Numerical simulation and optimization method of sports teaching and training based on embedded wireless communication network

In physical education training, IoT devices such as wearable sensors[24], smartwatches[6], and smart clothing[38] are widely used to monitor athletes’ physiological parameters and movement data in real time, including heart rate, respiratory rate, body temperature, acceleration, and posture. [27]For example, smartwatches can record data such as heart rate and calories burned, which can then be uploaded to the cloud for analysis. Smart clothing, equipped with embedded sensors, can capture muscle activity, posture, and respiratory rate, enabling coaches to monitor athletes’ physical conditions in real time.

The data collected through IoT technology can be used to analyze athletes’ performance, identify shortcomings in their movements, and provide optimization suggestions[32]. For instance, smart shoes used in football training can measure data such as running speed, distance, and step frequency, and transmit this information to a coach’s tablet or smartphone via wireless communication modules. Coaches can instantly access performance data and adjust training plans accordingly. Additionally, AI algorithms can further analyze these data to provide personalized training recommendations[4].

The use of IoT technology in physical education training presents several advantages. IoT devices enable real-time collection and transmission of exercise data, allowing for immediate feedback and timely adjustments to training plans. The integration of sensors and data analysis enhances the precision of data collection, capturing subtle changes often missed by traditional methods. Furthermore, IoT can offer personalized training suggestions by analyzing individual athlete data, improving outcomes and reducing injury risks [29]. Another major advantage is data-driven decision-making[10]; the extensive data collected by IoT devices can be analyzed using AI and machine learning models, leading to more informed decisions. Additionally, IoT technology facilitates remote monitoring and guidance, enabling coaches to provide support regardless of athletes’ locations or training schedules, which is especially valuable.

However, the application of IoT technology in physical education training also faces challenges. Firstly, the transmission of large amounts of personal physiological data over wireless networks may pose risks of data breaches or unauthorized access, making privacy protection a pressing issue. Secondly, overreliance on IoT devices for training may cause athletes to neglect their own perception and experience, which are crucial in training. Furthermore, device malfunctions or data errors could lead to incorrect training plans. Cost is another challenge; high-quality IoT devices and the necessary technical support often come with a high price, which could pose economic pressure on some small- to medium-sized sports organizations or individual athletes. Additionally, as the dimensions and volume of collected data increase, efficiently processing and analyzing these data becomes a significant challenge, potentially requiring higher computational resources and specialized technical support.

Recent research has continued to innovate in the application of IoT technology. For example, intelligent training systems that combine Augmented Reality (AR) [28] with IoT are emerging. Recent studies have shown that such systems can provide athletes with an immersive training experience. AR glasses display real-time data feedback, allowing athletes to see their physiological data and movement metrics during training and make immediate adjustments. Moreover, IoT-based injury prediction models are under development, analyzing historical training data and real-time monitoring data to predict potential future injuries and provide corresponding warnings. For example, machine learning models can analyze data from knee pressure sensors to predict the risk of knee joint injuries. Additionally, multi-modal sensor fusion training monitoring systems have gained widespread attention. Researchers are exploring methods to integrate data from multiple sensors (e.g., heart rate sensors, accelerometers, electromyography) and using deep learning models to comprehensively analyze these multi-modal data, thus providing a more holistic assessment of an athlete’s status, including fatigue levels, exercise intensity, and recovery conditions. These recent research advances highlight the broad prospects of IoT technology in physical education training and point the way for future research and applications.

The use of deep learning in sports data analysis has rapidly emerged as a forefront area in modern sports science research. By leveraging sophisticated algorithms, these technologies empower coaches, trainers, and athletes to make more data-driven and informed decisions by automatically identifying and extracting crucial features and patterns from extensive and often complex sports datasets. For instance, CNNs [44] are widely employed in image and video analysis, facilitating tasks such as action recognition and posture estimation. On the other hand, Recurrent Neural Networks (RNNs) and their advanced versions, such as Long Short-Term Memory (LSTM) networks, are particularly adept at handling sequential data, including time-series data and sensor-based information. These applications are especially beneficial in sports where continuous monitoring and real-time feedback are critical to enhancing performance and preventing injuries.

