Jet pool sports training adjustment method, device, equipment and storage medium

Abstract

The present invention discloses a method, device, equipment and storage medium for adjusting sports training in a jet pool, the method comprising: obtaining a swimming posture image of a trainee in a jet pool captured by a camera, and obtaining the swimming speed and heart rate of the trainee detected by a wearable device; extracting swimming posture features from the swimming posture image to obtain swimming posture features, and extracting features from the swimming speed and heart rate to obtain swimming speed features and heart rate features; performing multimodal feature fusion on the swimming posture features, swimming speed features and heart rate features to obtain fusion features; analyzing the fusion features using a reward-based reinforcement learning algorithm to obtain a water jet mode and water jet speed currently suitable for the training of the trainee; and controlling the water jet operation of the jet pool according to the water jet mode and the water jet speed. The present invention can improve the training effect and training experience of the trainee in the jet pool.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a jet swimming pool sports training adjustment method, device, equipment and storage medium. Background Art

[0002] A jet pool (also called a spa or whirlpool) is a device that uses nozzles to accelerate water into a pool to create different water flow patterns and water flow intensities to provide users with a massage, exercise or leisure experience. Jet pools are usually equipped with a closed water circulation system, which includes a water pump, a filtration system, a heating system (optional) and nozzles. The water pump draws water from the pool, filters and heats it (if any), and then sprays it back into the pool through the nozzles. The nozzles are used to adjust the direction and intensity of the water flow to provide different massage effects. By installing the nozzles at different positions and angles, various water flow patterns can be created, such as jet flow, pulse flow, vortex flow, etc. These water flow patterns can provide massage for different parts of the body, help relieve muscle tension and promote blood circulation. Current jet pools are usually equipped with electronic control systems, through which users can adjust the water flow intensity of the nozzles and turn on / off the nozzles to personalize their experience.

[0003] Current jet pool sports training systems often use fixed or preset water jet patterns and speeds, and lack the ability to perceive and respond to the real-time status of trainees. In these systems, water flow control is mainly based on preset programs or the subjective judgment of coaches, and fails to fully consider the individual differences and real-time physiological changes of trainees during training. For example, current jet pools may set water flow intensity according to general training plans, while ignoring the trainee's current heart rate, fatigue level, and swimming efficiency, which may lead to training intensity that is too high or too low, and cannot effectively promote the improvement of training results.

[0004] Therefore, the existing jet pool sports training system lacks accurate perception and response to the real-time status of trainees, and cannot provide personalized training conditions, resulting in low training efficiency and inability to effectively improve the training effect and training experience of trainees. Summary of the invention

[0005] The embodiments of the present invention provide a jet swimming pool sports training adjustment method, device, equipment and storage medium, which can accurately perceive and respond to the real-time status of trainees and provide personalized training conditions, thereby improving the training effect and training experience of trainees in the jet swimming pool.

[0006] An embodiment of the present invention provides a method for adjusting a jet pool exercise training, comprising:

[0007] Obtain an image of a trainee's swimming posture in a jet pool captured by a camera, and obtain the trainee's swimming speed and heart rate detected by a wearable device;

[0008] Extracting swimming posture features from the swimming posture image by a posture feature extraction algorithm to obtain swimming posture features of the trainee, and extracting features from the swimming speed and the heart rate by a time series feature extraction algorithm to obtain swimming speed features and heart rate features of the trainee;

[0009] Performing multimodal feature fusion on the swimming posture feature, the swimming speed feature and the heart rate feature through a multimodal feature fusion algorithm to obtain a fusion feature;

[0010] The fusion features are analyzed by a reward-based reinforcement learning algorithm to obtain a water jet mode and a water jet speed currently suitable for training the trainee; the water jet mode includes a jet flow, a pulse flow, and a vortex flow;

[0011] The water jetting operation of the jet swimming pool is controlled according to the water jetting mode and the water jetting speed.

[0012] As an improvement of the above solution, the step of extracting swimming posture features from the swimming posture image using a posture feature extraction algorithm includes:

[0013] Input the swimming posture image to the input layer of the DenseNet-121 densely connected network;

[0014] Perform a convolution operation through the initial convolutional layer of the DenseNet-121 densely connected network and output the feature map;

[0015] Processing the feature map through the maximum pooling layer of the DenseNet-121 densely connected network;

[0016] The output of the maximum pooling layer is further processed by the Dense Block layer of the DenseNet-121 densely connected network;

[0017] The output of the last Dense Block is globally averaged pooled through the global average pooling layer of the DenseNet-121 densely connected network;

[0018] The output of the global average pooling layer is input into the fully connected layer of the DenseNet-121 densely connected network to output the final swimming posture features.

[0019] As an improvement of the above scheme, the feature extraction of the swimming speed and the heart rate by a time series feature extraction algorithm to obtain the swimming speed feature and the heart rate feature of the trainee includes:

[0020] Input the swimming speed data sequence and the heart rate data sequence into the input layer of the Bi-LSTM bidirectional long short-term memory network;

[0021] For the swimming speed data sequence and the heart rate data sequence, starting from the first element of each sequence, they are sequentially passed through each unit of the forward LSTM of the Bi-LSTM; the forward LSTM includes an input gate, a forget gate, an output gate, and a cell state;

[0022] Starting from the last element of the respective sequences, reverse through the units of the backward LSTM of the Bi-LSTM; the units of the backward LSTM are the same as the units of the forward LSTM, but process the sequence in the opposite direction;

[0023] Concatenate the hidden states of the forward LSTM and the backward LSTM at each time step to form the final feature vector;

[0024] Through the fully connected layer of Bi-LSTM, the swimming speed feature vector and heart rate feature vector are output respectively.

