Physical training model construction method and device, equipment, storage medium and product

Abstract

The invention discloses a physical training model construction method and device, equipment, a storage medium and a product. Training data is divided into a plurality of training scene modules, and a data link network is constructed based on the training scene modules; performing entity recognition and association relationship quantification on the training scene module to obtain implicit field semantic features; carrying out dimension reduction processing on the data link network to obtain dimension reduction semantic features; performing semantic element aggregation on the data link network to obtain deep semantic association relationship features; fusing the data chain network, the implicit field semantic feature, the dimension reduction semantic feature and the deep semantic association relationship feature, and then performing implicit time sequence feature recognition to obtain a dynamic semantic feature; and training a generative adversarial network module by adopting the dynamic semantic features and the training data to obtain a physical training model. According to the method, the semantic features of the physical training data can be maximized and used for training of the physical training model, so that the performance of the physical training model is improved.

Description

Technical Field

[0001] The present invention relates to the field of model construction and training technology, and in particular to a sports training model construction method, device, equipment, storage medium and product. Background Art

[0002] With the rapid rise of national fitness, the shortage of sports coach resources is particularly prominent. Faced with the reality of the shortage of sports coach resources, artificial intelligence technology is an effective way to supplement it. The intelligent training model can collect the sports training data of athletes through high-precision sensors and specific terminals, and analyze the sports training data to generate personalized training feedback and suggestions to help athletes optimize the training effect. However, sports training data is of various types and huge quantities, and the structure and source are different. The current intelligent training model cannot effectively establish the association between multi-source heterogeneous data, resulting in the inability to fully tap the value of sports training data and provide users with accurate sports training guidance. Summary of the invention

[0003] The purpose of the embodiments of the present invention is to provide a sports training model construction method, device, equipment, storage medium and product, which can maximize the semantic features of sports training data and use them for the training of sports training models, thereby fully tapping the value of sports training data and providing users with more accurate sports training guidance.

[0004] To achieve the above objectives, an embodiment of the present invention provides a method for constructing a sports training model, comprising:

[0005] Acquire a plurality of pre-divided training scene modules; wherein the training scene modules include a plurality of types of training data;

[0006] Vertically link the same type of training data between different training scenario modules to obtain a data link network;

[0007] Respectively perform entity recognition in each of the training scene modules in the data link network and quantify the association relationship between different entities recognized to obtain implicit domain semantic features of each of the training scene modules;

[0008] Performing dimensionality reduction processing on the data link network to obtain dimensionality reduction semantic features;

[0009] The graph attention network module is called to aggregate the semantic elements of the training data within the data link network to obtain deep semantic association relationship features;

[0010] The data link network, the implicit domain semantic features, the dimension reduction semantic features and the deep semantic association relationship features are integrated to obtain enhanced data link semantic features;

[0011] Performing implicit temporal feature recognition on the enhanced data link semantic feature to obtain a dynamic semantic feature;

[0012] The dynamic semantic features and the training data are used to train a generative adversarial network module to obtain a sports training model for guiding the sports training process.

[0013] As an improvement of the above solution, the vertical linking of the same type of training data between different training scenario modules to obtain a data link network includes:

[0014] Extracting semantic elements of each of the training data, and converting the semantic elements into semantic vectors;

[0015] Summarizing the semantic vectors in the training scenario modules to obtain a horizontal data subchain;

[0016] Based on the degree of interaction between the semantic elements in different training scenario modules, an interaction matrix is ​​constructed;

[0017] Multiplying the semantic vector and the interaction matrix to obtain a vertical data subchain;

[0018] Based on the horizontal data sub-chain and the vertical data sub-chain, a data chain network is constructed.

[0019] As an improvement of the above scheme, the entity recognition is performed in each training scene module in the data link network and the association relationship between the different entities recognized is quantified to obtain the implicit domain semantic features of each training scene module, including:

[0020] Performing entity recognition on the training data within the training scenario module and quantifying the relationship between entities;

[0021] A knowledge network is constructed based on the identified entities and relationships to obtain implicit domain semantic features of the training scenario module.

[0022] As an improvement of the above solution, the data link network is subjected to dimensionality reduction processing to obtain dimensionality reduction semantic features, including:

[0023] Calculating a first similarity between first training data and other training data in the data link network; wherein the first training data is any one of all the training data;

[0024] Based on the dimensional mapping relationship, the training data in the high-dimensional space is mapped to the low-dimensional space to obtain the dimensionality reduction semantic features of the training data, and the second similarity between the first training data and other training data in the low-dimensional space is calculated;

[0025] Substituting the first similarity and the second similarity into a pre-constructed loss function to calculate a loss value;

[0026] The gradient descent method is used to adjust the dimensional mapping relationship according to the loss value, and return to the step of calculating the first similarity until the loss value meets the set adjustment end condition or the number of adjustments to the dimensional mapping relationship reaches a set number, and the reduced dimensionality semantic feature is output.

[0027] As an improvement of the above solution, the implicit temporal feature recognition of the enhanced data link semantic feature to obtain the dynamic semantic feature includes:

[0028] Hidden Markov modules are used to perform implicit temporal feature recognition on the enhanced data link semantic features to obtain dynamic semantic features.

