FlankPix: An Image Segmentation Algorithm for Flank Wear Analysis in Monel K500 Turning

1.1 Background

Many experts have written volumes and articles on the subject of tool wear. On the other hand, flank wear has received a relatively minor amount of attention, even though both online and offline methods for monitoring the state of machining tools have been offered. In their study, Danesh and Khalili [1] used a non-decimated wavelet transform and statistical data on surface flaws to look at how tool wear compared to the surface texture of the turned piece. As a result, they could detect how tool wear affected the texture of the workpiece's surface. Researchers [2, 3] used edge recognition and morphological component analysis to find blunted edges in industrial environments. Researchers [4, 5] used tungsten carbide inserts to measure the depth of the crater wear during quasi-orthogonal cutting tests on AISI 1045 steel. They used a synthetic neural network to help them accomplish this, which allowed them to succeed. Matlab served as the primary modelling tool in developing an image-processing method by Xiong and his colleagues to assess the degree of tool wear. The image capture system comprises high-frequency linear fluorescent light, a high-resolution CCD camera, and a data-gathering module [6, 7].

1.2 Literature review

Researchers [8-10] developed an automated tool wear monitoring system using the active contour approach and neural networks to analyse tool flank wear. Urbicain and Trejo [11] created a system to measure milling and insert cutting-edge flaws in real-time without interfering with production. An edge-preserving smoothing filter, picture gradient computation, and a cutting-edge damage assessment based on geometrical features make up a three-stage technique. The ANOVA (Analysis of Variance) and fuzzy rule methods [11, 12] were used to determine how much the AISI 5140 steel alloy wore down and how much noise it made. The L27 orthogonal array method, the MSER algorithm, and the deep pattern network [13] were used to measure the alloy's surface roughness and flank wear when dry. To determine how long a tool made from AISI 4340 steel will last, the study [14] ran tests on the material with an L9 orthogonal array, analysed the results with an ANOVA, and relied on Gilbert's method. For Gamma-Prime-reinforced alloys, the study [15] employed Kalman filtering to estimate tool wear and reported an 18% filter rate of error. Of these papers, the study [16] MSER Algorithm and DPN, published under the name FlankNet, are the most promising candidates for the method used to assess tool wear, in particular flank wear.

Ant Colony Optimization (ACO) and Canny's Edge Detection (Canny) have been used together to find skin lesions much more effectively in a study of optimisation and edge detection methods [17]. As a result, Fuzzy Logic's performance ranges from roughly 87% to 91%, with an average inaccuracy of 87%. The cutting forces and temperature acting cumulatively on the cutting tool throughout the machining operation create normal tool wear. Tool wear would have a significant impact on the surface quality of the machining workpiece. Improving workpiece surface quality is required to increase economic efficiency. A precise tool wear value, which may be utilised to schedule cutting tool replacement properly, can assist in providing optimal machining settings. The ideal machining parameter values can be found. As a result, much emphasis has been placed on developing the tool wear monitoring system [18-21]. Using the FlankPix algorithm, enhancing the accuracy of the Average Error Prediction of Flank Wear in Monel K500 to achieve the target sum (in percentage terms) is essential. If the proposed method succeeds, the typical error rate should decrease by over 90%.

A comprehensive review of existing literature reveals the importance of monitoring cutting tool technology, specifically in assessing the machinability of difficult-to-machine materials. The significance of flank wear in mining and various industries is underscored. Different ways of measuring tool wear have been looked into in the past, including non-decimated wavelet transform, edge recognition, morphological component analysis, neural networks, image processing tools like Matlab, and automated monitoring systems that use neural networks, edge-preserving smoothing filters, and real-time analysis methods.

