key: cord-0330032-8ux7bw7h authors: Schorr, Sebastian; Möller, Matthias; Heib, Jörg; Bähre, Dirk title: Quality Prediction of Drilled and Reamed Bores Based on Torque Measurements and the Machine Learning Method of Random Forest date: 2020-12-31 journal: Procedia Manufacturing DOI: 10.1016/j.promfg.2020.05.127 sha: 61ad792be2af4022df5156373ae7fc8cc4f36586 doc_id: 330032 cord_uid: 8ux7bw7h Abstract Increasing amount of available process data in the manufacturing industry together with analysis methods like machine learning provides new possibilities to increase the manufacturing efficiency and to rethink existing process structures. For example, assessing the workpiece quality in an early machining stage can be used to alter the quality control strategy, to increase the product quality, and to reduce the number of scrap parts which leads to savings in terms of time, resources, and cost. In this work, torque measurements obtained from the numerical control of a milling-machine in the serial production of hydraulic valves are used to predict the concentricity as well as the diameter of drilled and reamed bores of the valves. Statistical features are determined out of the torque measurements, which are presented as time series. The prediction of the quality is achieved with the machine learning method of random forest (RF) on the basis of the extracted features. The Pearson correlation between the features and the quality characteristics as well as the learning curves of the RF method are studied. It turns out, that a strong correlation comes along with a fast decreasing learning curve of the RF but gives no information over the achievable prediction accuracy with the RF. The obtained predictions are very precise and evaluated with four statistical criteria. A mean absolute error of 17.1 µm for the concentricity and only 0.27 µm for the diameter is achieved. In addition, the coefficients of determination of the concentricity and the diameter are 96.3% and 94.1% respectively, very high. It can be stated, that for the considered use case in this paper, a precise quality prediction on the basis of torque measurements and RF could be implemented which would make a much faster and efficient quality control possible. Through the integration of more and more sensors into machine tools as well as the accessibility of data from numerical controllers (NC), an increasing amount of process data for each workpiece is available [1] . The rising availability of affordable database storage and computing power as well as increasing knowledge about machine learning allow manifold usage of the process data in the manufacturing industry [2] . The analysis of the gathered process data can contribute for example to a more sustainable manufacturing by saving energy and avoiding waste [2] , improving process planning [1] or guarantee product quality [3] . Manufacturing companies are forced to search for new solutions and approaches to overcome challenges like decreasing competitiveness of production locations in high-wage countries, shortened product life cycles, increased product variances, and enhanced requirements on the product quality. Machine learning is seen as a contributing solution for the mentioned challenges and a way to increase the overall efficiency of production. This paper focuses on the prediction of the workpiece quality (drilled and reamed bores) in an early machining stage by using torque measurements from a milling machine and applying machine learning algorithms on it. Knowing the product quality close to real time not only guarantees high product quality or the possibility to adjust process parameters but also helps to avoid unnecessary machining steps and costs by removing a scrap workpiece directly after revealing its low quality [2] . The idea of the prediction of the final quality at early manufacturing stages has been discussed for example by Through the integration of more and more sensors into machine tools as well as the accessibility of data from numerical controllers (NC), an increasing amount of process data for each workpiece is available [1] . The rising availability of affordable database storage and computing power as well as increasing knowledge about machine learning allow manifold usage of the process data in the manufacturing industry [2] . The analysis of the gathered process data can contribute for example to a more sustainable manufacturing by saving energy and avoiding waste [2] , improving process planning [1] or guarantee product quality [3] . Manufacturing companies are forced to search for new solutions and approaches to overcome challenges like decreasing competitiveness of production locations in high-wage countries, shortened product life cycles, increased product variances, and enhanced requirements on the product quality. Machine learning is seen as a contributing solution for the mentioned challenges and a way to increase the overall efficiency of production. This paper focuses on the prediction of the workpiece quality (drilled and reamed bores) in an early machining stage by using torque measurements from a milling machine and applying machine learning algorithms on it. Knowing the product quality close to real time not only guarantees high product quality