In recent years, the development of specialized systems, such as the language processing-based automatic news generation and fact-checking system proposed by Peng et al. [20], has further highlighted the vast potential of deep learning technologies beyond traditional sports analysis, extending their utility into areas like sports reporting and data verification. Moreover, the integration of traditional machine learning algorithms, such as Random Forests and Support Vector Machines (SVMs), continues to offer robust and reliable solutions for classification and regression tasks within the sports data context, complementing the strengths of deep learning approaches.

However, despite the significant advancements, several challenges persist in the application of deep learning to sports data. Sports datasets often exhibit high dimensionality and intricate temporal dependencies, making data preprocessing and model training more complicated. Additionally, deep learning models generally require substantial amounts of labeled data, which can be particularly difficult to acquire in certain niche sports domains where data availability is limited. The process of annotating large-scale datasets is often labor-intensive and time-consuming, further complicating the application of these models. Training these models also demands significant computational resources, and when working with large-scale datasets, the prolonged training times may not align with the needs of real-time applications or quick decision-making. For example, the model proposed by Hengmin Zhang et al. [41] is heavily reliant on computational resources, limiting its effectiveness in real-time data processing, particularly in fast-paced sports environments where decisions need to be made quickly.

Moreover, the challenge of interpretability in deep learning models remains a significant barrier. Coaches and athletes often require clear and understandable insights to make effective use of the predictions generated by these models. Therefore, as deep learning continues to advance, there is a growing need to develop techniques that improve the transparency and interpretability of these models, making their outputs more actionable for practitioners in the sports field.

In summary, the application of deep learning and machine learning technologies in sports data analysis has shown great potential, particularly in action recognition, posture estimation, and performance analysis. While these technologies offer significant advantages in handling complex data, automating feature extraction, and providing personalized recommendations, challenges remain in data labeling, high computational resource requirements, and the interpretability of models. As research progresses, more and more improvements are being proposed, such as using transfer learning to address small sample problems or employing model compression and optimization techniques to reduce computational complexity. Additionally, improving model interpretability to make training results more understandable and applicable is becoming an important direction for future research. In conclusion, the future of deep learning and machine learning in sports data analysis is promising, and ongoing research will further promote the application of these technologies in practical sports training and competition.

IoT advanced systems harness real-time monitoring, comprehensive data analysis, and feedback mechanisms to deliver personalized training programs, thereby improving the overall effectiveness of training outcomes. For example, Wang and Park [30] developed an intelligent training system specifically designed to improve college students’ mental health and physical fitness through targeted exercise regimens. Typically, these systems incorporate a wide array of sensors, wearable devices, and sophisticated data analysis platforms, which work in concert to collect and process real-time physiological and movement data from athletes. The data are then analyzed using AI algorithms to generate specific and actionable training recommendations tailored to individual athletes.

Despite the numerous advantages that intelligent sports training systems offer, their design and implementation are not without challenges. One significant hurdle is the effective integration of multi-source data to ensure the accuracy and consistency of the information provided by these systems. For example, Wei et al. [34] introduced an AI-based training system that leveraged big data to optimize training, but the system faced considerable challenges due to limitations in data integration and processing capabilities. These challenges ultimately hindered the system’s scalability, leading to its retraction from broader application. Moreover, the necessity for real-time data processing adds another layer of complexity, as it requires the system to function seamlessly without disrupting the athletes’ training routines. This demands a careful balance between generalization and personalization to deliver training programs that cater to athletes with varying skill levels and needs.

Furthermore, the application of intelligent sports training systems in physical education offers a multitude of advantages. Firstly, these systems are capable of monitoring athletes’ training status in real-time by collecting vast amounts of physiological and movement data through sensors and wearable devices. This allows both coaches and athletes to receive feedback promptly, enabling them to make necessary adjustments to training plans with immediate effect. Luo et al. [15]conducted a study on an intelligent sports training system based on Triboelectric Nanogenerators (TENGs), which demonstrated significant improvements in energy harvesting and motion monitoring efficiency through innovative technological applications. This study underscores the potential of integrating intelligent sports devices into training systems. However, these advanced systems also present certain drawbacks. The high costs associated with their design and implementation, particularly concerning the setup of hardware devices and data processing platforms, can be prohibitive. Additionally, the complexity of these systems often increases the difficulty of maintenance and operation, necessitating additional training for non-technical personnel to use the systems effectively. This may further complicate the widespread adoption of such systems in various sports training environments.