1.3 Research gap

The literature survey revealed a crucial research gap concerning precise and reliable tool wear measurement techniques, particularly in high-temperature alloy applications like Monel K500. Existing methods, while informative, displayed limitations in accuracy and suitability for such specialized materials. There was a clear need for an advanced tool wear assessment approach that could effectively address the challenges posed by background noise, provide higher precision in edge detection, and offer more nuanced measurements of tool wear extent. The research gap made it clear that we need a new and improved way to measure tool wear that can provide higher accuracy, lower sensitivity to image noise, and work with the unique properties of high-temperature alloys like Monel K500. This gap laid the foundation for the development of the FlankPix algorithm as a solution to address these shortcomings and provide an advanced, technically enhanced tool wear measurement method for precise evaluation in challenging industrial settings.

1.4 Research objectives

The main research objective is determining the average flank wear error percentage when turning the Monel K500. According to the study [22], steel alloys have several exceptional properties, including excellent corrosion resistance in chloride and Sulphur dioxide environments. In recent years, monitoring cutting tool technology has been applied to evaluate the attributes of cutting tool materials to identify the machinability of cutting tool materials that are notoriously difficult to machine [23-25]. In addition, stress and cracking are extremely unlikely in these environments. According to the study [26], monitoring tool wear reduces workpiece errors by preventing unexpected breakage or overuse. If wear is not considered, the tool wear may cause damage to the workpiece or the machine itself. The importance of flank wear in the mining sector must be emphasized. According to the study [27], alloys are utilized in various global industries, including manufacturing, aerospace, the military, and medicine. According to the study [28], crater and flank wear are two additional types of tool wear, with the latter being crucial for machine stability, dependability, and diametric accuracy.

1.5 Significance of the study

Improving tool wear assessment accuracy is crucial for optimizing machining parameters and workpiece surface quality, enhancing economic efficiency. The development of the FlankPix algorithm is anticipated to fill the existing research gap and provide a technically advanced tool wear measurement method for high-temperature alloy applications.

The core steps of the FlankPix algorithm rely on the Canny edge detection technique, a multi-stage process that identifies abrupt changes in intensity within grayscale images. The steps in this method are

•Noise reduction using Gaussian blur;

•Gradient calculation to find out the strengths and directions of edges;

•Non-maximum suppression keeps track of the ridges representing edges, and hysteresis thresholding is used to create edge maps based on high and low thresholds.

FlankPix stands out for its ability to precisely identify tool edges while disregarding background noise, resulting in accurate measurements of tool wear. Its effectiveness lies in the combination of pixel distance measurement and edge detection techniques, which contribute to maintaining a smooth and precise tool contour, which is critical for evaluating the tool's condition. This method demonstrates an average prediction error of just 1.29%, showcasing its high accuracy in assessing tool geometry changes and enhancing product quality in manufacturing processes.

The images of the tool wear were first processed using the Innovative FlankPix algorithm, then submitted to the edge detection algorithm and pixel measurement to estimate the amount of flank wear that each of the respective images has. The % error is the ratio of the difference between the measured value and predicted value to the measured value, as in Eq. (1).

$\%$ Error $=\frac{M-P}{M}$   (1)

where, M is experimental value of Flank wear, P is predicted value using FlankPix method.

3.1 Canny edge detection

A canny edge detector is a multi-stage operator for edge detection that can identify various image edges. Moreover, Canny provided a computational edge detection theory to clarify the efficacy of his approach. It was intended to create a superior edge detector that the Canny operator developed. It takes in a grayscale image and spits out a second image labelled with the coordinates of any abrupt changes in intensity that it detects. There are five stages to the Canny Edge detection method. Methods of reducing noise, calculating gradients, employing non-maximal suppression, utilizing a double threshold, and using hysteresis to track edges are discussed.

Several steps are required for the Canny operator to complete his work. Gaussian convolution is first used to smooth the image. A straightforward 2-dimensional first derivative operator was applied to the smoothed image to draw attention to areas with high first spatial derivatives. With the magnitude of the gradient, edges create ridges. The algorithm then follows the crests of these ridges, performing non-maximal suppression by setting all pixels that are not actually on the ridge top to zero. This results in a narrow line in the final output. Two thresholds, T1 and T2, govern the hysteresis shown in the tracking process, with T1 being higher than T2. Assuming T1 is the starting location, tracking can only start at a higher ridge. From there, they tracked in both directions until the ridge's elevation was lower than T2. Using hysteresis like this helps to prevent the splintering of loud edges into smaller pieces.

