Dynamic graph convolution swimming stroke recognition system and method based on mutual information

Abstract

The invention belongs to the technical field of action recognition, and particularly relates to a dynamic graph convolution swimming stroke recognition system and method based on mutual information, and the method comprises the steps: carrying out the equally-divided rating of a sequence value interval of an obtained six-axis IMU signal, and obtaining the rated and quantized sequence data; calculating a standardized mutual information matrix of the input IMU signal according to the sequence data, and estimating an adjacent matrix through Laplacian normalization; the adjacent matrix is transmitted to a pre-trained graph convolution model to serve as a calculation unit of graph convolution, and feature representation of the six-axis IMU signals in real-time classified monitoring of different swimming strokes is obtained; the graph convolution model is a dynamic graph convolution neural network MIC-GCN model constructed based on a mutual information matrix; and based on the obtained features, classifying and predicting the output swimming stroke category as an output result label. The method can overcome the defect that the traditional deep learning model is low in swimming stroke category discrimination accuracy in the monitoring process of the six-axis IMU signal.

Description

Technical Field

[0001] The present invention relates to the technical field of action recognition, and in particular to a dynamic graph convolution swimming posture recognition system and method based on mutual information. Background Art

[0002] At present, the principle of the inertial measurement unit (IMU) sensor for swimming stroke recognition technology is mainly as follows: a device integrating sensors such as accelerometers and gyroscopes is used to measure the acceleration, angular velocity and amplitude of an object. In swimming, IMU sensors can be worn on the swimmer's body parts (such as wrists, calves, sacrum, etc.) to monitor and record the swimmer's posture, body posture and movements in real time. By tracking the swimmer's motion signals, the acceleration signal is combined with the angular velocity measurement of the gyroscope to analyze the swimmer's technical movements and thus identify the swimming stroke. As a daily wearable device, smart watches are portable and easy to use. Users of smart watches with integrated IMU sensors can perform swimming stroke recognition without carrying or installing other devices. In order to monitor the swimmer's movements and postures in real time and provide instant feedback. It is of practical significance to use smart watches for real-time swimming stroke monitoring to assist swimmers in timely adjusting their technical movements and improving swimming efficiency.

[0003] The current methods of using IMU signals for real-time monitoring of swimming postures mainly include three aspects: one is the discriminator method based on statistical rules; the second is to use the machine learning model to classify the feature information extracted from the IMU signal as input; the third is to use the traditional Euclidean distance-based DNN model (represented by CNN) to classify the IMU signal.

[0004] In this type of real-time monitoring method that uses IMU signals for swimming style recognition, the amount of information describing the swimming style categories of IMU signals is relatively incomplete, and there is a high possibility of misjudging and missing the swimming style categories during the actual monitoring process; in the process of using machine learning to classify and predict the swimming style of IMU signals, the key to obtaining a classification model with high accuracy is to extract discriminative features, but the extraction of such discriminative features requires a lot of prior knowledge, and such feature engineering is very time-consuming and laborious; in the process of using the traditional CNN model based on Euclidean distance to classify and predict the swimming style of IMU signals, since CNN describes the correlation between the time axis and sensor channel of the IMU signal through Euclidean distance, the relationship between different sensors and different axes of sensors for swimming style using IMU signals is an irregular one, and it is difficult to describe this correlation using Euclidean distance, that is, CNN at this time cannot maintain displacement invariance on such data, which leads to low accuracy in distinguishing the swimming style categories during the actual monitoring process.

[0005] Although deep learning represented by CNN is highly applicable to real-time monitoring of swimming postures using six-axis IMU signals, and also has the ability of automatic feature engineering, it can automatically extract discriminative features through training, and finally perform swimming posture recognition based on the extracted features. No other manual intervention is required in the middle, which is an end-to-end classification method.

[0006] However, the six-axis IMU signal is a two-dimensional signal, one dimension of which is time and the other dimension is the number of axes. For example, the IMU signal of a smart watch that integrates a three-axis accelerometer and a gyroscope includes six axes: x, y, z, roll, pitch, and yaw. The relationship between the axes is an irregular relationship, that is, a non-Euclidean distance relationship. At this time, CNN cannot maintain displacement invariance in the dimension of the direction axis.