or the possibility to adjust process parameters but also helps to avoid unnecessary machining steps and costs by removing a scrap workpiece directly after revealing its low quality [2] . The idea of the prediction of the final quality at early manufacturing stages has been discussed for example by Through the integration of more and more sensors into machine tools as well as the accessibility of data from numerical controllers (NC), an increasing amount of process data for each workpiece is available [1] . The rising availability of affordable database storage and computing power as well as increasing knowledge about machine learning allow manifold usage of the process data in the manufacturing industry [2] . The analysis of the gathered process data can contribute for example to a more sustainable manufacturing by saving energy and avoiding waste [2] , improving process planning [1] or guarantee product quality [3] . Manufacturing companies are forced to search for new solutions and approaches to overcome challenges like decreasing competitiveness of production locations in high-wage countries, shortened product life cycles, increased product variances, and enhanced requirements on the product quality. Machine learning is seen as a contributing solution for the mentioned challenges and a way to increase the overall efficiency of production. This paper focuses on the prediction of the workpiece quality (drilled and reamed bores) in an early machining stage by using torque measurements from a milling machine and applying machine learning algorithms on it. Knowing the product quality close to real time not only guarantees high product quality or the possibility to adjust process parameters but also helps to avoid unnecessary machining steps and costs by removing a scrap workpiece directly after revealing its low quality [2] . The idea of the prediction of the final quality at early manufacturing stages has been discussed for example by Through the integration of more and more sensors into machine tools as well as the accessibility of data from numerical controllers (NC), an increasing amount of process data for each workpiece is available [1] . The rising availability of affordable database storage and computing power as well as increasing knowledge about machine learning allow manifold usage of the process data in the manufacturing industry [2] . The analysis of the gathered process data can contribute for example to a more sustainable manufacturing by saving energy and avoiding waste [2] , improving process planning [1] or guarantee product quality [3] . Manufacturing companies are forced to search for new solutions and approaches to overcome challenges like decreasing competitiveness of production locations in high-wage countries, shortened product life cycles, increased product variances, and enhanced requirements on the product quality. Machine learning is seen as a contributing solution for the mentioned challenges and a way to increase the overall efficiency of production. This paper focuses on the prediction of the workpiece quality (drilled and reamed bores) in an early machining stage by using torque measurements from a milling machine and applying machine learning algorithms on it. Knowing the product quality close to real time not only guarantees high product quality or the possibility to adjust process parameters but also helps to avoid unnecessary machining steps and costs by removing a scrap workpiece directly after revealing its low quality [2] . The idea of the prediction of the final quality at early manufacturing stages has been discussed for example by 48th SME North American Manufacturing Research Conference, NAMRC 48 (Cancelled due to WAGNER ET AL. [4] and WUEST ET AL. [5] regarding machining operations or by ARIF ET AL. [6] for the semiconductor manufacturing. These works presented only concepts but no application. NETO ET AL. [7] predicted the diameter of bores, CALEB ET AL. [8] diagnosed faulty drills and MATHEWS ET AL. [9] predicted the hole quality of reamed bores. In each case, the predictions were achieved by using process data and artificial neural networks but their research was neither integrated into a manufacturing line nor had any specific application. We will show in this research paper that it is possible to implement an effective and economic quality control for an existing use case in the manufacturing industry by using only machine learning along with already available torque values of a NC. Hence, in this paper we describe an approach to gathering process data directly from a NC as well as determine the required data to achieve good prediction results by studying the Pearson correlation and the learning curves of the applied machine learning method. The machine learning method of random forest (RF) was used to predict the diameter and concentricity of drilled and reamed bores. The achieved prediction results were very precise. The use case considered in this research paper is about quality prediction of drilled and reamed bores of hydraulic valves at Bosch Rexroth AG in Homburg (Germany). The process data were obtained from a milling-machine in the serial production during the machining of the pre-casted valve housings made of gray cast iron. Hydraulic