Smart Sports Outward Bound Training Assistant System Based on WSNs

The outward-bound training has been a popular manner to exercise in daily life. However, there lacks an intelligent assistant system to monitor the real-time status of users to avoid accidents during training. In order to fill this gap, this paper established an intelligent system to monitor fatigue status during outward-bound training by using surface electromyography (sEMG) signals. The system consists of three parts: a wearable device, edge node, and cloud server. First, the wearable device collects sEMG signals. Second, the edge node processes the collected sEMG signals and sends the sEMG signal features to the cloud server. Finally, the cloud server returns the results to edge node according to a stored classification model that learnt from massive historical sEMG signals. The experimental results show the effectiveness of the proposed system.


INTRODUCTION
With the development of knowledge intensive industries, the concept of high-tech research and development has been updated and the pace of work has been sped up.Brain work has greatly enhanced the intensity of competition.With the increasing pressure, contemporary persons are facing various problems.How to improve their ability to withstand various pressures and deal with different kinds of crises has become an emerging topic in the research and has the top priority (Holsti 2019).Outward bound training (Arachchige & Sathsara 2020;Warner et al.2020) takes the natural environment as the basis, adopts specially designed challenging training programs to help people to experience psychological challenges, temper their perseverance, overcome difficulties, adjust their physical and mental status, and recognize the role of the group to enhance the sense of participation and responsibility for the collective, and improve interpersonal relationships.The outward bound training is an effective way to improve the psychological quality and social adaptability by more harmoniously cooperating with the group to improve their ability of solving problems (Hodgkinson et al. 2021).
The outward bound training refers to the development of unusual and unexpected experiences and work (Seaman et al. 2020).The training refers to controlled and repeated practice as a means to teach or complete a technology or procedure.Outward bound training is a new learning and training method for modern people and organizations.It extends the basic content of walking, running, jumping, throwing, climbing, leaping, etc., which is originally limited to the school physical education classroom and highly competitive to society and nature.
In the field of outward bound training (Kirwin et al. 2019), it is to experience some new things, which are strange, challenging, different from the usual way of life, and contain certain skills.In the experience process, people may face difficulties and frustrations, and even have to deal with the abrupt crisis or real danger.By combining with the characteristics and definition of outward bound training, the connotation of outward bound training is a new training method which makes use of the natural environment and various artificial complex environments, focuses on psychological challenges to achieve the training of "stimulating potential, refining teams, changing mental models".The outward bound training is a well-designed practice activity with development and challenge.It focuses on the educational concept of "people-oriented" and extends the social functions of sports.
With the continuous development of information technology, the emergence of computers with stronger computing power and more accurate sensors has also promoted the development of human beings.Cloud computing, big data, Internet of Things, and artificial intelligence have quickly rushed into various industries and disciplines to bring changes (Ghosh et al. 2018).At the same time, a more natural and harmonious human-computer interaction mode was also desired by the world (Shu et al. 2020).A large number of people have invested in the relevant research.Motion Sensing technology (Ahn et al. 2019), also known as somatosensory interaction technology and motion sensing control technology, is a way to directly use body movements to interact with surrounding devices or environments, and then recognize users' actions by machines to analyze the human-computer interaction technology for making corresponding feedback.The interaction mode of somatosensory technology and people also connects it with "human motion recognition" which is one of the most active research.
Human motion recognition (Wang et al. 2018) involves many disciplines, such as sports physiology, cognitive psychology, computer vision, machine learning, artificial intelligence, and so on.The research has also made some phased achievements in video surveillance, somatosensory games, gesture control and other fields, to make the research on human motion recognition based on somatosensory technology become a hot topic in related fields.However, there still is no related research in outward bound training.This paper combines Internet of Things (IoT) with wireless sensor networks (WSNs) to design an outward bound training assistant system.First, the surface electromyography (sEMG) signals are collected by using a wearable device.Second, the wearable device removes the noises in sEMG.Third, the denoised sEMG signals are transmitted through WSNs to an edge node.In the edge node, the sEMG signals are converted as meaningful features which are further analyzed in cloud server.Fourth, the cloud server receives the sEMG features from edge node and returns the associated results to edge node.In the cloud server, the assistant system implements fatigue status detection which is an important issue during outward bound training.
The main contributions of this paper are summarized as follows: first, an intelligent training assistant system is proposed for outward bound training; second, an sEMG signals based fatigue detection strategy is proposed and merged into the intelligent assistant system to control the amount of outward bound training; third, the proposed assistant system and fatigue detection method are evaluated by experiments.
The remains of this paper is organized as follows: the related work is reviewed in Section 2; the details of proposed assistant system for outward bound training is explained in Section 3; the experiments and simulations are provided in Section 4; the last section is the conclusion and discussion.

