The paper introduces AttCLHAR, a novel model designed for human activity recognition (HAR) using ambient sensor data in smart homes. The model leverages self-supervised learning, specifically the SimCLR framework, and incorporates self-attention and sharpness-aware minimization (SAM) to enhance generalization and performance. The AttCLHAR model is structured with an encoder that includes two convolutional layers, an LSTM layer, and a self-attention layer, followed by a projection head and a classification layer. Extensive experiments on three CASAS smart home datasets (Aruba-1, Aruba-2, and Milan) demonstrate the model's superior performance in semi-supervised and transfer learning scenarios compared to other models such as SimCLR, SimCLR with SAM, and SimCLR with self-attention. The results highlight the effectiveness of the AttCLHAR model in learning meaningful feature representations from unlabeled data, making it a promising tool for real-world HAR applications in smart homes. Future work will focus on addressing challenges like imbalanced activity data and expanding the model's applicability to diverse smart home environments.The paper introduces AttCLHAR, a novel model designed for human activity recognition (HAR) using ambient sensor data in smart homes. The model leverages self-supervised learning, specifically the SimCLR framework, and incorporates self-attention and sharpness-aware minimization (SAM) to enhance generalization and performance. The AttCLHAR model is structured with an encoder that includes two convolutional layers, an LSTM layer, and a self-attention layer, followed by a projection head and a classification layer. Extensive experiments on three CASAS smart home datasets (Aruba-1, Aruba-2, and Milan) demonstrate the model's superior performance in semi-supervised and transfer learning scenarios compared to other models such as SimCLR, SimCLR with SAM, and SimCLR with self-attention. The results highlight the effectiveness of the AttCLHAR model in learning meaningful feature representations from unlabeled data, making it a promising tool for real-world HAR applications in smart homes. Future work will focus on addressing challenges like imbalanced activity data and expanding the model's applicability to diverse smart home environments.