Dynamic risk early warning method and system based on student exercise data

Abstract

The invention discloses a dynamic risk early warning method and system based on student exercise data. The method comprises the steps that S1, original exercise data of student exercise is collected through wearable sensing equipment; s2, obtaining mapping motion data through an improved near-end depth Q network based on the original motion data and the normal motion data; s3, according to the mapping relation data, obtaining an action risk point through threshold comparison and considering duration time; and S4, performing matching detection on the obtained action risk points through an abnormal action chain type detection mechanism based on motion time-space association to obtain a risk prediction result, thereby avoiding misjudgment of a single tiny deviation as a risk, capturing a damage trend of a continuous abnormal action, scientifically fitting a motion damage occurrence rule, and accurately adapting to various campus motion scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sports risk prediction, in particular to a risk dynamic early warning method and system based on student sports data. BACKGROUND

[0002] In recent years, the development of wearable sensing technology (such as smart bracelets integrated with three-axis accelerometers, gyroscopes) and reinforcement learning algorithms has provided a technical basis for multi-dimensional collection of sports data and dynamic risk modeling, and has promoted the technical development of student sports risk early warning from pre-forecasting.

[0003] Currently, the existing sports risk prediction system usually uses single sensor threshold judgment, resulting in insufficient dimensions, only capturing single features such as acceleration or angular velocity, and being unable to represent complex motion patterns (such as the coordinated trajectory of the knee joint and the ankle joint when running) such as joint linkage and action timing, resulting in incomplete risk feature extraction, and it is difficult to adapt to individual differences through static threshold prediction, such as different students' joint range of motion, different exercise habits, so a risk dynamic early warning method and system based on student sports data is proposed. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, in order to achieve the above-mentioned purpose, the present application proposes the following technical scheme:

[0005] A risk dynamic early warning method based on student sports data, comprising:

[0006] S1: collecting original sports data of students' sports through a wearable sensing device;

[0007] S2: obtaining mapped action data based on original action data in combination with normal sports data through an improved proximal deep Q network;

[0008] The improved proximal deep Q network combines the policy optimization mechanism of PPO with the value learning of deep Q network, and uses a clipped target function to constrain the PPO policy update step length to obtain;

[0009] S3: obtaining action risk points by threshold comparison and considering the duration according to the mapping relationship data;

[0010] S4: matching and detecting the obtained action risk points through an abnormal action chain detection mechanism based on the spatiotemporal correlation of sports to obtain a risk prediction result.

[0011] The wearable sensing device is a smart sports bracelet with a built-in three-axis accelerometer, gyroscope, and magnetometer;

[0012] The original sports data includes three-dimensional linear acceleration data and three-dimensional angular velocity data.

[0013] The experience replay pool construction process of the improved proximal deep Q network is:

[0014] Collect normal motion data, collect three-dimensional linear acceleration data and three-dimensional angular velocity data in normal motion state through wearing sensing equipment to form normal motion data;

[0015] Set the experience replay pool R, and the experience replay pool sample includes original motion data and normal motion data;

[0016] Divide the original motion data and the normal motion data by 1 second window and 0.5 second step at the same time;

[0017] Then the experience replay pool sample is completed by the priority replay mechanism management.

[0018] The training process of the improved proximal deep Q network is:

[0019] From the experience replay pool R, extract the original motion data and the normal motion data according to the proportion of 60% of the normal motion data sample and 40% of the preprocessed original motion data sample, and construct a mixed training set;

[0020] Extract the motion data features of the mixed training set sample through the convolutional neural network, and map them to the action value space by the fully connected layer;

[0021] Initialize the DQN and PPO policy network, and obtain the policy optimization loss L of PPO CLIP and the value learning loss L of DQN Tz Fuse the policy optimization loss of PPO and the value learning loss of DQN based on the composite loss weights λ1 and λ2 to obtain the composite loss L total , the formula is:

[0022] L total =λ1L TD +λ2L CLIP

[0023] And use the Adam optimizer to update the network parameters, calculate the composite loss L total The gradient of network parameters θ Then update the network parameters until the loss function L total tends to be stable, and the training is completed.

