AI-based Framework for Deep Learning Applications in Grinding
Improving manufacturing resilience with virtual sensors
Authors: Tobias Kaufmann, S. Sahay, Philipp Niemietz, Daniel Trauth, W. Maaß, T. Bergs
Preamble
This paper has originally been published in proceedings of IEEE 18th World Symposium on Applied Machine Intelligence and Informatics 2020.
Abstract
Rejection costs for a finish-machined gearwheel with grinding burn can rise to the order of 10,000 euros each. A reduction in costs by reducing rejection rate by only 5–10 pieces per year already amortizes costs for data-acquisition hardware for online process monitoring. The grinding wheel wear, one of the major influencing factors responsible for the grinding burn, depends on a large number of influencing variables like cooling lubricant supply, feed rate, circumferential wheel speed and wheel topography. In the past, machine learning algorithms such as Support Vector Machines (SVM), Hidden Markov Models (HMM) and Artificial Neural Networks (ANN) have proven effective for the predictive analysis of process quality. In addition to predictive analysis, AI-based applications for process design and control may raise the resilience of machining processes. In a manufacturing context, resilience describes the ability to machine a constant quality at reasonable costs and production time despite disturbances. Using machine learning methods may also lead to a heavy reduction of cost amassed due to a physical inspection of each workpiece still required by law today. With this contribution, information from previous works is leveraged and an AI-based framework for adaptive process control for machining technologies using a cylindrical grinding process is introduced. For the development of such a framework, three research objectives have been derived: First, the dynamic wheel wear needs to be modelled and measured, because of its strong impact on the resulting temperature and mechanical stresses and thus the workpiece quality. Second, models to predict the quality features of the produced workpieces depending on process setup parameters and materials used have to be established. Here, special focus is set on deriving models that are independent of a specific wheel-workpiece-pair. The opportunity to use such a model in a variety of grinding configurations gives the production line consistent process support. Third, the resilience of analytical models regarding graceful degradation of sensors needs to be tackled, since the stability of such systems has to be guaranteed to be used in productive environments. Process resilience against human errors and sensor failures leads to a minimization of rejection costs in production. To do so, a framework is presented, where virtual sensors, upon the failure or detection of an erroneous signal from physical sensors, will be activated and provide signals to the downstream smart systems until the process is completed or the physical sensor is changed.
Motivation and state of research
In 2009 the European Research area of the European Commission published a document about the strategy for a sustainable European machine tools industry (European Commission, 2009) that divides the research in the machine tools industry into five sub-projects. One of those five sub-projects, is “The Manufacturing Break-through” and it is written that “In this case four demonstrators are envisaged, illustrating self-calibration, predictive maintenance, versatile configuration, and improved control of accuracy and acceleration at high operating speeds.” [1]. It is observed that 20–25% of total cost towards any machining job comes from grinding processes [2]. Metallic high-performance alloys such as titanium or high-strength steels, require effective and efficient machining due to their heavy machinability. Because of the high demands placed on the geometric shape, surface and surface area, grinding is a highly effective manufacturing process in the finishing of camshafts or gearwheels, for example. The increasing production pressure and decreasing cycle times force production employees to process the workpieces faster in less time that inevitably leads to errors, as presented by the seven types of muda [3]. Since the majority of manufacturing processes finishes with grinding, it is highly important to safeguard the workpiece during this process as significant machining has already been applied beforehand. If there is an error or damage to the workpiece during grinding, the previous investment to the workpiece is lost. A consistent process support by an adaptive AI-based framework gives a significant benefit for the improvement of process resilience and thus to the aimed minimization of rejection costs in manufacturing processes. The high demand placed on process control regarding dimensional and form accuracy is a key factor, which enhance the development of an assistance system for improved process control in grinding. With the support of an AI-based framework for predicting process results, currently prescribed time-consuming and cost-intensive 100%-physical inspections in quality assurance can be significantly reduced. The mass information stream (big data) from the acquisition of direct and indirect process signals, which results from the investigation of cause-effect-relationships in the grinding process, sets special challenges for the data processing architecture that needs to consider both online and historical data. Previous studies have shown that the amount of data generated from production processes in series production is already around 19.4 GB, or 12 billion data items per hour, when only the machine-internal data is considered. If additional, external sensors are used, this data volume for an exemplary fine blanking process quickly reaches ranges of 2.78 TB per hour [5], while acquiring data at frequencies between 2.5 and 1000 kHz, resulting up to 6.2 Gbit/s [28]. In previous work, force and current consumption, as well as acoustic emission signals have been intensively used to investigate process states, cause-effect-relationships and evaluate quality features of the process output. Zhiqiang gives an overview about the application of AI methods regarding data mining and analytics in manufacturing processes [4]. As an additional process evaluation parameter, the wheel wear in grinding processes and especially the capability of monitoring systems to online measure the wheel wear turned out to be of significant value for the process outcome [6]. Thereby process state values measured by acoustic emission sensors, force sensors and current consumption were used to monitor wheel wear and to identify cause-effect-relationships between input parameters, e.g. wheel speed and feed rate, and output parameters, e. g. surface quality, geometry and form. Therefore, discrete wavelet analysis [7] and support vector machines have been applied on the provided data [8] for discretization of the process signal and outlining correlations between direct and indirect process signals.
Over the years multiple approaches both theoretical [9] and intelligent systems-based approach has been applied, with the aim to streamline and monitor the grinding process [10]. These studies have used different algorithms such as Fuzzy logic, ANN+ Genetic Algorithm and Hidden Markov Models among others. The majority of the previous scientific work provide a specific wheel-workpiece-pair based system, which implies that the system cannot be generalized. An extensive literature survey, do not show resilience of machine to singular/multiple sensor failure, i.e. if a singular physical sensor fails the whole system breaks down as outlined in the following Fig. 1.
Since the modelling of wheel degradation during the process is a limiting factor [14] due to the semi-handmade production of the wheel and the alternating unique grain shapes, a monitoring and modelling system for degradation is inevitable for an effective process control. Fig. 2 shows the three different phases of wheel wear.
The attempted techniques that model wheel wear have used technologies such as acoustic emissions [15], optical sensors [16] among others. These technologies are functional but do not provide for live monitoring but in most cases, rather a binary output i.e. whether the wheel is dull or not. A summary of the state-of-the-art research dealing with wheel wear is given in Table 1.
The attempted techniques that model wheel wear have used technologies such as acoustic emissions [15], optical sensors [16] among others. These technologies are functional but do not provide for live monitoring but in most cases, rather a binary output i.e. whether the wheel is dull or not. A summary of the state-of-the-art research dealing with wheel wear is given in Table 1.
The remainder of this contribution is organized as follows. First a concept of an edge computing approach for the acquisition, sustainable storage and analysis of process data streams is presented. Second, an outlook on a demonstrating proof of concept using various data sources for a cylindrical grinding process is sketched. The acquired process data contains machine control states and external sensor data to monitor forces, coolant lubrication pressure and temperature. For the development of AI-applications to assist the grinding process, three main research objectives in the field of AI-application in grinding were identified: i) modelling dynamic wheel wear, ii) developing a generalized model independent of the wheel-workpiece-pair, iii) developing a resilient predictive layer for a graceful degradation. The dynamic wheel wear analysis is focused on the simultaneous surveillance of lubrication and force parameters and machine states. In the following sections challenges, methods and approaches for each research question will be discussed.
Concept for an AI-based framework in grinding
The current section outlines the three mentioned researched objectives to verify the research hypothesis:
An AI-based framework for adaptive process control will improve the process resilience by embedding a fault resistant virtual sensor approach into the machine data acquisition to continuously predict quality and process features and to empower the operator with real time advices.
Achieving the proof of work for the virtual sensor approach, the planned concept for Deep Learning applications in grinding technology, which uses modern sensor data provision based on the Industrial Internet of Things will be investigated and implemented.
