KR0173676B1 - Inspect apparatus and method of on-line non destructive testing for spot welding - Google Patents

Abstract

The present invention relates to an online non-destructive testing device and an inspection method using a neural network, in particular in an electric resistance welding inspection device and an inspection method, comprising: a first electrode connected to one side of a power source to provide an online non-destructive inspection device; In the electrical resistance welding comprising a second electrode connected to the other side of the power source and a base material inserted between the first electrode and the second electrode, the first analog for detecting a voltage applied to both ends of the base material during welding; A voltage waveform measurement system having a digital converter (A / D converter), and a second analog-to-digital converter for inputting sensor means for detecting a displacement between the first electrode and the second electrode and an output of the sensor means. An electrode expansion waveform measuring system comprising: a neural network inspection system having an input of an output of the voltage waveform measuring system and the electrode expansion waveform measuring system; By providing a computer system with a built-in system, it is a non-destructive method and an online inspection is performed, which greatly contributes to the improvement of product robustness and automation of the process, and can greatly reduce the conventional physical and human losses. An online non-destructive testing device and test method for resistance welding is provided.

Description

On-line nondestructive testing device and method of electric resistance welding

The present invention relates to an inspection apparatus and an inspection method for electric resistance welding in the welding technology, and more particularly, to an on-line non-destructive inspection device and method for performing inspection in a non-destructive manner at the same time by welding using a neural network. will be.

Electrical resistance welding refers to welding in which the base material is melted and applied by applying external force as heat of electric resistance generated in the contact portion through current to the base material. The electrical resistance heat generated at this time is represented by the following formula (1).

(Q: calorie (cal), I: current (a), R: resistance (Ω), t: time (sec))

In general, a material with a large resistance is used as a load side to pass a large amount of low-voltage current from a power supply to use a heat of resistance generated at this time.

At this time, the current is used up to about 80,000 (a), and the no-load terminal voltage at the load side that sends this is very low, about 1 to 10 (V).

In order to obtain such low voltage and high current, it is most convenient to use transformer by adopting AC power.

Types of the electric resistance welding include butt welding, spot welding, seam welding, projection welding, and the like.

They have the advantage of low welding temperature, fast working speed and high stability of welded parts.

Moreover, the spot welding is inexpensive, productive in mass, has high efficiency in bonding strength, and has advantages in performance such as weight reduction, material saving, and structure simplification.

In addition, the spot welding is generally used in the metal processing field because the welding conditions are automatically determined according to the machine and there is no need to rely on the skill or function of the operator.

Hereinafter, spot welding will be described in detail.

1 is a schematic diagram of a spot welder.

In FIG. 1, the welding transformer 1, the control unit 2 connected to the primary side of the welding transformer 1, and the electrode (or welding rod) connected to the secondary side of the welding transformer 1 3a, 3b), the press part 5 for pressing the said electrode, and the base material 7 mounted between the said electrodes 3a and 3b are shown.

As shown in the drawing, when the base material 7 is inserted between the anodes 3a and 3b, and lightly presses the pressurizing part 5 and passes the current, the joint part heats red locally due to resistance heat generated in the weld part.

At this time, if the pressure is pressed again, the spot welding is performed only by the area of the electrode to be contacted. The welds form a goggle, which is called a nugget.

As described above, since spot welding is made in a point shape, it is also called spot welding, and unlike rivet joints in which a hole is immediately punched and joined to a plate, it can be bonded without a hole.

In this case, the defect of the weld portion decreases the reliability of welding, and in particular, care is required when a cyclic load or an impact load is applied.

In addition, safety is important in the case of high pressure and high load, so there should be no defect.

There are two types of crime prevention methods for testing weld defects: non-destructive testing and non-destructive testing.

Fracture inspection is a test of specimens or special test welds from a number of products, such as the failure of welds and the examination of their quality.

Types of destructive testing include puncture, wavefront, macroscopic biopsy and microscopic biopsy.

In general, a direct fracture inspection method using a specimen, which performs a fracture inspection on a weld selected as a specimen after performing a certain number of spot welding, is commonly used.

If a problem is found after the test, all previous work must be destroyed or re-created.

Therefore, in the case of the destruction inspection method, there is a problem that not only a problem of the robustness of the product, but also a great obstacle in the automation of the process due to the inefficiency of the inspection process.

In addition, there is a problem that a huge physical and human loss.

Another method of checking the weld status is nondestructive testing or inspection.

Non-destructive testing means a method of investigating the integrity of a material or product without changing the material, shape, or numerical value of the material.

The types include radiographic examination, ultrasonic examination, magnetic examination, paint penetration method, and the like.

In case of using X-ray inspection, which is one of the radiographic inspection methods, defects and segregation of impurities are easily transmitted because X-rays are easily transmitted. The failure state of the welding can be detected.

However, the above-mentioned X-ray inspection method is not only expensive, but also poses a problem to the stability of the factory, and there is an inconvenience that an X-ray technician must always reside.

On the other hand, in the case of the ultrasonography method, it is difficult to inspect that the surface is very rough or complicated in appearance, and the inspection is performed by the contact of the probe transducer. There is a problem.

In addition, the magnetic inspection method is limited to ferromagnetic materials such as steel, and only theoretical studies are currently performed.

On the other hand, in the case of the paint penetration method, there is a problem that the surface must be polished in advance without being detected unless it is a defect extending to the surface.

In addition, the various non-destructive inspection methods, like the destruction inspection method, have to undergo a separate post-welding inspection process, which causes a lot of time loss and human loss, and is currently rarely used.

In order to solve the problems of current and non-destructive inspection methods, it is desirable to perform the inspection at the same time as the welding, not after the spot welding.

As mentioned above, in order to test simultaneously with welding, the development of a new non-destructive inspection method is calculated | required.