[0007] Therefore, the CNN model that only uses convolution as a means of spatial feature extraction has room for improvement in the real-time monitoring of swimming strokes using the input six-axis IMU signal. A good approach is to introduce a method that can describe the non-Euclidean distance relationship between the axes of the six-axis IMU signal. The typical algorithm GCN in GNN with graph as input can achieve this function.

[0008] Usually, the correlation matrix or covariance matrix is ​​used in graph convolution as the adjacency matrix that represents the connection relationship of the nodes in the graph. The correlation coefficient in the correlation matrix can only describe the linear relationship between the variables, and the calculation of the correlation matrix coefficient usually assumes that the data is normally distributed. In addition, the signals of each axis of the six-axis IMU sensor are usually nonlinear, and considering the real-time monitoring of swimming posture, the input signal length cannot be too long, which makes it difficult for the input signal data to meet the normal distribution requirements. The mutual information matrix is ​​not affected by the data distribution and can more comprehensively describe the linear and nonlinear relationships between the variables; the covariance matrix mainly describes the linear relationship and variance between the variables, and is affected by the dimension of the variables, while the mutual information itself is a dimensionless statistic. Therefore, the mutual information matrix is ​​not affected by the dimension and can capture a wider range of relationship types, including nonlinear relationships and interdependent relationships. As the adjacency matrix of graph convolution, the mutual information matrix has the advantages of capturing nonlinear relationships, strong robustness to noise, dimensionlessness, avoiding the limitations of the correlation coefficient, and being applicable to non-normally distributed data. These advantages enable the mutual information matrix to more accurately describe the relationship between nodes in the graph convolution network, thereby improving the performance and accuracy of the model.

[0009] However, the mutual information matrix is ​​calculated by the probability of signal values, and the IMU signal data is usually of floating point type, and the probability of the same value is small, resulting in the mutual information value usually calculated being zero. In view of this, the present invention is proposed. Summary of the invention

[0010] In view of the technical defects mentioned in the background technology, the purpose of the embodiments of the present invention is to provide a dynamic graph convolution swimming style recognition system and method based on mutual information, aiming to overcome the defect that the traditional deep learning model has low accuracy in distinguishing the swimming style category during the monitoring process of the six-axis IMU signal.

[0011] To achieve the above objectives, in a first aspect, an embodiment of the present invention provides a dynamic graph convolution swimming style recognition system based on mutual information, the system comprising:

[0012] A preprocessing unit, used for performing average rating of the sequence value interval of the acquired six-axis IMU signal to obtain sequence data after rating quantization; wherein the sequence data includes sequences corresponding to the six axes respectively;

[0013] An adjacency matrix construction unit is used to calculate a standardized mutual information matrix of the input IMU signal according to the sequence data, and then estimate the adjacency matrix through Laplace normalization to fully express the irregular relationship between the axes of the six-axis IMU;

[0014] A feature extraction unit, used to transmit the adjacency matrix to a pre-trained graph convolution model as a graph convolution calculation unit to obtain a feature representation of the six-axis IMU signal for real-time classification monitoring of different swimming styles; wherein the graph convolution model is a dynamic graph convolution neural network MIC-GCN model constructed based on a mutual information matrix;

[0015] The classification result output unit is used to classify and predict the output swimming style category as an output result label based on the obtained features.

[0016] As a specific implementation of the present application, the average rating specifically includes:

[0017] ‌Determine the range‌, find the minimum and maximum values ​​in the sequence;

[0018] ‌Calculation interval‌, calculate the value range of the entire sequence, that is, the maximum value to the minimum value;

[0019] ‌Equally divide the interval‌, based on the difference between the maximum and minimum values, into N equal parts; where N is a positive integer and is the set value;

[0020] ‌Determine the level boundaries‌ and the value range of each equal part;

[0021] ‌Assign rank‌, assign each point in the sequence to the corresponding rank.