[0025] As an improvement of the above solution, the multimodal feature fusion algorithm is used to fuse the swimming posture feature, the swimming speed feature and the heart rate feature to obtain a fusion feature, including:

[0026] Performing feature standardization on the swimming posture feature, the swimming speed feature, and the heart rate feature;

[0027] The swimming posture feature, the swimming speed feature and the heart rate feature after feature standardization are fused by the following multimodal feature fusion formula to obtain a fused feature vector F fused :

[0028]

[0029] Among them, α i is the attention weight of the ith modality, F i is the eigenvector of the ith mode;

[0030]

[0031] Among them, W i is the weight vector of the corresponding mode, used to learn the feature vector F i The importance of and relationship between modes; b i is a bias term; j is an index variable used to traverse all modes; This means that for each modality (j=1, 2, 3) its unnormalized attention score is calculated These scores are then summed and the result of this sum is used as the denominator to ensure that the final calculated attention weights are normalized, that is, the sum of the attention weights of all modalities is 1.

[0032] As an improvement of the above scheme, the reward-based reinforcement learning algorithm analyzes the fusion features to obtain the current water jetting pattern suitable for training the trainer, including:

[0033] The fused features are used as the state input of the agent of the reward-based reinforcement learning algorithm to identify the current state of the trainer; the current state includes swimming efficiency, heart rate changes, and fatigue level;

[0034] Based on the current state, the agent selects an action, which includes the water jet pattern and speed. The action selection is based on the strategy π learned by the agent, which can predict the water jet pattern and speed that are most likely to bring the highest cumulative reward based on the current state;

[0035] The construction process of the reward-based reinforcement learning algorithm includes:

[0036] Define the state space S and define the action space A; the state space S is composed of the fused feature vector F fused Composition; the action space A includes three water flow injection modes: jet flow, pulse flow and vortex flow, and the water flow injection speed range of each water flow injection mode; A = {a|a∈jet flow, pulse flow, vortex flow}×[V min ,V max ]}; V min and V max are the minimum and maximum values ​​of the water jet velocity, respectively;

[0037] The reward function for designing a reward-based reinforcement learning algorithm is:

[0038] R(s,a)=γ 1 ΔHR+γ 2 ·η+γ 3 ·(1-τ); where ΔHR represents heart rate change; η represents swimming efficiency; τ is the ratio of swimming distance to energy consumption; γ represents fatigue level, which is heart rate recovery time or subjective fatigue score; 1 , γ 2 and γ 3 It is the weight coefficient of each item in the reward function, which is used to adjust the influence of different factors; the reward function is used to reflect the training effect and experience of the trainer under different water flow modes and speeds;

[0039] The loss function for designing a reward-based reinforcement learning algorithm is:

[0040]

[0041] Among them, r t (θ) is the strategy π θ The action probability ratio under A t is the advantage function, which measures the value of taking action a relative to the average action; the clip function is used to limit the range of action probability ratios to avoid drastic changes in the strategy during the update process; L VF (θ) is the loss of the value function, which is used to predict the expected return of state S; S[π θ ] is the policy entropy, which is used to encourage policy exploration; c 1 and c 2 is a hyperparameter used to balance policy loss, value loss, and policy entropy;

[0042] During the training process, iterations are performed based on the collected empirical data of state, action, reward, and next state. In each round of training, the agent selects actions based on the current state and strategy, and observes the next state and immediate reward after executing the action, thereby updating the strategy parameter θ until the preset training round is reached.

[0043] Another embodiment of the present invention provides a jet swimming pool sports training adjustment device, including:

[0044] A data acquisition module, used to acquire the image of the training personnel's swimming posture in the jet pool captured by the camera, and to acquire the swimming speed and heart rate of the training personnel detected by the wearable device;

[0045] A feature extraction module is used to extract swimming posture features from the swimming posture image by using a posture feature extraction algorithm to obtain swimming posture features of the trainee, and to extract features from the swimming speed and the heart rate by using a time series feature extraction algorithm to obtain swimming speed features and heart rate features of the trainee;

[0046] A feature fusion module, used for performing multimodal feature fusion on the swimming posture feature, the swimming speed feature and the heart rate feature through a multimodal feature fusion algorithm to obtain a fusion feature;

[0047] An analysis module, used for analyzing the fusion features based on a reward-based reinforcement learning algorithm to obtain a water jet mode and water jet speed currently suitable for training the trainee; the water jet mode includes a jet flow, a pulse flow, and a vortex flow;

[0048] The control module is used to control the water jetting operation of the jet swimming pool according to the water jetting mode and the water jetting speed.

[0049] As an improvement of the above solution, the feature fusion module is specifically used for:

[0050] Performing feature standardization on the swimming posture feature, the swimming speed feature, and the heart rate feature;

[0051] The swimming posture feature, the swimming speed feature and the heart rate feature after feature standardization are fused by the following multimodal feature fusion formula to obtain a fused feature vector F fused :

[0052]

[0053] Among them, α i is the attention weight of the ith modality, F i is the eigenvector of the ith mode;

[0054]

[0055] Among them, W i is the weight vector of the corresponding mode, used to learn the feature vector F i The importance of and relationship between modes; b i is a bias term; j is an index variable used to traverse all modes; This means that for each modality (j=1, 2, 3) its unnormalized attention score is calculated These scores are then summed and the result of this sum is used as the denominator to ensure that the final calculated attention weights are normalized, that is, the sum of the attention weights of all modalities is 1.

[0056] As an improvement of the above solution, the analysis module is specifically used for:

[0057] The fused features are used as the state input of the agent of the reward-based reinforcement learning algorithm to identify the current state of the trainer; the current state includes swimming efficiency, heart rate changes, and fatigue level;

[0058] Based on the current state, the agent selects an action, which includes the water jet pattern and speed. The action selection is based on the strategy π learned by the agent, which can predict the water jet pattern and speed that are most likely to bring the highest cumulative reward based on the current state;

[0059] The construction process of the reward-based reinforcement learning algorithm includes:

[0060] Define the state space S and define the action space A; the state space S is composed of the fused feature vector F fused Composition; the action space A includes three water flow injection modes of jet flow, pulse flow and vortex flow and the water flow injection speed range of each water flow injection mode; A = {a|a∈{jet flow, pulse flow, vortex flow}×[V min , Vmax ]}; V min and V max are the minimum and maximum values ​​of the water jet velocity, respectively;

