Intelligent Monitoring and Compensation between EDM and ECM

Abstract

Electric discharge machining (EDM) is a type of high-precision machining usually applied to hard-material machining for mold manufacturing and in the aerospace industry. Longer process times typically reduce facility efficiency. The use of electrochemistry machining (ECM) can overcome this challenge to efficiently machine large workpieces. Some industries have adopted and combined these two processes for Inconel 718 material machining. However, the use of coordinate-measuring machine times to determine the machining accuracy of these two processes is difficult. This study matched process features by analyzing the electric driving pulses of ECM and EDM. Fitting intelligent sensing signals that respond to dimensional measurements can be used to analyze electrical pulse signals. For analyzing a cross-process model using extracted key features of the process, our feedback-based system determines lower machining measurement errors and improves geometric size. Finally, the processing time of experiments can be reduced by 80%, and our proposed model has a prediction accuracy of approximately 0.01 ${\mathrm{mm}}^2$.

1. Introduction

The micromachining and manufacturing of large workpieces is challenging in industries that require high efficiency and precision, particularly the aerospace industry. Although demand for structural materials with high strength, hardness, thermal conductivity, and corrosion resistance has been growing in industries such as health care, aerospace, and the automotive industry, the application of special materials in traditional machining is difficult. Thus, traditional drilling, milling, and machining processes are gradually being replaced by electrical machining [1]. Nickel-based materials have excellent creep, corrosion, and thermal fatigue resistance at high temperatures, and so nickel-based alloys (e.g., Inconel 718) are widely used in the aerospace industry [2,3,4]. The development of new materials and processes in the aerospace industry still presents numerous challenges, making it difficult to improve efficiency; therefore, the development of this technology warrants investigation.

The manufacturing industry is rapidly changing. Following technological advancements in manufacturing, supply chain restructuring, and numerous specific production processes, production lines are gradually becoming more automated. When technical barriers to the research and development of new products are reduced and the product life cycle is shortened, faster and more flexible production methods are required to satisfy consumer demand. The objectives of addressing such problems typically involve (1) the minimization of trim loss, (2) avoidance of production overruns, and (3) avoidance of unnecessary slitter setups. This is particularly crucial in processes that have multiple applications [5]. The approach to implementing these changes is typically somewhat conservative.

Electrochemical machining (ECM) is a highly cost-effective, high-strength machining method. The processing of heat-resistant materials into complex shapes through ECM is easy. However, this process is influenced by material and workpiece structure because the hydrodynamic instability of the anode boundary layer affects surface roughness [6]. In 1928, Gusseff et al. proposed the ECM method, which entails converting the positive and negative ions of two conductive materials into electrons by using electrolyte solutions to achieve the mutual conversion of chemical and electrical energy between the electrolyte solutions [7]. ECM has been actively used in the aerospace industry since the 1950s, but the development of ECM has been slowed by the increasing popularity of electrical discharge machining (EDM). ECM has received renewed attention because ECM provides high hardness and composite material for the rapid machining of complex geometries [8,9]. Accuracy and surface roughness are technical difficulties of ECM that require further research.

In the process of ECM refinement, the boundary conditions during machining can be analyzed through simulation, and the optimal machining conditions can be determined from the input changes of key parameters and simulation processes [10,11,12]. The material removal rate (MRR) is calculated using a thermal model, and the temperature distribution in the affected processing zone is analyzed using the finite element method to estimate its MRR [13]. Four parameters such as feed rate, electrolyte, flow rate, voltage, MRR, roughness, and overcut can be analyzed [14,15]. The neural network model is used for MRR prediction. In our training set, the control variables of voltage, feed rate, electrolytic concentration, and conductivity had a mean percentage of error (MAPE) of 0.0067, indicating that the control variable factor of the training set was highly correlated with the self-variable factor [16,17]. Processing can be predicted and product precision can be maintained using protocols described in the literature.

EDM is a nontraditional mechanical process that uses a pulsed electrical discharge to generate thermoelectric energy between a workpiece and an electrode, thereby melting conductive materials. The four mainstream processing methods are ultrasonic-vibration EDM, dry EDM, powder-mixed dielectric EDM, and submergence EDM. These methods are used to improve processing performance, develop new material processing technologies, and achieve better quality [18,19,20]. Enhanced processes, materials, controls, and monitoring are also goals of such methodology.

