KR20020051286A - Flash butt weld estimation method using multi-layer error back - Google Patents

Abstract

PURPOSE: A method is provided which judges welding quality by predicting tensile strength using a multi-layer error backpropagation neural net. CONSTITUTION: The method for judging welding quality comprises the steps of providing with a neural network which is consisted of an input layer into which heat input measured from flash butt welding process, upset pressing force, acceleration of moving stand and measured value of flash lightness are inputted, an output layer from which tensile strength of a material welded in the process is outputted, and first and second concealment layers arranged between the two layers, wherein each of the layers are connected by connection weight which is a correlation between input and output; calculating an error between tensile strength which is an output from the neural network and an actual tensile strength which is a target value; correcting the connection weight of the neural network in a direction of reducing the error; and predicting tensile strength by inputting heat input, upset pressing force, acceleration of moving stand and measured value of flash lightness of a material a welding quality of which is judged into the input layer of the neural network according to the corrected connection weight.

Description

FLASH BUTT WELD ESTIMATION METHOD USING MULTI-LAYER ERROR BACK}

The present invention relates to a flash butt welding quality determination method using a multilayer error back propagation neural network, and more particularly, to a method for determining welding quality by predicting tensile strength using a multilayer error back propagation neural network.

In general, flash butt welding refers to a typical butt welding method of resistance welding. Such flash butt welding is a welding method using resistance heat generation due to high current density short-circuit current by generating local contact portions by adjoining sections for welding at an appropriate speed. Specifically, the flash butt welding, as shown in Figure 1, in the pickling and cold rolling line to fix the end portion 6 of the preceding strip (S) to the holder 12 and the front end portion 5 of the trailing strip (T) moving table ( 14), after the end of the preceding strip and the leading end of the trailing strip by shearing at the beginning of welding, a certain distance (G) is formed, and then energizing both ends of the strip while moving the movable table toward the side of the stationary base. do. At this time, both ends of the strip are ruptured by melting to generate an arc and the vicinity thereof is heated. The molten metal is scattered by this arc, which is called flash. When the flash is generated continuously and the welding cross section is uniformly heated, the movable stand side is pushed forward to the stationary side and pressurized to form upsets (also referred to as protrusions) 7 at both ends.

In such flash butt welding, defects may be deteriorated in welding quality. The defects of the flash butt welding can be largely divided into metallic defects and mechanical defects. Metallic defects include infiltration of inclusions, softening of heat-affected zones, and cracks, and mechanical defects include insufficient weld strength, misalignment, and poor trimming. There may be.

Detecting the defect of the flash butt welding and determining the welding quality is an important basis for improving the actual process. Conventionally, US Pat. No. 4,484,057 discloses a method for determining welding quality to detect welding defects, and to compare voltage, current, and clamp transfer amount, which are process variables during flash butt welding, with comparison with an optimum welding method. That is, if the process variable is within the upper and lower limits of the reference value, it is a method that can determine that the welding quality is appropriate. In addition, US Patent No. 5,439, 157 discloses a method for determining the presence or absence of weld cracks and the integrity of a weld by using two EMAT sensors for generating ultrasonic waves after flash butt welding.

However, the determination method using the process variable disclosed in US Pat. No. 4,484,057 is generally related to metal defects, which can be solved when setting welding conditions. An important object in the welding quality determination method is mechanical defects. In other words, welding defects that occur during actual operation are mostly caused by misalignment of the welding target material. Therefore, the above conventional technique is not suitable as an approach for determining welding quality.

In addition, the determination method using the EMAT disclosed in the above-mentioned US Patent No. 5,439,157 is less effective in practical line application. That is, the site where the actual line is present, it is difficult to use the ultrasonic wave due to the occurrence of noise, vibration, electric noise, etc., and there is a problem in that it is impossible to detect other kinds of defects other than cracks.

In order to solve this problem, there is a need for a welding quality determination method that can be applied to actual lines for mechanical defects. The present invention pays attention to the fact that misalignment has the greatest influence on tensile strength, and has been studying a new method for determining weld quality using the same.

In order to overcome the above problems, an object of the present invention is to provide a method for determining the weld quality by predicting the tensile strength using a multi-layer error back propagation neural network.

1 is a schematic diagram showing step by step a flash welding process.

2A to D show experimental results of correlation between input data and tensile strength employed in the present invention.

