Pick ball sport injury prediction method based on deep learning

Abstract

The invention discloses a deep learning-based Pick ball sport injury prediction method, which belongs to the technical field of sport injury prediction, and comprises the following steps of: constructing deep training sample data by collecting historical sport data of a Pick ball athlete and injury types corresponding to the historical sport data; and then the deep learning model is used for learning the deep training sample data, so that the sport injury of the Pick ball athlete can be predicted by using the deep learning technology, the potential injury risk is predicted, the incidence rate of the sport injury is reduced, and the health of the athlete is guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of sports injury prediction, and specifically relates to a pickleball sports injury prediction method based on deep learning. Background Art

[0002] Pickleball is an emerging racket sport that combines elements of tennis, badminton and table tennis, so it is figuratively called a hybrid of these three sports. The rules of pickleball are relatively simple. When serving, the ball needs to be thrown in, and the contact surface between the ball and the racket is below the waist. The charm of this sport lies in its moderate activity and exercise. It is suitable for table tennis and badminton enthusiasts who want to find more intense sports, and for those who can't play badminton or tennis as a daily exercise. Pickleball is gradually becoming popular in my country, and athletes are prone to sports injuries during training and competition. At present, the prediction method for pickleball sports injuries mainly relies on the experience and judgment of coaches, which lacks scientificity and accuracy. In order to reduce the risk of athletes' injuries and improve the effect of sports training, it is necessary to study an effective method for predicting pickleball sports injuries. Summary of the invention

[0003] The present invention provides a method for predicting sports injuries of pickleball players based on deep learning, so as to solve the problem that the prior art cannot accurately predict sports injuries of pickleball players through the experience judgment of coaches.

[0004] A deep learning-based pickleball sports injury prediction method, comprising: Collecting historical sports data of pickleball players and injury types corresponding to the historical sports data, and obtaining deep training sample data after preprocessing the historical sports data and the injury types corresponding to the historical sports data; Constructing a deep learning model, and based on the deep training sample data, using an adaptive multi-group collaborative optimization algorithm to train the deep learning model, to obtain a trained deep learning model; The corresponding motion data to be identified of the pickleball operator is collected in real time, and the trained deep learning model is used to analyze the motion data to determine the prediction results of pickleball sports injuries.

[0005] Furthermore, after determining the pickle ball sports injury prediction result, it also includes: when the pickle ball sports injury prediction result meets the preset injury warning condition, injury warning information is generated, and a warning is issued according to the injury warning information.

[0006] Furthermore, a deep learning model is constructed, including: constructing a convolutional neural network model to obtain a deep learning model.

[0007] Furthermore, based on the deep training sample data, an adaptive multi-group collaborative optimization algorithm is used to train the deep learning model to obtain a trained deep learning model, including: The hyperparameters to be trained of the deep learning model are randomly initialized, and the initialized hyperparameters to be trained are encoded into vectors to obtain a hyperparameter vector, and multiple different hyperparameter vectors are obtained at the same time; For any hyperparameter vector, after applying the hyperparameters contained in the hyperparameter vector to the deep learning model, the historical motion data in the deep training sample data is used as the input of the deep learning model, and the expected output of the deep learning model is constructed based on the damage type corresponding to the historical motion data, and the fitness value corresponding to the hyperparameter vector is obtained; Traverse all hyperparameter vectors, determine the fitness value corresponding to each hyperparameter vector, and determine the optimal hyperparameter vector based on the fitness value corresponding to each hyperparameter vector; Divide all hyperparameter vectors into the first group, the second group, and the third group according to the preset ratio between the first group, the second group, and the third group; For the first vector in the first population, a neighborhood search mechanism based on fitness value decision is used to search for a better region of the first vector, and the first vector after the search is obtained; For the second vector in the second population, according to the optimal hyperparameter vector, a sine-cosine search mechanism based on fitness value decision is used to search for the optimal direction of the second vector, and the second vector after the search is obtained; For the third vector in the third population, an adaptive multi-vector learning mechanism is used to quickly search the third vector to obtain the third vector after the quick search; For the first vector after the search, the second vector after the search, and the third vector after the quick search, obtain the fitness value corresponding to each hyperparameter vector, and determine the mutation probability of the hyperparameter vector according to the fitness value corresponding to the hyperparameter vector; For each hyperparameter vector, according to the mutation probability corresponding to the hyperparameter vector, the annealing simulation rebirth mechanism is used to perform mutation processing to obtain the hyperparameter vector after mutation processing; Determine whether the current number of training times is greater than or equal to the preset maximum number of training times. If so, obtain the final hyperparameters of the deep learning model according to the hyperparameter vector after mutation processing to obtain the deep learning model after training. Otherwise, return to the step of dividing the first population, the second population, and the third population. Furthermore, the historical motion data in the deep training sample data is used as the input of the deep learning model, the expected output of the deep learning model is constructed based on the damage type corresponding to the historical motion data, and the fitness value corresponding to the hyperparameter vector is obtained, including: The historical motion data in the deep training sample data is used as the input of the deep learning model, the expected output of the deep learning model is constructed based on the damage type corresponding to the historical motion data, and the root mean square error value corresponding to the hyperparameter vector is obtained; The root mean square error value corresponding to the hyperparameter vector is added to a positive number less than 0.001 and then the inverse is taken to obtain the fitness value corresponding to the hyperparameter vector.