A. Modelling dynamic wheel wear
Modelling dynamic wheel wear is an integral aspect in grinding as the surface quality of the workpiece and the geometric accuracy depends heavily on the sharpness and the condition of the wheel. By using a sharpened wheel the highest specific removal rates can be achieved, while fulfilling both, the surface roughness and quality requirements [24]. Thus, in-process monitoring of the wheel is a preliminary to predict the surface roughness and the surface thermal damage. With the intention to minimize high rejection costs, the use of AI-based methods for in-process optimization of the various influencing variables is necessary. Motivated by previous work, c.f. section 1, the combination of different sensors such as sound and acoustic emission in the modelling of wear using various machine learning approaches is researched.
For modelling the wheel wear, proper data sets that, in the best case, represent a continuous change of wear online during the process are needed. In future work, the possibility to scan and check conventional grinding wheels with non-ionising terahertz radiation, also called millimetre waves, for cracks and inhomogeneities will be researched. The approach follows the research hypothesis, that the precise detection and analysis of the global and local density distribution within the grinding wheel by means of millimetre wave technology enables the prevention of cracks due to inhomogeneities and to investigate the influence of the manufacturing process on the formation of inhomogeneities and thus on crack formation in grinding wheels. Scanning grinding wheels with terahertz radiation was lately enabled by an experimental setup by BeckerPhotoniks GmbH and used in previous work at the research institute [17]. In combination with the acoustic sensors and inline laser monitoring of radial grinding wheel wear and macrotopography, this could lead to a highly reliable wheel monitoring system. Upon integration of an online datastream of wheel wear, the information can be set in correlation with a model of the grinding wheel structure based on the volumetric composition according to Barth [18]. Following Barth, the macro- and microstructure of the dressed grinding wheel has a significant influence on the thermo-mechanical load collective in the grinding process and therefore also on the workpiece and surface quality. Thus, it is necessary to consider and integrate the knowledge about grinding wheel surface parameters into the machine learning environment of the framework system to be developed. If influencing variables are identified for a specific machining step, a real-time correlation between known input and previously unknown output variables can be generated in future steps. This knowledge can also be used to optimize machine tools according to the thermo-energetic design. At the same time, costs by not-in-order parts and unnecessary dressing cycles can be significantly reduced.
B. Development of a generalized model independent of wheel-workpiece-pair
For the development of effective monitoring systems, it is important to understand the basic wheel-workpiece relation independent of the specificity of the pair. To achieve this, it is necessary to analyse different wheel behaviour on various workpiece materials and geometries. Based on this data, a generic model can be developed and combined with the concepts of transfer learning to be fine-tuned to achieve the model performance for specific wheel-workpiece-pairs. The primary objective for developing these models is to predict the workpiece quality at the end of the grinding phase (spark-out), reducing the production and assembly cycle times by cutting down on the prescribed physical quality assurance to virtual quality assurance.
The implementation of such generic models requires methods that accurately model the process, but at the same time offer flexibility for fine-tuning the general model to specific use cases. Apart from classical time series analysis and prediction models, artificial neural network has proven to be promising in this area [27] and especially recurrent neural networks and convolutional neural networks, often used for human activity recognition, have already been applied in similar context for sensory data analysis. Both methods either could be used standalone for the analysis of the raw data signal, or combined with prior feature engineering pipelines [25]. In context of the research project, both methods will be implemented and evaluated to model the quality of workpieces for generic wheel-workpiece pairs.
