With the rapid development of the artificial intelligence field today, sensor-based human activity recognition (HAR) has become mainstream, especially playing a crucial role in medical rehabilitation. However, existing HAR algorithms based on deep learning often ignore the complexity of human activities during rehabilitation and the scarcity of samples due to limited access to rehabilitation training data. To solve these problems, this paper proposes an innovative algorithm called 1DCLA, which is based on 1D convolutional neural network (1DCNN), long short-term memory (LSTM) and attention mechanism. Also, it combines the synthetic minority over-sampling technique (SMOTE) and a combinatorial loss function superimposed by central loss and cross-entropy loss. Experimental results demonstrate that the accuracy of this algorithm reaches 98.74% and 97.87% on UCI HAR and WISDM, outperforming comparative experimental models and improving the recognition accuracy. This study effectively solves the problems of uneven distribution of activity samples and difficulty in activity recognition during rehabilitation, providing more reliable support for rehabilitation therapy.