[0024] The process of using the clipped target function to constrain the PPO policy update step is:

[0025] Collect sample data from the mixed training set, which includes motion state, action executed corresponding to the motion state, reward obtained after executing the action, and next motion state after executing the action;

[0026] inputting the motion state into the value network to obtain the value of performing the action in the current state, and inputting the next state into the target value network to obtain the target value;

[0027] inputting the motion state into the policy network to obtain a new policy probability of performing the action in the current state, and obtaining an old policy probability of performing the action in the same state from the old policy network, dividing the new policy probability by the old policy probability to obtain a policy ratio;

[0028] based on the value calculation result of the DQN, obtaining an advantage estimate by subtracting the value of the current action from the target value;

[0029] clipping the policy ratio to limit it in the interval [1-, 1+], and combining it with the advantage estimate to obtain a policy optimization loss L of PPO CLIP completing the PPO constraint policy update step.

[0030] The action risk point acquisition process is:

[0031] collecting standard motion parameters, and determining a risk trigger threshold according to motion injury case statistics;

[0032] extracting the deviation type, deviation value and deviation duration in the current motion state from the mapping relationship data;

[0033] when the deviation value is greater than the risk trigger threshold and the deviation duration is greater than the risk duration threshold, the corresponding action data is determined as an action risk point.

[0034] The process of matching and detecting the obtained action risk point is:

[0035] defining a motion space-time correlation relationship, including a time dimension and a space dimension, the time dimension covering the start time, end time and duration of the action risk point, and the space dimension including the coordinate range and the coordinated motion trajectory involved in the action;

[0036] based on historical motion injury data, obtaining a motion space-time correlation relationship, the motion space-time correlation relationship including a time sequence correlation rule, a concurrent correlation rule and a random correlation rule;

[0037] expressing the above rules as space-time correlation rules, and storing all rules to build a space-time correlation rule library;

[0038] inputting the action risk point data to arrange them in time sequence to form a time sequence of action risks, and combining the spatial distribution of body parts to construct a space-time sequence structure of the risk points, traversing the rules in the space-time correlation rule library to perform chain matching on the space-time sequence structure to obtain a risk prediction result.

[0039] The process of performing chain matching is:

[0040] From the time dimension, check whether the duration and the order of occurrence of the action risk point conform to the time condition in the space-time correlation rule library, and from the space dimension, verify whether the involved body parts and the deviation value range match the space condition in the space-time correlation rule library;

[0041] When the space-time sequence structure matches any space-time correlation rule in the rule library, the corresponding risk prediction result is output, including the action risk point appearing simultaneously in the time dimension and the space dimension, and if no rule is matched, it is marked as normal.

[0042] A risk dynamic early warning based on student motion data, comprising:

[0043] The data acquisition module acquires original motion data of the student motion through a wearable sensing device;

[0044] The action mapping module acquires mapping action data based on the original action data and normal motion data through an improved proximal deep Q network;

[0045] The risk acquisition module acquires action risk points by threshold comparison and considering the duration according to the mapping relationship data;

[0046] The prediction output module matches and detects the acquired action risk points through an abnormal action chain detection mechanism based on motion space-time correlation to obtain a risk prediction result.

[0047] The present application has the following beneficial effects:

[0048] In the present application, first, three-dimensional motion data is acquired through fusion of multiple sensors (three-axis accelerometer, gyroscope) of a wearable device, and space-time features (such as joint coordinated trajectory and action timing) are extracted by combining an improved proximal deep Q network (PPO-DQN), which greatly improves the risk identification accuracy compared to traditional single threshold judgment;

[0049] Secondly, the PPO policy optimization mechanism is combined with the DQN value learning, and the policy update step is constrained by a clipped target function, so that the model can adapt to different student action modes (such as the difference in running posture of students with different body types), and the generalization in the diverse motion scenarios in the campus (basketball, gymnastics) is improved;

[0050] Finally, a space-time correlation rule library is constructed based on historical injury data, and chain matching is combined with the time dimension (deviation duration) and the space dimension (joint coordinate deviation and coordinated trajectory), which avoids misjudgment of a single small deviation as a risk and captures the injury trend of continuous abnormal action (such as knee joint inward buckling for more than 3 seconds when changing direction in basketball), which not only conforms to the motion injury occurrence rule, but also accurately adapts to the diverse motion scenarios in the campus. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the steps of a risk dynamic early warning method and system based on student exercise data proposed in this invention.

