A kayaking action recognition method, system, storage medium and program product based on data analysis

Abstract

The present invention relates to the technical field of kayak motion recognition, and discloses a kayak motion recognition method, system, storage medium and program product based on data analysis, wherein the method comprises obtaining kayak motion data; the kayak motion data comprises kayak mechanical data, kayak kinematic data and kayak environmental data; a fully connected neural network based on sparse constraints is used to extract features of the kayak motion data to obtain kayak motion features; a high-order neural network based on self-similar features is used to perform kayak motion recognition on the kayak motion features. The present invention can improve the classification accuracy of kayak motion recognition.

Description

Technical Field

[0001] The invention relates to the technical field of kayaking action recognition, and in particular to a kayaking action recognition method based on data analysis. Background Art

[0002] Canoeing is a complex competitive and technical activity. The core movements and technical evaluations involved require comprehensive analysis based on multi-dimensional data of mechanics, kinematics, and environmental factors. These data include mechanical parameters of the paddle in water (such as lift, drag, reaction force, etc.), kinematic parameters (such as acceleration, angular velocity, change in direction of movement, etc.), and environmental factors (such as water flow velocity, water temperature, etc.). Effective data collection and analysis are of great significance to the evaluation of the standardization of athletes' movements and technical improvements. However, in the existing technology, canoeing data analysis still faces many technical bottlenecks.

[0003] The Chinese invention patent with publication number CN118861838A proposes a distribution network data analysis method and system based on machine learning, the method comprising: obtaining a distribution network data set to be processed; using discrete wavelet transform to perform data enhancement and feature extraction on the distribution network data set to be processed, and obtaining data features of multiple power node data; inputting the data features of multiple power node data into an integrated learning data classification model, and outputting distribution network data classification results; wherein the integrated learning data classification model is obtained by combining multiple classifiers through an integrated learning algorithm; performing time series analysis and numerical anomaly analysis on the distribution network data classification results to obtain the operating status of each power node; realizing data classification and status analysis of multiple types of distribution network data, improving the efficiency and accuracy of distribution network data classification, and realizing status analysis of power nodes.

[0004] The Chinese invention patent with publication number CN118863740A proposes a material intelligent estimation and management system based on the Internet of Things and machine learning. The system monitors the material weight and image in real time through the Internet of Things technology, uses machine learning algorithms to recognize images, automatically classifies materials and calculates inventory. The system includes a perception layer, a data analysis layer, an application layer, a user interaction layer and an infrastructure layer to realize functions such as intelligent inventory management, automatic procurement and inventory reminders. Compared with traditional methods, the present invention significantly improves material management efficiency and estimation accuracy, provides reliable data support for project progress control and cost-benefit analysis, and promotes the transformation of construction site material management to intelligence and efficiency.

[0005] The above-mentioned prior arts all utilize the traditional SMOTE algorithm to generate corresponding data, which makes it difficult to ensure the diversity and uniformity of the data. In particular, when the traditional SMOTE algorithm recorded in the above-mentioned prior arts is used to generate kayaking motion data, it is limited to the simple interpolation of minority samples of kayaking motion data, lacks sufficient coverage of the distribution of kayaking motion data samples, and is prone to cause model deviation.

[0006] At the same time, the above-mentioned existing technologies use traditional neural networks to process corresponding data features, which is prone to gradient vanishing. Especially when applied to kayaking motion data features, gradient explosion or local optimal solution problems may occur. Generally, the number of attributes of kayaking data may reach dozens or even hundreds. The extracted kayaking motion data features may be redundant and require large amounts of calculation, which may easily lead to poor model stability and generalization ability. Summary of the invention

[0007] In view of the above-mentioned deficiencies in the prior art, the present invention provides a kayaking action recognition method based on data analysis.

[0008] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0009] In a first aspect, a kayaking action recognition method based on data analysis comprises the following steps:

[0010] Acquire kayak motion data; the kayak motion data includes kayak mechanics data, kayak kinematics data and kayak environment data;

[0011] A fully connected neural network based on sparse constraints is used to extract features of kayaking motion data and obtain the kayaking motion characteristics. On the basis of traditional neural networks, neuron constraint sparsification and dynamic clipping activation strategies are adopted. During the training process of the neural network, the connection of each neuron is constrained, forcing the neural network to retain only the most informative connections in each layer. This can not only reduce the amount of calculation, but also avoid overfitting, thereby improving the robustness and generalization ability of the neural network model.

