Session 2 Enabling Technologies II

Wednesday, 11.11.2020, 11:00-13:00 o'clock


Ivan Vishev: Data-based approach for identifying repeatability, reproducibility, and stability of industrial processes in series production

11:00-11:20 o'clock

(Paper ID: 1112)

The automotive industry sets very high standards for the quality of the manufactured components in the area of body production. The assembly of faulty components can lead to serious accidents and human injury. To ensure compliance with production standards and component quality, the manufacturing processes for the same components are designed to be repeatable and reproducible across all plants. The existing plant monitoring systems use control limits to avoid major consequential damage to the plants by sporadic production process changes, but do not provide information on process repeatability and stability. Any manufacturing process can undergo an unintended change due to many factors. Even planned actions, such as replacement of wearing parts or start of a new material batch, may have delayed and not always predictable effects. A significant change in the manufacturing process is only noticed by system failures or by the accumulation of quality defects detected afterward. Early detection of the changes in the process repeatability and stability allows timely intervention in the process to avoid system failures and quality defects in components.

This paper presents a new data-driven approach, which allows the evaluation of the repetition accuracy aiming to identify process changes along the time or geometric production sequence and evaluation of the stability of the manufacturing process. The approach makes use of the fact that the majority of the manufactured products correspond to the production standard and the required quality. For the analysis, the manufacturing process data is processed for a defined time frame, e.g. two production shifts. The high-dimensional process raw data are divided into sections according to process time or process geometry, standardized within these sections and processed accordingly using the Local Outlier Factor anomaly score. The results are displayed using an interactive heat map, which shows changes, their degree and their geometrical or time position in the course of the manufacturing process.

To conduct the experiments, we recorded and analyzed production data in welding series production. Using the data sets from different welding processes, it is shown that the developed approach can highlight process changes quickly without an initial training phase. To analyze the processes after targeted changes a new dataset is not required. Through the visual evaluation of the heat maps, plant engineers can decide whether the correction of a process is needed and can estimate the effects of changes. The approach also allows us to evaluate the relevance of selected features for process analysis. We demonstrate that with an appropriate choice of features, it is even possible to derive the quality of the manufactured components.

Paper PDF

Wjatscheslav Baumung: Application of Machine Learning and Vision for real-time condition monitoring and acceleration of product development cycles

11:20-11:40 o'clock

(Paper ID: 1192)

Development work within an experimental environment, in which certain properties are investigated and optimized, requires many test runs and is therefore often associated with long execution times, costs and risks. This can affect product, material and technology development in industry and research. New digital driver technologies offer the possibility to automate complex manual work steps in a cost-effective way, to increase the relevance of the results and to accelerate the processes many times over. In this context, this article presents a low-cost, modular and open-source machine vision system for test execution and evaluates it on the basis of a real industrial application. For this purpose a methodology for the automated execution of the load intervals, the process documentation and for the evaluation of the generated data by means of machine learning to classify wear levels. The software and the mechanical structure are designed to be adaptable to different conditions, components and for a variety of tasks in industry and research. The mechanical structure is required for tracking the test object and represents a motion platform with independent positioning by machine vision operators or machine learning. An evaluation of the state of the test object is performed by the transfer learning after the initial documentation run.

The manual procedure for classifying the visually recorded data on the state of the test object is described for the training material. This leads to an increased resource efficiency on the material as well as on the personnel side since on the one hand the significance of the tests performed is increased by the continuous documentation and on the other hand the responsible experts can be assigned time efficiently. The presence and know-how of the experts are therefore only required for defined and decisive events during the execution of the experiments. Furthermore, the generated data are suitable for later use as an additional source of data for predictive maintenance of the developed object.

Paper PDF

Stefan Wellsandt: Interacting with a Digital Twin using Amazon Alexa

11:40-12:00 o'clock

(Paper ID: 1101)

The Digital Twin is an evolving concept with many facets and applications in, for instance, engineering simulation, system control, and product-centric information management. This article focuses on the latter where literature uses the Product Avatar concept to refer to a product's digital counterpart. Such an avatar used to have one or more graphical interfaces to support user interactions with information about a product item. Over the last few years, voice user interfaces became more mature, and companies, such as Amazon and Google, used them to create digital assistants that support their users during tasks or by taking them over directly.