However, in spot welding using mostly AC power, it is difficult to develop a non-destructive tester because the welding state changes according to the current and time during welding, the state of the electrode, the adjustment of the pressure state, and the material of the base material.

Accordingly, the present invention has been made in order to solve the above problems, to provide a welding inspection apparatus and inspection method by a non-destructive method in the inspection apparatus and inspection method of electrical resistance welding, and to the inspection apparatus and inspection method of electrical resistance welding. The present invention provides an on-line welding inspection apparatus and an on-line nondestructive inspection method for simultaneously performing a welding inspection by a non-destructive method.

In order to achieve the above, the present invention is characterized in that the waveform of the voltage applied to both ends of the base material during welding and the displacement of the gap between the electrodes according to the expansion of the base material during welding is used for the welding inspection, and the present invention In an electrical resistance welding comprising a first electrode connected to a high potential, a second electrode connected to a low potential of the power source, and a base material inserted between the first electrode and the second electrode, which is applied to both ends of the base material during welding. A voltage waveform measurement system having a first analog-to-digital converter for detecting a voltage, a sensor means for detecting a displacement between the first electrode and a second electrode, and an output of the sensor means. An electrode expansion waveform measurement system having a second analog-to-digital converter as an input, and a neural circuit having an output of the voltage waveform measurement system and the electrode expansion waveform measurement system as inputs And it characterized in that it comprises a computer system built in a network testing system.

Thus, the welding voltage waveforms of the signals digitalized by the analog-to-digital converter are inputted at the same time as the welding and at the same time, the voltage applied to both ends of the substrate and the displacement between the electrodes are input to the computer system through the analog-to-digital converter. And a first step of detecting and storing the filtered electrode expansion waveform by a predetermined software, a second step of performing a welding test by the post-welding fracture inspection method, and a predetermined number of times of the first step and the second step. By repeating the data of the welding voltage waveform and the data of the electrode expansion waveform and the desired output value corresponding to each welding state input to the multilayer neural network, the third step of determining the connection strength by offline learning sequentially Process and the data by the same process as the first step by performing welding It characterized by entering the multi-layer neural network having a connection weights determined during the check constituted by any process of performing the welding in the welding and inspection line.

1 is a schematic diagram of a spot welder.

2 is a voltage waveform diagram of a spot dish.

Fig. 3 Voltage waveform diagram for inspection of spot welding.

4 is an electrode expansion waveform.

5 is an electrode expansion waveform through a digital filter.

Figure 6 modified electrode expansion waveform.

Figure 7 Electrode expansion waveform for spot welding inspection.

8 is a schematic diagram of a welding state monitoring system using a neural network according to the present invention.

9A to 9C are block diagrams of a voltage waveform input system according to the present invention.

Fig. 10 (a) A schematic diagram of an electrode expansion waveform input system according to the present invention.

(b) A block diagram of an electrode expansion waveform input system according to the present invention.

11 is a flow chart for data input according to the present invention.

Figure 12 (a) Neural network structure consisting of a multilayer perceptron.

(b) Graph showing sigmoid function.

(c) Neural network structure according to the present invention.

Figure 13 is a flow chart of the neural network inspection system used in the present invention.

14 is a learning data according to an embodiment of the present invention.

Fig. 15 (a) A voltage waveform diagram according to a first embodiment of the present invention.

(b) Standard data according to the first embodiment of the present invention.

(c) Voltage waveform diagram for inspection according to the first embodiment of the present invention.

(d) Electrode expansion waveform for inspection according to the first embodiment of the present invention.

Fig. 16 (a) A voltage waveform diagram according to a second embodiment of the present invention.

(b) Standard data according to the second embodiment of the present invention.

(c) Voltage waveform diagram for inspection according to the second embodiment of the present invention.

(d) Electrode expansion waveform for inspection according to a second embodiment of the present invention.

Figure 17 (a) Voltage waveform diagram according to a third embodiment of the present invention.

(b) Standard data according to the third embodiment of the present invention.

(c) Voltage waveform diagram for inspection according to the third embodiment of the present invention.

(d) Electrode expansion waveform for inspection according to a third embodiment of the present invention.

* Explanation of symbols for main parts of the drawings

1: welding transformer 2: control unit

3a, 3b: electrode 4: pressing part

7: base material 10: base material

11 spot welder 11a first electrode

11b: second electrode 12: voltage waveform measurement system

13 noise reduction device 14 electrode expansion waveform measurement system

15,28: analog-to-digital converter 16a: neural network inspection system

16b: welding state monitoring system 17: computer system

21a: input layer 21: input node

23a: hidden layer 25a: output layer

27 non-contact optical sensor 29 digital filter

31: Board initialization step 33: Channel selection step

35: sampling time determination step 37: hardware selection step

39: computer memory storage step 41: data storage step

45: step of saving in the form of a file 55: neural network structure setting step

57,59: reading data 61: obtaining actual output

63: step of adjusting the connection strength 65: inspection step

71: Neural network initialization stage 73: Test preparation stage

75: preparing to read the waveform step 77: reading the input

79: output generation step 81: screen display

95: Waveform with excessive weld 97: Waveform with good weld

99: Insufficient weld waveform

Before describing the configuration of the present invention, the waveforms of the voltage applied to the both ends of the base material during welding by the spot welder, and the factors affecting the electrode expansion waveform and welding will be described.

The spot welding method is classified into a DC power supply using an inverter and a DC power supply. In the DC power system, the state of the welding can be determined by immediately reading the voltage value and examining the waveform.

2 is a voltage waveform diagram during welding in a spot welder driven by an AC power source. As an example, the voltage waveform of the welding machine driven by an AC power source has a sine wave of 60 hertz during welding.

Low voltage values during welding are controlled by a thyristor of the control unit so that a voltage is applied for several msec representing the maximum or minimum voltage of the sine wave because it does not greatly affect the welding.