[0008] Furthermore, for the first vector in the first population, a neighborhood search mechanism based on fitness value decision is used to search for a better region of the first vector, and the first vector after the search is obtained, including: For the first vectors in the first population, obtain the fitness value corresponding to each first vector, and arrange the first vectors in descending order of the fitness values ​​to obtain the arranged first vectors; According to the fitness value corresponding to each first vector and the first vector after arrangement, the degree of goodness of the position corresponding to the first vector is obtained as follows:

[0009] in, Indicates t During the training i The position quality of the first vector after the arrangement is i =1,2,…,M1, M1 represents the total number of the first vector, represents the absolute value function, represents the fitness value corresponding to the optimal parameter vector, represents the fitness value corresponding to the worst parameter vector, Indicates i +1 The positional superiority of the first vector after permutation, and i =1, the formula Replace with i The fitness value corresponding to the first vector after permutation ; Indicates i -1 The positional quality of the first vector after the permutation, and i= When M1, the formula Replace with i The fitness value corresponding to the first vector after permutation ; For any first vector, randomly match the first vector with another first vector to obtain a competition vector corresponding to the first vector; Obtain the Euclidean distance between the first vector and the corresponding competition vector, and determine whether the Euclidean distance is less than the preset neighborhood radius R. If so, perform a positional superiority judgment, otherwise perform a fitness value judgment; The step of determining the superiority of the position is as follows: determining whether the superiority of the position corresponding to the first vector is greater than the superiority of the position corresponding to the competition vector; if so, determining the neighborhood search factor , otherwise determine the neighborhood search factor: ; in, Indicates i The fitness value corresponding to the competition vector corresponding to the first vector after permutation, represents a preset positive term, and =0.0001; It represents the exponential function with the natural constant e as the base; The fitness value judgment step is as follows: judging whether the fitness value corresponding to the first vector is greater than the fitness value corresponding to the competition vector; if so, determining the neighborhood search factor , otherwise determine the neighborhood search factor: ; According to the neighborhood search factor, a better region search is performed on the first vector, and the first vector after the search is obtained as follows: ;in, Indicates t During the training process i The d-th dimension parameter of the first vector, d =1,2,…,D, where D represents the total dimension of hyperparameters, Indicates after the search , It is represented by a mean of 0 and a variance of Generates random numbers from a normal distribution.

[0010] Furthermore, for the second vector in the second population, the optimal direction of the second vector is searched according to the optimal hyperparameter vector and the sine-cosine search mechanism of fitness value decision, and the second vector after the search is obtained, including: For the second vector in the second population, a decision factor between (0, 1) is randomly generated for the second vector, and it is determined whether the decision factor is less than 0.5. If so, a sine search is performed on the second vector in the optimal direction according to the optimal hyperparameter vector to obtain the second vector after the search. Otherwise, a cosine search is performed on the second vector in the optimal direction according to the optimal hyperparameter vector to obtain the second vector after the search. According to the optimal hyperparameter vector, a sine search is performed on the second vector, and the second vector after the search is obtained as follows:

[0011] in, Indicates t During the training process m The d-th dimension hyperparameter of the second vector, m =1,2,…,M2, M2 represents the total number of the first vector, After search ; represents the adaptive step size control factor, and ; represents the preset step size factor, T represents the preset maximum number of training times, represents the optimal parameter vector d Dimension parameters, represents the first random number between (0,1), represents the second random number between (0,1), sin represents the sine function; According to the optimal hyperparameter vector, a cosine search is performed on the second vector, and the second vector after the search is obtained as:

[0012] Here, cos represents the cosine function.

[0013] Furthermore, for the third vector in the third population, an adaptive multi-vector learning mechanism is used to quickly search the third vector to obtain the third vector after the quick search, including: For a third vector in the third population, two different second vectors are randomly matched to the third vector to obtain a first learning vector and a second learning vector corresponding to the third vector; According to the first learning vector and the second learning vector, a third vector is quickly searched, and the third vector obtained after the quick search is:

[0014]

[0015]

[0016] in, Indicates t During the training process n The third vector d Dimension parameters, Indicates the number of seconds after a quick search. , n=1,2,…,M3, M3 represents the total number of the first vector, represents the first adaptive learning factor, represents the second adaptive learning factor, represents the third random number between (0,1), represents the fourth random number between (0,1), represents the first learning vector d Dimension parameters, represents the second learning vector d Dimension parameters, Indicates n The fitness value of the third vector, represents the fitness value of the first learning vector, represents the fitness value of the second learning vector, represents an exponential function with the natural constant e as the base, represents a preset positive term, and =0.0001.

