AI identification physical exercise trajectory data analysis method and system

Abstract

The invention discloses an analysis method and system for AI recognition of physical exercise trajectory data. The method comprises the following steps: step 1, constructing a space-time-environment-biological feature fusion network, integrating collected multi-source data, capturing association of trajectory coordinates, skeleton joint angles and electromyographic signals, constructing a space-time dynamic graph, and comprehensively sensing the influence of a dynamic environment on a motion trajectory; 2, introducing a generative adversarial network to simulate extreme environment interference, and improving model robustness; forcing the model to learn a defensive decision through physical layer disturbance and behavior layer disturbance; distinguishing a real track from a generated track by using a discriminator, and trying to cheat the discriminator by a generator to enable the model to adapt to a noise environment; 3, cross-athlete skill migration is conducted, a parameter freezing strategy and dynamic weight adjustment are adopted, the basic exercise mode recognition capacity is reserved, and meanwhile individual differences are adapted; and step 4, in combination with deep reinforcement learning and an interpretable rule engine, carrying out instant tactical adjustment, and generating an action through a DRL strategy network.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of motion trajectory data analysis, in particular to an AI-identified sports motion trajectory data analysis method and system. BACKGROUND

[0002] Modern sports have entered a new stage of high intensity, fast pace and multi-variable interaction. The real-time position, motion trajectory and ball movement data of athletes become the core basis for tactical analysis, training optimization and injury prevention. Through trajectory data analysis, the coaching team can accurately identify technical weaknesses, predict opponent strategies, and optimize physical distribution.

[0003] However, the traditional data analysis has the problem of insufficient adaptability to dynamic environment. For example, in a football match, the trajectory analysis system originally planned based on dry ground training is used for real-time tactical guidance. When the match is 30 minutes, it suddenly rains heavily, and the grass on the field is seriously waterlogged. The rolling speed and rebound height of the ball are significantly different from those on the dry ground. The ball speed on the dry ground is 15 m / s, and the wet and slippery ground is reduced to 10 m / s. At this time, the forward tries to break through the opponent's defense line through short ground passes, but the traditional trajectory model still predicts the ball passing path based on dry ground data, without considering the speed attenuation and trajectory deviation caused by wetness, which causes a large deviation in motion trajectory data analysis, seriously affecting the normal analysis arrangement. Therefore, an AI-identified sports motion trajectory data analysis method and system are proposed. SUMMARY

[0004] The purpose of the application is to solve the problems in the prior art and provide an AI-identified sports motion trajectory data analysis method and system.

[0005] In order to achieve the above purpose, the application adopts the following technical scheme:

[0006] An AI-identified sports motion trajectory data analysis method, comprising:

[0007] Step 1: Build a space-time-environment-biological feature fusion network, integrate GPS / IMU trajectory, camera motion capture (such as skeletal point data), field three-dimensional coordinates, real-time weather / lighting conditions, and multi-source data such as electromyographic signals and heart rate variability collected by wearable devices; perform feature-level fusion through a cross-modal Transformer encoder to capture the cross-modal correlation between trajectory coordinates, skeletal joint angles and electromyographic signals, such as the corresponding relationship between sudden stop actions and quadriceps muscle electromyographic bursts, and build a space-time dynamic graph to model the real-time interaction between athletes, balls and opponents (the "triangle position" in the football passing path), so as to comprehensively perceive the influence of dynamic environment on motion trajectory;

[0008] Step two: Introduce a generative adversarial network to simulate extreme environmental disturbances and improve model robustness; through physical layer disturbances such as Gaussian noise for GPS signal loss, affine transformation distortion of the field coordinate system, and behavior layer disturbances such as "adversarial trajectories" generated by LSTM, force the model to learn defensive decisions; use the discriminator to distinguish between real and generated trajectories, and the generator to try to deceive the discriminator, ultimately adapting the model to a noisy environment;

[0009] Step three: Develop a meta-learning framework for cross-athlete skill transfer; through meta-training on multiple athlete datasets, learn the basic ability to "quickly adapt to new individuals"; use parameter freezing strategy (only adjust the last two fully connected layers) and dynamic weight adjustment (adjust attention weight according to athlete fitness test data such as VO2Max), retain the basic ability of motion pattern recognition while quickly adapt to individual differences, such as the difference in trajectory features between sprinters and long-distance runners;

[0010] Step four: Combine deep reinforcement learning with an interpretable rule engine for real-time tactical adjustments; define state space (athlete position, speed, fatigue index, ball status), action space (discrete actions such as change of direction, long pass, continuous actions such as pass strength, shooting angle) and reward function (immediate rewards such as successful dribbling, being tackled, delayed rewards such as scoring, running out of energy); generate actions through the DRL policy network and combine them with the rule engine, such as "prefer to pass when the opponent is less than 1m away", to make the strategy comply with the principles of kinematics.

[0011] The above technical solutions further include:

[0012] Further, the cross-modal correlation of the captured trajectory coordinates, skeletal joint angles, and electromyographic signals includes the following steps:

[0013] Data acquisition:

[0014] Collect spatiotemporal coordinates (x, y, z, t) of athletes through GPS and IMU sensors, and extract skeletal joint coordinates such as elbow and knee angles through motion capture with cameras; collect electromyographic signals through wearable devices such as electromyographic sensors, and simultaneously record heart rate variability and fatigue index; use frame synchronization algorithm (PTP clock protocol) to synchronize the timestamps of trajectory, skeletal, and electromyographic signals;

[0015] Cross-modal feature encoding:

[0016] Use 1D-CNN to extract spatiotemporal features of the trajectory, such as speed, acceleration, and direction change rate; model the topological relationship of skeletal joints through graph neural networks, such as "right knee bending angle > 30°" associated with running motion; use LSTM to analyze EMG time series patterns (quadriceps muscle EMG burst corresponds to emergency stop action);

[0017] Cross-modal fusion:

[0018] By using a cross-modal Transformer encoder, trajectory features, skeletal features, and electromyographic features are mapped to a unified latent space, and attention weights for each modality are calculated. For example, the weight of electromyographic signals in the "sudden stop action" is increased by 40%. High-frequency cross-modal associations are extracted based on the Apriori algorithm. For example, the defensive action is "trajectory sudden stop + knee flexion angle greater than 30° and quadriceps EMG burst".

[0019] Furthermore, the construction of the spatiotemporal dynamic graph model to depict the real-time interactive relationships between the athlete, the ball, and the opponent includes the following steps:

[0020] Graph structure definition:

[0021] Players (own / opponent) and the ball; each node contains state features (such as position, speed, fatigue index); edge weights are calculated based on real-time distance, speed difference, and line-of-sight angle.

[0022]

[0023] Where K1 is the edge weight of the player and the ball, and L is the distance;

[0024] K2=v1×cosα

[0025] Where K2 is the side weight of the opponent player and the player on our side, v1 is the speed difference, and cosα is the cosine of the line of sight angle.

