Sports Injury Risk Detection Method, System and Sports Injury Risk Location Method, System

Abstract

The present invention discloses a method and system for detecting sports injury risks, and a method and system for locating sports injury risk sites. The method collects multi-type biological signals of a human body at preset time points in daily life and at preset time points before and after target sports training to form an original signal data set; uniformly encodes the original signal data set to obtain a biological signal encoding sequence; constructs a Gaussian noise adversarial generation network model, including a generator and a discriminator; inputs the biological signal encoding sequence into the Gaussian adversarial generation network model, sets a total optimization objective for adversarial training, and obtains a trained generator and discriminator; uses the trained discriminator as a sports injury risk detection model, obtains a to-be-identified human biological signal, inputs the to-be-identified human biological signal into the sports injury risk detection model, and obtains a sports injury risk detection result. This method and model can objectively detect and locate sports injury sites with high detection accuracy.

Description

Technical Field

[0001] The present invention relates to the field of sports injury detection, and more specifically, to a method and system for detecting sports injury risks, and a method and system for locating sports injury risk sites. Background Art

[0002] With the continuous improvement of sports competition levels and the increasing requirements for athlete performance, how to effectively prevent injuries has become one of the key issues of concern in the sports world. Many athletes, due to the lack of scientific injury risk warnings, have unscientific training methods or engage in high-intensity competitions with high injury risks, resulting in serious injuries, greatly shortening their careers, and even endangering their lives and health.

[0003] Traditionally, the detection of sports injuries mainly relied on the reports of athletes themselves and the experience judgment of medical teams. This method has the disadvantages of strong subjectivity and obvious lag. In addition, due to the large differences in the physical conditions, technical levels, and training loads of athletes in different fields, it is difficult to meet personalized needs with a unified standard, resulting in a significant reduction in the accuracy and timeliness of injury prediction. Currently, there are also some methods for detecting sports injuries using statistical methods or machine learning algorithms. Although they can reflect the changing trends of physiological parameters to a certain extent, it is difficult to capture the subtle changes in physiological parameters and the potential injury risks behind them. At the same time, such models cannot predict the current injury risks of athletes in real time online, which makes it impossible for athletes and coaching staff to handle injury accidents that occur in real time during competitions.

[0004] The prior art discloses a method, device, and medium for alternating training of a generative adversarial network and medical image classification, which measures the convergence degrees of a generator and a discriminator, and adaptively determines whether to train the generator or the discriminator next according to the current convergence degrees of the generator and the discriminator; finally, disease classification of medical images is performed through an adaptive generative adversarial network. This method mainly relies on medical images for classification and does not combine multiple biological signals, so it is difficult to capture subtle changes and the potential injury risks behind them. At the same time, this method is complex to train and cannot predict the current injury risks of athletes in real time online. Summary of the Invention

[0005] In view of the deficiencies of the prior art in the low accuracy of injury detection and the inability to reflect the injury sites, the present invention provides a method and system for detecting sports injury risks, and a method and system for locating sports injury risk sites. This method can objectively detect and locate sports injury sites with high detection accuracy.

[0006] The primary object of the present invention is to solve the above technical problems, and the technical solution of the present invention is as follows:

[0007] A method for detecting sports injury risks, comprising:

[0008] S101: Collect multi - type biological signals of the human body at preset time points in daily life and at preset time points before and after target sports training to form an original signal data set;

[0009] S102: Perform unified encoding on the original signal data set to obtain a biological signal encoding sequence;

[0010] S103: Construct a Gaussian noise adversarial generation network model, including a generator and a discriminator;

[0011] S104: Input the biological signal encoding sequence into the Gaussian adversarial generation network model, set the total optimization objective for adversarial training, and obtain a trained generator and discriminator;

[0012] S105: Use the trained discriminator as a sports injury risk detection model, obtain the biological signal of the human body to be recognized, input it into the sports injury risk detection model, and obtain the sports injury risk detection result.

