Session 9 Future of Manufacturing III

Friday, 13.11.2020, 08:45-11:00 o'clock

Introduction

08:45 - Intro: Conference & Day, Dirk Lehmhus

Presentations:


Vanessa Kirchner: Detecting Process Instabilities in Industrial Gas Metal Arc Welding Time Series

09:00-09:20 o'clock

(Paper ID: 1194)

Interdependencies in gas metal arc welding process parameters lead to changing quality outcomes of a seam. To minimize process instabilities, the process needs to be stable and predictable having an enormous process repeating stability. By preventing instabilities leading to an increased process repeating stability, the welding result is optimized and rejects can be inhibited.

Aiming to optimize the Overall Equipment Effectiveness, a method is introduced to automatically detect process instabilities within the recorded variables. An autoencoder architecture using long short-term memory cells is implemented to estimate the underlying distributions for multivariate time series. Characterizing reliably the important features, the latent space of the autoencoder as well as the reconstruction of the time series are used to gain information about process instabilities. By applying the trained model, abnormal deviations are located point precisely. A safety window is further enhancing the reliability and therefore the suitability to increase the transparency within the industrial process. The results initiate the possibility of stabilizing the welding process by reducing the occurred anomalies. Therefore, the detected process instabilities are visualized in a heatmap being a highly supportive visualization method including metrics to describe the deviations from the expected behaviour. Once trained, the neural network can be used continuously to analyse further measured live data from the welding process.

The transparency and predictability of the machine output massively increases by monitoring and interpreting the unexpected data points. The model can automatically detect unseen patterns within multivariate industrial welding data taken from the processing line, which turns out to be a complex task for the human brain. The approach implements a successful detection of process instabilities with an autoencoder architecture in an enormous and diverse real-world dataset taken from industrial welding machines and being invariant towards changing frequencies. The proposed architecture is valuable for engineers to support optimization of the process variables leading to a more reasonable welding outcome.

Paper PDF

Sascha Lauer: Data driven approach for Robot-assisted multi-pass-welding thick sheet metal connections

09:20-09:40 o'clock

(Paper ID: 1125)

In the field of shipbuilding and large steel construction, steel sheets with wall thicknesses ranging from 20 mm to more than 100 mm are processed. Due to the masses and geometric dimensions, the occurring workpieces usually cannot be moved. Manual and direct work at the joint is currently required. Joining can only be carried out economically using processes such as arc welding, which is only possible in multi-pass welding due to the required filling volumes. This process therefore occurs not only in the flat position to be produced, but also in the technically demanding welding positions in vertical and overhead orientation. If the individual welding positions vary during a contour, further challenges arise.

The welding bead doesn‘t form uniformly due to changes in seam geometry and the earth's gravitational force acting on the melt. The resulting discontinuities in the formation of the weld bead geometry add up with increasing number of beads in the multi-pass weld and lead to defects in the joint. In a manual welding process, this variance is compensated by the experience and skills of the welder. The torch alignment or the type of movement is adapted to the real geometry. Although an experienced welder has the ability to react to the process. He can’t keep up the constant welding speeds or torch orientations and thus with the reproducibility of an industrial robot. The robot lacks the ability of the welder to adapt.

The robot-assisted multi-pass welding combines the strengths of manual welding with the accuracy and speed fidelity of an industrial robot. The course of the weld seam is planned by a design program on ideal geometry and converted into a robot-program by a specially developed postprocessor. This robot-program serves as a reference movement. At previously defined path support points, which are determined by a change of direction, gradient or type of movement, measurements are carried out and the actual seam cross-sections are determined.

On the base of these measurements, a welder positions points on the respective seam cross section on a graphical user interface at which he would have manually set the individual weld beads. These offline marked points are used to generate the welding program for the next layer. The method is currently being validated on the base of a application example from the manufacture of large tubular connections. Two tubes are aligned at defined angles to each other. The large number of welds to be produced is associated with a high number of welding bead positions.

The programming and process data collected in the process form the foundation for a database regarding the multi-pass welding. Depending on the geometry of the components to be joined, there are different contours, seam geometries and welding positions, which ensure that process data is obtained with each weld. The long-term aim of the research is to automatically generate multi-pass welds using the welding technology-based database for thick sheets and to focus on automatically generating the layer build-up based on the actual geometry data of the welding beads.