[0061] The reward function for designing a reward-based reinforcement learning algorithm is:

[0062] R(s,a)=γ 1 ΔHR+γ 2 ·η+γ 3 ·(1-τ); where ΔHR represents heart rate change; η represents swimming efficiency; τ is the ratio of swimming distance to energy consumption; γ represents fatigue level, which is heart rate recovery time or subjective fatigue score; 1 , γ 2 and γ 3 It is the weight coefficient of each item in the reward function, which is used to adjust the influence of different factors; the reward function is used to reflect the training effect and experience of the trainer under different water flow modes and speeds;

[0063] The loss function for designing a reward-based reinforcement learning algorithm is:

[0064]

[0065] Among them, r t (θ) is the strategy π θ The action probability ratio under A t is the advantage function, which measures the value of taking action a relative to the average action; the clip function is used to limit the range of action probability ratios to avoid drastic changes in the strategy during the update process; L VF (θ) is the loss of the value function, which is used to predict the expected return of state S; S[π θ ] is the policy entropy, which is used to encourage policy exploration; c 1 and c 2 is a hyperparameter used to balance policy loss, value loss, and policy entropy;

[0066] During the training process, iterations are performed based on the collected empirical data of state, action, reward, and next state. In each round of training, the agent selects actions based on the current state and strategy, and observes the next state and immediate reward after executing the action, thereby updating the strategy parameter θ until the preset training round is reached.

[0067] Another embodiment of the present invention provides a jet pool sports training adjustment device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, and when the processor executes the computer program, the jet pool sports training adjustment method described in the above-mentioned embodiment of the invention is implemented.

[0068] Another embodiment of the present invention provides a storage medium, wherein the computer-readable storage medium includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the jet swimming pool sports training adjustment method described in the above-mentioned embodiment of the invention.

[0069] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0070] The embodiment of the present invention uses a camera and a wearable device to obtain the swimming posture image, swimming speed and heart rate data of the trainee in real time, and then extracts the swimming posture features, swimming speed features and heart rate features respectively through the relevant feature extraction algorithm. Then, a multimodal feature fusion algorithm is used to integrate these features into a comprehensive fusion feature vector to fully reflect the real-time status of the trainee. Finally, a reward-based reinforcement learning algorithm analyzes the fusion features and intelligently decides the water flow injection mode and speed that are most suitable for the current trainee, including jet flow, pulse flow and vortex flow, so as to optimize the training environment and improve the training effect. From the above analysis, it can be seen that the embodiment of the present invention is based on the real-time physiological and behavioral characteristics of the trainee, and by dynamically adjusting the water flow injection mode and speed of the jet swimming pool, the personalized level and effect of the trainee's sports training in the jet swimming pool are significantly improved, and the effective stimulation of the trainee's potential and the significant improvement of the training quality are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a flow chart of a jet pool sports training adjustment method provided by one embodiment of the present invention;

[0072] Figure 2 It is a structural schematic diagram of a jet swimming pool sports training adjustment device provided by one embodiment of the present invention;

[0073] Figure 3 The present invention is a schematic structural diagram of a jet swimming pool sports training adjustment device provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0074] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0075] See also Figure 1 , is a flow chart of a jet pool sports training adjustment method provided by an embodiment of the present invention. The jet pool sports training adjustment method comprises steps S10 to S14:

[0076] S10, obtaining an image of a training person's swimming posture in a jet pool captured by a camera, and obtaining a swimming speed and a heart rate of the training person detected by a wearable device;

[0077] S11, extracting swimming posture features from the swimming posture image by a posture feature extraction algorithm to obtain swimming posture features of the trainee, and extracting features from the swimming speed and the heart rate by a time series feature extraction algorithm to obtain swimming speed features and heart rate features of the trainee;

[0078] S12, performing multimodal feature fusion on the swimming posture feature, the swimming speed feature and the heart rate feature through a multimodal feature fusion algorithm to obtain a fusion feature;

[0079] S13, analyzing the fusion features based on the reward-based reinforcement learning algorithm to obtain a water jet mode and a water jet speed currently suitable for training the trainee; the water jet mode includes a jet flow, a pulse flow, and a vortex flow;

[0080] S14, controlling the water jetting operation of the jet swimming pool according to the water jetting mode and the water jetting speed.

[0081] The embodiment of the present invention uses a camera and a wearable device to obtain the swimming posture image, swimming speed and heart rate data of the trainee in real time, and then extracts the swimming posture features, swimming speed features and heart rate features respectively through the relevant feature extraction algorithm. Then, a multimodal feature fusion algorithm is used to integrate these features into a comprehensive fusion feature vector to fully reflect the real-time status of the trainee. Finally, a reward-based reinforcement learning algorithm analyzes the fusion features and intelligently decides the water flow injection mode and speed that are most suitable for the current trainee, including jet flow, pulse flow and vortex flow, so as to optimize the training environment and improve the training effect. From the above analysis, it can be seen that the embodiment of the present invention is based on the real-time physiological and behavioral characteristics of the trainee, and by dynamically adjusting the water flow injection mode and speed of the jet swimming pool, the personalized level and effect of the trainee's sports training in the jet swimming pool are significantly improved, and the effective stimulation of the trainee's potential and the significant improvement of the training quality are achieved.

[0082] It is understood that after determining the optimal water jetting mode and water jetting speed, controlling the water jetting operation of the jet pool may include the following steps:

[0083] 1. Mode selection and configuration

[0084] Jet Stream Mode: If the algorithm recommends using Jet Stream Mode, the system adjusts the nozzle opening and angle to ensure that the water is sprayed in a continuous, high-intensity manner, providing stable resistance, suitable for training to enhance muscle strength and endurance.

[0085] Pulse flow mode: If the pulse flow mode is selected, the system will control the water flow to spray in a periodic pulse form, by changing the spray frequency and intensity, simulating irregular water flow resistance to improve the training personnel's reaction ability and adaptability.

[0086] Vortex Flow Mode: In vortex flow mode, the system needs to adjust the nozzle layout and rotation direction to create a rotating water flow, increase lateral resistance, and place higher demands on the trainee's sense of balance and core stability.

