Combination of one-dimensional convolutional neural network and negative correlation learning on spectral calibration

The advantage of data-sensitive deep learning methods used in spectral calibration is not obvious when the amount of available data is insufficient. To solve this problem, this paper proposes a new method that combines one-dimensional convolution neural network (1-dim CNN) with negative correlation learning (NCL). First, we create several identical one-dimensional convolutional neural networks as subnetworks of the NCL system. Second, we add the error function of each subnetwork to a negative correlation penalty term that is related to the correlation between the networks and then use this composite error function to back-propagate these networks for parameter adjustment. Finally, after the model has converged, we take the average of the results of all subnetworks as the result of the whole model. We compare CNN_NCL with PLS,creating diversity partial least squares (CDPLS) and a single 1-dim CNN on the pharmaceutical tablet dataset and diesel fuels dataset. The experimental results show that CNN_NCL performs better than PLS and CDPLS when the number of samples is sufficient. Additionally, CNN_NCL can always be more effective than a single CNN regardless of the data scale. Therefore, in the context of the era of big data, CNN_NCL is a fairly efficient model for spectral calibration.