Abstract
This paper presents a time-efficient approach to the drill wear classification problem that achieves a similar accuracy rate compared to more complex and time-consuming solutions. A total of three classes representing drill state are recognized: red for poor state, yellow for elements requiring additional evaluation, and green for good state. Images of holes drilled in melamine faced chipboard were used as input data, focusing on evaluating differences in image color values to determine the overall drill state. It is especially important that there are as few mistakes as possible between the red and green class, as these generate the highest loss for the manufacturer. In green samples presented in gray-scale, most pixels were either black (representing the hole) or white (representing the chipboard), with very few values in between. The current method was based on the assumption that the number of pixels with intermediate values, instead of extreme ones, would be significantly higher for the red class. The presented initial approach was easy to implement, generated results quickly, and achieved a similar accuracy compared to more complex solutions based on convolutional neural networks.
Download PDF
Full Article
Time-efficient Approach to Drill Condition Monitoring Based on Images of Holes Drilled in Melamine Faced Chipboard
Albina Jegorowa,a,* Izabella Antoniuk,b Jarosław Kurek,b Michał Bukowski,c Wioleta Dołowa,d and Paweł Czarniak a
This paper presents a time-efficient approach to the drill wear classification problem that achieves a similar accuracy rate compared to more complex and time-consuming solutions. A total of three classes representing drill state are recognized: red for poor state, yellow for elements requiring additional evaluation, and green for good state. Images of holes drilled in melamine faced chipboard were used as input data, focusing on evaluating differences in image color values to determine the overall drill state. It is especially important that there are as few mistakes as possible between the red and green class, as these generate the highest loss for the manufacturer. In green samples presented in gray-scale, most pixels were either black (representing the hole) or white (representing the chipboard), with very few values in between. The current method was based on the assumption that the number of pixels with intermediate values, instead of extreme ones, would be significantly higher for the red class. The presented initial approach was easy to implement, generated results quickly, and achieved a similar accuracy compared to more complex solutions based on convolutional neural networks.
Keywords: Chipboard machining; Melamine faced chipboard; Tool condition monitoring; Drill wear classification
Contact information: a: Institute of Wood Sciences and Furniture, Warsaw University of Life Sciences WULS-SGGW, Poland; b: Institute of Information Technology, Warsaw University of Life Sciences WULS-SGGW, Poland; c: no affiliation, Poland; d: no affiliation, Poland;
* Corresponding author: albina_jegorowa@sggw.edu.pl
INTRODUCTION
Tool condition monitoring is a research area dealing with evaluating and assessing how long different machine elements can be used before they wear out. Algorithms can incorporate different input signals, and various methods can be used for data collection depending on tool-specific properties. Many of the algorithms used for tool condition monitoring consider drill condition monitoring specifically. For the general evaluation of a drill state, it is very important for the manufacturer to determine the exact moment when a drill is starting to dull out, resulting in a product that is not satisfactory. If this point in time is not detected quickly, then such poor quality elements will generate losses for the company. While the drill state can be evaluated manually, it is a tiresome and time-consuming process, hence the need for automatization.
Existing solutions for drill condition monitoring vary greatly in data collection and preprocessing methodologies, and they usually incorporate dedicated sensors and devices. Signals that are measured can be related to feed force, noise and vibrations, cutting torque, acoustic emission, or other parameters (Kurek et al. 2016). The main disadvantage of this approach is that a complicated and quite expensive arrangement of diverse sensors needs to be installed to register the relevant signals. Furthermore, many preprocessing stages are required to obtain actual, usable data. During preprocessing, it needs to be ensured that chosen sensors are appropriate for the current environment; in addition, registered signals need to be checked to ensure that they are the correct signals for generation of usable diagnostic features. Finally, the best features need to be selected and used for building the final classification model. Error at any of these stages can result in a classifier that is not usable. Many different features are generated on the basis of the registered signals. However, the challenge is that the classification accuracy stays below the 90% threshold, while the entire process is lengthy and difficult to implement in an actual work environment (Kuo and Cohen 1999; Panda et al. 2006; Jemielniak et al. 2012). The initial setup is quite expensive and does not guarantee that satisfactory results will be achieved in a reasonable time (or at all if, for example, the wrong sensors are chosen), and often not compensating in any way for the amount of time spent during those initial stages.
In the presented approach, the main focus was on accelerating and simplifying this process in order to make it more applicable for the furnishing companies, without adding high initial costs. Three classes were defined for drill condition monitoring: green, red, and yellow. The green class described tools that were in good condition and could be used further, the red class denoted tools in poor condition, that should be immediately replaced, and the yellow class was for tools that were suspected of being too worn out for further use and, which required manual evaluation by a human expert. From the manufacturer’s perspective, mistakes between green and red classes are far more undesirable than the others because they result in the highest possibility of financial loss; this distinction is more important than overall accuracy. The second important element concerns time required to prepare usable solution, which should be minimized. It is especially undesirable if a long time is required to set up and calibrate used sensors, since such setups cannot always remain unfolded in the actual work environment. The usability of a solution is further diminished if both a long time for initial setup is needed and lengthy computations are done before first results can be obtained (i.e. a prolonged training process is needed).