3 is a block diagram of a multilayer error backpropagation neural network according to the present invention.

4 is a step explanatory diagram illustrating neural network learning and execution according to the present invention.

5 is a graph showing the relationship between the tensile strength predicted by the present invention and the actual tensile strength.

<Code Description of Main Parts of Drawing>

20: input data

30: input layer

40,50: first and second hidden layers

60: output layer

35,45,55: strength of connection

The present invention is an input layer to input measured heat input, upset pressing force, acceleration of the mobile platform and the amount of flash light measured from the flash butt welding process, an output layer outputting the tensile strength of the material welded in the process, arranged between the two layers Comprising a first and second concealed layers, each of which is provided between the layers connected by the connection strength of the correlation between the input and output, the tensile strength of the output value from the neural network and the error of the actual tensile strength which is the target value Calculating the connection strength of the neural network in a direction of reducing the error, the heat input amount of the material to determine the welding quality of the neural network according to the modified connection strength, the upset pressing force, It includes the step of predicting the tensile strength by inputting the measurement of the acceleration of the mobile platform and the flash light amount.

In one embodiment of the present invention, the step of providing a neural network further includes a step of presetting a tolerance range between predicted tensile strength, which is the output value of the neural network, and actual tensile strength, which is a target value, and calculating the error. The method may further include determining whether the calculated error corresponds to the tolerance range, and correcting the connection strength comprises modifying the connection strength until the calculated error corresponds to the tolerance range. It may be made of.

This plays a role of limiting the appropriate number of times of learning by setting the tolerance range in estimating the tensile strength using the learned neural network.

Furthermore, the present invention preferably applies the measurement value of the flash light amount to the ratio of the FFT coefficient of 120 Hz in which the flash light amount is converted and the FFT coefficient of 240 Hz in which the coefficient and the flash light amount are converted.

In order to accurately predict the tensile strength as a process variable of the tensile strength as much as possible, the FFT coefficient of 120 Hz in which the flash light amount was converted as described above and the 240 Hz FFT in which the coefficient and the flash light amount were converted It is preferable to set it as ratio of coefficients.

As described above, the welding quality can be easily determined by predicting the tensile strength using the multilayer error backpropagation neural network. The FFT coefficients of heat input, upset pressing force, acceleration of moving platform, and flash light amount input to the input layer of these neural networks are selected by repeating the correlation test between welding process variable and tensile strength.

It can be seen that the tensile strength and correlation with respect to the selected process variable. The welding process has a difficulty in that quantitative analysis results are not accurately calculated according to the process variables. Therefore, in order to determine the welding quality, it is impossible to select an accurate process variable capable of predicting the tensile strength, which is a strong determination factor of the welding quality, without repeated experiments. Accordingly, the present inventors have endeavored to provide a method for determining welding quality by selecting process variables correlated with tensile strength and predicting the tensile strength through numerous experiments.

As a result, the basis for predicting the tensile strength from the measured values of the heat input amount, the upset pressing force, the acceleration of the moving table, and the flash light amount was prepared.

Hereinafter, the present invention will be described in more detail with reference to the drawings.

Referring to Fig. 2, the correlation between the input data and the tensile strength employed in the present invention is shown graphically. It can be seen from FIG. 2A that the heat input consumed in the flash butt welding process is roughly proportional to the tensile strength. Fig. 2B shows the relationship between the upset pressing force and the tensile strength in the flash butt welding process. It can be seen that there is an irregular change depending on the upset pressing force. Figure 2 (C) shows the relationship with the tensile strength according to the acceleration of the moving table. Through this, it can be seen that the variable affects the tensile strength according to the movement acceleration. In addition, Figure 2 (D) can be predicted the correlation between the flash light quantity and the defect length through the signal analysis results according to the moving table acceleration of the flash butt welding process and the change result of the defect length during welding. It can be seen that the amount of flash light affects the tensile strength, as it affects the change of defect length by the amount of flash light.

Hereinafter, a neural network applied in the present invention will be described.

The neural network applied in the present invention is composed of a connection weight connected to a plurality of inputs, a node that is a neural unit, and one output. The signal output from the neural unit may represent the following relationship.

Here, f (x) is an activation function used to perform a nonlinear operation using the sum of the connection strength and the input value product at the node. In general, it has a relationship in the form of the sigmoid below.