[0017] Further, the mutation probability of the hyperparameter vector is determined according to the fitness value corresponding to the hyperparameter vector, including:

[0018] in, represents the probability of mutation, represents a natural constant, represents the fifth random number between (0,1), represents pi, represents the sine function, T represents the maximum number of training times; For each hyperparameter vector, according to the mutation probability corresponding to the hyperparameter vector, the annealing simulation rebirth mechanism is used to perform mutation processing to obtain the hyperparameter vector after mutation processing, including: For each hyperparameter vector, a mutation decision factor between (0,1) is randomly generated, and it is determined whether the mutation decision factor is less than the mutation probability corresponding to the hyperparameter vector. If so, the hyperparameter vector is regenerated to obtain the regenerated hyperparameter vector, and the annealing simulation algorithm is used to selectively accept the regenerated hyperparameter vector to obtain the hyperparameter vector after mutation. Otherwise, the original hyperparameter vector is directly used as the hyperparameter vector after mutation.

[0019] Furthermore, the hyperparameter vector is regenerated, and the hyperparameter vector after regeneration is obtained as follows: The generated mutation factor is:

[0020] in, represents the factor of variation, It represents the rounding function, and directly takes 1 for numbers less than 1; represents the mutation quantity control parameter, Indicates k The fitness value of the hyperparameter vector to be mutated, represents the fitness value corresponding to the worst parameter vector, Indicates a preset positive number. k =1,2,…,L, L represents the total number of hyperparameter vectors; The generated variation range factor is:

[0021] Where F represents the variation range factor, represents the variation range control parameter, Represents the fitness value corresponding to the optimal parameter vector; According to the variation range factor, the variation parameter vector is generated as:

[0022] in, represents the mutation parameter vector d Dimension parameters, Indicates t During the training k A vector of hyperparameters to be mutated, express A random number between Generate multiple different mutation parameter vectors until the number of mutation parameter vectors reaches the mutation number factor, and determine the mutation parameter vector with the largest fitness value as the new parameter vector; According to the new parameter vector, the hyperparameter vector after regeneration is obtained as:

[0023] in, represents the d-th dimension parameter of the new parameter vector, represents the d-th dimension parameter of the worst parameter vector, Represents the d-th dimension parameter of the hyperparameter vector after the k-th rebirth.

[0024] The present invention provides a method for predicting pickle ball sports injuries based on deep learning. The method collects historical sports data of pickle ball players and the injury types corresponding to the historical sports data to construct deep training sample data, and then uses a deep learning model to learn the deep training sample data. In this way, deep learning technology can be used to predict the sports injuries of pickle ball players, predict potential injury risks, reduce the incidence of sports injuries, and protect the health of athletes. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0026] Figure 1 A flowchart of a method for predicting pickleball sports injuries based on deep learning provided in an embodiment of the present invention.

[0027] The above drawings have shown clear embodiments of the present invention, which will be described in more detail below. These drawings and text descriptions are not intended to limit the scope of the present invention in any way, but to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0028] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0029] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0030] like Figure 1 As shown, a method for predicting sports injuries in pickleball based on deep learning is provided in an embodiment of the present invention, comprising: S101, collecting historical sports data of pickleball players and injury types corresponding to the historical sports data, and obtaining deep training sample data after preprocessing the historical sports data and the injury types corresponding to the historical sports data; Historical sports data may include physiological data and sports pattern data of pickleball players. For example, physiological data of pickleball players may include weight, height, three-dimensional and various physical fitness test data, and sports pattern data may include the two most commonly used hitting methods in each pickleball game and the standard degree of each hitting method. These historical sports data and the types of injuries corresponding to the historical sports data are factual data that have already occurred, so they can serve as the basis for deep learning. However, it is worth noting that historical sports data may also include other data related to injuries.

[0031] After preprocessing the historical motion data and the injury types corresponding to the historical motion data, the method may include normalizing the data to obtain deep training sample data.

[0032] S102, constructing a deep learning model, and based on the deep training sample data, using an adaptive multi-group collaborative optimization algorithm to train the deep learning model, to obtain a trained deep learning model; In the prior art, a gradient descent optimization algorithm is often used to train a deep learning model, which results in inadequate training and makes it difficult to achieve accurate prediction. Therefore, an embodiment of the present invention provides an adaptive multi-group collaborative optimization algorithm, and uses the adaptive multi-group collaborative optimization algorithm to train a deep learning model, which can effectively improve the algorithm's ability to learn data, thereby improving the accuracy of sports injury prediction.

[0033] S103: Collect the to-be-recognized motion data corresponding to the pickleball operator in real time, and use the trained deep learning model to analyze the to-be-recognized motion data to determine the pickleball sports injury prediction result.

