Abstract: A multi-input convolutional neural network (CNN) incorporating bidirectional long short-term memory neural network(BiLSTM) and attention mechanism is presented to address the issues of local convergence and poor heterogeneous fusion performance of deep learning algorithms in multi-sensor fusion positioning. Firstly, 5G channel state information (CSI) and geomagnetic data arepreprocessed separately. Then each of them is trained offline based on an independent branch network, and the spatial and temporal features of the fingerprint data are extracted at the sametime to append the attention mechanismlayer. Finally, the fusion of heterogeneoussensor data for localization is achieved at the fully connected layer.The experimental results in the conference room and the teaching building hall show that the average positioning error is 0.95 and 1.84 m respectively, which is 48.9% and 42.7% higher than that of the error backpropagation network(BPNN), and that positioning accuracy and system stability are both greatly improved.

Key words: indoor positioning, convolutional neural network(CNN), bidirectional long short-term memory neural network(BiLSTM), attention mechanism, channel state information(CSI)