[0029] As an improvement of the above solution, the training scenario module includes a sports project library data module, an athlete data module, a training equipment and system data module and an intelligent sports terminal data module;

[0030] The types of training data in the athlete data module, the training equipment and system data module and the smart sports terminal data module include basic attribute data, training videos, physical data and training records;

[0031] The types of training data of the sports project library data module include basic attribute data, training videos, physical data and training skills.

[0032] To achieve the above-mentioned purpose, an embodiment of the present invention further provides a sports training model construction device, comprising:

[0033] A data acquisition module, used to acquire a plurality of pre-divided training scene modules; wherein the training scene modules include a plurality of types of training data;

[0034] A data link network building module is used to vertically link the same type of training data between different training scenario modules to obtain a data link network;

[0035] An implicit domain semantic feature acquisition module is used to perform entity recognition in each of the training scene modules in the data link network and quantify the association relationship between the different entities recognized to obtain the implicit domain semantic features of each of the training scene modules;

[0036] A dimension reduction semantic feature acquisition module is used to perform dimension reduction processing on the data link network to obtain dimension reduction semantic features;

[0037] A deep-level semantic association relationship feature acquisition module is used to call the graph attention network module to aggregate the semantic elements of the training data within the data link network to obtain deep-level semantic association relationship features;

[0038] An enhanced data link semantic feature acquisition module is used to fuse the data link network, the implicit domain semantic features, the dimension reduction semantic features and the deep semantic association relationship features to obtain enhanced data link semantic features;

[0039] An implicit temporal feature recognition module is used to perform implicit temporal feature recognition on the enhanced data link semantic feature to obtain a dynamic semantic feature;

[0040] The model training module is used to train the adversarial generative network module using the dynamic semantic features and the training data to obtain a sports training model for guiding the sports training process.

[0041] To achieve the above-mentioned objectives, an embodiment of the present invention also provides a sports training model construction device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, and when the processor executes the computer program, it implements the sports training model construction method as described in any of the above embodiments.

[0042] To achieve the above objectives, an embodiment of the present invention further provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the sports training model construction method as described in any of the above embodiments.

[0043] To achieve the above objectives, an embodiment of the present invention further provides a computer program product, including a computer program / instructions, which, when executed by a processor, implements the sports training model construction method as described in any of the above embodiments.

[0044] Compared with the prior art, the sports training model construction method, device, equipment, storage medium and product provided by the embodiments of the present invention obtain several pre-divided training scene modules; wherein the training scene module includes several types of training data; vertically link the same type of training data between different training scene modules to obtain a data link network; respectively perform entity recognition in each of the training scene modules in the data link network and quantify the association relationship between the different entities identified to obtain the implicit domain semantic features of each of the training scene modules; perform dimensionality reduction processing on the data link network to obtain dimensionality reduction semantic features; call the graph attention network module to aggregate the semantic elements of the training data in the data link network to obtain deep-level semantic association relationship features; fuse the data link network, the implicit domain semantic features, the dimensionality reduction semantic features and the deep-level semantic association relationship features to obtain enhanced data link semantic features; perform implicit temporal feature recognition on the enhanced data link semantic features to obtain dynamic semantic features; use the dynamic semantic features and the training data to train a generative adversarial network module to obtain a sports training model for guiding the sports training process. Compared with the prior art, the embodiments of the present invention can maximize the semantic features of sports training data and use them for training sports training models, thereby fully tapping the value of sports training data and providing users with more accurate sports training guidance. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of a method for constructing a sports training model provided by an embodiment of the present invention;

[0046] Figure 2 is a schematic diagram of a data link network provided by an embodiment of the present invention;

[0047] Figure 3 is a flow chart of a method for constructing a sports training model provided by another embodiment of the present invention;

[0048] Figure 4 It is a structural schematic diagram of a sports training model building device provided by one embodiment of the present invention;

[0049] Figure 5 It is a structural schematic diagram of a sports training model building device provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0050] 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.

[0051] See also Figure 1 , is a flow chart of a method for constructing a sports training model provided by an embodiment of the present invention, comprising steps S1 to S8:

[0052] S1. Acquire a plurality of pre-divided training scene modules; wherein the training scene modules include a plurality of types of training data;

[0053] S2, vertically linking the same type of training data between different training scenario modules to obtain a data link network;

[0054] S3, respectively performing entity recognition within each of the training scene modules in the data link network and quantifying the association relationship between the different entities recognized to obtain the implicit domain semantic features of each of the training scene modules;

[0055] S4, performing dimensionality reduction processing on the data link network to obtain dimensionality reduction semantic features;

[0056] S5, calling the graph attention network module to aggregate the semantic elements of the training data within the data link network to obtain deep semantic association relationship features;

[0057] S6, fusing the data link network, the implicit domain semantic features, the dimension reduction semantic features and the deep semantic association relationship features to obtain enhanced data link semantic features;

[0058] S7, performing implicit temporal feature recognition on the enhanced data link semantic feature to obtain a dynamic semantic feature;

[0059] S8. Use the dynamic semantic features and the training data to train a generative adversarial network module to obtain a sports training model for guiding the sports training process.

