AI and Statistical Theory–Driven Wearable System for Intelligent Recognition of Professional Table Tennis Skills

Abstract

In traditional table tennis instruction, beginners typically rely on professional coaches to guide their motor skills, a process that depends on experience and is time‐consuming and labor‐intensive. Addressing this issue, this paper proposes a sensor‐based wearable system for automatically monitoring and identifying six table tennis motor skills. The system employs an embedded platform equipped with inertial sensors to collect multidimensional data from the athlete′s wrist during gameplay. Feature engineering and principal component analysis (PCA) are applied to preprocess and extract features from the raw data, effectively reducing dimensionality while preserving critical information. For skill recognition, an improved support vector machine (SVM) model is proposed. Its performance is compared against traditional convolutional neural network (CNN) models. Experimental results demonstrate the following: (1) The designed table tennis player skill monitoring system effectively captures in‐game data and enables athlete health monitoring; (2) the proposed improved SVM model demonstrates outstanding performance in technical skill recognition, achieving an average recognition accuracy of 97.77%. This represents a 5.07% improvement over the 92.70% accuracy of traditional CNN models, enabling more precise identification of various table tennis skills. The successful development of this system provides an effective data support and analysis tool for scientific athlete training and technical enhancement.

1. Introduction

As physical exercise increasingly becomes a core means of maintaining health, leveraging cutting-edge technology to enhance athletic performance has emerged as a significant trend [1]. The integration of artificial intelligence and wearable sensor technology enables objective monitoring of athletes′ performance metrics [2], powerfully advancing sports science [3]. As a globally popular traditional sport, table tennis demands exceptional skill due to its intricate techniques and high-speed nature [4, 5]. However, beginners often face skill development bottlenecks without professional guidance, leading to the ingraining of incorrect movements [6]. Traditional coaching heavily relies on subjective evaluations by instructors. This approach not only lacks objective, quantifiable real-time feedback, compromising assessment accuracy, but also renders self-directed training inefficient [7]. Therefore, developing an objective, intelligent table tennis coaching support system to overcome the limitations of traditional coaching models has become an urgent priority.

To address this challenge, Faubert believes that a system capable of accurately identifying professional athletic skills is needed [8]—yet this represents a gap in current technological applications. On one hand, Fuller et al. argue that consumer-grade wearable devices currently available on the market can only track generic health metrics such as heart rate [9], falling far short of the refined analytical demands for technically demanding sports like table tennis. On the other hand, even within academia, established Human Activity Recognition (HAR) frameworks primarily focus on distinguishing macrolevel activities like walking or running [10]. Their technical architecture struggles to directly identify the subtle, high-speed, and technically similar specialized movements inherent in sports.

In recent years, research in the field of HAR has shifted significantly toward deep learning approaches [11]. Models represented by convolutional neural networks (CNNs) [12] and long short-term memory networks (LSTMs) [13] have become the mainstream techniques for tackling such problems. Leveraging their powerful end-to-end learning capabilities, these models can automatically extract complex abstract features from raw sensor data, achieving outstanding performance across numerous general activity recognition tasks. However, when directly applying these advanced models to the fine-grained recognition of table tennis skills, He et al. suggest that certain limitations also exist [14]. Shi et al. designed a system for identifying and stratifying the athletic skills of table tennis players, demonstrating the feasibility and research value of this approach [15]. First, deep learning models [16] heavily rely on large-scale, high-quality annotated datasets, which are costly and difficult to obtain in professional sports domains [17]. Second, the kinematic differences between distinct stroke techniques are extremely subtle. Third, the kinematic differences between various hitting techniques are extremely subtle. Shaheen et al. argue that generic deep learning architectures are prone to overfitting when data is limited, making it difficult to consistently capture decisive key features [18].

It is precisely these challenges that have driven us to explore an alternative technical approach. The core innovation of this study lies in designing and implementing a HAR technology pipeline deeply optimized for recognizing the complex skills involved in table tennis. This field utilizes wearable inertial measurement unit (IMU) sensors [19] and machine learning [20] techniques to classify human movements. In other related fields, such as autonomous unmanned aerial vehicle (UAV) systems, similar challenges in using multisensor data for state recognition, context monitoring, and trajectory prediction have been extensively explored [21–24]. More broadly, intelligent systems based on sensor data and optimization algorithms have also demonstrated their application value in complex scenarios such as intelligent firefighting and emergency logistics [25, 26]. We first designed a lightweight wearable data acquisition system to capture key kinematic parameters during an athlete′s stroke [27]. Subsequently, through specific data processing workflows and targeted feature engineering, we constructed multidimensional representations capable of capturing the unique kinematic characteristics of each stroke. Principal component analysis (PCA) [28] was then employed for efficient dimensionality reduction. Finally, we deployed an enhanced support vector machine (SVM) model [29] for skill prediction, demonstrating superior recognition performance over traditional CNNs [30] for this specific task. This research culminated in a practical system specifically designed for the demanding application of table tennis skill recognition. By providing objective, quantifiable feedback on technical execution, it not only overcomes the limitations of traditional subjective evaluations but also paves the way for data-driven modern training methodologies.

