TWI628021B - Method for extracting intelligent features for predicting precision of electrical discharge machine and predicting method - Google Patents

Abstract

A feature extraction method for accuracy prediction of an electric discharge machine includes the following steps. Obtain the array process data when the EDM machine processes several workpiece samples. Each set of process data includes a discharge voltage signal and a discharge current signal. Process characteristics are established using process data. Array sample data associated with processing features during operation of the electrical discharge machine is collected. The probability distribution method is used to analyze the sample data associated with each processing feature. Obtain several measured values of the workpiece sample measured by the measuring machine. A correlation analysis step is performed to obtain a plurality of correlation coefficients between the processing features and the measured values. A preset machining feature is selected from the machining features using a correlation coefficient.

Description

Feature extraction method and prediction method for accuracy prediction of electric discharge machine

The invention relates to a feature extraction method for processing precision prediction of a processing machine and a prediction method for processing precision, and in particular to a feature extraction method for precision prediction of an electric discharge machine and prediction of machining accuracy of the electric discharge machine method.

In general electric discharge machining, when the electric discharge machining is performed, the workpiece size is not as expected and the surface roughness of the workpiece is large due to factors such as poor slagging, abnormal short circuit, or electrode consumption, resulting in poor workpiece quality. Moreover, the general EDM machine cannot effectively predict the processing quality of each workpiece during the machining process, and only the measuring device can measure the quality of the processed workpiece for the finished workpiece. This practice is not only time-consuming but also inefficient.

In order to solve the problem that the workpiece processing process cannot immediately detect the lack of processing quality, there is a prediction system that can predict the processing quality of the workpiece during the processing operation of the processing machine. However, how to find out the electrical discharge machining The pre-set characteristics of the machine that affect the processing quality and processing accuracy to provide a prediction system to predict the machining accuracy have become the goal of the relevant industry.

SUMMARY OF THE INVENTION Accordingly, it is an object of the present invention to provide a feature extraction method for accuracy prediction of an electric discharge machine and a prediction method for processing accuracy of the electric discharge machine, thereby effectively establishing a machining feature from process parameters of the electric discharge machine, and The preset features affecting the processing quality and the processing precision are selected from the processing features to provide a prediction system to predict the machining accuracy, thereby improving the quality of the workpiece.

According to the above object of the present invention, a feature extraction method for accuracy prediction of an electric discharge machine is proposed, which comprises the following steps. The multi-array process data is obtained when the EDM machine processes a plurality of workpiece samples respectively, wherein each set of process data includes a discharge voltage signal and a discharge current signal. A plurality of processing features are created using process data. Obtaining a plurality of measured values of the workpiece samples measured by a measuring machine, wherein each of the measured values is a measured value of the workpiece processed by the electric discharge machine according to the process data. A correlation analysis step is performed to obtain a plurality of correlation coefficients between the processed features and the measured values. A correlation coefficient is utilized to select at least one predetermined processing feature from the processing features.

According to the above object of the present invention, another feature extraction method for accuracy prediction of an electric discharge machine is proposed, which comprises the following steps. Obtain the complex array process data when the EDM machine processes a plurality of workpiece samples respectively. Each set of process data includes a discharge voltage signal and a discharge current signal. A plurality of processing features are created using process data. Complex array sample data associated with processing features during operation of the electrical discharge machine is collected. The probability distribution method is used to analyze the sample data associated with each processing feature to determine whether the average number and standard deviation of the sample data of each processing feature can represent the processing characteristics, and then obtain the judgment result. Obtaining a plurality of measured values of the workpiece sample measured by the measuring machine, wherein each of the measured values is a measured value of the workpiece sample processed by the electric discharge machine according to the process data. A correlation analysis step is performed to obtain a plurality of correlation coefficients between the processed features and the measured values. A correlation coefficient is utilized to select at least one predetermined processing feature from the processing features.

According to the above object of the present invention, there is further provided a method for predicting the processing accuracy of an electric discharge machine, comprising the following steps. At least one predetermined processing feature corresponding to each of the measured values is obtained by the feature extraction method described above. A prediction model for each measurement is established using each of the measurements and at least one predetermined machining feature corresponding to each of the measurements. The electrical discharge machine is operated in accordance with the process data to process the workpiece and collect a set of detected data of the workpiece associated with the process data during operation of the electrical discharge machine. The group detection data of the converted workpiece is at least one set of feature data. At least one set of feature data of the workpiece is input into the predictive model, and at least one predicted accuracy value of the workpiece for the measured value is estimated.

It can be seen from the above that the present invention collects process parameters (such as discharge voltage signals and discharge current signals) of the electric discharge machine, and establishes machining features according to the process parameters, and extracts preset features affecting the machining precision from the machining features, and further Provide predictive systems to effectively predict machining accuracy and improve machining quality.

On the other hand, the present invention can extract appropriate data from a large amount of sample data to represent the corresponding processing features by means of adaptive analysis. Moreover, by fitting the analysis, it is possible to further verify that the sample data uses its mean and standard deviation to represent the reliability of the corresponding processing features.

