Techniques for Monitoring the Physical Activities and Predicting the Performance of the Students

Abstract

One of the essential competitive problems today is the primary healthy lifestyle. Selfactiveness is the most important part in leading a healthy life. Practicing and exercising regularly is the only way that will help us to improve our health. There is many positivity, like sinking the risk of chronic diseases and enhancing overall health and fitness. Doing regular exercises in our daily schedule makes our life healthy that overcomes health issues. This work monitors the simple physical actions of the students. The widely used UCI-HAR (Human Activity Recognition) dataset is used in this work for analysis. The major goal of this strategy is to track physical activities using the collected data by the gadget, which allows them to examine their performance. Long Short Term Memory (LSTM) along with Recurrent Neural Network (RNN) are used in this approach that ensembles to fit the model that achieves best result 93% accuracy. The activity of all the 30 volunteers, are trained and tested with the total train data of 7352, and test data of 2947 records. To test the dataset's accuracy, the ensemble technique is tested with 10 key attributes of their mean.

-

Physical

1. Introduction

2022 Copyright for this paper by its authors.

Students' physical activities like lying, standing, running and walking is being monitored as everyday activities and tracked as physical activities in HAR dataset. In this work, ensemble deep learning method is used to trail the physical activity coaching among students and also to track the daily actions. Huge applications like healthy ageing, medical screening are included in physical monitoring. Research on medical field proves that there are more advantages on undergoing physical activities. If enough physical activities is not done on a daily basis, would result in inactive behaviour associated from all origins [ 1 ]. In addition to the increased long-term risk of premature death, insufficient physical exercise also has an impact on short-term quality of life, work ability, and social involvement [ 2 ].

The research study proves that school kids and teen-agers require physical exercise at their initial phase of physical human changes [ 3 ]. Weekly activities on sports by teen-agers gets more benefits that improves their mental as well as physical health [ 4 ]. Absence of physical exercise at the young age also leads to mental problems during their physical development [ 5 ]. Recognition of physical activities is accessible by products like wearable gadgets and smart phones. These gadgets contain a set of sensors, such as accelerometers and gyroscopes, as well as GPS, which, according to [ 6 ], can give the basic data required for activity detection, offering a full picture of the past. More technologies, patents and research are carried out on monitoring physical activities [ 7-9 ]. In some cases, these provide individualized personal training functionality [ 10,11 ]. Monitoring the physical activity is one of the major issues established few decades ago [ 12-14 ]. Development of new gadgets helps to create new applications in variety of industries, including healthcare, sports and defense. Recent surveys [15-17] have looked into the most important past research in this field in depth. This is the case, the scikit-learn library has been used to identify physical activity. The majority of contemporary research work, among other things, relies on WEKA [ 12, 18-20 ], ATLAB[21], or patentable developments, such as the work presented by [22]. Scikit-learn package has many algorithms peer-reviewed by professionals and it’s frequently utilised and grown popular nowadays.

In this work, Scikit-learn package is used with Ageing Population Physical Activity Monitoring PAMAP2 [23] dataset for monitoring that can be easily replicated. Recent study predicts that deep learning approaches are used to classify and extract features with the dataset [24-30]. In this study, the results obtained with deep learning approaches are exceeded by those obtained with other classifiers, indicating that multiple analysis and evaluation is required. The reason behind using deep learning techniques is to compare with existing classification algorithms working with signals that have already been translated in to frequency domain [30]. Finally, when working with sensors incorporated in battery-operated systems, one key study topic is to maximise the usage of energy [ 31-34 ].

Wearable sensors for monitoring the activities is related to precision and battery consumption. Collect and transfer all data gathered from all accessible sensors via the data network. Last but not least, when dealing with sensors incorporated in battery-operated devices, one essential study topic is how to maximise their energy consumption [ 31-34 ]. The UCI-HAR dataset collects all the activity labels such as sitting, normal walking, standing, walking upstairs and downstairs. In this database, 57, 85,091 records having 561 features, 7352 training data, and 2947 test data for 30 volunteers. The training and testing dataset are used to train and fit the LSTM and RNN models, which ensemble the results numerous times for the best accuracy.

