adversarial samples. Specifically, a generator and a discriminator are constructed. The function of the generator is to generate adversarial samples similar to the distribution of real data, and the discriminator is responsible for distinguishing between real samples and generated adversarial samples. During the training process, the generator and the discriminator are trained adversarially, the generator continuously optimizes itself to generate more realistic adversarial samples, and the discriminator continuously improves its own discriminative ability. The generated adversarial samples are added to the training set, so that the improved OpenPose model learns these adversarial samples, enhances the robustness and generalization ability of the model, and enables it to accurately identify and extract the coordinate position information of the upper limb bone nodes of the human body when facing various complex situations.

[0043] The skeleton node processing module 4 performs smoothing processing on the acquired skeleton node coordinate data and removes abnormal values.

[0044] Specifically, the Kalman filter algorithm is first used to smooth the coordinate data of the bone nodes obtained from the continuous images. The kalman_smoothing function is defined, which receives the bone node coordinate data, and its shape is (number of frames, number of nodes, 2). Inside the function, the number of frames and nodes of the data is first obtained, and a zero matrix with the same shape as the input data is initialized to store the smoothed data. For each bone node, a Kalman filter is initialized, and the state transfer matrix, measurement matrix, process noise covariance and measurement noise covariance, as well as the initial state and covariance are set. Then each frame is traversed, the data is predicted and updated, and the first two elements of the updated state are stored in the result matrix as the smoothed coordinates, and finally the smoothed data is returned.

[0045] Then define the dbscan_outlier_removal function, which is used to remove outliers using the density-based spatial clustering algorithm (DBSCAN). The function receives the skeleton node coordinate data, the neighborhood radius eps, and the minimum number of samples min_samples as parameters. In the function, first get the number of frames and nodes of the data, and initialize an empty list to store the data of each node after removing outliers. For the data of each node, initialize all points to be unclassified, then traverse each point, calculate its distance to other points, and find the points in the neighborhood. If the number of points in the neighborhood reaches the minimum number of samples, mark the point as a core point, and mark the unclassified points in its neighborhood. Finally, extract the non-outlier points and add them to the list. In order to make the processed data shape match the original data, the maximum length of the processed data of all nodes is found, create a zero matrix, fill the processed data of each node into the matrix, and finally return the data after removing outliers.

[0046] Functional assessment module 5 adopts a deep learning method based on graph convolution, takes the processed bone node data as input, performs quantitative assessment of upper limb function, and outputs the upper limb function assessment results.

[0047] Specifically, a deep learning method based on graph convolution is used to use historical bone node coordinate data, timestamps, and historical upper limb function evaluation scores as a data set. The historical bone node coordinate data and timestamps use the bone node data and timestamps of the human upper limb key point detection data set in the bone node acquisition module 3. The historical upper limb function evaluation scores are obtained by medical personnel evaluating the upper limb activity. The collected historical bone node coordinate data, timestamps, and historical upper limb function evaluation scores are sorted and associated to construct a complete data set.

[0048] The input layer receives the processed bone node data and converts it into graph structure data suitable for network processing. According to the physiological structure and movement logic of the human upper limbs, the connection relationship between nodes is determined. For example, the shoulder joint is connected to the elbow joint, the elbow joint is connected to the wrist joint, etc. These connection relationships constitute the edges of the graph, reflecting the spatial and movement associations between the bone nodes. After completing the above steps, the adjacency matrix is ​​generated. The adjacency matrix is ​​used to represent the connection relationship between nodes. According to the defined node connection relationship, a matrix of size (number of nodes, number of nodes) is created. If there is a connection between two nodes, the value of the corresponding matrix position is 1, otherwise it is 0. Then the feature matrix is ​​created. The feature vectors of each node are arranged in node order to form a feature matrix of size (number of nodes, feature dimension). Finally, data integration and format adaptation are performed. The normalized adjacency matrix and feature matrix are combined and encapsulated into a Data object to form a complete graph structure data.

