This paper proposes a novel model, AttCLHAR, for human activity recognition (HAR) in smart homes using self-supervised learning and self-attention mechanisms. The model is designed to work with limited or no annotated data, making it suitable for real-world scenarios where labeled data is scarce. AttCLHAR is based on the SimCLR framework and incorporates a self-attention mechanism to focus on relevant input segments. It also uses sharpness-aware minimization (SAM) to enhance model generalization. The model consists of an encoder with two convolutional layers, one LSTM layer, and a self-attention layer. The encoder is pre-trained on unlabeled data to learn robust features, which are then fine-tuned for specific tasks. The model was evaluated on three CASAS smart home datasets (Aruba-1, Aruba-2, and Milan). The results show that AttCLHAR outperforms existing models, especially in semi-supervised and transfer learning scenarios. The model demonstrates strong performance in capturing and recognizing human activities in real-world environments. The paper also discusses related work, including self-supervised learning, the SimCLR framework, and self-attention mechanisms. The methodology includes pre-training and fine-tuning phases, with extensive experiments on the CASAS datasets. The results highlight the effectiveness of AttCLHAR in HAR tasks using ambient sensor data. The model's ability to learn meaningful representations from unlabeled data makes it a promising approach for real-world applications in smart homes.This paper proposes a novel model, AttCLHAR, for human activity recognition (HAR) in smart homes using self-supervised learning and self-attention mechanisms. The model is designed to work with limited or no annotated data, making it suitable for real-world scenarios where labeled data is scarce. AttCLHAR is based on the SimCLR framework and incorporates a self-attention mechanism to focus on relevant input segments. It also uses sharpness-aware minimization (SAM) to enhance model generalization. The model consists of an encoder with two convolutional layers, one LSTM layer, and a self-attention layer. The encoder is pre-trained on unlabeled data to learn robust features, which are then fine-tuned for specific tasks. The model was evaluated on three CASAS smart home datasets (Aruba-1, Aruba-2, and Milan). The results show that AttCLHAR outperforms existing models, especially in semi-supervised and transfer learning scenarios. The model demonstrates strong performance in capturing and recognizing human activities in real-world environments. The paper also discusses related work, including self-supervised learning, the SimCLR framework, and self-attention mechanisms. The methodology includes pre-training and fine-tuning phases, with extensive experiments on the CASAS datasets. The results highlight the effectiveness of AttCLHAR in HAR tasks using ambient sensor data. The model's ability to learn meaningful representations from unlabeled data makes it a promising approach for real-world applications in smart homes.