Neural network models are non-direct which have a high difference that makes the final model difficult to make prediction. There are many models in neural networks that are combined using ensemble learning to decrease the prediction variance and generalization error. Ensemble learning strategies can be grouped based on variables such as training data, model, and how predictions are integrated. The outcome of the neural network model achieves above 93 percent precision by using ensembling LSTM and the RNN model with the main 10 features like tBodyGyro-Mean-X, tBodyGyroMean-Y, tBodyGyro-Mean-Z, tBodyAccMag-Mean-X, tBodyAccMag-Mean-Y, tBodyAccMagMean-Z, tGravityAccMag-Mean-X, tGravityAccMag-Mean-Y, tGravityAccMag-Mean-Z, tBodyGyroJerkMag-Mean-X, tBodyGyroJerkMag-Mean-Y and tBodyGyroJerkMag-Mean-Z [22].

2. Methodology 2.1.

Ensemble Learners

To achieve better results, these models incorporate the predictions of a large number of base estimators built using a specific learning technique [ 11 ]. The random forest technique is made up of 112 decision-making algorithms [ 35 ] with predictions that use feature subset that are selected randomly [ 36 ]. Since the decision trees has many different weights, analysis becomes too complicated. It requires complex parameter settings or domain knowledge to perform with high-dimensional data. The number of random forest decision trees and the important features are the key parameters to be updated by using these techniques. The decision trees are fixed as 30 with a limit of 10 by testing some values. Maximum n features are used in these classifications [ 49 ] that are similar to the randomized random forest algorithms for achieving better computational splits [ 37 ]. There are two divisions that an additional tree divides nodes and builds trees using the all-learning samples.

In this study, Adaptive boosting, couples all types of algorithms that helps to enhance the efficiency [ 38 ]. Voting classifier algorithms like adaptive boosting and bagging algorithm produce better outcomes on any datasets [ 39 ]. The adaptive boosting algorithm, is selected as the top classifier algorithm [ 40 ]. This algorithm such as bagging and AdaBoost have been very successful in improving outcomes for various different classifiers on both simulated and real-world datasets [ 39 ]. AdaBoost, in particular, has been voted the top out-of-the-box classifier in the world [ 40 ]. The AdaBoost algorithms are used and recommended for constant multilevel-class domains. AdaBoost was used to evaluate the situation [ 41 ]. We used AdaBoost to assess random forest output and other randomization. 2.2.

Deep Learning

Deep learning is one of the important techniques used in the area of neural networks [ 42 ]. This technique is used by processing multiple layers to form non-linear transformations. With the help of progressive learning, the outputs of each layer are grouped. RNN, convolutional deep neural networks, Deep and Deep belief neural networks are used in layers processing. As a result of these strategies, the data inputs will be mechanically built, offering an additional generic answer because the feature generation approach will be completely done automatically. These techniques are applied to a variety of fields like pc vision, tongue processing, and speech recognition and so on, with progressive results on numerous tasks [ 43–45 ]. Deep learning techniques have been used in a few publications to recognize activities; a representative set may be found in [26-31]. The TensorFlow framework are recently introduced by Google (Mountain View, CA, USA) [ 46 ] are used in DNN. The essential need for applying these strategies is to test 280 attributes into classification.

Varieties of topologies are tested that revealed two things: (1) Only exact topology styles have a significant impact on the produced results; and (2) Simple architectures are closer to the optimum results than the complex architecture. 2.3.

Parameter Setup

Data

Testing

Data

Total 2 1 0 9 8 6 6 5 2 8 7 2 8 6 2 7 6 9 689012 278256 967268 603026 264231 867257 554132 235620 789752 722732 275451 998183 772188 298452 1070640 790734 301257 1091991 A C T I V I T Y L A B E L S

Training Data

Testing Data 6 2 0 3 0 6 1 3 2 4 6 2 7 5 2 7 6 8 2 3 1 4 5 complex unit with Input gate (i), Output gate (o), Memory cell (c) and Forget gate (f). The LSTM technique is also used to transfer the data within the gates that decides the number of new inputs and also the memory to be maintained by the forgotten gate. Equations (1)-(6) gives the details about the functioning of LSTM.