3 is a block diagram of an error back propagation multilayer neural network employed in the present invention. Referring to FIG. 3, the error back propagation multilayer neural network has a structure consisting of an input layer, two hidden layers, and an output layer. The hidden and output layers are composed of nodes and connection strengths that are output according to the active function. The process of receiving input variables from the input layer and generating output values in the final output layer proceeds in a feed forward. The process of generating the output of the hidden layer by using the output of the input layer and the connection strength, and processing the output of the hidden layer and the corresponding connection strength to produce the output of the result layer is as follows.

Here, z is an input vector of the input layer, and y, x, o are output vectors of the hidden layer and the output layer. v, u, w represents the arrangement of the connection strength which is the relationship of each layer, j, t represents the number of nodes in the first and second hidden layers.

The error level of the target layer is determined by calculating the error of the target value from the output value of the connection strength process. Next, the error back propagation learning algorithm for correcting the connection strength is adopted to train the connection strength of the error back propagation multilayer neural network to the learning pattern.

The least square method is used for the error calculation. Where P is the number of learning patterns.

In describing the process of learning the connection strength, the following equation is used to correct the sensitivity of the connection strength, that is, the opposite direction of the change in the connection strength minute change compared to the error change amount, thereby learning in the direction of decreasing the error. That's the learning rate: Is controlled by

In this way, the connection strength is modified repeatedly so that the error is lowered to an acceptable level using the learning pattern.

Hereinafter, a welding quality determination method for predicting tensile strength using the multilayer error back propagation neural network as described above will be described.

3 shows an example of a neural network applied to the present invention. The neural network applied to the present invention includes an input data (20) consisting of an FFT coefficient at 120 Hz and a ratio at 240 Hz (240 Hz / 120 Hz) in which heat input, upset force, moving stage acceleration, and flash light amount are converted. ) Is composed of an input layer (30) for inputting the first and second hidden layers (40, 50) and an output layer (60). In addition, the neural network includes connection strengths 35, 45, and 55 which are input / output relationships between the layers. The input data 20 of the target material obtained through a plurality of welding processes may be input to the input layer to obtain the output data 70 from the output layer 60 by the preset connection strength. Then, the error between the obtained output data and the actual tensile strength which is the learning target is calculated and the preset connection strength is corrected in the direction of reducing the error. By learning the neural network several times in the same process as described above it can provide a neural network that can effectively determine the weld quality by predicting the tensile strength.

4 is an explanatory diagram of a neural network learning and execution step according to the present invention.

First, the neural network learning step is performed as described above. The learning step calculates the error between the output value (tensile strength by neural network before learning) and the target value (actual tensile strength) from the output layer 60 as shown in FIG. 4, and if it is not within the tolerance range, As described above, the learning step of correcting the connection strength is performed. By repeating the process continuously through the data of different materials, when the error between the output value and the target value is within the tolerance range, the learned connection strength is stored in the database. Subsequently, when it is necessary to judge the quality of the welded portion of the material obtained by the flash butt welding process, the FFT coefficient of the heat input amount, the pressing force, the acceleration of the moving table, and the flash light amount constituting the process conditions applied to the material (120㎐ ) And the FFT coefficient ratio (240 ㎐ / 120 ㎐) of the flash light amount can be easily predicted. Therefore, as described above, the welding quality can be easily determined by evaluating the tensile strength resulting from misalignment.

As described above, the neural network was configured to predict the tensile strength value for 80kg high tensile steel material. For each material subjected to a total of 41 welds, 24 of them were used for neural network learning, and the remaining 17 materials were tested with the learned nerve according to the present invention to predict tensile strength.

Table 1 shows the data used for learning, and Table 2 shows the results of predicting tensile strength for 17 experiments. 5 is a graph showing the relationship between the predicted tensile strength and the actual tensile strength which are the results of Table 2.

