Keywords

1 Introduction

The metal forming business is facing challenges worldwide due to new materials and processing processes. Because the process necessitates the employment of several resources, a computerized method for assessing the blanking procedure is required. Improved approaches for studying the behavior of the sheet metal forming sector are in high demand.

Automation is required to increase production. Given the automation level, deploying software applications in the sheet metal processing industry is advantageous. Different sensing techniques for fault diagnosis and the re-use of damaged component replace systems rapidly appearing as automation develops. However, a suitable process model is required to manage the complexity of sheet metal.

The next industrial revolution is being ushered in by advancements in automation. There has been growing research in sensing systems for defect detection and identifying significant parameters. The intricacy of the sheet metal working method, on the other hand, makes building a self-learning model challenging. With advent of mechanization, the industrial positioning in assessing defect is rapidly moving toward developing self-sustained systems. However, a suitable process model is required to manage the complexity of sheet metal processing.

While studying the process parameters in machining, research is conducted utilizing ANN modeling to increase the precision of intellectual structures [14, 21, 23] and as well as sheet metal blanking [11, 16, 17, 19]. The approach has been influenced by sheet thickness and tool wear [6, 7].

Classification model is proposed using decision tree modeling in current work, for IS277GI material for predicting fracture surface quality.

2 Literature Review

2.1 Blanking Process Setup

The process model of blanking consists of blanking die, sheet metal, blank holder, and punch as shown in Fig. 1. Suitable clearance is selected between blank and punch for obtaining smooth fracture surface. A punch with velocity shears the work-piece placed between blank holder and die. In this way, a slug, called blank, separates from the work-piece.

Fig. 1
A schematic diagram depicts the procedure system of blanking consists of a blanking die, sheet metal, blank holder, and punch.

Blanking procedure system [18]

Full size image

Figure 2 depicts blank part after shearing and the clean-cut surface. Strain at rupture is the most crucial element in determining when a fracture will begin and the propagation circumstances. From the inside die corner to the punch corner, a fracture line develops as a result of shearing. As per literature, the clearance is defined in terms of percentage of sheet thickness. Figure 3 shows mechanism of metal fracture at punch and die. Smooth sheared surface is obtained when the crack path joins the fracture line.

Fig. 2
An illustration of the workpiece depicts the burr, fracture, clean-cut surface, and roll-over, which are indicated by the arrows.

Slug work-piece [18]

Full size image
Fig. 3
A diagram represents the mechanism of metal fracture at the punch and die, and a sheet metal connects the punch and die with the diagonal distances P 1, P 2, and P 3, respectively. The thickness of the sheet is t, and the horizontal distance between punch and pie is c.

Diagonal fracture line diagram [16]

Full size image

The previous studies on clearance identification for sheet metal employed FEM [9, 11] and FEM simulation findings were consistent with experimental investigations. They concluded that surface roughness improves when the cracks beginning at the punch and die coincides [9]. However, punch velocity and heat production during processing have a significant role on metal behavior [16]. Investigations were done into how the punch geometry affected the features of the cut surface and the cutting forces [21].

$$ \begin{array}{*{20}c} {c = 100\frac{{D_{{\text{m}}} - D_{{\text{p}}} }}{2t}\left( \% \right) } \\ \end{array} $$
(1)

where Dm—die diameter, Dp—punch diameter, t—sheet thickness.

ANN is powerful methodology for studying the behavior of the turning process [12, 23], as well as the bending of sheet metal [3, 23]. The simulated system uses an algorithm that replaces the conventional judgment system. The NN is chosen to construct suitable model for predicting optimal punch-die clearance. A backpropagation neural network was used for the prediction of optimal clearance [16,17,18]. The data on experimental fracture angles was utilized to train the algorithm, and model is developed for a given data. While assessing the defects in sheet metal forming, [5] uses a CART, MLP, SVM, RF techniques for predicting the coil back, and utmost thinning result achieved the accurateness varying from 87.39 to 94.98%.

2.2 Decision Tree Classifier (DTC)

The decision tree (DT) algorithm is a supervised machine learning technique to solve classification and regression problems. This approach aims to develop a model that predicts the value of a targeted variable; for that, the decision tree solves the issue using the tree representation, where the leaf node belongs to a class label, and characteristics are expressed on the inner node of the tree. A classification strategy continually separates data using decision rules. Data is classified at each node for optimization of decision-making for information gain:

$$ {\text{Gain}} = \mathop \sum \limits_{j = 1}^{m} \frac{{n_{j} }}{{N_{j} }}H\left( {D_{j} } \right) $$

where N represents the total number of data points for node j, n represents the number of data points for node j of the expected class, D represents node values, and H represents impurities [5]. The formula for impurity Gini Index is as follows:

$$ I_{{\text{G}}} = 1 - \mathop \sum \limits_{i = 1}^{C} p_{i}^{2} $$

where pi share of sample for node.

$$ I_{{\text{H}}} = - \mathop \sum \limits_{i = 1}^{C} p_{i} \log_{2} \left( {p_{i} } \right) $$

where pi—the proportion node fitting to a class [2]. If all node samples are of the same kind, then entropy is zero. This step is followed until the same label remains in a sample from each terminating node. A stop condition can also be established to avoid over fitting.