The time for which the voltage is applied is allowed from several msec to several hundred msec by adjustment of the control unit having the thyristor.

In the figure, the magnitude of the voltage is 2.5V to 3.0V, and it can be seen that the voltage decreases little by little as the welding proceeds.

The analog-digital conversion time is changed in software according to the welding machine voltage application time, and the start time of the analog-digital conversion is started by the trigger of the input voltage.

For example, when the voltage application time is 150 msec and the sampling time is 0.1 msec, the number of data at this time is 1500 pieces.

However, it is difficult to review 1500 data in real time online, and only the maximum and minimum voltages are used for inspection because the data directly affecting spot welding is the peak voltage of each sine wave, that is, the maximum and minimum voltage. do.

FIG. 3 is a waveform diagram for inspection of spot welding, and is a voltage waveform diagram showing an absolute value by extracting only the maximum minimum value of each period of FIG.

If the voltage application time is 150msec, a sine wave of 60 hertz has 9 cycles for one spot welding, and 18 characteristic data can be obtained accordingly.

In order to obtain 18 data as shown in the figure, the analog signal during welding is converted into a digital signal, and all negative values of the digital signal are taken by the software in the computer system.

Then, the characteristic data of each half cycle is selected and the amount of characteristic data change is graphed.

As can be seen from the figure, since the welding between the base materials was not performed at the initial stage of welding, the contact resistance is high and the voltage value is the highest, but the resistance decreases as the welding proceeds, so the voltage value gradually decreases.

As shown in the figure, the voltage waveform showing the amount of change in the characteristic data of the voltage value when welding is normally performed is distinguished from the voltage waveform when over welding or under welding.

4 is an electrode expansion waveform showing the movement of the electrode during welding. The electrode expansion waveform uses a phenomenon in which the gap between the electrodes expands by initial welding and the electrode shrinks after melting.

The figure shows only the displacement of the gap by setting the interval of the initial welding to 0 (zero). The gap is increased by heating at the beginning of welding and shrinks after the medium is melted by heating.

The figure reads the output of the optical sensor system into a computer system through an analog-to-digital conversion by using an optical sensor system that outputs the measured interval in the form of voltage, indicating that the vibration and noise of the welder itself are severe. Can be.

5 is an electrode expansion waveform diagram in which vibration and noise of the welder itself are removed by a digital filter.

FIG. 6 is a diagram in which the waveform of FIG. 5 is changed into a smooth curve by a moving average method.

FIG. 7 is a diagram showing the overall contour of the waveform of FIG. 6 after the waveform of FIG. 6 is synchronized with the frequency of the voltage waveform. The waveform of FIG. 7 is used as an electrode expansion waveform for distinguishing welding states.

In the present invention, a voltage waveform for each welding state as shown in FIG. 3 and an electrode expansion waveform for each welding state as shown in FIG. 7 are used for the welding inspection.

On the other hand, the factors influencing the state of welding in electric resistance welding can be divided into the effect of the change of the current, the change of the pressing force, and the change of the cross-sectional area by the continuous use of the electrode.

The welding conditions are changed by the three conditions described above, which causes welding defects. The change of welding conditions according to the change of each condition is distinguished by the voltage waveform and the electrode expansion waveform applied to both ends of the base material during the welding time. do.

First, let's look at the effect of the change of current on the welding.

Currently, the spot welding machine is mainly driven by a constant current method, so the current has a fixed value during welding.

Therefore, as can be seen from the following equation (2), the change in voltage during welding is proportional to the change in resistance.

(E: voltage (V), I: current (a), R: resistance (Ω)

Although the driving method of most spot welders adopts a constant current method, the amount of current supplied can be changed by the operation error of the controller.

When all other conditions are constant, there is no change in the resistance of the welding object due to the change in current.

On the other hand, the amount of heat generated during the electric resistance welding is the same as the formula (1), but if you write this again, it is the same as the formula (3).

As can be seen from Equation (3), an increase in the current I causes excessive welding by generating a large amount of heat, and a decrease in the current causes an under welding by generating a small amount of heat.

Secondly, the influence of the change of the pressing force on the welding is as follows. Spot welding is performed by applying a current after applying pressure by a welding rod.

At this time, the pressing force is always required to maintain a constant level, but because of the physical quantity is difficult to control and there may be ups and downs over time.

This change in pressure has a close relationship with the contact resistance of the welding end, which is the most influential factor in the welding state.

Increasing the pressing force decreases the contact resistance between the two metals and decreasing the pressing force increases the contact resistance, so the pressing force and the contact resistance have an inverse relationship.

In addition, the pressing force is inversely proportional to the thickness of the welding target.

The resistance R can be expressed as follows.

(ρ: resistance coefficient, t: thickness, A: electrode tip contact area)

In the above formula (4), the increase in the thickness t results in an increase in the resistance.

An increase in resistance due to an increase in contact resistance and an increase in resistance due to an increase in thickness eventually lead to an increase in voltage.

Therefore, increasing the pressure (i.e., pressing force) applied to the electrode results in a decrease in resistance resulting in weak welding. On the contrary, a decrease in the pressure applied to the electrode increases resistance, resulting in excessive welding.

Finally, the effect of the change in cross-sectional area of the electrode, that is, the electrode tip contact area, will be examined.

In spot welding, a rod of a certain area is used at first, but the welding rod is gradually worn out by continuous welding, so that the cross-sectional area becomes wider than the beginning.

The increase in the cross-sectional area of the electrode is caused by two conflicting factors: the increase in contact resistance due to the decrease in pressure per unit area and the decrease in resistance due to the change in cross-sectional area.

The relation between the cross section of the electrode and the pressing force per unit area is as follows.

(F: Press force per unit area (force), P: Press force, A: Area of cross sectional area)

As can be seen from Equation (5), the area a of the cross-sectional area is inversely proportional to the change in the pressing force F per unit area.