Gaussian Blur Kernel Size (σ): The size of the Gaussian blur kernel influences the amount of smoothing applied to the image. A larger kernel (higher σ) produces more blur, reducing noise but potentially blurring edges. Common σ values range from 1 to 3.

High and Low Thresholds: The Canny edge detector uses two thresholds to determine which edge pixels to consider as strong, weak, or non-relevant:

High Threshold (T₁): Determines the minimum gradient value required for a pixel to be considered a strong edge pixel.

Low Threshold (T₂): Sets the minimum gradient for a pixel to be considered a weak edge pixel. Any pixel with a gradient value between T₁ and T₂ is marked as a weak edge pixel.

Hysteresis Thresholding: This step helps in connecting edge pixels into continuous edges. It involves:

Strong Edge Pixel: Pixels above the high threshold (T₁).

Weak Edge Pixel: Pixels between the high (T₁) and low (T₂) thresholds.

Non-relevant Edge Pixel: Pixels below the low threshold (T₂) are considered non-relevant.

Edge Connectivity: Weak edge pixels are considered part of the edge if they are connected to strong edge pixels. This connectivity step forms continuous edges.

3.1.1 Noise reduction

One way to remove noise from an image is to apply a Gaussian blur. It is achieved using the Gaussian Kernel (3×3, 5×5, 7×7, etc.) image convolution technique. The degree of blur is the most critical factor when choosing a blur kernel. The blur becomes less apparent as the size of the kernel decreases. This research uses a 5×5 Gaussian kernel.

$\begin{gathered}\text { lessto }= \\ \frac{1}{2 \pi \sigma^2} \exp \left(-\frac{(e-(k+1))^2+(f-(k+1))^2}{2 \sigma^2}\right) \\ 1 \leq e, f \leq(2 k+1)\end{gathered}$ (2)

$\mathrm{S}=\mathrm{I} * \mathrm{~g}(\mathrm{x}, \mathrm{y})$ (3)

where,

$\begin{gathered}g(x, y)=\frac{1}{\sqrt{2 \pi \sigma^1}} e^{\frac{x^2+y^2}{2 \sigma^2}} \\ \mathrm{H}_{\text {ef }}=\text { Filter function }\end{gathered}$（4）

Eq. (2) shows the Gaussian Kernel filter equation.

3.1.2 Gradient calculation

The gradient size throughout the dimensions can be determined by computing the filter's derivative for the X and Y dimensions and then converging that result with I. It is also possible to determine the image's orientation by calculating the tangent of the angle between the two measurements. A gradient vector, complete with magnitude and direction, results from the convolution. In the end, the Gaussian derivatives are responsible for the staleness of the final edges.

3.1.3 Non maximum suppression

Consider the three edge points on edge shown in Figure 1. Consider the coordinates (x,y) as an example, assuming the edge gradient is at its highest level. It must be proven that the gradient perpendicular to the edge is less than (x,y). If the values are less than the (x,y) gradient, the non-maximal locations along the curve can be skipped.

1.png

Figure 1. Non – maximum suppression

3.1.4 Hysteresis thresholding

When creating an edge map, hysteresis thresholding considers both high and low thresholds before settling on a solution (see Figure 2). A significant improvement over single thresholding can be achieved by appropriately computing both the low and high points and then processing the resulting thresholder images. In contrast to single thresholding, hysteresis thresholding uses low and high thresholds to find edges via a linking procedure.

