JPH10235490A - Method for evaluating weld state of electric welding machine and device therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method for evaluating weld state of automatic electric welding machine without infrared camera. SOLUTION: With a welding abnormality detector 7, frequency analytic power spectrum of at least one of the welding current I, welding voltage V and welding arc sound S in each of the welding normal state and abnormal state is found beforehand, and then the difference between normality and the distinction between normality and abnormality is learnt by neural network. By means of this learnt neural network, the power spectrum of at least one of the welding current I, welding voltage V and welding arc sound S is evaluated and normality or abnormality is decided, thus at the same time abnormal state is discriminated.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a method for evaluating a welding condition of an electric welding machine, and more particularly to a method for evaluating a welding condition suitable for use in an automatic electric welding machine.

[0002]

2. Description of the Related Art Conventionally, as a method for evaluating a welding state of an electric welding machine, for example, a method for inspecting a welding defect using an infrared camera described in Japanese Patent Application Laid-Open No. 63-68268 is known. The content is as follows: the infrared camera detects the radiant energy distribution from the molten pool of narrow groove welding and the weld solidified part in the vicinity as an energy image signal, and converts the detection signal into a temperature distribution image by an image processing device.
The temperature distribution image is transmitted to an image identification device, the image identification device determines the presence or absence of a welding defect based on the temperature distribution image, and generates a defect occurrence signal when a welding defect is detected. Things.

[0003] In addition, a technique disclosed in Japanese Patent No. 2606497 is known as an arc welding robot system having a function of detecting the occurrence of blowholes from an arc sound. This arc welding robot system includes an arc sound database storing arc sound spectrum data in an arbitrary frequency range included in an arc sound during welding when a blow hole is generated and when a blow hole is not generated, and an arc sound sampling. Unit, an arc sound analysis unit, an arc sound determination unit, a robot control device having a welding control unit, and an arc welding robot,
This is a system consisting of an arc welding machine.

[0004]

However, according to the conventional method described in Japanese Patent Application Laid-Open No. 63-68268, when the infrared camera is interfered depending on the groove shape or the shape of the work, the molten pool cannot be properly photographed. Because there is
There is a drawback that a case where the radiant energy distribution cannot be measured accurately occurs.

[0005] Further, welding defects include not only blow holes but also various defects such as poor fusion, remaining grooves, slag entrapment, and excessive welding amount. Patent No. 2606497 corresponds to only blow holes, and other defects include Not supported. Also,
Since a method is used in which a power spectrum of each of the abnormal and normal power spectra is collated in a database, a large amount of data accumulation and collation time are required to deal with a large number of defect types. Furthermore, when the sheet thickness and welding conditions change, data must be accumulated for each of them.

In general, the conventional determination of the welding state can be applied only to distinguish between normal and abnormal types, and the cause of the abnormality is identified by the knowledge and manual operation of the operator. Was not effective.
The present invention has been made in order to solve the problems of the prior art as described above, and does not require a database, and is capable of determining whether the welding state is normal or abnormal. An object is to provide an evaluation method and an apparatus.

It is a further object of the present invention to provide a method and an apparatus for evaluating a welding state of an electric welding machine capable of determining and presenting not only normal / abnormal but also various welding abnormal states.

[0008]

SUMMARY OF THE INVENTION The present invention relates to a method for evaluating a welding state when a workpiece is welded by using an electric welding machine. At least one of the voltage and the welding arc sound is subjected to frequency analysis to obtain a power spectrum, and a neural network is used to learn the distinction between normal and abnormal. Using the learned neural network, a welding current during welding, welding is performed. A method for evaluating a welding state of an electric welding machine, comprising: determining whether a welding state is normal or abnormal by evaluating at least one power spectrum of a voltage and a welding arc sound.

Further, at least one power spectrum of a welding current, a welding voltage and a welding arc sound in each of a normal state of welding and a plurality of abnormal states (of abnormal states) specified in advance is obtained to obtain a neural network. And learning at least one power spectrum of welding current, welding voltage and welding arc sound during welding by using the learned neural network, By determining whether the state is normal or any of several types of specified abnormalities, the method for evaluating the welding state of the electric welding machine can perform even the type of abnormality determination.