In the EDM process, discharge energy influences surface roughness and topography. The influence of the powder particles and capacitors connected in parallel in EDM must be determined, particularly in the gap, on the thermal spread in the dielectric, and on the influenced zone. A larger workpiece surface exerts greater discharge energy [21,22]. Discharge energy and contact angle directly affect shape in the texturing process [23].

EDM’s long processing time and fast pulse render defect detection in manufacturing difficult. Sensors can be used to collect voltage and current for analysis, but a high sampling rate of 100 MHz presents a problem; thus, an effective waveform is used for extraction [24], feature calculation [25], and dimensionality reduction to reduce data volume. Neural networks can be used to classify and predict processes [26,27]. This is an effective method for a single process, but how the results are input in subsequent processes is crucial.

Skrabalak et al. [28] compared ECM, EDM, and electrochemical discharge machining (ECDM). They reported the advantages of applying composite ECM and EDM to machining diamond tools combined with the effect of ECM erosion auxiliary discharge. However, this is difficult to control [29]. In industry, the controllability of a single process must be maintained using a fixture to regulate the process [30,31,32].

How to complete the processing and match the signal characteristics through preprocessing and postprocessing requires investigation. The same method can analyze the pulsed electrical signal and the two process models through the extraction of key process features. This study effectively lowered machining measurement errors and maintained more accurate geometric feedback.

2. Materials and Methods

To match the composite electrical machining, the process illustrated in Figure 1 was established. The electrical signals of ECM and EDM are collected, and the electrical signals and dimensions are simulated through a machine learning model to synchronously match the two process errors.

• Current estimate

To estimate the amount of material removed to bridge the two processes, the value of the initial material (before machining) is defined as ${\mathrm{x}}_{\mathrm{i}}$, the ECM process removal amount is defined as ${\mathrm{A}}_{\mathrm{ECM}}$, the volume after ECM processing is defined as ${\mathrm{x}}_{\mathrm{c}}$, the EDM process removal amount is defined as ${\mathrm{A}}_{\mathrm{EDM}}$, and the volume after EDM processing is defined as ${\mathrm{x}}_{\mathrm{c}}$.

$${\mathrm{x}}_{\mathrm{i}}-{\mathrm{A}}_{\mathrm{ECM}}={\mathrm{x}}_{\mathrm{c}}$$

(1)

$${\mathrm{x}}_{\mathrm{c}}-{\mathrm{A}}_{\mathrm{EDM}}={\mathrm{x}}_{\mathrm{d}}$$

(2)

When designing an ECDM machine, Skrabalak [28] et al. defined the equation for estimating the current as Equations (3), (4) and (9), where ${\mathrm{I}}_{\mathrm{ECM}}$ is the current of the electrochemical process, ${\mathrm{I}}_{\mathrm{EDM}}$ is the current of the electrochemical process, $\mathrm{H}$ is the width of the electrode tool, $\mathrm{b}$ is the electrode tool length, ${\mathrm{v}}_{\mathrm{f}}$ is the spindle feed rate, ${\mathrm{A}}_{\mathrm{ECM}}$ is an ECM dissolving surface, ${\mathrm{k}}_{\mathrm{ECM}}$ is the factor for estimating the ECM dissolution rate, and ${\mathrm{k}}_{\mathrm{EDM}}$ is the spark factor of EDM.

$${\mathrm{I}}_{\mathrm{ECM}}={\mathrm{A}}_{\mathrm{ECM}}\frac{{\mathrm{v}}_{\mathrm{f}}}{{\mathrm{k}}_{\mathrm{ECM}}}$$

(3)

$${\mathrm{I}}_{\mathrm{EDM}}=(\mathrm{Hb}-{\mathrm{A}}_{\mathrm{ECM}})\frac{{\mathrm{v}}_{\mathrm{f}}}{{\mathrm{k}}_{\mathrm{EDM}}}$$

(4)

To determine the removal amount, Equations (3) and (4) are transformed, and the electrode is regarded as a constant value in the ideal condition as $(\mathrm{Hb}-{\mathrm{A}}_{\mathrm{ECM}})={\mathrm{A}}_{\mathrm{EDM}}$. Formulas (5) and (6) are as follows:

$${\mathrm{A}}_{\mathrm{ECM}}={\mathrm{I}}_{\mathrm{ECM}}\frac{{\mathrm{k}}_{\mathrm{ECM}}}{{\mathrm{v}}_{\mathrm{f}}}$$