[0052] Figure 2 This invention presents a system block diagram of a dynamic risk warning method and system based on student movement data. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Example 1

[0055] like Figure 1 As shown, the present invention proposes a dynamic risk warning method based on student exercise data, comprising:

[0056] S1: Collect raw motion data of students' movements through wearable sensing devices;

[0057] Wearable devices integrating multiple types of sensors were selected to collect raw motion data D1 for a smart sports bracelet with a built-in three-axis accelerometer, gyroscope, and magnetometer;

[0058] Accelerometers are used to capture linear acceleration changes during movement, gyroscopes detect angular velocity to reflect limb rotation state, and magnetometers assist in calibrating direction to determine the orientation of movement posture;

[0059] Specifically, the device wearing position is optimized for different sports (such as running, basketball, and gymnastics). When running, the wristband is worn on the wrist. When playing basketball, the wristband is worn on the wrist. Additionally, a small inertial measurement unit (IMU) device is worn on the knee joint to accurately collect joint motion data. During deployment, Bluetooth technology is used to achieve wireless connection between the device and the data receiving terminal (sports monitoring base station, mobile phone).

[0060] The data types in the raw motion data D1 include three-dimensional linear acceleration data A(t) and three-dimensional angular velocity data Ω(t);

[0061] Three-dimensional linear acceleration A(t)=[I x (t), I y (t), I z[(t)], where t is the timestamp, I is the acceleration data, x represents the forward and backward direction, i.e. the direction of motion, y represents the left and right direction, i.e. the horizontal direction of motion, and z represents the up and down direction of motion, i.e. the direction of gravity of motion;

[0062] For example, capturing the acceleration changes of rapid limb movements (such as the force of a basketball shot or the arm swing during running) can be used to calculate the explosive force of a movement (such as the force during a shot). z (t) Peak value corresponds to the intensity of hand force exertion;

[0063] 3D angular velocity Ω(t)=[ω x (t), ω y (t), ω z [(t)], where ω is the angular velocity data;

[0064] For example, by monitoring limb rotation status (such as knee flexion angular velocity, gymnastic trunk rotation angular velocity) using three-dimensional angular velocity Ω(t), try to obtain joint angular velocity to judge movement deviation (such as whether the knee joint angle exceeds the normal range of [60°, 120°] when dribbling in basketball).

[0065] S2: Based on the original motion data combined with normal motion data, mapped motion data is obtained through an improved proximal deep Q-network;

[0066] Normal motion data is collected by collecting three-dimensional linear acceleration data A(t) and three-dimensional angular velocity data Ω(t) under normal motion conditions through wearable sensing devices to form normal motion data D2;

[0067] An improved proximal deep Q-network is constructed. The basic framework adopts a deep Q-network (DQN), which includes a convolutional neural network (CNN) and a fully connected layer (FC). The CNN is used to extract the spatiotemporal features of motion data.

[0068] The construction process of the improved near-end deep Q network experience replay pool is as follows:

[0069] Set up an experience replay pool R to store training samples. The experience replay pool samples contain the original motion data D1 and normal motion data D2.

[0070] The raw motion data D1 is generated by collecting raw data from students' daily movements (unfiltered real motion scenarios) during the initial stage of actual system operation, and dividing the data according to time windows (e.g., 1 second as the window, 0.5 seconds step).

[0071] Specifically, normal exercise data D2 is constructed and defined by collecting a large amount of standard student movement data (such as standardized running arm swing, standard radio gymnastics movements, etc.) and clarifying movement standard parameters (such as joint angle range, movement timing characteristics), and dividing the data according to time windows (1 second window, 0.5 seconds step length);

[0072] Finally, the samples are managed by a priority playback mechanism, which assigns different priorities to the samples according to their contribution to the model training, to ensure that the training process makes full use of effective samples and completes the experience replay pool construction.

[0073] Construct a mixed training set to balance the standard and real scene data:

[0074] From the experience replay pool R, extract the original motion data D1 and normal motion data D2 according to the proportion of 60% normal motion data and 40% preprocessed original motion data, and construct a mixed training set

[0075] Specifically, the preprocessing process of 40% original motion data is to remove obvious abnormal data (such as noise data with acceleration far exceeding the human motion limit) through initial threshold screening (based on historical normal motion parameter range) to ensure that the original data included in the mixed training set is within the normal motion range and unlabeled data.