On the other hand, the resistance (R) considering the contact resistance coefficient (ρ ') is as follows.

In the above formulas (6) and (7), the contact resistance coefficient ρ 'is proportional to the cross-sectional area (a), and serves to increase the total resistance (R). Therefore, the increase in the cross-sectional area leads to a decrease in the resistance value by the above equation (4).

Therefore, the change of the voltage value according to the cross-sectional area of the electrode shows that the effects are canceled by 1) the increase in contact resistance due to the decrease in pressure per unit area and 2) the decrease in resistance value due to the increase in the cross-sectional area.

However, the increase in resistance due to the decrease in pressure per unit area is indirect as shown in Eq. (6), whereas the decrease in resistance due to the increase in cross-sectional area is direct, so the decrease in resistance is actually larger.

However, in measuring the defects according to the change in the cross-sectional area of the electrode, there is a big difference between the voltage waveform when the welding is in a normal state and the voltage waveform when the welding is inferior, and thus it is difficult to actually distinguish it.

Thus, one embodiment of the present invention introduced an inspection method using an electrode expansion waveform to detect defects caused by changes in the cross-sectional area of the electrode.

The electrode expansion waveform utilizes a phenomenon in which the gap between the electrodes expands by initial welding and the electrode contracts after melting. If the welding rod cross-sectional area gets larger, the heating area becomes wider, so the expansion of the gap is fine, and there is no melting phenomenon, so no shrinkage phenomenon can be seen.

Using this principle, the relationship between the electrode expansion waveform and the area of the electrode can be established. Since the expansion of the gap between the electrodes is very fine, about 0.1 mm in the case of 1 mm iron plate, the change of the gap is measured using a non-contact optical sensor.

As described above, the state of spot welding changes according to the sudden change in welding current, the change in the pressing force of the electrode, and the change in the contact cross-section due to the continuous use of the electrode, which is the voltage waveform and the electrode expansion waveform at both ends of the base material. Can be distinguished by.

In the present invention, by adopting the learning method of the multi-neural neural network which can adapt to variously changed signals by various factors, it is possible to recognize the pattern for each welding state.

As such, by analyzing the voltage waveform and the electrode expansion waveform in the neural network using the neural network, the cause of the defect can be detected in real time, and the welding state can be easily determined in the industrial field.

Hereinafter, the present invention will be described in detail.

8 is a schematic diagram of a welding state monitoring system using a neural network. The spot welder 11, the voltage waveform measuring system 12, the electrode expansion waveform measuring system 14, the neural network inspection system 16a and the welding state monitoring system ( 16b).

Each output of the voltage waveform measurement system and the electrode expansion waveform measurement system undergoes a pattern inspection process by a neural network to indicate a state of welding and a cause of failure in the case of a welding defect. Hereinafter, each process will be described in detail.

9 (a) to 9 (c) are block diagrams when the voltage waveforms at both ends of the base material are used for welding inspection.

9 (a) is a block diagram according to an embodiment of the present invention, the base material 10 inserted between the first electrode 11a and the second electrode 11b connected to one side of a power source (not shown). A spot welder 11 comprising: an analog-digital converter for inputting a voltage across the base material, and a computer system 17 for inputting an output signal of the analog-digital converter 15.

In one embodiment of the present invention, the second electrode, which is the low potential point, is grounded.

The analog-to-digital converter 15 is activated by the computer system to convert the signal into a digital signal in order to process the voltage value, which is an analog signal, into the computer system.

In an embodiment of the present invention, Axom's Ax5611c-L High Performance DAC Carrier Board was used as the analog-to-digital converter 15.

The board system has a maximum of 16 inputs and can perform sampling of 1 M㎑ by a direct memory access (DMA) method.

However, in the exemplary embodiment of the present invention, one channel was used to read a voltage waveform corresponding to one welding point, and the sampling time was 10 ms through DMA.

In addition, the computer system 17 used in the embodiment of the present invention is a neural network inspection system that inspects a welding state by processing a voltage signal at the time of welding, and a welding state by analyzing the cause of the welding state and the welding failure. It has a surveillance system.

In order to process the voltage signal during welding, all the digitally processed signals after the analog-digital change are subjected to the software. The negative value is taken to take the absolute value and only the minimum and minimum values are provided.

The neural network inspection system stores the standard voltage waveform by adjusting the connection strength of the neural network by a predetermined learning method.

Thus, the weld inspection is carried out according to the stored connection strength.

9 (b) is a block diagram according to another embodiment of the present invention, and may further include a noise canceling device 13 between the spot welder 11 and the analog-to-digital converter 15.

The noise canceling device 13 inputs a voltage across the base material, and in one embodiment of the present invention, the second electrode 11b, which is a low potential point, is charged.

In fact, the spot welder not only generates a lot of high frequency noise, but also generates a lot of noise in the factory environment.

In addition, since the shape of the voltage waveform supplied is about 60 Hz, in one embodiment of the present invention, the low frequency band pass filter is used as the noise removing device 13 to remove the high frequency band noise.

However, in another embodiment of the present invention, an appropriate noise canceling device may be used depending on the type of voltage waveform and the type of noise supplied.

FIG. 9 (c) shows a block diagram of FIG. 9 (b).

10 (a) to 10 (b) are blocks in the case where the electrode expansion waveforms at both ends of the base material are used for welding inspection.

FIG. 10 (a) is a schematic diagram of an electrode expansion waveform measuring system showing that a non-contact optical sensor 27 is provided to measure the displacement of the gap during welding.

FIG. 10 (b) shows another embodiment of the electrode expansion waveform measuring system, in which a non-contact optical sensor 27 for measuring expansion of a gap between electrodes and an analog-digital input for inputting the output of the optical sensor 27 are shown. The converter 28 and the digital filter 29 which inputs the output of the said converter 28 are provided.