(5)

$${\mathrm{A}}_{\mathrm{EDM}}={\mathrm{I}}_{\mathrm{EDM}}\frac{{\mathrm{k}}_{\mathrm{EDM}}}{{\mathrm{v}}_{\mathrm{f}}}$$

(6)

Equations (1) and (2) into Equations (5) and (6) to obtain (7) and (8). It is assumed that $\frac{{\mathrm{k}}_{\mathrm{EDM}}}{{\mathrm{v}}_{\mathrm{f}}}$ and $\frac{{\mathrm{k}}_{\mathrm{EDM}}}{{\mathrm{v}}_{\mathrm{f}}}$ are both in the optimal control state and can be regarded as fixed values so that the removal amount is proportional to the current.

$${\mathrm{x}}_{\mathrm{i}}-{\mathrm{I}}_{\mathrm{ECM}}\frac{{\mathrm{k}}_{\mathrm{EDM}}}{{\mathrm{v}}_{\mathrm{f}}}={\mathrm{x}}_{\mathrm{c}}$$

(7)

$${\mathrm{x}}_{\mathrm{c}}-{\mathrm{I}}_{\mathrm{EDM}}\frac{{\mathrm{k}}_{\mathrm{EDM}}}{{\mathrm{v}}_{\mathrm{f}}}={\mathrm{x}}_{\mathrm{d}}$$

(8)

The sum of ECM and EDM machining currents can be defined as the total current. $I_{total}$ can be regarded as the total removal.

$${\mathrm{I}}_{\mathrm{total}}={\mathrm{I}}_{\mathrm{ECM}}+{\mathrm{I}}_{\mathrm{EDM}}$$

(9)

Equations (7)–(9) describe the proportional relationship between the current and the removal amount: when ${\mathrm{I}}_{\mathrm{ECM}}$ increases, ${\mathrm{x}}_{\mathrm{c}}$ also increases. To ensure the accuracy of the final ${\mathrm{x}}_{\mathrm{d}}$, the current range of ${\mathrm{I}}_{\mathrm{EDM}}$ can be adjusted according to the change of ${\mathrm{x}}_{\mathrm{c}}$.

b.

Average ignition delay time (AIDT)

The ignition delay time is defined as the time difference between the finished time of charging voltage $t_{d,i}$ and the effective start time of discharge current $t_{e,i}$ at the ith spark. The AIDT is the average ignition delay times of all sparks, as presented in Equation (10).

$$\mathrm{AIDT}=\frac{1}{{\mathrm{N}}_{\mathrm{t}}}{\displaystyle \sum }_{\mathrm{i}=1}^{{\mathrm{N}}_{\mathrm{t}}}\left({\mathrm{t}}_{\mathrm{e},\mathrm{i}}-{\mathrm{t}}_{\mathrm{d},\mathrm{i}}\right)$$

(10)

c.

Average spark frequency (ASF)

A spark is defined as follows: when a tool (positive electrode) is charged during processing, the tool approaches a workpiece (negative electrode), and a current loop is formed between the tool and workpiece until complete discharge. Hence, ASF can be defined as the total number of sparks $N_t$ generated during time period Tt, as presented in Equation (11).

$$\mathrm{AS}F=\frac{{\mathrm{N}}_{\mathrm{t}}}{{\mathrm{T}}_{\mathrm{t}}}$$

(11)

d.

Heatmap

A heatmap is used to display the correlation coefficient matrix of variables. This can intuitively depict the difference between given values. Each variable was determined by assigning a score from −1 to 1.

e.

Neural network

A neural network model is a model based on a biological nervous system and is composed of multiple neurons; the neurons mimic the function of biological neurons. To evaluate the nonlinear relationship between input and output, neural networks with supervised learning artificial neural networks have been proven to be effective in numerous fields. Neural networks are composed of multiple nodes, where X = [${\mathrm{x}}_1$,…,${\mathrm{x}}_{\mathrm{n}}$] is the input vector, W = [${\mathrm{w}}_1$,…${\mathrm{w}}_{\mathrm{n}}$,] is the weight of each input, and b is the partial weight, which is a neural network algorithm modification input value. The activation function F equation is then inputted, and the result Y is outputted. The overall neural network algorithm is expressed by Equation (12).

$$\mathrm{y}=\mathrm{f}\left(\mathrm{WX}+\mathrm{b}\right)$$

(12)

f.