valves are characterized by narrow tolerances to allow sealless fits and to prevent oil leakage. Even slight quality deviations can cause high scrap rates and financial losses. Therefore, quality control close to real time that does not increase cost is preferred. First, each housing is machined, then the valve is assembled and finally an end of line testing is carried out as depicted in Fig. 1 . Sample inspection with industrial metrology is used to control the machining process in the present stage. The sample inspection covers only a low percentage value of all valves but still results in high costs. In addition, the latency between the machining of a valve and its measurement results allows no direct feedback about the machining process and causes the risk to manufacture waste housings until the measurement results are known. In addition, a lot of resources, time, and money is lost if a valve is first detected as a waste part during the end of line testing. To increase the transparency of the machining process and to discover the quality of the housings close to real time, a quality prediction based on process data and machine learning methods is pursued. The aim for the data collection is to use the process data (e.g. torque, current, speed) already available in a NC without the integration of any additional sensors. This allows a fast transfer of the technology to further machining centers and to reduce maintenance cost. Hence, a feasible and cost-effective in-process quality surveillance solution for machining processes under industrial conditions is achieved. Drilling is a machining method to create bores by a rotating motion of a drill along a single feed motion of the rotary axis. It is a prerequisite for reaming and has an influence on the quality of the reamed bore. The kinematics of reaming is equal to drilling but it is a bore finishing method to reach high dimensional and surface quality [9, 10] . It is the last machining process after drilling by widening the bore by a few tenths of a millimeter. Wear on the reamer and the drill that occurs during operation causes geometrical, dimensional, and surface faults [10] . The tool wear is detectable through increased cutting force, torque, power, vibration, or acoustic emission [7, 8, 11] . These values can be used to control and to predict the quality of a bore during machining. Direct and indirect techniques are used to monitor machining processes. A high degree of accuracy is achieved with direct techniques because the variable of interest is measured directly with integrated measuring equipment in the machining center. Under industry conditions, direct measurement methods are often not applicable due to limited accessibility of the workpiece and tool or use of cooling lubricant. Hence, indirect monitoring solutions are preferred in industry albeit they are less accurate. Auxiliary quantities like torque or current of the drives are measured which correlate with the desired variable. [9, 12, 13] Increasing tool wear or process instabilities cause fluctuation of cutting forces, which is noticeable by changing current and torque of the drives. The current and torque obtained from the drive controller can be used to predict the cutting forces of a milling process as proved in the ReffiZ research project [14] . The deviation between the predicted cutting forces out of the current of the drives and the measured forces at the workpiece was less than 10%. As described in many research papers, process data obtained directly from the drive can be used to monitor machining conditions [12, 13, 14, 15, 16, 17] and in combination with machine learning e.g. tool wear, chatter marks, or workpiece quality can be predicted [1, 7, 9, 18] . The analysis of torque measurements with machine learning in this paper aims not only to monitor the machining process Machining Quality control with industrial metrology Assembling Testing Casted housing Product but also to predict the quality to reduce the number of sample inspections and to establish a 100% quality check at an early machining stage close to real time and with minimal cost. There are many different types of machine learning methods that can be applied for various applications. For each application, the most suitable method has to been determined for example during a grid search. In our use case a machine learning method is needed which can predict the desired quality characteristics with a minimal error. We ran a grid search to find the best machine learning method and its parameter configuration. We tested convolutional neural networks (CNN), artificial neural network (ANN), support vector regression (SVR), Random Forests Regression (RF), and Ada Boost Regression (ABR) for the prediction of the bore quality. In Fig. 2 the mean absolute errors (MAE) achieved with a 10-fold cross validation for the best configuration of the machine learning methods are shown. The MAE values of the diameter from the CNN and ANN are not good and only depicted for the sake of completeness. The best prediction results (lowest MAE values) were obtained from the ensemble method RF and it was therefore used in this paper to predict the diameter and the concentricity of the bores. BREIMAN developed the machine