A layer-wise neural network for multi-item single-output quality estimation

A layer-wise neural network architecture is proposed for classification and regression of time series data where multiple instances have a single output. This data format is encountered in the manufacturing industry where parts are produced in batches—due to the short production cycle—and labelled as a whole for defects. The end-to-end neural network approach is benchmarked against a previously proposed feature engineering method based upon mean shift clustering and K nearest neighbours with dynamic time warping, and a naive approach of flattening the instances and training a support vector machine. An ablation study is performed on a layer-wise 1D-convolutional neural network (CNN) to understand which of the architectural design choices are critical for prediction performance. Based on a transfer moulding production dataset, it is found that the layer-wise 1D-CNN and multilayer perceptron (MLP) have the best performance across most of the common classification and regression metrics, but the layer-wise MLP has a lower computational cost. Finally, it is shown that the proposed parameter sharing in the dense layers of both networks is key to reducing the number of parameters and improving prediction performance.