[0034] The prediction result of the injury of the pickle ball sports may include that the pickle ball player may be injured or the pickle ball player will not be injured for the time being. When the pickle ball player may be injured, it can be transmitted to the coach for analysis and advice, so as to avoid the injury of the pickle ball player.

[0035] The present invention provides a method for predicting pickle ball sports injuries based on deep learning. The method collects historical sports data of pickle ball players and the injury types corresponding to the historical sports data to construct deep training sample data, and then uses a deep learning model to learn the deep training sample data. In this way, deep learning technology can be used to predict the sports injuries of pickle ball players, predict potential injury risks, reduce the incidence of sports injuries, and protect the health of athletes.

[0036] In the embodiment of the present invention, after the pickle ball sports injury prediction result is determined, it also includes: when the pickle ball sports injury prediction result meets the preset injury warning condition, injury warning information is generated, and a warning is issued according to the injury warning information.

[0037] In an embodiment of the present invention, constructing a deep learning model includes: constructing a convolutional neural network model to obtain a deep learning model.

[0038] It is worth noting that the convolutional neural network is only a preferred implementation of the embodiment of the present invention, and other neural networks can also be used as deep learning models.

[0039] In an embodiment of the present invention, based on the deep training sample data, an adaptive multi-group collaborative optimization algorithm is used to train the deep learning model to obtain the trained deep learning model, including: The hyperparameters to be trained of the deep learning model are randomly initialized, and the initialized hyperparameters to be trained are encoded into vectors to obtain a hyperparameter vector, and multiple different hyperparameter vectors are obtained at the same time; For example, for the connection weights of a deep learning model, they can be randomly initialized between the upper and lower limits of the connection weights, and the initialized connection weights can be formed into a vector to obtain a hyperparameter vector.

[0040] For any hyperparameter vector, after applying the hyperparameters contained in the hyperparameter vector to the deep learning model, the historical motion data in the deep training sample data is used as the input of the deep learning model, and the expected output of the deep learning model is constructed based on the damage type corresponding to the historical motion data, and the fitness value corresponding to the hyperparameter vector is obtained; It is worth noting that for any pickleball sport, its most recent K historical sports data should be constructed into a K-row data matrix (where each row of data represents a historical sports data), and then the data matrix is ​​used as the input of the deep learning model.

[0041] Traverse all hyperparameter vectors, determine the fitness value corresponding to each hyperparameter vector, and determine the optimal hyperparameter vector based on the fitness value corresponding to each hyperparameter vector; Divide all hyperparameter vectors into the first group, the second group, and the third group according to the preset ratio between the first group, the second group, and the third group; For example, the total number of all hyperparameter vectors can be set to an integer multiple of 10, and after all hyperparameter vectors are arranged in descending order according to fitness values, they are divided into the first population, the second population, and the third population in a ratio of 2:6:2.

[0042] For the first vector in the first population, a neighborhood search mechanism based on fitness value decision is used to search for a better region of the first vector, and the first vector after the search is obtained; For the second vector in the second population, according to the optimal hyperparameter vector, a sine-cosine search mechanism based on fitness value decision is used to search for the optimal direction of the second vector, and the second vector after the search is obtained; For the third vector in the third population, an adaptive multi-vector learning mechanism is used to quickly search the third vector to obtain the third vector after the quick search; For the first vector after the search, the second vector after the search, and the third vector after the quick search, obtain the fitness value corresponding to each hyperparameter vector, and determine the mutation probability of the hyperparameter vector according to the fitness value corresponding to the hyperparameter vector; For each hyperparameter vector, according to the mutation probability corresponding to the hyperparameter vector, the annealing simulation rebirth mechanism is used to perform mutation processing to obtain the hyperparameter vector after mutation processing; Determine whether the current number of training times is greater than or equal to the preset maximum number of training times. If so, obtain the final hyperparameters of the deep learning model according to the hyperparameter vector after mutation processing to obtain the deep learning model after training. Otherwise, return to the step of dividing the first population, the second population, and the third population. The adaptive multi-group collaborative optimization algorithm provided by the embodiment of the present invention can not only effectively improve the global search capability of the algorithm, but also effectively improve the convergence speed of the algorithm, thereby achieving better data learning effects and improving sports injury prediction capabilities.

[0043] In an embodiment of the present invention, the historical motion data in the deep training sample data is used as the input of the deep learning model, the expected output of the deep learning model is constructed based on the damage type corresponding to the historical motion data, and the fitness value corresponding to the hyperparameter vector is obtained, including: The historical motion data in the deep training sample data is used as the input of the deep learning model, the expected output of the deep learning model is constructed based on the damage type corresponding to the historical motion data, and the root mean square error value corresponding to the hyperparameter vector is obtained; The root mean square error value corresponding to the hyperparameter vector is added to a positive number less than 0.001 and then the inverse is taken to obtain the fitness value corresponding to the hyperparameter vector.