3 Methodology

For investigating the quality of surface, decision tree classifier is employed. The independent factors dependent, and categorical variable is chosen. Figure 4 depicts a correlation analysis using Python code to investigate the association between the independent variables through heat map generation.

Fig. 4
A 6 by 6 square matrix table represents a heat map of the correlation using Python code. The rows and columns have the surqual index, punch pen, fracture angle, shear angle, bur height, and sure rough. A variance scale from negative 0.4 to positive 0.8.

Analysis of correlation

Full size image

The control factors are derived though literature and total of 140 data points (training and testing) and 42 (validation) data points. This study develops a DTC model by splitting data: 70% for training and 30% for testing. An estimator with a variable maximum depth is used to train the resulting model, while parameter adjustment is applied. At the depth of 5, we obtained highest accuracy for accuracy measurements in terms of training and testing, recall, precision, f1-score, and confusion matrix. At each node, we employed Gini and Entropy as impurity indices, with a maximum depth of 5 and minimum sample leaf 5. The research flowchart is shown in Fig. 5.

Fig. 5
A flowchart of research starts from the identification of process parameters, then to experimental data collection, tunning of D T C model estimators, output, favorable results, and next experimental validation followed by result consistence and finally ends in results acceptance.

Research chart

Full size image

During training phase, progressive categorization of samples and visualization of the decision tree is developed using Python code for the detailed functioning of the Gini and progressive calculation Entropy indexes.

Figure 6 shows gradual classification of samples based on Gini index, and Fig. 7 shows iterative steps till samples are classified based on entropy.

Fig. 6
A flowchart of the gradual classification of samples based on the Gini index starts from X subscript 2 less than or equals 0.049, Gini equals 0.595, samples equals 98, value equals 15, 31, 52, and this is classified into two segments, such as true and false. The steps are repeated until samples are classified.

Functioning of Gini index visualization

Full size image
Fig. 7
A flowchart of the iterative steps till samples are classified based on entropy, starts from X subscript 2 less than or equal to 0.096, Gini equal to 1.425, samples equals 98, value equals 15, 31, 52, and this is classified into two segments such as true and false.

Functioning of Entropy index visualization

Full size image

The outcome of classification is approved if the validation (experimental) provides consistent results; the findings are directed to the DTC model training and testing for parametric optimization.

4 Discussion and Result

Experiments are carried out for blanking operation using power press with a punch (hallow circular). Table 1 displays the input parameter and properties of ‘IS277GI’ for the experiment. Engineering strain and real strain are calculated using material characteristics [5].

Table 1 ‘IS277GI’ mechanical properties
Full size table

We trained the DTC model using the Gini and Entropy index of impurity criterion, and both models obtained 93% accuracy. Figure 8 depicts a comparison of prediction performance with various classes.

Fig. 8
A bar graph with error lines depicts the comparison of prediction performance with various classes. The Y axis ranges from 0 to 1.5. The X-axis has precision, recall, and f 1 score support, and each has three bars. The indexes 1 and 3 have the highest values, around 1 in precious and 1 in recall.

Accuracy measures for different classes

Full size image

Confusion matrix is an another approach employed for measuring the accuracy of model as shown in Fig. 9, to validate the performance of DTC model. During the testing phase, class 3 has all of its sample points properly identified; however, classes have a deviance of sample points (one and two).

Fig. 9
A 3 by 3 square of the confusion matrix depicts the correlation between actuals and prediction, which ranges from 0 to 2. The diagonal is filled with 7, 5, and 27. The upper triangular matrix is filled with 1, 0, and 2, whereas the lower is filled with 0, 0, and 0.

Confusion matrix

Full size image

One more performance measure for the model is receiver operating characteristic (ROC), which presents graphically performance of classification. Two variations are plotted: the true positive rate (TPR) and the false positive rate (FPR), as shown in Fig. 10. The probability for the ROC curve of class 2 is 0.85, which is less than the probability for classes 1 and 3, which are 0.93 and 0.92, respectively.

Fig. 10
A graph represents the decision tree classifier of the true positive against false positive rate. 5 lines, 3 solid lines for the classes 1.0, 2.0, and 3.0, follow an increasing trend, and 2 dotted lines refer to the micro and macro average R O C curves follow an increasing trend. All the lines start at the origin and ends at the (1, 1).

ROC curve

Full size image

5 Conclusion

The machine learning approach (decision tree classifier method) is used to estimate quality of the fracture surface. The result shows that punch penetration has a direct impact on burr height creation and surface roughness. When employing DTC, we discovered that both the Gini as well as Entropy index impurities provide precise model correctness for the training, testing, and validation sets. The precision metric for class one is 1, whereas the recall measure for class two is 1. With k-fold cross-validation, additional sample points would imply greater model accuracy. Other classification models, such as support vector machines and random forest, can also be investigated using the significant data points.

Current study provides sheet metal production engineers with a decision-making solution for selecting the underlying factors for calculating precision for right fit requirement for processing IS277GI.