[0022] As a specific implementation of the present application, the training process of the graph convolution model specifically includes:

[0023] The collected historical data are labeled with samples for model training; wherein, the six-axis IMU signal with a fixed length is selected as the sample data and the corresponding swimming style category is selected as the sample label;

[0024] The fixed-length six-axis IMU signal collected in real time is normalized using the normalization parameters shared by the training set using the sliding window method and then used as input;

[0025] The classification prediction outputs the swimming style category as a label.

[0026] As a specific implementation of the present application, the feature extraction unit is also used for:

[0027] By sliding the IMU signal window and synchronously updating the calculated adjacency matrix and graph convolution to aggregate IMU feature information, the model avoids the problem of insufficient expression of discriminative features of different swimming styles, while improving the generalization performance of the model.

[0028] As a specific implementation method of the present application, the graph convolution model sets two graph convolution layers, uses ReLU as the activation function, and then uses a fully connected layer to linearly map the features extracted by GCN and input them into the classification layer.

[0029] In a second aspect, an embodiment of the present invention further provides a method for dynamic graph convolution swimming stroke recognition based on mutual information, which is applied to a dynamic graph convolution swimming stroke recognition system based on mutual information described in the first aspect, and the method comprises the following steps:

[0030] The acquired six-axis IMU signal is evenly graded in the sequence value interval to obtain the sequence data after the rating is quantified; wherein the sequence data includes the sequences corresponding to the six axes respectively;

[0031] The standardized mutual information matrix of the input IMU signal is calculated according to the sequence data, and then the adjacency matrix is ​​estimated by Laplace normalization to fully express the irregular relationship between the axes of the six-axis IMU;

[0032] The adjacency matrix is ​​transmitted to a pre-trained graph convolution model as a graph convolution calculation unit to obtain a feature representation of the six-axis IMU signal for real-time classification monitoring of different swimming styles; wherein the graph convolution model is a dynamic graph convolution neural network MIC-GCN model constructed based on the mutual information matrix;

[0033] Based on the obtained features, the classification prediction outputs the swimming style category as the output result label.

[0034] The technical solution provided by the embodiment of the present invention is to more comprehensively describe the relationship between the axes of the IMU sensor, and construct a dynamic graph convolutional neural network MIC-GCN model based on the mutual information matrix; the solution quantizes and grades the input IMU signal by averaging the ratings in the sequence numerical interval, and then calculates the standardized mutual information matrix of the input IMU signal, and then estimates the adjacency matrix by Laplacian normalization, which is used as the calculation unit of the graph convolution, fully expressing the irregular relationship between the axes of the six-axis IMU, and obtaining the feature representation of the real-time classification monitoring of the six-axis IMU signal in different swimming styles;

[0035] And by sliding the IMU signal window, synchronously updating the calculated adjacency matrix and graph convolution aggregation IMU feature information, the model avoids the problem of insufficient expression of discriminative features of different swimming strokes, and at the same time improves the generalization performance of the model, making the model proposed in the present invention more suitable for real-time classification and monitoring of different swimming strokes based on six-axis IMU signals, solving the problems of low accuracy and poor generalization performance of the existing technology in distinguishing swimming stroke categories during the monitoring process of six-axis IMU signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the specific implementation of the present invention or the technical solution in the prior art, the drawings required for use in the specific implementation or the description of the prior art will be briefly introduced below.

[0037] Figure 1 It is a principle block diagram of a dynamic graph convolution swimming style recognition system based on mutual information provided by an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of an equal-rating process provided by an embodiment of the present invention;

[0039] Figure 3 It is a flow chart of a dynamic graph convolution swimming style recognition method based on mutual information provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0040] 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 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.

[0041] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0042] It should be noted that, unless otherwise specified, the technical terms in this embodiment have the common meanings understood in the relevant technical field.

[0043] CNN: Convolutional Neural Networks, convolutional neural network;

[0044] DNN: Deep Neural Networks, deep neural network;

[0045] GNN: Graph Neural Networks, graph neural network;

[0046] GCN: Graph Convolutional Neural Networks, graph convolutional neural network;

[0047] MIC: Mutal Information Coefficient, mutual information coefficient;

[0048] IMU: Inertial measurement unit, inertial measurement unit;

[0049] ReLU: Rectified Linear Unit, linear rectification function;

[0050] CORR: Pearson Correlation Coefficient Matrix, Pearson correlation coefficient matrix;

[0051] COV: Covariance Matrix, covariance matrix.