[0026] Updated animated GIF:

[0027] The graph structure is updated every 100ms, and the edge weights are dynamically calculated based on real-time data. For example, when the opponent is less than 1m away, the edge weight is increased by 30%, triggering a "defensive decision". When the ball enters the penalty area, the edge weights are redistributed, prioritizing the connection between the forward and the ball.

[0028] Graph Neural Network Inference:

[0029] The GAT graph attention network aggregates information from neighboring nodes. For example, the friendly player node aggregates the position and speed information of the opponent players to predict the interception path; the ball node aggregates the position information of all players to predict the passing / shooting direction; and based on graph embedding vectors, an MLP classifier outputs action labels, such as "change of direction" and "long pass".

[0030] Furthermore, the introduction of generative adversarial networks to simulate extreme environmental interference and improve model robustness includes the following steps:

[0031] Adopt LSTM network, input random noise, output simulated trajectory coordinates; Through nn.LSTM and nn.Linear layers, generate adversarial trajectories, such as sudden changes in direction or high-speed sprinting abnormal paths; Adopt Conv1d and fully connected layers, input trajectory coordinates, output the probability of real trajectory; Through convolutional layers, extract the spatio-temporal features of the trajectory to judge its authenticity; The generator generates simulated trajectories, and the discriminator calculates the true probability; Through the binary cross-entropy loss function, update the parameters of the generator and the discriminator, so that the generator gradually generates more realistic adversarial samples; Finally, the discriminator cannot distinguish between real trajectories and generated trajectories, and the generator has the ability to simulate extreme environmental disturbances; Through repeated training, the accuracy of the discriminator gradually decreases, and the loss of the generator decreases, proving that the GAN successfully simulates extreme environmental disturbances, such as signal loss and opponent attacks.

[0032] Further, the model is forced to learn defensive decisions through physical layer disturbance and behavior layer disturbance, including the following steps:

[0033] Physical layer disturbance: By injecting noise and abnormal data, simulate sensor errors and environmental mutations in real games, force the model to adapt from "ideal data" to "noisy data"; Add Gaussian noise to the trajectory data to simulate GPS signal loss or sensor accuracy decline; For example, randomly offset the real-time position data of the player by 1-3 meters, forcing the model to still infer reasonable actions when the position is inaccurate; Modify the three-dimensional coordinates of the field (compress the football field half by 20%), simulate environmental changes such as wet and slippery field and blurred markings, and require the model to adapt to trajectory prediction under non-standard field conditions; Dynamically adjust weather parameters, increase trajectory slip coefficient in rainy weather, and change ball speed and direction in windy weather, requiring the model to maintain action reasonableness when the environment changes, for example, in a strong wind simulation scenario, the pass trajectory needs to automatically adjust the force and angle to offset the wind effect;

[0034] Behavior layer disturbance: By simulating the intelligent behavior of the opponent, such as sudden change of direction and encirclement, force the model to change from "passive prediction" to "active defense", and learn reasonable decisions under high pressure; Use LSTM network to analyze the historical trajectory data of the opponent player, generate a simulated path with aggressive behavior; For example, according to the sprinting habits of the opponent forward, generate a "sudden acceleration and change of direction" trajectory, forcing the model to learn how to predict in advance and adjust the defense strategy; Combine the current game state such as score, time, and player fatigue, dynamically adjust the opponent's strategy; For example, in the last stage of the game when the fatigue index of the home player exceeds 0.8, generate a "opponent encirclement" simulation scenario, requiring the model to still be able to escape the defense through passing or changing direction when the physical strength decreases;

[0035] Physical and behavioral layer disturbance synergy: Inject physical layer noise (such as GPS signal offset) and behavioral layer disturbance (adversary raid) simultaneously in the training scene; the model needs to generate action strategy under double disturbance, such as judging whether the adversary is really close when the signal is offset, or choosing safe pass instead of risky shot in high wind speed environment; through the reward function (such as "successfully escape from the adversary" +0.5 points, "invalid shot" -1 point) to optimize the strategy, so that the model gradually learns to prefer low-risk, high-yield actions in noisy environments.

[0036] Further, the parameter freezing strategy and dynamic weight adjustment are adopted, including the following steps:

[0037] Parameter freezing: preserve the basic motion pattern recognition ability, reduce the calculation resource consumption in cross-individual adaptation; divide the basic model in the meta-learning framework (such as cross-modal Transformer or spatio-temporal dynamic graph model) into bottom feature extraction module and top decision module; the bottom module includes trajectory encoding, skeletal joint analysis, and electromyographic signal processing general feature extraction layer; the top module includes cross-modal fusion, dynamic graph reasoning, and action classification task related layer (such as fully connected layer, attention mechanism); freeze the bottom module, fix the parameters of the bottom module such as CNN convolution kernel weight and LSTM hidden layer parameter when migrating across athletes, only open the parameters of the top module for fine-tuning; the bottom module extracts general kinematic features such as speed, acceleration, and joint angle, which have commonality among different athletes; the top module needs to adjust the decision logic according to individual differences (short-distance runner vs. long-distance runner's trajectory preference);

[0038] Dynamic weight adjustment: dynamically adjust the model's attention to different modalities or features according to the individual characteristics of the athletes such as physical data and technical style; obtain the individualized data of the athletes through wearable devices or physical tests, for example: physiological data: maximum oxygen uptake, muscle fiber type ratio (fast muscle / slow muscle); technical data: turning frequency in historical trajectory, pass success rate, shooting angle preference; manually set the initial weights according to the athlete type (such as short-distance defender vs. long-distance midfielder), for example, the trajectory speed feature weight of short-distance athletes is increased by 20%, and the physical consumption feature weight of long-distance athletes is increased by 15%; establish a mapping relationship between individualized data and model weights through a regression model (such as linear regression, decision tree); for example, for every 1 unit increase in VO2Max value, the trajectory persistence feature weight is increased by 5%; dynamically adjust the weight according to the real-time state of the athlete (such as fatigue index) in the game or training, for example, when the electromyographic signal shows that the quadriceps muscle is fatigued, the trajectory turning ability feature weight is reduced by 10% to avoid over-reliance on actions of fatigued muscle groups;

[0039] Collaborative work: freezing the underlying module ensures that the model retains the general motor pattern recognition ability, such as the biomechanical characteristics of the emergency stop action, dynamically adjusting the weights of the top-level module to make the model focus on the most relevant features of the current player (such as the explosive force index of a sprinter); only need to fine-tune the top-level module, usually accounting for 10%-20% of the total parameters of the model, to complete cross-individual migration, and the adaptation time is shortened from hours to minutes; the model can generate more "straight sprint breakthrough" strategies for sprint-type guards and more "ball control and energy consumption" strategies for long-distance midfielders.