Preferably, the power spectrum is obtained by an autoregressive analysis method. The apparatus of the present invention includes a means for frequency analysis of at least one of a welding current, a welding voltage, and a welding arc sound when welding in an electric welding machine to obtain a power spectrum, and a means for obtaining the power spectrum. Neural network means which has learned in advance the distinction between a normal state and an abnormal state of welding based on the input of the neural network means, and an output of the neural network means. It is preferable to provide a device for evaluating the welding state of an electric welding machine, characterized by having a determination means for determining whether or not the above condition is satisfied.

Similarly, at least one of a welding current, a welding voltage and a welding arc sound at the time of welding in an electric welding machine is frequency-analyzed to obtain a power spectrum, and an input from the means for obtaining the power spectrum is provided. The normal state of welding and the specified abnormal state are distinguished in advance from the neural network means, and the output of the neural network means is used to determine the normal state and specified welding state from the at least one power spectrum. It is preferable to provide a device for evaluating a welding state of an electric welding machine, characterized by having a determination means for determining which of several abnormal states corresponds to the above.

In the above apparatus, it is preferable that the power spectrum is obtained by an autoregressive analysis method. Further, it does not prevent the means for obtaining the power spectrum from having a function of storing the obtained power spectrum.

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will be described below in detail with reference to the drawings. FIG.
It is a schematic diagram showing the composition of one embodiment of the electric welding machine of the present invention. In the figure, reference numeral 1 denotes a work to be welded, and reference numeral 2 denotes a welding electrode controlled by a robot (not shown). Reference numeral 3 denotes a welding power supply unit that supplies welding power to the work 1 and the welding electrode 2 via the ground cable 4 and the power cable 5. Reference numeral 6 denotes an acoustic sensor such as a microphone provided near the welding electrode 2. Numeral 7 denotes a welding abnormality detecting device. Respectively.

Here, the configuration of the welding abnormality detecting device 7 will be described in detail with reference to FIG. In FIG. 2, 11 is a welding current waveform sampler, 12 is a welding voltage waveform sampler, and 13 is a welding arc acoustic waveform sampler. Reference numeral 14 denotes a frequency analyzer, which performs frequency analysis by inputting each waveform signal sampled by the welding current waveform sampler 11, the welding voltage waveform sampler 12, and the welding arc acoustic waveform sampler 13, and outputs a power spectrum. .

15 is a neural network (Neural Net).
work: a neural network), which receives a power spectrum signal from the frequency analyzer 14 and determines whether the welding state is normal or abnormal. The structure of the neural network 15 will be described. For example, as shown in FIG.
The input layer 16 has, for example, 512 neurons (Neuron: neuron unit, blk 1 to blk).
8) In the middle layer 17, for example, 16 neurons (blk 1)
The output layer 18 is composed of, for example, one neuron (blk 1).

The output of the neural network is as follows. In the example of FIG. 4, the intensity (dB) of the power spectrum corresponding to each frequency obtained by dividing 1 to 512 kHz into 512 by 1 kHz is input to each neuron of the input layer 16. Then, a value obtained by adding a predetermined threshold value different for each neuron of the input layer to the above-described input value is output from each neuron of the input layer 16 to each neuron of the intermediate layer 17. In each neuron of the intermediate layer 17, the output from each neuron of the input layer 16 is multiplied by an individual coefficient (connection coefficient) and added, and then a predetermined threshold value different for each neuron of the intermediate layer is further added. , Each neuron of the output layer 18 (1 in this example)
Output).

It is preferable that the input layer is composed of 300 neurons or more. If the number of neurons is less than 300, it is difficult to capture the characteristics of the power spectrum, and the accuracy rate decreases. The upper limit is not particularly required in principle, but if it exceeds 800 neurons, it takes a very long time for learning (optimization calculation such as threshold value), so that it is inefficient.

The intermediate layer is preferably composed of 10 neurons or more. If the number is less than 10 neurons, the degree of freedom is too small to perform a sufficient optimization, so that an appropriate learning effect cannot be obtained and the correct answer rate decreases. On the other hand, if the number of neurons in the intermediate layer is too large, the degree of freedom is too high, so that proper learning cannot be performed. The appropriate upper limit of the number of neurons in the hidden layer depends on the number of output neurons.
It is better to have a large number of neurons in the hidden layer.)
In the case of about 8 to 8 neurons, 40 or less neurons are preferable.