[0052] Please refer to Figure 1 , an embodiment of the present invention provides a dynamic graph convolution swimming style recognition system based on mutual information, the system comprising:

[0053] The preprocessing unit is used to perform average rating of the sequence value interval of the acquired six-axis IMU signal to obtain the sequence data after the rating is quantified; wherein the sequence data includes the sequences corresponding to the six axes respectively; Figure 1 The standardization in is the rating quantification;

[0054] An adjacency matrix construction unit is used to calculate a standardized mutual information matrix of the input IMU signal according to the sequence data, and then estimate the adjacency matrix through Laplace normalization to fully express the irregular relationship between the axes of the six-axis IMU;

[0055] A feature extraction unit, used to transmit the adjacency matrix to a pre-trained graph convolution model as a graph convolution calculation unit to obtain a feature representation of the six-axis IMU signal for real-time classification monitoring of different swimming styles; wherein the graph convolution model is a dynamic graph convolution neural network MIC-GCN model constructed based on a mutual information matrix;

[0056] The classification result output unit is used to classify and predict the output swimming style category as an output result label based on the obtained features.

[0057] In this embodiment, if Figure 2 As shown, the average rating specifically includes:

[0058] ‌Determine the range‌, find the minimum and maximum values ​​in the sequence; that is, calculate the maximum and minimum values ​​of the sequence;

[0059] ‌Calculate the interval‌, calculate the value range of the entire sequence, that is, the maximum value to the minimum value; that is, subtract the minimum value from the maximum value;

[0060] ‌Equally divide the interval‌, divide the difference between the maximum value and the minimum value into N equal parts; where N is a positive integer and is a set value; that is, divide the difference into N equal parts;

[0061] ‌Determine the level boundaries‌ and the value range of each equal part;

[0062] ‌Assign levels‌, assign each point in the sequence to the corresponding level; thus obtaining the distribution of points in each interval.

[0063] This embodiment takes 5 levels as an example for illustration.

[0064] The construction process of the adjacency matrix is:

[0065] First, we obtain the quantized sequence of the six-axis IMU signal ratings: qx, qy, qz, qroll, qpitch, and qyaw; that is, the distribution of each interval;

[0066] Then calculate the normalized mutual information coefficient between qx, qy, qz, qroll, qpitch and qyaw ; That is, using entropy as the denominator The value is adjusted to between 0 and 1; the mutual information , and It represents the entropy of sequences X and Y, and the calculation formula is , where x and y represent points in the sequences X and Y, the base of log in the formula is b=2, and n represents the number of elements in the sequence. represents the probability of the i-th point in the sequence X; represents the joint probability of x and y, represents the marginal probability of x, represents the marginal probability of y.

[0067] Considering that there is a correlation between the six axes of the IMU, and the self-information of the same axis is 1, that is, the mutual information of the two same axes is equal

[0068] ;

[0069] in, represents the mutual information of sequences f and g, where f and g represent any sequence among qx, qy, qz, qroll, qpitch and qyaw, respectively. Represents the mutual information of sequences g and f. Therefore, the matrix constructed using the standardized maximum mutual information coefficient is as follows:

[0070] ;

[0071] Among them, the standardized mutual information matrix is ​​obtained by Represented, and then the adjacency matrix is ​​estimated using Laplacian symmetric normalization , the specific process is as follows:

[0072] ;

[0073] in Indicates The same order of identity matrix, where The value of is 6;

[0074] in is the unnormalized Laplacian matrix, is the degree matrix, is equal to the degree of node i, Indicated by The resulting diagonal matrix represents the inverse matrix of the square root of the degree matrix. i and j represent the matrix subscripts, which are positive integers from 1 to N. represents the symmetric normalized Laplacian matrix.