learning method of RF that can be applied for regression and classification problems [19] . It can process huge amounts of data, learns fast, is robust against overfitting, and can reach very precise predictions [20, 21] . To predict numerical values for the quality characteristics, a RF approach suitable for regression tasks is used in this paper. A RF consists of N single decision trees. Each tree is grown on an individual training data set (bootstrap sample) obtained from the original training data set. The generation of these samples is called bagging and is an important characteristic of RF to avoid overfitting. At each node of a tree only a subset of all available variables are considered for splitting the node. The split of the node is done for the variable that minimizes errors e.g. the mean squared error. The splitting is stopped when the number of remaining samples comes under a predefined threshold. After each of the N trees is grown on the basis of the bootstrap samples, each tree predicts the output variable yn on basis of the same input data xnew. The final prediction result ynew of the RF is the average of all single predictions yn [22, 23] . Fig. 3 shows the architecture of a RF. The best configuration of a RF (e.g. number of trees N, the splitting conditions, etc.) is obtained by a design of experiments or in the language of machine learning a so-called grid search. The machining of the contemplated bore in this paper is done with one spiral drill and one reamer. The spiral drill drills up the pre-casted bore close to the final diameter and shapes the geometrical form of the bore. The reamer creates the final diameter and surface. We gathered process data from 216 bores belonging to three batches (data gathering was carried out on three different days) to get a more heterogeneous data set by considering the variability of the product quality in serial production. The casted housings themselves belonged to one batch and the machining conditions (e.g. milling machine, NC program, temperature, cooling lubricant) were the same on the three days of the machining of the 216 housings. Only the remaining tool life of the used tools at the beginning of the machining process differed between the three batches as depicted in Fig. 4 . The tool life is stated in percentage and was determined empirically regarding how many bores could have been machined with the required quality in the past. In this use case we did not consider the tool life as a separate feature in the machine learning model because it was not directly accessible from the NC. We wanted to develop a prediction model, which can achieve fully automated predictions on basis of the available data of the NC without any human interaction. Therefore, we focused only on the actual torque values. 5 shows the whole approach to the quality prediction applied in this paper at a glance. First, the actual torque values of the z-axis and spindle were gathered for each bore and eventually diameter and concentricity were measured as described in Section 4.1 and 4.2. Subsequently, features were calculated out of the torque values (Section 4.3) and the Pearson correlation coefficients of the features were determined. Studying the learning curves of the RF method shows which axis contributes most to the learning progress of the RF (Section 4.4) . Finally, all the data were split into training and testing data sets and the quality prediction with RF was done (Section 5). To avoid downtime of a highly utilized four-axis milling machine in a manufacturing line and to use an easy transferable solution for data collection no additional sensors for data acquisition were integrated into the machine. Therefore, the already available actual torque values of the motor controllers, which are sent over the field bus to the NC to control the machining process, were collected. The machine integrated NC system IndraMotion MTX 14V16 works with an interpolation cycle frequency of 1,000Hz. This is the highest frequency to collect machine internal data with the used NC. The software MTX ewb Recorder installed on the NC built-in PC was used to collect and to store the data. The actual torque values of the z-axis and the spindle were collected during drilling and reaming of each bore. In total, we collected data from 216 bores. The duration of the drilling operation was 20 seconds and the reaming operations took 7.3 seconds for each bore. Using a sampling frequency of 1,000Hz, each bore is represented by a 27,300 dimensional vector for the z-axis as well as for the spindle. Hence, the raw data are in a shape of a 216 x 54,600 dimensional matrix. Feature extraction will be applied in this paper (as described in Section 4.3) to reduce the amount of data. The distribution of the mean torque values for the drilling operation carried out by the spindle is depicted by Fig. 6 for each batch. All three batches can be well distinguished because of the different ranges of the torque values for each batch. The second batch has the widest spread where some values are in the range of batch 1 and batch 3 but most of the values are between the two batches. The diameter and concentricity of each of the 216 bores in the three batches were determined with the coordinate measuring