[0087] 2. Flow and pressure regulation

[0088] Based on the optimal water jet speed calculated by the algorithm, the control system needs to accurately adjust the pump speed or valve opening to achieve the target flow rate. The water flow rate can be stabilized at the expected value by dynamically adjusting the pump power output or changing the cross-sectional area of ​​the water flow channel through the electronic valve.

[0089] 3. Real-time monitoring and feedback

[0090] During the water jetting process, the system should monitor the actual speed and pattern execution of the water flow in real time, collect data through sensors (such as flow meters and pressure sensors), compare them with the target values, and make fine adjustments when necessary to ensure that the training conditions always meet the recommendations of the intelligent algorithm.

[0091] In one embodiment, extracting swimming posture features from the swimming posture image using a posture feature extraction algorithm includes:

[0092] Input the swimming posture image to the input layer of the DenseNet-121 densely connected network;

[0093] Perform a convolution operation through the initial convolutional layer of the DenseNet-121 densely connected network and output the feature map;

[0094] Processing the feature map through the maximum pooling layer of the DenseNet-121 densely connected network;

[0095] The output of the maximum pooling layer is further processed by the Dense Block layer of the DenseNet-121 densely connected network;

[0096] The output of the last Dense Block is globally averaged pooled through the global average pooling layer of the DenseNet-121 densely connected network;

[0097] The output of the global average pooling layer is input into the fully connected layer of the DenseNet-121 densely connected network to output the final swimming posture features.

[0098] In this embodiment, DenseNet-121 is used for swimming posture image feature extraction. Specifically, the structure and function of each layer of DenseNet-121 are introduced as follows:

[0099] 1. Input layer

[0100] Input: swimming posture image, usually an RGB three-channel image.

[0101] Preprocessing: Standardize the images, such as subtracting the mean and dividing by the standard deviation, to improve the training process.

[0102] 2. Initial Convolutional Layer

[0103] Operation: Use a 7x7 convolution kernel with a stride of 2 to convolve the input image.

[0104] Function: Reduce the image size while increasing the number of channels to provide more feature representations for subsequent layers.

[0105] Output: The feature map size is about 112×112, and the number of channels is increased to 64.

[0106] 3. Max Pooling Layer

[0107] Operation: Applies a 3x3 max pooling operation with a stride of 2.

[0108] Function: Further reduce the size of the feature map, reduce the amount of calculation, and retain important features.

[0109] Output: The feature map size becomes 56×56.

[0110] 4.Dense Block

[0111] Composition: Contains multiple densely connected convolutional layers, and the input of each convolutional layer is the concatenation of the outputs of all previous layers.

[0112] Operation: Batch normalization normalizes the input features to stabilize the training process.

[0113] ReLU activation function: introduces nonlinearity, enabling the model to learn complex features.

[0114] 1x1 convolution layer (bottleneck layer): reduces the dimension of input features and reduces computational cost.

[0115] 3x3 convolutional layer: learn more complex features and expand the receptive field.

[0116] Output: The feature map output by each convolutional layer will be connected to the input of all subsequent layers to achieve feature reuse.

[0117] 5. Transition layer

[0118] Operations: Contains batch normalization, ReLU activation, 1x1 convolution, and 2x2 average pooling operations.

[0119] Function: Reduce the number of feature maps, perform downsampling, control the number of network parameters, and avoid overfitting.

[0120] Output: The feature map size is reduced and the number of channels is appropriately reduced.

[0121] 6. Global Average Pooling Layer

[0122] Operation: At the end of the network, the output of the last Dense Block is globally average pooled.

[0123] Function: Convert the feature map into a fixed-length vector as a compact feature representation of the swimming posture.

[0124] Output: A fixed-length vector, typically 1024 dimensions.

[0125] 7. Fully connected layer

[0126] Operation: Connect the output of the global average pooling layer to a fully connected layer.

[0127] Output: Generate the final swimming posture feature vector, which contains the comprehensive information of the swimming posture in the image.

[0128] In this embodiment, the output of each layer of DenseNet-121 is not only passed to the next layer, but also serves as the input of all subsequent layers, which greatly increases the reuse rate of features and also helps the propagation of gradients, solving the gradient vanishing problem in deep networks. Through carefully designed convolutional layers, pooling layers, and fully connected layers, DenseNet-121 can extract highly generalized and descriptive features from swimming posture images, providing strong support for subsequent multimodal fusion and training strategy optimization.

[0129] In one embodiment, extracting the swimming speed and the heart rate by using a time series feature extraction algorithm to obtain the swimming speed feature and the heart rate feature of the trainee includes:

[0130] Input the swimming speed data sequence and the heart rate data sequence into the input layer of the Bi-LSTM bidirectional long short-term memory network;

[0131] For the swimming speed data sequence and the heart rate data sequence, starting from the first element of each sequence, they are sequentially passed through each unit of the forward LSTM of the Bi-LSTM; the forward LSTM includes an input gate, a forget gate, an output gate, and a cell state;

[0132] Starting from the last element of the respective sequences, reverse through the units of the backward LSTM of the Bi-LSTM; the units of the backward LSTM are the same as the units of the forward LSTM, but process the sequence in the opposite direction;

[0133] Concatenate the hidden states of the forward LSTM and the backward LSTM at each time step to form the final feature vector;

[0134] Through the fully connected layer of Bi-LSTM, the swimming speed feature vector and heart rate feature vector are output respectively.

[0135] In this embodiment, the forward LSTM unit is specifically introduced as follows:

[0136] Operation: Traverse the sequence data from front to back. The forward LSTM unit of each time step receives the input of the current time step (which can be swimming speed or heart rate) and updates the state based on the hidden state and cell state of the previous time step.

[0137] Gating mechanism:

[0138] Input gate: determines how much of the current input and the hidden state of the previous time step are written into the new cell state.

[0139] Forget gate: determines which information in the cell state is retained or forgotten.

[0140] Output gate: determines which information in the cell state is output as the hidden state of the current time step.