Linear regression

The use of linear regression is suitable because the outcomes are linear. Linear regression is suitable for determining the relationship between the response variable Y and the explanatory variable.

$${\mathrm{Y}}_{\mathrm{i} =}{\mathsf{\beta }}_0+{\mathsf{\beta }}_1{\mathrm{X}}_{\mathrm{i}1}+\dots +{\mathsf{\beta }}_{\mathrm{p}}{\mathrm{X}}_{\mathrm{ip}}+\mathsf{\epsilon }, \mathrm{i}=1,2,\dots ,\mathrm{n}$$

(13)

${\mathrm{X}}_1,{\mathrm{X}}_{2,\dots ,}{\mathrm{X}}_{\mathrm{n}}$, the formula is presented in Equation (13).

Among Equation (13), i is the total number of samples, p is the number of features, $\left(X_1,X_2,\dots ,X_i\right)$ is the explanatory variable, Y is the output value, and $({\beta }_0,\dots ,{\beta }_p)$ is the parameter to be estimated.

3. Case Results

3.1. ECM

ECM is a chemical reaction that alters the energy of the electrons on the electrode surface by influencing electric level to transfer electrons between the conductive workpiece and electrodes. Moving electrons from workpiece to electrode is the result of an oxidation reaction, with the electrode serving as the anode. In ECM (Figure 2a), the feed rate is a crucial factor for changing the size [6,7]. A slower feed speed at a single position may lead to excessive erosion of the chemical liquid. This study established the effect of different sizes, designs, and experimental parameters, as presented as

Table 1. ECM Experiment.

No.	Voltage (V)	Machining Depth (mm)	Feed Rate (µm/s)
1	18	10	7
2	18	10	9
3	18	10	11
4	18	10	13
5	18	10	16

. Constant voltage was used to alter the feed rate, and we collected the spindle voltage and current for the subsequent step of analysis.

The method involved moving the Y axis for processing, as presented in Figure 2b. After five experiments, the workpiece was measured using the coordinate-measuring machine (CMM), as illustrated in Figure 3a, and measured according to its machining direction, as presented in Figure 3b. The measurement results are summarized in

Table 2. Measurements after ECM.

	No.1	No.2	No.3	No.4	No.5
	$({\mathbf{mm}}^2)$	$({\mathbf{mm}}^2)$	$({\mathbf{mm}}^2)$	$({\mathbf{mm}}^2)$	$({\mathbf{mm}}^2)$
Level 1	186.737	184.648	182.246	182.633	179.452
Level 2	174.828	170.764	167.566	167.369	163.481
Level 3	173.274	169.486	166.121	166.293	162.794
Level 4	167.916	168.147	165.297	165.208	161.791
Level 5	165.757	169.146	167.441	167.067	164.066
Level 6	171.004	168.132	164.518	164.154	161.284

. Because the EDM study involved the contact area, the measurement area was used for subsequent model establishment.

Different designs led to unique contact areas, as presented in Table 2. The electrical signals measured using a current hook meter and voltage probe are presented in Figure 4 and Figure 5. The currents were affected by processing. Periodic results of the behavior changes were obtained. The faster the input was, the higher the current and the shorter the processing time were.

An electrical machining (ECM and EDM) matching method was developed to maintain the machining accuracy of the approximate machining process and the postprocessing finishing process, shorten the process connection processing time, analyze the correlations across the process, and calculate the electrical signal and the removal area. The mechanism achieved process matching.

Heat map analysis was performed with voltage, current, and postprocessing measurements. The correlation between each variable was determined by assigning a score from −1 to 1. Figure 6a presents the current versus area. A high correlation was observed between current.

A pair plot was used to analyze the distribution relationship between the two methods, indicating all the distributions between each variable. Figure 6b demonstrates that the relationship between current and size was linear (green color frame) and that between voltage and size (red color frame) was clustered.