[0044] In the embodiment of the present invention, for the first vector in the first population, a neighborhood search mechanism based on fitness value decision is used to search for a better region of the first vector to obtain the first vector after the search, including: For the first vectors in the first population, obtain the fitness value corresponding to each first vector, and arrange the first vectors in descending order of the fitness values ​​to obtain the arranged first vectors; According to the fitness value corresponding to each first vector and the first vector after arrangement, the degree of goodness of the position corresponding to the first vector is obtained as follows:

[0045] in, Indicates t During the training i The position quality of the first vector after the arrangement is i =1,2,…,M1, M1 represents the total number of the first vector, represents the absolute value function, represents the fitness value corresponding to the optimal parameter vector, represents the fitness value corresponding to the worst parameter vector, Indicates i +1 The positional superiority of the first vector after permutation, and i =1, the formula Replace with i The fitness value corresponding to the first vector after permutation ; Indicates i -1 The positional quality of the first vector after the permutation, and i= When M1, the formula Replace with i The fitness value corresponding to the first vector after permutation ; For any first vector, randomly match the first vector with another first vector to obtain a competition vector corresponding to the first vector; Obtain the Euclidean distance between the first vector and the corresponding competition vector, and determine whether the Euclidean distance is less than the preset neighborhood radius R. If so, perform a positional superiority judgment, otherwise perform a fitness value judgment; The step of determining the superiority of the position is as follows: determining whether the superiority of the position corresponding to the first vector is greater than the superiority of the position corresponding to the competition vector; if so, determining the neighborhood search factor , otherwise determine the neighborhood search factor: ; in, Indicates i The fitness value corresponding to the competition vector corresponding to the first vector after permutation, represents a preset positive term, and =0.0001; It represents the exponential function with the natural constant e as the base; The fitness value judgment step is as follows: judging whether the fitness value corresponding to the first vector is greater than the fitness value corresponding to the competition vector; if so, determining the neighborhood search factor , otherwise determine the neighborhood search factor: ; According to the neighborhood search factor, a better region search is performed on the first vector, and the first vector after the search is obtained as: ;in, Indicates t During the training i The d-th dimension parameter of the first vector, d =1,2,…,D, where D represents the total dimension of hyperparameters, Indicates after the search , It is represented by a mean of 0 and a variance of Generates random numbers from a normal distribution.

[0046] In the neighborhood search mechanism of fitness value decision provided by the embodiment of the present invention, if the competition vector is located in the neighborhood range of the first vector, it proves that the two positions are relatively close, so the position with a large fitness value change is more worthy of development, so the neighborhood search can be performed according to the degree of position quality. When the two are far apart, the position of the vector with a larger fitness value is more worthy of development, so the development of the adaptive optimal area can be achieved, and the development efficiency of the algorithm can be improved.

[0047] In the embodiment of the present invention, for the second vector in the second population, according to the optimal hyperparameter vector, a sine-cosine search mechanism of fitness value decision is used to search the second vector in the optimal direction to obtain the second vector after the search, including: For the second vector in the second population, a decision factor between (0, 1) is randomly generated for the second vector, and it is determined whether the decision factor is less than 0.5. If so, a sine search is performed on the second vector in the optimal direction according to the optimal hyperparameter vector to obtain the second vector after the search. Otherwise, a cosine search is performed on the second vector in the optimal direction according to the optimal hyperparameter vector to obtain the second vector after the search. According to the optimal hyperparameter vector, a sine search is performed on the second vector, and the second vector after the search is obtained as follows:

[0048] in, Indicates t During the training process m The d-th dimension hyperparameter of the second vector, m =1,2,…,M2, M2 represents the total number of the first vector, After search ; represents the adaptive step size control factor, and ; represents the preset step size factor, T represents the preset maximum number of training times, represents the optimal parameter vector d Dimension parameters, represents the first random number between (0,1), represents the second random number between (0,1), sin represents the sine function; According to the optimal hyperparameter vector, a cosine search is performed on the second vector, and the second vector after the search is obtained as:

[0049] Here, cos represents the cosine function.

[0050] The sine-cosine search mechanism for fitness value decision provided in the embodiment of the present invention can effectively enable all second vectors to search for the optimal position, which can not only improve the search speed of the algorithm, but also provide a chance to find a better position during the search process.

[0051] In the embodiment of the present invention, for the third vector in the third population, an adaptive multi-vector learning mechanism is used to quickly search the third vector to obtain the third vector after the quick search, including: For a third vector in the third population, two different second vectors are randomly matched to the third vector to obtain a first learning vector and a second learning vector corresponding to the third vector; According to the first learning vector and the second learning vector, a third vector is quickly searched, and the third vector obtained after the quick search is:

[0052]

[0053]

[0054] in, Indicates t During the training process n The third vector d Dimension parameters, Indicates the number of seconds after a quick search. , n =1,2,…,M3, M3 represents the total number of the first vector, represents the first adaptive learning factor, represents the second adaptive learning factor, represents the third random number between (0,1), represents the fourth random number between (0,1), represents the first learning vector d Dimension parameters, represents the second learning vector d Dimension parameters, Indicates n The fitness value of the third vector, represents the fitness value of the first learning vector, represents the fitness value of the second learning vector, represents an exponential function with the natural constant e as the base, represents a preset positive term, and =0.0001.