[0060] It is worth noting that the existing sports training models only analyze the training data of an athlete's training session and can only process a single type of data. For example, they can only analyze physiological indicators. This can easily lead to the final training quality assessment results being inaccurate and the training suggestions provided being not suitable enough.

[0061] In order to fully explore the value of sports training data, improve the performance of sports training models and provide users with more accurate sports training guidance, an embodiment of the present invention pre-collects training data under different training scenarios, for example, training data under training scenarios such as physical fitness tests, daily training and daily activities, and does not limit the data type of the training data. For example, the training data can be image videos or text materials, etc., and then the above-mentioned multi-source heterogeneous training data are divided into different training scenario modules based on the training scenarios.

[0062] The embodiment of the present invention constructs a data link network by acquiring training scene modules from multiple sources, performs entity recognition and entity relationship quantification on each training scene module in the data link network, and obtains implicit domain semantic features of each training scene module; while maintaining the local structural relationship of the original data of the training scene module, the high-dimensional data of the data link network is mapped to a low-dimensional space, and rich semantic mutual information features between modules are obtained as dimensionality reduction semantic features; in order to make full use of the internal and external association features of each training scene module in the data link network, a graph attention network module is used to learn the attribute features of the data link network to obtain deep-level semantic association relationship features; enhanced data link semantic features are obtained by fusing the data link network, implicit domain semantic features, dimensionality reduction semantic features and deep-level semantic association relationship features, and implicit temporal features of the enhanced data link semantic features are identified to obtain dynamic semantic features, thereby strengthening the data link semantic information, giving full play to the role of the data link network that maximizes the mutual information of each training scene module, and fully exploring the value of the training data; finally, adversarial learning is performed using the dynamic semantic features and the training data of each training scene module to train a sports training model, which can provide users with accurate sports training guidance.

[0063] As an optional implementation, the training scenario module includes a sports project library data module, an athlete data module, a training equipment and system data module, and an intelligent sports terminal data module;

[0064] The types of training data in the athlete data module, the training equipment and system data module and the smart sports terminal data module include basic attribute data, training videos, physical data and training records;

[0065] The types of training data of the sports project library data module include basic attribute data, training videos, physical data and training skills.

[0066] See also Figure 2 , is a schematic diagram of a data link network provided by an embodiment of the present invention, Figure 2 In the illustrated embodiment, the training scenario module includes a sports project library data module, an athlete data module, a training equipment and system data module, and an intelligent sports terminal data module, wherein body base, equipment data, terminal data, and project data belong to "basic attribute data", self-test videos, video data, and action videos belong to "training videos", and training records and training data belong to "training records".

[0067] It is worth noting that, in this embodiment, the "equipment data" in the "training equipment and system data module" can be the type, size and model of the equipment, while the "self-test video", "physical data" and "training record" can be the training data about the training process recorded by the training equipment, for example, the "training record" can be the training duration and number of training sessions; the "terminal data" in the "intelligent sports terminal data module" can be the type and version of the intelligent sports terminal, and the "video data", "physical data" and "training data" can be the training data about the training process recorded by the intelligent sports terminal; the training data in the "athlete data module" can be other training data about the training process other than the above two training scenario modules; the "sports project library data module" can be the official training data related to each sports project, for example, the "project data" can be the basic attribute data of the basketball project, the "action video" can be the action demonstration video of the basketball project, the "physical data" can be the training indicators recommended for different genders and ages, and the "training skills" can be the summary of the skills of different training actions.

[0068] As an optional implementation, in step S2, vertically linking the same type of training data between different training scenario modules to obtain a data link network includes:

[0069] S21, extracting semantic elements of each of the training data, and converting the semantic elements into semantic vectors;

[0070] S22, summarizing the semantic vectors in the training scenario modules to obtain a horizontal data subchain;

[0071] S23, constructing an interaction matrix based on the degree of interaction of the semantic elements between different training scenario modules;

[0072] S24, multiplying the semantic vector and the interaction matrix to obtain a vertical data subchain;

[0073] S25. Construct a data chain network based on the horizontal data sub-chain and the vertical data sub-chain.

[0074] Specifically, in step S21, the semantic elements of each training data in each training scenario module (such as athlete's physical fitness, athletic ability, venue parameters and training equipment attributes, etc.) are determined, and each semantic element is represented as a semantic vector v using a pre-trained semantic model. i , where i is the index of the semantic element. Then, in step S22, the semantic vectors of each training scene module are summarized to obtain the horizontal data sub-chain V of each training scene module. X = {v1, v2, ... v n}, where V Xrepresents the horizontal data subchain, and n represents the total number of semantic elements.

[0075] Furthermore, in step S23, an interaction matrix M′ is constructed to represent the interaction intensity and mode of training data between different training scenario modules. For example, for the i-th semantic element and the j-th semantic element, the element m in the interaction matrix is ij Indicates the degree of interaction between them. Furthermore, the correlation strength of different training data can be obtained by analyzing the dynamic data of different training scenario modules, and an interaction matrix can be constructed.

