Introduction

Research background and motivations

In the field of competitive sports today, aerobics stands out as a unique and challenging discipline, attracting a significant number of athletes1,2. This sport emphasizes both physical flexibility and coordination, requiring athletes to demonstrate a high level of technical proficiency in their performances3,4,5. Existing research explores the effects of factors such as exercise training, nutrition intake, and technical support on athlete performance. However, most current studies rely on small sample sizes or data collected in specific environments, limiting their applicability across diverse populations and varied settings. Furthermore, there is a significant research gap in the existing literature regarding the integration of personalized nutrition with advanced artificial intelligence (AI) technologies, particularly deep learning (DL) algorithms, to optimize athlete training and recovery.

DL is one of the AI algorithms. It is applied to enhance the skills of aerobics athletes primarily to make them more intelligent and personalized6. DL provides real-time feedback and personalized suggestions by thoroughly analyzing movement data. It offers athletes and coaches more comprehensive and accurate information, thereby facilitating effective skill improvement7,8,9,10. The application of this technology in the training process enhances the athletes’ competitive level and holds the potential to drive technological innovation across the entire field of aerobics.

Research objectives

This work aims to comprehensively analyze the factors influencing the competitive level of aerobics athletes by integrating advanced technologies, including sports nutrition assistance and DL algorithms in AI, to enhance their overall performance. Therefore, the innovation of this work lies in combining personalized sports nutrition assistance with AI neural network technology, proposing a novel method for researching factors influencing aerobics athletes’ performance. By comprehensively utilizing DL algorithms, particularly optimizing the ShuffleNet V3 network structure and introducing channel attention mechanisms, this work has significantly improved the accuracy of classifying and recognizing athletes’ movements. Ultimately, the combined impact of different factors on the performance of aerobics athletes is comprehensively understood. This provides more scientific and personalized training programs for aerobics athletes, propelling the entire field forward. The contributions of this work are as follows:

Integration of Personalized Nutrition and AI Technology: this work develops a method that combines athletes’ personalized nutritional needs with AI neural network technology, providing customized training and nutrition plans for aerobics athletes.

Innovative Application of DL Models: this work improves the accuracy and efficiency of classifying and recognizing athletes’ movements by optimizing the ShuffleNet V3 network structure and introducing channel attention mechanisms.

Creation and Validation of a Self-Built Dataset: this work constructs a large-scale self-built dataset to enhance the model’s generalization ability and validates the proposed method’s effectiveness in diverse environments.

Real-Time Feedback and Dynamic Adjustment Mechanism: this work designs a real-time feedback system to monitor athletes’ training and nutritional status, allowing dynamic adjustments to training and nutrition plans to meet athletes’ immediate needs.

Construction of an Interdisciplinary Methodology: this work integrates theories and technologies from sports science, nutrition, data science, and AI to provide a new interdisciplinary perspective for research on improving athlete performance.

  Literature review

Over the past few years, notable advancements have been achieved in the field of sports nutrition research. Numerous studies have explored the roles of different nutrients in athletes’ physical fitness, recovery, and overall health. For instance, Jonvik et al. (2022)11 identified new opportunities driving the development of sports nutrition, emphasizing personalized nutritional plans and advanced technological means. Lowery et al. (2023)12 specifically highlighted the positive effects of coffee on athletic performance, emphasizing the role of caffeine in enhancing alertness and endurance. Jagim et al. (2023)13 explored the impact of energy drinks and energy shots, advocating cautious management of their dosage and timing. Vázquez-Espino et al. (2022)14 studied the nutritional knowledge, attitudes, information sources, and dietary habits of sports team athletes, emphasizing the importance of personalized guidance. Jang et al. (2024)15 investigated the hierarchical structure of sustainable sports coaching competency in South Korea.

The application of AI in sports science has shown a growing trend, providing powerful tools for processing and analyzing sports data, with numerous scholars conducting relevant research. Qing et al. (2024)16 achieved motion recognition in a low-sample situation using DL and time-set matching, enhancing the accuracy of action recognition. Kumar Jain et al. (2023)17 improved the performance of a DL model for human action recognition through hyperparameter tuning. Afsar et al. (2023)18 focused on using wearable sensors in Exergaming to identify sports activities through DL models, demonstrating high accuracy in sports recognition.