Embodiment of neural network training data according to the present invention number Heat input Pressing force F2 acceleration Light quantity FFT coefficient 120㎐ Light FFT Factor 240㎐ / 120㎐ The tensile strength One 258783 40 0.3 3852 0.0607 69.64 2 241895 40 0.5 4188 0.0704 66.13 3 241297 40 0.7 3334 0.0494 71.23 4 228336 40 0.4 3032 0.0471 73.6 5 225778 40 0.6 3435 0.0978 59.5 6 231313 40 0.3 4026 0.2006 59.8 7 228681 40 0.5 2966 0.1732 65.24 8 229801 40 0.7 3704 0.1906 65.63 9 222496 40 0.4 2714 0.3669 55.06 10 225252 40 0.6 3210 0.3554 54.02 11 220317 40 0.3 2699 0.3656 26.09 12 220506 40 0.5 2716 0.3136 68.15 13 221953 40 0.7 3317 0.3681 54.04 14 217785 20 0.5 2147 0.2394 55.86 15 221508 40 0.5 2736 0.243 57.11 16 237739 20 0.5 2507 0.1727 58.4 17 241644 40 0.5 3378 0.1817 67.78 18 233253 20 0.5 2803 0.23 50.65 19 220048 40 0.5 2923 0.2699 68.9 20 228160 20 0.5 2295 0.1385 64.33 21 233649 40 0.5 2936 0.1542 67.68 22 232868 40 0.5 2836 0.3455 46.26 23 227261 10 0.5 3022 0.1575 34.87 24 237294 30 0.5 3006 0.1676 48.83

Example of Determination of Tensile Strength of Flash Butt Weldment Using Neural Network Experiment number Actual tensile strength Predicted tensile strength error Percentage error One 72.24 51.35 20.89 28.92 2 69.62 63.51 6.11 8.77 3 62.43 58.81 3.62 5.80 4 61.35 70.62 -9.27 -15.10 5 53.80 68.94 -15.14 -28.14 6 60.29 71.94 -11.65 -19.32 7 67.07 52.72 14.35 21.40 8 61.73 62.62 -0.89 -1.44 9 57.83 61.04 -3.21 -5.55 10 44.38 52.64 -8.26 -18.61 11 49.13 69.31 -20.18 -41.08 12 62.93 52.86 10.07 16.00 13 37.43 26.44 10.99 29.37 14 48.11 41.53 6.58 13.68 15 31.20 33.43 -2.23 -7.14 16 51.64 45.58 6.06 11.73 17 54.82 37.69 17.13 31.24

As shown in FIG. 5, the difference between the predicted tensile strength and the actual tensile strength was less than 20% during the 17 experiments, and the maximum error was 41%.

As such, it can be seen that the determination of the welding quality can be made relatively accurately by approximating the actual tensile strength.

As described above, the flash butt welding quality determination method using the multilayer error back propagation neural network according to the present invention provides an excellent method for determining the butt welding quality of poor weldable materials such as high tensile strength steel and electrical steel sheet. This can be used in the site where the flash butt welder is installed, it may be provided in a way to determine the welding quality after the welding. Therefore, it is possible to prevent sheet breakage and improve productivity in the steelmaking line.

Embodiments of the detailed description of the invention have been presented for purposes of illustration and description. It is not intended to be exhaustive or to limit the invention to the precise form disclosed. Many modifications and variations are possible in light of the above-described expressions and the like. It is intended that the scope of the invention be limited not by this detailed description, but rather by the claims. The above specification, examples and calculations are intended to provide a complete description of the manufacture and use of the invention. As there may be many embodiments without departing from the spirit and scope of the invention, the invention is based on the appended claims.

Claims (3)

An input layer into which measured values of heat input, upset pressing force, acceleration of a mobile platform, and flash light amount are input from a flash butt welding process, an output layer which outputs tensile strength of a material welded in the process, and a first layer arranged between the two layers And a second hidden layer, wherein each layer provides a neural network connected with connection strength that is a correlation between an input and an output; Calculating an error between the tensile strength as an output value and the actual tensile strength as a target value from the neural network; Modifying the connection strength of the neural network in a direction of reducing the error; The welding quality determination method comprising the step of predicting the tensile strength by inputting the measured values of the heat input, the upset pressing force, the acceleration of the moving table and the flash light amount of the material to determine the welding quality in the input layer of the neural network according to the modified connection strength . The method of claim 1, Providing the neural network further includes presetting a tolerance range between the predicted tensile strength which is the output value of the neural network and the actual tensile strength which is the target value, The calculating of the error may further include determining whether the calculated error corresponds to the tolerance range. And correcting the connection strength comprises modifying the connection strength until the calculated error falls within the tolerance range. The method according to claim 1 or 2, The measurement of the flash light amount An FFT coefficient of 120 Hz obtained by converting the amount of flash light, And a ratio of a 240 Hz FFT coefficient obtained by converting the coefficient and the amount of flash light.