The study is structured into three main segments:

Firstly, the acquisition and preprocessing of table tennis players′ physical parameters.

Secondly, feature extraction of table tennis players′ movement skills.

Finally, recognition and prediction of table tennis players′ professional movement skills.

2. Materials and Methods

2.1. Data Acquisition

During the experimental phase, motion data was collected from a total of 10 table tennis athletes. Our data acquisition system continuously recorded each participant′s motion data at a sampling rate of 100 Hz for 10 s. The data acquisition and preprocessing section includes the design of the athlete data acquisition device and the data and processing flow.

2.1.1. Data Acquisition Device Design

The wearable device utilized in this research must encompass several crucial features, such as offline functionality, shock resistance, and a lightweight design. The table tennis sports information acquisition device introduced in this paper is composed of three primary modules: a power module, an IMU sensor (a gyroscope module, an accelerometer module), and a central processor. The specific specifications and parameters are shown in Table 1. The power supply module fuels the entire data acquisition system, whereas the gyroscope and accelerometer modules are tasked with gathering angle, velocity, and acceleration data along the XYZ axis during the utilization of the table tennis racket. The central processing unit (CPU) module oversees the operation of each component and handles data storage.

Table 1 Wearable device hardware component specifications.

Module	Model/type	Key specifications
Power module	3.7 V LiPo battery with a TP4056 management board	Battery chemistry: Lithium polymer Nominal voltage: 3.7 V Charging: 5 V micro-USB input via TP4056 IC Output: Regulated to 3.3 V
Gyroscope module	MPU-6050 (embedded gyroscope)	Sensing axes: Three axes (X, Y, Z) Measurement range: Configured to ±1000°/s (degrees per second) ADC resolution: 16-bit Output data rate: 100 Hz
Accelerometer module	MPU-6050 (embedded accelerometer)	Sensing axes: Three axes (X, Y, Z) Measurement range: Configured to ±8 g ADC resolution: 16-bit Output data rate: 100 Hz
Central processor	Atmel ATmega328P	Architecture: 8-bit AVR Clock speed: 16 MHz Program memory: 32 KB Flash Communication interfaces: I2C, SPI, UART

Figure 1 illustrates the structural block diagram of the table tennis information collection system developed in this research.

[IMAGE OMITTED. SEE PDF]

As illustrated in Figure 1, this table tennis motion capture system consists of two main components: the “wearable sensor unit” and the “data processing terminal.” The wearable unit centers around a CPU, which manages the power system and coordinates the IMU sensor within the angular module. This IMU sensor integrates an accelerometer and gyroscope, enabling precise capture of key kinematic parameters such as acceleration, angular velocity, and posture in three-dimensional space. After processing the raw data, the CPU transmits it via the communication system to the personal computer (PC) serving as the data processing terminal. The PC then handles subsequent data collection, analysis, and motion pattern recognition. This integrated design ensures high efficiency and precision throughout the entire chain—from motion capture to intelligent analysis.

2.1.2. Data Preprocessing

During the data acquisition process, the system is prone to electromagnetic interference (EMI), resulting in the introduction of noise into the collected signals and subsequent deviation of the data from its true values. Such inaccurate data can significantly impair the precision of the model. Hence, it is crucial to carry out data preprocessing on the raw data that has been gathered. A set of raw data collected is shown in Figure 2.

[IMAGE OMITTED. SEE PDF]

This diagram displays the acceleration, velocity, and angular data of a table tennis player′s racket swing along the X, Y, and Z axes over approximately 10 s. Overall, the object exhibits a primary, periodic reciprocating motion or rotation along the z-axis, which may represent its dominant motion pattern. Secondary periodic motion is also present along the x-axis, while movement along the y-axis is relatively minor.

Data cleaning in this paper primarily consists of three parts: anomalous data rejection, missing data interpolation, and high-frequency noise filtering.