The factors influencing the performance of aerobics athletes involve various aspects, including technical proficiency, physical fitness, and psychological traits. Numerous scholars have conducted relevant research. For example, Wang et al. (2023)19 found that aerobic fitness had a protective effect on the harmful impact of extreme exercise, particularly in information processing abilities. Abraham et al. (2023)20 proposed the recovery of epigenetic reprogramming through diet and exercise, offering a new perspective for preventing and managing metabolic diseases. Mendez-Gutierrez et al. (2023)21 studied the impact of exercise on hormone secretion, emphasizing the potential role of exercise in regulating brown adipose tissue function. McKenzie et al. (2023)22 explored factors influencing community fitness participation for cerebral palsy patients, guiding personalized rehabilitation and exercise plans. Schaffarczyk et al. (2023)23 validated the effectiveness of non-linear indicators of heart rate variability in determining aerobic and anaerobic thresholds during cycling. Jia et al. (2024)24 found that exercise self-efficacy and health-related quality of life mediated the relationship between social support and relapse tendency. Zaman et al. (2022)25 focused on the importance of monitoring driver health and emotions in intelligent vehicle systems. They combined improved Faster R-CNN and NasNet-Large CNN architectures with transfer learning for the recognition of real-time driver emotions. This model achieved high recognition accuracy on both self-built datasets and public datasets. Hussain et al. (2022)26 utilized a hybrid phantom method to enhance privacy and reduce energy consumption by combining phantom nodes and multipath routing. Ullah et al. (2022)27 addressed the issues of spectrum reuse and interference management in modern cellular systems. They proposed a partition-based Fractional Frequency Reuse (FFR) scheme based on game theory auction mechanisms to optimally allocate bandwidth resources to individual users in practical multi-cell network deployments. Zaman et al. (2023)28 focused on the design and optimization of wireless body sensor networks in medical environments, and employed sensor nodes for remote operations.

Upon examining the studies conducted by these researchers, it becomes apparent that they have delved into various aspects of the factors influencing aerobics athletes’ performance, focusing on sports nutrition, the application of AI in sports science, and various factors affecting aerobics athletes’ performance. However, there is limited exploration of specific cases of these applications in the enhancement of aerobics athletes’ skills, and the combination of sports nutrition and DL is relatively rare in research. The research innovation lies in the integration of sports nutrition and AI, providing more personalized and scientific training programs for aerobics athletes.

  Research methodology

This work employs a comprehensive approach, combining sports nutrition supplementation and AI technology to explore the factors influencing aerobic athletes’ performance comprehensively. The research methodology mainly includes three aspects. They are personalized assessment and analysis of athletes’ nutritional needs, the application of AI technology in optimizing neural network analysis, and analysis of aerobic athlete movement classification and recognition models based on ShuffleNet V3. First, in the personalized assessment and analysis of athletes’ nutritional needs, various data collection methods are utilized. They include fitness tests, physiological monitoring devices, surveys, and big data analysis techniques. A personalized nutritional needs model for athletes is established based on these data, providing accurate personalized nutritional guidance for each athlete. The data collection stage is optimized to further understand the characteristics and requirements of aerobic athletes. Athletes’ exercise data are integrated with nutritional data to better analyze and identify athletes’ behavior patterns and nutritional needs trends. Moreover, in terms of AI technology, ShuffleNet V3 is chosen as the basic network structure, and a channel attention mechanism is introduced to focus on key channels. In order to further optimize the network structure, improvements to ShuffleNet V2 and Inception V3 are described to adapt to the complexity and characteristics of aerobic exercise. In the optimization analysis of neural networks, detailed explanations are provided regarding the basic principles of Convolutional Neural Network (CNN) and how it is optimized using algorithms such as ShuffleNet V2 and Inception V3. Finally, researchers propose a new model in the analysis of the aerobic athlete movement classification and recognition model based on ShuffleNet V3. This model achieves accurate classification and recognition of aerobic athlete movements by integrating sports nutrition, ShuffleNet V3, and attention mechanisms. The structure and working principles of the model are elaborated, and a pseudocode flow using ShuffleNet V3 and attention mechanisms is provided.