[0055] The adaptive multi-vector learning mechanism provided by the embodiment of the present invention can enable the third vector to search towards a more optimal position and in a non-optimal direction, which can effectively improve the global search capability of the algorithm and the fine search capability of the algorithm in the later stage.

[0056] In an embodiment of the present invention, determining the mutation probability of the hyperparameter vector according to the fitness value corresponding to the hyperparameter vector includes:

[0057] in, represents the probability of mutation, represents a natural constant, represents the fifth random number between (0,1), represents pi, represents the sine function, T represents the maximum number of training times; For each hyperparameter vector, according to the mutation probability corresponding to the hyperparameter vector, the annealing simulation rebirth mechanism is used to perform mutation processing to obtain the hyperparameter vector after mutation processing, including: For each hyperparameter vector, a mutation decision factor between (0,1) is randomly generated, and it is determined whether the mutation decision factor is less than the mutation probability corresponding to the hyperparameter vector. If so, the hyperparameter vector is regenerated to obtain the regenerated hyperparameter vector, and the annealing simulation algorithm is used to selectively accept the regenerated hyperparameter vector to obtain the hyperparameter vector after mutation. Otherwise, the original hyperparameter vector is directly used as the hyperparameter vector after mutation.

[0058] In the embodiment of the present invention, the hyperparameter vector is regenerated, and the hyperparameter vector after regeneration is obtained as follows: The generated mutation factor is:

[0059] in, represents the factor of variation, It represents the rounding function, and directly takes 1 for numbers less than 1; represents the mutation quantity control parameter, Indicates k The fitness value of the hyperparameter vector to be mutated, represents the fitness value corresponding to the worst parameter vector, Indicates a preset positive number. k =1,2,…,L, L represents the total number of hyperparameter vectors; The generated variation range factor is:

[0060] Where F represents the variation range factor, represents the variation range control parameter, Represents the fitness value corresponding to the optimal parameter vector; According to the variation range factor, the variation parameter vector is generated as:

[0061] in, represents the mutation parameter vector d Dimension parameters, Indicates t During the training k A vector of hyperparameters to be mutated, express A random number between Generate multiple different mutation parameter vectors until the number of mutation parameter vectors reaches the mutation number factor, and determine the mutation parameter vector with the largest fitness value as the new parameter vector; According to the new parameter vector, the hyperparameter vector after regeneration is obtained as:

[0062] in, represents the d-th dimension parameter of the new parameter vector, represents the d-th dimension parameter of the worst parameter vector, Represents the d-th dimension parameter of the hyperparameter vector after the k-th rebirth.

[0063] The annealing simulation regeneration mechanism provided in the embodiment of the present invention can effectively realize mutation and find a better position while avoiding the worst position. The annealing simulation algorithm controls the entire process. Even if the fitness value of the generated new parameter vector is worse, there is a certain probability of accepting it, which helps the algorithm to jump out of the local optimal solution and improve the global search capability.

[0064] Optionally, after each update of the hyperparameter vector, an out-of-bounds process may be performed on the hyperparameter vector to ensure the effectiveness of the algorithm.

[0065] Through the above mechanisms, the global search capability of the algorithm can be improved while ensuring the algorithm search speed, ultimately improving the accuracy of sports injury prediction.

[0066] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art that are not disclosed by the present invention. It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. A method for predicting injuries in pickleball based on deep learning, characterized in that: include: Collecting historical sports data of pickleball players and injury types corresponding to the historical sports data, and obtaining deep training sample data after preprocessing the historical sports data and the injury types corresponding to the historical sports data; Constructing a deep learning model, and based on the deep training sample data, using an adaptive multi-group collaborative optimization algorithm to train the deep learning model, to obtain a trained deep learning model; The corresponding motion data to be identified of the pickleball operator is collected in real time, and the trained deep learning model is used to analyze the motion data to determine the prediction results of pickleball sports injuries.

2. The method for predicting sports injuries in pickleball based on deep learning according to claim 1, characterized in that: After determining the pickle ball sports injury prediction result, it also includes: when the pickle ball sports injury prediction result meets the preset injury warning condition, injury warning information is generated, and a warning is issued according to the injury warning information.

3. The method for predicting sports injuries of pickleball based on deep learning according to claim 1, characterized in that: Constructing a deep learning model, including: constructing a convolutional neural network model to obtain a deep learning model.