Based on statistical principles [31], the first-order differences X_i of the collected motion data samples x_i should follow a normal distribution. We calculated the mean (EX) and standard deviation (σ) of the first-order difference sequence. Data points X_i falling within the 3σ criterion interval (EX −3σ, EX +3σ) were retained as normal data. Conversely, X_i data points exceeding this interval were treated as outliers and discarded. Simultaneously, the x_i + 1 data point associated with these anomalous X_i values is also removed, ensuring all anomalous data is completely eliminated. After removing outliers, we employed Newton′s interpolation method [32] to fill data gaps. This approach excels at capturing the inherent characteristics of rapid ping-pong ball motion by defining local curves using multiple surrounding data points. It provides a more physically plausible reconstruction of the athlete′s trajectory, ensuring signal continuity and fidelity post-preprocessing [33].

In the realm of table tennis sports data acquisition, there often exists a degree of high-frequency noise during the data collection process. To address this challenge, researchers have employed an adaptive smoothing filtering technique for data processing. This method involves the derivation of a new output value by assigning a weight to both the current sampling value and the previous output value. Through this weighted approach, the current output value is obtained, thereby mitigating the impact of high-frequency noise on the acquired data. This adaptive smoothing filtering method serves to enhance the quality of the processed data, ensuring a more accurate representation of the underlying trends and patterns in the table tennis sports data. The smoothing coefficient m in the formula is a critical parameter that balances noise suppression and tracking of true signal variations. After multiple comparative experiments, we ultimately selected m = 0.5. This value effectively filters out high-frequency noise while maximally preserving the waveform characteristics of the athlete′s actual movements [34]. The filtered output is shown in Equation (1): 1 $\begin{array}{}Y\left(n\right)=m\times X\left(n\right)+\left(1-m\right)\times Y\left(n-1\right).\end{array}$

In Equation (1), Y(n) is the filtered output value of the filter. X(n) is the current sample value; m is the filter coefficient, whose value range is [0, 1]. Y(n − 1) is the output value of the previous filter.

Use the data preprocessing in this section to process athlete data, so that the data can better reflect the actual situation of athletes when making professional movements, and also make the data smoother. The effect is shown in Figure 3.

[IMAGE OMITTED. SEE PDF]

This graph compares the raw motion data (blue curve) before filtering with the filtered data (orange curve) across the X, Y, and Z axes. It is evident that the filtering algorithm effectively eliminates high-frequency noise and abnormal spikes present in the original signal, yielding smoother data that more accurately reflects the true motion trends.

Athlete movement information is typically transmitted in the form of a data stream, necessitating preprocessing before direct processing. Data segmentation is a crucial step in this process, often achieved by applying a sliding window to the collected data. By incorporating window processing, the collected data can be divided into distinct segments based on the window function, laying the groundwork for subsequent feature engineering processes. Notably, researchers have observed that setting the overlap rate of the window function to 50% proves effective in numerous experimental studies involving sliding window functions. Martin et al. argue that sampling the sliding window spectrum on a specific time-frequency grid alone suffices to fully reconstruct the original signal [35]. Christoph et al. compare the upper bounds of scores across different window sets against the currently found optimal solution, thereby significantly enhancing search efficiency [36]. The window setting of the window function directly affects the recognition of actions. When the window setting is too large, it may cause a delay in action recognition or include different types of actions, causing training errors. When the window function is set too small, it may not include a complete action, thus introducing recognition errors.

A comprehensive statistical analysis was conducted on the duration of each action in the single table tennis sport. A default length of 2 s was chosen as the duration for each action segment, equivalent to a sliding window. Considering a motion data sampling frequency of 100 Hz, the width of the sliding window was set to 200 samples, with a 50% overlap rate. This configuration is depicted in Figure 4, illustrating the sliding window function employed in this study. By standardizing the length of action data through the sliding window function, uniformity and consistency in action data were achieved, facilitating subsequent statistical analysis and feature extraction for motion recognition.

[IMAGE OMITTED. SEE PDF]

2.2. Motion Feature Extraction Based on PCA

In table tennis, the standard benchmark movements predominantly encompass the forehand attack, forehand push, forehand loop, backhand attack, backhand push, and backhand chop. Feature extraction is carried out for these six movement skills.

2.2.1. Characteristic Engineering Structural Features

In feature engineering for action recognition, the core challenge lies in distinguishing between “action patterns” and “action intensity.” The same technical movement often exhibits vastly different signal magnitudes due to variations in an athlete′s exertion. If a model learns only magnitude-dependent features like peak acceleration, it is highly prone to misclassifying action intensity as action type, leading to insufficient generalization capabilities.

