Wearable Sensor Network‐Driven Human Activity Recognition via Philippine Eagle Optimized Statistics‐Infused Neural Network

ABSTRACT A crucial area of research in behavior analysis, computer communication, and widespread technology is human activity recognition (HAR). Deep learning (DL) approaches are effectively deployed to forecast human behaviors with periodic information from mobile phones as well as wearable sensors. DL‐based methods are very good at identifying activities, but struggle with time series data. In this research work, wearable sensor network‐driven human activity recognition via Philippine eagle optimized statistics‐infused neural network (WSN‐HAR‐PEO‐SINN) is suggested. First, the input information is gathered from the Wireless Sensor Data Mining (WISDM) dataset. In preprocessing, the flat files are combined to form a data frame using confidence partitioning sampling filtering (CPSF). After preprocessing, the data is segmented using unpaired multiview graph clustering (UMGC). It is used to segment the number of human activities. The divided output is sent to the feature extraction system using quadratic phase S‐transform (QPST), which is utilized to gather local characteristics such as skewness, mean, and deviation from the mean, energy, mel‐frequency cepstral coefficients, and entropy. Then, the extracted characteristics are entered into a neural connection termed statistics‐informed neural network (SINN) to recognize the human activity with wearable sensor data and categorize the data as ambulation‐oriented, hand‐oriented activity general, and hand‐oriented activity eating. Generally speaking, SINN does not incorporate any optimization technique adaptations for identifying the optimal parameters that guarantee precise identification of human activities. To enhance the weight parameter of the SINN approach, which accurately detects human activity using wearable sensor data, Philippine eagle optimization (PEO) is suggested. The accuracy, precision, recall, F1 score, ROC, computing time, and error rate are among the performance metrics used to assess the effectiveness of the suggested SINN‐HAR‐WSD‐PEO strategy. In contrast to previous approaches, such as CNN‐GRU‐HAR‐WSD, which uses wearable sensor data for deep learning‐based human activity recognition (HAR), and CNN‐BILSTM‐HAR‐WSD, which uses a multiple branch CNN‐BiLSTM model for employing wearable sensor data for human activity recognition, the suggested SINN‐HAR‐WSD‐PEO method achieves higher accuracy, recall, and precision, which were 22.36%, 25.42%, and 18.27%, 21.36%, 26.42%, and 18.27%, and a multi‐task deep learning approach for sensor‐based human activity recognition and segmentation (MDNN‐SB‐HAR) was 21.36%, 22.42%, and 19.27%, respectively.