4. The method for predicting sports injuries in pickleball based on deep learning according to claim 1, characterized in that: Based on the deep training sample data, the deep learning model is trained using an adaptive multi-group collaborative optimization algorithm to obtain a trained deep learning model, including: The hyperparameters to be trained of the deep learning model are randomly initialized, and the initialized hyperparameters to be trained are encoded into vectors to obtain a hyperparameter vector, and multiple different hyperparameter vectors are obtained at the same time; For any hyperparameter vector, after applying the hyperparameters contained in the hyperparameter vector to the deep learning model, the historical motion data in the deep training sample data is used as the input of the deep learning model, and the expected output of the deep learning model is constructed based on the damage type corresponding to the historical motion data, and the fitness value corresponding to the hyperparameter vector is obtained; Traverse all hyperparameter vectors, determine the fitness value corresponding to each hyperparameter vector, and determine the optimal hyperparameter vector based on the fitness value corresponding to each hyperparameter vector; Divide all hyperparameter vectors into the first group, the second group, and the third group according to the preset ratio between the first group, the second group, and the third group; For the first vector in the first population, a neighborhood search mechanism based on fitness value decision is used to search for a better region of the first vector, and the first vector after the search is obtained; For the second vector in the second population, according to the optimal hyperparameter vector, a sine-cosine search mechanism based on fitness value decision is used to search for the optimal direction of the second vector, and the second vector after the search is obtained; For the third vector in the third population, an adaptive multi-vector learning mechanism is used to quickly search the third vector to obtain the third vector after the quick search; For the first vector after the search, the second vector after the search, and the third vector after the quick search, obtain the fitness value corresponding to each hyperparameter vector, and determine the mutation probability of the hyperparameter vector according to the fitness value corresponding to the hyperparameter vector; For each hyperparameter vector, according to the mutation probability corresponding to the hyperparameter vector, the annealing simulation rebirth mechanism is used to perform mutation processing to obtain the hyperparameter vector after mutation processing; Determine whether the current number of training times is greater than or equal to the preset maximum number of training times. If so, obtain the final hyperparameters of the deep learning model according to the hyperparameter vector after mutation processing to obtain the deep learning model after training. Otherwise, return to the step of dividing the first population, the second population, and the third population.

5. The method for predicting sports injuries of pickleball based on deep learning according to claim 4, characterized in that: The historical motion data in the deep training sample data is used as the input of the deep learning model, and the expected output of the deep learning model is constructed based on the damage type corresponding to the historical motion data, and the fitness value corresponding to the hyperparameter vector is obtained, including: The historical motion data in the deep training sample data is used as the input of the deep learning model, the expected output of the deep learning model is constructed based on the damage type corresponding to the historical motion data, and the root mean square error value corresponding to the hyperparameter vector is obtained; The root mean square error value corresponding to the hyperparameter vector is added to a positive number less than 0.001 and then the inverse is taken to obtain the fitness value corresponding to the hyperparameter vector.

6. The method for predicting sports injuries in pickleball based on deep learning according to claim 5, characterized in that: For the first vector in the first population, a neighborhood search mechanism based on fitness value decision is used to search for a better region for the first vector, and the first vector after the search is obtained, including: For the first vectors in the first population, obtain the fitness value corresponding to each first vector, and arrange the first vectors in descending order of the fitness values ​​to obtain the arranged first vectors; According to the fitness value corresponding to each first vector and the first vector after arrangement, the degree of goodness of the position corresponding to the first vector is obtained as follows: in, Indicates t During the training process i The position quality of the first vector after the arrangement is i =1,2,…,M1, M1 represents the total number of the first vector, represents the absolute value function, represents the fitness value corresponding to the optimal parameter vector, represents the fitness value corresponding to the worst parameter vector, Indicates i +1 The positional superiority of the first vector after permutation, and i =1, the formula Replace with i The fitness value corresponding to the first vector after permutation ; Indicates i -1 The positional quality of the first vector after the permutation, and i= When M1, the formula Replace with i The fitness value corresponding to the first vector after permutation ; For any first vector, randomly match the first vector with another first vector to obtain a competition vector corresponding to the first vector; Obtain the Euclidean distance between the first vector and the corresponding competition vector, and determine whether the Euclidean distance is less than the preset neighborhood radius R. If so, perform a positional superiority judgment, otherwise perform a fitness value judgment; The step of determining the superiority of the position is as follows: determining whether the superiority of the position corresponding to the first vector is greater than the superiority of the position corresponding to the competition vector; if so, determining the neighborhood search factor , otherwise determine the neighborhood search factor: ； in, Indicates i The fitness value corresponding to the competition vector corresponding to the first vector after permutation, represents a preset positive term, and =0.0001; It represents the exponential function with the natural constant e as the base; The fitness value judgment step is as follows: judging whether the fitness value corresponding to the first vector is greater than the fitness value corresponding to the competition vector; if so, determining the neighborhood search factor , otherwise determine the neighborhood search factor: ； According to the neighborhood search factor, a better region search is performed on the first vector, and the first vector after the search is obtained as follows: ;in, Indicates t During the training process i The d-th dimension parameter of the first vector, d =1,2,…,D, where D represents the total dimension of hyperparameters, Indicates after the search , It is represented by a mean of 0 and a variance of Generates random numbers from a normal distribution.