The current solution is based on some existing ones (Bengio 2009; Deng and Yu 2014; Schmidhuber 2015) as well as on the authors’ previous research in this area (Kurek et al. 2017a,b, 2019a,b). The first major improvement was removing any specialized equipment for initial data collection. A camera, which was used to take pictures of drilled holes, was the only external element used for data collection. This approach simplified the entire solution and allowed much easier assembly in the working environment. Furthermore, this does not require a large financial investment from the furniture company. During initial approaches, different methods were tested, in which the algorithms were based on convolutional neural networks (CNN); this is a good solution that does not require specialized diagnostic features (Goodfellow et al. 2016). Limited training data was used, based on Kurek et al. (2017a), and an artificially extended initial set of samples, based on Kurek et al. (2017b) was used while recognizing two classes. The extended approach combined data augmentation methodology with transfer learning (Kurek et al. 2019a), achieving higher accuracy than the results obtained by using complicated setups with diverse sensors (over 93%). In the final solution (Kurek et al. 2019b), a classifier ensemble was prepared using different pre-trained networks (Krizhevsky et al. 2012; Russakovsky et al. 2015; AlexNet Model online); this time over 95% accuracy was achieved.
While the above approaches showed high accuracy, two problems were encountered. Firstly, the methods did not take into consideration the additional manufacturer requirement regarding green-red misclassification rate, which should be minimized (high overall accuracy rate is inadequate if the number of mistakes between those two classes is significant). Secondly, although the training process was simplified and faster than solutions based on more complicated sensor setups, it still was not fast enough to meet manufacturer requirements. Therefore, in the current approach the main focus was placed on improving these aspects, while maintaining high overall accuracy. To achieve this, image color distribution was evaluated, with the assumption that samples classified as red will have a much higher number of pixels having values between black and white (in an image represented in gray-scale) than those not classified as red. In order to select the best possible setup, recursive feature elimination and cross-validated selection was used, along with applied Bayesian optimization of hyperparameters.
The rest of this work is organized as follows: The next section contains a description of the overall data preparation process and analysis (in regards to additional manufacturer requirements), as well as the prepared algorithm (which is capable of precise and efficient classification). The section after this then describes the results obtained during performed experiments. Conclusions are presented in the final section.
EXPERIMENTAL
Materials and methods
Holes were drilled using a standard CNC (Busellato Jet 100, Thiene, Italy) vertical machining center. The drillings were made using tungsten carbide through-hole drill bits Faba WP-01 (Faba SA, Baboszewo, Poland). An example drill (diameter of chosen tool equaled 12 mm) is presented in Fig. 1. Drills used in presented experiments can be used on glued wood, chipboard, and derivatives. A total of five drills were used to obtain input images, while holes drilled by each of them were separately stored in a time series fashion (representing the exact order in which the successive drillings were made).
Fig. 1. General view of drill bit used in experiments (FABA WP – 01)
The tool was periodically monitored using a standard workshop microscope (TM – 505; Mitutoyo, Kawasaki, Japan) for manual evaluation of the drill state. Wear of the outer corner was observed and defined as W (mm) (Jegorowa et al. 2019, 2020). This was determined separately for each of the blades, and the arithmetic mean of the wear of the cutting edges was calculated. Three classes for drill wear were selected based on obtained values. The upper limit of the “green” class (0.2) was defined by the manufacturer of the tools used in current experiment. Remaining classes: “yellow” (0.2-0.35) and “red” (above 0.35) were arbitrarily determined based on expert observation of the machining quality to distinguish the “worn” state of the tool from the one that is “completely worn out”. Those classes were also used for drill wear definition in the current, automated approach. Figure 2 shows the method used for manual evaluation of tool state.
Fig. 2. Method used during manual tool state evaluation
During the drilling process the spindle speed was set at 4500 rpm, while the feed speed equaled 1,35 m/min. The above parameters were selected according to the recommendations of the tool manufacturer. The drillings were made using three-layer melamine faced chipboard (Kronopol U 511 SM; Swiss Krono Sp. z o. o., Żary, Poland), an example of which is presented in Fig. 3. The test piece size equaled 2500 x 300 x 18 mm. The processed material is characterized by a different density resulting from a multilayer structure. The density profiles of the melamine faced chipboard (Fig. 4) were measured on a GreCon DAX 5000 device (Fagus-GreCon Greten GmbH & Co. KG, Alfeld/Hannover, Germany) (Sala et al. 2020). Laminate thickness was measured using an Olympus BX40 microscope (Olympus Corporation, Shinjuku, Tokyo, Japan) and equaled 0.092 ± 0.004 mm.
Previous work incorporated images of drilled holes in different solutions (Kurek et al. 2016, 2017a,b, 2019a,b) using various approaches to classification with convolutional neural networks (both training them from scratch as well as using transfer learning methodologies). Data samples with images of drilled holes were collected in cooperation with the Institute of Wood Sciences and Furniture at the Warsaw University of Life Sciences.
After the drilling process was finished, the obtained test piece was cut into smaller samples and photographed using a Nikon D810 (Nikon Corporation, Shinagawa, Tokyo, Japan) single-lens reflex digital camera with 35.9 x 24.0 mm CMOS image sensor. The images were stored in png format. Since slices of laminated chipboard contained sets of holes, individual samples still needed to be extracted. An additional method was used to prepare images for the current algorithm, while avoiding their manual preparation. In this case, image processing procedures were adapted for the task of extracting consecutive holes from the original picture and storing them in separate images.