[0012] A high-order neural network based on self-similar features is used to identify kayaking motion features.

[0013] Furthermore, a fully connected neural network based on sparse constraints is used to extract features from the kayaking motion data to obtain kayaking motion features, including:

[0014] Initialize the parameters of the fully connected neural network;

[0015] Standardize kayaking action data;

[0016] The kayaking data input into the neural network is forward propagated through each layer of the neural network. The output of each layer will be used as the input of the next layer, and finally a set of high-dimensional feature representations will be output through layer-by-layer transmission;

[0017] According to the error of the feature extraction result, the total loss function is established for training by using the sparsity constraint term and the activation function constraint term;

[0018] The use of adjustable sparsity factors and activation parameters during training allows the neural network to automatically adjust its structure according to the different stages of the kayaking data;

[0019] Repeat the above steps until the preset stop iteration condition is met.

[0020] Furthermore, according to the error of the feature extraction result, the total loss function is established for training by using the sparsity constraint term and the activation function constraint term, which is:

[0021]

[0022] in, is the total loss function; The number of samples input to the neural network for the current batch; For the The true labels of samples; is the prediction output of the model; is the L2 norm; is the L2 regularization coefficient; is the Frobenius norm of the weight matrix; is the Frobenius norm; is the sparsification constraint coefficient; is the sparsity control term; Compute the norm for non-zero connectivity numbers; is the total number of layers in the neural network.

[0023] Furthermore, during the training process, an adjustable sparse factor and activation parameter are used to enable the neural network to automatically adjust its structure according to the different stages of the kayaking data, specifically:

[0024]

[0025]

[0026] in, is the sparse factor in the updated warp network; is the sparse factor in the neural network; and is a hyperparameter that controls the rate of change of sparsity and shear strength; is the error term of the neural network; For the neural network The activation output of the layer; is the shear factor in the updated neural network; is the shear factor in the neural network; is the L2 norm.

[0027] Furthermore, a high-order neural network based on self-similar features is used to perform kayaking action recognition on kayaking action features, including:

[0028] Initialize the parameters of the high-order neural network model;

[0029] During the training of the high-order neural network, the self-similarity structure in the kayaking data is calculated using the self-similarity loss function;

[0030] The error is propagated back through each layer through the back-propagation algorithm, and the gradient of the loss function of the high-order neural network with respect to the weight parameters of the high-order neural network is calculated;

[0031] Adopting adaptive learning rate mechanism to dynamically adjust the learning rate of high-order neural network during each gradient update process;

[0032] Repeat the above steps until the preset stop iteration condition is met.

[0033] Furthermore, before the feature data reaches the decision layer, during the training of the high-order neural network, the self-similarity structure in the kayaking data is calculated by the self-similarity loss function, specifically:

[0034]

[0035]

[0036] in, is the self-similarity loss function; is the number of samples before expansion; is the self-similarity weight; For the The sample input to the high-order neural network and the Similarity measure between samples input to high-order neural network; is the first sample input to the high-order neural network The weight of For the The first sample of the input to the high-order neural network Features For the The first sample of the input to the high-order neural network Features is the L2 norm; is the number of features.

[0037] Furthermore, after obtaining the kayaking motion data, the kayaking motion data is expanded based on the quantum probability-guided SMOTE method, including:

[0038] Set quantum states based on the existing minority class samples, where each sample corresponds to a quantum bit, and the superposition of these quantum states represents all possible sample states;

[0039] The probability amplitude of each quantum bit is adjusted through the probability amplitude modulation strategy in quantum computing;

[0040] Perform quantum measurement, randomly select states and collapse them according to the modulated quantum state probability distribution to generate new kayaking data points;

[0041] The newly generated kayak data points are corrected using a topological invariance strategy;

[0042] Repeat the above steps until the preset stop iteration condition is met.