Reference numeral 19 denotes a display device for outputting and displaying the evaluation result of the neural network 15. The procedure for evaluating the welding state in the electric welding machine configured as described above will be described below. In advance, waveform signals in a normal state and an abnormal state of the welding current, welding voltage, and welding arc sound when the work 1 is welded by the electric welding machine are collected. These welding current, welding voltage, and welding arc sound waveform signals are frequency-analyzed by the frequency analyzer 14 using, for example, an autoregressive analysis method, to obtain respective power spectra. The neural network 15 learns the discrimination between the normal pattern and the abnormal pattern of each of the obtained power spectra.

That is, for example, when the power spectrum of the normal pattern is input, the output of the output layer 18 is 0, and when the power spectrum of the abnormal pattern is input, the output layer 18 is output.
The threshold value at the time of output of each neuron and the coefficient by which the input value is multiplied are optimized so that the 18 outputs are each closest to 1 (for example, a back propagation method is used). Next, the welding current, welding voltage, and welding arc sound waveform signals during welding work are sampled by a welding current waveform sampler.
11, a welding voltage waveform sampler 12 and a welding arc acoustic waveform sampler 13 are respectively detected and learned by the neural network 15 to learn whether the welding condition is normal or not.
Determine the abnormality. The normal / abnormal discrimination result is output to the display device 19 and displayed. If the determination result is normal, the welding abnormality detection device 7
A signal is output, and in the case of an abnormality, a STOP signal is output. It should be noted that such GO / STOP information can also be used for determining a flaw detection inspection method and repair.

It should be noted that in the above-mentioned embodiment, the welding current,
Although the description has been made assuming that three elements of welding voltage and welding arc sound are used at the same time, the present invention is not limited to this, and any one of these elements alone or any two elements may be used.
Even when the elements are used in combination, the same operation and effect can be obtained.

In the above example, the normal / abnormal discrimination of the power spectrum is performed in real time.
The power spectrum may be stored for a while, and after that, normality / abnormality may be determined and evaluated. FIG. 3 is a schematic diagram illustrating an example of a configuration of a welding control device used in the present invention. In FIG. 3, FIG.
As in the above, the welding current I is sampled by a welding current waveform sampler, the welding voltage V is sampled by a welding voltage waveform sampler, and the welding arc sound S is sampled by a welding arc sound waveform sampler. In this embodiment, 8000 points are sampled at a rate of 1 Hz. And this
8000 × 3 = 24000 points of sampling data are recorded and stored in the data memory 22. Next, the frequency analyzer 14
Find the power spectrum of each species data. The obtained power spectrum is evaluated by the neural network 15, and a GO / STOP signal is transmitted via the signal transmitter 23 to the electric welding machine 26 operated by a robot or the like.

After this series of processing, a new welding current,
The welding voltage and welding arc sound are sampled and stored in the data memory 22, and the frequency analyzer 14 recognizes the latest data in the data memory 22 and obtains a power spectrum. Also,
It is also possible to select arbitrary data from the stored data, perform frequency analysis and organize the results into a file, and perform re-learning of the neural network. These operations are performed via the operation device 25, and the results and the like can be displayed on the display device 19.

The CPU 24 manages these operations. FIG. 5 shows a neural network used in the present embodiment.
Shown in The input layer 16 is composed of 512 neurons, the intermediate layer 17 is composed of 16 neurons, and the output layer 18 is composed of 5 neurons. In this way, for example, by setting the output layer to five neurons and assigning the welding state to each neuron, it is possible to specify not only whether the welding state is normal or abnormal, but also the type of abnormal state.

Here, in the example of FIG. 6, the welding state is “normal”, “target of the torch” is abnormal, “torch height”.
Five states of abnormal, "groove wet" abnormal and "no gas" abnormal are assigned and used as teacher data. Here, as a method of performing a frequency analysis to obtain a power spectrum,
It is preferable to use an autoregressive analysis method.