[0075] In this embodiment, the training process of the graph convolution model specifically includes:

[0076] The collected historical data are labeled with samples for model training; the fixed-length six-axis IMU signal is selected as the sample data and the corresponding swimming style category is selected as the sample label; the swimming style category includes five defined categories: freestyle, breaststroke, backstroke, butterfly stroke and others;

[0077] The fixed-length six-axis IMU signal collected in real time is normalized using the normalization parameters shared by the training set using the sliding window method and then used as input;

[0078] The classification prediction outputs the swimming style category as a label.

[0079] Specifically, the standardization method is to assume that a variable sequence is , L represents the window length; the mean and standard deviation are: and , then the standardized sequence is: ; Among them, the standardized sequence values ​​follow the normal distribution and are stored and The value of .

[0080] Based on the principle of GCN spatial feature extraction, compared with the convolution operation in CNN, the biggest difference of the convolution operation in GCN is that it has an explicit adjacency matrix expression. The process is: first, the attribute information is affine transformed, the interaction pattern between attribute features is learned, and then the features of neighboring nodes are iteratively aggregated to update the features of the current node. With each additional layer of convolution operation, the central node can integrate more information from the outer circle, which can more fully extract the feature information in the six-axis IMU signal and filter out the information with discrimination. While improving the accuracy of the six-axis IMU signal in the real-time classification monitoring method of different swimming styles, it also ensures the generalization performance of the classification system;

[0081] The feature extraction unit is also used for:

[0082] By sliding the IMU signal window and synchronously updating the calculated adjacency matrix and graph convolution to aggregate IMU feature information, the model avoids the problem of insufficient expression of discriminative features of different swimming styles, while improving the generalization performance of the model.

[0083] In this embodiment, the constructed GCN model (i.e., the graph convolution model) sets two graph convolution layers, uses ReLU as the activation function, and then uses a fully connected layer to linearly map the features extracted by GCN and input them into the classification layer.

[0084] In order to more intuitively demonstrate the effect of this solution, the IMU signal data of the actual swimming stroke was collected by a watch for verification, and accurate performance indicators were used: accuracy, recall rate, specificity and Matthew correlation coefficient.

[0085] The calculation formula of accuracy (Ac for short) is: ;

[0086] The calculation formula of recall rate (abbreviated as Re) is: ;

[0087] The calculation formula of specificity (Sp) is: ;

[0088] The calculation formula of Matthew correlation coefficient (Mcc) is:

[0089] ;

[0090] Here, TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative instances, respectively. The value range of the Matthew correlation coefficient MCC is from -1 to 1, where MCC -1 represents the worst possible prediction, and 1 represents the best possible prediction scheme; in addition, MCC 0 represents random prediction. The 20,000 labeled sample points are divided into training set, validation set, and test set in a ratio of 8:1:1; the number of sample points in the training set, validation set, and test set are 16,000, 2,000, and 2,000, respectively. The classification results of the trained MIC-GCN model and the CNN model with similar structure, CORR-GCN model (using correlation matrix to construct adjacency matrix) and COV-GCN model (using covariance matrix to construct adjacency matrix) on the test set are expressed by performance indicators as shown in Table 1 below.

[0091] It can be seen that compared with the CNN model, CORR-GCN model and COV-GCN model, the MIC-GCN model has obvious performance improvement on the test set used for real-time classification monitoring of six-axis IMU signals in different swimming styles.

[0092] Table 1 Performance data of each model

[0093]

[0094] It can be seen from Table 1 that under the same dataset, MIC-GCN has higher accuracy and correctness.

[0095] The IMU signal window is updated by sliding, and the input IMU signal is quantified and graded using the average rating of the sequence numerical interval, and the corresponding mutual information is calculated. The adjacency matrix is ​​dynamically calculated to extract graph information features.

