On the other hand, the ratio of the result of commercialized product to the reference value, that is γ/α column, ranges 74.0% to 169.4%. Table 3 shows that the ratio of the result of developed algorithm to the reference value, that is β/α column, ranges 95.5% to 115.7%. Table 3 below shows that the developed algorithm outperforms the compared commercial product. We compared the assessed calorie consumption results for walking fast and slow, running and jumping rope. They were equipped with an accelerometer device on their waist which was applied with the developed algorithm and a commercialized target device on their wrist. Thirty-one boys and girls who did not participate in the previous classification algorithm development session have participated for this experiment. To demonstrate the validity of the developed algorithm, we compared it with the commercially available product, UP move™ (JAWBONE, USA). The developed algorithm also provides the consumed calories for the four types of continuous physical activity. Then the training process was over to prevent overfitted network. The repetition of training and validation process continued until the performance of the verification process deteriorated six consecutive times. The performance of trained network is defined here as a cross-entropy. After each training session, we provided data of the verification group to the training network to see if network performance improved. Like the previous research, scaled conjugate gradient algorithm was used in the data group training for fast supervised learning. All data patterns were shuffled randomly and devided into three groups, those were training, validation, and test group, with the ratio of 70:15:15, respectively. The hidden layer’s nodes were determined by examining the best performance among the various numbers of nodes. Each three network was designed for four intermittent physical activity, four continuous activity, and calorie consumption estimation. The overall algorithm has three different ANNs. The ANN is composed of input layer, hidden layer with tangent sigmoid transfer function, and output layer with softmax transfer function. The ANN determines what kind of physical activity is detected based on the input pattern. We extracted and selected features from the physical activity data and made input patterns of the ANN for classification. For this algorithm study, we adopted Artificial Neural Network (ANN) method which is one of the supervised machine learning algorithm to attain physical activity classification.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |