Process force determines productivity, quality, and safety in milling. Current approaches of process design often focus on a priori optimization. In order to enable online optimization, the establishment of active force controllers is required. Due to fast-changing engagement conditions of the tool in conjunction with the slower machine dynamics, classic control is not suited. A promising approach is the application of model predictive control (MPC) for force control, which is proposed in this contribution. The model predictive force controller (MPFC) explicitly takes into account a model to predict the immediate future. It consists of a model of the machine tool and a separate model of the process. The process model describes the relation between feed velocity of the tool, force, and geometric properties of the tool, such as the radial deviation, and of the tool/workpiece engagement. The feedback loop of the controller is closed by an online identification of the process model to account changes in the material properties or of the tool wear state. For this identification an ensemble Kalman filter (EnKF) is applied. The MPFC solves an optimization problem on the future behavior in each sampling step to determine the optimal controller output enabling high dynamic control. The proposed control system is validated experimentally and compared with a conventionally designed process with constant feed. It can be shown that the manufacturing time is reduced by 50%. The system enables a paradigm shift in the design of milling processes operating the manufacturing process at its technological limit.

The productivity of CNC machining and milling in specific is limited by the endurable forces of the working tools. An increased feed velocity also induces higher active forces on the tool, such that the feed velocity has to be maximized within dynamical limits for an optimal operation of the CNC machining center. Model predictive control (MPC) strategies, together with an according target selector enable this kind of optimized operation. The optimization is possible due to models, which describe the feed dynamics and the relationship between the feed and maximum force. However in such a scenario, an accurate modeling and control of the feed dynamics become more crucial, as the built-in routines within the machining center induce unknown nonlinearities. Thus, to achieve the aforementioned objectives, the authors propose a practical nonlinear model predictive control (PNMPC) strategy for milling, incorporating Support Vector Machines (SVM) for both, the target selector and the identification of the unknown nonlinearities. The results show an improved overall control performance with PNMPC, compared to a linear time-varying MPC (LTV-MPC) with a successively linearized model.

Model-based predictive control (MPC) describes a set of advanced control methods, which make use of a process model to predict the future behavior of the controlled system. By solving a—potentially constrained—optimization problem, MPC determines the control law implicitly. This shifts the effort for the design of a controller towards modeling of the to-be-controlled process. Since such models are available in many fields of engineering, the initial hurdle for applying control is deceased with MPC. Its implicit formulation maintains the physical understanding of the system parameters facilitating the tuning of the controller. Model-based predictive control (MPC) can even control systems, which cannot be controlled by conventional feedback controllers. With most of the theory laid out, it is time for a concise summary of it and an application-driven survey. This review article should serve as such. While in the beginnings of MPC, several widely noticed review paper have been published, a comprehensive overview on the latest developments, and on applications, is missing today. This article reviews the current state of the art including theory, historic evolution, and practical considerations to create intuitive understanding. We lay special attention on applications in order to demonstrate what is already possible today. Furthermore, we provide detailed discussion on implantation details in general and strategies to cope with the computational burden—still a major factor in the design of MPC. Besides key methods in the development of MPC, this review points to the future trends emphasizing why they are the next logical steps in MPC.

Data-driven learning methods represent a promising field of research to complement classical approaches in the area of control theory. Within the German cluster of excellence “Internet of Production” (IoP), model-based control strategies are researched using collective knowledge accumulated in shared databases, and adapted online according to sensor acquired data. With their inherent generalization ability and affinity for greybox modeling, Support Vector Machines (SVM) are very suitable for such online identification and adaption. However, the computational efficiency of the identification, while maintaining accuracy, is crucial for the real-time capability of the overall framework. This work compares different definitions of the learning problem with SVM for the identification of dynamic systems. Computational efficiency within the given framework is thereby of particular interest. In addition, an extension of existing libraries by transfer learning capabilities is investigated to further speed up the recurrent online identification scheme. The results suggest that SVM with “Sequential Minimal Optimization” (SMO) qualify as a real-time capable general purpose identification approach for model-based control of the derived framework. The addition of transfer learning heavily contributes to the real-time capability.

Nonlinear friction is the limiting factor in using motor current signals to estimate the load of machine tools. The inertia of the axis and the positional dependency of the friction add another degree of complexity. The work focuses on industrial machining centers with ball-screw driven stages as they are used in metal cutting. The approach uses Internal low-frequency signals from the NC controller to keep the barriers for an industrial application at a minimum. The contribution of this study is twofold: First, it extends conventional analytic friction models so that they incorporate positional dependency of friction, as well as the contribution of the inertia of the axis. Second, it proposes how to model the both effects jointly through support vector regression. This data-driven model outperforms the extended Stribeck and the generalized Maxwell-slip friction models, which serve as a representative benchmark for static and dynamic friction models respectively. However, this comes with the need for a careful selection of the data, on which the support vector machine is trained, in order to obtain an accurate and general model.