In an embodiment of the present invention, a low-cost optical sensor is used, and since only displacement is measured in a non-contact manner, measurement is possible without prior information on a reference interval.

Most of the studies in the present experiments are not precise due to the use of contact sensors, but also require advance information on the reference spacing of the electrodes.

This reference distance can be set in the laboratory, but in practice it is not possible because the parameters change due to the continuous use of the electrode. In addition, the distance measurement using the laser can be precisely measured, but the measuring device of the micrometer unit is very expensive.

The measured waveform of electrode expansion is used as input of neural network to learn neural network.

The electrode expansion waveform continuously reads the voltage generated by the reflection type optical sensor spacing system consisting of the optical sensor 27 of FIGS. 10A to 10B using the channel of the analog-to-digital converter 28. It is configured by lifting.

In the embodiment of the present invention, PT-165 optical sensor system of Japan Keyence was used as the reflective optical sensor gap measurement system.

The measuring range of the sensor is 2.5 cm and the precision is 5 micro meters. At this time, the output is linear from -5V to + 5V.

The electrode expansion waveform of reading the output signal of the sensor into the analog-to-digital converter 28 is shown in FIG.

The electrode expansion waveform shown in FIG. 4 shows noise due to vibration of the welding machine itself, and the signal itself also shows irregular vibration.

In order to eliminate such vibration, a digital filter 29 is placed between the analog-digital converter 28 and the computer system 17 which input the output of the optical sensor as shown in FIG. 10 (b). The electrode expansion waveform of is as shown in FIG.

The digital filter 29 is for removing sudden shaking of the welding machine. The filtered waveform is transformed into a smoother curve by a moving average method, in which the waveform is as shown in FIG.

Since the data read from the modified electrode expansion waveform as described above is a large amount of data to be processed in real time using a neural network, it is sampled at a time corresponding to one wavelength of the welding frequency. The sampled electrode expansion waveform is as shown in FIG.

In addition, the reading time of the entire data was adjusted to twice the welding time to confirm the shrinkage process. Therefore, in case of 6 cycles, 12 data are read to examine the characteristics of electrode expansion waveform.

As shown in the drawing, by connecting the voltage waveform measurement system and the electrode expansion waveform measurement system to the spot welder, the voltage change and the displacement of the gap applied to both ends of the base material at the time of welding are processed by the computer system with the neural network. Welding method is available. Online means the change of voltage and gap during welding is processed by computer system simultaneously with welding.

Hereinafter, a description will be given of the software for activating the analog-digital board and the neural network which checks the welding by inputting the signal processed by the software, which is built in the computer system.

11 is a flowchart for data input according to the present invention, which is a flowchart for operating hardware of an analog-digital board.

Referring to the flowchart, in order to use the analog-digital board, first, the step of initializing the board (31), the step (33) of determining the channel to be read from the board, the step of determining the sampling time (35), Selecting the hardware so that the analog-to-digital conversion starts only when the voltage is above a predetermined value, storing the signal converted into a digital signal into a computer memory (39), and selecting predetermined data. (41) and the step (45) of storing the signal selected in the above process in the form of a file for use in the learning process of the neural network.

In the data selection step 41, predetermined data required for the voltage waveform measurement system or the electrode expansion waveform measurement system are selected, respectively. That is, in the case of the voltage waveform measuring system, the negative value among the signals converted in the above process takes only the absolute value and selects only the welding voltage waveform which affects the welding. In the case of the electrode expansion waveform measuring system, the signal converted in the above process Using digital filter and moving average method, noise and vibration are eliminated and electrode expansion waveform data is obtained by synchronizing with frequency of voltage waveform.

In the channel selection step 33, in the case of the present invention, one channel is selected, but in the case of the AX 5611C board used as an embodiment of the present invention, up to 16 channels can be input.

In the step 35 of determining the sampling time, the sampling time can be selected up to 1 MHz. However, if the sampling time is too small, the accuracy of the data is increased, but it is difficult to process a lot of data. Therefore, in the exemplary embodiment of the present invention, a sampling time of 0.1 msec (10 ms) is selected for the voltage waveform measuring system, and a sampling time corresponding to one wavelength of the welding frequency is selected for the electrode expansion waveform measuring system.

On the other hand, since the method of continuously reading signals not only uses the memory of the computer but also interferes with performing other tasks, the analog-to-digital conversion is started above a predetermined voltage as in step 37.

In one embodiment of the present invention, an input trigger of 0,7V is used. The analog-digital board is set through the steps (31), (33), (35) and (37) above. Next, in step 39, the analog-digital board is operated using the direct memory access (DMA) method for fast data conversion.

FIG. 12 (a) shows a neural network composed of multilayered perceptrons as a neural network structure for performing learning and online inspection.

The multi-layer perceptron model is composed of an input layer 21a, a hidden layer 23a, and an output layer 25a to solve a problem that cannot be separated linearly. As a learning method applicable to the multi-layer perceptron, PDP such as Rumelheart in 1986 A backpropagation learning algorithm was developed by the group.

Here, the question of how many hidden layers should be made and how many treatment units of each hidden layer should be appropriate should be repeated to find a suitable structure for each application.

Hereinafter, a reverse propagation algorithm capable of learning the multilayer perceptron as described above will be described.

Backpropagation learning algorithm is designed to minimize the error between neural network's actual output value and our desired output value by iterative slope reduction method.

Differential nonlinear functions are used here as output functions. In most cases, the sigmoid function is

12 (b) is a graph showing the sigmoid function.

A simple backpropagation algorithm is as follows.

Initialize the connection weight. Then we give the input pattern and the output we want. Thus, the actual output value is calculated.

The connection strength is minimized to minimize the overall error between the calculated actual output value and the desired output value.

The strength of the connection is adjusted by the following equation.

In Equation (9), the connection strength between the i and j layers immediately before Wij (n), and Wij (n + 1) refers to the new connection strength between i and j.