100‧‧‧Electrical discharge equipment

110‧‧‧Electric discharge machine

111‧‧‧ Spindle

111a‧‧‧electrode

120‧‧‧Processing module

121‧‧‧Sensor unit

122‧‧‧Capture unit

123‧‧‧Processing unit

200‧‧‧Measuring machine

300‧‧‧Virtual Measurement System

400‧‧‧Characteristic extraction method

410‧‧‧Steps

420‧‧ steps

430‧‧ steps

440‧‧‧Steps

450‧‧‧Steps

460‧‧ steps

470‧‧‧ steps

500‧‧‧Characteristic extraction method

510‧‧ steps

520‧‧‧Steps

530‧‧‧Steps

540‧‧‧Steps

550‧‧ steps

P1‧‧‧ workpiece

For a more complete understanding of the embodiments and the advantages thereof, the following description is made with reference to the accompanying drawings, wherein: FIG. 1 is a schematic view showing an apparatus of an electric discharge machining apparatus according to an embodiment of the present invention; 2] is a schematic flow chart showing a feature extraction method for accuracy prediction of an electric discharge machine according to an embodiment of the present invention; [Fig. 3] is a graph of voltage and current versus time; [Fig. 4] is simplified Schematic diagram of voltage waveform and current waveform; [Fig. 5] is a graph of current versus time; [Fig. 6] is a graph of current versus time; [Fig. 7] is a graph of current versus time; [Fig. 8] The relationship between voltage and current and time; [Fig. 9A] - [Fig. 9D] respectively, the histogram, empirical distribution map, QQ map (Quantile-Quantile Plot) and PP map obtained by matching the probability distribution of the spark frequency sample data. (probability-probability plot); [Fig. 10A] - [Fig. 10D] respectively, a histogram, an empirical distribution map, a QQ map, and a PP map obtained by fitting the open ratio to the sample data for the probability distribution; [Fig. 11A] - [Fig. 11D] are respectively for the short circuit ratio The sample data is used as the probability distribution, the empirical distribution map, the QQ map and the PP map; [Fig. 12A] - [Fig. 12D] are the histograms obtained by fitting the probability distribution of the average short-circuit current time sample data, The empirical distribution map, the QQ map, and the PP map; [Fig. 13A] - [Fig. 13D] are respectively a histogram, an empirical distribution map, a QQ map, and a PP map obtained by fitting the probability distribution data of the average short-circuit current sample data; [Fig. 14A] ]-[Fig. 14D] respectively, the histogram, empirical distribution map, QQ map and PP map obtained by fitting the probability data of the average delay time sample data; [Fig. 15A] - [Fig. 15D] are the average discharge peak current sample data respectively. The histogram distribution, the empirical distribution map, the QQ map and the PP map are obtained by the probability distribution; [Fig. 16A] - [Fig. 16D] respectively, the average discharge duration sample data is used as the probability distribution matching system. Obtained histogram, empirical distribution map, QQ map and PP map; [Fig. 17A] - [Fig. 17D] respectively, the histogram, empirical distribution map, QQ map and PP obtained by matching the probability distribution data of the average discharge energy sample data. Figure 18 is a schematic flow chart showing a method for predicting the processing accuracy of an electric discharge machine according to an embodiment of the present invention; [Fig. 19] is a test result generated by using a virtual measurement system to model the roughness of the bottom surface of the hole and its preset features; [Fig. 20] is a virtual circle for the opening of the hole under the hole using the virtual measurement system. The measured value of the measured value and its preset characteristics are used to model the test result; and [Fig. 21] is a test generated by modeling the measured value of the diameter of the opening under the hole with the virtual measuring system and its preset feature. result.

Please refer to FIG. 1 , which is a schematic diagram of an apparatus for electrical discharge machining apparatus according to an embodiment of the present invention, and FIG. 2 is a schematic diagram of accuracy prediction for an electric discharge machine according to an embodiment of the present invention. Schematic diagram of the extraction process. The feature extraction method 400 of the accuracy prediction of the present embodiment can be realized by the electric discharge machining apparatus 100 shown in FIG. 1. The electric discharge machining apparatus 100 of the present embodiment includes an electric discharge machine 110 and a processing module 120 electrically connected to the electric discharge machine 110. The electric discharge machine 110 has a main shaft 111, and the main shaft 111 is provided with an electrode 111a through which the workpiece sample P1 can be processed. The processing module 120 includes a sensing unit 121, a capturing unit 122, and a processing unit 123. The sensing unit 121 is mainly used to generate the sensing signal to provide the capturing unit 122 to determine whether to process the processing data when the electric discharge machine 110 is processed online. The processing unit 123 can establish a plurality of processing features according to the processed process data, and can collect the measured values of the different measurement items generated by the measuring machine 200, and then extract the processing precision from the processing features. The preset features are provided to provide a virtual metrology system 300 to predict the machining accuracy of the EDM machine 110.