[0043] Furthermore, the probability amplitude of each quantum bit is adjusted through the probability amplitude modulation strategy in quantum computing, specifically:

[0044]

[0045] in, After adjustment The probability amplitude of the samples; is a parameter that controls the sharpness of the distribution; For the The distance between a sample and its nearest neighbor; For the The complex probability amplitude of the samples; For the The complex probability amplitude of samples; is the number of samples before expansion.

[0046] In a second aspect, a kayaking action recognition system based on data analysis includes:

[0047] A data acquisition module, wherein the data acquisition module is used to obtain kayaking action data;

[0048] A first data feature extraction module, wherein the first data feature extraction module uses a fully connected neural network based on sparse constraints to extract features from the kayaking motion data to obtain kayaking motion features;

[0049] A data recognition module, the data recognition module is used to perform kayaking action recognition based on kayaking action features;

[0050] An output module is used to output the result of the kayaking action recognition.

[0051] In a third aspect, a computer-readable storage medium storing instructions is provided, wherein the storage medium stores a computer program or instructions, and when the computer program or instructions are executed by an image processing device, a kayaking action recognition method based on data analysis of the first aspect is implemented.

[0052] In a fourth aspect, a computer program product is provided, comprising: a computer program code, which, when executed, enables a processor to execute a kayaking motion recognition method based on data analysis according to the first aspect.

[0053] The present invention has the following beneficial effects:

[0054] 1. The present invention adopts the SMOTE algorithm guided by quantum probability, uses the superposition of quantum states and probability amplitude modulation to realize the expansion of minority class samples, generates more uniform and diverse samples, and uses the topological invariance strategy to correct the newly generated samples to ensure the consistency of the expanded samples with the original data in terms of numerical value and topological structure.

[0055] 2. The present invention proposes a neural network algorithm based on sparse constraints. By dynamically cutting activation functions and neuron sparsification constraint strategies, the feature extraction process is optimized, the computational complexity is reduced, overfitting is avoided, and the stability and generalization ability of the network are improved. A mechanism for dynamically adjusting the sparsity factor and activation parameters is adopted to automatically optimize the feature selection ability of the neural network according to the training process.

[0056] 3. The present invention proposes a high-order neural network classification algorithm based on self-similar features, combines the self-similarity loss function to enhance the classifier's ability to recognize kayaking motion patterns, and uses a weighted similarity metric to combine local and global information to dynamically adjust the classifier's calculation of feature similarity and improve classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flow chart of a kayaking action recognition method based on data analysis;

[0058] Figure 2 This is a comparison chart of the accuracy of different regularization methods;

[0059] Figure 3 The figure shows the comparison of model performance under different sparsity. DETAILED DESCRIPTION

[0060] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0061] like Figure 1 As shown, the embodiment of the present invention provides a kayaking action recognition method based on data analysis, comprising the following steps S1 to S3:

[0062] S1, obtaining kayak motion data; the kayak motion data includes kayak mechanical data, kayak kinematic data and kayak environmental data;

[0063] In an optional embodiment of the present invention, kayaking data involves mechanics, kinematics and environmental factors of kayaking motion, and sources of kayaking data collection mainly include kayaking mechanics data, kayaking kinematics data and kayaking environmental data.

[0064] The kayak mechanical data is collected from a mechanical sensing module installed on the paddle blade, specifically including information such as the lift, resistance, reaction force in the water, and resultant force of the paddle blade in the water;

[0065] The kayak kinematic data uses a three-axis acceleration sensor and an angular velocity sensor installed in the paddle shaft to collect kayak kinematic data, including acceleration, angular velocity, movement direction change and other information;

[0066] The kayaking environmental data uses water velocity sensors, temperature sensors and other equipment to collect information such as environmental variables.

[0067] The kayak data is transmitted to the kayak data storage platform in real time through the wireless sensor network. Specifically, the wireless module is connected to the transportation sensors to monitor and transmit the mechanical, kinematic and kayak environmental data in real time.

[0068] All collected kayaking data is stored in a standardized structured format, specifically CSV format.

