Optimization of the Plasma Arc Cutting Process Through Technological Forecasting

Despite the extensive application of plasma arc cutting (PAC) in heavy-duty manufacturing cycles, there are challenges and drawbacks that require careful consideration prior to its implementation. Traditional control methods falter due to the unique nature of the input parameters, necessitating advanced technological approaches for process enhancement.

PAC is renowned for its efficiency and versatility in cutting a wide range of materials within the industrial sector. A significant hurdle for PAC is the formation of dross [1]. The primary influential factors in the PAC process are identified as arc current, standoff distance, and cutting speed [2], with each factor carrying varying degrees of significance. A study on the heat-affected zone (HAZ) generated during PAC of stainless steel (Grade 304) revealed optimal cutting conditions utilizing response surface methodology (RSM): a voltage of 240 volts, a current of 65 amps, and a speed of 1.5 millimeters per second [3]. Employing these input parameters resulted in an output of a material removal rate (MRR) of 5.79 millimeters per second over a period of 246 seconds, achieving a surface roughness of 65 micrometers [4]. Analysis of the cut surface's microstructure uncovered the presence of a recast layer, microstriations, and dross accumulation [5]. Experimental results indicate that the proposed control method surpasses traditional PI control in enhancing accuracy, finish, ripple reduction, and overall quality of the workpiece, alongside other comprehensive metrics. Additionally, the power supply for PAC demonstrates exceptional control performance, attributed to its fuzzy-neural network framework [6].

The primary objective of this study is to pinpoint the variables that significantly influence the PAC process when applied to stainless steel [7]. The investigation proceeds through the evaluation and application of six distinctive output characteristics. The adaptive neurofuzzy inference system (ANFIS) [8], [9], [10] is an extraordinarily effective tool for identifying and categorizing critical input parameters essential to the progression of the PAC process. This system is characterized by robustness, resilience, immunity to interference, fault tolerance, conformity with ANFIS standards, and proficiency in normalization and generalization. Consequently, variations in the characteristics are unlikely to significantly impact the performance of the PAC method as refined by ANFIS [11], [12], [13].

To conduct the cutting tests, stainless steel AISI 304 specimens measuring 250×70×50mm were prepared. The input parameters include standoff distance, cutting current, and cutting speed. The outputs, such as material removal rate (MRR), chamfer, surface roughness, dross, kerf width, and heat-affected zone (HAZ), are measured. The MRR is efficiently determined using electronic weighing equipment. Surface roughness is assessed with a Talysurf instrument, while the remaining outputs are measured using an automated Vernier caliper. Figure 1 displays the experimental setup of the plasma arc cutting system. Table 1 presents sample experimental data related to the input and output variables [14], [15], [16].

Figure 2 depicts the five distinct layers of the ANFIS network. Within the ANFIS framework, the fuzzy inference system functions analogously to a central nervous system. The first layer receives inputs and employs membership functions to convert them into fuzzy values. For this study, the bell-shaped membership function was selected due to its superior efficacy in handling nonlinear data regression.

The membership functions, characterized by a bell-shape, are defined in the following manner:

where, $\left\{a_i, b_i, c_i\right\}$ denote the parameters set, while $x$ is the input.

The calculation of rule firing strength involves multiplying the fuzzy signals generated by the first layer within the second layer. The third layer, known as the rule layer, is responsible for normalizing the signals emanating from the second layer. The fourth layer undertakes the task of rule inference and transforms all signals into crisp values. The final layer produces an impeccable output through the aggregation of all incoming signals.

By employing the ANFIS approach, the most precise predictors for each type of defect were identified. The process of selecting and preprocessing input parameters to remove extraneous information is critical. The MATLAB command below is used to split the dataset into a training set (odd-indexed samples) and a checking set (even-indexed samples):

>>[data] = plasma;

>> trn_data = data(1:2:end,:);

>> chk_data = data(2:2:end,:);

The function “exhsrch" determines the set of inputs that have the greatest impact on the six output variables through an exhaustive search of all available inputs. The first argument of the function specifies the number of input combinations to be evaluated during the selection process. “Exhsrch" first constructs an ANFIS model for all possible combinations. The model is then trained for a single epoch, after which the outcomes are evaluated. The command line used to identify the one and two most important factors in predicting outputs is as follows:

>> exhsrch(1,trn_data,chk_data);

>> exhsrch(2,trn_data,chk_data);

Table 2 displays the correlations between the effects of single and dual input combinations on the Material Removal Rate (MRR). These effects are quantified using the Root Mean Square Error (RMSE) for both training and error checking. The standoff distance has the most significant impact on the MRR and is associated with the smallest training error. Essentially, even minor adjustments to the standoff distance can lead to considerable changes in the MRR.