[0076] Furthermore, in step S24, the semantic vector and the interaction matrix are multiplied to obtain a comprehensive representation of each semantic element in the data chain, as shown in formula (1):

[0077]

[0078] Among them, l i represents the comprehensive representation of the i-th semantic element in the data chain, including its own semantic information and the interaction information with other semantic elements, m ij represents the element in the i-th row and j-th column of the interaction matrix, v j Represents the semantic vector of the jth semantic element.

[0079] Further, in step S25, the longitudinal data sub-chain V Y and horizontal data subchain V X Connect them together to get the data link network V:

[0080] V={V X , V Y} (2)

[0081] See also Figure 2 In this embodiment, the longitudinal data sub-chains constituting the data chain network include a video training data sub-chain, a body dynamics data sub-chain and a training dynamics data sub-chain.

[0082] Compared with the prior art, the embodiment of the present invention vertically links the same type of training data in different training scenario modules to obtain multiple vertical data sub-chains and construct a data chain network, which can enhance the semantic association of training data between training scenario modules.

[0083] As one optional implementation, in step S3, entity recognition is performed in each training scenario module in the data link network and the association relationship between different entities identified is quantified to obtain implicit domain semantic features of each training scenario module, including:

[0084] S31, performing entity recognition on the training data within the training scenario module and quantifying the relationship between entities;

[0085] S32. Construct a knowledge network based on the identified entities and relationships to obtain implicit domain semantic features of the training scenario module.

[0086] It is worth mentioning that there are two problems that need to be solved within the training scenario module. First, the training data within the module is discrete and lacks sufficient contextual information, which makes it impossible to accurately understand the constructed target preferences; second, there are semantic differences between different information. For example, in the athlete data module, both physical data and training records reflect training characteristics, but these two types of information naturally have semantic differences. Therefore, in order to increase the contextual semantic relationship of each discrete entity and attribute in the training data, the embodiment of the present invention constructs a knowledge network to obtain implicit domain semantic features and splices them onto the data link network in step S6.

[0087] Specifically, in step S31, the big model is used to identify entities in the training data, such as training movements, training equipment, technical terms, athlete roles, training methods and physiological indicators, etc., to improve the semantic understanding of the training data; further, the associations between entities are quantified, such as the relationship between movements and muscle activation, the impact of training methods on specific indicators, the effect of technology on skill improvement, etc. Further, in step S32, the identified entities and relationships are input into the sports big model to construct a knowledge network, so as to intuitively represent the associations between concepts, methods and skills.

[0088] Specifically, step S31 and step S32 can be expressed as shown in formulas (3) to (6):

[0089] Entity=match(b,Dictionary)(3)

[0090] Relation(e′1,e′2)=match(b,Pattern(e′1,e′2))(4)

[0091]

[0092] Attribute(e′,a)=extract(b,e′,a)(6)

[0093] In formulas (3) to (6), Entity means identifying entities in the input text; match means the match function; b means the input text; Dictionary means the dictionary containing predefined entities; Relation means the relationship between quantitative entities; e′1, e′2, e′ i , e′ jand e′ are extracted entities; Pattern(e′1,e′2) represents the pattern of the relationship between e′1 and e′2; A ij Represents the adjacency matrix; SLLM (sLLM) represents the small large language model; Attribute represents the attribute of the extracted entity; extract represents the attribute extraction function; r represents the relationship; a represents the attribute of the entity.

[0094] Among them, formula (5) indicates whether there is an association relationship between entities. If there is an association relationship, a value of 1 is assigned, and if there is no association relationship, a value of 0 is assigned. By constructing the adjacency matrix in formula (5), it can be ensured that the influence of irrelevant factors is eliminated in the subsequent calculation of the implicit domain semantic features.

[0095] Compared with the prior art, the embodiments of the present invention can enhance the contextual semantic relationship between entities and attributes of each discrete data within the training scenario module by calculating implicit domain semantic features.

[0096] Furthermore, the embodiment of the present invention constructs a data link network in step S2 to enhance the semantic connection between training scenario modules, and calculates implicit domain semantic features in step S3 to enhance the contextual relationship between each training data within the training scenario module, thereby being able to more completely represent the data features and relationships of each training data generated during the training process, that is, to maximize the internal semantic features of each training scenario module for training the sports training model.

[0097] As an optional implementation manner, in step S4, the data link network is subjected to dimensionality reduction processing to obtain dimensionality reduction semantic features, including:

[0098] S41, calculating a first similarity between the first training data and other training data in the data link network; wherein the first training data is any one of all the training data;

[0099] S42, based on the dimensional mapping relationship, mapping the training data in the high-dimensional space to the low-dimensional space, obtaining the dimension reduction semantic features of the training data, and calculating the second similarity between the first training data and other training data in the low-dimensional space;

[0100] S43, substituting the first similarity and the second similarity into a pre-constructed loss function to calculate a loss value;

[0101] S44. Use the gradient descent method to adjust the dimensional mapping relationship according to the loss value, and return to the step of calculating the first similarity until the loss value meets the set adjustment end condition or the number of adjustments to the dimensional mapping relationship reaches a set number, and output the reduced dimensionality semantic feature.