machine ZEISS PRISMO ultra. The measuring machine was calibrated before starting the measurements of the workpieces of each batch. Fig. 7 shows the box plots of data from the three batches based on the concentricity and diameter. It is observed, that the concentricity values are decreasing from batch to batch. A very small interquartile range of the third batch is remarkable. The diameters from the first and the third batch are nearly within the same range compared to the second batch, where the diameters are much bigger. Determination of features is recommended to obtain relevant information for the quality prediction and to reduce the amount of the raw data to enable data processing close to real time [24] . The aim is to represent the time series data by relevant features to achieve a precise and effective quality prediction. Features can be divided into time domain based statistical features and frequency domain based features. VUNUNU ET AL. used frequency domain based features of collected sound signals to distinguish between healthy and faulty drills by an artificial neural network [8] . NETO ET AL. predicted diameters in drilling processes by using quadratic means of various sensor signals [7] . According to TETI ET AL. the most commonly used time domain features for monitoring machining operation are: arithmetic mean, effective value, standard deviation, skewness, kurtosis, signal power, range, peak values, and crest factor [12] . We determined standard deviation, mean, skewness, kurtosis, minimum, and maximum value from the gathered torque values from the spindle and the z-axis. A look at the calculated features reveals that the minimum and maximum values have very little statistical spread over all 216 bores. These comes, due to a peak of the torque signal at the beginning of the machining process and a value close to zero at the end. We neglect them because irrelevant features can reduce the prediction accuracy of the machine learning model [25] . Hence, we obtained finally 16 features {four time domain features * two tools * (spindle + z-axis)} which lead to a data matrix with the size of 216 x 16 because of the 216 bores. The Pearson correlation coefficient for each feature were calculated and depicted in a correlation matrix (Fig. 8) . It is a measure for the strength of linear dependency between two values but it ignores any existing nonlinear dependency. The number of features can be reduced by the ones that are highly correlated but the risk rises to lose any important information which is essential for the prediction of the quality [24] . The green color in Fig. 8 indicates high positive and the red color high negative correlation. It can be determined, that the standard deviation, mean, and kurtosis from the drilling operation of the spindle and the standard deviation of the reaming operation of the spindle is very positively correlated. The skewness and kurtosis for both machining operations regarding the z-axis is strongly correlated, too. In addition, it can be seen, that the features obtained from the torque values of the spindle are mainly strongly positively correlated with the concentricity (except the skewness for drilling), but are weakly correlated with the diameter of the bore. Regarding the features for the z-axis no clear linear correlation with the both quality characteristics can be determined. Hence, the concentricity is mainly linearly correlated with the torque measurements from the spindle but for the diameter no clear linear correlation with the features of the spindle and the z-axis exists. Based on the revealed correlations the features of the z-axis are less relevant for the predictions and can be removed from the overall number of features because irrelevant information can deteriorate the prediction accuracy. Before reducing the number of features we will study the learning curves to understand the impact of these features on the prediction results obtained with RF. Now, all the required data are available to run a machine learning project. The input data are the calculated features and the output data are the measured quality characteristics. For the machine learning method the RF is used as explained in Section 3.3. We trained an individual RF model for each quality characteristic. However, the configuration (e.g. number of trees, splitting conditions, etc.) of both RF models determined during grid search were the same. Each RF consists of 100 trees, at each node the square root of all available features was considered for splitting, the mean absolute error was used to determine the feature to split the node, and the minimum number of samples to split a node was set to three. To train the model 85% of the data were used and with the remaining 15% the testing was carried out. Regarding the calculated correlation of Fig. 8 , the features from the spindle and z-axis might contribute with a different weight to the prediction of the quality characteristics. To determine if both axis (spindle and z-axis) are necessary we trained three RF models for each quality characteristic. One considers only the features of the z-axis, other considers only the features of the spindle, and the third model considers the features of both axis. To evaluate the