In an embodiment, the sensing unit 121 can be a laser ranging device and disposed on the main shaft 111. Thereby, the capturing unit 122 can select whether to process the process data (ie, the voltage signal and the current signal) of the electric discharge machine 110 through the position of the laser ranging device sensing electrode 111a. For example, the laser ranging device determines whether the moving direction of the spindle 111 is the feeding direction. If the determination result is YES, it indicates that the spindle 111 is being fed, so the capturing unit 122 can further obtain the processing during processing. data. Conversely, if the result of the determination is no, the capture unit 122 does not need to retrieve the process data. Therefore, the sensing unit 121 can effectively filter the required amount of data captured by the capturing unit 122. In other embodiments, the sensing unit 121 can also be an optical scale. Alternatively, the capturing unit 122 can also directly learn the coordinate position of the spindle 111 from the controller of the electric discharge machine 110 to determine whether the relevant process data is to be retrieved, and the required data captured by the screening and capturing unit 122 can also be obtained. The purpose of quantity.

Please refer to FIG. 1 to FIG. 3 at the same time, wherein FIG. 3 is a graph of voltage and current versus time. The feature extraction method 400 of the present embodiment mainly collects process data (for example, the discharge voltage signal and the discharge current signal in FIG. 3) in the discharge process of the electric discharge machine 100 according to predetermined processing instructions and workpiece sample characteristics, and then according to These process data establish a plurality of processing features, and the data fusion technology method is used to extract preset features that can be used to estimate the processing accuracy. These preset features provide a predictive system to predict the machining accuracy of the EDM.

2 and FIG. 3, the feature extraction method 400 of the present embodiment includes the following steps. First, proceed to step 410 to obtain a discharge. The complex array process data when the processing machine processes a plurality of workpiece samples respectively. In one example, each set of process data mainly includes a discharge voltage signal and a discharge current signal. In one example, the process data is the processing conditions (eg, open circuit voltage and pulse on/off time, etc.) of the electrical discharge machine during the electrical discharge machining process, and the machine state (eg, actual discharge voltage waveform and discharge current waveform, etc.). In this embodiment, the high voltage probe and the current hook can be used to sense the discharge voltage signal and the discharge current signal of the electric discharge machine.

After obtaining the process data of the electric discharge machine, step 420 is performed to establish a plurality of processing features using the process data. In this embodiment, a data fusion technique can be applied to create processing features from the discharge voltage signal and the discharge current signal, and these processing features can be used to estimate the processing accuracy. Before the processing feature is established, the discharge voltage waveform and the discharge current waveform as shown in FIG. 3 can be used to establish the threshold value and capture the effective waveform time. Please refer to FIG. 4, which is a schematic diagram of a voltage waveform and a current waveform. For the sake of clarity, FIG. 4 illustrates a simplified version of the voltage waveform and current waveform diagram, and t _d shown in FIG. 4 represents an ignition delay time or a charging time, and t _e represents a discharge duration time. ), t ₀ represents a pulse interval time, u ₀ represents an open voltage, and i _e represents a discharge current. The threshold value is established according to the experimental requirements. As shown in FIG. 4, for example, the line L1 can be established as the threshold value for discriminating the number of peaks, the line L2 is used as the threshold value for judging the short circuit, and the line L3 is used to discriminate the threshold value of the peak. The operation of capturing the effective waveform time includes extracting the start point and the end point of each current peak, calculating the time, extracting the start point and the end point of each voltage peak, and calculating the time and the time of the drawoff time. Start and end points.

In this embodiment, the processing features include a spark frequency, an open circuit ratio, a short circuit ratio, an average short circuit time, a short circuit time standard deviation, an average short circuit current, and a short circuit current standard deviation. Average delay time, standard deviation of delay time, Average discharge peak current, peak current standard deviation, average discharge time, discharge time standard deviation, average discharge energy, and discharge energy standard deviation.

In the processing characteristics, the spark frequency, the average discharge peak current, the peak current standard deviation, the average discharge time, and the discharge time standard deviation are mainly established from the discharge current signal. Among them, the spark frequency is defined as the total number of sparks that appear during the signal period. Please refer to FIG. 5 , which is a graph of current versus time. In order to identify the spark in the current signal, the pulse time range and the minimum threshold peak (for example, line L3 in Fig. 4) can be defined from the relationship between current and time in Fig. 5, and the peak value of the current wave exceeds the minimum in the pulse time range. When the threshold is peaked, a current spark can be defined.

Please refer to FIG. 6, which is a graph of current versus time. The discharge peak current is the maximum current value that reaches the workpiece through the electrode during the pulse period. The average discharge peak current is defined as the average value of the discharge peak currents measured over a period of time (for example, the period from the start of the pulse to the end of the pulse shown in FIG. 6). The peak current standard deviation is also calculated from the peak discharge current. As shown in Figure 6, the discharge duration (discharge duration) It is the time difference between the start point and the end point of the discharge current waveform (for example, the period from the start of the pulse to the end of the pulse shown in Fig. 6).