[0102] It is worth noting that in order to prevent the "crowding phenomenon" of redundant data outside the data link, in some embodiments, after maximizing the internal semantic features of each training scenario module, the t-SNE (-distributed Stochastic Neighbor Embedding) algorithm is used to map the high-dimensional data in the data link network to the low-dimensional space. By using the t-SNE algorithm for dimensionality reduction processing, it can be ensured that the training data after dimensionality reduction still retains the original data local structural relationship within its training scenario module, that is, similar training data are still close in the low-dimensional space.

[0103] Specifically, in step S41, the first similarity is calculated by formula (7):

[0104]

[0105] Among them, P j∣i represents the first similarity; x i 、x j and x k Represents training data in high-dimensional space; ∥x i -x j ∥ represents x i and x j The Euclidean distance of i Represents the scope of the Gaussian kernel function, which is determined by the training data x i The surrounding density is determined dynamically; e represents a natural constant.

[0106] Furthermore, in step S42, the second similarity is calculated by formula (8):

[0107]

[0108] Among them, Q j∣i Represents the second similarity; y i ,y j and k Represents training data in low-dimensional space; ∥y i -y j ∥ means y i and i The Euclidean distance of .

[0109] Furthermore, in step S43, the loss function is as shown in formula (9):

[0110]

[0111] Among them, C represents the loss function; P j∣i Indicates the first similarity; Q j∣irepresents the second similarity; log represents the logarithmic function.

[0112] Compared with the prior art, the embodiment of the present invention obtains reduced-dimensional semantic features by performing dimensionality reduction processing on the data link network, so as to perform implicit timing feature recognition after subsequent splicing with the data link network. This can prevent the "crowding phenomenon" of redundant data outside the data link and obtain rich semantic mutual information features between training scenario modules.

[0113] Further, in step S5, the graph attention network module is called to aggregate the semantic elements of the training data within the data link network to obtain deep semantic association relationship features. Specifically, the graph attention network module, also known as the graph attention network, is a special type of graph neural network (GNN), which introduces a learnable attention mechanism. By allocating weights between source nodes and target nodes, the node can decide which neighbor node is more important when aggregating messages from local neighbors. Finally, through multi-layer aggregation and updating, the deep semantic association relationship features of the data link network are learned, as shown in formulas (10) to (11):

[0114]

[0115] In formulas (10) to (11), represents the semantic element feature of the (l′+1)th layer. Specifically, the features of all neighboring nodes of the l′th layer are weighted summed based on attention to update the features and obtain the node (semantic element) features of the (l′+1)th layer. ; σ′ represents the Sigmoid activation function; W (1′) represents the weight matrix; l′ represents the current layer; l′+1 represents the next layer; ρ(i) represents z i The set of adjacent semantic attributes of ; represents the attention coefficient; LeakyReLU represents the LeakyReLU activation function; a T represents a learnable attention parameter; || represents a concatenation operation.

[0116] Furthermore, in step S6, the data link network, implicit domain semantic features, dimension reduction semantic features and deep semantic association relationship features are integrated to obtain enhanced data link semantic features, so as to perform implicit temporal feature recognition on the enhanced data link semantic features in step S7.

[0117] As an optional implementation manner, in step S7, the implicit temporal feature recognition is performed on the enhanced data link semantic feature to obtain a dynamic semantic feature, including:

[0118] Hidden Markov modules are used to perform implicit temporal feature recognition on the enhanced data link semantic features to obtain dynamic semantic features.

[0119] Specifically, let M represent the training scene module; J represent the dimension reduction semantic feature; H represent the change of the training scene module over time; V represent the data link network; I(H; L) represent the feature of the module data link semantic mutual information, as shown in formula (12):

[0120]

[0121] Among them, P(h,l″) represents the joint probability distribution; P(h) represents the marginal probability distribution of H; P(l″) represents the marginal probability distribution of V; h represents the hidden state; l″ represents the state.

[0122] Further, calculate the forward probability α t (i) and the backward probability β t (i) Then, based on the forward probability and the backward probability, the optimal hidden state sequence H is calculated using the Viterbi algorithm.

[0123] Specifically, the forward probability α t (i) is the state l″ observed at scene t t And the hidden state is h i The probability of is as shown in formula (13):

[0124] α t (i)=P(l″1,…,l″ t ,h t =h i ∣λ) (13)

[0125] Among them, λ represents the model parameter.

[0126] Further, we can recursively obtain α1(i), ..., α t+1 (j), the recursive formula is shown in equations (14) to (15):

[0127] α1(i)=π i b i (l″1) (14)

[0128]

[0129] Among them, b i (l″1) represents the observation probability; π i represents the initial probability state; N represents the total number of hidden states in the training scenario module. It can be understood that formula (14) accumulates the probabilities of all possible hidden states to obtain the joint probability of all possible hidden state sequences under a given data link network.