[0013] Further, in step S102, performing unified encoding on the original signal data set includes:

[0014] S10201: Resample the multi - type biological signals in the original signal data set to obtain a resampled original signal data set;

[0015] S10202: Segment the biological signals in the resampled original signal data set according to the same preset time window to obtain a corresponding original signal matrix sequence;

[0016] S10203: Calculate the position coordinate encoding and electrode channel encoding matrix of each original signal matrix in each original signal matrix sequence;

[0017] S10204: Unify the dimensions of the original signal matrix sequence to obtain a corresponding dimension - unified original signal matrix sequence;

[0018] S10205: Add the corresponding position coordinate encoding and the corresponding electrode channel encoding matrix to each original signal matrix in each dimension - unified original signal matrix sequence to obtain a corresponding original signal matrix with added position coordinate encoding and electrode channel encoding;

[0019] S10206: Concatenate the original signal matrices with added position coordinate encoding and electrode channel encoding corresponding to the same time to obtain a biological signal encoding sequence.

[0020] Further, in step S10203, the formula for the position coordinate encoding is as follows:

[0021]

[0022] t represents the unit time serial number, i represents the number of electrode channels, and dim represents the maximum number of electrode channels in the original signal matrix sequence segmented by time.

[0023] Further, in step S10203, the calculation of the electrode channel coding matrix includes:

[0024] S1020301: Take the average value of each of the original signal matrices according to the electrode channels to obtain the corresponding electrode channel average matrix;

[0025] S1020302: Unify the dimensions of each of the electrode channel average matrices to obtain the corresponding electrode channel coding matrix.

[0026] Further, the generator includes a first perceptron layer, a first noise perturbation layer, a second perceptron layer, a second noise perturbation layer, a third perceptron layer, a third noise perturbation layer, a fourth perceptron layer, a fourth noise perturbation layer, a fifth perceptron layer, and a fifth noise perturbation layer arranged in sequence;

[0027] The discriminator includes a first convolutional layer, a first BN layer, a second convolutional layer, a second BN layer, a third convolutional layer, a third BN layer, a flattening layer, and a fully connected layer arranged in sequence.

[0028] Further, the total optimization objective is:

[0029]

[0030] G represents the generator, D represents the discriminator, V(D,G) represents the loss function, E represents the expectation, P x (x) represents the input distribution, N P represents the noise, N(x) represents the noise distribution, and x represents the input;

[0031] The loss function of the generator is:

[0032]

[0033] E represents the expectation, N P represents the noise, N(x) represents the noise distribution, x represents the input, P x (x) represents the input distribution, G represents the generator, and D represents the discriminator;

[0034] The loss function of the discriminator is:

[0035]

[0036] E represents the expectation, N Pdenotes noise, N(x) denotes the noise distribution, x denotes the input, and P x (x) denotes the input distribution, G denotes the generator, and D denotes the discriminator.

[0037] A sports injury risk detection system, comprising:

[0038] An original signal acquisition module: acquiring multiple types of biological signals of a human body at preset time points in daily life and at preset time points before and after target sports training to form an original signal data set;

[0039] An original signal encoding module: uniformly encoding the original signal data set to obtain a biological signal encoding sequence;

[0040] An adversarial generation network model construction module: constructing a Gaussian noise adversarial generation network model, including a generator and a discriminator;

[0041] An adversarial training module: inputting the biological signal encoding sequence into the Gaussian adversarial generation network model, setting a total optimization target for adversarial training, and obtaining a trained generator and discriminator;

[0042] An injury risk detection module: using the trained discriminator as a sports injury risk detection model, obtaining a human biological signal to be recognized, inputting the human biological signal into the sports injury risk detection model, and obtaining a sports injury risk detection result.

[0043] A method for locating sports injury risk parts, comprising:

[0044] S201: Acquiring multiple types of biological signals of multiple target parts of a human body at a preset time point after target sports training to form a disease source location data set;

[0045] S202: Modifying the sports injury risk detection model to obtain a sports injury risk feature extraction model;

[0046] S203: Inputting the disease source location data set into the sports injury risk feature extraction model to obtain corresponding disease source risk features;

[0047] S204: Clustering the disease source risk features to obtain multiple final clustering centers;

[0048] S205: Obtaining a human biological signal to be located, inputting the human biological signal into the trained sports injury risk feature extraction model, and obtaining an injury risk feature to be located;

[0049] S206: Calculating the Mahalanobis distance between the injury risk feature to be located and each final clustering center respectively and sorting them to obtain the clustering center corresponding to the minimum Mahalanobis distance as the minimum clustering center;

[0050] S207: Use the target part corresponding to the pathogen risk feature with the largest proportion in the cluster corresponding to the minimum cluster center as the sports injury risk part.