2.png

Figure 2. Hysteresis thresholding

3.2 Flank wear measurement using distancing method

The FlankPix algorithm incorporates distance measurement through the utilization of Canny edge detection. After the Canny edge detection technique identifies edges within grayscale images, the algorithm calculates the distance or extent of flank wear based on these detected edges. Canny edge detection, within the context of FlankPix, helps pinpoint the abrupt changes in intensity that signify the edges of the worn tool. The outcome of the canny edge detection can be seen in Figure 3.

3.png

Figure 3. (a) Tool wear SEM image, (b) Edge detected image of original SEM image

Once the edges are identified, FlankPix utilizes distance measurement techniques to determine the amount of wear along these edges precisely. It involves calculating the distance between specific points on the edges, allowing for an accurate estimation of the flank wear's extent. By integrating distance measurement with the information obtained through Canny Edge detection, FlankPix ensures a detailed assessment of the tool's wear, improving accuracy in evaluating tool conditions and ultimately enhancing manufacturing quality.

4.png

Figure 4. Proposed methodology

The process followed in tool wear detection from SEM images is illustrated in Figure 4, as given below. The input test image is pre-processed by converting it to grayscale binary. The bwconncomp command calculates the binary cumulative area. Moreover, the orientation, minor axis length, and significant axis length at their maximums can be obtained. The pixels of the captured image are determined, as illustrated in Figure 5. After examining the collected image of the interconnected construction blocks, we calculated the size of each pixel in the x and y directions using the number of pixels placed in the x and y directions. The capture event command reads the coordinates of the point in the X and Y directions.

5.png

Figure 5. FlankPix analysis

The measured parameters are as follows:

Major Axis Length 583.54 mm

Minor Axis Length 200.36 mm

3.2.1 Steps in FlankPix algorithm

1. Open CV

2. Create a new file

3. Launch a Python program

4. Import the files that contain the headers

5. Download and install the necessary Python interpreters for the algorithms, such as OpenCV, matplotlib, and NumPy

6. Change the color mode of the image to grayscale

7. To detect edges in the photos, run Canny's Edge Detection Algorithm

8. Fix the threshold value (Edge detection)

9. Accumulate Pixel coordinates in Px and Py

10. Calculate the distance from the pixel coordinates

12. Convert the value from Pixels to mm

13. Compare the predicted value with the respective experimental data

15. Calculate the % Error and average error percentage

3.2.2 Sensitivity analysis of the vision system

The precision of the suggested vision system depicted in Figure 6 was determined by comparing the measured values for flank wear.

Error Sensitivity = [ Experimental Value]-[Predicted Value]/

[Experimental value]* 100

For the given input image,

Experimental measured value = 0.43 mm

FlankPix predicted value = 0.4296 mm

Error Sensitivity = (0.43-0.4296)/0.43 = 0.00093×100 = 0.093%

Average Prediction error = 1.29%

6.png

Figure 6. Flank wear measurement using FlankPix

The industrial sector will find this approach of measuring tools wear much more helpful than using a single metric. The Edge Detection Method was developed to measure the amount of wear on the Cubic Boron Nitride insert used to turn Monel K500. It is essential to preserve the integrity of the machined surface and keep tool wear to a minimum.

The subsequent findings are given below:

• The FlankPix method may be utilized to measure flank wear.
• It is a promising and practicable method for measuring flank wear in-process for industrial applications.
• Minimal flank wear was attained at medium and moderate cutting speeds with a low feed rate.

Compared to the standard way of evaluating flank wear using a toolmaker's microscope, the flank wear measured with FlankPix yielded greater accuracy, with an average percentage of 1.29 percent. FlankPix algorithm's applications in precision tool wear measurement across aerospace, manufacturing, defense, and medical sectors.

Some potential limitations include susceptibility to image quality variations, sensitivity to lighting conditions, challenges with complex tool geometries, and the need for further validation in diverse industrial settings to confirm its generalizability and performance under varying conditions. Improvements for FlankPix could involve refining the algorithm, integrating advanced technologies like AI, ensuring robustness across varied conditions, validating in different industries, and developing methodologies for quantifying tool life, enhancing its reliability and wider applicability.