Assessment and analysis of personalized athlete nutrition needs

Energy and nutrient requirements for training and competition may vary significantly among athletes29. Assessing personalized athlete nutrition needs allows better catering to each athlete’s unique training demands, enhancing their physical performance and skill levels30,31. Figure 1 illustrates the specific process of assessing personalized athlete nutrition needs.

Fig. 1
figure 1

Schematic diagram of the personalized assessment process for athlete nutrition needs.

Full size image

Figure 1 illustrates the entire process of assessing personalized athlete nutrition needs. The initial stage involves data collection using diverse methods, including fitness tests, physiological monitoring devices, surveys, and big data analysis techniques32,33. Subsequently, the integrated analysis results are used to establish a personalized athlete nutrition needs model. This model considers data from fitness tests and physiological monitoring, providing accurate and personalized nutrition guidance for each athlete.

Further optimization of the data collection stage has been implemented to comprehensively understand the individual characteristics and requirements of aerobics athletes. This optimization involves integrating the athletes’ movement data with nutritional data, as depicted in Fig. 2.

Fig. 2
figure 2

The integrated process diagram of exercise and nutrition data for aerobics athletes.

Full size image

Figure 2 illustrates the process of integrating athletes’ exercise and nutrition data. Initially, diverse data from multiple sources, including exercise metrics, body indicators, and personalized nutritional information, are collected34,35,36. These data are then stored in a dedicated database through advanced integration techniques, forming a self-established dataset. Subsequently, neural network algorithms are employed for DL analysis to recognize patterns in athletes’ behavior and trends in nutritional requirements. Finally, the analysis results are presented through visual tools such as charts, offering coaches and athletes an intuitive representation of the data.

Optimization analysis of neural networks in AI technology

CNN is a type of feedforward neural network and a prominent algorithm in DL. In the analysis of movements in aerobics, CNN has been widely applied in the field of image recognition37,38. Further optimization of the CNN model is crucial for better adapting to the complexity and characteristics of aerobics movements. ShuffleNet is a high-performance and lightweight CNN. It employs pointwise group convolutions, significantly enhancing the computational efficiency of the convolution process39,40,41. Additionally, the introduction of channel attention mechanisms facilitates information exchange among different channels, aiding in more comprehensive encoding of information in motion data42. The traditional calculation process for the convolution parameter F is depicted in Eq. (1):

$$F=Kernel\;size * Kernel\;size * C * C^{\prime}$$
(1)

Equation (2) represents the calculation process for the parameter quantity \(F^{\prime}\) of grouped convolutions:

$$F^{\prime}=Kernel\;size * Kernel\;size * \left( {\frac{C}{P}} \right) * \left( {\frac{{C^{\prime}}}{P}} \right) * P$$
(2)

In this context, the term Kernel size represents the size of the convolutional kernel. C and \(C^{\prime}\) refer to the input and output channel quantity in the network, and P indicates the number of groups. In order to further optimize this network structure, ShuffleNet V2 initially eliminates grouped convolution operations in all structures. Subsequently, improvements are made to each unit in the network. Channel split operations are performed on the basic units to partition input channels effectively, and memory usage and memory access are minimized. Finally, channel shuffling operations are performed to ensure information exchange between different branches43. Figure 3 illustrates the foundational and spatial downsampling units of the ShuffleNet V2 module.

Fig. 3
figure 3

Flowchart of the ShuffleNet V2 module.