This paper focuses on the hypothesis that a company could use a voice-enabled digital assistant to interact with item-level information. Our study used product tracking and tracing, and quality control in the production as a realistic application case. The design of the assistant bases on the information needs outlined in the Electronic Product Code Information Services (EPCIS) standard. We implemented this design in a small-scale demonstrator on an Echo Show 5 smart speaker with an integrated touch display and an embedded Amazon Alexa assistant. This paper concludes that significant technological barriers, such as low transcription accuracy for object identifier information and the handling of factory noise, remain. A significant non-technological barrier is the mistrust regarding the closed voice assistant technologies from companies, such as Amazon and Google. An approach to address the latter barrier is to use open technologies, such as the privacy-focused assistant Mycroft or Mozilla's transcription solution DeepSpeech. Further research and experiments with these technologies are useful to identify how they can support the interaction with Digital Twins.

Paper PDF

David Barton: Self-aware LiDAR sensors in autonomous systems using a convolutional neural network 

12:00-12:20 o'clock

(Paper ID: 1159)

Autonomous systems, as found in autonomous driving and highly automated production systems, require an increased reliability in order to achieve their high economic potential. Self-aware sensors are a key component in highly reliable autonomous systems. In this paper we highlight a proof of concept (PoC) of a deep learning method that enables a LiDAR (Light detection and ranging) sensor to detect functional impairment. More specifically, a deep convolutional neural network (CNN) is developed and trained with labelled LiDAR data in the form of point clouds to classify the degree of impairment of its functionality.

The results are statistically significant and can be regarded as a general classifier for objects within LiDAR data, applied to selected cases of sensor impairment. In detecting impairment and evaluating the correctness of the captured data, the sensor gains a basic form of self-awareness. The presented methods and insights pave the way for improved safety of autonomous systems by the means of more sophisticated “self-aware” neural networks.

Paper PDF

Thorsten Wuest: Data-driven context awareness of smart products in discrete smart manufacturing systems

12:20-12:40 o'clock

(Paper ID: 1138)

Traditionally, smart-connected products are predominantly utilized during the usage phase of the product lifecycle. However, we argue that there are distinct benefits of system-integrated sensor systems during the beginning of life, more specifically in manufacturing and assembly. In this paper, we analyze the ability of a smart-connected product with an integrated sensor system to recognize and label different manufacturing processes, generating a distinct process fingerprint within a discrete smart manufacturing system. The ability of the smart-connected product to detect distinct manufacturing process patterns (‘process fingerprint’) enables the production planner and operator, e.g., to optimize the scheduling, improve part quality, and/or reduce the energy footprint. The experimental setup is based on a FestoDidactics CPlab with eight different manufacturing processes. The smart-connected product is equipped with a sensor system providing data from eight different sensors (e.g., temperature, humidity, acceleration). We used an Artificial Neural Network (ANN) algorithm to create a model to detect specific events/patterns within the dataset after labelling it manually over the course of a complete production cycle.

The focal manufacturing process was the heating tunnel where the smart-connected product was exposed to a heat treatment process and sequence. The results of this prototypical implementation indicate that a smart-connected product can reliably recognize specific process patterns with a system-integrated sensor system during a simulated manufacturing process. While this work is only a first step, the potential applications and benefits are promising and further research should focus on the potential quality implications within smart manufacturing of product-integrated sensor readings compared to machine tool-based sensors, both of which monitored during the beginning of life. Smart products’ integrated sensor systems provide the means to obtain measurements relevant for smart manufacturing systems that are not obtainable with common external sensor systems today.

Paper PDF

Stefan Bosse: Self-adaptive Traffic and Logistics Flow Control using Learning Agents and Ubiquitous

12:40-13:00 o'clock

(Paper ID: 1201)

Traffic flow optimisation is a distributed complex problem. Traditional traffic and logistics flow control algorithms operate on a system level and address mostly switching cycle adaptation of traffic signals and lights. This work addresses traffic flow optimisation by self-adaptive micro-level control by combining Reinforcement Learning and rule-based agent models for action selection with a new hybrid agent architecture. I.e., long-range routing is performed by agents that adapt their decision making for re-routing on local environmental sensors. Agent-based modelling and simulation are used to study emergence effects on urban city traffic flows with learning agents. The approach and the proposed agent architecture can be generalised and applied to a broader range of application fields, e.g., logistics and general transport phenomena.

Paper PDF

Discussion and Comments