[0069] In one embodiment, the properties of the kayak data include:

[0070] Ra represents lift (FL), which is the force perpendicular to the water surface exerted on the paddle blade in water; Da represents drag (FD), which is the water flow resistance exerted on the paddle blade in water; Fa represents reaction force (F), which is the reaction force exerted by water on the paddle blade; Fa represents resultant force (F), which is the resultant force of lift and drag; Acca represents acceleration, which is the acceleration of the athlete's paddle shaft; ωa represents angular velocity, which is the rotation speed of the paddle shaft; Δθa represents change in direction, which is the change in angle of the paddle shaft; Va represents water flow velocity, which is the water flow velocity in the water area where the athlete is located; Ta represents temperature, which is the water temperature in the sports area; SLa represents posture angle, which is the athlete's rowing posture angle (such as paddle shaft angle, torso posture, etc.).

[0071] It should be noted that this embodiment is only intended to illustrate a kayak data format and type of the present invention. In actual applications, the attributes of kayak data are usually more than 10 attributes, and the number of attributes of kayak data may reach dozens or even hundreds.

[0072] Furthermore, the collected kayaking data is annotated. The annotating method of the present invention is manual annotating. In one embodiment, the annotated categories include: kayaking standard training status, kayaking non-standard training status, kayaking no operation status, and kayaking other status.

[0073] It is understandable that in the task of the present invention, the collection, acquisition, labeling and preprocessing of kayaking training data are time-consuming and labor-intensive, and insufficient training samples are likely to lead to poor generalization ability of the model, while affecting the accuracy of the model. It is understandable that in the task of the present invention, the collection, acquisition, labeling and preprocessing of kayaking training data are time-consuming and labor-intensive, and insufficient training samples are likely to lead to poor generalization ability of the model, while affecting the accuracy of the model.

[0074] The present invention adopts the SMOTE algorithm guided by quantum probability for sample generation. The traditional SMOTE algorithm generates new samples by interpolating between minority class samples, but is often limited by the uniformity and diversity of sample distribution. The present invention uses quantum probability distribution to guide the interpolation step, so that the newly generated samples can better cover the potential kayaking data space, maintain the diversity of kayaking data, and enhance the model's generalization ability for new samples.

[0075] Specifically, the method of generating samples based on the quantum probability-guided SMOTE algorithm is expressed as follows:

[0076] S101. Set quantum states according to the existing minority class samples, where each sample corresponds to a quantum bit. The superposition of these quantum states represents all possible sample states. The initial quantum state is defined as:

[0077]

[0078] In the formula, represents the quantum state of the entire sample set, is the number of samples before expansion, For the The complex probability amplitude of samples, is the corresponding quantum ground state.

[0079] S102. The probability amplitude of each quantum bit is adjusted through the probability amplitude modulation strategy in quantum computing to reflect the probability distribution of sample generation. Specifically, the generation probability of new sample points is optimized according to the local density and category imbalance of each sample point to more accurately reflect the distribution of the minority class in the feature space. The probability amplitude modulation method is expressed as:

[0080]

[0081] In the formula, is the adjusted probability amplitude, For the The distance between a sample and its nearest neighbor; No. The Euclidean distance between a sample and its nearest neighbor; For the The complex probability amplitude of the samples; is a parameter that controls the sharpness of the distribution; here e is a natural constant.

[0082] In one embodiment, the distance between a sample and its nearest neighbor reflects the similarity and difference between the adjacent samples. By using a sine term to increase the nonlinear characteristics of the distance calculation, the probability amplitude modulation is made more sensitive to the subtle changes between samples. The calculation method is expressed as:

[0083]

[0084] In the formula, and They are The sample and its neighbors The sample in The eigenvalue of dimension, is the dimension of the features of the kayak sample dataset to be expanded, is a parameter that adjusts the nonlinear effects.

[0085] S103, perform quantum measurement, randomly select states and collapse according to the modulated quantum state probability distribution, and generate new kayaking data points. In this process, the random nature of quantum computing helps to explore and generate diverse samples, overcoming the possible overfitting and sample bias problems in the traditional SMOTE algorithm. The process of quantum state measurement and collapse is expressed as:

[0086]

[0087] In the formula, Indicates the measurement after The probability that the state of a sample is selected.

[0088] Furthermore, after quantum measurement, the state will be Collapse to the corresponding quantum ground state, based on which new kayaking data points are generated. The way to generate new sample points is expressed as:

[0089]

[0090] In the formula, is the newly generated sample point, and They are The sample and its neighbors samples, are interpolation coefficients randomly drawn from a uniform distribution in [0, 1].