ReLATeD WORK
With paying attention to the health, more and more intelligent sports devices have been developed and used in daily life, such as smart sports watches, smart sports shoes, intelligent fitness equipment and intelligent sports data collection device (Lu et al. 2020).The professional sports data collection device helps basketball players to correct their movements during basketball training (Ma et al. 2018).The intelligent fitness equipment has high-definition displays and can connect to the Internet to guide people's daily exercises (Song et al. 2022).
The "Play Pure" launched in the intelligent tennis racket tracks the swing strength, hitting speed, hitting point, etc. (Takeuchi 2020).The collected data is uploaded to supporting applications for further analyzing to guide user daily training.Meanwhile, users can view the sports data of players around the world through the Internet.As a similar product, the "smart basketball" has 6 built-in sensors which can track and record the user's dribble speed, ball control strength, shooting time, shooting angle and other data.
Rope skipping is a common and effective way to keep fit.However, in the process of rope skipping, due to the influence of movement, users often forget the amount, which makes it difficult to control the amount of movement.In order to solve this issue, an intelligent skipping rope is produced to accurately count and calculate the amount of skipping (Dong et al. 2018).Additionally, the intelligent skipping rope can calculate the calories consumed by users according to their body mass index combined with the rope movement trace, and recommend a reasonable exercise cycle.The intelligent fitness equipment and cloud platform can provide more refined sports training equipment for professional fitness athletes.
It can be seen that the existing intelligent sports equipment is either for professional fields or for professional athletes.For the popular intelligent sports training equipment, it is difficult to provide comprehensive and more balanced sports training for the public (Kenney at al. 2021).After a period of investigation and research, it is found that the use of somatosensory technology combined with motion recognition algorithm (Nakamura et al. 2005) can bring more scientific and comprehensive fitness activities to the masses with low cost and efficiency, especially for schools and communities in which individuals have little demand for sports, but the base number is large.At present, most scholars' motion recognition based on somatosensory technology only stays at the theoretical stage and has not been developed and utilized, but it has been applied in related medical rehabilitation fields, such as rehabilitation equipment with virtual scene stimulation (Xie et al. 2022), virtual treatment environment for upper limb stroke rehabilitation (Burke et al. 2010), rehabilitation robot with adjustable posture that can design motion tracks and virtual reality rehabilitation training system based on machine vision (Chen et al. 2009).For the outward bound training, there still lacks related works.