Full size image

Figure 3 illustrates that in the ShuffleNet V2 module network, X(\(X \in {R^{H \times W \times C}}\)) undergoes some encoding or decoding blocks F to generate the output feature map Y(\(Y \in {R^{H \times W \times C}}\)). When\({X_i}\) (\({X_i} \in {R^{H \times W \times C}}\)) is considered as the input to the i-th ShuffleNet V2 module, the inputs to the two branches in this structure are \({X_{i1}}\) and \({X_{i2}}\). For the basic unit, after the channel split operation, there is \({X_{i1}},{X_{i2}} \in {R^{H \times W \times C/2}}\). For the spatial downsampling unit, there is \({X_{i1}},{X_{i2}} \in {R^{H \times W \times C}}\), resulting in the output feature \({Y_i}\) (\({Y_i} \in {R^{H \times W \times C}}\)). This procedure can be represented by the Eq. (3):

$${Y_i}=\left\{ {\begin{array}{*{20}{c}} {Concat\left( {H\left( {{X_{i1}}} \right),F\left( {{X_{i2}},{W_{i2}}} \right)} \right)} \\ {Concat\left( {F\left( {{X_{i1}},{W_{i1}}} \right),F\left( {{X_{i2}},{W_{i2}}} \right)} \right)} \end{array}} \right.$$
(3)

where W(.) denotes the weight matrix in convolutional calculations, F(.) represents the residual function, Concat indicates the fusion operation, and \(H\left( {{X_i}} \right)\) signifies the identity mapping function. In the basic unit of the ShuffleNet V2 module, it is expressed as Eq. (4):

$$H\left( {{X_{i1}}} \right)={X_{i1}}$$
(4)

\({Y_i}\) (\({Y_i} \in {R^{H \times W \times C}}\)) is then adjusted to a shared \({U_i}\) (\({U_i} \in {R^{H \times W \times C}}\)) after the channel shuffling operation, as indicated by Eq. (5):

$$C^{\prime}=P \times N$$
(5)

A pooling layer is used to reduce the spatial dimensions of feature maps. The common max pooling operation can be expressed as Eq. (6):

$$\hbox{max} pool\left( f \right)\left[ {i,j} \right]={\hbox{max} _{m,n \in window}}f\left[ {i - m,j - n} \right]$$
(6)

f denotes the input feature map, and window defines the local region for the pooling operation.

In order to enhance the representational capacity of Inception V2 for complex data patterns, this work further expands it by introducing the Inception V3 network44,45. This network optimizes the design of V2 convolutional kernels by incorporating larger-sized kernels, aiding in capturing more extensive features. Simultaneously, combining the Inception V3 network with channel attention mechanisms46 allows for the capture of internal relationships in the data, facilitating the recognition of long-distance dependencies in the movement data of aerobics. Figure 4 illustrates the channel attention mechanism.

Fig. 4
figure 4

Schematic diagram of the channel attention mechanism structure.

Full size image

Figure 4 illustrates that this module effectively enhances the feature extraction capability of the ShuffleNet V3 network model. This further improves the recognition results for classifying the movements of aerobics athletes.

Analysis of aerobics movement classification and recognition model based on ShuffleNet V3 with integrated attention mechanism under sports nutrition assistance

In order to delve deeper into the impact of athletes’ performance during the exercise process, this work proposes an aerobics movement classification and recognition model based on ShuffleNet V3. The model integrates sports nutrition assistance and attention mechanisms, as illustrated in Fig. 5.

Fig. 5
figure 5

Schematic diagram of the aerobics movement classification and recognition model based on ShuffleNet V3 with an integrated attention mechanism under sports nutrition assistance.

Full size image

In this model, the first step involves assessing and analyzing the personalized nutritional needs of athletes. Multiple dimensions of data, including physical fitness tests, body metric analysis, and nutritional surveys, are collected to establish an accurate personalized nutritional needs model. Next, ShuffleNet V3 is chosen as the foundational network structure, leveraging its lightweight and high-performance characteristics to process complex exercise data more effectively. A graph attention module is also introduced, assigning different weights to each channel to enhance attention to crucial channels. By integrating sports nutrition, ShuffleNet V3, and attention mechanisms, this DL-based model achieves accurate classification and recognition of aerobics movements.