C. Development of a resilient predictive layer for graceful degradation of systems
The concept of a resilient predictive layer is inspired by the ability of the human brain to be resilient in case of the loss of functionality of a particular region in the brain. It reacts against the lack of information or other faults, first by heightened processing of other sensory signals and secondly by allowing for relearning the lost functionality by several other parts of the brain [26]. In a sensor-based predictive model, this phenomenon is envisioned by exploiting the relationship between different sensors, for example, the values from force sensors are known to be related to the spindle load. Using this particular example and exploring other relationships between sensors, a set of virtual sensors whose output models the behaviour of a specific sensor, based on the input of other sensor signals, can be implemented. These virtual sensors can be thought of as a predictive algorithm which overtime learns from other sensors and act as a replica of an actual physical sensor in case of failure. These virtual sensors can, in case of failure of a certain physical sensor, kick-in to provide input to the actual predictive analysis model. Regarding the example of the spindle load and the resulting forces applied on the workpiece, if the force sensor fails, the input from the spindle load can be used to predict the force applied on the workpiece and thus the thermo-mechanical load, which the workpiece has seen. This information can then be used for the ongoing prediction of the workpiece quality and resulting wheel wear induced by the force. A schematic representation is provided in Fig. 3 (a) and Fig. 3 (b). Virtual sensors that model the behaviour of real sensor with a high accuracy also unveil that the sensors data may be redundant and can be discarded, or a weighting of various sensors data that reflect a cause-effect relationship can be determined by data-driven determination. In this context, the value of the actual sensor signal in comparison to the virtual sensor will be evaluated. Further on, the possibility of using these virtual sensors, once trained, as a way to detect anomalies in the physical sensor data, informing the operator of a possible breakdown of any physical sensor can also be explored.
The described test system equipped with a data acquisition system will be used to acquire versatile process data including varying wheel and workpiece material as well as different process setups during the grinding of workpieces known in the automotive industry. Based on the expected findings in the research environment, the customer can then use the acquisition of indirect process signals via modern machine controls to derive the surface properties, quality and shape tolerances of the workpiece as well as an adaptive dressing cycle to machine at highest ressource efficiency. A cost- and time-intensive 100 %-inspection of needed quality features, as is prescribed for safety-critical components, can thus be reduced to statistical process control (SPC) or the enabled virtual quality assurance. The concept of the virtual sensors and the concept of graceful degradation described above, if they perform to the expected objectives, can thus be expanded to a production line, where multiple sensors from multiple machines can be explored for relationships based on the output of one machine, which will serve as the input to the other.
References
[1] P. Zelinski, “Labor Savings in Grinding,” 2013.
[2] S. Malkin and C.Guo, “Grinding Technology and Applications of Machining with Abrasives,” New York, NY: Industrial Press Inc., 2008.
[3] A. Gnanavelbabu, and P. Arunagiri, “Ranking of MUDA using AHP and Fuzzy AHP algorithm,” in Materials Today: Proceedings, vol. 5, pp. 13406–13412, March 2017.
[4] Z. Ge, Z. Song, S. X. Ding and B. Huang, “Data Mining and Analytics in the Process Industry. The Role of Machine Learning”, in IEEE Access, vol. 5, pp. 20590–20616, September 2017.
[5] R. Glebke, et al, “A Case for Integrated Data Processing in Large-Scale Cyber-Physical Systems,” Proceedings of the 52nd Hawaii International Conference on System Sciences, pp.7252–7261, 2019.
[6] D. Zhang, G. Bi, Z. Sun and Y. Guo, “Online monitoring of precision optics grinding using acoustic emission based on support vector machine,” in The International Journal of Advanced Manufacturing Technology, vol. 80, pp.761–774, September 2015.
[7] Z. Yang, Z. Yu, “Grinding wheel wear monitoring based on wavelet analysis and support vector machine,” in The International Journal of Advanced Manufacturing Technology, vol. 62, pp.107–121, September 2012.
[8] P. Wang, Q. Meng, J. Zhao, J. Li and X. Wang, “Prediction of Machine Tool Condition Using Support Vector Machine,” in Journal of Physics: Conference Series, vol. 305, July 2011.
[9] I. Marinescu, W. Rowe, B. Dimitrov, and I. Inasaki, “Tribology of Abrasive Machining Processes,” Norwich, NY: William Andrew Pub, 2004.
[10] A. Nandi, and D. Pratihar, “Design of a genetic-fuzzy system to predict surface finish and power requirement in grinding,” in Fuzzy Sets and Systems, vol. 148, pp.487–504, December 2004.
[11] M. Sedighi and D. Afshari, ”Creep feed grinding optimization by an integrated GANN system,” J. Intell. Manuf., vol. 21, pp. 657–663, December 2010
[12] Q. Yang and J. Jin, “Study on Machining Prediction in Plane Grinding based on Artificial Neural Network,” Proceedings of the IEEE International Conference on Intelligent Systems and Knowledge Engineering, p. 15–16, 2010.