[0076] Further, the advantage of the mixed training set is that it allows the model to learn both standard action specifications (based on normal motion data D2) and adapt to the diversity and uncertainty of real motion scenes (based on normal motion data D1), avoiding model overfitting to standard data and improving the adaptability to actual student motion actions.

[0077] Input the sample data (including normal motion data and original motion data) in the mixed training set D into the basic network. First, extract the motion data features through CNN, and the multi-layer convolution layer operation follows the formula y l = f(W l *y l-1 +b l ));

[0078] After convolution layer processing, the feature vector is input into the fully connected layer and mapped to the action value space to calculate the output Q value Q(s, a; θ), i.e. the action value function, which represents the expected value of performing action a in state s.

[0079] Combine the policy optimization mechanism of PPO (Proximal Policy Optimization algorithm) with the value learning of deep Q network, and constrain the PPO (Proximal Policy Optimization algorithm) policy update step through the clipped target function to obtain an improved proximal deep Q network. The implementation process is as follows:

[0080] First, initialize the DQN and PPO policy network, the process is as follows:

[0081] Initialize the value network Q θ(s, a), which acts as learning "in the state of motion s, the value of performing action a", the output result is the quantitative evaluation of the value of action, and at the same time, reinitialize a target value network

[0082] Initialize the policy network π of PPO θ (s, a), learn the probability distribution of selecting each action a in the state of motion s, decide the output action preference, set the policy old network π θold (s, a), the initial policy old network parameter and the policy network π θ (s, a) is consistent;

[0083] Then set the key hyperparameters in advance, including:

[0084] Clip parameter ∈: set to 0.2, which is the core parameter in PPO to restrict the amplitude of policy update, to prevent the policy update from being too aggressive;

[0085] Discount factor γ: take 0.95, used to balance the weight of the reward generated by the current motion action and the reward brought by the future action, so that the model can reasonably weigh the short-term and long-term value.

[0086] Compound loss weight λ1, λ2, set to 0.6 and 0.4 respectively, to control the influence of PPO policy optimization loss and DQN value learning loss in the overall optimization process;

[0087] Learning rate α: set to 1e-4, to determine the step size of each parameter update;

[0088] The process of using the clipped target function to constrain the PPO policy update step is as follows:

[0089] Collect sample data from the mixed training set D, and the structure of each sample is (s i , a i , r i , s i+1 ), where i is the index, s i is the state of motion, a i is the action performed in the state of motion, r i is the reward obtained after performing the action (the more standard the action, the higher the reward; the action is risky, and the reward is negative), s i+1 is the next state of motion after performing action a i ;

[0090] Input the state of motion s i to the value network Q θ (s, a), the network outputs the current state si Next, execute action a i Value Q θ (s i r i );

[0091] Next state s i+1 Input target value network calculate:

[0092] in, The next state action value is the output of the target value network, and γ is the discount factor.

[0093] Then the motion state s i Input Policy Network π θ (s, a), calculate the state in the current state s i Next, execute action a i The new policy probability P, and simultaneously, the policy from the old network π. θold In (s, a), obtain the state s in the same state. i Next, execute action a i The probability P of the old strategy old ;

[0094] Specifically, by comparing these two probabilities, we can reflect the direction and magnitude of the strategy update;

[0095] Divide the new policy probability by the old policy probability to obtain the policy ratio r. i The strategy ratio measures the performance of the new strategy relative to the old strategy for action a. i The degree of change in preferences;

[0096] Based on the value calculation results of DQN, the target value y is... i Subtract the current action value Q θ (s i r i Obtain advantage estimation

[0097] Specifically, advantage estimation Reflects the execution of action a i The advantage of the action over the average action makes the network more inclined to choose the action with the higher advantage when optimizing the policy;

[0098] Specifically, by using advantage estimation to reflect the advantage of a movement relative to the average movement, and combining it with the clipped strategy ratio, it can accurately identify high-risk movement patterns that may cause injury during student sports, such as incorrect force application and abnormal joint linkage, and detect potential risks in advance.