In summary, intelligent sports training systems, by incorporating advanced technologies such as AI, big data, and IoT, offer athletes scientific and personalized training programs that significantly enhance training effectiveness and efficiency. Nevertheless, challenges remain in areas such as data integration, real-time performance, cost management, and user-friendliness. As these systems continue to evolve, it is crucial to address these challenges to fully realize their potential in transforming sports training and physical education. Ongoing research and development efforts are essential to overcoming these obstacles, with a particular focus on improving the scalability and accessibility of intelligent training systems while ensuring they remain cost-effective and easy to use for a broad range of users.

In this study, we employ embedded wireless sensors to capture real-time motion data from athletes, including metrics such as heart rate, acceleration, and speed. After transmission to a central processing system via a wireless network, the data undergoes preprocessing before being fed into the ASPP module for extracting multi-scale features. ASPP leverages dilated convolution filters with varying sampling rates to process data across multiple scales, enabling the capture of spatial information at different resolutions. This effectively enlarges the receptive field of the neural network, allowing it to gather richer contextual details without escalating computational load.

The core of the ASPP module is the dilated convolution, mathematically expressed as: (1)z[i]=∑j=1Kx[i+s⋅j]⋅u[j]$$z\left[ i \right] = \mathop \sum \nolimits_{j = 1}^K x\left[ {i + s \cdot j} \right] \cdot u\left[ j \right]$$

where z[i] denotes the convolution output, x[i] is the input feature map, u[j] is the convolution filter, and s represents the dilation factor, with j marking the filter index.

ASPP achieves feature extraction across different scales by varying the dilation rate s. Suppose the ASPP module comprises M dilated convolution layers, each with dilation rates s₁, s₂, …, s_M, then the convolution output is expressed as: (2)z(m)[i]=∑j=1Kx[i+sm⋅j]⋅u(m)[j],m=1,2,…,M$${z^{\left( m \right)}}\left[ i \right] = \mathop \sum \nolimits_{j = 1}^K x\left[ {i + {s_m} \cdot j} \right] \cdot {u^{\left( m \right)}}\left[ j \right],m = 1,2, \ldots ,M$$

Here, z^(m)[i] is the output of the m-th dilated layer, and u^(m)[j] is the corresponding filter.

Finally, the ASPP module aggregates the outputs of the various dilated layers, typically through concatenation: (3)zASPP[i]=concat(z(1)[i],z(2)[i],…,z(M)[i])$${z_{{\text{ASPP}}}}\left[ i \right] = {\text{concat}}\left( {{z^{\left( 1 \right)}}\left[ i \right],{z^{\left( 2 \right)}}\left[ i \right], \ldots ,{z^{\left( M \right)}}\left[ i \right]} \right)$$

where concat(·) indicates concatenation along the channel dimension, yielding the ASPP module’s final output.

In these formulas, z[i] is the output feature, x[i] is the input map, u[j] represents filter weights, s is the dilation rate, and M is the number of dilated layers. By aggregating the outputs from different scales, the ASPP module enhances the network’s capability to recognize intricate motion patterns, forming a robust basis for LSTM-based temporal analysis.

The ASPP module captures contextual information at multiple scales within the input features, thus enhancing the recognition of complex motion patterns. This provides a solid foundation for the subsequent use of LSTM for timing analysis.

LSTM is a recurrent neural network tailored for processing sequential data, effectively capturing long-term dependencies through its distinct architecture. In this research, spatial features derived from the ASPP module are input into the LSTM network to manage temporal relationships within motion data. The LSTM, utilizing memory and forget mechanisms, accurately reflects variations in an athlete’s condition across different training stages (see Figure 2 for the LSTM structure following convolution).

