Marco Carratù, Vincenzo Gallo, Antonio Pietrosanto, Paolo Sommella
An LSTM based soft sensor for rear motorcycle suspension

The increasing development of neural networks for classification and prediction of temporal sequences has opened the way for a new development of mathematical models for soft sensor design. In particular, Long Short-Term Memory (LSTM) networks have greatly improved execution time and reduced error in both single-step and multi-step prediction. In this context, it is therefore possible to improve on the current concept of Instrument Fault Detection and Isolation (IFDI), reducing costs and footprint by not using physical redundancies of sensitive elements but by employing virtual sensors themselves. Therefore, the work aims to develop a soft sensor for rear suspension stroke using an LSTM network. This new approach was trained on over 50000 samples acquired in a real-world environment, and the results were compared with ground truth on a total of over 100000 samples. The results of the analysis showed excellent potential of the method and wide room for improvement in future developments.

Doris Schadler, Michael Wohlthan, Andreas Wimmer
Adaptive Methods for Fault Detection on Research Engine Test Beds

In the case of test beds for research engines, fault detection methods that use models based on historical data face a particular challenge. Due to the experimental design of the test bed, offline training of statistical models with a data set containing all possible variations is simply not possible. The methods must adapt to the current data situation directly on-site. But this involves risks. First, computational time and memory requirements can become extremely large with high data volumes. Second, the data may be faulty and thus negatively affecting the models. To avoid both, a selection of data is made before it is used to build the fault-free reference model. For this purpose, a new statistic is presented as the combination of the Mahalanobis distance and the forecast residual. With it, it is possible to reduce the update frequency and to increase the rate of detected faulty points, since the models are no longer manipulated by faulty data points and thus the residuals provide a better structure for fault detection.

Chao-Ching Ho, Wei-Ming Su, Sankarsan Mohanty
Based on Deep Convolutional Neural Network and Machine Vision Applied to the Surface Defect Detection of Hard Disk Metal Gaskets

This Study aims at the surface defects of aluminum gaskets as the detection targets. The types of defects are yellow spots, incomplete grinding and bump damages. The detection method will select image processing or deep learning according to the characteristics of the defects. The characteristic of yellow spots has many variables of random shapes and different shades of color, it is difficult to use image processing to detect defects, therefore, this Study selects deep learning as the detection method of yellow spot and the detection network architecture is a modified architecture based on U-Net. It also proposes the preprocess of removing the background of the image before the model training, by removing the outer pixel value out side the gasket area on the image. It was found that the preprocess can improve the Intersection over Union (IoU) by 0.041. The experiment results showes that using the proposed network architecture the evaluation of yellow spot IoU is 0.611 which is better than the original U-Net with a model accuracy of 99.56%.

Viola Gallina, Zsolt János Viharos, Gabor Nick, Maik Frye, Andreas Kluth, Adam Szaller, Robert H. Schmitt
Factory of the Year Prize – A Benchmarking

The first factory of the year prize was granted more than 60 years ago in the USA. Since then, a considerable number of countries joined this way and several best factory assessment methods and awards have been developed on national, regional, and international levels. These competitions give the possibility for benchmarking the companies. However, in the era of industry 4.0 maturity models have emerged for evaluating individual enterprises' readiness. These models support the companies in the individual strategy development. But the companies are always interested in their results and achievements compared to their competitors. But setting up a benchmark for a part of an industrial sector might be challenging. Therefore, combining the factory of the year evaluation concept with the maturity assessment might be advantageous. In the paper both of the approaches are analysed and it is discussed how they might be linked in a meaningful way.

Paul-Alexander Vogel, Anh Tuan Vu, Hendrik Mende, Shrey Gulati, Tim Grunwald, Robert H. Schmitt, Thomas Bergs
Machine learning-based predictions of form accuracy for curved thin glass by vacuum assisted hot forming process

Thin glass is applied in numerous applications, appearing as three-dimensional smartphone covers, displays, and in thin batteries. Nonisothermal glass molding has been developed as a hot forming technology that enables to fulfil demands of high quality yet low-cost production. However, finding optimal parameters to a new product variant or glass material is highly demanding. Accordingly, manufacturers are striving for efficient and agile solutions that enable quick adaptations to the process. In this work, we demonstrate that machine learning (ML) can be utilized as a robust and reliable approach. ML-models capable of predicting form shapes of thin glass produced by vacuum-assisted glass molding were developed. Three types of input data were considered: set parameters, sensor values as time series, and thermographic in-process images of products. Different ML-algorithms were implemented, evaluated, and compared to reveal random forest and gradient boosting regressors as best performing on the first frame of the thermographic images.