[0040] Further, the combination of deep reinforcement learning and explainable rule engine for real-time tactical adjustment; define state space, action space and reward function, including the following steps:

[0041] Define space:

[0042] State space: player state: real-time position (x, y, z), speed, acceleration, fatigue index (calculated by integrating the value of electromyographic signal EMG); ball state: position, rotation speed, relative distance from the player; environmental state: field characteristics (such as football field restricted area, basketball court three-point line), opponent position, real-time score; every 100ms refresh by sensor and camera data, synchronized with real game;

[0043] Action space: discrete actions: pre-defined tactical actions such as "change direction and break through", "long pass", "shoot", "man-to-man defense"; continuous actions: quantifiable parameters such as pass strength, shooting angle, defense stance adjustment (forward and backward movement distance); through expert interviews and game video analysis, the action space covers more than 90% of real tactical scenarios;

[0044] Reward function:

[0045] Instant reward: positive reward: successful dribble (+0.5 points), accurate pass (+0.3 points), goal (+10 points); negative reward: being broken (-1 point), offside (-2 points), physical exhaustion (-5 points);

[0046] Delayed reward: tactical coherence: trigger additional reward (+2 points) after 3 consecutive successful passes; strategic goal: keep the possession rate above 60% and reward +5 points per hour; punishment mechanism: violate sports ethics actions (such as malicious foul) directly end the round and deduct points (-20 points);

[0047] Deep reinforcement learning strategy network training:

[0048] Network architecture: actor network, input state space, output action probability distribution (e.g. Softmax layer handles discrete actions); critic network, input state space, output state value function (estimates long-term expected return); uses LSTM to handle temporal state (player speed changes), CNN to handle spatial state, e.g. ball position distribution;

[0049] Generate a large number of state-action pairs by simulating game environments, covering normal games, overtime, penalty shootouts, etc.

[0050] Policy iteration: use the proximal policy optimization algorithm, update network parameters every round, set a high exploration rate (ε = 0.3) in the initial stage, gradually decay to 0.05 to avoid premature convergence to a suboptimal strategy;

[0051] Integrate an interpretable rule engine:

[0052] Based on coach experience and kinematics principles, such as "prefer to pass when opponent distance is less than 1m" and "recommend shooting when shooting angle is greater than 30° in the penalty area", establish tactical rules; prevent unreasonable actions, such as "prevent up-charging when defensive player distance to own goal is less than 5m", establish safety rules; adjust according to game progress, such as "allow higher risk actions (such as long shots) when trailing", establish dynamic rules;

[0053] The candidate actions generated by DRL are first checked by the rule engine, and actions that violate tactical or safety rules are filtered out; for legal actions, parameters are adjusted, such as "shooting strength 100%" is corrected to "80%" (reserve strength); at the same time, the action is output with the rule basis, such as "recommend short pass because opponent clamping angle > 120°";

[0054] Real-time tactical adjustment:

[0055] Collect player position, ball state and environmental feature data in real time through sensors and cameras, dynamically update state space parameters (refresh every 100ms); the deep reinforcement learning strategy network generates candidate actions and corresponding expected return values based on the current state, then the interpretable rule engine filters and corrects the actions to make the strategy comply with the preset tactical logic (prefer short pass when opponent clamping) and safety constraints (such as avoid high-risk shooting in the penalty area); the verified action instructions are executed to the players or intelligent training equipment, according to the action results, such as whether to score or be tackled, the reward function parameters are optimized in reverse, forming a closed-loop learning mechanism.

[0056] An AI-identified sports trajectory data analysis system based on an AI-identified sports trajectory data analysis method, comprising:

[0057] Multi-modal data collection: Real-time trajectory and kinematics data are captured by GPS and IMU sensors worn by athletes, 4K camera arrays extract skeletal joint coordinates based on OpenPose algorithm, and miniature weather stations record site temperature, humidity, and wind speed. Wearable armband electromyography captures muscle activity signals; edge nodes clean and fuse raw data, including Butterworth low-pass filtering of IMU data, wavelet denoising of electromyography signals, and normalization of environmental data, to generate structured data packets; data packets are input into a cross-modal Transformer encoder, which captures the spatiotemporal correlation between trajectory coordinates, skeletal angles, and electromyography signals (such as sudden stop actions corresponding to quadriceps muscle electromyography bursts) to generate a fusion feature vector; based on the fusion features, a spatiotemporal dynamic graph is constructed, with athletes, balls, and opponents as nodes, and edge weights dynamically calculated from real-time distance, speed difference, and line-of-sight angle, capturing team tactical interactions (such as football "triangle position" passing paths);

[0058] Dynamic environment simulation: GAN network is used to generate adversarial trajectories, generator generates virtual paths for opponents "sudden change of direction interception" through LSTM, discriminator distinguishes between real and generated trajectories, and through game training, improves the realism of generated trajectories; add Gaussian noise to simulate GPS signal loss in original data, or distort the field coordinate system through affine transformation (such as compressing the basketball court half by 20%), forcing the model to adapt to non-ideal environments; mix the adversarial trajectories generated by the generator with the real data and input them into the model, punish the model's misjudgment of abnormal trajectories through loss function, force the model to learn defensive decisions, such as predicting the opponent's interception path in advance in ice hockey games;

[0059] Meta-learning adaptation: Train the MAML framework on datasets containing different projects and roles (forward, defender) athletes, learn the basic ability of "how to quickly adapt to new individuals"; deploy only the last two layers of fully connected layers, freeze the bottom feature extraction layer to retain the basic movement pattern recognition ability, and fine-tune the top layer parameters through a small amount of target athlete data (such as 10 minutes of running trajectory); according to the athlete's physical test data, including VO2Max, adjust the model's attention weight, and strengthen the feature dimensions related to individual characteristics (such as the step frequency stability of long-distance runners);

[0060] Real-time decision: Real-time receive fused multi-modal data, construct state vector including player position, speed, fatigue index (EMG integral value), ball position, rotation speed, opponent dynamic graph node state; DRL strategy network based on PPO algorithm outputs candidate action (including "change direction breakthrough" or "long pass") according to state vector, action space contains discrete action and continuous action (intensity, angle parameter); Explainable rule engine (such as "when the opponent distance is less than 1m, prefer to pass the ball") carries out legality check on candidate action, filters the action that does not conform to kinematics principle, such as offside path in football match; Combined with immediate reward (successfully overtake +0.5 points, broken ball-1 point) and delayed reward (goal +10 points, physical exhaustion-5 points), long-term income is optimized through Monte Carlo tree search;

[0061] Blockchain storage: The original data is encrypted according to the CP-ABE attribute encryption scheme, cut into k fragments and applied Reed-Solomon encoding to generate r redundant fragments, improve data privacy and fault tolerance; The encrypted fragments are stored in IPFS, only the file hash value is stored on the chain, the blockchain contract is associated with a trusted timestamp, forming an isolated architecture of "on-chain metadata storage + off-chain ciphertext storage"; Supervision nodes trigger CP-ABE decryption through attribute matching, and combine ZK-SNARK zero-knowledge proof verification to verify the compliance of the detection report, so that the audit process does not expose commercial secrets.