The core of an LSTM consists of three gates: forget, input, and output gates. These gates collaborate to regulate information flow within the cell state. The operations are mathematically defined as follows: (4)fn=σ(Wf⋅[hn−1,xn]+bf)$${f_n} = {\rm \sigma }\left( {{W_f} \cdot \left[ {{h_{n - 1}},{x_n}} \right] + {b_f}} \right)$$ (5)in=σ(Wi⋅[hn−1,xn]+bi)$${i_n} = {\rm \sigma }\left( {{W_i} \cdot \left[ {{h_{n - 1}},{x_n}} \right] + {b_i}} \right)$$ (6)Sn˜=tanh(WS⋅[hn−1,xn]+bS)$$\tilde {{S_n}} = \tanh \left( {{W_S} \cdot \left[ {{h_{n - 1}},{x_n}} \right] + {b_S}} \right)$$ (7)Sn=fn⋅Sn−1+in⋅Sn˜$${S_n} = {f_n} \cdot {S_{n - 1}} + {i_n} \cdot \tilde {{S_n}}$$ (8)on=σ(Wo⋅[hn−1,xn]+bo)$${o_n} = {\rm \sigma }\left( {{W_o} \cdot \left[ {{h_{n - 1}},{x_n}} \right] + {b_o}} \right)$$ (9)hn=on⋅tanh(Sn)$${h_n} = {o_n} \cdot \tanh \left( {{S_n}} \right)$$

In these equations, f_n is the output of the forget gate, i_n is the output of the input gate, Sn˜$$\tilde {{S_n}}$$ represents the candidate state, S_n is the current cell state, o_n is the output of the output gate, and h_n is the hidden state at time step n. The matrices W_f, W_i, W_S, W_o correspond to the weights linked with the gating units, while b_f, b_i, b_S, b_o are the associated bias terms. The function σ is typically the sigmoid, and tanh denotes the hyperbolic tangent.

Figure 2.

Multi-group LSTM network structure after convolution.

These equations explain how the LSTM controls the retention of information from the previous state S_n−1 through the forget gate f_n, updates the current state S_n via the input gate i_n and candidate state Sn˜$$\tilde {{S_n}}$$, and generates the hidden state h_n at time step n using the output gate o_n and the updated state S_n. These mechanisms enable the LSTM to effectively manage long-term dependencies, making it ideal for analyzing shifts in an athlete’s state across various training phases.

In these formulas, x_n represents the input features, typically derived from the ASPP module. The variables h_n−1 and h_n are the hidden states from the previous and current time steps, respectively. Similarly, S_n−1 and S_n refer to the cell states from the previous and current steps. The parameters W_f, W_i, W_S, W_o, and biases b_f, b_i, b_S, b_o are the weight matrices and bias vectors optimized during training.

In the model optimization phase, the Gradient Descent (GD) algorithm is widely utilized for iterative updates of model parameters, aimed at minimizing the loss function. The integrated ASPP and LSTM model undergoes several training cycles and parameter tuning, ultimately achieving high-precision predictions for motion data. GD operates by computing the gradient of the loss function concerning model parameters and adjusting these parameters in the direction opposite to the gradient, steering the model toward optimal performance.

The fundamental formula for Gradient Descent is: (10)ϕn+1=ϕn−η⋅∇ϕL(ϕn)$${\phi _{n + 1}} = {\phi _n} - {\rm \eta} \cdot {\nabla _\phi }L\left( {{\phi _n}} \right)$$

where ϕ_n represents the model parameters at iteration n, η is the learning rate, and ∇_ϕL(ϕ_n) denotes the gradient of the loss function L(ϕ) with respect to ϕ.

The gradient is calculated as: (11)∇ϕL(ϕ)=∂L(ϕ)∂ϕ$${\nabla _\phi }L\left( \phi \right) = \frac{{\partial L\left( \phi \right)}}{{\partial \phi }}$$

where ∂L(ϕ)∂ϕ$$\frac{{\partial L\left( \phi \right)}}{{\partial \phi }}$$ is the partial derivative of the loss function.

During each iteration, the parameter update formula can be further expanded as: (12)ϕn+1=ϕn−η⋅1p∑j=1p∇ϕLj(ϕn)$${\phi _{n + 1}} = {\phi _n} - {\rm \eta} \cdot \frac{1}{p}\mathop \sum \nolimits_{j = 1}^p {\nabla _\phi }{L_j}\left( {{\phi _n}} \right)$$

where p denotes the number of training samples, and ∇_ϕL_j(ϕ_n) is the gradient for the j th sample. Iteratively, GD refines the model parameters, reducing the loss function and enhancing the model’s accuracy.

