A human activity recognition model based on an improved Transformer-BiLSTM network is proposed to address the problem of decreased accuracy in activity recognition methods due to the high dimensionality and large noise of time series collected by wearable sensors. The model leverages the advantages of Transformer encoder in handling long-range dependencies and parallelized computations to enhance the efficiency of sequence feature extraction. Subsequently, the features are passed to a bidirectional long short-term memory network with skip residual connections, where two residual connections replace numerous convolutional layers while retaining essential information. Additionally, an attention layer integrated with time information encoding is proposed to enhance the model′s expressive power and understanding of temporal data. Experimental results show that the model achieves an accuracy of 98.38% on public datasets, effectively improving the accuracy of human activity recognition.