= ℎ = = . −1 + .

ℎ( ). = ( + ℎℎ −1 + ) , = ( 1 +

ℎℎ −1 + ) = ( + ℎℎ −1 + ) ℎ( + ℎ 1 + ) (1) (2) (3) (4) (5) (6)

To aggregate the predictions of k LSTM spanning from 1 to 3, we have used ensemble algorithm predictive system involves the usage of an ensemble algorithm. An ensemble can be easily executed by taking the average of individual predictions. As a result, the ensemble's prediction for Activity A and instance t at Equation (7), for example,

To compare the ensemble LSTM, the total error reduction are calculated based on the average percentage reduction [ 48 ]. The

are calculated using Mean Squared Error or Mean Absolute Error. The error is formed from ith time series and jth ensemble method by _ error_reductionj, denoted as, jth ensemble prediction method is defined as follows in Equation (8): then, the 100 ( = 1 ) (8)

Therefore, next step after training on first data set (n train), is to test the models and then forecasting the next observation is represented in Equation (9): (7) (9) ( , ) = = 1 ∑ =1 ) = 1 ∑ Where ( ,

Activity Mean tBodyAcc-mean()-X tBodyAcc-mean()-Y tBodyAcc-mean()-Z

tBodyAcc-XYZ tGravityAcc-mean-X tGravityAcc-mean-Y tGravityAcc-mean-Z tBodyAccJerk-mean-X tBodyAccJerk-mean-Y tBodyAccJerk-mean-Z

WALKING 0.287 -0.155 0.205 0.426 0.205 ), indicates the AUC score of the modeli, at time period t at 128 element time that results in best result using LSTM ensemble.

3. Results

After applying the ensemble LSTM models, the accuracy of the model could be visualized clearly. the three-dimensional data are trained and tested to monitor the physical activity. Figure 2, shows the graphical representation of these mean values of all the cumulative outputs of all the 30 volunteers. By considered the result of these activities, one could able to monitor the students and analyse the performance and actions of all individuals. -0.155 0.278

1 WALKING

4 SITTING

0.4 0.3

4. Conclusion

In the present work, we train our deep learning algorithms on activity labels at all-time intervals. The monitoring parameters, indicates the movements of all the students. The features selected are used to check the accuracy based on the mean values of the data collected that indicates the essential parameter to analyse the monitoring benefits of all the individuals. The ensemble learning model are tested to find the best combination models. The multiple LSTM RNN are ensembled to fit the model. The selected features improve the monitoring system, to analyse the physical activity. The study indicates that the multiple ensembled model with LSTM RNN neural network improves the prediction accuracy about 93%.

5. Acknowledgements

We would like to thank the reviewers for providing suggestions to improve our work. We would also like to thank our parents and family for supporting us in all our future growth.

6. References

[13] S. Pirttikangas, K. Fujinami and T. Nakajima, “Feature Selection and Activity Recognition from Wearable Sensors,” Ubiquitous Computing Systems, Lecture Notes in Computer Science; Springer: Berlin, Germany, Vol. 4239, pp. 516–527, 2006. [14] E. Antonsson and R. Mann, “The frequency content of gait,” J. Biomech., Vol. 18, pp. 39–47, 1985. [15] O.D. Lara, M.A. Labrador, “A Survey on Human Activity Recognition using Wearable Sensors,”

IEEE Commun. Surv. Tutor, Vol. 15, pp. 1192–1209, 2013. [16] S. Mukhopadhyay, “Wearable Sensors for Human Activity Monitoring: A Review,” IEEE Sens.