[0099] Finally, the strategy ratio r will be compared. iThe clipping restricts the interval to [1-, 1+], and the loss L of the PPO network is obtained by combining the advantage estimation CLIP The PPO constraint policy update step is completed

[0100] Specifically, the principle of the clipped target function is to combine the advantage estimation with the clipped policy ratio to utilize the new policy improvement direction and limit the update step. By updating around the policy gradient, the constraint of the policy update step can make the network more stable when learning and optimizing the motion strategy, help the student to avoid risks caused by improper actions, and dynamically monitor the student's motion state with the help of the cooperation of the value network and the policy network. In combination with the rules formed by the historical injury data, the motion risk is dynamically warned, which provides a basis for timely intervention (such as adjusting the action and reducing the intensity);

[0101] Then the loss function of the improved proximal deep Q network is calculated, and the process is as follows:

[0102] Based on the target value, the DQN value learning loss is obtained:

[0103] The PPO policy optimization loss L CLIP and the DQN value learning loss L TD are fused based on the composite loss weights λ1 and λ2 to obtain the composite loss L total , which is represented as: L total = λ1L TD + λ2L CLIP ;

[0104] The parameter update adopts the Adam optimizer, and the gradient of the composite loss L total is calculated with respect to the network parameters θ (including the parameters of the value network and the policy network) Then the network parameters are updated according to (α is the learning rate) until the loss function L total tends to be stable (the loss change rate is less than (1e-5) in continuous several rounds of iterations, and the Q value and the actual action value fitting degree meet the standard (MSE<0.01 on the validation set)), and the training is completed. At this time, the model realizes the combination of the PPO policy optimization mechanism and the DQN value learning, and the policy update step is constrained by the clipped target function;

[0105] Specifically, the improved proximal deep Q network is trained to accurately evaluate the risk level of various actions in student movement, identify risk actions such as abnormal joint angles and incorrect force timing that are prone to injury, and improve the stability of the model in learning student movement data through techniques such as experience replay, target network, and PPO clipping strategy to constrain the update step. This allows the model to continuously mine effective movement risk patterns, real-time combine spatiotemporal data of student movement (such as joint trajectory and action duration), dynamically match the rule library, and timely warn of movement risks, providing a basis for intervention in student movement and preventing injuries, and ensuring movement safety.

[0106] At this time, the original movement state s (original movement data collected) is input into the model, which outputs the action mapping result through Q θ (s, a) and the policy θ (s, a) to output the deviation value of the original action from the standard specification (such as a 20° deviation in knee joint angle), forming a mapping relationship data M.

[0107] Specifically, the mapping relationship data M is the structured output of the model after training, which compares the differences between the original movement state s and the standard action. For example, by comparing the current joint angle with the standard joint angle, the mapping relationship data M obtained through the improved proximal deep Q network includes the joint angle deviation (such as knee joint and elbow joint).

[0108] S3: Obtain the action risk point by threshold comparison and considering the duration based on the mapping relationship data.

[0109] Collect standard movement parameters (such as the knee joint flexion angle range during normal running is 110°-130°)), and determine the risk trigger threshold T risk and the risk duration threshold t min (For example, a knee joint angle deviation of more than 20° for 5 seconds is determined as a risk) based on the statistics of movement injury cases. From the mapping relationship data M, extract the deviation type (such as joint angle deviation and action frequency deviation) under the current movement state, the deviation value d (such as a knee joint angle deviation of 25°), and the duration t of the deviation (by continuously collecting movement data, the cumulative duration of the deviation state is calculated).

[0110] Determine the risk abnormality by threshold comparison, for example, there is a joint angle deviation, compare the deviation value d of the joint angle deviation with the risk trigger threshold T risk , that is, determine whether d>T risk is true.

[0111] At the same time, determine whether the duration t of the deviation satisfies the risk duration threshold t≥t min .

[0112] When d>T risk and t≥tmin At this time, the corresponding action data is determined as an action risk point;

[0113] The information determined as a risk point is standardized and arranged, including deviation type (whether it is joint angle or action trajectory), deviation specific value (such as 25°, 30 times / minute, etc.), duration, corresponding sports project (such as running, basketball), involved body part (such as knee joint, waist), to form structured action risk point data R p ;

[0114] Specifically, the standardized action risk point data R p includes sports project, body part, deviation type, deviation value, duration, three core features of deviation type, deviation value d, and duration t are extracted from the mapping relationship data M, the key indicators required for risk determination are accurately focused, and double condition constraints (deviation value breaks through risk threshold and duration meets damage accumulation condition) are adopted to improve the comprehensiveness of risk comparison.