To overcome this issue, our feature engineering strategy suppresses the influence of magnitude while enhancing pattern recognition. We extract scale-invariant, shape-descriptive features. These features precisely capture the inherent kinematic patterns of movements by calculating the relative proportions and distribution shapes of signal waveforms, rather than their instantaneous force. The extracted features can be conceptualized into two categories:

• a.

  Foundational statistical features:

This category includes fundamental indicators describing the central tendency and dispersion of signals. Although some indicators are inherently amplitude-related, they provide a benchmark for understanding overall movement, specifically including the mean value of motion data in the XYZ direction and its composite direction, the variance of motion data in the XYZ direction and its composite direction, the maximum and minimum values of motion data in the XYZ direction and its composite direction, the peak and valley values of motion data in the XYZ direction and its composite direction, the mean square value of motion data in the XYZ direction and its composite direction, and the root mean square of motion data in the XYZ direction and its composite direction.

• b.

  Scale-invariant shape-descriptive features:

This category of features includes those that mitigate the impact of amplitude by capturing the inherent shape of the signal waveform. These features are typically calculated as ratios of different statistical measures, making them less sensitive to overall signal scaling. They primarily include the crest factor, pulse factor, margin factor, kurtosis factor, waveform factor, skewness, and kurtosis. These factors are computed for both the XYZ direction and the synthetic direction of motion information. Additionally, the crest factor is evaluated separately for the XYZ direction and the synthetic direction. The list of extracted motion features is shown in Table 2.

Table 2 List of extracted motion features.

Feature name as per original text	Calculation scope	Calculation formula
(Mean value)	Intensity of factor fluctuations in the signal	$\overline{X}=\frac{1}{N}\sum _{i=1}^N{X}_i$
(Variance)	Average level of the mean factor in the signal	$S^2=\frac{1}{N}\sum _{i=1}^N{\left(X_i-\overline{X}\right)}^2$
(Maximum values)	Maximum value factor in the signal	Xmax = max(Xi)
Minimum values	Minimum value factor in the signal	Xmin = min(Xi)
(Peak and valley values)	Fluctuation range of a certain factor in the signal	Xpv = Xmax − Xmin
(Mean square value)	Signal energy	$X_{ms}=\frac{1}{N}\sum _{i=1}^N{X}_i^2$
(Root mean square)	RMS value of the signal	$X_{rms}=\sqrt{\frac{1}{N}\sum _{i=1}^N{X}_i^2}$
(Crest factor)	Extremity of peaks in the waveform	$X_{cf}=\frac{X_{\max }-X_{\min }}{\sqrt{1/N{\sum }_{i=1}^N{X}_i^2}}$
(Pulse factor)	Extremity of peaks in the waveform	$X_{pf}=X_{\max }-X_{\min }/1/N{\sum }_{i=1}^N\left|X_i\right|$
(Margin factor)	Extremity of peaks in the waveform	$X_{mf}=\frac{X_{\max }-X_{\min }}{{\left(1/N{\sum }_{i=1}^N\sqrt{\left|X_i\right|}\right)}^2}$
(Kurtosis factor)	Flatness of the waveform	$X_{kf}=\frac{1/N{\sum }_{i=1}^N{X}_i^4}{\sqrt{1/N{\sum }_{i=1}^N{X}_i^2}}$
(Waveform factor)	Flatness of the waveform	$X_{wf}=\frac{\sqrt{1/N{\sum }_{i=1}^N{X}_i^2}}{1/N{\sum }_{i=1}^N\left|X_i\right|}$
(Skewness)	The position of the wave peak relative to the center	$SK=\frac{1/N{\sum }_{i=1}^N{\left(X_i-\overline{X}\right)}^3}{{\left(1/N{\sum }_{i=1}^N{\left(X_i-\overline{X}\right)}^2\right)}^{32/}}$
Kurtosis	The steepness of the waveform	$K=\frac{1/N{\sum }_{i=1}^N{\left(X_i-\overline{X}\right)}^4}{\left(1/N{\sum }_{i=1}^N{\left(X_i-\overline{X}\right)}^2\right)}-3$

2.2.2. PCA Method Feature Extraction

Within the scope of machine learning, overfitting has a direct impact on the precision of classification and heightens the workload of machine learning tasks. Consequently, implementing data dimensionality reduction techniques becomes exceedingly important. PCA stands as a quintessential unsupervised dimensionality reduction technique. It condenses numerous indicators into several principal components, which are linear combinations of the original variables and orthogonal to each other. These components capture the majority of valuable information present in the original dataset. Given that the motion feature data extracted in this study—acceleration, angular velocity, and angle—along with multiple statistics (such as mean, variance, extreme values, and root mean square) and dimensionless features (such as peak factor and skewness) in the X, Y, Z, and composite directions collectively form a 180-dimensional feature vector, it becomes imperative to employ the PCA model to reduce the dimensionality of the extracted athletic features, thereby constructing efficacious athletic features.