7. The method for predicting sports injuries in pickleball based on deep learning according to claim 6, characterized in that: For the second vector in the second population, the optimal direction of the second vector is searched according to the optimal hyperparameter vector and the sine-cosine search mechanism of fitness value decision, and the second vector after the search is obtained, including: For the second vector in the second population, a decision factor between (0, 1) is randomly generated for the second vector, and it is determined whether the decision factor is less than 0.

5. If so, a sine search is performed on the second vector in the optimal direction according to the optimal hyperparameter vector to obtain the second vector after the search. Otherwise, a cosine search is performed on the second vector in the optimal direction according to the optimal hyperparameter vector to obtain the second vector after the search. According to the optimal hyperparameter vector, a sine search is performed on the second vector, and the second vector after the search is obtained as follows: in, Indicates t During the training process m The d-th dimension hyperparameter of the second vector, m =1,2,…,M2, M2 represents the total number of the first vector, After search ; represents the adaptive step size control factor, and ; represents the preset step size factor, T represents the preset maximum number of training times, represents the optimal parameter vector d Dimension parameters, represents the first random number between (0,1), represents the second random number between (0,1), sin represents the sine function; According to the optimal hyperparameter vector, a cosine search is performed on the second vector, and the second vector after the search is obtained as: Here, cos represents the cosine function.

8. The deep learning-based pickleball sports injury prediction method according to claim 7, characterized in that: For the third vector in the third population, an adaptive multi-vector learning mechanism is used to quickly search the third vector, and the third vector after the quick search is obtained, including: For a third vector in the third population, two different second vectors are randomly matched to the third vector to obtain a first learning vector and a second learning vector corresponding to the third vector; According to the first learning vector and the second learning vector, a third vector is quickly searched, and the third vector obtained after the quick search is: in, Indicates t During the training process n The third vector d Dimension parameters, Indicates the number of seconds after a quick search. , n =1,2,…,M3, M3 represents the total number of the first vector, represents the first adaptive learning factor, represents the second adaptive learning factor, represents the third random number between (0,1), represents the fourth random number between (0,1), represents the first learning vector d Dimension parameters, represents the second learning vector d Dimension parameters, Indicates n The fitness value of the third vector, represents the fitness value of the first learning vector, represents the fitness value of the second learning vector, represents an exponential function with the natural constant e as the base, represents a preset positive term, and =0.0001.

9. The deep learning-based pickleball sports injury prediction method according to claim 8, characterized in that: The mutation probability of the hyperparameter vector is determined according to the fitness value corresponding to the hyperparameter vector, including: in, represents the probability of mutation, represents a natural constant, represents the fifth random number between (0,1), represents pi, represents the sine function, T represents the maximum number of training times; For each hyperparameter vector, according to the mutation probability corresponding to the hyperparameter vector, the annealing simulation rebirth mechanism is used to perform mutation processing to obtain the hyperparameter vector after mutation processing, including: For each hyperparameter vector, a mutation decision factor between (0,1) is randomly generated, and it is determined whether the mutation decision factor is less than the mutation probability corresponding to the hyperparameter vector. If so, the hyperparameter vector is regenerated to obtain the regenerated hyperparameter vector, and the annealing simulation algorithm is used to selectively accept the regenerated hyperparameter vector to obtain the hyperparameter vector after mutation. Otherwise, the original hyperparameter vector is directly used as the hyperparameter vector after mutation.

10. The deep learning-based pickleball sports injury prediction method according to claim 9, characterized in that: The hyperparameter vector is regenerated, and the hyperparameter vector after regeneration is obtained as follows: The generated mutation factor is: in, represents the factor of variation, It represents the rounding function, and directly takes 1 for numbers less than 1; represents the mutation quantity control parameter, Indicates k The fitness value of the hyperparameter vector to be mutated, represents the fitness value corresponding to the worst parameter vector, Indicates a preset positive number. k =1,2,…,L, L represents the total number of hyperparameter vectors; The generated variation range factor is: Where F represents the variation range factor, represents the variation range control parameter, Represents the fitness value corresponding to the optimal parameter vector; According to the variation range factor, the variation parameter vector is generated as: in, represents the mutation parameter vector d Dimension parameters, Indicates t During the training process k A vector of hyperparameters to be mutated, express A random number between Generate multiple different mutation parameter vectors until the number of mutation parameter vectors reaches the mutation number factor, and determine the mutation parameter vector with the largest fitness value as the new parameter vector; According to the new parameter vector, the hyperparameter vector after regeneration is obtained as: in, represents the d-th dimension parameter of the new parameter vector, represents the d-th dimension parameter of the worst parameter vector, Represents the d-th dimension parameter of the hyperparameter vector after the k-th rebirth.

Images