[0062] User layer interaction: Receive the tactical suggestions generated by the DRL engine through the Web front end, such as "prefer short pass breakthrough", combine GIS map to show the real-time position and compliance state (color marking) of the players, and risk data automatically triggers emergency plan pop-up window; Manage the confrontation sample library, label the difficulty level of the trajectory ("primary change direction" to "advanced fake action"), and monitor the prediction accuracy rate change of the model in the noise environment through the dashboard; Mobile terminal shows personalized training suggestions (including "strengthening the quadriceps muscle strength"), combined with the historical trajectory playback function, supports filtering key actions according to time range, for example "three times of sudden stop in last week's match".

[0063] The present application has the following beneficial effects:

[0064] In the present application, a space-time-environment-biological feature fusion network is constructed, multi-source data is integrated, and cross-modal Transformer is used to capture the cross-modal correlation of trajectory coordinates, skeletal joint angles and electromyographic signals. The interaction relationship among athletes, balls and opponents is modeled in real time by combining a space-time dynamic graph, and the influence of the dynamic environment on the trajectory is comprehensively perceived. A generative adversarial network is introduced to simulate extreme environmental disturbances, and through physical layer disturbance and behavior layer disturbance, the model is forced to learn defensive decisions and adapt to noisy environments. A meta-learning framework is developed to quickly adapt to the basic abilities of new individuals through meta-training of multi-athlete datasets, combined with a parameter freezing strategy to retain basic motion pattern recognition capabilities, and dynamically adjusting the weights to adapt to individual differences. Deep reinforcement learning and an interpretable rule engine are combined to define the state space, action space and reward function, generate actions through DRL and verify them through the rule engine to ensure that the strategy conforms to the kinematic principle, and realize real-time tactical adjustment. From environmental perception, noise adaptation, individual adaptation to real-time decision-making, the problem of insufficient dynamic environment adaptability in traditional data analysis is effectively solved. BRIEF DESCRIPTION OF DRAWINGS

[0065] Fig. 1 A method step diagram of an AI identification sports movement trajectory data analysis method is provided for the present application.

[0066] Fig. 2 A system block diagram of an AI identification sports movement trajectory data analysis system is provided for the present application. DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0068] Please refer to Figs. 1-2 The present application is an AI identification sports movement trajectory data analysis method, which includes:

[0069] Step one: build a spatio-temporal-environment-biological feature fusion network, integrate GPS / IMU trajectory, camera motion capture (such as skeletal point data), site three-dimensional coordinates, real-time weather / lighting conditions, and multi-source data such as electromyographic signals and heart rate variability collected by wearable devices; through cross-modal Transformer encoder for feature-level fusion, capture the cross-modal correlation of trajectory coordinates, skeletal joint angles, and electromyographic signals, such as the corresponding relationship between sudden stop action and quadriceps muscle electromyographic burst, and build a spatio-temporal dynamic graph to model the real-time interaction of athletes, balls, and opponents (the "triangle stance" in the football passing path), so as to comprehensively perceive the influence of dynamic environment on movement trajectory;

[0070] Step two: introduce a generative adversarial network to simulate extreme environmental disturbances and improve model robustness; through physical layer disturbances such as Gaussian noise of GPS signal loss, affine transformation distortion of site coordinate system, and behavior layer disturbances such as "adversarial trajectory" generated by LSTM, force the model to learn defensive decisions; use the discriminator to distinguish between real and generated trajectories, and the generator tries to deceive the discriminator, ultimately making the model adapt to noisy environments;

[0071] Step three: develop a meta-learning framework for cross-athlete skill transfer; through meta-training on multi-athlete datasets, learn the basic ability of "how to quickly adapt to new individuals"; use parameter freezing strategy (only adjust the last two fully connected layers) and dynamic weight adjustment (adjust attention weight according to athlete physical test data such as VO2Max), retain the basic movement pattern recognition ability while quickly adapt to individual differences, such as the trajectory feature differences between sprinters and long-distance runners;

[0072] Step four: combine deep reinforcement learning with an interpretable rule engine for real-time tactical adjustment; define state space (athlete position, speed, fatigue index, ball status), action space (discrete actions such as change direction, long pass, continuous actions such as pass strength, shooting angle), and reward function (immediate rewards such as successful dribbling, being tackled, delayed rewards such as scoring, physical exhaustion); generate actions through the DRL policy network and combine with the rule engine, such as "prefer to pass when opponent distance is less than 1m", to make the strategy comply with kinematics principles.

[0073] In one embodiment, the cross-modal correlation of trajectory coordinates, skeletal joint angles, and electromyographic signals includes the following steps:

[0074] Data collection:

[0075] Collect the spatiotemporal coordinates (x, y, z, t) of the athletes through GPS and IMU sensors, and extract the coordinates of the skeletal joint points such as the elbow and knee angles through camera motion capture; collect the electromyographic signals through wearable devices such as electromyographic sensors, and synchronously record the heart rate variability and fatigue index; use a frame synchronization algorithm (PTP clock protocol) to make the timestamps of the trajectory, skeleton, and electromyographic signals consistent;

[0076] Cross-modal feature encoding:

[0077] Use 1D-CNN to extract the spatiotemporal features of the trajectory, such as speed, acceleration, and direction change rate; model the topological relationship of the skeletal joint points through a graph neural network, such as "right knee bending angle > 30°" associated with the running action; use LSTM to analyze the EMG time series pattern (quadriceps muscle EMG burst corresponding to emergency stop action);

[0078] Cross-modal fusion:

[0079] Map the trajectory features, skeletal features, and electromyographic features to a unified latent space through a cross-modal Transformer encoder, calculate the attention weights of each modality, such as the muscle signal weight in the "emergency stop action" increasing by 40%; extract high-frequency cross-modal associations based on the Apriori algorithm, for example: "trajectory emergency stop + knee bending angle greater than 30° and quadriceps muscle EMG burst" corresponds to the defense action.

[0080] In one embodiment, the construction of the spatiotemporal dynamic graph models the real-time interaction relationship of the athletes, the ball, and the opponent, including the following steps:

[0081] Graph structure definition:

[0082] Athletes (self / opponent), ball, each node contains state features (such as position, speed, fatigue index); calculate the edge weight based on real-time distance, speed difference, and line of sight angle:

[0083]

[0084] Wherein, K1 is the edge weight between the self player and the ball, and L is the distance;

[0085] K2 = v1 x cos a

[0086] Wherein, K2 is the edge weight between the opponent player and the self player, v1 is the speed difference, and cos a is the cosine value of the line of sight angle;

[0087] Update the dynamic graph:

[0088] Update the graph structure every 100ms, dynamically calculate the edge weight according to real-time data, for example: when the opponent distance is less than 1m, the edge weight is increased by 30%, triggering "defensive decision"; when the ball enters the penalty area, the edge weight is redistributed, and the forward is connected with the ball;

[0089] Graph neural network inference:

[0090] Aggregate neighbor node information through GAT graph attention network, for example, aggregate opponent player position and speed information for own player node to predict interception path; aggregate all player position information for ball node to predict pass / shoot direction; based on graph embedding vector, output action label through MLP classifier, such as "breakthrough" and "long pass".

