Performance enhancement of deep neural network using fusional data assimilation and divide-and-conquer approach; case study: earthquake magnitude calculation.

The presence of "ill-posed samples" specifically in low-volume datasets leads to accuracy decrement in the learning procedure and the generalization of neural networks. Such samples can be caused by various reasons such as noise contamination, corrupted sensors, or even, the complex distribution of physical properties governing the problem. The peak ground acceleration (PGA) datasets are definitely among the last mentioned. Focusing on speed and accuracy, a method for calculating earthquake magnitude based on the PGA data recorded at a single station along with hypocentral information has been presented in this research. Here, after training a deep neural network, the regression errors of the training data samples are clustered into two groups, namely well and ill posed using the grey wolf optimization algorithm. Instead of being removed, the data samples with low learning rates are then modified using samples selected from the other cluster in a fusional form. Then, two separate models are used and trained independently for the clusters. Next, in addition to the routine procedure of network generalization, every new sample is first checked whether is more likely to belong to which group of the clustered data, and after processing, the corresponding trained model is used. The results of the experiments show that using the proposed method results in magnitude calculation with an error order of less than 0.212 units of moment magnitude with a probability of more than 99.7%, which is superior to the conventional methods some of which were reviewed in this research. [ABSTRACT FROM AUTHOR]