ASSISTANT SySTeM FOR OUTWARD BOUND TRAINING
Muscle fatigue usually refers to the decline of muscle function caused by long-time or excessive exertion of force, which is a common physiological phenomenon (Rampichini et al. 2020).The muscle fatigue is not only an important supplement to sports analysis research, but also helps to promote the understanding of the motion control mechanism of the neuromuscular system.It has important theoretical and application value in sports, rehabilitation medicine, biomechanics and other fields.
In order to understand the cause of muscle fatigue, many scholars have analyzed the changes of structure, function and physical and chemical indexes of central nerve and motion unite (MU) before and after fatigue using various methods, including near-infrared spectroscopy, ultrasound, biochemical detection and electromyography.By summarizing these changes, it is generally believed that the occurrence of muscle fatigue involves the active inhibition of the central nerve and the metabolic accumulation of the surrounding muscle receptors.Active inhibition of the central nervous system refers to the protective response triggered by the central nervous system to avoid muscle damage in excessive exertion.It is mainly achieved by reducing the recruitment of MU and the frequency of MU release.Metabolic accumulation of peripheral muscle receptors refers to the decline of muscle function caused by the continuous accumulation of various metabolites and the continuous consumption of adenosine triphosphate in cells during muscle contraction.It has been reported that the metabolic accumulation of muscle cells may lead to the decline of muscle fiber conduction velocity and contractile force.In the above-mentioned means of studying muscle fatigue, sEMG does not only have the advantages of non-invasive and in vivo measurement, but also can directly reflect the exercise intensity and muscle strength level.It is suitable for exercise analysis and muscle fatigue analysis (Toro et al. 2019).
Until now, there is no standard for determining the degree of muscle fatigue, especially during the dynamic exercise.According to the broad consensus of fatigue, when the subject has been unable to complete the target action, it can be considered that the relevant muscle groups have reached exhaustion for the target action at this time, which is considered as a specific fatigue state.In general, under the condition of static constant force contraction, the sEMG signal amplitude will gradually increase and the frequency spectrum will shift to low with the deepening of muscle fatigue.Under the condition of dynamic contraction, the amplitude change of sEMG is more complicated due to the constant change of muscle force and poor signal stability, but its frequency spectrum still shifts to a low value.This section designs an assistant system for outward bound training to monitor the fatigue status by using sEMG signals.The structure of assistant system is illustrated in the following figure.
In Figure 1, the assistant system consists of three parts: wearable device which contains a sensor to collect sEMG signals of users, edge node which receives the sEMG signals and extracts sEMG signal features, and cloud server which analyzes the sEMG signal features from edge node.
During the process of collecting sEMG signals, the collected signals may be corrupted by noises which are produced by the environment, equipment and cables in motion.The noises would have the side effect on the following analysis.Generally, the main energy of sEMG signals is concentrated between 20-500 Hz, while the low-frequency noises caused by motion are mostly within 20 Hz.A Butterworth band-pass filter is adopted to remove the noised in the signals.The 50 Hz power frequency interference is removed by a notch filter.
Since sEMG signals have the characteristics of randomness, nonstationarity and nonlinearity, the raw signal cannot accurately reflect the fatigue status of muscles.The raw data should be processed by feature selection (Zheng et al. 2021;Zheng et al. 2018) or feature extraction (Shi et al. 2020) method.This paper adopts the time domain and frequency domain features to reduce the redundancy of information in raw data.
The signal in time domain changes the amplitude and shape by muscle contraction.The amplitude of sEMG signal increases with the deepening of fatigue.The commonly used time domain features include: mean absolute value (MAV), root mean square (RMS), integral electromyography (IEMG), variance (VAR) and zero crossing rate (ZC).The average absolute value refers to the average of the absolute value of valid data in the active segment.The root mean square reflects the effective value of the signal in the active segment.The integral electromyography is the sum of absolute values of sEMG amplitude.The more obvious the muscle contraction, the greater its amplitude.The iEMG decreases when muscle fatigue occurred.The variance is generally used to reflect the degree of dispersion of data.The zero crossing rate refers to the number of times the amplitude passes through the zero axis.With the appearance of fatigue, the zero crossing rate decreases.Because of the randomness of the sEMG signal, frequency domain analysis can be used for indepth analysis of muscle fatigue.Frequency domain analysis converts time-domain signal to frequencydomain signal and extract features according to the spectrum distribution.With the deepening of fatigue, the dominant frequency moves to the low frequency.The commonly used frequency domain features include average frequency, total power, median frequency and average power frequency.The average frequency is the sum of the power spectral density divided by the sum of the frequencies.The total power refers to the sum of power spectral density.The median frequency can be regarded as the frequency corresponding to half of the total power, which has good anti-interference characteristics.The average power frequency is the frequency corresponding to the average power spectral density.The sensitivity to spectrum change is higher than that of median frequency.
For an active segment of sEMG signal, it is represented as a 1 9 ´ vector which consists 5 time domain features and 4 frequency domain features.Let x i represent the associated feature vector, y i denote whether the person is under fatigue state during outward bound training ( i l = … 1, , , l is the number of active segments).When y i = +1 , the associated active segment is under fatigue status.When y = −1 , the associated active segment is under non-fatigue status.The collected feature vectors of sEMG segments are used to learn a classifier, which is uploaded into cloud server, to further analyze the fatigue status during outward bound training.The commonly used classification methods include linear discriminant analysis (Zhu et al. 2022), logistic regression (Venugopal & Ramakrishnan 2014), support vector machine (Zhu et al. 2017a;Zhu et al. 2017b) etc.In this paper, the minimum class variance margin support vector machine (MCVSVM) (Zafeiriou et al. 2007)) is adopted as classification model.The linear decision function of MCVSVM is represented as d b T ( ) x w x = + .Let S w be the within-class scatter of the training set, which is represented as follows: Here, P and N are the positive sample set and negative sample set, respectively; x p and x n are the mean of positive samples and negative samples, respectively.Compared with canonical SVM, the MCVSVM maximizes the minimum margin along with the direction of minimum variance to obtain the better performance.An illustration of canonical SVM and MCVSVM is shown in Figure 2.
In Figure 2, the support vector machine (SVM) merely maximizes the minimum margin between positive samples and negative samples, while the minimum class variance margin support vector machine (MCVSVM) maximizes the minimum margin along with the minimum variance direction.Compared with SVM, MCVSVM considers both minimum margin and data distribution of the training set.
The formulation of MCVSVM is written as follows: min: . .: , l as the Lagrangian multipliers for the constraints in Eq. ( 2), the dual form of Eq. ( 2) is written as follows: Here, 1 l is a 1´l vector with all elements 1, Q x S x ( , ) i j y y i j i w j ( ) , where b is determined by Eq. ( 4) as follows: For the nonlinear MCVSVM, the within-class scatter is rewritten as follows: The formulation of nonlinear MCVSVM is written as follows: min: . .: Let P be the matrix with columns eigenvectors of S w F and  S P S P w T w The formulation of Eq. ( 6) can be rewritten as follows: Here, AE • = P .Introducing the Lagrangian multipliers for the constraints in Eq. ( 7), the dual form of Eq. ( 7) is the same as that of Eq. ( 3) except the definition of matrix Q whose element is defined as follows:

