This paper surveys recent advances in deep learning for sensor-based activity recognition (HAR). It discusses three key aspects: sensor modality, deep models, and applications. The study highlights the advantages of deep learning in automatically extracting high-level features, improving performance, and addressing challenges in unsupervised and incremental learning. It also summarizes public HAR datasets and presents grand challenges for future research. The paper emphasizes the importance of sensor deployment, preprocessing, and model selection for effective HAR. It compares various deep learning models, including deep neural networks, convolutional neural networks, autoencoders, restricted Boltzmann machines, and recurrent neural networks, and discusses their suitability for different tasks. The study also addresses challenges such as online and mobile HAR, unsupervised activity recognition, recognition of high-level activities, lightweight models, non-invasive sensing, and applications beyond activity recognition. The paper concludes that deep learning offers significant improvements over traditional methods in HAR, but further research is needed to overcome existing challenges.This paper surveys recent advances in deep learning for sensor-based activity recognition (HAR). It discusses three key aspects: sensor modality, deep models, and applications. The study highlights the advantages of deep learning in automatically extracting high-level features, improving performance, and addressing challenges in unsupervised and incremental learning. It also summarizes public HAR datasets and presents grand challenges for future research. The paper emphasizes the importance of sensor deployment, preprocessing, and model selection for effective HAR. It compares various deep learning models, including deep neural networks, convolutional neural networks, autoencoders, restricted Boltzmann machines, and recurrent neural networks, and discusses their suitability for different tasks. The study also addresses challenges such as online and mobile HAR, unsupervised activity recognition, recognition of high-level activities, lightweight models, non-invasive sensing, and applications beyond activity recognition. The paper concludes that deep learning offers significant improvements over traditional methods in HAR, but further research is needed to overcome existing challenges.