A Quaternion Gated Recurrent Unit Neural Network for Sensor Fusion

Recurrent Neural Networks (RNNs) are known for their ability to learn relationships within temporal sequences. Gated Unit (GRU) networks have found use in challenging time-dependent applications such as Natural Language Processing (NLP), financial analysis and sensor fusion due capability cope with the vanishing gradient problem. GRUs also be more computationally efficient than variant, Long Short-Term Memory neural network (LSTM), less complex structure such, suitable requiring management of computational resources. Many require a stronger mapping features further enhance prediction accuracy. A novel Quaternion (QGRU) is proposed this paper, which leverages internal external dependencies quaternion algebra map correlations across multidimensional features. The QGRU can used efficiently capture inter- intra-dependencies unlike GRU, only captures sequence. Furthermore, performance method evaluated on problem involving navigation Global Navigation Satellite System (GNSS) deprived environments well human activity recognition results obtained show that produces competitive almost 3.7 times fewer parameters compared GRU. code available at.