[0130] Furthermore, the joint probability P(V) of the observation sequence is shown in formula (16):

[0131]

[0132] Furthermore, the backward probability β t (i) is the state l″ observed at scene t t And the hidden state is h i , and in the subsequent scenes t+1,…,T, L is observed t+1:T The probability of is as shown in equations (17) to (18):

[0133] β T (i)=1 (17)

[0134]

[0135] Furthermore, the optimal hidden state sequence H is calculated using the Viterbi algorithm, as shown in equations (19) to (25):

[0136] P * =max H P(V,H) (19)

[0137] H * = argmax H P(V,H) (20)

[0138] δ t (i) = max 1≤j＜t {α j (i)β t (i)} (21)

[0139]

[0140] Among them, P * Denotes the optimal hidden state sequence H for a given data link network V * P(V,H) represents the joint probability distribution of the data link network V observed under a specific hidden state sequence H; H * represents the optimal hidden state sequence that maximizes the scene semantic mutual information, i.e., the dynamic semantic feature; argmax represents the maximum independent variable point set; ψ t (i) indicates the state l″ observed at scene t j The node passed by the previous scene of the path with the maximum probability; represents the optimal hidden state sequence at scene t; δ t (i) indicates the state l″ observed at scene t i The maximum probability value of all nodes.

[0141] Further, in step S8, the dynamic semantic features and training data are used to train the generative adversarial network module to obtain a sports training model for guiding the sports training process.

[0142] Specifically, the generative adversarial network module, also known as the generative adversarial network (GANs), consists of a generator D and a discriminator G, where the generator D attempts to generate training data close to the real data distribution, while the discriminator G attempts to distinguish between real data and virtual data generated by the generator, and the goal of the generator D is to maximize the error of the discriminator G, as shown in equations (25) to (27):

[0143]

[0144] Among them, X represents real data; Z represents virtual data; p data (x) represents the distribution of real data; p z (z) represents the distribution of virtual data; θ D represents the parameters of the generator; θ G Represents the parameters of the discriminator.

[0145] Furthermore, the sports training model trained in step S8 is SAH-GANs ("tSNE-GraphAttention-HMM-GANs", t-distributed Stochastic Neighbor Embedding-GraphAttention-Hidden Markov Model-Generative Adversarial Networks, t-distributed stochastic neighbor embedding-graph attention-hidden Markov-generative adversarial network), including a graph attention network module, a hidden Markov module and a generative adversarial network module.

[0146] In order to more clearly understand the method of building a sports training model, a specific example is used below to illustrate. Figure 3 , is a flow chart of a method for constructing a sports training model provided by another embodiment of the present invention, comprising Figure 3It can be seen that after the embodiment of the present invention constructs a data link network based on the training scenario module, it also performs entity recognition inside the training scenario module and quantifies the association relationship of the entities, thereby obtaining implicit domain semantic features; and also performs dimensionality reduction processing on the data link network to obtain dimensionality reduction semantic features; and calls the graph attention network module to aggregate the semantic elements of the training data inside the data link network to obtain deep-level semantic association relationship features; then the above-mentioned implicit domain semantic features, dimensionality reduction semantic features and deep-level semantic association relationship features are spliced ​​together with the data link network and then input into the hidden Markov module, that is, the spliced ​​enhanced data link semantic features are input into the hidden Markov module to obtain dynamic semantic features; finally, the dynamic semantic features and training data are input into the generative adversarial network module to train them, and finally a sports training model is obtained.

[0147] Compared with the prior art, the sports training model construction method provided by the embodiment of the present invention obtains several pre-divided training scene modules; wherein the training scene module includes several types of training data; vertically links the same type of training data between different training scene modules to obtain a data link network; performs entity recognition in each of the training scene modules in the data link network and quantifies the association relationship between the different entities identified to obtain the implicit domain semantic features of each of the training scene modules; performs dimensionality reduction processing on the data link network to obtain dimensionality reduction semantic features; calls a graph attention network module to aggregate the semantic elements of the training data in the data link network to obtain deep-level semantic association relationship features; integrates the data link network, the implicit domain semantic features, the dimensionality reduction semantic features and the deep-level semantic association relationship features to obtain enhanced data link semantic features; performs implicit temporal feature recognition on the enhanced data link semantic features to obtain dynamic semantic features; uses the dynamic semantic features and the training data to train a generative adversarial network module to obtain a sports training model for guiding the sports training process. Compared with the prior art, the embodiments of the present invention can maximize the semantic features of sports training data and use them for training sports training models, thereby fully tapping the value of sports training data to provide users with more accurate sports training guidance.

[0148] See also Figure 4 The embodiment of the present invention further provides a sports training model construction device 10, comprising:

[0149] The data acquisition module 11 is used to acquire a plurality of pre-divided training scene modules; wherein the training scene modules include a plurality of types of training data;

[0150] A data link network construction module 12, used to vertically link the same type of training data between different training scenario modules to obtain a data link network;

[0151] The implicit domain semantic feature acquisition module 13 is used to perform entity recognition in each of the training scene modules in the data link network and quantify the association relationship between the different entities recognized to obtain the implicit domain semantic features of each of the training scene modules;

[0152] A dimension reduction semantic feature acquisition module 14 is used to perform dimension reduction processing on the data link network to obtain dimension reduction semantic features;

[0153] A deep-level semantic association relationship feature acquisition module 15 is used to call the graph attention network module to aggregate the semantic elements of the training data inside the data link network to obtain deep-level semantic association relationship features;

[0154] An enhanced data link semantic feature acquisition module 16 is used to fuse the data link network, the implicit domain semantic features, the dimension reduction semantic features and the deep semantic association relationship features to obtain enhanced data link semantic features;

[0155] An implicit temporal feature recognition module 17 is used to perform implicit temporal feature recognition on the enhanced data link semantic feature to obtain a dynamic semantic feature;

[0156] The model training module 18 is used to train a generative adversarial network module using the dynamic semantic features and the training data to obtain a sports training model for guiding the sports training process.