[0141] The backward LSTM unit is introduced as follows:

[0142] Operation: Similar to forward LSTM, but in the opposite direction, starting from the last element of the sequence and traversing backwards.

[0143] Output: For each time step, a hidden state is produced, which reflects the backward information of the sequence.

[0144] The hidden states of the forward LSTM and the backward LSTM at each time step are concatenated to form a larger feature vector. The feature vector of each time step is mapped to a lower-dimensional space through a fully connected layer, usually using an activation function (such as ReLU or tanh). The final output is the swimming speed feature vector and the heart rate feature vector, which represent the comprehensive features of the swimming speed and heart rate in the sequence, respectively.

[0145] Therefore, this embodiment can effectively capture the long-term dependencies and contextual information of sequence data such as swimming speed and heart rate through the unique gating mechanism and bidirectional processing method of Bi-LSTM, thereby maintaining the integrity of time series data in the feature extraction process of swimming speed and heart rate.

[0146] In one embodiment, the multimodal feature fusion algorithm is used to perform multimodal feature fusion on the swimming posture feature, the swimming speed feature and the heart rate feature to obtain the fused feature, including:

[0147] Performing feature standardization on the swimming posture feature, the swimming speed feature, and the heart rate feature;

[0148] The swimming posture feature, the swimming speed feature and the heart rate feature after feature standardization are fused by the following multimodal feature fusion formula to obtain a fused feature vector F fused :

[0149]

[0150] Among them, α i is the attention weight of the ith modality, F i is the eigenvector of the ith mode;

[0151]

[0152] Among them, W i is the weight vector of the corresponding mode, used to learn the feature vector F i The importance of and relationship between modes; b i is a bias term; j is an index variable used to traverse all modes; This means that for each modality (j=1, 2, 3) its unnormalized attention score is calculated These scores are then summed and the result of this sum is used as the denominator to ensure that the final calculated attention weights are normalized, that is, the sum of the attention weights of all modalities is 1.

[0153] In this embodiment, since different features may have different dimensions and ranges, feature normalization is required to ensure that the importance of each feature is relatively balanced during the fusion process. Common normalization methods include minimum-maximum normalization and Z-score normalization. In addition, through the fusion algorithm of the above-mentioned multimodal attention mechanism, different modal features from swimming posture, swimming speed and heart rate can be effectively integrated to obtain a richer and more expressive fusion feature vector, thereby improving the performance and generalization ability of the overall model. It can be understood that the weight α obtained by learningi , the model can automatically adjust the importance of different modal features, which is very important for processing the complementarity and redundancy of inter-modal information. When a modality is more critical in the current task, the corresponding α i will be larger, thus giving this modality a higher weight in the fusion process.

[0154] In one embodiment, the reward-based reinforcement learning algorithm analyzes the fusion features to obtain a water jetting pattern currently suitable for training the trainer, including:

[0155] The fused features are used as the state input of the agent of the reward-based reinforcement learning algorithm to identify the current state of the trainer; the current state includes swimming efficiency, heart rate changes, and fatigue level;

[0156] Based on the current state, the agent selects an action, which includes the water jet pattern and speed. The action selection is based on the strategy π learned by the agent, which can predict the water jet pattern and speed that are most likely to bring the highest cumulative reward based on the current state;

[0157] The construction process of the reward-based reinforcement learning algorithm includes:

[0158] Define the state space S and define the action space A; the state space S is composed of the fused feature vector F fused The action space A includes three water jet modes, namely, jet flow, pulse flow and vortex flow, and the water jet velocity range of each water jet mode; A = [a|a∈{jet flow, pulse flow, vortex flow}×[V min , V max ]}; V min and V max are the minimum and maximum values ​​of the water jet velocity, respectively;

[0159] The reward function for designing a reward-based reinforcement learning algorithm is:

[0160] R(s,a)=γ 1 ΔHR+γ 2 ·η+γ 3 ·(1-τ); where ΔHR represents heart rate change; η represents swimming efficiency; τ is the ratio of swimming distance to energy consumption; γ represents fatigue level, which is heart rate recovery time or subjective fatigue score; 1 , γ 2 and γ 3 It is the weight coefficient of each item in the reward function, which is used to adjust the influence of different factors; the reward function is used to reflect the training effect and experience of the trainer under different water flow modes and speeds;

[0161] The loss function for designing a reward-based reinforcement learning algorithm is:

[0162]

[0163] Among them, r t (θ) is the strategy π θ The action probability ratio under A t is the advantage function, which measures the value of taking action a relative to the average action; the clip function is used to limit the range of action probability ratios to avoid drastic changes in the strategy during the update process; L VF (θ) is the loss of the value function, which is used to predict the expected return of state s; S[π θ ] is the policy entropy, which is used to encourage policy exploration; c 1 and c 2 is a hyperparameter used to balance policy loss, value loss, and policy entropy;

[0164] During the training process, iterations are performed based on the collected empirical data of state, action, reward, and next state. In each round of training, the agent selects actions based on the current state and strategy, and observes the next state and immediate reward after executing the action, thereby updating the strategy parameter θ until the preset training round is reached.

[0165] In this embodiment, by using a reward-based reinforcement learning algorithm and combining multimodal fusion features, intelligent decision-making of the water jet pattern and speed in a jet pool is achieved. Specifically, first, a comprehensive feature vector is formed by fusing multimodal data (swimming posture, swimming speed, heart rate) as the state input of the reinforcement learning agent to comprehensively identify the current state of the trainee, including swimming efficiency, heart rate changes, and fatigue level. Based on the strategy learned by the agent, according to the current state, the agent selects actions including water jet pattern and speed, with the goal of predicting the water jet pattern and speed that are most likely to bring the highest cumulative reward, thereby optimizing the training effect.

[0166] The reinforcement learning framework construction process includes: defining the state space and action space, where the state space is composed of fused feature vectors, and the action space includes multiple water jet modes and speed ranges. Designing reward functions and loss functions to ensure that the algorithm can dynamically adjust the strategy based on the feedback from trainers (such as heart rate changes, swimming efficiency, and fatigue level) to maximize the training effect. During the training process, the agent collects experience data (state, action, reward, and next state) to iterate the strategy and gradually optimize the strategy parameters until the predetermined training goal is achieved.