In the processing characteristics, the average delay time, the delay time standard deviation, the short circuit ratio, the average short circuit time, and the short circuit time standard deviation are established from the discharge voltage signal. Referring again to FIG. 4, the average delay time is defined as the time difference from the point in time when a sufficient open circuit voltage has been established to the gap between the electrode and the workpiece, and the start of the discharge current (for example, as shown in FIG. 4). _d ).

The short circuit ratio is defined as the number of short circuit pulses divided by the number of discharge pulses. Please refer to FIG. 7, which is a graph of current versus time. When the short circuit pulse is within a discharge pulse period and the open circuit voltage value continues to be less than the specified voltage threshold (for example, line L2 of FIG. 4), the discharge pulse period of the time is recorded as a short circuit pulse. As shown in Figure 7, the average short-circuit time and short-circuit time standard deviation are related to the short circuit duration, and the short-circuit duration is defined as multiple consecutive occurrences during a period of discharge pulse (more than two consecutive pulses are required). Short circuit, this short circuit duration is calculated as the time difference from the short circuit peak to the last short pulse peak during the short circuit.

In the processing characteristics, the open circuit ratio, the average discharge energy, the discharge energy standard deviation, the average short circuit current, and the short circuit current standard deviation are jointly established based on the discharge current signal and the discharge voltage signal. Please refer to FIG. 8 , which is a graph of voltage and current versus time. As shown in Figure 8, when a voltage peak (Ignition Voltage) ends and does not follow a current peak (Discharge Current), it is called an open circuit (open). Circuit). An open circuit indicates that the voltage peak fails to lead to a subsequent current peak, which is an invalid pulse. The open circuit ratio is defined as the number of open circuits divided by the total number of discharge pulses.

The average discharge energy is mainly used to maintain the stability of the electric discharge machining process to ensure the processing quality, and the discharge energy (Ei) formula of the i-th discharge is as shown in the following formula (1):

Where t _ei is the discharge duration, U _i is the discharge voltage, and I _pi is the discharge peak current. This formula assumes that the discharge voltage remains constant during the discharge process.

The short-circuit current is mainly defined as the average value of the maximum peak current (beyond the minimum threshold) and the minimum valley current (below the maximum threshold) during each pulse duration. The Average Short Current is defined as the value of all pulse short-circuit currents during the duration of the short circuit.

Referring to both FIG. 1 and FIG. 2, after a plurality of processing features are established, step 430 can be performed to collect the complex array sample data associated with the processing features during operation of the electrical discharge machine 110. In this embodiment, a mold steel having a length, a width, and a height of 30 mm, 30 mm, and 10 mm, respectively, is used as a workpiece, and an electrode having a diameter of 3 mm is used, wherein the depth of the hole is 100 μm and the current is 4 A. In the embodiment, the hole drilling process is taken as an example, and a hole is formed in the workpiece by drilling the workpiece through the electrode. During the processing, when the spindle 111 is in different feeding positions, the capturing unit 122 can capture a plurality of different current and voltage signals. Therefore, the complex array sample corresponding to the processing feature of the spindle 111 in different feed position intervals can be further obtained. material. For example, in the process of processing the workpiece sample P1 by the spindle 111, the processing unit 123 can obtain a plurality of spark frequency sample data, a plurality of open circuit ratio sample data, a plurality of short circuit ratio sample data, a plurality of average short circuit time sample data, and a number. Average short-circuit current sample data, several average delay time sample data, several delay time standard deviation sample data, several average discharge peak current sample data, several peak current standard deviation sample data, several average discharge time sample data, Several average discharge time standard deviation sample data, several average discharge energy sample data, and several discharge energy standard deviation sample data.

After obtaining the array sample data associated with the processing feature, step 440 is performed to perform a fit analysis on each sample data by using the probability distribution method to determine whether the average number and standard deviation of each sample data are representative. The overall processing characteristics, and then the judgment results. In the present embodiment, the normal distribution method, the Weibull distribution method, and the Gamma distribution method are used to perform fitting analysis on all sample data of each processing feature to determine whether the sample data conforms to the normal distribution. In some examples, the P-P map and the Q-Q map can be used to compare whether the distribution generated by the probability distribution method conforms to the ideal normal distribution curve. If the judgment result is yes, it means that the average number and standard deviation of the sample data of each processing feature can represent the corresponding processing features. Appropriate analysis can be used to extract the appropriate data from a large number of sample data to represent the corresponding processing features. Moreover, by fitting the analysis, it is possible to further verify that the sample data uses its mean and standard deviation to represent the reliability of the corresponding processing features. Conversely, if the result of the determination is no, it indicates that the correlation analysis of the processing feature can be omitted in the subsequent correlation analysis step, but this is not intended to limit the present invention.