impact of the axis on the performance and the behavior of the RF model, we studied the learning curves which are shown in Fig. 9 . Learning curves depict the learning progress (y-axis) over the training set size (x-axis). They are divided into a training curve which depicts how well a model is learning on the basis of the training dataset and a testing curve that shows how well a model generalized regarding the test dataset. The learning curves in Fig. 9 were obtained with a 10-fold cross validation. The mean absolute error (MAE) was chosen to evaluate the learning progress of the RFs regarding the training set size and represents the training as well as the testing error. The training curves were received from the training data set after it was used to train the RF. The testing curve was obtained from the data, which the RF has not seen before (test data set) using the trained RF on the basis of the individual training set size. The learning curves for the concentricity (Fig. 9 a) converge very fast and with only 40 training samples the training error is very close to its minimum of 10 µm. With more training samples the testing error decreases further until it reaches a value of 22 µm. This is achieved if the features from the spindle and z-axis or only the features from the spindle are used. For both combinations, the testing errors only differ slightly. The errors are much greater (40 µm) and the decreasing is much slower only if the features from the z-axis are considered. A similar picture can be drawn for the prediction of the diameter, but it needs much more training samples (approx. 150) to reach a training error that not further decreases. The testing curves are decreasing as long as more training samples are used to train the RF. Relying only on the features from the z-axis leads to a testing error, which is much greater than any other combination of axis. The lowest testing error of 0.33 µm is reached only if the features from the spindle are considered for predicting the diameter. To sum up, the features from the spindle and z-axis should be considered together to predict the concentricity. To predict the diameter of a bore, the features from the spindle might be sufficient. Adding more training data will likely contribute to a further decreasing testing error where the effect might be stronger for the diameter than for the concentricity. These findings will be used in this paper to analyze the prediction performance in more detail. To predict the concentricity and the diameter of the bores the same configurations of the RFs were used as for the determination of the learning curves. For the prediction of the concentricity, the features from the z-axis and the spindle were considered, and for the prediction of the diameter, only the features from the spindle were used. The RFs were trained with the training data set (85% of the data) and after the training the predictions were obtained from the testing data set (15% of the data). The prediction accuracy was evaluated using the mean absolute error (MAE), the mean absolute percentage error (MAPE), the maximum error, and the coefficient of determination (R 2 ). Comparing the statistical evaluation of the prediction results (Table 1) for the concentricity with the diameter shows that the diameter could be predicted with a higher accuracy. The maximum single prediction error of all the predicted diameters is only 0.72µm, an MAPE of only 0.002% is reached, and the MAE is with 0.27µm very little, too. For all these statistical criteria the results for the concentricity are little worse but still sufficient and precise enough for the quality prediction in this use case. R 2 value is higher only for the concentricity than for the diameter but both the values of 96.3% and 94.1% obtained respectively, are equally good. For both quality characteristics is the MAE lower than the corresponding value in the charts of the learning curves. This is due to the applied 10-fold cross validation and the averaging of MAEs to create the learning curves. The standard deviation or maximum and minimum values for MAE were not depicted in the learning curves. Hence, the MAE of a set of predictions can be better than the given mean value by the learning curve, which is the case in this example. In addition to Table 1 , some graphical representations of the prediction results are depicted in Fig. 10 . For each individual bore of the testing data set the measured quality as well as the predicted quality is depicted. The three batches can be very well distinguished (also compare Fig. 6 and Fig. 7) , on the basis of the measured as well as the predicted values. Though all the three batches were combined to one data set to train the RF and to predict the quality characteristics, the RF was able to learn on the basis of the features and to predict the concentricity and diameter with minor errors. Fig. 10 . Representation of the measured and predicted concentricity as well as the diameter for the test data set of each batch. The scattering of the measured values within each batch influences the accuracy of the predictions. For example, the measured values of the concentricity of the third batch are very close to each other compared to the both other batches. Hence, the