[0157] In one implementation, the data link network construction module 12 is specifically used to:

[0158] Extracting semantic elements of each of the training data, and converting the semantic elements into semantic vectors;

[0159] Summarizing the semantic vectors in the training scenario modules to obtain a horizontal data subchain;

[0160] Based on the degree of interaction between the semantic elements in different training scenario modules, an interaction matrix is ​​constructed;

[0161] Multiplying the semantic vector and the interaction matrix to obtain a vertical data subchain;

[0162] Based on the horizontal data sub-chain and the vertical data sub-chain, a data chain network is constructed.

[0163] In one implementation, the implicit domain semantic feature acquisition module 13 is specifically used to:

[0164] Performing entity recognition on the training data within the training scenario module and quantifying the relationship between entities;

[0165] A knowledge network is constructed based on the identified entities and relationships to obtain implicit domain semantic features of the training scenario module.

[0166] In one implementation, the dimension reduction semantic feature acquisition module 14 is specifically used to:

[0167] Calculating a first similarity between first training data and other training data in the data link network; wherein the first training data is any one of all the training data;

[0168] Based on the dimensional mapping relationship, the training data in the high-dimensional space is mapped to the low-dimensional space to obtain the dimensionality reduction semantic features of the training data, and the second similarity between the first training data and other training data in the low-dimensional space is calculated;

[0169] Substituting the first similarity and the second similarity into a pre-constructed loss function to calculate a loss value;

[0170] The gradient descent method is used to adjust the dimensional mapping relationship according to the loss value, and return to the step of calculating the first similarity until the loss value meets the set adjustment end condition or the number of adjustments to the dimensional mapping relationship reaches a set number, and the reduced dimensionality semantic feature is output.

[0171] In one embodiment, the enhanced data link semantic feature acquisition module 16 is specifically used to:

[0172] Hidden Markov modules are used to perform implicit temporal feature recognition on the enhanced data link semantic features to obtain dynamic semantic features.

[0173] In one embodiment, the training scenario module includes a sports project library data module, an athlete data module, a training equipment and system data module, and an intelligent sports terminal data module;

[0174] The types of training data in the athlete data module, the training equipment and system data module and the smart sports terminal data module include basic attribute data, training videos, physical data and training records;

[0175] The types of training data of the sports project library data module include basic attribute data, training videos, physical data and training skills.

[0176] The sports training model construction device 10 provided in the embodiment of the present invention can implement all the process steps of the sports training model construction method described in the above embodiment. The functions of each module and unit in the device and the technical effects achieved are respectively the same as the functions and technical effects achieved by the sports training model construction method described in the above embodiment. The specific implementation method will not be repeated here.

[0177] See also Figure 5 The embodiment of the present invention further provides a sports training model construction device 20, comprising a processor 21, a memory 22, and a computer program stored in the memory 22 and configured to be executed by the processor 21. When the processor 21 executes the computer program, the steps in the above-mentioned sports training model construction method embodiment are implemented, for example Figure 1 Steps S1 to S8 described in the above; or, when the processor 21 executes the computer program, the functions of the modules in the above-mentioned device embodiments are realized.

[0178] The sports training model construction device can be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The sports training model construction device can include, but is not limited to, a processor 21 and a memory 22. Those skilled in the art can understand that the sports training model construction device can also include input and output devices, network access devices, buses, etc.

[0179] The processor 21 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor 21 is the control center of the sports training model construction device, and uses various interfaces and lines to connect various parts of the entire sports training model construction device.

[0180] The memory 22 can be used to store the computer program and / or module, and the processor 21 realizes various functions of the sports training model building device by running or executing the computer program and / or module stored in the memory 22, and calling the data stored in the memory 22. The memory 22 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function, etc.; the data storage area can store data created according to the use of the mobile phone, etc. In addition, the memory 22 can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0181] Wherein, if the module integrated in the sports training model construction device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor 21. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0182] An embodiment of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the sports training model construction method as described in any of the above embodiments.

[0183] Compared with the prior art, the sports training model construction device, equipment, storage medium and product provided by the embodiments of the present invention obtain several pre-divided training scene modules; wherein the training scene module includes several types of training data; vertically link the same type of training data between different training scene modules to obtain a data link network; respectively perform entity recognition in each of the training scene modules in the data link network and quantify the association relationship between the different entities identified to obtain the implicit domain semantic features of each of the training scene modules; perform dimensionality reduction processing on the data link network to obtain dimensionality reduction semantic features; call the graph attention network module to aggregate the semantic elements of the training data in the data link network to obtain deep-level semantic association relationship features; fuse the data link network, the implicit domain semantic features, the dimensionality reduction semantic features and the deep-level semantic association relationship features to obtain enhanced data link semantic features; perform implicit temporal feature recognition on the enhanced data link semantic features to obtain dynamic semantic features; use the dynamic semantic features and the training data to train a generative adversarial network module to obtain a sports training model for guiding the sports training process. Compared with the prior art, the embodiments of the present invention can maximize the semantic features of sports training data and use them for training sports training models, thereby fully tapping the value of sports training data to provide users with more accurate sports training guidance.