[0091] In one embodiment, the introduction of the generative adversarial network simulates extreme environmental interference and improves model robustness, including the following steps:

[0092] Use LSTM network, input random noise, output simulated trajectory coordinates; generate adversarial trajectories through nn.LSTM and nn.Linear layers, such as sudden changes in direction or high-speed sprinting abnormal paths; use Conv1d and fully connected layers, input trajectory coordinates, output the probability of real trajectory; extract the spatiotemporal features of the trajectory through the convolution layer to judge its authenticity; the generator generates simulated trajectories, and the discriminator calculates the real probability; update the parameters of the generator and the discriminator through the binary cross-entropy loss function, so that the generator gradually generates more realistic adversarial samples; finally, the discriminator cannot distinguish between real and generated trajectories, and the generator has the ability to simulate extreme environmental interference; through repeated training, the accuracy of the discriminator gradually decreases, and the loss of the generator decreases, proving that the GAN successfully simulates extreme environmental interference, such as signal loss and opponent attack.

[0093] In one embodiment, the model is forced to learn defensive decisions by physical layer disturbance and behavior layer disturbance, including the following steps:

[0094] Physical layer disturbance: By injecting noise and abnormal data, simulate sensor errors and environmental mutations in real matches, force the model to adapt from "ideal data" to "noisy data"; Add Gaussian noise to the trajectory data, simulate the loss of GPS signal or the decrease of sensor accuracy; For example, randomly offset the real-time position data of the player by 1-3 meters, force the model to infer reasonable actions even when the position is inaccurate; Modify the three-dimensional coordinates of the field (compress the half court by 20%), simulate the change of environment such as wet and slippery field and blurred markers, require the model to adapt to the trajectory prediction under non-standard field conditions; Dynamically adjust the weather parameters, increase the trajectory slip coefficient in rainy days, change the speed and direction of the ball in windy days, require the model to maintain the rationality of the action when the environment changes, for example, in the strong wind simulation scene, the pass trajectory needs to automatically adjust the force and angle to offset the wind force;

[0095] Behavioral layer disturbance: By simulating the intelligent behavior of the opponent, such as sudden change of direction, encircle, forcing the model to change from "passive prediction" to "active defense", learning reasonable decision-making under high pressure; Use LSTM network to analyze the historical trajectory data of the opponent player, generate an aggressive simulation path; For example, according to the sprinting habit of the opponent forward, generate a "sudden acceleration and change of direction" trajectory, force the model to learn how to predict in advance and adjust the defense strategy; Combine the current game state such as score, time, and player fatigue, dynamically adjust the opponent's strategy; For example, in the last stage of the game when the fatigue index of the home player exceeds 0.8, generate a "opponent encircle" simulation scene, require the model to still be able to get rid of the defense through pass or change of direction when the physical strength decreases;

[0096] Physical and behavioral layer disturbance synergy: Inject physical layer noise (such as GPS signal offset) and behavioral layer disturbance (opponent attack) in the training scene at the same time; The model needs to generate action strategies under double disturbance, for example, judge whether the opponent is really close when the signal is offset, or choose safe pass instead of risky shooting in high wind speed environment; Through the reward function (such as "successfully get rid of the opponent" +0.5 points, "invalid shooting" -1 point) to optimize the strategy, make the model gradually learn to choose low-risk and high-yield actions in a noisy environment.

[0097] In one embodiment, the parameter freezing strategy and dynamic weight adjustment are adopted, including the following steps:

[0098] Parameter freezing: retain basic motion pattern recognition ability, reduce computational resource consumption when adapting across individuals; divide the base model in the meta-learning framework (such as cross-modal Transformer or spatio-temporal dynamic graph model) into a bottom feature extraction module and a top decision module; the bottom module contains trajectory encoding, skeletal joint analysis, and general feature extraction layers for electromyographic signal processing; the top module contains cross-modal fusion, dynamic graph reasoning, and action classification task-related layers (such as fully connected layers and attention mechanisms); freeze the bottom module, fix the parameters of the bottom module such as the convolution kernel weights of CNN and the hidden layer parameters of LSTM when migrating across athletes, and only open the parameters of the top module for fine-tuning; the bottom module extracts general kinematic features such as velocity, acceleration, and joint angles, which are common among different athletes; the top module needs to adjust the decision logic according to individual differences (trajectory preference of sprinters vs. long-distance runners);

[0099] Dynamic weight adjustment: dynamically adjust the model's attention to different modalities or features based on individual characteristics such as physical data and technical style; obtain personalized data of athletes through wearable devices or physical tests, such as: physiological data: maximum oxygen uptake, muscle fiber type ratio (fast / slow); technical data: turning frequency in historical trajectory, pass success rate, shooting angle preference; manually set initial weights according to athlete types (such as sprinter vs. long-distance runner), for example, increase the trajectory speed feature weight of sprinters by 20%, and increase the physical consumption feature weight of long-distance runners by 15%; establish a mapping relationship between personalized data and model weights through regression models (such as linear regression and decision trees); for example, for every 1 unit increase in VO2Max value, the trajectory persistence feature weight increases by 5%; dynamically adjust the weight according to the real-time state of the athlete (such as fatigue index) during the game or training, for example, when the electromyographic signal shows that the quadriceps muscle is fatigued, the trajectory turning ability feature weight is reduced by 10% to avoid relying too much on the actions of the fatigued muscle group;

[0100] Collaborative work: freezing the bottom module ensures that the model retains general motion pattern recognition ability, such as the biomechanical features of sudden stop actions, and dynamically adjusting the weights of the top module makes the model focus on the most relevant features of the current athlete (such as the explosive power indicator of sprinters); only the top module needs to be fine-tuned, which usually accounts for 10%-20% of the total model parameters, so cross-individual migration can be completed in minutes instead of hours as in traditional methods; the model can generate more "straight sprint breakthrough" strategies for sprinters and more "ball control and energy consumption" strategies for long-distance runners.