[0051] Furthermore, the sports injury risk feature extraction model removes the fully connected layer in the sports injury risk detection model.

[0052] A sports injury risk part positioning system includes:

[0053] Pathogen location data collection module: Collect multiple types of biological signals of multiple target parts of the human body at a preset time point after target sports training to form a pathogen location data set;

[0054] Feature extraction model construction module: Modify the sports injury risk detection model to obtain a sports injury risk feature extraction model;

[0055] Pathogen risk feature extraction module: Input the pathogen location data set into the sports injury risk feature extraction model to obtain the corresponding pathogen risk features;

[0056] Clustering module: Cluster the pathogen risk features to obtain multiple final cluster centers;

[0057] Injury risk feature extraction module: Obtain the biological signal of the human body to be located and input it into the trained sports injury risk feature extraction model to obtain the injury risk feature to be located;

[0058] Mahalanobis distance calculation module: Calculate the Mahalanobis distance between the injury risk feature to be located and each final cluster center respectively and sort them to obtain the cluster center corresponding to the minimum Mahalanobis distance as the minimum cluster center;

[0059] Target part positioning module: Use the target part corresponding to the pathogen risk feature with the largest proportion in the cluster corresponding to the minimum cluster center as the sports injury risk part.

[0060] Compared with the prior art, the beneficial effects of the present invention are:

[0061] The present invention forms an original signal data set by collecting multiple types of biological signals of the human body at preset time points in daily life and at preset time points before and after target sports training; uniformly encodes the original signal data set to obtain a biological signal coding sequence; constructs a Gaussian noise adversarial generation network model, including a generator and a discriminator; inputs the biological signal coding sequence into the Gaussian adversarial generation network model, sets the total optimization objective for adversarial training, and obtains the trained generator and discriminator; uses the trained discriminator as a sports injury risk detection model, obtains the biological signal of the human body to be identified, inputs it into the sports injury risk detection model, and obtains the sports injury risk detection result; can objectively detect sports injury risks with high detection accuracy.

[0062] By collecting multiple types of biological signals of multiple target parts of the human body at preset time points after target sports training, a disease source location data set is formed; the sports injury risk detection model is modified to obtain a sports injury risk feature extraction model; the disease source location data set is input into the sports injury risk feature extraction model to obtain corresponding disease source risk features; the disease source risk features are clustered to obtain multiple final clustering centers; the biological signal of the human body to be located is obtained, input into the trained sports injury risk feature extraction model to obtain the injury risk feature to be located; the Mahalanobis distances between the injury risk feature to be located and each final clustering center are calculated and sorted respectively, and the clustering center corresponding to the minimum Mahalanobis distance is obtained as the minimum clustering center; the target part corresponding to the disease source risk feature with the largest proportion in the clustering cluster corresponding to the minimum clustering center is used as the sports injury risk part; thus, the location of the sports injury risk is accurately determined. Description of the Drawings

[0063] Figure 1 It is a flowchart of a sports injury risk detection method provided for Example 1.

[0064] Figure 2 It is a flowchart of the acquisition method of the original signal data set provided for Example 1.

[0065] Figure 3 It is a flowchart of uniformly encoding the original signal data set provided for Example 1.

[0066] Figure 4 It is a schematic diagram of the principle of uniformly encoding the original signal data set provided for Example 1.

[0067] Figure 5 It is a flowchart of the calculation of the electrode channel coding matrix provided for Example 1.

[0068] Figure 6 It is a structural diagram of the generator provided for Example 1.

[0069] Figure 7 The structural diagram of the discriminator provided for Embodiment 1.

[0070] Figure 8 The training flow chart of the Gaussian adversarial generation network model provided for Embodiment 1.

[0071] Figure 9 The structural diagram of a sports injury risk detection system provided for Embodiment 1.

[0072] Figure 10 The flow chart of a method for locating sports injury risk sites provided for Embodiment 1.

[0073] Figure 11 The structural diagram of a system for locating sports injury risk sites provided for Embodiment 1.