[0091] S104. The newly generated samples are corrected by the topological invariance strategy to ensure that each new sample is not only reasonable in value, but also consistent in the overall structure of the kayak data. Specifically, the topological mapping of the kayak data is used to check and correct any new samples that may destroy the original topological structure of the kayak data. The topological correction method is expressed as:

[0092]

[0093] In the formula, is the new sample point after topology correction, is the correction intensity parameter; is a mapping function that aims to adjust the new sample to make it more consistent with the topological structure of the original kayaking dataset. Preferably, Set to 0.2.

[0094] In one embodiment, the mapping function is calculated as follows:

[0095]

[0096] In the formula, is the number of reference samples, It is The influence weight of each sample on the new sample, is a parameter used to adjust the sensitivity of the activation function. is the learning rate for sample generation, is the hyperbolic tangent function.

[0097] S105, merging the corrected newly generated samples with the original kayaking data set to ensure that the expanded set of kayaking data can be effectively used for subsequent model training, expressed as:

[0098]

[0099] In the formula, For the expanded kayaking dataset, This is the original kayaking dataset.

[0100] S2, using a fully connected neural network based on sparse constraints to extract features from kayaking motion data to obtain kayaking motion features;

[0101] In an optional embodiment of the present invention, the present invention uses a 6-layer fully connected neural network for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, they may encounter problems such as gradient vanishing, gradient explosion, or falling into local optimal solutions, which affects the stability of training and the performance of the model. The present invention uses a neural network algorithm based on sparse constraints as a feature extraction model. On the basis of traditional neural networks, neuron constraint sparsification and dynamic clipping activation strategies are adopted. During the training process of the neural network, the connection of each neuron is constrained, forcing the neural network to retain only the most informative connections in each layer. This can reduce the amount of calculation and avoid overfitting, thereby improving the robustness and generalization ability of the neural network model.

[0102] Specifically, the training process of the neural network algorithm based on sparse constraints is as follows:

[0103] S201, initialize the parameters of the neural network. Each layer of the neural network contains multiple neurons, and the number of connections of each layer of neurons is subject to sparse constraints. Represents the neural network The weight matrix of the layer, Represents the neural network The bias term of the layer is initialized as:

[0104]

[0105]

[0106] In the formula, For the neural network The weight matrix of the layer; It means that the mean is zero and the variance is Normal distribution of is a normal distribution; Initialize the variance for the neural network's parameters; Indicates compliance with a specific distribution; For the neural network Preferably, Set to 0.001.

[0107] S202, the expanded kayaking training data is standardized to scale the range of the kayaking data input into the neural network to a uniform scale to avoid gradient explosion or disappearance during the training process. The standardized method is expressed as:

[0108]

[0109] In the formula, is the kayaking data input to the neural network, that is, the expanded kayaking data; is the mean of the kayaking data input to the neural network; is the standard deviation of the kayaking data input to the neural network; The normalized kayaking data.

[0110] S203, the kayaking data input into the neural network is forward propagated through each layer of the neural network, and the output of each layer will be used as the input of the next layer. After being transmitted layer by layer, a set of high-dimensional feature representations is finally output. The forward propagation method is expressed as:

[0111]

[0112] In the formula, For the neural network The activation output of the layer, i.e., the neural network The activation input of the layer; For the neural network The activation output of the layer, i.e., the neural network The activation input of the layer; is the dynamic clipping activation function.

[0113] In one embodiment, the dynamic clipping activation function adjusts the sparsity and activation value according to the output of each layer of the neural network, thereby optimizing the activation mode of the neural network at different training stages. The calculation method is expressed as:

[0114]

[0115] In the formula, is the result of linear transformation, specifically the input of the dynamic shear activation function; is the shear strength factor; is the clipping threshold; is the indicator function, which means if If yes, it is 1, otherwise it is 0; is the maximum value function. Preferably, Set to 0.3, Set to 0.1.

