(199a) Image Classification in Manufacturing Analytics: Improving a Pellet Classification System with Deep Neural Networks

Manufacturing analytics is of paramount importance in many plants today, and its significance increases in the current context of big data analytics and industry 4.0 initiatives[1]. The fields of statistics, chemometrics, and machine learning are expected to provide tools that effectively handle data characteristics, such as process measurements (e.g. temperature, pressure), spectra from online analyzers and quality control labs, pictures taken with standard or hyper-spectral cameras, and all other data sources that provide insight into the state of product processes and quality.

The task of image classification[2] is considered in this paper. A supervised learning problem is presented whereby an image is the input and the output is a unique label attributed to the image from a set of available classes. Image classification is one of the main tools used for quality assurance and control at Dow Chemical, and developing a suitable classifier is a relevant industrial challenge due to accuracy and robustness requirements. In this context, recent developments in deep learning[3] that have proven successful in increasing image classification accuracy and providing state-of-the-art results in computer vision. Traditional approaches utilized for image classification are based on prior knowledge and pre-defined features. However, in this work, we leverage deep neural networks’ (DNN) ability to automatically learn features from images[4] and test their performance in a real industrial context of pellet shape prediction. Moreover, other less complex techniques such as partial least squares discriminant analysis (PLS-DA) and random forests (RF) are explored in order to assess the benefits of adopting DNN as opposed to a more traditional classifier.

PLS-DA, RF, and DNN models were developed for two classification tasks: pellet shape classification (distinguishing good and bad pellets), and detecting tails in a pellet (distinguishing whether a pellet contains tails or not). After developing these models, the results were consistent for both classification objectives. Compared to the in situ classification system currently used at Dow Chemical, RF were able to better utilize the same pre-defined features and improve prediction accuracy significantly, while PLS-DA had the worst performance. DNN obtained the highest accuracy overall (higher than 96%), and the main advantage is that there is no need to specify a priori features because they are extracted from the raw image itself. Furthermore, visualizing the output of some layers of the network showed that activations occurred in regions that are meaningful for the classification tasks, further supporting that DNN effectively modeled the relevant features of the pellet.

References

1. Reis, M. and G. Gins, Industrial Process Monitoring in the Big Data/Industry 4.0 Era: from Detection, to Diagnosis, to Prognosis. Processes, 2017. 5(3): p. 35.
2. Duchesne, C., J.J. Liu, and J.F. MacGregor, Multivariate image analysis in the process industries: A review. Chemometrics and Intelligent Laboratory Systems, 2012. 117(Supplement C): p. 116-128.
3. LeCun, Y., Y. Bengio, and G. Hinton, Deep learning. Nature, 2015. 521(7553): p. 436-444.
4. Simonyan, K. and A. Zisserman, Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.