[0101] In one embodiment, the combination of deep reinforcement learning and explainable rule engine is used for real-time tactical adjustment; define the state space, action space, and reward function, including the following steps:

[0102] Definition space:

[0103] State space: Athlete state: Real-time position (x, y, z), velocity, acceleration, fatigue index (calculated by electromyographic signal EMG integral value); Ball state: Position, rotational speed, relative distance from the athlete; Environment state: Court features (such as football field penalty area, basketball court three-point line), opponent position, real-time score; Refreshed every 100ms through sensor and camera data, synchronized with real games;

[0104] Action space: Discrete actions: Predefined tactical actions such as "change direction and break through", "long pass", "shoot", "man-to-man defense"; Continuous actions: Quantifiable parameters such as pass strength, shooting angle, defensive stance adjustment (forward and backward movement distance); Through expert interviews and game video analysis, the action space covers more than 90% of real tactical scenarios;

[0105] Reward function:

[0106] Instant reward: Positive reward: Successful dribbling (+0.5 points), accurate passing (+0.3 points), goal scoring (+10 points); Negative reward: Being intercepted (-1 point), offside (-2 points), physical exhaustion (-5 points);

[0107] Delayed reward: Tactical coherence: Trigger additional reward (+2 points) after 3 consecutive successful passes; Strategic goal: Keep the possession rate above 60% and reward +5 points per hour; Punishment mechanism: Violate sports ethics actions (such as malicious fouls) directly end the round and deduct points (-20 points);

[0108] Deep reinforcement learning strategy network training:

[0109] Network architecture: Actor network, input state space, output action probability distribution (such as Softmax layer processing discrete actions); Critic network, input state space, output state value function (estimate long-term expected return); Use LSTM to process time-series state (athlete speed change), CNN to process spatial state such as ball position distribution;

[0110] Generate a large number of state-action pairs by simulating game environments, covering normal games, overtime, penalty shootouts, etc.

[0111] Policy iteration: Use the proximal policy optimization algorithm, update network parameters every round, set a higher exploration rate (ε = 0.3) in the initial stage, gradually decay to 0.05 to avoid premature convergence to a suboptimal strategy;

[0112] Integrate an interpretable rule engine:

[0113] Based on the coach's experience and kinematic principles, such as "prefer to pass when the opponent is less than 1m away" and "recommend shooting when the shooting angle is greater than 30° in the penalty area", tactical rules are established; to prevent unreasonable actions, such as "prohibit up-throw when the defender is less than 5m away from the goal", safety rules are established; to adjust according to the progress of the game, such as "allow higher-risk actions (such as long shots) when trailing", dynamic rules are established;

[0114] The candidate actions generated by DRL are first checked by the rule engine, and the actions that violate the tactical or safety rules are filtered out; for legal actions, parameters are adjusted, such as "shooting strength 100%" is corrected to "80%" (reserve physical strength); at the same time, the rule basis is output when the action is output, such as "recommend short pass because the opponent's clamping angle is >120°";

[0115] Instant tactical adjustment:

[0116] Through sensors and cameras, real-time collection of athlete position, ball state and environmental feature data, dynamic updating of state space parameters (refreshed every 100ms); the deep reinforcement learning strategy network generates candidate actions and corresponding expected return values based on the current state, and then the interpretable rule engine filters and corrects the actions, making the strategy comply with the preset tactical logic (prefer short pass when the opponent is clamped) and safety constraints (such as avoiding high-risk shooting in the penalty area); the verified action instructions are executed to the athletes or intelligent training equipment, and according to the action results, such as whether the ball is scored or broken, the reward function parameters are optimized in the opposite direction, forming a closed-loop learning mechanism.

[0117] An AI-identified sports movement trajectory data analysis system based on an AI-identified sports movement trajectory data analysis method, comprising:

[0118] Multi-modal data acquisition: real-time trajectory and kinematic data are captured through GPS and IMU sensors worn by athletes, 4K camera arrays extract bone joint coordinates based on the OpenPose algorithm, and miniature weather stations synchronously record site temperature, humidity, and wind speed. Wearable myoelectric arm rings collect muscle activity signals; edge nodes clean and fuse the original data, including Butterworth low-pass filtering of IMU data, wavelet denoising of myoelectric signals, and normalization processing of environmental data, to generate structured data packets; data packets are input into a cross-modal Transformer encoder, which captures the spatio-temporal correlation of trajectory coordinates, bone angles, and myoelectric signals (such as the quadriceps muscle myoelectric burst corresponding to the sudden stop action) through self-attention mechanisms, generating a fusion feature vector; a spatio-temporal dynamic graph is constructed based on the fusion features, treating athletes, balls, and opponents as nodes, and edge weights are dynamically calculated from real-time distance, speed difference, and line-of-sight angle, capturing team tactical interactions (such as the "triangle formation" pass path in football);

[0119] Dynamic environment simulation: Use GAN network to generate adversarial trajectories, the generator generates virtual paths for the opponent's "sudden change of direction interception" through LSTM, and the discriminator distinguishes between real and generated trajectories. Through game training, improve the authenticity of the generated trajectory; Add Gaussian noise to the original data to simulate GPS signal loss, or distort the court coordinate system through affine transformation (such as compressing the basketball court half by 20%), forcing the model to adapt to non-ideal environments; Mix the generated adversarial trajectories with real data and input them into the model, and punish the model for misjudging abnormal trajectories through the loss function, forcing the model to learn defensive decisions, such as predicting the opponent's interception path in advance in a hockey game;

[0120] Meta-learning adaptation: Train the MAML framework on a dataset containing different projects and role (forward, guard) athletes, learn the basic ability of "how to quickly adapt to new individuals"; When deployed, only unfreeze the last two layers of the model, freeze the bottom feature extraction layer to retain the basic movement pattern recognition ability, and fine-tune the top layer parameters through a small amount of target athlete data (such as 10 minutes of running trajectory); According to the athlete's physical test data, including VO2Max, adjust the model's attention weight, and strengthen the feature dimensions related to individual characteristics (such as the step frequency stability of long-distance runners);

[0121] Real-time decision-making: Real-time receive fused multi-modal data, construct state vector including athlete position, speed, fatigue index (myoelectric integral value), ball position, rotation speed, opponent dynamic graph node state; The DRL policy network based on PPO algorithm outputs candidate actions (including "change direction and break through" or "long pass") according to the state vector, the action space includes discrete actions and continuous actions (intensity, angle parameters); The explainable rule engine (such as "when the opponent is less than 1m away, prefer to pass the ball") checks the legality of the candidate actions, filters actions that do not conform to the kinematics principle, such as offside paths in soccer games; Combine immediate rewards (successfully breaking through +0.5 points, being blocked +1 point) and delayed rewards (scoring +10 points, exhausting physical energy -5 points), and optimize long-term returns through Monte Carlo tree search;

[0122] Blockchain storage: Encrypt the original data according to the CP-ABE attribute encryption scheme, cut it into k slices and apply Reed-Solomon encoding to generate r redundant slices, improve data privacy and fault tolerance; Store the encrypted slices in IPFS, only return the file hash value on the chain, and associate the blockchain contract with a trusted timestamp to form a "on-chain metadata storage + off-chain ciphertext storage" isolation architecture; Supervision nodes trigger CP-ABE decryption through attribute matching, and combine ZK-SNARK zero-knowledge proof verification to verify the compliance of the detection report, so that the audit process does not expose commercial secrets;

[0123] User layer interaction: Receive tactical suggestions generated by DRL engine through web frontend, such as "prefer short pass to break through", combine with GIS map to show real-time position and compliance status of players (color marking), risk data automatically trigger emergency plan pop-up window; Manage confrontation sample library, label difficulty level of generated trajectory ("primary change direction" to "advanced fake action"), monitor prediction accuracy rate change of model in noise environment through dashboard; Mobile end shows personalized training suggestions (including "strengthening the strength of the stock four heads"), combined with historical trajectory playback function, support filtering key actions by time range, such as "three times of sudden stop in last week's game".