[0074] Figure 12 The flow chart of a method for real-time sports injury risk detection provided for Embodiment 2. Detailed implementation manners

[0075] The attached drawings are only for illustrative purposes and should not be construed as a limitation to this patent;

[0076] For better illustration of this embodiment, some components in the attached drawings are omitted, enlarged or reduced, which do not represent the dimensions of the actual product;

[0077] For those skilled in the art, it is understandable that some well-known structures and their descriptions in the attached drawings may be omitted.

[0078] The technical solutions of the present invention will be further described below with reference to the attached drawings and embodiments.

[0079] Embodiment 1

[0080] As Figure 1 shown, a method for detecting sports injury risks includes:

[0081] S101: Collect multi-type biological signals of the human body at preset time points in daily life and at preset time points before and after target sports training to form an original signal data set;

[0082] S102: Perform unified encoding on the original signal data set to obtain a biological signal encoding sequence;

[0083] S103: Construct a Gaussian noise adversarial generation network model, including a generator and a discriminator;

[0084] S104: Input the biological signal encoding sequence into the Gaussian adversarial generation network model, set the total optimization objective for adversarial training, and obtain the trained generator and discriminator;

[0085] S105: Use the trained discriminator as a sports injury risk detection model, obtain the human biological signals to be recognized, input them into the sports injury risk detection model, and obtain the sports injury risk detection result.

[0086] In a specific embodiment, if the sports injury risk detection result is greater than a preset threshold, preferably 0.6, it indicates that the athlete's physical condition may reach the limit, and continued training or competition is likely to cause injuries.

[0087] In a specific embodiment, taking the use of two biological signals as an example: electromyogram and electrocardiogram. First, collect the user's biological signals at different times through an electromyograph and an electrocardiograph, such as three times in the morning, noon, and evening, three minutes before training, three minutes after training, or after a certain special training (such as sprint training in sprinting). These biological signals form the original signal dataset.

[0088] In a specific embodiment, as Figure 2 shown, for the basketball field, select six times: 8:00 and 11:00 in the morning, 13:00 and 15:00 at noon, and 18:00 and 20:00 in the evening, and for the special training items in the basketball field: eight basic skill trainings of vigorously dribbling in place, dribbling around obstacles, fixed-point shooting, jump shooting while stopping suddenly, bounce passing, running layup, sliding step, and crossover step; five endurance trainings of 5000-meter long-distance running, sprinting, box jump, squat jump, and quick reaction; eight tactical and psychological quality trainings of meditation, breathing control, man-to-man defense, full-court press, one-on-one offense and defense practice, fast break, pick-and-roll, and triangle offense.

[0089] Two minutes before and two minutes after the above training, as well as the three-minute electrocardiogram, electromyogram, and electroencephalogram signals of the athletes are collected respectively at the six times mentioned above to construct the original signal database.

[0090] Furthermore, as Figure 3 、 Figure 4 shown, in step S102, the unified encoding of the original signal dataset includes:

[0091] S10201: Resample the multi-type biological signals in the original signal dataset to obtain the resampled original signal dataset;

[0092] S10202: Segment the biological signals in the resampled original signal dataset according to the same preset time window to obtain the corresponding original signal matrix sequence;

[0093] S10203: Calculate the position coordinate encoding and electrode channel encoding matrix of each original signal matrix in each original signal matrix sequence;

[0094] S10204: Unify the dimensions in the original signal matrix sequence to obtain the corresponding original signal matrix sequence with unified dimensions;

[0095] S10205: Add the corresponding position coordinate encoding and the corresponding electrode channel encoding matrix to each original signal matrix in the original signal matrix sequence with each dimension unified, to obtain the corresponding original signal matrix after adding the position coordinate encoding and the electrode channel encoding;

[0096] S10206: Concatenate the original signal matrices after adding the position coordinate encoding and the electrode channel encoding corresponding to the same time to obtain a biological signal encoding sequence.

[0097] It should be noted that due to the different sampling rates of each signal, resampling operations need to be performed to unify the sampling rate. Due to the different numbers of channels of each signal, dimension unification is required to facilitate subsequent processing.