In the ShuffleNet V3 network model, each layer often produces new channels with varying relevance to key information. Therefore, corresponding weights can be assigned to signals on each channel. The output \({f_{out}} \in {R^{H \times W \times C}}\) of the graph attention module is used as the input for a squeeze operation in this module, achieving global information embedding. Average pooling operations are conducted in the time and spatial dimensions, as shown in Eq. (7):

$${z_q}=\frac{1}{{H \times W}}\sum\limits_{{i=1}}^{H} {\sum\limits_{{j=1}}^{W} {{m_q}\left( {i,j} \right)} }$$
(7)

\({m_c} \in {R^{H \times W}}\) represents the elements of the matrix Z, which is the output of this step. \({m_q}\left( {i,j} \right)\) denotes the output of the graph attention module. Next, a transformation is applied to the output Z, as shown in Eq. (8):

$$S=\sigma \left( {{W_2}\delta \left( {{W_1}Z} \right)} \right)$$
(8)
$$\sigma \left( x \right)=\frac{1}{{1+{e^{ - x}}}}$$
(9)

In Eq. (8), \({W_1}\) and \({W_2}\) correspond to the two weight matrices for the compression and reconstruction of the fully connected layer, respectively. \(\sigma\) indicates the Sigmoid activation function, as shown in Eq. (9). \(\delta\) represents the PReLu activation function.

Finally, the matrix S is multiplied with the input feature map \({f_{out}}\), and the result is added to the original input feature map in a residual manner. This process yields the ultimate output of the channel attention module. By enhancing attention to critical channel information, the model can extract the spatiotemporal features of aerobics athlete movement samples better.

Ultimately, Algorithm 1 presents the pseudocode for the application of ShuffleNet V3 with an integrated attention mechanism to the classification and recognition of aerobics athlete movements.

Algorithm 1
figure a

The pseudocode flow for applying ShuffleNet V3 with integrated attention mechanism to the classification and recognition of aerobics athlete movements.

Full size image

In this section, the studies involving human participants were reviewed and approved by the School of Science of Physical Culture and Sports, Kunsan University Ethics Committee (Approval Number: 2021.022384). The participants provided their written informed consent to participate in this study.

The studies involving human participants were reviewed and approved by the School of Science of Physical Culture and Sports, Kunsan University Ethics Committee (Approval Number: 2021.022384). The participants provided their written informed consent to participate in this study.

  Experimental design and performance evaluation

Datasets collection

In order to validate the effectiveness of the proposed model, the data for this work is sourced from the MultiSports dataset (https://github.com/MCG-NJU/MultiSports) and a self-built dataset. The MultiSports dataset comprises four sports: soccer, basketball, volleyball, and aerobics, with a total of 66 types of movements. The aerobics movements are selected from this dataset, and 2879 samples are obtained after anonymization. Furthermore, the MultiSports and self-built datasets are categorized into training and testing sets in an 8:2 ratio based on data categories for further analysis.

Experimental environment

This work utilizes the TensorFlow 2.0 DL framework Google developed, CUDA 10.2, and a GTX 1060 Ti graphics card to accelerate the program and establish the experimental environment. The operating system used is Windows 10 64-bit, and the programming language is Python 3.7. The Anaconda distribution version 3.0 is chosen. Before installing TensorFlow, Anaconda is installed as a pre-requisite. Anaconda is an open-source environment management tool that comes with the Conda package manager, allowing compatibility with multiple Python environments and easy environment switching.

Parameters setting

The initial learning rate before the optimization is determined as 0.001, Momentum is configured to 0.8, the batch size is configured to 64, and the loss function employed is the cross-entropy loss function. The number of iterations is set to 90, and the batch training size for each iteration (Batch size) is configured to 100.