[0167] From the above analysis, it can be seen that this embodiment dynamically adjusts the water jet mode and speed of the jet pool through an intelligent algorithm to adapt to the real-time status and needs of the trainees, thereby achieving personalized, efficient and safe sports training. By integrating multimodal data, the reinforcement learning algorithm can respond to the physiological and behavioral changes of the trainees in real time, automatically optimize the training environment, and improve the training effect and experience. The ultimate goal is to maximize the potential development and skill improvement of the trainees, while ensuring the safety and comfort of the training process.

[0168] See also Figure 2 , is a schematic diagram of a structure of a jet swimming pool sports training adjustment device provided by an embodiment of the present invention. The jet swimming pool sports training adjustment device comprises:

[0169] The data acquisition module 10 is used to acquire the swimming posture image of the trainee in the jet pool captured by the camera, and acquire the swimming speed and heart rate of the trainee detected by the wearable device;

[0170] The feature extraction module 11 is used to extract swimming posture features from the swimming posture image by using a posture feature extraction algorithm to obtain swimming posture features of the trainee, and to extract features from the swimming speed and the heart rate by using a time series feature extraction algorithm to obtain swimming speed features and heart rate features of the trainee;

[0171] A feature fusion module 12 is used to perform multimodal feature fusion on the swimming posture feature, the swimming speed feature and the heart rate feature through a multimodal feature fusion algorithm to obtain a fusion feature;

[0172] The analysis module 13 is used to analyze the fusion features based on the reward-based reinforcement learning algorithm to obtain the water jet mode and water jet speed currently suitable for the training of the trainee; the water jet mode includes a jet flow, a pulse flow and a vortex flow;

[0173] The control module 14 is used to control the water jetting operation of the jet swimming pool according to the water jetting mode and the water jetting speed.

[0174] The embodiment of the present invention uses a camera and a wearable device to obtain the swimming posture image, swimming speed and heart rate data of the trainee in real time, and then extracts the swimming posture features, swimming speed features and heart rate features respectively through the relevant feature extraction algorithm. Then, a multimodal feature fusion algorithm is used to integrate these features into a comprehensive fusion feature vector to fully reflect the real-time status of the trainee. Finally, a reward-based reinforcement learning algorithm analyzes the fusion features and intelligently decides the water flow injection mode and speed that are most suitable for the current trainee, including jet flow, pulse flow and vortex flow, so as to optimize the training environment and improve the training effect. From the above analysis, it can be seen that the embodiment of the present invention is based on the real-time physiological and behavioral characteristics of the trainee, and by dynamically adjusting the water flow injection mode and speed of the jet swimming pool, the personalized level and effect of the trainee's sports training in the jet swimming pool are significantly improved, and the effective stimulation of the trainee's potential and the significant improvement of the training quality are achieved.

[0175] As an improvement of the above solution, the feature fusion module is specifically used for:

[0176] Performing feature standardization on the swimming posture feature, the swimming speed feature, and the heart rate feature;

[0177] The swimming posture feature, the swimming speed feature and the heart rate feature after feature standardization are fused by the following multimodal feature fusion formula to obtain a fused feature vector F fused :

[0178]

[0179] Among them, α i is the attention weight of the ith modality, F i is the eigenvector of the ith mode;

[0180]

[0181] Among them, W i is the weight vector of the corresponding mode, used to learn the feature vector F i The importance of and the relationship between modes; b i is a bias term; j is an index variable used to traverse all modes; This means that for each modality (j=1, 2, 3) its unnormalized attention score is calculated These scores are then summed and the result of this sum is used as the denominator to ensure that the final calculated attention weights are normalized, that is, the sum of the attention weights of all modalities is 1.

[0182] As an improvement of the above solution, the analysis module is specifically used for:

[0183] The fused features are used as the state input of the agent of the reward-based reinforcement learning algorithm to identify the current state of the trainer; the current state includes swimming efficiency, heart rate changes, and fatigue level;

[0184] Based on the current state, the agent selects an action, which includes the water jet pattern and speed. The action selection is based on the strategy π learned by the agent, which can predict the water jet pattern and speed that are most likely to bring the highest cumulative reward based on the current state;

[0185] The construction process of the reward-based reinforcement learning algorithm includes:

[0186] Define the state space S and define the action space A; the state space S is composed of the fused feature vector F fused Composition; the action space A includes three water flow injection modes of jet flow, pulse flow and vortex flow, and the water flow injection speed range of each water flow injection mode; A = {a|a∈{jet flow, pulse flow, vortex flow}×[V min , V max ]}; V min and V max are the minimum and maximum values ​​of the water jet velocity, respectively;

[0187] The reward function for designing a reward-based reinforcement learning algorithm is:

[0188] R(s,a)=γ 1 ΔHR+γ 2 ·η+γ 3 ·(1-τ); where ΔHR represents heart rate change; η represents swimming efficiency; τ is the ratio of swimming distance to energy consumption; γ represents fatigue level, which is heart rate recovery time or subjective fatigue score; 1 , γ 2 and γ 3 It is the weight coefficient of each item in the reward function, which is used to adjust the influence of different factors; the reward function is used to reflect the training effect and experience of the trainer under different water flow modes and speeds;

[0189] The loss function for designing a reward-based reinforcement learning algorithm is:

[0190]

[0191] Among them, r t (θ) is the strategy π θ The action probability ratio under A t is the advantage function, which measures the value of taking action a relative to the average action; the clip function is used to limit the range of action probability ratios to avoid drastic changes in the strategy during the update process; L VF(θ) is the loss of the value function, which is used to predict the expected return of state S; S[π θ ] is the policy entropy, which is used to encourage policy exploration; c 1 and c 2 is a hyperparameter used to balance policy loss, value loss, and policy entropy;

[0192] During the training process, the agent iterates based on the collected empirical data of state, action, reward, and next state. In each round of training, the agent selects an action based on the current state and strategy, and observes the next state and immediate reward after executing the action, thereby updating the strategy parameter θ until the preset training round is reached. Figure 3 , is a schematic diagram of a jet pool sports training adjustment device provided by an embodiment of the present invention. The jet pool sports training adjustment device of this embodiment includes: a processor 100, a memory 101, and a computer program stored in the memory 101 and executable on the processor 100, such as a jet pool sports training adjustment program. When the processor 100 executes the computer program, the steps in the above-mentioned jet pool sports training adjustment method embodiments are implemented. Alternatively, when the processor 100 executes the computer program, the functions of the modules / units in the above-mentioned device embodiments are implemented.