eXPeRIMeNTS AND SIMULTATIONS
This section will evaluate the proposed outward bound training assistant system.The sEMG signals are collected from testers through the SX230-1000 sensors.There are 20 volunteers as testers in the experiments, 10 males and 10 females.Each tester executes the target actions for many times during outward bound training from non-fatigue status to fatigue status.After removing the noises, each sEMG signal is split by a fixed window lasting 30 seconds.Two adjacent windows overlap 10 seconds.The sEMG segments are used as training set to learn classification model which is uploaded into the cloud server.
In order to verify the effectiveness of the proposed scheme, the classification model is learnt by using raw data, time-domain feature, frequency domain feature, and time & frequency domain features.For the classification model, MCVSVM is compared with k-nearest neighbors (kNN) (Al-Faiz et al. 2010), linear discriminant analysis (LDA) (Negi et al. 2016), radial basis function neural network (RBF-NN) (Hamedi et al. 2014), canonical SVM (Sui et al. 2019), and weighted SVM (WSVM) (Zhu et al. 2016).The associated results are reported in term of precision, recall and F1-score, which are defined in Eq. ( 9), ( 10) and (11), respectively.
Here, TP represents true positive which means the true label is positive and is predicted as positive, FP represents the false positive which means the true label is negative but is predicted as positive, and FN represents the false negative which means the true label is positive but is predicted as negative.The associated performance is reported in Table 1, 2 and 3 for precision, recall, and F1-score, respectively.
From the results in Table 1, it can be found that when using MCVSVM as classification model, the precision reaches 64.06%, 83.73%, 82.56% and 88.72% for raw data, time domain features, frequency domain features, and time-frequency domain features, respectively; when using time-frequency domain features, the precision reaches 84.35%, 85.67%, 86.32%, 85.93%, 87.18% and 88.72% for kNN, LDA, RBF-NN, SVM, WSVM and MCVSVM, respectively.For precision, the time-frequency domain features perform better than raw data, time domain features, and frequency domain features; MCVSVM performs better than kNN, LDA, RBF-NN, SVM and WSVM.The proposed scheme performs better than previous ones.
From the results in Table 2, it can be found that when using MCVSVM as classification model, the recall reaches 63.38%, 84.03%, 83.47% and 88.46% for raw data, time domain features, frequency domain features, and time-frequency domain features, respectively; when using time-frequency domain features, the recall reaches 85.16%, 86.39%, 85.91%, 85.74%, 86.37% and 88.46% for kNN, LDA, RBF-NN, SVM, WSVM and MCVSVM, respectively.From the results in Table 3, it can be found that when using MCVSVM as classification model, the F1-score reaches 0.618, 0.827, 0.821 and 0.879 for raw data, time domain features, frequency domain features, and time-frequency domain features, respectively; when using time-frequency domain features, the F1 score reaches 0.849, 0.853, 0.849, 0.851, 0.864 and 0.879 for kNN, LDA, RBF-NN, SVM, WSVM and MCVSVM, respectively.It can obtain similar conclusions for recall and F1 score.

CONCLUSIONS
This paper designs an assistant system for the outward bound training.The assistant system can monitor the fatigue status of users during outward bound training to control the quantity of training.

Figure 1 .
Figure 1.The architecture of assistant system for outward bound training

Figure 2 .
Figure 2. The comparison of canonical SVM and MCVSVM.The dotted lines are margin hyperplanes.Left: SVM in which the margin hyperplanes maximize the minimum margin between positive samples and negative samples; right: MCVSVM in which the margin hyperplanes maximize the minimum margin along with the minimum variance direction

Table 3 . The F1-score results of fatigue recognition during outward bound training by using sEMG signals
(Zhu et al. 2021)tem consists of three parts: a wearable device to collect user's sEMG signals during outward bound training, edge nodes to process the collected sEMG signals from wearable devices, and a cloud server to store classification model which is learnt by using collected sEMG signals.In the edge node, it extracts the time and frequency domain features of sEMG signals and sends the extracted features to cloud server.In the cloud server, it predicts the fatigue state according to the stored classification model which is learnt by using massive labeled sEMG signals and returns the results to edge node.The final fatigue recognition results can achieve 88.72%, 88.46% and 0.879 for precision, recall and F1 score, respectively.In the future, more functions should be added into the proposed intelligent training assistant system and more machine learning method are used, such as the ordinal regression(Zhu et al. 2021)for fatigue level analysis.ACKNOWLeDGMeNT This research is supported by the Basic Scientific Research Business Cost Project of Colleges and Universities in Heilongjiang Province in 2022 (No. 145209605) and the Education Science Research Project of Qiqihar University (No. GJQTZX2021023).