This paper extends the ensemble Kalman filter (EnKF) for inverse problems to identify trending model coefficients. This is done by repeatedly inflating the ensemble while maintaining the mean of the particles. As a benchmark serves a classic EnKF and a recursive least squares (RLS) on the example of identifying a force model in milling, which changes due to the progression of tool wear. For a proper comparison, the true values are simulated and augmented with white Gaussian noise. The results demonstrate the feasibility of the approach for dynamic identification while still achieving good accuracy in the static case. Further, the inflated EnKF shows a remarkably insensitivity on the starting set but a less smooth convergence compared to the classic EnKF.

Force determines the product quality, the productivity and the safety of a milling process. Mechanistic force models are the key to understand, optimize or control the cutting process. They combine the undeformed chip parameter with empiric tuning coefficients in a gray-box model. Identifying those coefficients is costly in both, time and number of experiments. This paper introduces two recursive identification methods for force model identification: recursive least squares and ensemble Kalman filters. The model is nonlinear. The ensemble Kalman filter shows an extraordinary robustness against measurement noise and a fast convergence time -depending on the selection of the ensemble size and the measurement noise. The recursive least squares fit serves as a benchmark but is highly sensitive to measurement noise. It is the first time that a continuous identification is examined for mechanistic force models in milling.

Advanced learning methods enable the model-based control of systems with complex unknown dependencies. Within the German cluster of excellence “Internet of Production”, a configuration for an interconnected data-base is proposed, where data-driven model-based control strategies can be applied using the collective knowledge and adapted online according to data. For the exchanged data it is imperative to establish a generalizing learning technique for the controller design. A machine learning technique with inherent generalization ability is the Support Vector Machines (SVM) algorithm, where the choice of kernel is crucial for the resulting model quality. In the related literature, usually a radial basis function (RBF) is chosen as kernel, although many studies show the necessity of a more sophisticated kernel selection. This work tackles the point of a kernel selection based on composite kernel search in context of data-driven model-based control of a CNC machining center. The results support the capability of the presented approach to further automate and improve the identification of the controller model for the machining center.

Mechanistic force models are popular to describe the force in cutting technology. Process simulation, process optimization, and process control rely on the accuracy of these models. Standard identification techniques are not capable of identifying a mechanistic force model on-line and in hard real-time. However, it is necessary to adjust the model to increasing tool wear, e.g. in a model predictive controller for force control in milling. This work introduces the ensemble Kalman filter to the field of force model identification in cutting technology - enabling for the first time a continuous parametrization of mechanistic force models. The approach shows high accuracy and fast convergence in spite of the presence of measurement noise. The novel approach is validated statistically using 1000 random initial distributions and different ensemble sizes. The ideal, simulated force signals is augmented with different levels of noise (signal-to-noise ratios of 50, 15, five). Nevertheless, the filter converges within three, eight, and 32 revolutions of the tool respectively.

This paper presents how to classify the wear state of an end-mill based on force and current signals of the linear axes. The data is divided in binary classes based on the maximum flank wear. A support vector machine and a random forest are trained on orthogonal cutting experiments, but the validation is performed on arbitrary tool paths. To achieve this unique level of generality the signals are transformed into the rotation tool coordinate system. The features are extracted over five cutter revolutions. Support vector machines outperform random forests achieving 99,8% and 97% accuracy in the two classes on the test data and 98% and 61% accuracy on the validation data.

Today’s manufacturing systems are either optimized for flexible or individualized manufacturing. The machine operator determines the optimal setup for the machine variables that are accurately implemented by the machine controllers. However, the overall objective is productivity under the restriction of product quality, where a model-based predictive controller is used to rather control the process than machine settings. This approach requires an accurate model of the dynamic behavior of the machine tool. Therefore, the Support Vector Machines algorithm is applied to identify and model the dynamic behavior under unknown nonlinearities. This model is compared to a classical modeling approach to predict the future system behavior in a model-based predictive controller. Based on the prediction, an optimization problem is solved in order to determine an optimal feed velocity. The presented approach outperforms the previous control strategy with 15 % shorter manufacturing time. Although the identified nonlinear SVM model is far more accurate for the analyzed system than previous models, further research has to be conducted regarding the application within a model predictive control strategy.