[0193] Exemplarily, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to implement the present invention. The one or more modules / units may be a series of computer program instruction segments capable of implementing specific functions, which are used to describe the execution process of the computer program in the jet pool sports training adjustment device.

[0194] The jet pool sports training adjustment device can be a computing device such as a desktop computer, a notebook, a PDA, and a cloud server. The jet pool sports training adjustment device can include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the schematic diagram is only an example of the jet pool sports training adjustment device and does not constitute a limitation of the jet pool sports training adjustment device. It can include more or less components than shown in the figure, or combine certain components, or different components. For example, the jet pool sports training adjustment device can also include input and output devices, network access devices, buses, etc.

[0195] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the jet pool sports training adjustment device, and uses various interfaces and lines to connect various parts of the entire jet pool sports training adjustment device.

[0196] The memory can be used to store the computer program and / or module, and the processor realizes various functions of the jet pool sports training adjustment device by running or executing the computer program and / or module stored in the memory and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0197] Wherein, if the module / unit integrated in the jet pool sports training adjustment device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by the processor, the steps of the above-mentioned various method embodiments can be implemented. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media does not include electrical carrier signals and telecommunication signals.

[0198] It should be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art may understand and implement it without paying any creative effort.

[0199] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A jet pool sports training adjustment method, characterized in that: include: Obtain an image of a trainee's swimming posture in a jet pool captured by a camera, and obtain the trainee's swimming speed and heart rate detected by a wearable device; Extracting swimming posture features from the swimming posture image by a posture feature extraction algorithm to obtain swimming posture features of the trainee, and extracting features from the swimming speed and the heart rate by a time series feature extraction algorithm to obtain swimming speed features and heart rate features of the trainee; Performing multimodal feature fusion on the swimming posture feature, the swimming speed feature and the heart rate feature through a multimodal feature fusion algorithm to obtain a fusion feature; The reward-based reinforcement learning algorithm analyzes the fusion features to obtain a water jetting pattern and a water jetting speed currently suitable for training the trainee; The water jet modes include jet stream, pulse stream and vortex stream; Controlling the water jetting operation of the jet swimming pool according to the water jetting mode and the water jetting speed; The reward-based reinforcement learning algorithm analyzes the fusion features to obtain a water jetting pattern currently suitable for training the trainer, including: The fused features are used as the state input of the agent of the reward-based reinforcement learning algorithm to identify the current state of the trainer; the current state includes swimming efficiency, heart rate changes, and fatigue level; Based on the current state, the agent selects an action, which includes the water jet pattern and speed. The action selection is based on the strategy π learned by the agent, which can predict the water jet pattern and speed that are most likely to bring the highest cumulative reward based on the current state; The construction process of the reward-based reinforcement learning algorithm includes: Define the state space S and define the action space A; the state space S is composed of the fused feature vector F fused The action space A includes three water jet modes, namely, jet flow, pulse flow and vortex flow, and the water jet velocity range of each water jet mode; A = {a|a∈{jet flow, pulse flow, vortex flow}×[V min , V max ]}; V min and V max are the minimum and maximum values ​​of the water jet velocity, respectively; The reward function for designing a reward-based reinforcement learning algorithm is: R(s,a)=γ1·ΔHR+γ2·η+γ3·(1-τ); where ΔHR represents heart rate change; η represents swimming efficiency; τ is the ratio of swimming distance to energy consumption; represents fatigue level, which is heart rate recovery time or subjective fatigue score; γ1, γ2 and γ3 are weight coefficients of each item in the reward function, which are used to adjust the influence of different factors; the reward function is used to reflect the training effect and experience of trainees under different water flow modes and speeds; The loss function for designing a reward-based reinforcement learning algorithm is: Among them, r t (θ) is the strategy π θ The action probability ratio under A t is the advantage function, which measures the value of taking action a relative to the average action; the clip function is used to limit the range of action probability ratios to avoid drastic changes in the strategy during the update process; L VF (θ) is the loss of the value function, which is used to predict the expected return of state S; S[π θ ] is the policy entropy, which is used to encourage policy exploration; c1 and c2 are hyperparameters used to balance policy loss, value loss, and policy entropy; During the training process, iterations are performed based on the collected empirical data of state, action, reward, and next state. In each round of training, the agent selects actions based on the current state and strategy, and observes the next state and immediate reward after executing the action, thereby updating the strategy parameter θ until the preset training round is reached.

2. The jet pool exercise training adjustment method according to claim 1, characterized in that: The step of extracting swimming posture features from the swimming posture image using a posture feature extraction algorithm comprises: Input the swimming posture image to the input layer of the DenseNet-121 densely connected network; Perform a convolution operation through the initial convolutional layer of the DenseNet-121 densely connected network and output the feature map; Processing the feature map through the maximum pooling layer of the DenseNet-121 densely connected network; The output of the maximum pooling layer is further processed by the Dense Block layer of the DenseNet-121 densely connected network; The output of the last Dense Block is globally averaged pooled through the global average pooling layer of the DenseNet-121 densely connected network; The output of the global average pooling layer is input into the fully connected layer of the DenseNet-121 densely connected network to output the final swimming posture features.