Please refer to FIG. 9A - FIG. 9D . FIG. 9A - FIG. 9D are respectively a histogram, an empirical distribution diagram, a Q-Q diagram and a P-P diagram obtained by fitting the probability distribution of the spark frequency sample data. The Q-Q diagram and the oblique line in the P-P diagram represent the ideal line, and the ideal line represents the line that appears when the spark frequency sample data satisfies the normal distribution under ideal conditions. The blue circle, the red circle and the green circle in the figure respectively represent the actual data of the spark frequency sample data according to the normal distribution, the Weibull distribution and the distribution generated by the gamma distribution method. As can be seen from Figures 9A-9D, the distribution of the spark frequency sample data falls near the ideal line, which indicates that these spark frequency sample data are close to the normal distribution. Therefore, the average and standard deviation of the data of these spark frequency samples can be used to represent the eigenvalues of the overall spark frequency.

Please refer to FIG. 10A - FIG. 10D . FIG. 10A - FIG. 10D are respectively a histogram, an empirical distribution diagram, a Q-Q diagram, and a P-P diagram obtained by fitting the probability ratio to the sample data. The Q-Q diagram and the oblique line in the P-P diagram represent the ideal line, and the ideal line represents the line that appears when the open circuit satisfies the normal distribution in the ideal case. The blue circle, red circle and green circle in the figure respectively represent the distribution of the actual data of the open-circuit ratio sample data according to the normal distribution, the Weibull distribution and the gamma distribution method. As can be seen from Figures 10A-10D, the distribution of open-circuit ratio sample data falls near the ideal line, which means that these open circuits are closer to the normal distribution than the sample data. Therefore, the average and standard deviation of these open circuit ratio sample data can be used to represent the eigenvalues of the overall open circuit ratio.

Please refer to FIG. 11A - FIG. 11D , and FIG. 11A - FIG. 11D are respectively a histogram and empirical distribution obtained by fitting the probability distribution of the short-circuit ratio sample data. Figure, Q-Q diagram and P-P diagram. The Q-Q diagram and the oblique line in the P-P diagram represent the ideal line, and the ideal line represents the line that appears when the short-circuit ratio of the sample data satisfies the normal distribution. The blue circle, the red circle and the green circle in the figure respectively represent the distribution of the actual data of the short-circuit ratio sample data according to the normal distribution, the Weibull distribution and the gamma distribution method. It can be seen from Fig. 11A - Fig. 11D that the distribution of the short circuit is larger than the ideal distribution of the sample data, which means that the average value and the standard deviation of the short circuit ratio sample data cannot be used to represent the characteristic value of the overall short circuit ratio.

Please refer to FIG. 12A - FIG. 12D . FIG. 12A - FIG. 12D are respectively a histogram, an empirical distribution diagram, a Q-Q diagram and a P-P diagram obtained by fitting the probability distribution of the average short-circuit current time sample data. The Q-Q diagram and the oblique line in the P-P diagram represent the ideal line, and the ideal line represents the line that appears when the sample data of the average short-circuit current time satisfies the normal distribution under ideal conditions. The blue circle, red circle and green circle in the figure respectively represent the actual data of the average short-circuit current time sample data according to the normal distribution, the Weibull distribution and the gamma distribution method. It can be seen from Fig. 12A to Fig. 12D that the distribution of the sample data of the average short-circuit current time differs greatly from the ideal distribution, which means that the average and standard deviation of the sample data of these average short-circuit current times cannot be used to represent the overall average short-circuit. Characteristic value of current time.

Please refer to FIG. 13A - FIG. 13D. FIG. 13A - FIG. 13D are respectively a histogram, an empirical distribution diagram, a Q-Q diagram and a P-P diagram obtained by fitting the probability distribution of the average short-circuit current sample data. The Q-Q diagram and the diagonal line in the P-P diagram represent the ideal line, and the ideal line represents the ideal case, the average short circuit The sample data of the current satisfies the line presented in the normal distribution. The blue circle, the red circle and the green circle in the figure respectively represent the actual data of the sample data of the average short-circuit current according to the normal distribution, the Weibull distribution and the distribution generated by the gamma distribution method. It can be seen from Fig. 13A to Fig. 13D that the distribution of the sample data of the average short-circuit current differs greatly from the ideal distribution, which means that the average and standard deviation of the sample data of these average short-circuit currents cannot be used to represent the overall average short-circuit current. Eigenvalues.

Please refer to FIG. 14A - FIG. 14D. FIG. 14A - FIG. 14D are respectively a histogram, an empirical distribution diagram, a Q-Q diagram and a P-P diagram obtained by fitting the probability distribution of the average delay time sample data. The Q-Q diagram and the oblique line in the P-P diagram represent the ideal line, and the ideal line represents the line that appears when the sample data of the average delay time satisfies the normal distribution under ideal conditions. The blue circle, the red circle and the green circle in the figure respectively represent the actual data of the sample data of the average delay time according to the normal distribution, the Weibull distribution and the distribution generated by the gamma distribution method. As can be seen from Figs. 14A - 14D, the distribution of the sample data of the average delay time falls near the ideal line, which means that the sample data of these average delay times is close to the normal distribution. Therefore, the average and standard deviation of the sample data that can be used for these average delay times can represent the eigenvalues of the overall average delay time.