In application, voltage and current are regarded as input data, and the removed area (dimension) is regarded as output to establish a linear regression model, as presented in Figure 7; to estimate the processing quality, regression analysis is used to determine the relationship between the reaction variable Y and the explanatory variables current and voltage.

The total number of samples was 200, and the training model was randomly cut at 90% (180) and 10% (20) for data verification. The mean absolute error (MAE) was used to evaluate the difference between the regression results and the experimental results. If no difference between the two was evident, the MAE was 0. As presented in Figure 8, the MAE of this model reached $0.308{ \mathrm{mm}}^2$.

To compare the prediction and experimental results, six points were extracted proportionally. The statistical table in Figure 9 shows the actual measurements and predictions, and the error between single-level points and points is expressed as the MAE. The minimum error of No. 2 MAE is $1.856{ \mathrm{mm}}^2$. According to the single-point error results in

Table 3. Discrepancies between model predictions and experimental results.

	No.1	No.2	No.3	No.4	No.5
	$({\mathbf{mm}}^2)$	$({\mathbf{mm}}^2)$	$({\mathbf{mm}}^2)$	$({\mathbf{mm}}^2)$	$({\mathbf{mm}}^2)$
Level 1	1.653	4.096	9.974	4.769	8.97
Level 2	2.576	0.284	1.119	0.66	2.696
Level 3	2.272	1.22	1.352	0.225	2.058
Level 4	2.561	1.351	0.954	0.192	1.66
Level 5	3.943	0.704	2.121	2.301	1.911
Level 6	1.514	0.124	0.638	0.104	0.02

, the minimum error of a single point is No. 5, the bottom six level of $0.02{ \mathrm{mm}}^2$. The resulting error was within the acceptable range. This model can be used for online real-time measurement (monitoring), and the processing quality per second can be obtained and input into EDM with the information of “predicted size and time,” which is convenient for two processes that are connected in series.

3.2. EDM

The EDM process uses a pulsed current to melt the electrode extremely close to the area where the workpiece is. It is often used to manufacture parts or molds with high complexity or hard materials. For the installation of voltage probes and current hooks on the EDM as presented in Figure 10, the electrical signals of the machine spindle were collected and signal analysis was conducted.

EDM requires a long time and has a high discharge frequency (>10 kHz) in Figure 11, and therefore, a large amount of data was collected. Feature calculation and extraction from the data are complex. For analyzing EDM process quality, typical steps include data collection, feature calculation, model building, and quality estimation in Figure 12.

The workpieces processed using ECM were processed through EDM in sequence. The final results are presented in Figure 13, and the three-dimensional measurement results are presented in

Table 4. Averages of the feature values.

Feature	ASF
	No.1	No.2	No.3	No.4	No.5
Average value	240.273	255.679	126.244	103.03	67.898
Feature	AIDT
	No.1	No.2	No.3	No.4	No.5
Average value	1.258	1.156	1.321	1.558	1.828

. The exit (the sixth level) was observed and is presented in EDM measurement results.

Different processing conditions were observed before EDM processing, and further improvement was required. After ECM, the shape was incorrect and had high roughness. Yet, this process saved much time, reducing machining time from hours to minutes. EDM was then used; it is superior for use with nickel-based alloys, which are difficult materials for machining.

As presented in Table 4 and Figure 14, EDM must be connected with different sizes after different processing parameters of ECM. To be closest to the actual processing state, we discussed the export. In data analysis, approximately 20% of the data collected before the end of processing was analyzed, and the results after calculating the characteristics were assessed subsequently. The model for estimating the machining effective area and protocol matching led to enhanced efficiency and path optimization of the ECM process. To estimate one diffuser with 88 polygon shape hole machining, the total working time provided the superior time of 7.3 h to combine ECM and EDM. In comparison, in the traditional process using only EDM, the time consumed is 4.8 days.

The discharge performance was indicated with high ASF values, suggesting high machining efficiency. In the observations, the removal area No. 1 > No. 2 > No. 3 > No. 4 > No. 5 was observed in the CMM results. After the calculation of the characteristics using this method, the ASF was observed to be No. 1 > No. 2 > No. 3 > No. 4 > No. 5. The larger contact areas and spark frequencies are presented in Figure 15.