scattering of the predictions is little for the third batch. In addition, the three batches can also be determined, for example from the ranges of the mean torque values of the spindle (see Fig. 6 ) which has a correlation of 86% with the concentricity (see Fig. 8 ). In addition, the measured diameters of the first and the third batch belong nearly to the same range of the scale. The RF had some difficulties to predict the small and wide diameters of the third batch. However, the RF discovered the decreasing trend of the diameters within all the three batches and delivered very good predictions. In this paper, an approach to predict the quality of drilled and reamed bores of hydraulic valves in an early machining stage in a serial production was presented. The machine learning method of random forest was used to predict the concentricity and the diameter of the bores on the basis of the torque measurements of the spindle and z-axis of a millingmachine. In total, 16 features were calculated for each of the 216 bores out of the gathered torque measurements. The Pearson correlation between the features and the quality 17 characteristics, the learning curves of the RF, and the prediction results were analyzed. A comparison of the results from the Pearson correlation and the learning curves revealed, that the features from the z-axis correlated only weakly with both quality characteristics and that the learning effect as well as the prediction results were very poor if only the features from the z-axis were used to train the RF. The features from the spindle were strongly correlated with the concentricity but weakly correlated with the diameter. If the features from the spindle were used to train the RF, the MAE for the concentricity decreased very fast to a lower limit but for the diameter the descent was less strong. The features from the spindle and the z-axis together delivered slightly better prediction results for the concentricity but no improvement for the diameter. Thus, the linear correlation between the features and the quality characteristic influences the descent of the MAE of the learning curves. This leads to a reduced number of required training data to reach the lowest possible limit for the MAE that can be achieved with the machine learning method RF. Drawing conclusions from the linear correlation to the prediction accuracy obtainable with the RF was not possible. Very good prediction results were achieved for both quality characteristics independent of the linear correlation. Finally, the accuracy of the predictions was evaluated with the performance measures MAE, MAPE, maximum error, and R 2 . Very precise predictions were achieved which demonstrated that process data obtained from a NC together with machine learning can be used to predict and monitor the quality of bores. In the future, we will research about which minimum data collection frequency can be used to achieve the same prediction accuracy. Furthermore, the number of extracted features will be increased by adding more statistical features and by also considering frequency domain based features as well as features like tool condition and machining settings. Finally, the quality prediction will be implemented in the serial production of highly precise hydraulic valves to detect failure at an early machining stage and to contribute to a more efficient, competitive, and sustainable manufacturing. Self-optimizing cutting process using learning process models A review of machine learning for the optimization of production processes A review of data mining applications for quality improvement in manufactruing industry Function-Oriented Quality Control Strategies for High Precision Products An approach to monitoring quality in manufacturing using supervised machine learning on product state data Cascade Quality Prediction Method Using Multiple PCA+ID3 for Multi-Stage Manufacturing System Neural models for predicting hole diameters in drilling processes Automatic Fault Diagnosis of Drills Using Artificial Neural Networks Neural-network approach for predicting hole quality in reaming Manufacturing Processes 1 Effect of tool wear on quality in drilling of titanium alloy Ti6Al4V, Part I: Cutting Forces, Burr Formation, Surface Quality and Defects Advanced monitoring of machining operations Monitoring and control of manufacturing processes: A review Realisierung effizienter Zerspanprozesse, Ergebnisbericht des BMBF Verbundprojekts ReffiZ Fault detection in machining via torque estimation. IFAC Manufacturing, Modeling Tool breakage detection using support vector machine learning in a milling process Real time monitoring of the CNC process in a production environment-the data collection & analysis phase A technical note on integrated product quality model using artificial neural networks Random Forests Analysis of a Random Forests Model Random Forests A Comparative Study on Machine Learning Algorithms for Smart Manufacturing: Tool Wear Prediction Using Random Forests Improvements to random forest methodology. Doctoral dissertation Feature-based Supervision of Shear Cutting Processes on the Basis of Force Measurements: Evaluation of Feature Engineering and Feature Extraction A Few Useful Things to Know About Machine Learning