[0098] In a specific embodiment, dimension unification expands the signal with a smaller dimension to the dimension of the signal with a larger dimension.

[0099] Further, in step S10203, the formula for the position coordinate encoding is as follows:

[0100]

[0101] t represents the unit time serial number, i represents the number of electrode channels, and dim represents the maximum number of electrode channels in the original signal matrix sequence after time segmentation.

[0102] Further, as Figure 5 shown, in step S10203, the calculation of the electrode channel encoding matrix includes:

[0103] S1020301: Take the average value of each original signal matrix according to the electrode channels to obtain the corresponding electrode channel average matrix;

[0104] S1020302: Unify the dimensions of each electrode channel average matrix to obtain the corresponding electrode channel encoding matrix.

[0105] Further, as Figure 6 shown, the generator includes a first perceptron layer, a first noise perturbation layer, a second perceptron layer, a second noise perturbation layer, a third perceptron layer, a third noise perturbation layer, a fourth perceptron layer, a fourth noise perturbation layer, a fifth perceptron layer, and a fifth noise perturbation layer arranged in sequence;

[0106] As Figure 7As shown, the discriminator includes a first convolutional layer, a first BN layer, a second convolutional layer, a second BN layer, a third convolutional layer, a third BN layer, a flattening layer, and a fully connected layer arranged in sequence.

[0107] It should be noted that different Gaussian noise inputs are added to each layer, and the output is perturbed multiple times to simulate injury conditions. In this way, the output of the generator will try to conform to the fluctuation distribution of the original output, but there is obvious perturbation noise to achieve an accurate simulation of the injury signal.

[0108] In a specific embodiment, the first noise perturbation layer, the second noise perturbation layer, the third noise perturbation layer, and the fourth noise perturbation layer add noise four times respectively, and the added intensities are 20%, 30%, 40%, and 50% respectively.

[0109] Furthermore, the total optimization objective is:

[0110]

[0111] G represents the generator, D represents the discriminator, V(D,G) represents the loss function, E represents the expectation, P x (x) represents the input distribution, N P represents the noise, N(x) represents the noise distribution, and x represents the input;

[0112] The loss function of the generator is:

[0113]

[0114] E represents the expectation, N P represents the noise, N(x) represents the noise distribution, x represents the input, P x (x) represents the input distribution, G represents the generator, and D represents the discriminator;

[0115] The loss function of the discriminator is:

[0116]

[0117] E represents the expectation, N P represents the noise, N(x) represents the noise distribution, x represents the input, P x (x) represents the input distribution, G represents the generator, and D represents the discriminator.

[0118] The training process of the Gaussian adversarial generation network model is as Figure 8 shown.

[0119] As Figure 9 shown, a sports injury risk detection system includes:

[0120] Original signal acquisition module: Collect multiple types of biological signals of the human body at preset time points in daily life and at preset time points before and after target sports training to form an original signal data set;

[0121] Original signal encoding module: Uniformly encode the original signal data set to obtain a biological signal encoding sequence;

[0122] Adversarial generation network model construction module: Construct a Gaussian noise adversarial generation network model, including a generator and a discriminator;

[0123] Adversarial training module: Input the biological signal encoding sequence into the Gaussian adversarial generation network model, set the total optimization objective for adversarial training, and obtain a trained generator and discriminator;

[0124] Injury risk detection module: Use the trained discriminator as a sports injury risk detection model, obtain the biological signal of the human body to be recognized, input it into the sports injury risk detection model, and obtain the sports injury risk detection result.

[0125] As Figure 10 shown, a method for locating sports injury risk sites includes:

[0126] S201: Collect multiple types of biological signals of multiple target sites of the human body at a preset time point after target sports training to form a disease source location data set;

[0127] S202: Modify the sports injury risk detection model to obtain a sports injury risk feature extraction model;

[0128] S203: Input the disease source location data set into the sports injury risk feature extraction model to obtain corresponding disease source risk features;

[0129] S204: Cluster the disease source risk features to obtain multiple final cluster centers;

[0130] S205: Obtain the biological signal of the human body to be located, input it into the trained sports injury risk feature extraction model, and obtain the injury risk feature to be located;

[0131] S206: Calculate the Mahalanobis distance between the injury risk feature to be located and each final cluster center respectively and sort them to obtain the cluster center corresponding to the minimum Mahalanobis distance as the minimum cluster center;

[0132] S207: Take the target site corresponding to the disease source risk feature with the largest proportion in the cluster corresponding to the minimum cluster center as the sports injury risk site.