Please refer to FIG. 15A - FIG. 15D . FIG. 15A - FIG. 15D are respectively a histogram, an empirical distribution diagram, a Q-Q diagram and a P-P diagram obtained by fitting the probability distribution of the average discharge peak current sample data. The oblique lines in the Q-Q diagram and the P-P diagram represent ideal lines, and the ideal line represents a line which is presented when the sample data of the average discharge peak current satisfies the normal distribution under ideal conditions. and The blue circle, the red circle and the green circle in the figure represent the distribution of the actual data of the sample data of the average discharge peak current according to the normal distribution, the Weibull distribution and the gamma distribution method, respectively. As can be seen from Figs. 15A to 15D, the Weibull distribution of the sample data of the average discharge peak current in the P-P diagram falls near the ideal line. Therefore, the average and standard deviation of the sample data that can be used for these average discharge peak currents can represent the eigenvalues of the overall average discharge peak current.

Please refer to FIG. 16A - FIG. 16D . FIG. 16A - FIG. 16D are respectively a histogram, an empirical distribution diagram, a Q-Q diagram and a P-P diagram obtained by fitting the probability distribution of the average discharge duration sample data. The Q-Q diagram and the oblique line in the P-P diagram represent the ideal line, and the ideal line represents the line that appears when the sample data of the average discharge duration satisfies the normal distribution under ideal conditions. The blue circle, the red circle and the green circle in the figure respectively represent the actual data of the sample data of the average discharge duration according to the normal distribution, the Weibull distribution and the distribution generated by the gamma distribution method. As can be seen from Figs. 16A - 16D, the normal distribution of the sample data of the average discharge duration shown in the P-P diagram falls near the ideal line, which means that the sample data of these average discharge durations are close to the normal distribution. Therefore, the average and standard deviation of the sample data that can be used for these average discharge durations can represent the eigenvalues of the overall average discharge duration.

Please refer to FIG. 17A - FIG. 17D. FIG. 17A - FIG. 17D are respectively a histogram, an empirical distribution diagram, a Q-Q diagram and a P-P diagram obtained by fitting the probability distribution data of the average discharge energy sample data. The Q-Q diagram and the diagonal line in the P-P diagram represent the ideal line, and the ideal line represents the ideal discharge under ideal conditions. The sample data of energy satisfies the lines presented in the normal distribution. The blue circle, the red circle and the green circle in the figure respectively represent the actual data of the sample data of the average discharge energy according to the normal distribution, the Weibull distribution and the distribution generated by the gamma distribution method. As can be seen from Figures 17A-17D, the distribution of the sample data of the average discharge energy falls near the ideal line, which means that the sample data of these average discharge energies are close to the normal distribution. Therefore, the average and standard deviation of the sample data that can be used for these average discharge energies can represent the eigenvalues of the overall average discharge energy.

After obtaining each processing feature representative, step 450 is performed to obtain a complex array measurement value of the workpiece sample measured by the measuring machine, wherein each set of measurement values respectively processes the workpiece according to the process data of the electrical discharge machining machine. The measured value of the sample. In this embodiment, the boring processing is taken as an example. Holes are formed in the workpiece sample by drilling the workpiece sample through the electrode, and each hole has an opening above the upper surface of the workpiece sample and an opening below the lower surface of the workpiece sample. The measurement items of the embodiment are respectively the roughness of the bottom surface of the hole, the roundness of the upper opening, the diameter of the upper opening, the true roundness of the lower opening, and the diameter of the lower opening, and each measurement item has a corresponding amount respectively. Measured value.

After obtaining the measured value, step 460 is followed to perform a correlation analysis step to obtain a plurality of correlation coefficients between the processed feature and the measured value. In one embodiment, the relationship between the processed features and the measured values can be observed using the MATLAB tool and using the following relation (2): Where i is used to indicate the i- th workpiece sample, X is the machining feature, and Y is the measurement value of the different measurement items.

After finding the correlation coefficient between the processing feature and the measured value, step 470 is performed to extract the preset features related to the measured values of the different measurement items by using the correlation coefficient. In one embodiment, when the equivalent measurement is a measure of the roughness of the bottom surface of the hole of the workpiece sample, the processing characteristics having a correlation coefficient greater than 0.2 or less than -0.2 are as shown in Table 1 below. Therefore, the first five processing features with higher correlation coefficients can be selected from Table 1 as preset features related to roughness, such as average short-circuit current average, average short-circuit current standard deviation, average delay time average, average Delay time standard deviation and average discharge energy standard deviation.

