Norbert Link

Gesture recognition with inertial sensors and optimized DTW prototypes

In this work our approach for human gesture recognition with inertial sensors is presented. The proposed method utilizes a dynamic time warping (DTW) algorithm for online time series recognition. Our DTW implementation is able to deal with gesture signals varying in amplitude and to resolve ambiguities in the recognition result when DTW is used for multiclass classification.

Indoor 3D position estimation using low-cost inertial sensors and marker-based video-tracking

In this paper, a system for indoor 3D position tracking with an inertial measurement unit and a marker-based video tracking system utilizing external cameras is presented. Similar to an integrated navigation system, 3D position, velocity and attitude are calculated from IMU measurements and aided by using position corrections from the video tracking system. The measurements from both sensor sources are fused with an extended Kalman filter model, which incorporates the estimation of IMU biases for drift compensation during video outages. The performance of the filter approach has been tested with simulated data and the whole system has been evaluated with real data from a hand tracking scenario. By means of the combination of inertial sensors and vision-based position tracking, the proposed system is able to overcome video measurement outages over short periods of time as well as drift problems of the IMU.

A System‐Oriented Approach for the Optimal Control of Process Chains under Stochastic Influences

Process chains in manufacturing consist of multiple connected processes in terms of dynamic systems. The properties of a product passing through such a process chain are influenced by the transformation of each single process. There exist various methods for the control of individual processes, such as classical state controllers from cybernetics or function mapping approaches realized by statistical learning. These controllers ensure that a desired state is obtained at process end despite of variations in the input and disturbances. The interactions between the single processes are thereby neglected, but play an important role in the optimization of the entire process chain. We divide the overall optimization into two phases: (1) the solution of the optimization problem by Dynamic Programming to find the optimal control variable values for each process for any encountered end state of its predecessor and (2) the application of the optimal control variables at runtime for the detected initial process stat...

Optimal Process Control through Feature‐Based State Tracking Along Process Chains

The optimal control of a manufacturing process aims at control parameters that achieve the optimal result with least effort while accepting and handling uncertainty in the state space. This requires a description of the process which includes a representation of the state of the processed material. Only few observable quantities can usually be measured from which the state has to be reconstructed by real-time capable and robust state tracker models. This state tracking is performed by a mapping of the measured quantities on the state variables which is found by nonlinear regression. The mapping also includes a dimension reduction to lower the complexity of the multi-stage optimization problem which is approximately solved online. The proposed generic process model provides a universal description that can be adapted to specific data from simulations or experiments. We show the feasibility of the generic approach by the application to two deep drawing simulation models.

Revealing effects of dynamic characteristics in process monitoring and their influences on quality prediction

In manufacturing, quality evaluation of final process results often involves destructive testing methods which lead to workpiece loss. In contrast, we observe process dynamics and relate them to final quality results as a nondestructive method for quality prediction. Once the relations between process dynamics and a quality criterion have been established, they can also be applied to process control. This is achieved by finding adequate process parameters to obtain predefined process dynamics and related final quality results. In this paper, we present a method for modeling observable process dynamics. Process dynamics are described by parametric functions derived from a priori process knowledge. The observed dynamics are reduced to a compact parameter representation by Nonlinear Curve Fitting of the parameters. These extracted characteristic quantities describe the crucial dynamic behavior and are related to a quality criterion by Principal Least Squares Regression. This enables the prediction of the fin...

Meta-process modeling methodology for process model generation in intelligent manufacturing

The present paper details a novel methodology called Meta-Process Model that is able to generate new data-based models for manufacturing processes when no experimental data is available. For that purpose, the concept of Hyper-Models was used to create a higher level of abstraction of these manufacturing processes, along with a Statistical Shape Model (SSM) that is able to capture the modes of shape variations and build up a deformable model to generate new shapes. The main premise of the present work is to interpret a process model as a n-dimensional shape and use SSM to capture the variations among a set of different process models. This methodology is evaluated by using two already existing process models for a model generalization, from which a new process model is derived just with new, given process conditions. This new process model is then compared with a process model, which was independently estimated using real experimental data acquired under the same process conditions. The results show that a previously nonexistent process model that captures the dynamics of the real process can be generated, even when theres no experimental data and only the new process conditions are available.

From models to hyper-models of physical objects and industrial processes

Feasible process parameters of a given task in industrial production can be extracted from mathematical process models. An appropriate general method is the task-to-method transform which is applied to Laser Welding. The process model is the function of task-defining quantities over the space of process parameters. Different classes of operating conditions and tasks are represented in different corresponding process models. In this paper we propose the construction of a common model representation. The parameters of these representations form a model space, in which a model of the models can be formed, the hyper-model. General ideas of hyper-model formation are proposed and how dedicated models for specific conditions and tasks can be derived from it. An outlook is given how such hyper-models will be used in different application fields and how the information represented by the hyper-models can be exploited.

Improving Expert Knowledge in Dynamic Process Monitoring by Symbolic Regression

To avoid destructive testing methods in the evaluation of final process quality results in production environments, we monitor and evaluate the process dynamics to make assumptions about the associated final process quality. This information can again be used in process control to adapt the process parameters according to the observed and given reference quantities. In our approach, we propose a method for modeling the observable process dynamics. This can be modeled by parametric functions which can either be determined solely by expert knowledge and Nonlinear Curve Fitting using an additional correction term that is found via Symbolic Regression. The obtained model parameters characterize the process dynamics and can be used to detect abnormal process behavior in order to adapt the process parameters by a control unit. For a proof of concept, we have applied the proposed approach to experimental resistance spot welding data.