Performance evaluation

The performance comparison of different models on the MultiSports dataset is shown in Table 1. The proposed model is significantly better than other models in terms of accuracy and F1 score. Specifically, the proposed model achieves an accuracy of 95.11% and an F1 score of 0.8981, while the performance of other models is relatively lower. The accuracy of the CNN model is 82.1%, with an F1 score of 0.75; The accuracy of the LSTM model is 85.6%, with an F1 score of 0.80; Even the lightweight ShuffleNet model has an accuracy and F1 score of only 88.0% and 0.83, respectively. In addition, although the proposed model performs better in terms of performance, its training time is not the longest, indicating that the proposed model can achieve high performance while maintaining high efficiency. This result emphasizes the superiority and adaptability of the proposed model in processing data from different motion scenarios, especially in accurately identifying complex motion actions. This may be attributed to the integrated attention mechanism in the proposed model, which can better capture the key features of motion actions, thereby improving the accuracy and efficiency of classification.

Table 1 Performance comparison of different models on the MultiSports dataset.
Full size table

The performance comparison of different models on the self built dataset is shown in Table 2. Similar to the results on the MultiSports dataset, the proposed model also demonstrates excellent performance on the self-built dataset, with an accuracy of 96.73% and an F1 score of 0.9041, all higher than those of CNN, LSTM, and ShuffleNet models (accuracy of 84.5%, 87.2%, and 89.5%, respectively, F1 scores of 0.78, 0.81, and 0.84, respectively). In addition, the training time of the proposed model is relatively short, which further proves its potential in practical applications. The results on the self-built dataset further validates the robustness and reliability of the proposed model. These datasets typically contain more diverse and challenging motion action samples, but the proposed model is still able to achieve high-precision recognition. This indicates that the proposed model is not only effective in theory, but also has high practical value in practical applications. Thse comparisons suggest that the proposed model provides a new and efficient solution for the field of motion recognition.

Table 2 Performance comparison of different models on self-built datasets.
Full size table

A comparison is made between the algorithm with CNN47, ShuffleNet V248, and model algorithms proposed by Afsar et al. (2023) and Saha et al. (2024). The evaluation is conducted based on accuracy and F1 values on both the MultiSports and self-built datasets.The prediction time required for action recognition by each algorithm is further compared. Figure 6 shows the results.

Fig. 6
figure 6

Prediction efficiency results of different algorithms.

Full size image

Figure 6 shows that with the increase in iteration cycles, the prediction time for all algorithms exhibits a downward trend, likely due to the optimization of model parameters during the training process. Specifically, at the 10th iteration cycle, the proposed model algorithm is 15 milliseconds faster than the model algorithm by Saha et al. (2024), 20 milliseconds faster than the model algorithm by Afsar et al. (2023), 34 milliseconds faster than ShuffleNet V2, and 50 milliseconds faster than CNN. As the iterations continue, this gap gradually decreases, but even at the 90th iteration cycle, our model algorithm remains 13 to 42 milliseconds faster than the other algorithms. This result indicates that the proposed model algorithm maintains high accuracy while also having low time complexity, which is a significant advantage for application scenarios requiring real-time feedback.

Ablation experiments are conducted to analyze model performance. The complete model proposed is compared with only the ShuffleNet V3 algorithm, the ShuffleNet V3 algorithm using smaller convolution kernels, and the traditional CNN algorithm, as shown in Table 3.

Table 3 Ablation experiment results.
Full size table

This work evaluates the impact of different components on the performance of the aerobics athlete action classification and recognition model through a series of ablation experiments. Experimental results show that the model performs best in terms of accuracy and F1 score, achieving 96.73% and 90.41%, respectively. When the attention mechanism is removed from the model, there is a performance decrease with accuracy dropping to 94.21% and F1 score to 87.65%. This indicates the importance of the attention mechanism in enhancing the model’s recognition capability. In further experiments, reducing the convolution kernel size slightly decreases the model’s performance, but it remains at a high level with accuracy at 95.45% and F1 score at 88.93%. This suggests the model has some robustness to changes in kernel size. However, replacing the core structure of the model with a traditional CNN further reduces performance, with accuracy and F1 score dropping to 76.90% and 67.93%, respectively, confirming the advantages of DL models in handling such tasks.