[13] T. Liao, G. Hua,J. Qu, and P. Blau, “Grinding wheel condition monitoring with boosted minimum distance classifiers,” in Mechanical Systems and Signal Processing, vol. 22, pp. 217–232, January 2008.
[14] B. Linke, “Review on Grinding Tool Wear With Regard to Sustainability,” in Journal of Manufacturing Science and Engineering, vol.137, December 2015.
[15] N. Ding, C. L. Zhao, X.C. Luo and J. Shi “An Intelligent Grinding Wheel Wear Monitoring System Based on Acoustic Emission,” in Solid State Phenomena, vol. 261, pp.195–200, August 2017.
[16] E. Brinksmeier, and F. Werner, “Monitoring of Grinding Wheel Wear,” in CIRP Annals, vol. 41, pp. 373–376, 1992.
[17] http://www.becker-photonik.de/produkte/3d-scanner [08.07.2019].
[18] S. Barth, and F. Klocke, “Influence of the Grinding Wheel Topography on the Thermo-Mechanical Stress Collective in Grinding,” in Inventions, vol. 2, p. 34, December 2017.
[19] S. Kamps, S. Goetz, T. Auerbach, G. Keitzel, S. Gierlings and D. Veselovac, “Vorsprung durch Wissen — Wertschöpfung durch Metainformationen in der Produktionstechnik,“ in Virtuelle Instrumente in der Praxis 2015, R. Jamal, R. Heinze, Begleitband zum 20. VIP-Kongress.
[20] http://zone.ni.com/reference/de-XX/help/370858P-0113/dlgscript/script_dlg_menu/tss_datapreprocessor_general/ [26.06.2019].
[21] T. Maksoud and M. Atia, “Applications of artificial intelligence to grinding operations via neural networks,” in Machining Science and Technology, vol. 7, pp. 361–387, February 2003.
[22] T. Liao, G. Hua, J. Qu and P. Blau, “Grinding wheel condition monitoring with hidden Markov model-based clustering methods,” in Machining Science and Technology, vol. 10, pp. 511–538, February 2007.
[23] A. Hosokawa, K. Mashimo, K. Yamada, and T. Ueda, “Evaluation of Grinding Wheel Surface by Means of Grinding Sound Discrimination,” in JSME International Journal Series C, vol. 47, pp. 52–58, 2004.
[24] F. Klocke and W. König, “Appropriate Conditioning Strategies Increase the Performance Capabilities of Vitrified-Bond CBN Grinding Wheels,” in CIRP Annals, vol. 44, pp. 305–310, 1995.
[25] N. Hammerla, S. Halloran and T. Plötz, “Deep, Convolutional, and Recurrent Models for Human Activity Recognition using Wearables,” in Proceedings of the International Joint Conference on Artificial Intelligence, vol. 25, pp.1533–1540, 2016.
[26] H. Duffau, “Brain Plasticity and Reorganization Before, During, and After Glioma Resection,” Glioblastoma, pp. 225–236, 2016.
[27] N. S. M. Sahid, M. M. Rahman, and K. Kadirgama, “Neural Network Modeling of Grinding Parameters of Ductile Cast Iron Using Minimum Quantity,” in International Journal of Automotive and Mechanical Engineering, vol. 11, pp. 2608–2621, July 2015.
[28] T.Bergs et al., “Stamping Process Modelling in an Internet of Production,” in CIRP TESConf, 2019.
[29] S. Haag, N. Pyschny and C. Brecher, “Flexible Assembly Robotics for Self-optimizing Production,” in Intelligent Robotics and Applications, Jeschke S., Liu H., Schilberg D., Eds. Berlin: Springer, 2011, pp. 189–198.
[30] R. Cleveland, W. Cleveland, J. McRae, I. Terpenning, “STL: A Seasonal-Trend Decomposition Procedure Based on Loess,“ in Journal of official statistics, vol. 6 no. 1, 1990, pp. 3–73.