[0049] Then, we build the graph convolution layer and define the Graph class, whose main function is to build the graph structure. We initialize the maximum hop count and dilation rate, and call the get_edge method to obtain the edge information of the graph. Then, we use get_hop_distance to calculate the hop distance between nodes, and finally call get_adjacency to obtain the adjacency matrix of the graph. The get_edge method determines the number of nodes, self-connected edges, and adjacent edges of the graph according to different layouts, and the get_adjacency method normalizes the adjacency matrix.

[0050] The get_hop_distance function is used to calculate the hop distance between nodes. It determines the reachability between nodes by constructing a power matrix of the adjacency matrix, thereby obtaining the hop distance. The normalize_digraph function is used to normalize the adjacency matrix of a directed graph. The normalized matrix is ​​obtained by calculating the inverse of the degree matrix and multiplying it with the adjacency matrix.

[0051] The STGCNBlock class consists of a graph convolution layer (ConvTemporalGraphical) and a temporal convolution layer (tcn), including residual connections. In the forward method, the input data first passes through the graph convolution layer, then the temporal convolution layer, and finally is added to the residual connection and activated by the ReLU function.

[0052] The STGCN class is a complete spatiotemporal graph convolutional network. In the initialization method, the graph structure is loaded and the adjacency matrix is ​​registered. A network composed of multiple STGCNBlocks is constructed, and the importance of the edges is weighted. In the forward method, the input data is first normalized, then passed through each STGCNBlock in turn, and finally globally pooled and predicted.

[0053] The UpperLimbFunctionEvaluator class is used to evaluate the upper limb function and is inherited from nn.Module. In the __init__ method, an STGCN model is initialized for feature extraction with an output channel number of 256. Then an adaptive average pooling layer AdaptiveAvgPool1d is defined for dimensionality reduction, and two fully connected layers fc1 and fc2 are defined. Dropout is used in between to prevent overfitting. In the forward method, the input data is first extracted through the STGCN model, then reduced in dimensionality through the pooling layer, and then obtained through the fully connected layer. The predicted score is then limited to the [0,100] interval using torch.clamp.

[0054] Finally, according to the scores in the [0,100] interval output by the fully connected layer, the upper limb function assessment results are divided into five levels: excellent, good, medium, passing, and unpassive. The excellent score interval is [90,100], the good score interval is [80,90), the medium score interval is [70,80), the passing score interval is [60,70), and the unpassive score interval is [0,60].

[0055] The loss function is the mean square error loss function MSELoss, the optimizer is Adam, and the model is trained for 20 rounds. The processed bone node data is used as input to the trained model, and the upper limb function evaluation results are output.

[0056] In the second embodiment, the present invention provides a data analysis system for upper limb function assessment, the system also includes a rehabilitation training suggestion module for generating a personalized rehabilitation training program according to the upper limb function assessment results.

[0057] Specifically, this module realizes intelligent recommendation by constructing a rehabilitation knowledge graph, which contains entity relationships such as upper limb dysfunction type, rehabilitation stage division criteria, exercise therapy parameters, and clinical effect data. When the functional evaluation module outputs the evaluation score, the score is first mapped to the corresponding dysfunction level, and 17 kinematic parameters such as joint range of motion and movement coordination are extracted in combination with the motion trajectory characteristics of the bone nodes. The best rehabilitation stage is matched through a pre-trained random forest classifier. According to the stage characteristics, the rehabilitation action combination that meets the ISO 13482 standard is retrieved from the knowledge graph, including 9 basic action libraries such as resistance training and mirror therapy. Each action is associated with constraints such as intensity, frequency, and duration. For cases with limited elbow joint activity, the dynamic time warping algorithm is used to calculate the similarity between the current motion trajectory and the standard rehabilitation action template, and the training amplitude threshold is automatically adjusted. The reinforcement learning mechanism is introduced in the training program generation process, and a Q-Learning model with patient compliance and functional improvement rate as reward functions is constructed. The recommendation strategy is iteratively optimized through historical rehabilitation data, and the personalized program containing action sequence, execution parameters, and progress tracking indicators is finally output, and converted into JSON format for visual interface call.