To enhance the accuracy of machine learning and mitigate the risk of overfitting associated with high-dimensional (180-dimensional) feature sets, we employ PCA for dimensionality reduction [37]. We set the information retention threshold for principal components at 95%. This means the number of principal components selected adequately explains 95% of the variance in the original 180-dimensional feature data. This data-driven approach ultimately constructs a new, optimized 23-dimensional feature vector that captures the most critical information while discarding redundant data and noise [38].

2.3. Motor Skill Recognition Model

In the SVM model, the dataset is partitioned by identifying a classification method, and the associated optimization equation is simplified as demonstrated in Equation (2): 2 $\begin{array}{}\left\{\begin{array}{c}\min {‖\omega ‖}^2\hfill \\ s.t.y_i\left({\omega }^{T}x+b\right)\ge 1\hfill \end{array}\right..\end{array}$

In Equation (2), ‖ω‖ denotes the L2 norm of the hyperplane′s normal vector, whose minimum value represents the maximization of the classification margin; y_i denotes the class label of sample x_i (typically +1 or −1 for binary classification problems); ωx + b = 1 and ωx + b = −1 define the classification hyperplane boundary of the SVM, where ω is the hyperplane′s normal vector, x is the input feature vector, and b is the bias term.

The SVM model hyperplane is shown in Figure 5.

[IMAGE OMITTED. SEE PDF]

In order to enhance the classification efficacy of SVMs when dealing with nonlinear datasets, we introduce a Gaussian kernel function. Although the SVM model can be applied to the classification and recognition of athletic status, the classification of athletic skills is not a simple binary division. In view of this, the optimization of the traditional SVM model is particularly necessary. Considering that the motor skills of table tennis players mainly involve six finite states, the adoption of an SVM model based on directed acyclic graphs plays a crucial role in overcoming the multiclassification difficulties inherent in the recognition of table tennis players′ motor skills. This study employs the directed acyclic graph support vector machine (DAGSVM) model to classify and recognize the six types of athlete movement skills during their performance. Figure 6 illustrates the classification approach utilizing directed acyclic graphs for identifying the six action skills of table tennis players.

[IMAGE OMITTED. SEE PDF]

As shown in Figure 6, the DAGSVM model employs a DAGSVM topology. Within this framework for classifying six table tennis skills, unknown samples are evaluated by traversing paths from nodes. Each node represents a binary SVM classifier, with the final classification determined after a sequence of five binary decisions.

3. Experimental Results and Analyses

Figure 7 illustrates the classification confusion matrix, contrasting the SVM model built upon feature construction with the traditional CNN model.

[IMAGE OMITTED. SEE PDF]

The accuracy rate of any classification in this classifier is defined as the ratio of correctly identified instances to the total predictions made for that classification. This accuracy rate is expressed by Equation (3) as follows: 3 $\begin{array}{}P=\frac{TP}{TP+FP}\times 100\%.\end{array}$

In Equation (3), TP represents the count of accurately predicted categories within the type, while FP denotes the count of predicted categories within the category that do not correspond to the true category.

The recognition accuracy of the conventional CNN for the six action skills is additionally contrasted with the recognition accuracy of the feature-based SVM developed in this study for the same six action skills. The comparative results are illustrated in Table 3.

Table 3 Classification results for fix states.

Name of professional skills	DAGSVM accuracy based on feature construction	Traditional convolutional neural network model accuracy
Forehand attack	100%	93.88%
Forehand push	98.99%	89.85%
Forehand shop	98.98%	88.57%
Backhand attack	93.46%	92.16%
Backhand push	100%	97.85%
Backhand chop	95.19%	93.88%
Mean	97.77%	92.70%

As depicted in Table 3, the model constructed in this article is better than the traditional CNN model in the recognition accuracy of all six action skills. In particular, it has achieved 100% accuracy in the recognition of forehand attack and backhand push action skills. However, among the two action skills of forehand push and forehand shop, which have poor recognition rates with the traditional CNN model, the model in this article has also been improved, but in the recognition of the two action skills of backhand attack and backhand chop, the recognition accuracy of the two is basically the same. Overall, the model constructed in this article improves the recognition accuracy by 5.07% compared with the traditional model.