In another embodiment, when the equivalent measurement is a measure of the true roundness of the opening below the workpiece sample, the processing characteristics having a correlation coefficient greater than 0.2 or less than -0.2 are as shown in Table 2 below. Therefore, the first five processing features with higher correlation coefficients can be selected from Table 2 as preset features related to the true roundness of the lower opening, such as the average value of the spark frequency, the standard deviation of the spark frequency, the average of the short circuit ratio, and the short circuit. The difference from the standard and the average of the average discharge time.

In another embodiment, when the equivalent measurement is a measure of the diameter of the opening below the workpiece sample, the processing characteristics having a correlation coefficient greater than 0.2 or less than -0.2 are as shown in Table 3 below. Therefore, the first five processing features with higher correlation coefficients can be selected from Table 3 as preset features related to the diameter of the lower opening, such as spark frequency standard deviation, average short circuit current time standard deviation, average short circuit current standard deviation, Average delay time standard deviation and average discharge time average.

For the sake of clarity, the aforementioned selection of predetermined features is to screen with processing features having a correlation coefficient greater than 0.2 or less than -0.2 and is not intended to limit the invention. Under other process conditions, the thresholds of different correlation coefficients can be determined according to different experimental requirements to select preset features.

After obtaining the preset features related to the measured values, each of the measured values and the preset features corresponding to each of the measured values may be used to be imported into the database of the virtual measuring system for each measurement. The value establishes a predictive model, which in turn provides a virtual measurement system to predict the machining accuracy. The prediction system currently used in the present embodiment may be one of the full disclosures as disclosed in the Republic of China Patent No. I349867. The Automatic Virtual Metrology (AVM) system is used as a prediction system for processing accuracy, but this is not intended to limit the present invention.

Please refer to FIG. 18 , which is a flow chart showing a method for predicting the processing accuracy of an electric discharge machine according to an embodiment of the present invention. The method 500 for predicting the processing accuracy of the electric discharge machine of the present invention comprises the following steps. First, step 510 is performed to obtain a preset processing feature corresponding to each measurement value by using the feature extraction method as shown in FIG. 2. Next, step 520 is performed to establish a prediction model for each measurement value using each measurement value and a preset processing feature corresponding to each measurement value. Next, step 530 is performed to operate the electrical discharge machine in accordance with the process data to process the workpiece and collect a set of detected data of the workpiece associated with the process data during operation of the electrical discharge machine. Then, step 540 is performed to convert the detected data of the workpiece into feature data. Next, step 550 is performed to input the feature data of the workpiece into the prediction model, and the prediction accuracy value of the workpiece for the measured value is estimated.

Please refer to FIG. 19, which is a test result generated by using a virtual measurement system to model the measured value of the roughness of the bottom surface of the hole and its preset features. Among them, the present embodiment uses a neural network (NN) and a partial least square method (Partial Least Square, PLS) as an accuracy prediction model to predict the roughness of the bottom surface of the hole. The prediction results are shown in Table 4 below. From Table 4, it can be observed that the Mean Absolutely Error (MAE) of NN and PLS is 0.010, and the 95% Max Error of NN and PLS is 0.017 and 0.016um, respectively, which is less than the actual amount. The measurement (Real Y) is twice the standard deviation of 0.015 um, which means that the two models can be used to predict the roughness of the bottom surface of the hole with the extracted preset features.

Please refer to Table 5 below and FIG. 20, wherein FIG. 20 is a test result generated by using a virtual measurement system to model the measured value of the true roundness of the opening under the hole and its preset characteristics. As shown in Table 5 and Figure 20, when the virtual measurement system is used to model and test the measured value of the true roundness of the lower opening of the hole and its preset characteristics, the MAE of NN and PLS are 0.005 and 0.004 um, respectively. The 95% Max Error of NN and PLS is 0.006 and 0.011 um, respectively, which is less than the standard deviation of Real Y by 0.01, which means that the extracted preset features can be used to predict the true roundness of the lower opening of the hole.

Please refer to Table 6 below and FIG. 21, wherein FIG. 21 is a test result generated by using a virtual measurement system to model the measured value of the diameter of the opening under the hole and its preset characteristics. As shown in Table 6 and Figure 21, when using the virtual measurement system to model and test the diameter of the lower opening of the hole and its preset characteristics, the MAE of NN and PLS is 0.007 and 0.008 um, respectively. The 95% Max Error with PLS is 0.012 and 0.012 um, respectively, which is less than the actual measurement (Real Y) twice the standard deviation of 0.017, which means that the extracted preset features can be used to predict the diameter of the lower opening of the hole.

Table 6. Prediction accuracy of the diameter of the lower opening of the hole

According to the embodiment of the present invention, the present invention collects process parameters (such as discharge voltage signals and discharge current signals) of the electric discharge machine, and establishes machining features according to the process parameters, and extracts preset features that affect machining precision from the machining features. In turn, the prediction system is provided to effectively predict the machining accuracy, thereby improving the processing quality.

On the other hand, the present invention can extract appropriate data from a large amount of sample data to represent the corresponding processing features by means of adaptive analysis. Moreover, by fitting the analysis, it is possible to further verify that the sample data uses its mean and standard deviation to represent the reliability of the corresponding processing features.

Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention, and any one of ordinary skill in the art can make some changes and refinements without departing from the spirit and scope of the present invention. The scope of the invention is defined by the scope of the appended claims.

Claims (9)

A feature extraction method for accuracy prediction of an electric discharge machine: obtaining a complex array process data when an electric discharge machine separately processes a plurality of workpiece samples, wherein each of the group process data includes a discharge voltage signal and a discharge current signal And using the set of process data to establish a plurality of processing features; obtaining a plurality of measured values of the workpiece samples measured by a measuring machine, wherein each of the measured values is respectively determined by the electrical discharge machine Processing data of the workpiece samples processed by the process data; performing a correlation analysis step according to a correlation coefficient formula to obtain a plurality of correlation coefficients between the processing features and the measurement values; and utilizing the correlations And a coefficient to select at least one predetermined processing feature from the plurality of processing features, wherein the step of selecting the at least one predetermined processing feature from the processing features by using the correlation coefficients comprises selecting a greater correlation among the correlation coefficients The processing characteristics of the coefficients are represented as a preset processing feature. A feature extraction method for accuracy prediction of an electric discharge machine includes: obtaining a complex array process data when an electric discharge machine separately processes a plurality of workpiece samples, wherein each of the group process data includes a discharge voltage signal and a discharge a current signal; using the set of process data to create a plurality of processing features; Collecting complex array sample data associated with the processing features during operation of the electric discharge machine; performing fitting analysis on the sample data associated with each processing feature by using a probability distribution method to determine each processing feature Whether an average number and a standard deviation of the sample data can represent the processing feature, thereby obtaining a determination result, wherein if the determination result is yes, performing the following steps: obtaining the quantity measured by a measuring machine a plurality of measured values of the workpiece samples, wherein each of the measured values is a measured value of the workpiece samples processed by the electrical discharge machine according to the process data; and a correlation is performed according to a correlation coefficient formula An analysis step of obtaining a plurality of correlation coefficients between the processing features and the measured values; and using the correlation coefficients to select at least one predetermined processing feature from the processing features, wherein the correlation coefficients are utilized The step of selecting at least one of the predetermined processing features includes selecting processing features having a larger correlation coefficient among the correlation coefficients Representatives to serve as pre-processing features. The feature extraction method for accuracy prediction of an electric discharge machine as described in claim 2, wherein the probability distribution method comprises a normal distribution method, a Weibull distribution method, and a Gamma distribution method. A feature extraction method for accuracy prediction of an electric discharge machine as described in claim 2, wherein the fitting analysis step The method is to compare whether the sample data of each processing feature conforms to a normal distribution. If the comparison result is yes, the average number and standard deviation of the sample data of each processing feature may represent the corresponding processing feature. A feature extraction method for accuracy prediction of an electric discharge machine according to claim 2, wherein the fitting analysis step is a ratio of a probability-probability plot and a quantification-quantity plot. Whether the distribution generated by the probability distribution method conforms to an ideal distribution curve, and if the comparison result is yes, the average number and standard deviation of the sample data of each processing feature may represent the corresponding processing feature. The feature extraction method for accuracy prediction of an electric discharge machine according to claim 1 or 2, wherein the processing features include a spark frequency, an open circuit ratio, a short circuit ratio, an average short circuit time, and a short circuit. Time standard deviation, an average short circuit current, a short circuit current standard deviation, an average delay time, a delay time standard deviation, an average discharge peak current, a peak current standard deviation, an average discharge time, a discharge time standard deviation, one Average discharge energy and a standard deviation of discharge energy. The feature extraction method for accuracy prediction of an electric discharge machine according to claim 1 or 2, wherein the measurement values include roughness measurement values of the workpiece samples. A feature extraction method for accuracy prediction of an electric discharge machine according to claim 1 or 2, wherein the electric discharge machine performs a drilling step for each of the workpiece samples, respectively, and each of the Forming a hole in the workpiece sample, wherein each of the holes has an opening on one of the upper surfaces of each of the workpiece samples and an opening under one of the lower surfaces of each of the workpiece samples; wherein the measured values include a measured value of a diameter of one of the upper openings, a measured value of one roundness of the upper opening, a measured value of a diameter of one of the lower openings, or a measured value of one roundness of the lower opening. A method for predicting the processing accuracy of an electric discharge machine, comprising: obtaining, by using the feature extraction method described in claim 1 or 2, the at least one predetermined processing feature corresponding to each of the measured values; And measuring the at least one predetermined processing feature corresponding to each of the measured values to establish a prediction model for each of the measured values; operating the electrical discharge machine according to the processing data to process a a workpiece, and collecting a set of detection data of the workpiece associated with the process materials during operation of the electrical discharge machine; converting the set of detection data of the workpiece to at least one set of feature data; and inputting the workpiece At least one set of feature data into the predictive model, and estimating at least one prediction of the workpiece for the at least one measured value Precision value.