[0058] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A data analysis system for upper limb function assessment, characterized in that: The system comprises: An image data acquisition module is used to collect real-time human upper limb activity image data; A data preprocessing module, used for performing data preprocessing on the collected image data; The skeleton node acquisition module uses the human upper limb key point detection algorithm to analyze the preprocessed image, identify and extract the coordinate position information of the human upper limb skeleton nodes; The skeleton node processing module performs smoothing processing on the acquired skeleton node coordinate data and removes abnormal values; The function assessment module adopts a deep learning method based on graph convolution, takes the processed bone node data as input, performs quantitative assessment of the upper limb function, and outputs the upper limb function assessment results.

2. The data analysis system according to claim 1, characterized in that: The human upper limb activity image data is color image data collected from multiple angles and multiple scenes, including images during stretching and rotation movements.

3. The data analysis system according to claim 1, characterized in that: The data preprocessing includes noise reduction, contrast enhancement and normalization; Use Gaussian filtering algorithm to reduce noise and remove noise interference in the image; Enhance contrast through histogram equalization method; Normalization maps the pixel value range of image data to the interval [0,1].

4. The data analysis system according to claim 1, characterized in that: The human upper limb key point detection algorithm adopts an improved OpenPose algorithm, which is trained using a human upper limb key point detection dataset, which is composed of historical human upper limb activity images and annotations; A transfer learning strategy is introduced during the training process, and the network parameters are initialized with the help of a pre-trained model of a large-scale general human posture dataset. An adversarial training mechanism is adopted, and a generative adversarial network structure is introduced to generate adversarial samples.

5. The data analysis system according to claim 4, characterized in that: The improved OpenPose algorithm specifically includes: Embedded attention mechanism module automatically learns the importance of different upper limb bone node features and pays more attention to key parts; The loss function is optimized, combining focal loss and L1 smoothness loss to give higher weights to difficult samples.

6. The data analysis system according to claim 1, characterized in that: Skeleton node processing includes: The Kalman filter algorithm is used to smooth the acquired bone node coordinate data; A density-based spatial clustering algorithm is used to remove outliers. The density of data points is calculated to remove outliers.

7. The data analysis system according to claim 1, characterized in that: The graph convolution-based deep learning method includes: The input layer receives the processed skeleton node data and converts it into graph structure data suitable for network processing; The graph convolution layer extracts features from graph structure data through spatiotemporal graph convolution operations to capture dynamic changes during upper limb movement. The pooling layer reduces the dimension of the features output by the graph convolution layer to reduce the amount of data; The fully connected layer maps the features output by the pooling layer to the final feature space; The output layer outputs the upper limb function assessment score and upper limb function assessment results based on the results of the fully connected layer.

8. The data analysis system according to claim 7, characterized in that: The graph convolution layer uses a graph convolution kernel approximated by Chebyshev polynomials, and performs convolution operations on irregular graph structures through the convolution kernel to extract local features between bone nodes; A batch normalization layer and ReLU activation function are added after each graph convolution layer. The batch normalization layer is used to normalize the input data to accelerate network convergence, and the ReLU activation function is used to increase the nonlinear expression ability of the network.

9. The data analysis system according to claim 8, characterized in that: The fully connected layer uses Dropout technology to randomly discard some neuron connections to prevent network overfitting; During the network training process, the Adam optimizer is used to adjust the network parameters, with the goal of minimizing the loss function, and the network parameters are updated through the back propagation algorithm; The loss function adopts a mean square error loss function, which is used to measure the difference between the predicted upper limb function assessment result and the true value.

10. The data analysis system according to claim 7, characterized in that: The upper limb function assessment score interval is [0,100], and the upper limb function assessment results are divided into five levels: excellent, good, medium, pass, and fail, and each level corresponds to a different preset assessment score interval.

Images