3. The jet pool exercise training adjustment method according to claim 1, characterized in that: The step of extracting the swimming speed and the heart rate by using a time series feature extraction algorithm to obtain the swimming speed feature and the heart rate feature of the trainee includes: Input the swimming speed data sequence and the heart rate data sequence into the input layer of the Bi-LSTM bidirectional long short-term memory network; For the swimming speed data sequence and the heart rate data sequence, starting from the first element of each sequence, they are sequentially passed through each unit of the forward LSTM of the Bi-LSTM; the forward LSTM includes an input gate, a forget gate, an output gate, and a cell state; Starting from the last element of the respective sequences, reverse through the units of the backward LSTM of the Bi-LSTM; the units of the backward LSTM are the same as the units of the forward LSTM, but process the sequence in the opposite direction; Concatenate the hidden states of the forward LSTM and the backward LSTM at each time step to form the final feature vector; Through the fully connected layer of Bi-LSTM, the swimming speed feature vector and heart rate feature vector are output respectively.

4. The jet pool exercise training adjustment method according to claim 1, characterized in that: The multimodal feature fusion algorithm is used to perform multimodal feature fusion on the swimming posture feature, the swimming speed feature and the heart rate feature to obtain a fusion feature, including: Performing feature standardization on the swimming posture feature, the swimming speed feature, and the heart rate feature; The swimming posture feature, the swimming speed feature and the heart rate feature after feature standardization are fused by the following multimodal feature fusion formula to obtain a fused feature vector F fused : Among them, α i is the attention weight of the ith modality, F i is the eigenvector of the ith mode; Among them, W i is the weight vector of the corresponding mode, used to learn the feature vector F i The importance of and the relationship between modes; b i is a bias term; j is an index variable used to traverse all modes; This means that for each modality (j=1, 2, 3) its unnormalized attention score is calculated These scores are then summed and the result of this sum is used as the denominator to ensure that the final calculated attention weights are normalized, that is, the sum of the attention weights of all modalities is 1.

5. A jet swimming pool sports training adjustment device, characterized in that: include: A data acquisition module, used to acquire the image of the training personnel's swimming posture in the jet pool captured by the camera, and to acquire the swimming speed and heart rate of the training personnel detected by the wearable device; A feature extraction module is used to extract swimming posture features from the swimming posture image by using a posture feature extraction algorithm to obtain swimming posture features of the trainee, and to extract features from the swimming speed and the heart rate by using a time series feature extraction algorithm to obtain swimming speed features and heart rate features of the trainee; A feature fusion module, used for performing multimodal feature fusion on the swimming posture feature, the swimming speed feature and the heart rate feature through a multimodal feature fusion algorithm to obtain a fusion feature; An analysis module, used for analyzing the fusion features based on a reward-based reinforcement learning algorithm to obtain a water jetting mode and a water jetting speed currently suitable for training the trainee; The water jet modes include jet stream, pulse stream and vortex stream; A control module, used for controlling the water jetting operation of the jet swimming pool according to the water jetting mode and the water jetting speed; The analysis module is specifically used for: The fused features are used as the state input of the agent of the reward-based reinforcement learning algorithm to identify the current state of the trainer; the current state includes swimming efficiency, heart rate changes, and fatigue level; Based on the current state, the agent selects an action, which includes the water jet pattern and speed. The action selection is based on the strategy π learned by the agent, which can predict the water jet pattern and speed that are most likely to bring the highest cumulative reward based on the current state; The construction process of the reward-based reinforcement learning algorithm includes: Define the state space S and define the action space A; the state space S is composed of the fused feature vector F fused The action space A includes three water jet modes, namely, jet flow, pulse flow and vortex flow, and the water jet velocity range of each water jet mode; A = {a|a∈{jet flow, pulse flow, vortex flow}×[V min , V max ]}; V min and V max are the minimum and maximum values ​​of the water jet velocity, respectively; The reward function for designing a reward-based reinforcement learning algorithm is: R(s,a)=γ1·ΔHR+γ2·η+γ3·(1-τ); where ΔHR represents heart rate change; η represents swimming efficiency; τ is the ratio of swimming distance to energy consumption; represents fatigue level, which is heart rate recovery time or subjective fatigue score; γ1, γ2 and γ3 are weight coefficients of each item in the reward function, which are used to adjust the influence of different factors; the reward function is used to reflect the training effect and experience of trainees under different water flow modes and speeds; The loss function for designing a reward-based reinforcement learning algorithm is: Among them, r t (θ) is the strategy π θ The action probability ratio under A t is the advantage function, which measures the value of taking action a relative to the average action; the clip function is used to limit the range of action probability ratios to avoid drastic changes in the strategy during the update process; L VF (θ) is the loss of the value function, which is used to predict the expected return of state S; S[π θ ] is the policy entropy, which is used to encourage policy exploration; c1 and c2 are hyperparameters used to balance policy loss, value loss, and policy entropy; During the training process, iterations are performed based on the collected empirical data of state, action, reward, and next state. In each round of training, the agent selects actions based on the current state and strategy, and observes the next state and immediate reward after executing the action, thereby updating the strategy parameter θ until the preset training round is reached.

6. The jet pool sports training adjustment device as claimed in claim 5, characterized in that: The feature fusion module is specifically used for: Performing feature standardization on the swimming posture feature, the swimming speed feature, and the heart rate feature; The swimming posture feature, the swimming speed feature and the heart rate feature after feature standardization are fused by the following multimodal feature fusion formula to obtain a fused feature vector F fused : Among them, α i is the attention weight of the ith modality, F i is the eigenvector of the ith mode; Among them, W i is the weight vector of the corresponding mode, used to learn the feature vector F i The importance of and the relationship between modes; b i is a bias term; j is an index variable used to traverse all modes; This means that for each modality (j=1, 2, 3) its unnormalized attention score is calculated These scores are then summed and the result of this sum is used as the denominator to ensure that the final calculated attention weights are normalized, that is, the sum of the attention weights of all modalities is 1.

7. A jet pool sports training and adjustment device, characterized in that: The invention comprises a processor, a memory and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the jet pool sports training adjustment method according to any one of claims 1 to 4 is implemented.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the jet pool sports training adjustment method according to any one of claims 1 to 4.

Images