4. Conclusion

This study addresses the critical need for monitoring and analyzing the diverse motor skills employed by table tennis players. To this end, a comprehensive motor skill recognition system is proposed, encompassing the acquisition, processing, and recognition of players′ motion data. The primary contributions of this study can be outlined as follows:

• 1.

  Development of a sophisticated data measurement system tailored specifically for table tennis players. This system enables precise and detailed measurement of athletes′ movement patterns and performance metrics during gameplay.

• 2.

  Establishment of a robust motion feature extraction model that leverages both feature engineering techniques and PCA. This model effectively captures and synthesizes the salient features inherent in table tennis players′ motion data, facilitating subsequent analysis and recognition processes.

• 3.

  Introduction of a novel motor skill recognition model built upon the foundation of the table tennis player′s motion information. By utilizing sophisticated machine learning algorithms and pattern recognition methodologies, this model facilitates precise and instantaneous identification of diverse motor skills demonstrated by table tennis players during gameplay.

In conclusion, the system proposed by this research represents a significant breakthrough in the field of sports technology. By providing deep insights into the complex dynamics of table tennis, it lays the foundation for enhanced training methods and athlete development strategies. However, while acknowledging its achievements, we must also recognize its current limitations. The system′s feature extraction and recognition models are specifically designed for the kinematic patterns of six fundamental table tennis strokes. This implies that its applicability to other racket sports, such as badminton or tennis, remains an open question awaiting verification. Therefore, future research should focus on expanding the system′s recognition capabilities and testing its generalization performance across broader athletic scenarios. This will ultimately mark a crucial step toward integrating wearable technology with machine learning to create more efficient and personalized athletic training tools.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.

Conflicts of Interest

The authors declare no conflicts of interes.

Author Contributions

Yafei Song and Shuning Zhang contributed equally to this work and are considered co-first authors.

Funding

This work is funded by the Natural Science Basic Research Program of Shaanxi (2022JQ-593) and the Key R&D Program of Shaanxi Provincial Department of Science and Technology (2022GY-089).

References

1 Birrer D., Röthlin P., and Morgan G., Mindfulness to Enhance Athletic Performance: Theoretical Considerations and Possible Impact Mechanisms, Mindfulness. (2012) 3, no. 3, 235–246, https://doi.org/10.1007/s12671-012-0109-2, 2-s2.0-84865786170.

2 Shajari S., Kuruvinashetti K., Komeili A., and Sundararaj U., The Emergence of AI-Based Wearable Sensors for Digital Health Technology: A Review, Sensors. (2023) 23, no. 23, https://doi.org/10.3390/s23239498, 38067871.

3 Bourdon P. C., Cardinale M., Murray A., Gastin P., Kellmann M., Varley M. C., Gabbett T. J., Coutts A. J., Burgess D. J., Gregson W., and Cable N. T., Monitoring Athlete Training Loads: Consensus Statement, International Journal of Sports Physiology and Performance. (2017) 12, no. s2, S2-161–S2-170, https://doi.org/10.1123/IJSPP.2017-0208, 2-s2.0-85019241404.

4 Zhang H., Fu Z., and Shu K.-I., Recognizing Ping-Pong Motions Using Inertial Data Based on Machine Learning Classification Algorithms, IEEE Access. (2019) 7, 167055–167064, https://doi.org/10.1109/ACCESS.2019.2953772.

5 Song Y., Sun X., and Shi Z., Improvement of Traditional Electrical Engineering System Based on Statistical Theory, Proceedings of the 2025 IEEE 2nd International Conference on Energy and Electrical Engineering (EEE), 2025, IEEE, 1–4, https://doi.org/10.1109/EEE64897.2025.11162895.

6 Yao H., de Siqueira A. G., Bafna A., Peterkin D., Richards J., Rogers M. L., Foster A., Galynker I., and Lok B., A Virtual Human Interaction Using Scaffolded Ping-Pong Feedback for Healthcare Learners to Practice Empathy Skills, Proceedings of the 22nd ACM International Conference on Intelligent Virtual Agents, 2022, ACM, 1–8, https://doi.org/10.1145/3514197.3549621.

7 Zhang P., Ward P., Li W., Sutherland S., and Goodway J., Effects of Play Practice on Teaching Table Tennis Skills, Journal of Teaching in Physical Education. (2012) 31, no. 1, 71–85, https://doi.org/10.1123/jtpe.31.1.71.