[0124] While the embodiments of the application have been illustrated and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and alterations can be made therein without departing from the spirit and scope of the application, which is defined by the appended claims and their equivalents.

Claims

1. An AI-identified sports movement trajectory data analysis method, characterized in that, Comprise: Step one: build a space-time-environment-biological feature fusion network, integrate GPS, IMU trajectory, camera motion capture, site three-dimensional coordinates, real-time weather, lighting conditions and wearable devices collected myoelectric signal, heart rate variability and multi-source data; through cross-modal Transformer encoder for feature-level fusion, capture the cross-modal correlation of trajectory coordinates, skeletal joint angles and myoelectric signals, and build a space-time dynamic graph to model the real-time interaction of athletes, balls and opponents, and comprehensively perceive the influence of dynamic environment on movement trajectory; Step two: introduce a generative adversarial network to simulate extreme environmental disturbances and improve model robustness; through physical layer disturbance and behavior layer disturbance, force the model to learn defensive decision-making; use the discriminator to distinguish between real trajectories and generated trajectories, and the generator tries to deceive the discriminator, ultimately making the model adapt to noisy environments; Step three: develop a meta-learning framework for cross-athlete skill transfer; Through meta-training on a multi-athlete dataset, learn the basic ability of "how to quickly adapt to new individuals"; use parameter freezing strategy and dynamic weight adjustment to retain the basic movement pattern recognition ability while quickly adapting to individual differences; Step four: combine deep reinforcement learning with an interpretable rule engine for real-time tactical adjustments; define the state space, action space and reward function; generate actions through the DRL policy network and combine them with the rule engine to make the strategy comply with the kinematic principle. 2.The AI-identified sports trajectory data analysis method of claim 1, wherein, The cross-modal correlation of trajectory coordinates, skeletal joint angles and myoelectric signals includes the following steps: Data collection: Collect the spatiotemporal coordinates (x, y, z, t) of the athletes through GPS and IMU sensors, and extract the skeletal joint coordinates in combination with camera motion capture; collect myoelectric signals through wearable devices, and simultaneously record heart rate variability and fatigue index; use frame synchronization algorithm to synchronize the timestamps of trajectory, skeleton and myoelectric signals; Cross-modal feature encoding: Use 1D-CNN to extract the spatiotemporal features of trajectory including speed, acceleration and direction change rate; model the topological relationship of skeletal joints through graph neural network; use LSTM to analyze EMG time series pattern; Cross-modal fusion: Through cross-modal Transformer encoder, map trajectory features, skeletal features and myoelectric features to a unified latent space and calculate the attention weights of each modality; extract high-frequency cross-modal correlations based on Apriori algorithm. 3.The AI identification sports movement trajectory data analysis method of claim 1, wherein, The construction of a space-time dynamic graph to model the real-time interaction of athletes, balls and opponents includes the following steps: Graph structure definition: Define nodes including athletes and balls, each node contains state features; calculate edge weights based on real-time distance, speed difference and line of sight angle: Where K1 is the edge weight between the home player and the ball, and L is the distance; K2 = v1 x cosα Where K2 is the edge weight between the opponent player and the home player, v1 is the speed difference, and cosα is the line of sight angle cosine value; Update the graph structure every 100ms, dynamically calculate the edge weight based on real-time data; Graph neural network inference: The neighbor node information is aggregated by a graph attention network (GAT), the position and speed information of the opponent player is aggregated by the home player node, and the interception path is predicted; the position information of all players is aggregated by the ball node, and the pass and shot direction is predicted; and the action label is output by an MLP classifier based on the graph embedding vector. 4.The AI identification sports movement trajectory data analysis method of claim 1, wherein, The introduced generative adversarial network simulates extreme environmental interference and improves model robustness, including the following steps: An LSTM network is used to input random noise and output simulated trajectory coordinates; a generative adversarial trajectory is generated through nn.LSTM and nn.Linear layers, a Conv1d and fully connected layer is used to input trajectory coordinates and output the probability of a real trajectory; the authenticity is judged by extracting the spatiotemporal features of the trajectory through a convolutional layer; the generator generates a simulated trajectory, and the discriminator calculates the real probability; the parameters of the generator and the discriminator are updated through a binary cross-entropy loss function, so that the generator gradually generates more realistic adversarial samples; and finally the discriminator cannot distinguish between real and generated trajectories, and the generator has the ability to simulate extreme environmental interference. 5.The AI identification sports movement trajectory data analysis method of claim 1, wherein, The physical layer disturbance and the behavior layer disturbance force the model to learn defensive decisions, including the following steps: Physical layer disturbance: noise and abnormal data are injected to simulate sensor errors and environmental mutations in real games, forcing the model to adapt from "ideal data" to "noisy data"; Gaussian noise is added to the trajectory data to simulate GPS signal loss or sensor accuracy degradation; the three-dimensional coordinates of the field are modified to simulate environmental changes such as slippery fields and blurred markers, requiring the model to adapt to trajectory prediction under non-standard field conditions; weather parameters are dynamically adjusted, including increasing the trajectory slip coefficient in rainy weather and changing the speed direction of the ball in windy weather, requiring the model to maintain reasonable actions in the face of environmental mutations, and in the simulation of strong wind scenarios, the pass trajectory needs to automatically adjust the force and angle to offset the wind force; Behavior layer disturbance: simulate the intelligent behavior of the opponent, including sudden changes in direction and encirclement, to force the model to switch from "passive prediction" to "active defense" and learn reasonable decisions under high pressure; use an LSTM network to analyze the historical trajectory data of the opponent player to generate an aggressive simulated path; combine the current game state to dynamically adjust the opponent's strategy; Physical and behavioral layer disturbance synergy: physical layer noise and behavioral layer disturbance are injected simultaneously in the training scene; the model needs to generate action strategies under double disturbance, and the strategy is optimized through a reward function, so that the model gradually learns to prioritize low-risk, high-reward actions in a noisy environment. 6.The AI identification sports trajectory data analysis method of claim 1, wherein, The parameter freezing strategy and dynamic weight adjustment are used, including the following steps: Parameter freezing: retain the basic motion pattern recognition ability, reduce the consumption of computing resources when adapting across individuals; divide the base model in the meta-learning framework into a bottom feature extraction module and a top decision module; the bottom module includes trajectory encoding, skeletal joint analysis, and myoelectric signal processing general feature extraction layers; the top module includes cross-modal fusion, dynamic graph reasoning, and action classification task related layers; freeze the bottom module, fix the parameters of the bottom module when migrating across athletes, and only open the parameters of the top module for fine-tuning; the bottom module extracts general kinematic features, which are common among different athletes; the top module needs to adjust the decision logic according to individual differences; Dynamic weight adjustment: dynamically adjust the model's attention to different modalities and features according to the individual characteristics of athletes, including physical data and technical style; obtain personalized data of athletes through wearable devices or physical tests, including physiological data: maximum oxygen uptake, muscle fiber type ratio, technical data: turning frequency in historical trajectory, pass success rate, shooting angle preference; manually set the initial weight according to the athlete's type, and establish the mapping relationship between personalized data and model weight through a regression model; dynamically adjust the weight according to the real-time state of the athlete during the game or training, including but not limited to when the electromyographic signal shows that the quadriceps muscle is fatigued, the feature weight of trajectory turning ability is reduced to avoid excessive reliance on actions of fatigued muscle groups; Collaborative work: freezing the bottom module ensures that the model retains general motion pattern recognition ability, and dynamically adjusting the weight of the top module makes the model focus on the most relevant features of the current athlete; only the top module needs to be fine-tuned, which shortens the adaptation time; the model generates more "straight sprint break" strategies for sprint-type guards and more "ball control to consume energy" strategies for long-distance midfielders.