ΔWij is the amount of change by learning.

The process of deriving the equation is as follows.

In Equation (10), Net j is a value obtained by multiplying the outputs of the lower layer by the connection strength, and Oj is the output of the j-th neuron, and when j is an input, Oj = Ij.

The output Oj is expressed as

Where f (Net j) is the sigmoid function

Θj in the above Equation 12 is a threshold.

On the other hand, δ i of the output layer is expressed by the following equation.

In the above formula (13)

to be.

Tj is the output we want at the jth nerve, and Oj is the calculated actual output at the jth nerve.

In addition, δj of the hidden layer is expressed by the following equation.

Where m is the number of processing units above the layer with j.

Thus, the connecting strength is adjusted by the above formula (9).

η is defined as '0η1' as the learning rate. The larger η is, the larger the change in connection strength is.

However, if η is greater than 0.75 degrees, the whole neural network becomes unstable, so the value between 0 and 1η 0.75 is generally taken.

Within this range, the larger the value of η, the faster the learning rate.

If the total error between the actual output value and the desired output value is minimized by adjusting the connection strength in the same way as above, the input pattern and the desired output value are given together to learn another pattern.

In this case, a moment term term or a bias term may be added to speed up convergence.

Figure 12 (c) shows a neural network structure according to the present invention.

As in the embodiment of the present invention, when a 60 Hz sine wave is applied for 150 msec, the number of data for input is 18, so 20 input nodes 21b are set with two margins. The node of the hidden layer 23b is set to twice the input node by experience.

In the output 25b, two nodes were set to over weld for '11', under weld for '00' and normal weld for '10'.

In one embodiment of the present invention, three states are set, but in another embodiment, the number of states may be increased or decreased as necessary.

In addition, in preparation for the change of the number of input nodes according to the change of the setting of the initial condition of the spot welder, the software system can be configured to change the number of input nodes, hidden nodes and output nodes.

The neural network inspection system uses a waveform inspector for classifying a pattern of a welding voltage waveform using the voltage waveform of FIG. 3 as an input or a voltage inspector of FIG. 7 for inputting a pattern of an electrode expansion waveform. Used for waveform inspector

13 is a flowchart of a neural network inspection system used in the present invention.

The neural network inspection system is composed of a learning portion 51 for learning the welding state, and an output generation portion 53 for the actual inspection after the learning is completed.

The selection of the learning part and the output generation part is made to be selected in the program software, and it is designed using pointers to select the number of input and output layers to be extended to the learning and inspection of other patterns. The learning method used in the embodiment of the present invention is a back propagation method.

First, the learning part 51 will be described.

The first step is to set the structure of the neural network (55). The number of input and output nodes and the number of nodes of the hidden layer can be set in the program, and the initial learning rate (η) and the momentum rate for learning are also set. Set in this step

After initializing the data and the connection strength, steps 57 and 59 are performed to read data for learning. At this stage, you can choose to start an entirely new learning, or you can choose to continue learning by adding to the previously learned connection strength.

Data is read from the file stored in step 45 of FIG.

Next, to obtain the actual output of the multilayer perceptron to perform the neural network learning (61), using the data file generated in step (45) of FIG. Obtain

The next step is to adjust the connection strength 63 using the error between the calculated output value obtained in step 61 and the desired output value. The improvement of the connection strength is continuously performed, and the step 65 of checking whether the error falls within a predetermined range is interrupted if the error falls within the predetermined range.

After learning, the improved connection strength values should be stored for use in the output generation process.

The values of the stored data are saved to a file in step 67 for permanent storage.

According to the learning process, it is possible to recognize patterns of standard voltage waveforms for each of normal welding, over welding, and under welding. The learning process is done off-line here.

Here, offline means a neural network learning method that determines the connection strength by obtaining data on each case such as normal state, over-welding and under-welding by performing welding inspection by direct fracture method after every welding. At this time, the voltage waveform in each case is stored in the computer at the same time as the welding in accordance with the procedure shown in FIG.

The learning data of neural network increases or decreases the overcurrent from the small current, and the pressure and cross-sectional area of the electrode, and collects the voltage waveforms through continuous experiments. Of the collected voltage waveforms, the range of welding is appropriately determined to determine the area of the standard waveform, and the types of over welding and under welding are determined to study neural networks.

Next, the output generating portion 53 will be described.

Initialization step 71 of the neural network uses data on the structure of the neural network stored in the learning process.

Next, in step 73, the values of the connection strengths improved in the learning process are read from the stored file and prepared for examination by the neural network.

Next, in step 75, the analog-digital board is activated to prepare to read the waveform of the digitized voltage according to the same procedure as the flowchart shown in FIG.

The next step is to read the input through the active analog-digital board (77), where the read data consists only of the maximum and minimum voltage values.

Then, a step 79 of generating the actual calculated output by the voltage signal data processed by the above process is compared with the calculated output value and the desired output value by learning to display the welding state and the cause on the screen (81). In case of failure, the inspection is terminated. In case of failure, the inspection returns to step 77 to continuously receive the input signal and continue the inspection.

Since the output portion is performed at the same time as the welding proceeds, the inspection is performed at the same time as the welding.

When the inspection process is set in the same condition as the learning process, only the inspection process may be repeatedly performed with the connection strength determined by the learning process without repeating the learning process anew.

FIG. 14 is learning data according to an embodiment of the present invention and is for offline learning of neural networks using the backpropagation method described in FIG.

Data sets 1, 2, and 3 are cases where welding is normal, data sets 4, 5, and 6 are under welded, and data sets 7, 8 and 9 are over welded.

For each data set inputted at the same time as welding, an output node value indicating the welding state is determined.

In one embodiment of the present invention, '10' for data sets 1, 2, and 3 when welding is normal, '00' for data sets 4, 5, and 6 when under welding, and over welding. The data set of 7, 8, and 9 was set as '11'.