Paper PDF

Julia Huuk: Simulation-based feed rate adaption considering tool wear condition 

09:40-10:00 o'clock

(Paper ID: 1162)

The process forces generated in machining are related to a deflection of the milling tool, which results in shape deviations. In addition to process parameters like feed rate, width and depth of cut or cutting speed, the wear condition of the tool has a significant influence on the shape deviation during flank milling. In process planning it is important to take the tool condition and the ideal time for tool change into account when selecting the process parameters. An assistance system is being researched at the Institute of Production Engineering and Machine Tools (IFW) in cooperation with Kennametal Shared Services GmbH to support this task. The assistance system adjusts automatically the feed rate considering a predefined maximum shape deviation. Additionally, it identifies an optimal moment for tool change.

The advantages of the system are particularly evident in planning of individual milling processes. The assistance system is based on a combination of a material removal simulation and empirical models of the shape error. For this purpose, spindle currents as well as measured shape errors are stored in a database. These data are extended by the actual local cutting conditions calculated by a process-parallel material removal simulation. Afterwards, the data is transferred into process knowledge via a Support Vector Machine (SVM).

Within a technological NC simulation before the start of manufacturing, the generated knowledge is applied to predict the shape error of the workpiece and to automatically adjust the feed rate. By adapting the feed rate, it is possible to control the tool life. The required tool change is defined by specifying a limit for the permitted width of flank wear land. The presented assistance system enables the prediction of the shape error parallel to the manufacturing process and the automatic determination of the feed rate as well as the ideal time for tool change.

Paper PDF

Dirk Lehmhus: A digital twin approach to predict and compensate distortion in a High Pressure Die Casting (HPDC) process chain 

10:00-10:20 o'clock

(Paper ID: 1186)

The HPDC industry is changing, and E-mobility is one reason for this – power train components traditionally produced by HPDC are eliminated as the industry abolishes internal combustion engines (ICE). The casting industry reacts to this challenge by extending its product range while keeping the customers. In practice, this means increasingly focusing on structural automotive components. These, however, tend to be large area/low thickness geometries prone to distortion as a result e. g. of inhomogeneous cooling. Compensating this effect by mechanical alignment, i.e. forming operations aimed at meeting the shape and positional tolerances, is done, but is costly.

The alternative proposed here is based on analysis of the full HPDC process chain and identification of the main process steps that control distortion. The assumption is that the respective parts undergo a solution heat treatment after casting and removal of the casting systems, followed by quenching and warm ageing. Among these individual steps, casting and cooling in the mold and after removal from it as well as the solution heat treatment and subsequent quenching carry highest potential for introduction of residual stresses and distortion.

Thus the approach suggested employs the final quenching step to realize an inhomogeneous cooling defined in terms of local heat extraction rates to

  • compensate previously introduced residual stresses and distortion
  • while still achieving the supersaturated solid solution required for warm ageing.

Practical realization of locally differentiated heat extraction rates is envisaged via multi-nozzle spray cooling.

The fundamental necessity behind this solution is process modelling and simulation which extends from casting all the way to spray cooling, covering all relevant intermediate steps in terms of their influence on part geometry and residual stresses.

Establishing this capability is a prerequisite for realizing the concept on production series level. The digital twin aspect comes in when taking into account systematic or stochastic variations in the controlling process parameters. To cover such variation, process simulation and derivation of parameter settings for an adaptive spraying system must be executed for each individual part’s production. The present study discusses general methodologies for establishing locally required cooling rates and their distribution over the part surface as a basis of spraying system control. The viability of the general concept is demonstrated based on initial numerical modeling and simulation of heat conservation and distortion potential.

Paper PDF

Laura Fabiola Palatto: A Methodology for the Optimization of Mechanical Properties of Automotive Iron-Casting Brakes using Artificial Neural Networks

10:20-10:40 o'clock

(Paper ID: 1202)

In the foundry process of automotive brakes, specialists use a simulation followed by a trial and error method to obtain desired mechanical properties of gray iron casting products. The difficulty of the problem increases due to the numerous process variables. Also, the accuracy of the prediction of mechanical properties using available solidification simulators is limited, which forbids a simulation-driven approach. The current strategy requires then a large amount of trial experiments, which raises the cost of the whole optimization process.

In this paper, we present a compound strategy using Artificial Neural Networks (ANN) along with Optimization Engineering to predict the mechanical properties of iron-casted products using a minimum number of physical experiments. A case study evaluates the proposed methodology to accurately predict the mechanical properties of an automotive iron-casting brake component using a minimum amount of physical experimental data.

Paper PDF

Discussion and Comments