In the CMM results, the obtained removal area was No. 1 > No. 2 > No. 3 > No. 4 > No. 5. After calculating the characteristics of this method, No. 5 was observed in the AIDT > No. 4 > No. 3 > No. 2 > No. 1. Thus, the contact area was larger, and the spark delay time was shorter. High AIDT values indicate that the capabilities of slag removal and motion control require improvement, as presented in Figure 16. Subsequently, 20% of the data before the completion of processing was captured for use as the subsequent model data. In data analysis, a staged change was observed. With the difference of the contact area, features were identified and predicted, and the training of the neural network was conducted in this manner.

As the contact area decreased, the spark frequency values changed proportionally. This was proportional to the AIDT, and the quantified data of the whole section of machining results were averaged using calculation features. The model has effective processing features and using a neural network can quickly obtain predictions, such as those presented in

Table 5. Model predictions and actual results error.

	No.1	No.2	No.3	No.4	No.5
	$({\mathbf{mm}}^2)$	$({\mathbf{mm}}^2)$	$({\mathbf{mm}}^2)$	$({\mathbf{mm}}^2)$	$({\mathbf{mm}}^2)$
Level 1	11.318	29. 237	28.181	24.714	19.499
Level 2	1.879	5.174	0.1	1.573	0.481
Level 3	3.765	1.582	2.316	0.3	1.473
Level 4	1.485	1.015	0.584	2.578	1.059
Level 5	5.34	1.1995	0.788	4.646	2.687
Level 6	2.388	1.134	1.117	6.44	0.288

, which correspond to the aforementioned results.

To establish a neural network model, the features calculated using EDM were the input, as presented in Figure 17a, to estimate the processing quality. This analysis was used to determine the relationship between the reaction variable Y and the explanatory variables. The maximum training number of the model was 1000. The learning rate was 0.01, and a seven-level neural network was established. In the proposed model, the number of hidden nodes in the proposed model was [10, 4, 1, 10, 7, 10, 7, 4, 1], the total Parameters was 263, as presented in Figure 17b, and the model was established with 90% of the data. The other 10% was used for testing the model. The model establishment result MAE was (1.32 mm²), as presented in Figure 17c.

The trained model was tested to compare its results with actual values after forecasting. The minimum error of No. 3 MAE is $1.232 {\mathrm{mm}}^2$, as presented in Figure 18. The minimum error of No. 3 single point was $0.1 {\mathrm{mm}}^2$, and its error result is presented in Table 5. Because the processing easily caused a cone shape, the prediction of level 1 was deemed poor.

By combining the sensing collection, signal analysis, feature calculation, and model prediction of the two processes, their respective advantages were preserved, and data could be effectively transmitted to intelligent monitoring to compensate size for differences between EDM and ECM.

4. Conclusions

Electrical processing is the most common mold-manufacturing method for metal products. The quality of each section may be affected by the design, manufacturing, materials, equipment, enhanced processing, assembly, and testing. Common electrical machining methods include ECM, EDM, and wire-EDM. Electrical processing has been actively used in the aerospace industry. It is a processing method for high-strength, heat-resistant materials, with advantages for rapid processing of materials into complex geometric shapes. However, the process is technically challenging.

This research integrated nontraditional processing parameter recommendations and regulations to establish parameter-adaptive adjustment recommendations based on the database table method. Process parameter status monitoring trains machine learning to obtain the correlations of processing quality goals to establish a suitable data inversion algorithm for nontraditional processing.

A multiple-process method is the proposed solution. A special-shaped structure-through-hole processing was used as a verification vehicle, which can be used in the original EDM. ECM was improved from the original 6-h processing time to 1 h, constituting an 80% reduction of processing time. We determined the key to controlling factor-pulse currents that affect the processing size before and after the experiment. After the evaluation results were adjusted, the expected processing accuracy requirements were met. In the decision model of the data inversion algorithm, the minimum model prediction accuracy was $0.01 {\mathrm{mm}}^2$. This study developed nontraditional processing quality monitoring and prediagnosis. The integrated data inversion algorithm powers process modeling, quality prediction, and parameter adjustment suggestions. Online actual processing parameter revision and feedforward compensation can be performed, and the process accuracy of the front and back stages is integrated. To address the matching problem, we have overcome the high-value processing technology gap in industry.