J., Vol. 15, pp. 1321–1330, 2013. [17] M. Shoaib, S. Bosch, O. Incel, H. Scholten and P.A. Havinga, “Survey of Online Activity

Recognition Using Mobile Phones,” Sensors, Vol. 15, pp. 2059–2085, 2015. [18] D. Maguire and R. Frisby, “Comparison of feature classification algorithm for activity recognition based on accelerometer and heart rate data,” In: Proc. 9th Information Technology and Telecommunications Conference (IT&T’09), Dublin, Ireland, 22–23 October 2009, p. 11. [19] M. Ayu, S. Ismail, A. Matin and T.A. Montoro, “Comparison Study of Classifier Algorithms for Mobile-Phone’s Accelerometer Based Activity Recognition,” Procedia Eng., Vol. 41, pp. 224– 229, 2012. [20] W. Wu, S. Dasgupta, E.E. Ramirez, C. Peterson and G.J. Norman, “Classification accuracies of physical activities using smartphone motion sensors,” J. Med. Internet Res., Vol. 14, pp. 130, 2012. [21] A. Mannini, S.S. Intille, M. Rosenberger, A.M Sabatini and W. Haskell, “Activity recognition using a single accelerometer placed at the wrist or ankle,” Med. Sci. Sports Exerc., Vol. 45, pp. 2193–2203, 2013. [22] A. Reiss, G. Hendeby and D. Stricker, “A novel confidence-based multiclass boosting algorithm for mobile physical activity monitoring,” Pers. Ubiquitous Comput.,Vol. 19, pp. 105–121, 2014. [23] A. Reiss, M. Weber and D. Stricker, “Exploring and extending the boundaries of physical activity recognition” In: Proc. 11th IEEE International Conference on Systems, Man, and Cybernetics (ICSMC’11), 2011, Anchorage, AK, USA, pp. 46–50. [24] H. J. Kim, J.S. Lee and H.S. Yang, “Human Action Recognition Using a Modified Convolutional Neural Network,” In: Proc. 4th Int’l Symp. on Neural Networks: Part II—Advances in Neural Networks (ISSN’07), Nanjing, China, June 2007, pp. 715–723. [25] M. Baccouche, F. Mamalet, C. Wolf, C. Garcia, and A. Baskurt, “Sequential Deep Learning for Human Action Recognition,” In: Proc. 2nd Int’l Conf. on Human Behavior Understanding (HBU’11), Amsterdam, The Netherlands, 2011, pp. 29–39. [26] S. Ji, W. Xu, M. Yang and K. Yu, 3D “Convolutional Neural Networks for Human Action

Recognition,” IEEE Trans. Pattern Anal. Mach. Intell., Vol. 35, pp. 221–231, 2013. [27] K. Charalampous and A. Gasteratos, “On-line deep learning method for action recognition,”

Pattern Anal. Appl., Vol. 19, pp. 337–354, 2014. [28] W. Jiang and Z. Yin, “Human Activity Recognition Using Wearable Sensors by Deep Convolutional Neural Networks,” In: Proc. 23rd ACM International Conference on Multimedia (MM’15), Brisbane, Australia, 2015, pp. 1307–1310. [29] F.J. Ordoñez and D. Roggen, “Deep convolutional and LSTM recurrent neural networks for multimodal wearable activity recognition,” Sensors, Vol. 16, pp. 115, 2016. [30] N.Y. Hammerla, S. Halloran, and T. Ploetz, “Deep, Convolutional, and Recurrent Models for Human Activity Recognition using Wearables,” In: Proc. 25th International Joint Conference on Artificial Intelligence (IJCAI’16), New York, NY, USA, 9–15 July 2016, pp. 1533–1540. [31] Y. Wang, J. Lin, M. Annavaram, Q. Jacobson, J. Hong, B. Krishnamachari and N. A. Sadeh, “Framework of Energy Efficient Mobile Sensing for Automatic User State Recognition. In: Proc. 7th International Conference on Mobile Systems, Applications, and Services (MobiSys’09), Kraków, Poland, 22–25 June 2009, pp. 179–192. [32] Z. Yan, V. Subbaraju, D. Chakraborty, A. Misra and K. Aberer, “Energy-Efficient Continuous Activity Recognition on Mobile Phones: An Activity-Adaptive Approach”, In: Proc. 16th Annual International Symposium on Wearable Computers (ISWC’12), Newcastle, UK, 18–22 June 2012, pp. 17–24.

