Embargoed until 2025-03-02
Author
Date
2022Type
- Doctoral Thesis
ETH Bibliography
yes
Altmetrics
Abstract
Thermal errors, which are among the most relevant sources of inaccuracies of produced workpieces, are a limitation for precision manufacturing. For this reason, a variety of energy-intensive measures, such as extensive machine cooling or precise air conditioning in the shop floor, are used to reduce the thermal errors of machine tools. This work develops a self-learning thermal compensation strategy that enables a transformation from resource-based to intelligence-based methods for reducing thermal errors of machine tools. The developed compensation strategy is based on the thermal adaptive learning control, which combines on-machine measurements with data-driven models. It includes adaptive model inputs, on-demand triggered on-machine measurements and automatically set action control limits to define the need for model updates. The use of the self-learning thermal error compensation increases the degree of automation and enables completely autonomous model generation, so that no specific prior knowledge of the thermal behaviour of the machine tool is required.
The developed method of adaptive input selection realises an automated and adaptive selection of inputs for data-driven models. This enables a situation- and location-specific adaptation of the used model inputs to create an optimal representation of the thermal errors. Two methods are developed to realise adaptive model inputs for data-driven models. The first method is based on a statistical signal comparison and is composed of a separate input selection and modelling step. The second method combines Group-LASSO for auto-regressive models with exogenous inputs and particle swarm optimisation for the simultaneous identification of the optimal model inputs, model structure and model parameters. To analyse these methods, a statistical analysis is conducted using a test series with approximately 600 hours of measurements from a 5-axis machine tool. The results show that the prediction precision evaluated by the peak-to-peak value of the volumetric errors at the workspace positions is improved between 18% and 20% compared to models with static inputs. The method comparison illustrates that the Group-LASSO based method reduces the root mean squared error by another 25% from 5.4 µm to 4.1 µm compared to the first method and has a 50% lower standard deviation within the statistical analysis. This shows that considering the model structure in the input selection process increases the performance of the resulting data-driven models.
The trade-off between precision and productivity of the thermal adaptive learning control is overcome by integrating a novelty detection model. The integrated novelty detection model is based on a one-class support vector machine and triggers on-machine measurements when previously unknown thermal states occur. Consequently, thermal states that are not included in the training data of the compensation models are detected by the novelty detection model. The statistical analysis shows that the non-productive time due to the on-machine measurements for checking the action control limits is reduced by up to 72% compared to hourly on-machine measurements without a significant loss of precision. Compared to the same total number of periodic on-machine measurements, the on-demand triggered on-machine measurements convince with a better compensation quality.
The developed self-learning thermal error compensation are applied to model thermal errors of the linear and rotary axes of 5-axis machine tools. These errors are identified with a specifically developed on-machine measurement cycle based on a precision sphere and a touch trigger probe. The precision and robustness of the compensation results of thermally compensated 5-axis machine tools are evaluated using two 5-axis machine tools with different kinematic chains. The analysed test series with a duration of approximately 700 and 800 hours include both single and multiple axis movements with and without the use of metal working fluid. The peak-to-peak values of the resulting thermal errors at the different working space positions are reduced from 76 µm to 17 µm and from 84 µm to 33 µm, respectively. This corresponds to a reduction of the thermal errors of 78% and 61%, respectively.
For the further evaluation, the self-learning thermal compensation is applied to a 5-axis test piece to analyse the compensation results under manufacturing conditions. The 5-axis thermal test piece enables the separation of the table-related thermal errors that change the effective direction during 5-axis machining. The compensation results for the analysed thermal load cases with and without metal working fluid illustrate that the self-learning thermal error compensation reduces the thermal position errors and the dominant orientation errors. The average thermal position errors are reduced by 71% from 26 µm to 8 µm for the load case without cooling lubricant and by 32% from 18 µm to 12 µm for the load case with metal working fluid.
Finally, the effectiveness of the self-learning thermal error compensation for reducing thermal influences in 5-axis simultaneous milling is shown by manufacturing two uncompensated and two compensated impellers. The self-learning thermal error compensation reduces the maximum root mean squared error of the impeller blades by up to 73% from 31.6 µm to 8.7 µm at ambient temperature variations of about 10 °C. Thus, the self-learning thermal error compensation significantly increases the process capability of 5-axis machine tool under fluctuating ambient temperatures. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000601225Publication status
publishedExternal links
Search print copy at ETH Library
Publisher
ETH ZürichOrganisational unit
03641 - Wegener, Konrad (emeritus) / Wegener, Konrad (emeritus)
More
Show all metadata
ETH Bibliography
yes
Altmetrics