[0116] S204, the training goal of the neural network is to minimize the error of generating feature representation in the feature extraction process. The loss function of the present invention not only considers the error of the feature extraction result, but also uses the sparsity constraint term and the activation function constraint term to focus on the optimal sparsity of the neural network structure. The calculation method of the loss function is expressed as:

[0117]

[0118] In the formula, is the total loss of the neural network; For the The true labels of samples; is the predicted output of the model, which is calculated by the feature vector obtained by the preset Softmax after neural network feature extraction; is the L2 norm; is the L2 regularization coefficient; is the Frobenius norm of the weight matrix, indicating the size of the weight; is the Frobenius norm; is the sparsification constraint coefficient; is the sparsity control term, indicating the The number of non-zero connections of neurons in the layer; Compute the norm for non-zero connectivity numbers; is the number of samples input to the neural network in the current batch. Preferably, Set to 0.3, Set to 0.4.

[0119] Furthermore, based on the loss function, the neural network parameters are updated by error back propagation. In the process of error back propagation, the gradient of the loss to each layer weight is calculated by the chain rule. For the weight and bias of each layer, the calculation method of the gradient update is expressed as:

[0120]

[0121]

[0122] In the formula, represents partial derivative; For the neural network The error term of the layer.

[0123] Furthermore, the calculation method of the error term of the neural network is expressed as:

[0124]

[0125] In the formula, is the derivative of the dynamic shear activation function; For the neural network The transpose of the layer's weight matrix; For the neural network The error term of the layer.

[0126] S205. As the training process progresses, the sparsity and activation strategy of the neural network will be dynamically adjusted. Specifically, when the kayaking training data becomes more complex, the neural network will adjust the number of activated features by dynamically cutting the activation function, allowing more features to participate in the learning process of the neural network. At the same time, the sparsity constraint will be dynamically relaxed or strengthened to ensure that the neural network maintains appropriate feature selectivity during the learning process. Specifically, the use of adjustable sparsity factors and activation parameters enables the neural network to automatically adjust its own structure according to different stages of the kayaking data. The calculation method is expressed as:

[0127]

[0128]

[0129] In the formula, and is a hyperparameter that controls the rate of change of sparsity and shear strength; It is the sparsity factor in the neural network, which enables the neural network to dynamically control the sparsity of each layer according to the current training progress, and enhance the adaptability of the model under specific tasks; is the sparse factor in the updated warp network; is the shear factor in the neural network; is the shear factor in the updated neural network; is the error term of the neural network. Preferably, and Set to 0.3 and 0.4 respectively.

[0130] S206, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0131] Figure 2As shown in the figure, in order to verify the effectiveness of the neural network regularization method and the advantage of sparse constraints in maintaining model performance, the experimental results are compared with traditional L2 regularization, Dropout method and non-regularization method. The accuracy curve of this technology shows a stable upward trend, eventually reaching about 93%, and the convergence speed is the fastest. Traditional L2 regularization has an over-smoothing problem, and its accuracy is lower than that of this technology. The Dropout method fluctuates greatly in the early stage of training and finally stabilizes at about 90%. The non-regularization method has the lowest accuracy and the slowest convergence, which verifies the necessity of regularization. Therefore, this technology effectively prevents overfitting while maintaining model capacity through dynamic sparse constraints.

[0132] Figure 3 As shown in the figure, in order to verify the improvement of parameter efficiency by dynamic sparse constraints, the model performance is compared under different sparsities. The experimental results show that the performance of conventional methods decreases when the sparsity is >0.6, while this technology maintains the highest accuracy at all sparsities, and has the best performance in the sparsity range of 0.5-0.7, reflecting the adaptive advantage.

[0133] S3. Use a high-order neural network classification model to evaluate the energy consumption of photovoltaic power generation equipment based on the dimensionality reduction features of the energy consumption of photovoltaic power generation equipment.

[0134] In an optional embodiment of the present invention, the present invention adopts a high-order neural network classification algorithm based on self-similar features as a classifier model. On the basis of the traditional high-order neural network, the present invention utilizes a self-similarity loss function to enable the model to strengthen the learning of self-similar features during the training process, thereby improving the model's ability to recognize action patterns.