8 Faubert J., Professional Athletes Have Extraordinary Skills for Rapidly Learning Complex and Neutral Dynamic Visual Scenes, Scientific Reports. (2013) 3, no. 1, https://doi.org/10.1038/srep01154, 2-s2.0-84873637318.

9 Fuller D., Colwell E., Low J., Orychock K., Tobin M. A., Simango B., Buote R., van Heerden D., Luan H., Cullen K., Slade L., and Taylor N. G. A., Reliability and Validity of Commercially Available Wearable Devices for Measuring Steps, Energy Expenditure, and Heart Rate: Systematic Review, JMIR mHealth and uHealth. (2020) 8, no. 9, e18694, https://doi.org/10.2196/18694.

10 Tang Y., Zhang L., Min F., and He J., Multiscale Deep Feature Learning for Human Activity Recognition Using Wearable Sensors, IEEE Transactions on Industrial Electronics. (2023) 70, no. 2, 2106–2116, https://doi.org/10.1109/TIE.2022.3161812.

11 LeCun Y., Bengio Y., and Hinton G., Deep Learning, Nature. (2015) 521, no. 7553, 436–444, https://doi.org/10.1038/nature14539, 2-s2.0-84930630277.

12 Kattenborn T., Leitloff J., Schiefer F., and Hinz S., Review on Convolutional Neural Networks (CNN) in Vegetation Remote Sensing, ISPRS Journal of Photogrammetry and Remote Sensing. (2021) 173, 24–49, https://doi.org/10.1016/j.isprsjprs.2020.12.010.

13 Sherstinsky A., Fundamentals of Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) Network, Physica D: Nonlinear Phenomena. (2020) 404, 132306, https://doi.org/10.1016/j.physd.2019.132306.

14 He J., Chen J. N., Liu S., Kortylewski A., Yang C., Bai Y., and Wang C., TransFG: A Transformer Architecture for Fine-Grained Recognition, Proceedings of the AAAI Conference on Artificial Intelligence. (2022) 36, no. 1, 852–860, https://doi.org/10.1609/aaai.v36i1.19967.

15 Shi Z., Jia Y., Shi G., Zhang K., Ji L., Wang D., and Wu Y., Design of Motor Skill Recognition and Hierarchical Evaluation System for Table Tennis Players, IEEE Sensors Journal. (2024) 24, no. 4, 5303–5315, https://doi.org/10.1109/JSEN.2023.3346880.

16 Janiesch C., Zschech P., and Heinrich K., Machine Learning and Deep Learning, Electron Markets. (2021) 31, no. 3, 685–695, https://doi.org/10.1007/s12525-021-00475-2.

17 Rajendran S., Meert W., Giustiniano D., Lenders V., and Pollin S., Deep Learning Models for Wireless Signal Classification With Distributed Low-Cost Spectrum Sensors, IEEE Transactions on Cognitive Communications and Networking. (2018) 4, no. 3, 433–445, https://doi.org/10.1109/TCCN.2018.2835460.

18 Shaheen F., Verma B., and Asafuddoula M., Impact of Automatic Feature Extraction in Deep Learning Architecture, Proceedings of the 2016 International Conference on Digital Image Computing: Techniques and Applications (DICTA), 2016, IEEE, 1–8, https://doi.org/10.1109/DICTA.2016.7797053, 2-s2.0-85011106060.

19 Zhao J., A Review of Wearable IMU (Inertial-Measurement-Unit)-Based Pose Estimation and Drift Reduction Technologies, Journal of Physics: Conference Series. (2018) 1087, 042003, https://doi.org/10.1088/1742-6596/1087/4/042003, 2-s2.0-85054921736.

20 Toreini E., Aitken M., Coopamootoo K., Elliott K., Zelaya C. G., and Van Moorsel A., The Relationship Between Trust in AI and Trustworthy Machine Learning Technologies, Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, 2020, ACM, 272–283, https://doi.org/10.1145/3351095.3372834.

21 Shi Z., Shi G., Zhang J., Wang D., Xu T., Ji L., and Wu Y., Design of UAV Flight State Recognition System for Multisensor Data Fusion, IEEE Sensors Journal. (2024) 24, no. 13, 21386–21394, https://doi.org/10.1109/JSEN.2024.3394883.

22 Zhang J., Shi Z., Zhang A., Yang Q., Shi G., and Wu Y., UAV Trajectory Prediction Based on Flight State Recognition, IEEE Transactions on Aerospace and Electronic Systems. (2024) 60, no. 3, 2629–2641, https://doi.org/10.1109/TAES.2023.3303854.