In these equations, ϕ_n signifies the model parameters at iteration n; η is the learning rate, determining the update magnitude; ∇_ϕL(ϕ_n) directs the update process; and p is the count of training samples. By appropriately selecting η and performing multiple iterations, the model progressively converges to optimal parameters, thus improving prediction accuracy on motion data.

The experiments employ two publicly available datasets: PAMAP2 [26] and MHEALTH [1].

The PAMAP2 dataset includes physiological and motion data from nine participants, collected using three IMUs and a heart rate monitor, capturing metrics like heart rate, acceleration, angular velocity, magnetic field, and temperature. This dataset is ideal for real-time monitoring and analysis in physical training, particularly when integrated with deep learning models for time-series prediction.

The MHEALTH dataset comprises physiological and motion data from ten volunteers engaged in various activities. Each volunteer is equipped with three IMUs placed on the chest, right wrist, and left ankle. The data encompass activities such as jogging, walking, and cycling, with detailed metrics including acceleration, angular velocity, and heart rate. These data are valuable for time-series prediction using LSTM and multi-scale spatial feature extraction via the ASPP module.

Each dataset is split into training, validation, and test sets in a 60%, 20%, and 20% ratio, respectively.

Throughout the training process, the model undergoes numerous iterations using the previously mentioned optimizer, loss function, batch size, and learning rate configurations. These iterations ensure that the model adapts and fine-tunes its performance progressively. During training, the model undergoes multiple iterations using the previously mentioned optimizer, loss function, batch size, and learning rate settings.

The Maximum F-measure F_β is calculated as: (13)Fβ=(1+β2)×Precision×Recallβ2×Precision+Recall$${F_\beta } = \frac{{\left( {1 + {\beta ^2}} \right) \times {\text{Precision}} \times {\text{Recall}}}}{{{\beta ^2} \times {\text{Precision}} + {\text{Recall}}}}$$

Here, the parameter β adjusts the balance between precision and recall, typically set to 0.5 to emphasize precision. This metric provides insights into the trade-off between these two important aspects of model accuracy.

The Mean Absolute Error (MAE) can be computed as: (14)MAE=1H×W∑i=1H∑j=1W|P(i,j)−G(i,j)|$$\text{MAE}=\frac{1}{H\times W}\sum\nolimits_{i=1}^{H}{\sum\nolimits_{j=1}^{W}{\ \left| \mathbf{P}\left( i,j \right)-\mathbf{G}\left( i,j \right) \right|}}$$

where H and W are the image height and width. P(i, j) is the predicted saliency score at position (i, j), normalized to [0,1], and G(i, j) is the ground truth saliency score, also normalized to [0,1]. while G(i, j) represents the corresponding ground truth saliency score, also normalized to the same range. MAE provides a straightforward measure of the average prediction error across the image.

The Weighted F-measure Fωβ$$F_{\rm \omega}^\beta$$ is calculated as: (15)Fβω=(1+β2)×Precisionω×Recallωβ2×Precisionω+Recallω$$F_\beta ^\omega = \frac{{\left( {1 + {\beta ^2}} \right) \times {\text{Precisio}}{{\text{n}}^\omega } \times {\text{Recal}}{{\text{l}}^\omega }}}{{{\beta ^2} \times {\text{Precisio}}{{\text{n}}^\omega } + {\text{Recal}}{{\text{l}}^\omega }}}$$

Here, Precision^ω and Recall^ω represent the weighted precision and recall, taking into account the importance of each pixel. The parameter β is similarly used to adjust the relative importance of precision and recall, typically set to 0.3.

The Structure Similarity Measure S_m can be computed as (16)Sm=αSo+(1−α)Sr$${S_m} = \alpha {S_o} + \left( {1 - \alpha } \right){S_r}$$

where S_o measures object-aware structure similarity between the predicted saliency map and the ground truth at the object level, and S_r measures region-aware similarity at the region level. The parameter α balances the importance of S_o and S_r, typically set to 0.5.

The Enhanced Alignment Measure E_m can be calculated as: (17)Em=1H×W∑i=1H∑j=1WE(P(i,j),G(i,j))$${{E}_{m}}=\frac{1}{H\times W}\sum\nolimits_{i=1}^{H}{\sum\nolimits_{j=1}^{W}}\ \ E\left( \mathbf{P}\left( i,j \right),\mathbf{G}\left( i,j \right) \right)$$

where E(P(i, j), G(i, j)) evaluates the alignment between the predicted saliency P and the ground truth saliency map G at pixel (i, j), and H and W representing the image dimensions.