References

[1] U. Ekelund , J. Steene-Johannessen , W.J. Brown , M.W. Fagerland , N. Owen; K.E. Powell , A. Bauman , and I.M. Lee , “ Does physical activity attenuate, or even eliminate, the detrimental association of sitting time with mortality? A harmonised meta-analysis of data from more than 1 million men and women,” Lancet, vol. 388 , pp. 1302 - 1310 , 2016 .

[2] E. Losina , H.Y. Yang , B.R. Deshpande , J.N. Katz , and J.E. Collins, “ Physical activity and unplanned illness-related work absenteeism: Data from an employee wellness program,” PLoS One , vol. 12 , pp. 1 - 8 , 2017 .

[3] I. Janssen and A.G. Leblanc , “ Systematic review of the health benefits of physical activity and fitness in school-aged children and youth,” Int. J. Behav. Nutr. Phys. Act. , Vol. 7 , pp. 40 , 2010 .

[4] A. Merglen , A. Flatz , R.E. Bélanger , P.A. Michaud and J.C. Suris , “ Weekly sport practice and adolescent well-being,” Arch. Dis. Child , Vol. 99 , pp. 208 - 210 , 2014 .

[5] S. J. Biddle and M. Asare , “ Physical activity and mental health in children and adolescents: a review of reviews,” British journal of sports medicine , https://doi.org/10.1136/bjsports-2011 - 090185 , Vol. 45 ( 11 ), pp. 886 - 895 , 2011 .

[6] J. Andreu-Perez , D.R. Leff , H.M.D. Ip and G.Z. Yang , “ From Wearable Sensors to Smart Implants-Toward Pervasive and Personalized Healthcare,” IEEE Trans. Biomed. Eng . Vol. 62 , pp. 2750 - 2762 , 2016 .

[7] C.A. Vock , P. Flentov and D.M. Darcy , “ Activity Monitoring Systems and Methods”, U.S. Patent 8 , 352 , 211B2 , 8 January 2013 .

[8] S.G.J. Yuen , J. Park and E.N. Friedman , “ Activity Monitoring Systems and Methods of Operating Same”, U.S. Patent 8 , 386 , 008 B2 , 24 September 2013 .

[9] P.R. Kahn , A. Kinsolving , M.A. Christensen , B.Y. Lee , and D. Vogel , “ Human Activity Monitoring Device with Activity Identification” , U.S. Patent 8 , 949 , 070 B1 , 3 February 2015 .

[10] D. Morris , I. Kelner , F. Shariff , D. Tom, T.S. Saponas and A. Guillory , “ Personal Training with Physical Activity Monitoring Device” , U.S. Patent 8 , 951 , 165 B2 , 10 February 2015 .

[11] K.L. White , M.L. Orenstein , J. Campbell, C.S. Self , E. Walker, M. Micheletti , G. McKeag , J. Zipperer and M. Lapinsky , “ Activity Recognition with Activity Reminders” , U.S. Patent 20 , 150 , 042 , 468 A1 , 12 February 2015 .

[12] N. Ravi , D. Nikhil , P. Mysore and M.L. Littman , “ Activity recognition from accelerometer data,” In: Proc. 17th Conference Innovative Applications Artificial Intelligence (IAAI'05) , Pittsburgh, PA, USA, pp. 1541 - 1546 , 2005 .

[33] A. Khan , M. Siddiqi , S.W. Lee , “ Exploratory Data Analysis of Acceleration Signals to Select Light-Weight and Accurate Features for Real-Time Activity Recognition on Smartphones,” Sensors , Vol. 13 pp. 13099 - 13122 , 2013 .