Further case analysis is to analyze the performance of a combined ShuffleNet V3 model with attention mechanism in aerobics athlete action classification and recognition, using data from a self-built dataset involving 7 athletes under the influence of sports nutrition assistance. Table 4 presents the details.

Table 4 Performance comparison of 7 athletes before and after using the model assistance.
Full size table

Table 4 presents an in-depth analysis of performance changes in 7 aerobics athletes before and after receiving personalized nutrition plans in this case study. These athletes vary in age, gender, and years of training, representing a diverse sample group. Using a self-built dataset, this work meticulously records baseline data and improvements in technical scores, physical fitness tests, and action recognition for each athlete post-nutrition.

The application of the personalized nutrition assistance model constructed is based on each athlete’s physiological characteristics, training needs, and personal preferences. Through scientific nutritional interventions, significant improvements are observed across all athletes in technical execution, physical performance, and action recognition accuracy. For instance, ATH001 improves her technical score by 7 points, reaching 92 points, indicating significant progress in the quality of her movement execution. ATH002’s physical fitness test score increases from 70 to 78 points, an improvement of 11.43%, demonstrating accelerated recovery in physical fitness. Notably, emerging athlete ATH004 and young athlete ATH007 both show steady improvement after receiving the nutrition plan, demonstrating their positive response to training and nutritional adjustments. Veteran athlete ATH006, after nutritional adjustments, successfully recovers physically and improves movement precision, underscoring the importance of personalized nutrition plans in maintaining athletes’ competitive states. Additionally, ATH003’s data shows top-tier performance in action recognition, likely attributable to the optimization of her physical condition through the nutrition plan, thereby enhancing training efficiency and performance. Overall, these results highlight the synergistic effects of personalized nutrition assistance and AI technology in enhancing athlete training and performance.

Discussion

Based on the above results, in the MultiSports dataset, the proposed model shows a trend of initially increasing and then stabilizing as the number of iterations grows, and demonstrates excellent performance in accuracy and F1 values. In contrast to alternative algorithms, the proposed model achieves an accuracy of 95.11% (2.66% higher than other algorithms), and an F1 value of 89.81% (significantly outperforming other algorithms). This indicates its more accurate classification recognition for aerobics movements performed by athletes. This aligns with the findings of Wu et al. (2022)49. Additionally, the experimental results on the self-built dataset further validate the superiority of the proposed model across various datasets. With the increase in data volume, the runtime of each algorithm increases, with the proposed model standing out on accuracy and F1 values. It reaches an accuracy of 96.73% and an F1 value of 90.41%, significantly surpassing other algorithms. This aligns with the observations made by Wang et al. (2023)50 and Fu et al. (2023)51. Therefore, this work demonstrates that the model has achieved a more significant performance improvement under sports nutrition assistance. It provides strong support for research in the field of sports science.

Compared to existing research, this work not only achieves breakthroughs in technical methods but also surpasses traditional CNN baseline algorithms in accuracy and F1 scores, achieving higher recognition accuracy. Additionally, by integrating athletes’ fitness tests, physiological monitoring, and survey data, this work establishes a comprehensive personalized nutrition model, providing athletes with more precise nutritional guidance. The work also constructs its own dataset, further enhancing the model’s generalizability and adaptability. The methods of this work also emphasize real-time feedback and personalized recommendations, offering athletes and coaches immediate performance monitoring and nutrition adjustment solutions. Moreover, the work enhances the interpretability of the model, making the decision-making process more transparent and increasing the model’s credibility. These innovations and improvements contribute significantly to the fields of sports nutrition and human activity recognition both theoretically and practically.