Milling is a manufacturing process for machining metals, where a milling tool cuts metal from a workpiece. Especially during rough milling the productivity is significantly affected by the process force that acts on the tool. The process force depends on the feed velocity and the engagement conditions. The latter are defined by the path that the tool takes through the metal. Consequently, the feed velocity of the cutter is a suitable manipulated variable. A force model describes the relation between the process force, the engagement conditions and the feed velocity. A separate machine model describes the behavior of the machine tool axis and its numerical controller. Based on these two models a model predictive controller (MPC) is examined, which controls the feed velocity with respect to the predicted process force. The model predictive force controller is applied and validated on a machining center. The mentioned approach is the first implementation of a MPC which explicitly controls the active force in milling.

Designing manufacturing systems requires a profound understanding of the manufacturing process and its challenges to meet final customer requirements. Considering future objectives already at an early design stage increases the flexibility of the manufacturing system and its robustness regarding changed boundary conditions. Today’s manufacturing systems rather control machine settings than process variables or even product quality. The major barrier for quality control is that in most manufacturing processes, quality cannot be measured on-line. Model-based self-sptimization (MBSO) has been developed to overcome this limitation. A combination of embedded process knowledge and tailored sensor integration enables for on-line quality estimation. The overall objective is to control key characteristics of product quality in a broad manufacturing landscape. This work describes a guideline of how to design an MBSO system with examples at each stage of the development process.

Mechanistic force models allow an estimation of the force components in cutting technology. This is essential for an accurate simulation of the force, an analysis of the tool load, or a model-based predictive force control. The model coefficients do not only depend on the material and the tool but also on the engagement condition. This requires an online identification of those coefficients. To determine those coefficients, this paper uses curve fitting based on the instantaneous uncut chip thickness, a method that has only recently gained momentum in milling. This work applies this technique to the well-established, nonlinear Kienzle force model. The paper reviews a broad range of nonlinear, derivative-free optimization algorithms for this least-squares curve-fitting problem evaluating accuracy and runtime. The evaluation is conducted on 121experiments with distinct process conditions resulting in a statistically verified conclusion. The results indicate that in spite of a heterogeneous optimization space, local constrained algorithms are most suited for model identification.

Support Vector Machines (SVM) is a machine learning algorithm with inherent generalization ability and a convex optimization problem. This paper studies the application of the SVM method for the online identification of the nonlinear dynamic behavior of the feed velocity in a CNC machining center. Both blackbox and greybox modeling approaches are tested for this purpose. Within the German Cluster of Excellence "Integrative Production Technology for High-Wage Countries", a modelbased predictive control (MPC) strategy with a linear state-space model is already implemented for the feed velocity of the CNC machining center. Due to nonlinearities, the model of the controlled system has to be identified and updated during the process. Therefore, the SVM method should be used to recurrently identify a model in every time-step. Additionally, the identified models should be capable of being formulated in a linear state-space model. The methodology is validated with measured data from the CNC machining center. The gained results for the blackbox and greybox approaches show only small deviations from the measured behavior of the system.

Milling is one of the most flexible and productive manufacturing processes for machining metals. In the case of rough milling as much material as possible should be removed in as little time as possible. Therefore, a high cutting force is desirable. The maximum force is thus a suitable control variable to reduce the manufacturing time. The cutting force is related to the feed velocity. The relationship can be described by force models. They in turn can be used to determine the maximum feed velocity for a given force. This maximum feed velocity can be used as a reference, which shall not be exceeded. Hence, a Model-based Predictive Controller (MPC) manipulates the desired feed velocity of the machine with respect to the machine behavior. This MPC shows good performance in the case of feed velocity references in time-domain. Though, the feed velocity reference depends on the position of the cutting tool through the workpiece. Hence, an approach shall be described which enables force control in position-domain. Therefore, the existing MPC is extended to a 2-Layer-MPC. An additional MPC works as a reference generator transforming the reference in position-domain into time-domain. Furthermore, an approach is described which allows position dependent feed velocity control without an additional MPC. Finally, the presented approaches are implemented and validated on a real machining center. (C) 2017, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.

Kurzfassung Die Zerspankraft liefert in spanenden Fertigungsverfahren eine wichtige Information über den Prozesszustand. Trotz optimal ausgelegter Prozesse wird selten eine konstante Zerspankraft erreicht, da Unsicherheiten, wie z. B. Werkzeugverschleiß oder Materialkenngrößen, diese beeinflussen. Aus Schutz vor Überlast, werden die Prozessparameter so gewählt, dass die vom Werkzeug ertragbare Last auch zum Standzeitende nicht überschritten wird. Damit einher geht ein Produktivitätsverlust bei schneidscharfem Werkzeug. Durch Regelung der Prozesskraft kann eine signifikante Produktivitätssteigerung erreicht werden. Die modellbasierte prädiktive Regelung (MPR) erlaubt die explizite Berücksichtigung von Nichtlinearitäten sowie zeitvariantem Übertragungsverhalten und ermöglicht so eine deutlich höhere Regelgüte als klassische PI-Regler.