[0133] In a specific embodiment, the electromyogram signals and electroencephalogram signals are particularly recorded three minutes after special training of body parts such as leg training, arm muscle training, and abdominal training, and a disease source localization data set is constructed.

[0134] In a specific embodiment, for the basketball field, nine parts that are extremely vulnerable to injury in the basketball project, namely the ankle, knee, finger, elbow, shoulder, head, thigh muscle, biceps brachii, and abdominal muscle, are selected, and a certain low-intensity pressure training is performed on them, that is, this part is forced to be used for one minute. The electrocardiogram, electromyogram, and electroencephalogram of the athlete are recorded respectively two minutes before and two minutes after the pressure training, and a disease source localization database is constructed.

[0135] In a specific embodiment, K-Means clustering is used, including:

[0136] S20401: Randomly select N disease source risk features as N initial clustering centers;

[0137] S20402: Assign each disease source risk feature to the cluster closest to the clustering center;

[0138] S20403: Update the clustering center for each of the N clusters;

[0139] S20404: Repeat steps S20402 to S20403 until the clustering center no longer changes or reaches the preset number of iterations;

[0140] S20405: Use the clustering center as the final clustering center.

[0141] In a specific embodiment, since a total of nine body parts have undergone low-intensity pressure training, it is stipulated that the number of clustering clusters is nine, and nine clustering centers can be obtained.

[0142] The formula for updating the clustering center is:

[0143]

[0144] C(k) represents the set of the kth clustering cluster, |C(k)| represents the number of disease source features in the kth clustering cluster, i represents the serial number, and X outputware,i represents the ith disease source risk feature.

[0145] In a specific embodiment, the calculation formula of the Mahalanobis distance is as follows:

[0146]

[0147] Δ is the covariance matrix of all final clustering centers and injury risk features, V outputbs represents the injury risk feature, μ center represents the final clustering center to be calculated, Δ-1 Denotes the inverse matrix of Δ.

[0148] It should be noted that the Mahalanobis distance takes into account the covariance matrix of the vectors and can better reflect the correlation between data points in a high-dimensional space.

[0149] Furthermore, the motion injury risk feature extraction model removes the fully connected layer in the motion injury risk detection model.

[0150] As Figure 11 shown, a motion injury risk location system includes:

[0151] Disease source location data collection module: Collect multiple types of biological signals of multiple target parts of the human body at a preset time point after target sports training to form a disease source location data set;

[0152] Feature extraction model construction module: Modify the motion injury risk detection model to obtain a motion injury risk feature extraction model;

[0153] Disease source risk feature extraction module: Input the disease source location data set into the motion injury risk feature extraction model to obtain corresponding disease source risk features;

[0154] Clustering module: Cluster the disease source risk features to obtain multiple final clustering centers;

[0155] Injury risk feature extraction module: Obtain the biological signal of the human body to be located, input it into the trained motion injury risk feature extraction model, and obtain the injury risk feature to be located;

[0156] Mahalanobis distance calculation module: Calculate the Mahalanobis distance between the injury risk feature to be located and each final clustering center respectively and sort them to obtain the clustering center corresponding to the minimum Mahalanobis distance as the minimum clustering center;

[0157] Target part location module: Take the target part corresponding to the disease source risk feature with the largest proportion in the clustering cluster corresponding to the minimum clustering center as the motion injury risk part.

[0158] Embodiment 2

[0159] Based on the motion injury risk detection method described in Embodiment 1, that is, this embodiment uses the same motion injury risk detection method as Embodiment 1.

[0160] As Figure 12As shown in the figure, in order to improve the practical effect of the model, the function of real-time detecting the injury risk of athletes is added by combining the Test Time Adaptation (TTA) technology. Using the data collected during the daily training stage, the model is pre-trained in advance to obtain a customized pre-trained model exclusive to this athlete. In stages that require real-time detection such as competitions, we equip the athletes with simple sensors, and at this time, various biological signals of the athletes can be obtained in real time, and real-time detection of sports injury risks and location of injury risk sites can be performed.