Through the validation of the constructed model, it is highlighted that the proposed model exhibits significant performance advantages across different datasets, achieving more accurate classification recognition for aerobics movements. This provides a new and more effective method for advancing the training and performance of aerobics athletes. Meanwhile, it offers valuable theoretical and empirical foundations for research in the field of sports science. However, this work also has some potential limitations, which provide directions for future research. The dataset used may have limitations in terms of sample size and representativeness. This suggests that the generalizability of the model and its applicability in different populations and environments may be limited. Future research should consider introducing larger and more diverse datasets to improve the model’s universality and accuracy. Although the work employs personalized nutritional assessments, it may still not cover all individual differences among athletes. Future research needs to comprehensively consider athletes’ individual characteristics such as age, gender, genetics, and training history, and how these factors interact with performance and nutritional needs. While the work combined sports nutrition assistance, there is still a need for further exploration into the specific mechanisms of nutritional factors, optimal nutrient ratios, and the best timing for nutritional interventions. Future research can delve into these factors more meticulously to provide more scientific sports nutrition plans. The current study primarily relies on the integration analysis of exercise and nutrition data. Future research can consider introducing various data sources such as video analysis, sensor data, and physiological monitoring data to improve the accuracy and comprehensiveness of athlete movement recognition and performance analysis through multimodal data fusion technology. Larger-scale and more representative datasets should be constructed to improve the model’s generalizability and accuracy. Additionally, big data and machine learning techniques could be utilized to deeply explore data to discover more key factors influencing athlete performance. More refined personalized training and nutrition plans can be developed. These plans should consider individual differences among athletes and changes in nutritional needs during different training stages to provide more precise assistance to athletes. Moreover, research methods and theories from multiple disciplines such as sports science, nutrition, psychology, and bioinformatics could be integrated to comprehensively analyze various factors affecting athlete performance and their interactions. Systems for real-time monitoring of athletes’ physiological status, performance, and nutritional intake could be developed to provide coaches and athletes with real-time feedback and adjustment suggestions through real-time data analysis. New DL architectures and algorithms could be explored to improve the model’s recognition capability and robustness in complex environments. Additionally, it is essential to research how to improve the interpretability of the model so that coaches and athletes can better understand the model’s decision-making process. Addressing these limitations and exploring future research directions can further advance the intersection of sports science and AI fields, providing more effective support for improving athlete performance and health levels.

Conclusion

Research contribution

This work has achieved satisfactory results by applying the aerobics movement classification and recognition model based on ShuffleNet V3 with an integrated attention mechanism and sports nutrition assistance. Experimental results on the MultiSports dataset and the self-built dataset demonstrate that the proposed model exhibits higher accuracy and F1 values in the classification and recognition of aerobics movements than traditional algorithms and models proposed by relevant scholars. It has an accuracy exceeding 95%. This accomplishment not only enhances the technical performance of the model but also provides a new paradigm for the integration of DL and sports nutrition.

Future works and research limitations

While the methods employed in the current study have improved the accuracy of aerobics athlete movement classification to some extent, significant limitations still exist. First, the work heavily relies on the analysis of specific datasets, potentially limiting the model’s applicability to different scenarios and populations. Second, despite considering personalized nutrition assessments, the work may not fully account for individual differences among athletes, such as genotype, phenotype, and lifestyle factors, which can significantly impact athlete performance and nutritional responses.

To address these limitations, future research can focus on several aspects. First, by expanding the scale and diversity of datasets, the model’s generalizability can be enhanced to adapt to broader application scenarios and populations. Second, in-depth research on the impact of individual differences on athlete performance can be conducted to develop more personalized training and nutrition plans that consider factors such as genotype, lifestyle, and psychological state. Third, the optimal timing and duration of nutritional interventions can be explored and nutrition strategies can be adjusted based on athletes’ physiological and psychological states. Finally, new DL architectures and algorithms can be explored to improve the model’s recognition capabilities and robustness, while enhancing its interpretability to better understand the decision-making process for coaches and athletes. Existing research explores the effects of factors such as exercise training, nutrition intake, and technical support on athlete performance. However, most current studies rely on small sample sizes or data collected in specific environments, limiting their applicability across diverse populations and varied settings. Furthermore, there is a significant research gap in the existing literature regarding the integration of personalized nutrition with advanced AI technologies, particularly DL algorithms, to optimize athlete training and recovery. Through these approaches, future research can provide more precise and personalized support for athletes, effectively enhancing their training outcomes and competitive performance.