[34] L. Morillo , L. Gonzalez-Abril , J. Ramirez and M. delaConcepcion , “ Low Energy Physical Activity Recognition System on Smart phones , ” Sensors , Vol. 15 , pp. 5163 - 5196 , 2015 .

[35] J. Mingers , “ An Empirical Comparison of Selection Measures for Decision-Tree Induction,” Mach. Learn . Vol. 3 , pp. 319 - 342 , 1989 .

[36] L. Breiman , “Random Forests,” Mach. Learn., Vol. 45 , pp. 5 - 32 , 2001 .

[37] P. Geurts , D. Ernst and L. Wehenkel , “Extremely Randomized Trees, ” Mach. Learn , Vol. 63 , pp. 3 - 42 , 2006 .

[38] Y. Freund and R.E. Schapire , “ A Decision-Theoretic Generalization of On-Line Learning and an Application to Boosting,” J. Comput . Syst. Sci. , Vol. 55 , pp. 119 - 139 , 1997

[39] E. Bauer and R. Kohavi , “ An Empirical Comparison of Voting Classification Algorithms: Bagging, Boosting, and Variants” Mach. Learn. , Vol. 36 , pp. 105 - 139 , 1999 .

[40] L. Breiman , Bagging Predictors . Mach. Learn , Vol. 24 , pp. 123 - 140 , 1996 .

[41] T. Hastie , S. Rosset , J. Zhu and H. Zou , “ Multi-class AdaBoost ,” Stat. Its Interface , Vol. 2 , pp. 349 - 360 , 2009 .

[42] D.E. Rumelhart , G.E. Hinton and R.J. Williams , “ Learning Internal Representations by Error Propagation,” Parallel Distributed Processing: Explorations in the Microstructure of Cognition; MIT Press: Cambridge, MA, USA, Vol. 1 , pp. 318 - 362 , 1986 .

[43] V. Le . Quoc, Y. Zou . Y. Yeung . Serena and Y. Ng . Andrew, “ Learning Hierarchical Invariant Spatio-temporal Features for Action Recognition with Independent Subspace Analysis ” In Proc. of the 2011 IEEE Conference on Computer Vision and Pattern Recognition (CVPR'11) , 20 - 25 June 2011, pp. 3361 - 3368 , Colorado Springs , CO, USA.

[44] J. Dean , G. Corrado , R. Monga , K. Chen , M. Devin , M. Mao , M. Ranzato , A. Senior , P. Tucker and K. Yang , “ Large Scale Distributed Deep Networks,” In: Advances in Neural Information Processing Systems; NIPS Foundation: La Jolla , CA, USA, Vol. 25 , pp. 1232 - 1240 , 2012 .

[45] L. Deng and D. Yu , “ Deep Learning: Methods and Applications,” Found. Trends Signal Process , Vol. 7 , pp. 197 - 387 , 2014 .

[46] M. Abadi , A. Agarwal , P. Barham , E. Brevdo , Z. Chen , C. Citro , G.S. Corrado , A. Davis , J. Dean and M. Devin , “ TensorFlow: Large-Scale Machine Learning on Heterogeneus Distributed Systems”, Tech. Report; Google Research: Mountain View, CA, USA, 2015 .

[47] S. Borovkova and I. Tsiamas , “ An ensemble of LSTM neural networks for high‐frequency stock market classification , ” Journal of Forecasting , Vol. 38 , pp. 600 - 619 , 2019 .

[48] J. Choi and B. Lee , “ Combining LSTM Network Ensemble via Adaptive Weighting for Improved Time Series Forecasting ,” Mathematical Problems in Engineering, Hindawi, Vol. 2018 , pp. 1 - 8 , 2018 .

[49] Sáez Achaerandio , Yago, Baldominos Gómez, Alejandro, Isasi, Pedro. "A Comparison Study of Classifier Algorithms for Cross-Person Physical Activity Recognition" , MDPI , 2017