23 Shi Z., Zhang J., Shi G., Ji L., Wang D., and Wu Y., Design of a UAV Trajectory Prediction System Based on Multi-Flight Modes, Drones. (2024) 8, no. 6, https://doi.org/10.3390/drones8060255.

24 Shi Z., Zhang J., Shi G., Zhu M., Ji L., and Wu Y., Autonomous UAV Safety Oriented Situation Monitoring and Evaluation System, Drones. (2024) 8, no. 7, https://doi.org/10.3390/drones8070308.

25 Li N., Shi Z., Jin J., Feng J., Zhang A., Xie M., Min L., Zhao Y., and Lei Y., Design of Intelligent Firefighting and Smart Escape Route Planning System Based on Improved Ant Colony Algorithm, Sensors. (2024) 24, no. 19, https://doi.org/10.3390/s24196438, 39409478.

26 Zhang Y., Xu W., Ye H., and Shi Z., A Two-Stage Optimization Framework for UAV Fleet Sizing and Task Allocation in Emergency Logistics Using the GWO and CBBA, Drones. (2025) 9, no. 7, https://doi.org/10.3390/drones9070501.

27 Kos M. and Kramberger I., A Wearable Device and System for Movement and Biometric Data Acquisition for Sports Applications, IEEE Access. (2017) 5, 6411–6420, https://doi.org/10.1109/ACCESS.2017.2675538, 2-s2.0-85028004789.

28 Hasan B. M. and Abdulazeez A. M., A Review of Principal Component Analysis Algorithm for Dimensionality Reduction, Journal of Soft Computing and Data Mining. (2021) 2, no. 1, https://doi.org/10.30880/jscdm.2021.02.01.003.

29 Lessmann S., Stahlbock R., and Crone S. F., Genetic Algorithms for Support Vector Machine Model Selection, The 2006 IEEE International Joint Conference on Neural Network Proceedings, IEEE, https://doi.org/10.1109/IJCNN.2006.247266.

30 Li Z., Liu F., Yang W., Peng S., and Zhou J., A Survey of Convolutional Neural Networks: Analysis, Applications, and Prospects, IEEE Transactions on Neural Networks and Learning Systems. (2022) 33, no. 12, 6999–7019, https://doi.org/10.1109/TNNLS.2021.3084827, 34111009.

31 Altman D. G. and Bland J. M., Statistics Notes: The Normal Distribution, BMJ. (1995) 310, no. 6975, 298–298, https://doi.org/10.1136/bmj.310.6975.298, 2-s2.0-84970860170.

32 Dontchev A. L., Qi H., and Qi L., Convergence of Newton′s Method for Convex Best Interpolation, Numerische Mathematik. (2001) 87, no. 3, 435–456, https://doi.org/10.1007/PL00005419, 2-s2.0-0035595452.

33 Polyak B. T., Newton′s Method and Its Use in Optimization, European Journal of Operational Research. (2007) 181, no. 3, 1086–1096, https://doi.org/10.1016/j.ejor.2005.06.076, 2-s2.0-33947685523.

34 Martin R., Noise Power Spectral Density Estimation Based on Optimal Smoothing and Minimum Statistics, IEEE Transactions on Speech and Audio Processing. (2001) 9, no. 5, 504–512, https://doi.org/10.1109/89.928915, 2-s2.0-0035396555.

35 Lampert C. H., Blaschko M. B., and Hofmann T., Beyond Sliding Windows: Object Localization by Efficient Subwindow Search, Proceedings of the 2008 IEEE Conference on Computer Vision and Pattern Recognition, 2008, IEEE, 1–8, https://doi.org/10.1109/CVPR.2008.4587586, 2-s2.0-51949099868.

36 Bastiaans M., On the Sliding-Window Representation in Digital Signal Processing, IEEE Transactions on Acoustics, Speech, and Signal Processing. (1985) 33, no. 4, 868–873, https://doi.org/10.1109/TASSP.1985.1164653, 2-s2.0-0021890796.

37 Ringnér M., What Is Principal Component Analysis?, Nature Biotechnology. (2008) 26, no. 3, 303–304, https://doi.org/10.1038/nbt0308-303, 2-s2.0-40449141013.

38 Song F., Guo Z., and Mei D., Feature Selection Using Principal Component Analysis, Proceedings of the 2010 International Conference on System Science, Engineering Design and Manufacturing Informatization, 2010, IEEE, 27–30, https://doi.org/10.1109/ICSEM.2010.14, 2-s2.0-78650554581.