Thus, each data set is input to the neural network and the connection strength is adjusted by comparing the corresponding output with our desired output ('11' '00' or '01').

The learning rate of 0.51 and the momentum rate of 0.5 showed a learning error of 0.001 at 1,500 times. The learning time of the 486DX-2 computer was about 6 minutes.

Values such as learning rate, momentum rate, etc. given in the neural network learning are the same in other embodiments of the present invention described below.

The inspection process of the learned system is online real time, and the inspection time for one welding spot is about 20 mses, which is enough time for real time inspection.

The trained neural network adopts an analog-to-digital conversion system to continuously receive the waveform of voltage and analyzes the waveform by computer.

Fig. 15 (a) is an AC voltage waveform diagram according to the first embodiment of the present invention. The type of the welding machine is a 19V-SCF spot welding machine for welding a TV CRT frame produced by Orion Electric Co., Ltd., and the setting of the welding machine is 6-3. It was -10 cycles, and the base metal used was a 1.2t SPC1 frame produced by Toshiba's 1.0t bimetal and Pohang Steel Coated on the surface.

When the waveform of the voltage applied as shown in the figure is composed of three alternating current waveforms in series, there is a case where there is a foreign matter by the coating or the like on the surface of the base material.

That is, the first half corresponding to 6 cycles is a preliminary current supplied before the welding current to burn off the coating area of the welded part when the base metal is coated, and the actual welding current is 10 cycles after passing through the cooler corresponding to 3 cycles. Supplied throughout.

Here, the voltage waveform corresponding to the above 10 cycles is used for the welding inspection.

FIG. 15 (b) is standard data according to the first embodiment of the present invention, and shows the effect of the current change on the welding.

The standard data shown in the figure was learned offline by a backpropagation learning algorithm according to the flowchart of the learning part 51 of FIG.

In the table, data sets 1 and 2 are in a normal state, and data sets 3, 4 and 5 are in an under welded state.

In the embodiment of the present invention, the current in the steady state is 4100A, and in the case of under welding, the current is 3000A.

FIG. 15 (c) is a voltage waveform diagram for inspection according to the first embodiment of the present invention, and shows the data shown in FIG. 15 (b) as a graph.

In the figure, waveform 85 represents a steady state and waveform 87 represents a low current state.

FIG. 15 (d) is an electrode expansion waveform diagram for inspection according to the first embodiment of the present invention, where waveform 89 is a low current (3000 (a)), waveform 91 is a steady current (4100 (a)), and waveform 93 Shows electrode expansion waveforms in the case of overcurrent 5000 (a), respectively.

Figure 16 (a) is an AC voltage waveform diagram according to a second embodiment of the present invention, the type of the welding machine is Orion Electric Co., Ltd. SP5361-CH spot welding machine, the welding machine setting was 9 cycles, the base material Toshiba (Toshiba) This is the case when 0,76t Bimetal Co., Ltd. and 0,65tSVS 304 Frame of Poongsan Metal Co., Ltd. are used.

16 (b) is standard data according to the second embodiment of the present invention.

In the figure, data sets 1 and 2 are cases where welding is well performed by a normal pressure, data sets 3 and 4 are under welded by pressure, and data sets 5 and 6 are over welded by decompression. If it is.

In one embodiment of the present invention, the normal pressure is 4 kg / m 2, the pressurized state is 5 kg / m 2, and the reduced pressure is 3 kg / m 2.

As described above, the data presented in the drawings are performed by the welding inspection in a direct fracture method every time the welding is completed to extract samples for the steady state, over welded state and under welded state.

At this time, the voltage waveform at the time of welding for each state is stored in the computer at the same time as welding.

Here, the voltage waveform corresponding to each sample becomes the output value we want when learning by neural network.

Thus, as shown in the learning part 51 of FIG. 13, the calculated output value is obtained by reading the voltage values for each state, and then compared with the desired output value, and the connection strength is continuously adjusted until the error is within a predetermined range. do.

FIG. 16 (c) is a voltage waveform diagram for tilting according to the second embodiment of the present invention. The data shown in FIG. 16 (b) is a graph.

As can be seen from the figure, the waveform 97 in the case of good welding shows a form in which the voltage is high and gradually decreases at the beginning of the welding.

In the case where the welding is underestimated, the waveform 99 has a low slope and shows a low voltage without large change. The waveform 95 in the case where the welding is excessive has a larger slope compared to the case where the welding is good, so that the voltage is overall. It is high.

FIG. 16 (d) is an electrode expansion waveform for inspection according to the second embodiment of the present invention, where waveform 101 is the normal pressure and waveform 103 is the electrode expansion waveform in the down state, respectively.

Figure 17 (a) is an AC voltage waveform diagram according to another embodiment of the present invention, the type of the welding machine is Orion Electric Co., Ltd. 20V CF spot welding machine, the welding machine setting is 12-4-12 cycles, the base material is Toshiba This is the case when the 1.0t bimetal of the company and the 0.8t SPC1 frame of Pohang Steel Co., Ltd. are used. Only 12 cycles of the latter half of the 12-4-12 cycles correspond to the actual welding current.

Figure 17 (b) is the standard data according to another embodiment of the present invention.

In the table, the data sets 1 and 2 are in a steady state. The diameter of the electrode tip is 5 mm (Φ 5) and the current is 3700-4100A.

The data sets 3 and 4 are in a low current state where the current is 2500 A, and the data sets 5 and 6 are in the case where the diameter of the electrode tip is widened, and the diameter is 6 mm (Φ 6). In addition, the data sets of 7, 8 are in the case of an overcurrent state where the current is 5500A or more.

FIG. 17 (c) is a voltage waveform diagram for inspection according to another embodiment of the present invention. The data shown in FIG. 17 (b) is a graph.