[0115] S4: Match and detect the obtained action risk points through an abnormal action chain detection mechanism based on the spatiotemporal correlation of sports, and obtain a risk prediction result;

[0116] The process of matching and detecting the obtained action risk points is as follows:

[0117] Define the spatiotemporal correlation relationship of sports, including time dimension and space dimension;

[0118] The time dimension covers the start time t start , end time t end , and duration Δ t of the action risk point;

[0119] The space dimension includes the spatial distribution of the body part where the action occurs (the coordinate range (x, y, z) of the knee joint in three-dimensional space and the coordinated motion trajectory involved in the action;

[0120] Specifically, the coordinated motion trajectory is represented by four-dimensional spatiotemporal vector quantization. Taking the coordinated motion of the knee joint and the ankle joint during running as an example:

[0121] Where: ωknee(t), ωankle(t) represent the angular velocity of the knee joint and the ankle joint at time t;

[0122] Project the of the continuous time sequence to a two-dimensional phase space (with ωknee(t) as the horizontal axis and ωankle(t) as the vertical axis). The normal coordinated motion trajectory forms an elliptical ring. When there is an action deviation, the trajectory deviates from the ring and forms a distorted trajectory;

[0123] For example, when the knee joint angle deviation exceeds 20°, the trajectory horizontal axis exceeds the range of the ellipse and the trajectory exhibits a divergent shape.

[0124] Based on historical sports injury data, spatiotemporal correlations of sports injuries are obtained, and data from historical sports injury data are represented as spatiotemporal correlation rules (RBs). ts ;

[0125] Specifically, spatiotemporal risks include action risks that occur simultaneously in both the time and space dimensions;

[0126] Furthermore, the spatiotemporal association rules include:

[0127] Temporal correlation rules: The pattern of risk points appearing in a specific time sequence (e.g., ankle joint deviation first, then knee joint deviation);

[0128] Concurrency association rule: The pattern of multiple body parts having risk points at the same time (such as simultaneous deviation of the knee joint and hip joint);

[0129] Random association rules: No fixed temporal sequence but satisfying the pattern of spatial coordination deviation (such as sudden abnormal coordination between the waist and shoulder);

[0130] For example:

[0131] In historical injury data of basketball, the most frequently occurring spatiotemporal risk is an injury pattern in which the knee joint angle deviation exceeds 20° (spatial characteristic) for 5-10 seconds (temporal characteristic) and is accompanied by an ankle joint force deviation exceeding 3 times the body weight (spatial correlation characteristic).

[0132] Store all the mined spatiotemporal correlation rules to build a spatiotemporal correlation rule library RB. ts ;

[0133] Specifically, the spatiotemporal association rule base uses a relational database such as MySQL to store the mined spatiotemporal association rules in a structured format (such as rule ID, condition set, conclusion, confidence score, etc.), thereby constructing the spatiotemporal association rule base RB. ts The database's storage and query functions are used to manage and invoke rules;

[0134] Input action risk point data R p Arrange the risks in chronological order to form a time series, and combine this with the spatial distribution of body parts to construct a spatiotemporal sequence structure R of the risk points. p_seq ;

[0135] Chain matching is performed on the spatiotemporal sequence structure, i.e.:

[0136] Traversing the Spatiotemporal Association Rule Base (RB) ts The rules in the spatiotemporal sequence structure R p_seq Perform chained matching:

[0137] From the time dimension, check whether the duration and the order of occurrence of the risk points conform to the time conditions in the space-time correlation rule library. From the spatial dimension, verify whether the involved body parts and the range of deviation values match the spatial conditions in the space-time correlation rule library;

[0138] For example, for the rule R ts , check whether the sports item in R p_seq is basketball, whether the knee joint angle deviation value is more than 20°, whether the duration is in the interval of 5-10 seconds, whether there is a risk point of ankle joint stress deviation more than 3 times the body weight, and whether the space-time order of these risk points conforms to the chain rule of injury occurrence (the knee joint deviation appears first, followed by the ankle joint deviation);

[0139] When the space-time sequence structure R p_seq matches any space-time correlation rule in the rule library, output the corresponding risk prediction result and give a warning, including the action risk points that appear simultaneously in the time dimension and the spatial dimension. If no rule is matched, mark it as normal;

[0140] Specifically, by storing space-time correlation rules, using the structured storage and efficient query characteristics of a relational database, the rules can be managed in order;