[0135] Specifically, the training process of the high-order neural network classification algorithm based on self-similar features is as follows:

[0136] S301, initialize the high-order neural network model, including the input layer of the high-order neural network, three hidden layers of the high-order neural network and the output layer of the high-order neural network, the input layer of the high-order neural network receives the kayaking data from the feature processing module, the hidden layer of the high-order neural network uses a specific activation function for nonlinear transformation, and the output layer of the high-order neural network is used for the probability output of the multi-classification problem. Specifically, the kayaking data input to the high-order neural network is defined as , is the number of samples input to the high-order neural network, is the feature dimension of each sample input to the high-order neural network. The neuron output of the layer is calculated as:

[0137]

[0138] In the formula, For high-order neural networks The weight matrix of the layer, For high-order neural networks The input of the layer (for the 0th layer, i.e., the input layer, ), For high-order neural networks The bias term of the layer, For high-order neural networks The pre-activation output of the layer.

[0139] Furthermore, the activation function is used to perform nonlinear transformation on the pre-activation output of the high-order neural network to obtain the layer output, which is expressed as:

[0140]

[0141] In the formula, For high-order neural networks The output of the layer, is the activation function of the high-order neural network.

[0142] It should be noted that in traditional high-order neural networks, the weight matrix It is randomly initialized from a certain distribution (such as normal distribution). In order to better converge in training, the present invention adopts a weight initialization strategy based on the characteristics of kayaking data, which is expressed as:

[0143]

[0144] In the formula, is the high-order neural network calculated by the kayaking data features. The weight matrix of the layer, It is a high-order neural network The characteristic dimension of the layer; It is a high-order neural network The first learning coefficient of the layer is a training parameter.

[0145] Moreover, when calculating the bias of each layer, the characteristic relationship with the input kayak data is adopted, which makes the calculation of the bias term more flexible. The bias term not only depends on the output of the previous layer, but also adjusts according to the specific characteristics of the input kayak data. The calculation method is expressed as:

[0146]

[0147] In the formula, It is a high-order neural network The second learning coefficient of the layer is a training parameter; is the weight coefficient associated with each feature, represents the first kayak data input to the high-order neural network Features.

[0148] Moreover, the activation functions (such as ReLU and Sigmoid) used in traditional high-order neural networks may cause gradient vanishing or gradient exploding problems in some cases. The present invention adopts an adaptive activation function, combined with the characteristics of the high-order neural network layer and the distribution of kayaking data, and the calculation method is expressed as:

[0149]

[0150] In the formula, is the hyperbolic tangent function; is the adaptive adjustment factor, which indicates the activation response degree of each layer; It is a high-order neural network The third learning coefficient of the layer is a training parameter; is the number of ReLU activations; It is the average similarity between all samples input into the high-order neural network.

[0151] S302, in the training process of the high-order neural network, the learning ability of the model is enhanced by calculating the self-similarity structure in the kayaking data, wherein the self-similarity feature in the kayaking data indicates that certain information has similar feature patterns at different scales or time periods, which is specifically achieved by optimizing the self-similarity loss function, wherein the self-similarity loss function is used to measure the similarity between different samples. Assume that for two samples and , is the first sample input to the high-order neural network, is the second sample input to the high-order neural network, and the similarity between the two in the feature space is expressed as , the similarity is calculated by the cosine similarity metric, and the calculation method of the self-similarity loss function is defined as:

[0152]

[0153] In the formula, is the self-similarity loss function; is the self-similarity weight, which represents the sample and samples The importance of similarity between For the The sample input to the high-order neural network and the A similarity measure between samples input to a high-order neural network.

[0154] In one embodiment, for the first sample input to the high-order neural network and the second sample input to the high-order neural network, in order to measure the similarity between the two samples, a weighted similarity metric is used in combination with local and global information in the kayaking data, so that the model can dynamically adjust the similarity metric according to the contribution of each feature during the training process, thereby more accurately capturing the self-similarity of the kayaking data. The calculation method is expressed as:

[0155]

[0156] In the formula, is the first sample input to the high-order neural network. The weights of the features indicate the importance of different features to the similarity calculation; is the first sample input to the high-order neural network. Features is the second sample input to the high-order neural network. Features.