When combining the cutting current and standoff distance, the training error decreases further, which indicates a greater influence on the MRR. Conversely, the most significant variation in MRR is expected to occur with simultaneous changes to the cutting current and standoff distance. Figure 3 illustrates the ANFIS decision surface generated by using the two optimal inputs.

Table 3 shows the relationships between the effects of single and combined input pairs on surface roughness. Among the input parameters, the cutting current exerts the greatest influence on surface roughness, as indicated by its status as the most accurate during training. In other words, even minimal adjustments to the cutting current can lead to significant variations in surface roughness.

The pairing of cutting current and standoff distance significantly affects surface roughness by reducing the training error. Specifically, alterations in surface roughness are most pronounced when both the cutting current and standoff distance are varied together. Figure 4 presents the ANFIS decision surface generated by employing the two most effective inputs.

Table 4 illustrates the relationships between the effects of single and paired input combinations on chamfer. The cutting current is the input parameter that most significantly influences the chamfer, as indicated by the least training error. Essentially, even small fluctuations in the cutting current can lead to considerable variations in the chamfer.

Combining the cutting current with the standoff distance minimizes the training error, which in turn exerts the most significant effect on the chamfer. More specifically, the chamfer is likely to show the greatest variation when changes in the cutting current and standoff distance occur simultaneously. Figure 5 shows the ANFIS decision surface that is derived from using the two optimal inputs.

Table 5 outlines the correlations between the effects of individual and combined input pairs on dross. The standoff gap, as an input feature, has the most profound impact on dross, associated with the least training error. Essentially, even minor adjustments to the standoff distance can lead to significant changes in dross.

When the cutting current and standoff distance are paired, there is a reduction in training error, which translates to a more substantial effect on dross. In practice, the most notable change in dross would occur when there are concurrent variations in both the cutting current and standoff gap. Figure 6 shows the ANFIS decision surface generated by employing the two optimal inputs.

Table 6 displays the relationships between the effects of single and combined input pairs on kerf width. The kerf width is most significantly influenced by the cutting current, which is the input parameter with the lowest training error. Essentially, even minor fluctuations in the cutting current can lead to considerable shifts in kerf width.

Table 7 shows the correlations between the effects of individual and combined input pairs on the Heat-Affected Zone (HAZ). The standoff distance is the input parameter that has the most substantial impact on the HAZ, as evidenced by the smallest training error. Essentially, even minor adjustments to the standoff distance can lead to significant changes in the HAZ.

The combination of cutting current and standoff distance reduces the training error, thereby having the most pronounced influence on kerf width. Practically speaking, the kerf width would experience the greatest variation when changes to the cutting current and standoff gap occur concurrently. Figure 7 illustrates the ANFIS decision surface created by utilizing the two optimal inputs.

The combination of cutting current and standoff gap minimizes the training error and, therefore, has the most profound influence on the HAZ. Put another way, the HAZ would undergo the greatest degree of variation when the cutting current and standoff distance are altered simultaneously. Figure 8 depicts the ANFIS decision surface generated using the two optimal inputs.

In the industrial sector, plasma arc cutting is utilized for particularly demanding tasks. Traditional control methods fall short due to the complex nature of the input parameters, rendering them unsuitable for advancing technology in this area. This experiment employed a data-driven optimization approach to enhance the plasma arc cutting process. The quality of the cut is evaluated by six output attributes. The input factors are classified as the standoff gap, cutting current, and cutting speed. The output variables include the Material Removal Rate (MRR), surface roughness, chamfer, dross, kerf width, and Heat-Affected Zone (HAZ).

The primary goal of this research was to identify the most significant characteristics that determine the most effective settings for plasma arc cutting. An adaptive neural fuzzy inference system (ANFIS) was employed to analyze the various input parameters of the plasma arc cutting process. According to ANFIS predictions, the variables of Material Removal Rate (MRR), chamfer, surface roughness, dross, kerf width, and Heat-Affected Zone (HAZ) are most significantly influenced by the combination of cutting current and standoff distance. In other words, the optimal configuration for predicting the dimensions of the Heat-Affected Zone, Material Removal Rate (MRR), surface roughness, chamfer, dross, and kerf width occurs when the cutting current and standoff gap are considered in tandem.