[0141] Secondly, by sorting the action risk point data sequence by time and combining the spatial distribution to construct a space-time sequence structure, the actual characteristics of the space-time correlation of sports risks are met, and the dispersed risk points are integrated into a logically correlated sequence, laying a foundation for accurate detection;

[0142] The implementation process of obtaining the risk prediction result is as follows:

[0143] Input the action risk point data to form a time sequence of action risks in chronological order, and construct a space-time sequence structure of risk points by combining the spatial distribution of body parts;

[0144] Traverse the rules in the space-time correlation rule library to perform chain matching on the space-time sequence structure:

[0145] For time sequence correlation rules: allow a time deviation of ±1 second, and match if the "main time sequence order" is met;

[0146] For concurrent correlation rules: judge whether the time overlap of multiple risk points is ≥50%, and consider it as matching;

[0147] For random correlation rules: only verify whether the spatial coordination deviation meets the threshold, without limiting the time order;

[0148] Obtain the risk prediction result through the above matching.

[0149] Finally, by traversing the rule base and matching the risk point sequence chain, the risk point and the rule are comprehensively checked from the time and space dimensions, the time and space logic of sports injury occurrence is simulated, potential risks can be effectively identified, and risk early warning can be performed in advance.

[0150] Embodiment two

[0151] As Figure 2 shown, the application provides a risk dynamic early warning based on student sports data, which comprises:

[0152] The data acquisition module acquires the original sports data of the students by wearing a sensing device;

[0153] The action mapping module obtains the mapping action data by an improved proximal deep Q network based on the original action data and the normal sports data;

[0154] The risk acquisition module acquires the action risk points by threshold comparison and considering the duration according to the mapping relationship data;

[0155] The prediction output module matches and detects the acquired action risk points by an abnormal action chain detection mechanism based on the space-time correlation of sports, and obtains the risk prediction result.

[0156] In the application, several formulas involved are dimensionless values, and the establishment of the formula is obtained by software simulation of a formula closest to the real situation by collecting a large amount of data, and part of the coefficients or weights in the formula are set by the person skilled in the art according to the actual situation, so no more will be described here.

[0157] The above embodiments can be realized by software, hardware, firmware or any combination thereof, in whole or in part. When realized by software, the above embodiments can be realized in the form of a computer program product in whole or in part. Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution.

[0158] Although the embodiments of the application have been shown and described, those skilled in the art can understand that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the application, and the scope of the application is defined by the appended claims and their equivalents.

Claims

1. A risk dynamic early warning method based on student motion data, characterized in that, The method comprises the following steps: S1: collecting original motion data of students' motion through a wearable sensing device; S2: obtaining mapping motion data through an improved proximal deep Q network based on original motion data and normal motion data; The improved proximal deep Q network combines the policy optimization mechanism of PPO and the value learning of deep Q network, and uses a clipped target function to constrain the PPO policy update step length; S3: obtaining motion risk points through threshold comparison and considering the duration according to the mapping relationship data; S4: obtaining risk prediction results through matching detection of the obtained motion risk points based on the abnormal motion chain detection mechanism of the motion space-time correlation.

2. The risk dynamic early warning method based on student motion data according to claim 1, characterized in that, The wearable sensing device is a smart motion bracelet with a built-in three-axis accelerometer, gyroscope and magnetometer; The original motion data includes three-dimensional linear acceleration data and three-dimensional angular velocity data.

3. The risk dynamic early warning method based on student motion data according to claim 2, characterized in that, The experience replay pool construction process of the improved proximal deep Q network is as follows: Collect normal motion data, and collect three-dimensional linear acceleration data and three-dimensional angular velocity data under normal motion state through the wearable sensing device to form normal motion data; Set an experience replay pool R, and the experience replay pool samples include original motion data and normal motion data; Divide the original motion data and normal motion data into windows of 1 second and steps of 0.5 second at the same time; Then, the experience replay pool samples are managed through a priority playback mechanism to complete the construction.