In the figure, waveform 105 represents an overcurrent state, waveform 107 represents a steady state, and waveform 109 represents a low current state when the cross-sectional area of the electrode tip increases.

FIG. 17 (d) is an electrode dispersion waveform diagram for inspection according to another embodiment of the present invention, where waveform 113 is a steady state and waveform 115 is an electrode expansion waveform when the electrode is widened.

As can be seen from above, the state of welding closes according to the initial condition of the welding machine and the sudden change of the welding machine itself. In particular, sudden changes in conditions, such as a change in current, a change in pressing force, and a change in the cross-sectional area due to the continuous use of the electrode, are the main causes of failure of the welding state.

Here, the change in the welding state according to the change of the current and the pressing force can be easily distinguished by the voltage waveform applied to both ends of the base material during welding.

However, it is difficult to distinguish the shape of the voltage waveform by increasing the cross-sectional area of the electrode due to the offsetting effect of two factors: the increase in contact resistance according to the pressure change per unit area and the decrease in resistance due to the increase in cross-sectional area.

On the other hand, the electrode expansion waveform has a good characteristic for discriminating whether the welding state is good, but if it is defective, the characteristic curve is simple to determine the cause.

Therefore, by simultaneously analyzing the voltage waveform and the electrode expansion waveform, the welding inspection apparatus that can be corrected immediately by presenting the cause of the defect as well as the good or bad of the welding was implemented.

Analyzing the output of the neural network including the voltage waveform measuring system and the electrode expansion waveform measuring system as shown in FIG.

Here, when the output of each neural network whose data is the output value of the voltage waveform measurement system and the output value of the electrode expansion waveform measurement system is '11', the over welding state is '10', the '10' is normal welding state, and '00' is under welding state It was set as.

In the embodiment of the present invention, the case of the spot welding machine has been described as an example, but those skilled in the art will appreciate that the present invention can be applied to other welding methods for welding using electric resistance heat in addition to the spot welding. .

As described above, the present invention focuses on the reduction in voltage as the contact resistance of the load side gradually decreases in welding and the expansion and contraction of the gap between electrodes during welding, and thus the welding is normally performed. An on-line nondestructive testing device and inspection method have been developed that can check the quality of welded state by the difference of voltage waveform and electrode expansion waveform.

Accordingly, the direct inspection method to check the abnormality at the time of welding can not only check whether each welding point is defective, but also indicate the cause of the defect.

In addition, since the non-contact sensor is used as the measurement method of the electrode expansion waveform, the measurement accuracy can be secured, and there is an advantage that the industry can measure the change of the electrode length without continuous change of the reference distance due to the continuous use of the electrode.

Accordingly, it is possible to secure the robustness of the product by significantly reducing the probability of welding failure, and it can be applied to various types and settings of welding machines because the inspection is performed through the learning of neural network.

This system can be inspected automatically by computer in real time online to prevent waste of human and material resources, greatly improve productivity and reduce product cost, and have been raised as a problem in the automation of current process. Improving the inspection method can also be a great help in the automation of the process.

In addition, since the waveform inspection method is performed by the neural network learning, it is easy to apply without an expert in the industry by saving only the initial value of the software without requiring a separate operation in other types of spot welding machines.

Claims (7)

An electric resistance welder having a first electrode connected to a high potential of a power source, a second electrode connected to a low potential of the power source, and a base material inserted between the first electrode and the second electrode; Noise canceling means comprising a low frequency bandpass filter for inputting the voltage applied across the base material, and voltage waveform measurement outputting a voltage waveform measurement result using the output of the noise removing means as an input of the first analog-to-digital converter. system; An electrode expansion waveform measurement system for outputting an electrode expansion waveform measurement result by using a sensor means for detecting a displacement between the first electrode and the second electrode during welding and an output of the sensor means as an input of a second analog-to-digital converter; And a computer system incorporating a neural network inspection system for inputting the voltage waveform measurement system and the output of the electrode expansion waveform measurement system. The on-line non-destructive of electric resistance welding according to claim 1, wherein the neural network is composed of a multilayer neural network composed of at least one hidden layer between an input layer and an output layer, and the learning of the neural network is performed offline. Inspection device. The on-line non-destructive inspection device for electric resistance welding according to claim 1, wherein the sensor means is a non-contact optical sensor. The computer system of claim 1, wherein the computer system comprises: first software for detecting characteristic data of signals digitized by the first analog-to-digital converter; And second software for digitizing by the second analog-to-digital converter and then removing the noise and vibration by using the digital filter and then obtaining the characteristic data. In the non-destructive testing method of electric resistance welding, the offline learning process includes: while performing welding, the voltage and electrode expansion waveforms at both ends of the base material at the time of welding are input to the computer system through an analog-to-digital converter, respectively. Detecting a welding voltage waveform and a modified electrode expansion waveform of voltage among the signals digitized by the digital converter and storing them in a computer system; Performing a welding inspection by the post-welding fracture inspection method; Repeat step 1 and step 2 to input data representing the welding voltage waveform and the electrode expansion waveform and the desired output value corresponding to the welding state to the multilayer neural network, and then determine the connection strength by using the back propagation algorithm. In three steps; The inspection process performs welding and inputs the data according to the same process as the first step into the multilayer neural network having the connection strength determined in the learning process, and examines the welding strength online with the welding, and the same as the learning process. On-line non-destructive testing method for electric resistance welding, characterized in that when the condition is set in the condition, it can be repeatedly tested with the connection strength determined by the learning process. The on-line non-destructive testing device for electric resistance welding according to claim 1, wherein the neural network adapts through learning to welding conditions such as welding time, welder type, weld thickness, and the like. The on-line non-destructive inspection device for electric resistance welding according to claim 6, wherein the learning uses data stored in the computer, and is specific pattern data extracted from a sample welding of a fracture test using the electric resistance welding machine.