7. The AI identification sports movement trajectory data analysis method of claim 1, wherein, The combination of deep reinforcement learning and explainable rule engine is used for real-time tactical adjustment; the state space, action space, and reward function are defined, including the following steps: Define the space: State space: athlete state: real-time position (x, y, z), velocity, acceleration, fatigue index; ball state: position, rotational velocity, relative distance from the athlete; environmental state: field characteristics, opponent position, real-time score; refresh every 100ms through sensor and camera data, synchronized with real games; Action space: discrete actions: pre-defined tactical actions; continuous actions: quantifiable parameters, including pass strength, shooting angle, defensive stance adjustment; through expert interviews and game video analysis, the action space covers more than 90% of real tactical scenarios; Reward function: Instant reward: positive reward: including but not limited to successful dribble bonus, accurate pass bonus, goal bonus; negative reward: including but not limited to being broken up for points, offside for points, and energy depletion for points; Delayed rewards: tactical consistency: including but not limited to triggering additional bonus points after 3 consecutive successful passes; strategic goals: including but not limited to rewarding bonus points per hour when maintaining a possession rate of over 60%; punishment mechanism: including but not limited to directly terminating the round and deducting points for violating sports ethics actions; Deep reinforcement learning policy network training: Network architecture: actor network, input state space, output action probability distribution; critic network, input state space, output state value function; use LSTM to process time series state, CNN to process spatial state; Generate a large number of state-action pairs by simulating the game environment, covering normal games, overtime, and penalty shootout scenarios; Policy iteration: use the proximal policy optimization algorithm, update network parameters every round, set a high exploration rate in the initial stage, and gradually decay to avoid premature convergence to a suboptimal strategy; Integrate the explainable rule engine: Establish tactical rules based on coach experience and kinematics principles, establish safety rules to prevent unreasonable actions, and establish dynamic rules to adjust according to the progress of the game; The candidate actions generated by DRL are first checked by the rule engine, and actions that violate tactical or safety rules are filtered out; legal actions are adjusted in parameters, and the rules are output when the action is output; Real-time tactical adjustment: Collect athlete position, ball state and environmental feature data in real time through sensors and cameras, dynamically update state space parameters; deep reinforcement learning policy network generates candidate actions and corresponding expected return values based on the current state, then the explainable rule engine filters and corrects the actions to make the strategy comply with the preset tactical logic and safety constraints; the verified action instructions are executed to the athletes or intelligent training equipment, and the reward function parameters are optimized in reverse according to the action results, forming a closed-loop learning mechanism. 8.The AI-identified sports trajectory data analysis system of claim 1, wherein Including: Multi-modal data acquisition: real-time trajectory and kinematics data captured by GPS and IMU sensors worn by athletes, 4K camera array extracts bone joint coordinates based on OpenPose algorithm, micro weather station synchronously records field temperature, humidity, wind speed, wearable myoelectric arm ring collects muscle activity signals; edge nodes clean and fuse raw data, including Butterworth low-pass filtering of IMU data, wavelet denoising of myoelectric signals, and normalization processing of environmental data, to generate structured data packets; data packets input cross-modal Transformer encoder, capture the spatio-temporal correlation of trajectory coordinates, bone angles and myoelectric signals through self-attention mechanism, generate fusion feature vectors; Based on the fusion features, construct a spatio-temporal dynamic graph, treat athletes, balls, and opponents as nodes, and dynamically calculate edge weights from real-time distance, speed difference, and line of sight angle to capture team tactical interaction; Dynamic environment simulation: Use GAN network to generate adversarial trajectories, generator generates virtual paths for opponent actions through LSTM, discriminator distinguishes between real and generated trajectories, and through game training, the authenticity of the generated trajectory is improved; Add Gaussian noise to the original data to simulate GPS signal loss, or distort the field coordinate system through affine transformation, forcing the model to adapt to non-ideal environments; Mix the generated adversarial trajectories with real data and input them into the model, and use the loss function to punish the model for misjudging abnormal trajectories, forcing the model to learn defensive decisions; Meta-learning adaptation: Train the MAML framework on a dataset containing different projects and role players, learning the basic ability of "how to quickly adapt to new individuals"; Deploy only the last two fully connected layers of the model, freeze the bottom feature extraction layer to retain the basic motion pattern recognition ability, and fine-tune the top layer parameters with a small amount of target player data; Adjust the model's attention weight based on the player's physical test data to strengthen the feature dimensions related to individual characteristics; Real-time decision-making: Real-time reception of fused multi-modal data, construction of state vectors including player position, speed, fatigue index, ball position, rotation speed, and opponent dynamic graph node state; The DRL strategy network based on PPO algorithm outputs candidate actions according to the state vector, and the action space includes discrete actions and continuous actions; The explainable rule engine checks the legality of the candidate actions and filters out actions that do not conform to the kinematics principle; Combine immediate rewards with delayed rewards and optimize long-term returns through Monte Carlo tree search; Blockchain storage: Encrypt the original data according to the CP-ABE attribute encryption scheme, cut it into k shards and apply Reed-Solomon encoding to generate r redundant shards, improving data privacy and fault tolerance; Store the encrypted shards in IPFS and only return the file hash value on the chain, associate the blockchain contract with a trusted timestamp to form a "on-chain metadata storage + off-chain ciphertext storage" isolation architecture; Supervisory nodes trigger CP-ABE decryption through attribute matching, and combine ZK-SNARK zero-knowledge proof verification to detect the compliance of the detection report, making the audit process not expose commercial secrets; User-level interaction: Receive tactical recommendations generated by the DRL engine through the Web front end, display player real-time position and compliance status on the GIS map, and automatically trigger emergency plan pop-up windows for risk data; Manage the adversarial sample library, label the difficulty level of the generated trajectory, and monitor the prediction accuracy of the model in a noisy environment through the dashboard; Mobile app displays personalized training recommendations, combined with historical trajectory playback function, supports filtering key actions by time range.

Images