4. The risk dynamic early warning method based on student motion data according to claim 3, characterized in that, The training process of the improved proximal deep Q network is as follows: From the experience replay pool R, extract original motion data and normal motion data according to a proportion of 60% of normal motion data samples and 40% of preprocessed original motion data samples, and construct a mixed training set; Extract motion data features of the mixed training set samples through a convolutional neural network, and map them to an action value space through a fully connected layer; Initialize the DQN and PPO policy networks, and obtain the policy optimization loss L of PPO CLIP and the value learning loss L of DQN TD , the policy optimization loss of PPO and the value learning loss of DQN are fused based on the composite loss weights λ1, λ2 to obtain the composite loss L total , and the formula is represented as: L total = λ1L TD + λ2L CLIP And the Adam optimizer is used to update the network parameters, and the composite loss L is calculated total The gradient of the network parameters θ Then update the network parameters until the loss function L total Tends to be stable, complete training.

5. The risk dynamic early warning method based on student motion data according to claim 4, characterized in that, The process of using a clipped target function to constrain the PPO policy update step length is as follows: Collect sample data from the mixed training set, which includes motion state, action performed in the motion state, reward obtained after performing the action, and next motion state after performing the action; Input the motion state into the value network to obtain the value of performing the action in the current state, and input the next state into the target value network to obtain the target value; Input the motion state into the policy network to obtain the new policy probability of performing the action in the current state, and simultaneously, obtain the old policy probability of performing the action in the same state from the old policy network, divide the new policy probability by the old policy probability to obtain the policy ratio; Based on the value calculation result of DQN, obtain the advantage estimation by subtracting the current action value from the target value; The strategy ratio is clipped to limit the interval to [1-, 1+], and the advantage estimate is combined to obtain the policy optimization loss L of PPO CLIP Complete the PPO constraint policy update step.

6. The risk dynamic early warning method based on student motion data according to claim 5, characterized in that, The process of obtaining motion risk points is as follows: Collect standard motion parameters, and determine risk trigger thresholds according to motion injury case statistics; Extract the deviation type, deviation value and deviation duration of the current motion state from the mapping relationship data; When the deviation value is greater than the risk trigger threshold and the deviation duration is greater than the risk duration threshold, the corresponding action data is determined as a motion risk point.

7. The risk dynamic early warning method based on student motion data according to claim 1, characterized in that, The process of matching detection of the obtained motion risk points is as follows: The motion space-time correlation relationship is defined, including a time dimension and a space dimension, the time dimension covers the start time, end time and duration of the action risk point, and the space dimension includes the coordinate range and the coordinated motion trajectory involved in the action; Based on historical motion injury data, the motion space-time correlation relationship is obtained, including time sequence correlation rules, concurrent correlation rules and random correlation rules; The above rules are expressed as space-time correlation rules, and a space-time correlation rule library is constructed by storing all rules; The action risk point data is arranged in time sequence to form a time sequence of action risks, and the space-time sequence structure of the risk points is constructed by combining the spatial distribution of the body parts, and the rules in the space-time correlation rule library are traversed to chain match the space-time sequence structure to obtain a risk prediction result.

8. The risk dynamic early warning method based on student motion data according to claim 7, characterized in that, The chain matching process is as follows: From the time dimension, check whether the duration and occurrence order of the action risk point meet the time conditions in the space-time correlation rule library, and from the space dimension, verify whether the body parts and deviation value range involved meet the space conditions in the space-time correlation rule library; When the space-time sequence structure matches any space-time correlation rule in the rule library, the corresponding risk prediction result is output, including the action risk points appearing simultaneously in the time dimension and the space dimension, and if no rule is matched, it is marked as normal.

9. The risk dynamic early warning method based on student motion data according to claim 8, characterized in that, The process of obtaining the risk prediction result is as follows: The action risk point data is arranged in time sequence to form a time sequence of action risks, and the space-time sequence structure of the risk points is constructed by combining the spatial distribution of the body parts; Traverse the rules in the space-time correlation rule library, chain match the space-time sequence structure to obtain a risk prediction result.

10. A risk dynamic early warning based on student movement data, according to any one of claims 1-9, wherein, It includes: Data acquisition module: collect original motion data of students' motion through wearable sensing devices; Action mapping module: based on original action data and combined with normal motion data, an improved proximal deep Q network is used to obtain mapping action data; Risk acquisition module: according to the mapping relationship data, the action risk point is obtained by threshold comparison and considering the duration; Prediction output module: through the abnormal action chain detection mechanism based on motion space-time correlation, the obtained action risk point is matched and detected to obtain a risk prediction result.

Images