Enhancing Human Activity Recognition using PSOOptimized 1D CNN-GRU Model on Wearable Sensor Data

Human Activity Recognition (HAR) is important in areas such as smart environments, health monitoring, behavior tracking, and Human–Computer Interaction (HCI). Although deep learning (DL) models have shown great results in recognizing human activities, handling time-series data from wearable sensors is still challenging. In this work, we introduce a new model that combines a 1D Convolutional Neural Network (1D CNN) with Gated Recurrent Units (GRU), and uses Particle Swarm Optimization (PSO) to fine-tune the model’s hyperparameters. The 1D CNN extracts useful patterns from the sensor data, while the GRU learns the sequence and time-based relationships between the activities. PSO helps improve the model’s performance and training speed by finding the best parameter settings. Batch normalization is used to make training more stable, and max pooling is added to reduce the number of parameters. A Softmax layer at the end is used to classify different types of activities. The model was tested on two widely used datasets, MHEALTH and WISDM, and achieved high classification accuracies of 99.17% and 98.55%, respectively. The results show that the proposed method performs better than many existing techniques. This work uniquely integrates PSO for automated hyperparameter optimization within a CNN-GRU framework, demonstrating a clear advancement over existing HAR models in terms of accuracy and computational efficiency.