[0184] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for constructing a sports training model, characterized in that: include: Acquire a plurality of pre-divided training scene modules; wherein the training scene modules include a plurality of types of training data; Vertically link the same type of training data between different training scenario modules to obtain a data link network; Respectively perform entity recognition in each of the training scene modules in the data link network and quantify the association relationship between different entities recognized to obtain implicit domain semantic features of each of the training scene modules; Performing dimensionality reduction processing on the data link network to obtain dimensionality reduction semantic features; The graph attention network module is called to aggregate the semantic elements of the training data within the data link network to obtain deep semantic association relationship features; The data link network, the implicit domain semantic features, the dimension reduction semantic features and the deep semantic association relationship features are integrated to obtain enhanced data link semantic features; Performing implicit temporal feature recognition on the enhanced data link semantic feature to obtain a dynamic semantic feature; The dynamic semantic features and the training data are used to train a generative adversarial network module to obtain a sports training model for guiding the sports training process.

2. The method for constructing a sports training model according to claim 1, characterized in that: The vertical linking of the same type of training data between different training scenario modules to obtain a data link network includes: Extracting semantic elements of each of the training data, and converting the semantic elements into semantic vectors; Summarizing the semantic vectors in the training scenario modules to obtain a horizontal data subchain; Based on the degree of interaction between the semantic elements in different training scenario modules, an interaction matrix is ​​constructed; Multiplying the semantic vector and the interaction matrix to obtain a vertical data subchain; Based on the horizontal data sub-chain and the vertical data sub-chain, a data chain network is constructed.

3. The method for constructing a sports training model according to claim 1, wherein: The method of performing entity recognition in each training scene module in the data link network and quantifying the association relationship between different entities recognized to obtain the implicit domain semantic features of each training scene module includes: Performing entity recognition on the training data within the training scenario module and quantifying the relationship between entities; A knowledge network is constructed based on the identified entities and relationships to obtain implicit domain semantic features of the training scenario module.

4. The method for constructing a sports training model according to claim 1, wherein: The performing dimensionality reduction processing on the data link network to obtain dimensionality reduction semantic features includes: Calculating a first similarity between first training data and other training data in the data link network; wherein the first training data is any one of all the training data; Based on the dimensional mapping relationship, the training data in the high-dimensional space is mapped to the low-dimensional space to obtain the dimensionality reduction semantic features of the training data, and the second similarity between the first training data and other training data in the low-dimensional space is calculated; Substituting the first similarity and the second similarity into a pre-constructed loss function to calculate a loss value; The gradient descent method is used to adjust the dimensional mapping relationship according to the loss value, and return to the step of calculating the first similarity until the loss value meets the set adjustment end condition or the number of adjustments to the dimensional mapping relationship reaches a set number, and the reduced dimensionality semantic feature is output.

5. The method for constructing a sports training model according to claim 1, wherein: The step of performing implicit temporal feature recognition on the enhanced data link semantic feature to obtain a dynamic semantic feature includes: Hidden Markov modules are used to perform implicit temporal feature recognition on the enhanced data link semantic features to obtain dynamic semantic features.

6. The method for constructing a sports training model according to claim 1, wherein: The training scenario module includes a sports project library data module, an athlete data module, a training equipment and system data module, and an intelligent sports terminal data module; The types of training data in the athlete data module, the training equipment and system data module and the smart sports terminal data module include basic attribute data, training videos, physical data and training records; The types of training data of the sports project library data module include basic attribute data, training videos, physical data and training skills.

7. A sports training model construction device, characterized in that: include: A data acquisition module, used to acquire a plurality of pre-divided training scene modules; wherein the training scene modules include a plurality of types of training data; A data link network building module is used to vertically link the same type of training data between different training scenario modules to obtain a data link network; An implicit domain semantic feature acquisition module is used to perform entity recognition in each of the training scene modules in the data link network and quantify the association relationship between the different entities recognized to obtain the implicit domain semantic features of each of the training scene modules; A dimension reduction semantic feature acquisition module is used to perform dimension reduction processing on the data link network to obtain dimension reduction semantic features; A deep-level semantic association relationship feature acquisition module is used to call the graph attention network module to aggregate the semantic elements of the training data within the data link network to obtain deep-level semantic association relationship features; An enhanced data link semantic feature acquisition module is used to fuse the data link network, the implicit domain semantic features, the dimension reduction semantic features and the deep semantic association relationship features to obtain enhanced data link semantic features; An implicit temporal feature recognition module is used to perform implicit temporal feature recognition on the enhanced data link semantic feature to obtain a dynamic semantic feature; The model training module is used to train a generative adversarial network module using the dynamic semantic features and the training data to obtain a sports training model for guiding the sports training process.

8. A sports training model construction device, characterized in that: The method comprises a processor, a memory and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the method for constructing a sports training model as described in any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the sports training model construction method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the sports training model construction method as described in any one of claims 1 to 6 is implemented.

Images