[0161] Since simple and lightweight sensors are worn during the athletes' competitions, and the competition state is different from the training state, the distribution of biological signals also changes. If the discriminator described in Embodiment 1 is directly used, there may be deviations. Therefore, if real-time testing is required, the first BN layer, the second BN layer, and the third BN layer use the moving average method to modify the statistics.

[0162] The principle of modifying the statistics by the moving average method is as follows:

[0163] The first BN layer, the second BN layer, and the third BN layer in Embodiment 1 use the global mean and variance for normalization, and the formula is as follows:

[0164]

[0165] represents the input of the i-th BN layer, represents the first processing result of the input of the i-th BN layer, represents the second processing result of the input of the i-th BN layer, μ and σ 2 are the accumulated global mean and variance respectively, ∈ is an extremely small value to prevent the denominator from being zero, γ and β are learnable parameters, and i represents the serial number.

[0166] When only one data point is passed in at a time, the batch size is too small to accurately estimate the mean and variance.

[0167] Therefore, the improved formulas for the mean and variance are as follows:

[0168]

[0169] μ ′ represents the improved mean, σ ′2 represents the improved variance, α represents the momentum parameter, x represents the i-th input, and i represents the serial number.

[0170] It should be noted that the moving average statistics μ ′ and σ ′2Perform normalization and scaling translation. After each segment of real-time signal is input and the detection result is obtained, the parameters of the model will return to the original pre-trained state. This means that the tests for each segment of signal are independent and do not affect each other. Slightly modifying the model parameters according to the input biological signal segment each time can help eliminate the signal distribution differences between different sensors or different occasions, and ensure that the original pre-trained model after being trained with a large amount of data still plays a discriminative role, guaranteeing the training accuracy and eliminating the distribution error.

[0171] The same or similar reference numerals correspond to the same or similar components;

[0172] The terms used to describe the positional relationship in the drawings are for illustrative purposes only and should not be construed as a limitation of this patent;

[0173] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, rather than limitations on the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for detecting the risk of sports injuries, characterized in that, Including: S101: Collect multiple types of biological signals of the human body at daily preset time points and preset time points before and after target sports training to form an original signal data set; S102: Uniformly encode the original signal data set to obtain a biological signal encoding sequence; S103: Construct a Gaussian noise adversarial generation network model, including a generator and a discriminator; S104: Input the biological signal encoding sequence into the Gaussian adversarial generation network model, set the total optimization objective for adversarial training, and obtain the trained generator and discriminator; S105: Use the trained discriminator as a sports injury risk detection model, obtain the biological signals of the human body to be identified, input them into the sports injury risk detection model, and obtain the sports injury risk detection result; In step S102, uniformly encoding the original signal data set includes: S10201: Resample multiple types of biological signals in the original signal data set to obtain the resampled original signal data set; S10202: Segment the biological signals in the resampled original signal data set according to the same preset time window to obtain the corresponding original signal matrix sequence; S10203: Calculate the position coordinate encoding and electrode channel encoding matrix of each original signal matrix in each original signal matrix sequence; S10204: Unify the dimensions in the original signal matrix sequence to obtain the corresponding original signal matrix sequence with unified dimensions; S10205: Add the corresponding position coordinate encoding and corresponding electrode channel encoding matrix to each original signal matrix in each original signal matrix sequence with unified dimensions to obtain the corresponding original signal matrix after adding position coordinate encoding and electrode channel encoding; S10206: Concatenate the original signal matrices after adding position coordinate encoding and electrode channel encoding corresponding to the same time to obtain a biological signal encoding sequence; In step S10203, the calculation of the electrode channel encoding matrix includes: S1020301: Take the average value of each original signal matrix according to the electrode channels to obtain the corresponding electrode channel average matrix; S1020302: Unify the dimensions of each electrode channel average matrix to obtain the corresponding electrode channel encoding matrix.

2. The method for detecting the risk of sports injuries according to claim 1, wherein In step S10203, the formula for the position coordinate encoding is as follows: Indicates the unit time serial number, Indicates the number of electrode channels, Indicates the maximum number of electrode channels in the original signal matrix sequence segmented by time.