[0096] In order to more comprehensively describe the relationship between the axes of the IMU sensor, the above scheme constructs a dynamic graph convolutional neural network MIC-GCN model based on the mutual information matrix; the scheme quantizes and grades the input IMU signal by averaging the ratings in the sequence numerical interval, and then calculates the standardized mutual information matrix of the input IMU signal, and then estimates the adjacency matrix through Laplacian normalization as the calculation unit of the graph convolution, which fully expresses the irregular relationship between the axes of the six-axis IMU and obtains the feature representation of the real-time classification monitoring of the six-axis IMU signal in different swimming styles;

[0097] And by sliding the IMU signal window, synchronously updating the calculated adjacency matrix and graph convolution aggregation IMU feature information, the model avoids the problem of insufficient expression of discriminative features of different swimming strokes, and at the same time improves the generalization performance of the model, making the model proposed in the present invention more suitable for real-time classification and monitoring of different swimming strokes based on six-axis IMU signals, solving the problems of low accuracy and poor generalization performance of the existing technology in distinguishing swimming stroke categories during the monitoring process of six-axis IMU signals.

[0098] Based on the same inventive concept, the embodiment of the present invention further provides a dynamic graph convolution swimming style recognition method based on mutual information, which is applied to a dynamic graph convolution swimming style recognition system based on mutual information described in the first aspect, referring to Figure 3 , the method comprises the following steps:

[0099] S101, performing average rating of the sequence value intervals on the acquired six-axis IMU signals to obtain sequence data after rating quantization; wherein the sequence data includes sequences corresponding to the six axes respectively;

[0100] S102, calculating a standardized mutual information matrix of the input IMU signal according to the sequence data, and then estimating an adjacency matrix through Laplace normalization to fully express the irregular relationship between the axes of the six-axis IMU;

[0101] S103, transmitting the adjacency matrix to a pre-trained graph convolution model as a graph convolution calculation unit to obtain a feature representation of the six-axis IMU signal for real-time classification monitoring of different swimming styles; wherein the graph convolution model is a dynamic graph convolution neural network MIC-GCN model constructed based on the mutual information matrix;

[0102] S104: Based on the obtained features, classify and predict the output swimming style category as an output result label.

[0103] Furthermore, the average rating specifically includes:

[0104] ‌Determine the range‌, find the minimum and maximum values ​​in the sequence;

[0105] ‌Calculation interval‌, calculate the value range of the entire sequence, that is, the maximum value to the minimum value;

[0106] ‌Equally divide the interval‌, based on the difference between the maximum and minimum values, into N equal parts; where N is a positive integer and is the set value;

[0107] ‌Determine the level boundaries‌ and the value range of each equal part;

[0108] ‌Assign rank‌, assign each point in the sequence to the corresponding rank.

[0109] In this embodiment, the training process of the graph convolution model specifically includes:

[0110] The collected historical data are labeled with samples for model training; wherein, the six-axis IMU signal with a fixed length is selected as the sample data and the corresponding swimming style category is selected as the sample label;

[0111] The fixed-length six-axis IMU signal collected in real time is normalized using the normalization parameters shared by the training set using the sliding window method and then used as input;

[0112] The classification prediction outputs the swimming style category as a label.

[0113] The method further comprises:

[0114] By sliding the IMU signal window and synchronously updating the calculated adjacency matrix and graph convolution to aggregate IMU feature information, the model avoids the problem of insufficient expression of discriminative features of different swimming styles, while improving the generalization performance of the model.

[0115] It should be noted that for a more specific description of the workflow of the method embodiment, please refer to the aforementioned system embodiment part, which will not be repeated here.

[0116] The entire solution is more suitable for real-time classification monitoring of different swimming styles based on six-axis IMU signals, which solves the problem of low accuracy and poor generalization performance of the existing technology in distinguishing swimming styles during the monitoring process of six-axis IMU signals.

[0117] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A dynamic graph convolution swimming style recognition system based on mutual information, characterized in that: The system comprises: A preprocessing unit, used for performing average rating of the sequence value interval of the acquired six-axis IMU signal to obtain sequence data after rating quantization; wherein the sequence data includes sequences corresponding to the six axes respectively; An adjacency matrix construction unit is used to calculate a standardized mutual information matrix of the input IMU signal according to the sequence data, and then estimate the adjacency matrix through Laplace normalization to fully express the irregular relationship between the axes of the six-axis IMU; A feature extraction unit, used to transmit the adjacency matrix to a pre-trained graph convolution model as a graph convolution calculation unit to obtain a feature representation of the six-axis IMU signal for real-time classification monitoring of different swimming styles; wherein the graph convolution model is a dynamic graph convolution neural network MIC-GCN model constructed based on a mutual information matrix; The classification result output unit is used to classify and predict the output swimming style category as an output result label based on the obtained features.