In Table 1, we present a comparative analysis of the performance of several deep learning models applied to the PAMAP2 and MHEALTH datasets, assessed using four key metrics: Maximum F-measure F_β, Mean Absolute Error (MAE), Weighted F-measure Fωβ$$F_{\rm \omega}^\beta$$, and Structure Similarity Measure S_m. Specifically, the F_β score provides an evaluation of the balance between precision and recall, while the MAE offers a measure of the average deviation between predicted values and ground truth. Additionally, the Fωβ$$F_{\rm \omega}^\beta$$ metric places emphasis on precision and recall after applying a weighting factor, and S_m focuses on evaluating structural similarity.

The experimental findings decisively show that our proposed ASPP+LSTM model outperforms all other models across both datasets, achieving superior performance on every metric considered. For instance, when tested on the PAMAP2 dataset, our model recorded an impressive F_β score of 0.740 and an MAE of 0.178. In comparison, alternative models such as AFNet and TSPOANet yielded F_β scores of 0.721 and 0.716, with corresponding MAE values of 0.184 and 0.185, respectively. These results underscore the effectiveness of our model, particularly its enhanced ability to capture and leverage multi-scale spatial features along with complex temporal patterns, thereby leading to highly accurate predictions of training outcomes.

Moreover, as depicted visually in Figure 3, these outcomes further illustrate the distinct advantages our ASPP+LSTM model holds across all evaluation metrics. The figure clearly demonstrates how our model consistently surpasses the other methods evaluated on the PAMAP2 and MHEALTH datasets, highlighting its remarkable capability and reliability in real-world applications.

Figure 3.

Multiple performance metrics comparison of different models on PAMAP2 and MHEALTH datasets.

In Table 2, we compared the performance of various advanced models on the PAMAP2 and MHEALTH datasets using the Enhanced Alignment Measure E_m. The E_m metric evaluates the alignment between predictions and ground truth at both the global and local pixel levels, reflecting the model’s ability to capture detail and maintain overall consistency. The results show that our ASPP+LSTM model achieved the highest E_m scores on both datasets, reaching 0.869 and 0.865, significantly outperforming other methods. For example, on PAMAP2, PiCA and PoolNet had E_m scores of 0.754 and 0.761, while on MHEALTH, DUCRF, though close, still fell short with an E_m of 0.821. These results indicate that our proposed model exhibits strong generalization capabilities across different datasets, maintaining superior alignment performance in various environments.

To validate the contributions of the ASPP and LSTM modules to the model’s performance, we designed an ablation study that included the following configurations:

M1: Removed both ASPP and LSTM modules, using only the base CNN model for feature extraction and prediction through fully connected layers. M2: Removed the ASPP module, using only LSTM for time-series prediction. M3: Replaced the ASPP module with CNN and performed predictions. M4: Removed the LSTM module, using only the ASPP module for spatial feature extraction and prediction through fully connected layers. M5: Complete ASPP+LSTM model.

Each model configuration was trained and evaluated, with results recorded for the test set on the metrics F_β, MAE, Fωβ$$F_\omega ^\beta$$, S_m, and E_m. The ablation study results were analyzed to clarify the contributions of each module to the final performance.

In Table 3, we assessed the contributions of the ASPP and LSTM modules to the model’s performance through an ablation study. The experiments were conducted on the PAMAP2 and MHEALTH datasets, using key metrics such as Maximum F-measure F_β, Mean Absolute Error (MAE), Weighted F-measure Fωβ$$F_\omega ^\beta$$, Structure Similarity Measure S_m, and Enhanced Alignment Measure E_m. These metrics respectively evaluate the model’s precision, error, weighted balance, structural similarity, and alignment. The results from five different model configurations show that the full ASPP+LSTM model (M5) consistently performed the best across all metrics. For instance, on the PAMAP2 dataset, the model achieved an F_β of 0.723 and an MAE of 0.140, significantly outperforming the CNN-only M1 configuration (F_β of 0.701 and MAE of 0.147). This highlights the critical role of the ASPP module in multi-scale spatial feature extraction and the significant enhancement of temporal prediction accuracy by the LSTM module.