[0157] S303. After completing the forward propagation and the adoption of the self-similar structure, the high-order neural network performs error feedback, and the prediction performance of the model is evaluated by using the loss function by calculating the loss between the predicted value and the actual label. The error is back-propagated through each layer through the back-propagation algorithm, and the gradient of the loss function of the high-order neural network with respect to the weight parameters of the high-order neural network is calculated as follows:

[0158]

[0159] In the formula, is the loss function of the high-order neural network; For high-order neural networks The transpose of the output of the layer, is the gradient effect of self-similarity loss on weights.

[0160] S304. In order to ensure the stability of the model training process, an adaptive learning rate mechanism is used. In each gradient update process, the learning rate will be dynamically adjusted according to the size and change of the current gradient. When the gradient changes greatly, the learning rate will be appropriately reduced to avoid training instability caused by too fast update. When the gradient changes slightly, the learning rate will be appropriately increased to accelerate convergence. The adjustment method of the learning rate of the high-order neural network is expressed as:

[0161]

[0162] In the formula, is the initial learning rate of the high-order neural network; For the The learning rate of the high-order neural network in iterations; For the The factor that controls the change of learning rate in the iteration is a training parameter; For the The L2 norm of the gradient of the iteration. Preferably, Set to 0.01.

[0163] S305, repeat the above steps until the preset stop iteration condition is met, which means the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0164] After the model training is completed, the trained model is used to classify the kayak training status. In one embodiment, the collected original kayak data is input into the model in the trained feature processing module for feature processing, and further, the processed features are input into the model in the decision module for classification, thereby obtaining the classification result. In this embodiment, the classification categories include: kayak standard training status, kayak irregular training status, kayak no operation status, and kayak other status.

[0165] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0166] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0167] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0168] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0169] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A kayaking action recognition method based on data analysis, characterized in that: The following steps are involved: Acquire kayak motion data; the kayak motion data includes kayak mechanics data, kayak kinematics data and kayak environment data; The fully connected neural network based on sparse constraints is used to extract the features of kayaking motion data and obtain the kayaking motion features; The feature extraction method comprises: Initialize the parameters of the fully connected neural network; Standardize kayaking action data; The kayaking data input into the neural network is forward propagated through each layer of the neural network. The output of each layer will be used as the input of the next layer, and finally a set of high-dimensional feature representations will be output through layer-by-layer transmission; According to the error of the feature extraction result, the total loss function is established for training by using the sparsity constraint term and the activation function constraint term; The use of adjustable sparsity factors and activation parameters during training allows the neural network to automatically adjust its structure according to the different stages of the kayaking data; Repeat the above steps until the preset stop iteration condition is met; Perform feature dimension reduction on kayaking action features to obtain kayaking action features; Performing kayaking action recognition on the kayaking action features; The error of the feature extraction result is based on the sparsity constraint term and the activation function constraint term to establish a total loss function for training, specifically: in, is the total loss function; The number of samples input to the neural network for the current batch; For the The true labels of samples; is the prediction output of the model; is the L2 norm; is the L2 regularization coefficient; is the Frobenius norm of the weight matrix; is the Frobenius norm; is the sparsification constraint coefficient; is the sparsity control term; Compute the norm for non-zero connectivity numbers; is the total number of layers of the neural network; The adjustable sparse factor and activation parameters are used in the training process to enable the neural network to automatically adjust its structure according to different stages of the kayaking data, specifically: , in, is the sparse factor in the updated warp network; is the sparse factor in the neural network; and is a hyperparameter that controls the rate of change of sparsity and shear strength; is the error term of the neural network; For the neural network The activation output of the layer; is the shear factor in the updated neural network; is the shear factor in the neural network; is the L2 norm.

2. A kayaking action recognition method based on data analysis according to claim 1, characterized in that: A high-order neural network based on self-similar features is used to identify the kayaking action features.

3. The kayaking action recognition method based on data analysis according to claim 1 is characterized in that: After obtaining the kayaking motion data, the kayaking motion data is expanded based on the SMOTE method guided by quantum probability.

4. A computer-readable storage medium storing instructions, characterized in that: The storage medium stores a computer program or instruction. When the computer program or instruction is executed by the image processing device, a kayaking action recognition method based on data analysis as described in any one of claims 1 to 3 is implemented.

5. A computer program product, characterized in that The computer program product comprises: a computer program code, which, when executed, enables a processor to execute a kayaking motion recognition method based on data analysis according to any one of claims 1 to 3.

Images