2. The system according to claim 1, characterized in that The average rating specifically includes: ‌Determine the range‌, find the minimum and maximum values ​​in the sequence; ‌Calculation interval‌, calculate the value range of the entire sequence, that is, the maximum value to the minimum value; ‌Equally divide the interval‌, based on the difference between the maximum and minimum values, into N equal parts; where N is a positive integer and is the set value; ‌Determine the level boundaries‌ and the value range of each equal part; ‌Assign rank‌, assign each point in the sequence to the corresponding rank.

3. The system according to claim 2, characterized in that The training process of the graph convolution model specifically includes: The collected historical data are labeled with samples for model training; wherein, the six-axis IMU signal with a fixed length is selected as the sample data and the corresponding swimming style category is selected as the sample label; The fixed-length six-axis IMU signal collected in real time is normalized using the normalization parameters shared by the training set using the sliding window method and then used as input; The classification prediction outputs the swimming style category as a label.

4. The system according to claim 3, characterized in that The feature extraction unit is also used for: By sliding the IMU signal window and synchronously updating the calculated adjacency matrix and graph convolution to aggregate IMU feature information, the model avoids the problem of insufficient expression of discriminative features of different swimming styles, while improving the generalization performance of the model.

5. The system according to claim 3, characterized in that The graph convolution model sets two graph convolution layers and uses ReLU as the activation function. Then, a fully connected layer is used to linearly map the features extracted by GCN and input them into the classification layer.

6. A dynamic graph convolution swimming style recognition method based on mutual information, characterized in that: The method for the mutual information-based dynamic graph convolution swimming style recognition system applied to claim 1 comprises the following steps: The acquired six-axis IMU signal is evenly graded in the sequence value interval to obtain the sequence data after the rating is quantified; wherein the sequence data includes the sequences corresponding to the six axes respectively; The standardized mutual information matrix of the input IMU signal is calculated according to the sequence data, and then the adjacency matrix is ​​estimated by Laplace normalization to fully express the irregular relationship between the axes of the six-axis IMU; The adjacency matrix is ​​transmitted to a pre-trained graph convolution model as a graph convolution calculation unit to obtain a feature representation of the six-axis IMU signal for real-time classification monitoring of different swimming styles; wherein the graph convolution model is a dynamic graph convolution neural network MIC-GCN model constructed based on the mutual information matrix; Based on the obtained features, the classification prediction outputs the swimming style category as the output result label.

7. The method according to claim 6, characterized in that The average rating specifically includes: ‌Determine the range‌, find the minimum and maximum values ​​in the sequence; ‌Calculation interval‌, calculate the value range of the entire sequence, that is, the maximum value to the minimum value; ‌Equally divide the interval‌, based on the difference between the maximum and minimum values, into N equal parts; where N is a positive integer and is the set value; ‌Determine the level boundaries‌ and the value range of each equal part; ‌Assign rank‌, assign each point in the sequence to the corresponding rank.

8. The method according to claim 7, characterized in that The training process of the graph convolution model specifically includes: The collected historical data are labeled with samples for model training; wherein, the six-axis IMU signal with a fixed length is selected as the sample data and the corresponding swimming style category is selected as the sample label; The fixed-length six-axis IMU signal collected in real time is normalized using the normalization parameters shared by the training set using the sliding window method and then used as input; The classification prediction outputs the swimming style category as a label.

9. The method according to any one of claims 6 to 8, characterized in that The method further comprises: By sliding the IMU signal window and synchronously updating the calculated adjacency matrix and graph convolution to aggregate IMU feature information, the model avoids the problem of insufficient expression of discriminative features of different swimming styles, while improving the generalization performance of the model.

Images