3. The method for detecting sports injury risks according to claim 1, wherein, The generator includes a first perceptron layer, a first noise perturbation layer, a second perceptron layer, a second noise perturbation layer, a third perceptron layer, a third noise perturbation layer, a fourth perceptron layer, a fourth noise perturbation layer, a fifth perceptron layer, and a fifth noise perturbation layer arranged in sequence; The discriminator includes a first convolutional layer, a first BN layer, a second convolutional layer, a second BN layer, a third convolutional layer, a third BN layer, a flattening layer, and a fully connected layer arranged in sequence.

4. The method for detecting the risk of sports injuries according to claim 1, wherein The total optimization objective is: Denotes a generator, Denotes a discriminator, Denotes a loss function, Denotes an expectation, Denotes an input distribution, Denotes noise, Denotes a noise distribution, Denotes an input; The loss function of the generator is: Indicates expectation Indicates noise Indicates noise distribution Indicates input Indicates input distribution Indicates generator Indicates discriminator; The loss function of the discriminator is: Indicates expectation, Indicates noise, Indicates noise distribution, Indicates input, Indicates input distribution, Indicates generator, Indicates discriminator.

5. A sports injury risk detection system, applied to the detection method described in any one of claims 1 to 4, characterized in that, Including: Original signal acquisition module: Collect multiple types of biological signals of the human body at daily preset time points and preset time points before and after target sports training to form an original signal data set; Original signal encoding module: uniformly encode the original signal dataset to obtain a biological signal encoding sequence; Adversarial generation network model construction module: construct a Gaussian noise adversarial generation network model, including a generator and a discriminator; Adversarial training module: input the biological signal encoding sequence into the Gaussian adversarial generation network model, set the total optimization objective for adversarial training, and obtain a trained generator and discriminator; Injury risk detection module: use the trained discriminator as a sports injury risk detection model, obtain the human biological signal to be recognized, input it into the sports injury risk detection model, and obtain the sports injury risk detection result.

6. A method for locating risk sites of sports injuries, applied to the sports injury risk detection model according to any one of claims 1 to 4, characterized in that, Including: S201: Collect multi-type biological signals of multiple target parts of the human body at a preset time point after target sports training to form a disease source location dataset; S202: Modify the sports injury risk detection model to obtain a sports injury risk feature extraction model; S203: Input the disease source location dataset into the sports injury risk feature extraction model to obtain corresponding disease source risk features; S204: Cluster the disease source risk features to obtain multiple final cluster centers; S205: Obtain the human biological signal to be located, input it into the trained sports injury risk feature extraction model, and obtain the injury risk feature to be located; S206: Calculate the Mahalanobis distance between the injury risk feature to be located and each final cluster center respectively and sort them to obtain the cluster center corresponding to the minimum Mahalanobis distance as the minimum cluster center; S207: Take the target part corresponding to the disease source risk feature with the largest proportion in the cluster corresponding to the minimum cluster center as the sports injury risk part.

7. The method for positioning a risk site of sports injuries according to claim 6, wherein The sports injury risk feature extraction model removes the fully connected layer in the sports injury risk detection model.

8. A positioning system for sports injury risk sites, applied to the positioning method described in claim 6 or 7, characterized in that, Including: Disease source location data collection module: collect multi-type biological signals of multiple target parts of the human body at a preset time point after target sports training to form a disease source location dataset; Feature extraction model construction module: modify the sports injury risk detection model to obtain a sports injury risk feature extraction model; Disease source risk feature extraction module: input the disease source location dataset into the sports injury risk feature extraction model to obtain corresponding disease source risk features; Clustering module: cluster the disease source risk features to obtain multiple final cluster centers; Injury risk feature extraction module: obtain the human biological signal to be located, input it into the trained sports injury risk feature extraction model, and obtain the injury risk feature to be located; Mahalanobis distance calculation module: calculate the Mahalanobis distance between the injury risk feature to be located and each final cluster center respectively and sort them to obtain the cluster center corresponding to the minimum Mahalanobis distance as the minimum cluster center; Target part location module: take the target part corresponding to the disease source risk feature with the largest proportion in the cluster corresponding to the minimum cluster center as the sports injury risk part.

Images