FF which is a typical method of spectrum analysis
In T (Fast Fourier Analysis), the sensitivity is too good. For example, if the curves in FIGS. 10 to 12 are taken as an example, even small fluctuations are reproduced, which makes learning with a neural network rather difficult. . The auto-regression analysis method is a method of spectrum estimation by linear prediction, and estimates a power spectrum using an auto-regression model (a model in which a current output is obtained by feedback of a past output). Specifically, the current time is n
Then, the output value x (n) at the present time is converted into the past time series i
The output value x (i) at (i <n) and the current input value u (n) are represented by the following equation (1). x (n) = u (n ) -Σa k x (n-k) ... (1) where, sigma represents the sum of k = 1 to p. Also, p is a predetermined integer, and a _k represents a coefficient. By setting the human input signal u (n) as white noise having an average value of 0 in the above model, a power spectrum can be obtained from the variance of the white noise and the value of a _i .

The method of determination and display based on the output from (the output layer of) the neural network is not specified. For example, the output of each neuron in the output layer may be directly displayed as a number or a graph. In the example of allocating each neuron of the output layer for each welding state, the welding state corresponding to the neuron closest to 1 may be displayed, or may be output as a signal or voice, or may be higher than a specific threshold (for example, 0.2 or higher). May be listed up and displayed, etc. Furthermore, when there is no neuron having a value equal to or more than a specific threshold value (for example, 0.8), "unrecognizable" may be displayed.

[0028]

[Working example] As shown in FIGS. 6 (a) and 6 (b), the work 1 has a standard of SM490 and dimensions of 300 mm length × 230 mm width × thickness.
Specimen 1a which is 19mm, standard is SM490 and dimensions are length
A test material 1b having a size of 300 mm x a width of 264 mm x a thickness of 19 mm and having a 35 ° groove on one side was used. The test materials 1a and 1b were opposed to each other at intervals of 6 mm on the backing material 20, and the method of the present invention was applied when performing a welding experiment under the standard welding conditions shown in Table 1. At this time, using a welding wire with a diameter of 1.2 mmφ, shield gas of 100% CO ₂ was used at 25 l / min.
Supplied at the rate of

[0029]

[Table 1]

First, when welding was performed by flowing a shield gas of 25 l / min, a normal state was set, and when the shield gas was not flown at 0 l / min, an abnormal state was defined. When the waveform signals in the normal and abnormal states of the voltage and the welding arc sound were detected, the characteristics shown in FIGS. 7A and 7B for the welding current, and FIGS. 8A and 8B for the welding voltage. When the characteristic shown in (b) is a welding arc sound, the characteristics shown in FIGS. 9 (a) and (b) were obtained, respectively.

Next, the power spectrum was obtained by frequency-analyzing the waveform signals in the normal and abnormal states by the autoregressive analysis method.
(a) and (b), the characteristics as shown in FIGS. 11 (a) and (b) for the welding voltage and the characteristics as shown in FIGS. 12 (a) and (b) for the welding arc sound. The illustration of the frequency power spectrum of 500 kHz or more is omitted.

After the neural network 15 learned these frequency power spectra, the welding current, welding voltage, and welding arc sound waveform signals during welding were similarly evaluated by obtaining a power spectrum by an autoregressive analysis method. . Table 2 shows the results of the evaluation.

[0033]

[Table 2]

As is clear from Table 2, the correct answer rate for the welding current excluding the welding voltage and the welding arc sound is 100%.
%, Indicating that normality / abnormality can be determined with a high probability. Further, as an abnormal state, the shielding gas is 0 l / m
In addition to the factor of "in", three factors such as a defective welding electrode aiming position, a poor wire protrusion length, and the occurrence of wetness in a groove were obtained in the same manner, and a frequency power spectrum was obtained by frequency analysis for learning of the neural network 15. Table 3 shows the results.

[0035]

[Table 3]

In the results of Table 3 as well, in any of the welding current, welding voltage, and welding arc sound, similar to Table 2 described above, normal / abnormal discrimination with a high probability for each abnormal state. You can see that you can do it. Next, a case will be described in which not only normal or abnormal but also the type of abnormal state is specified.

FIGS. 13A and 13B show power spectra in the welding state, respectively. FIG. 13A shows a normal state, FIG. 13B shows a torch aiming abnormal state in which the target position of the welding torch is abnormal, and FIG.
Is the abnormal torch height condition where the welding torch height position is abnormal, (d) is the abnormal groove wet condition where the welding groove is wet, and (e) is no shielding gas. FIG. 7 is a power spectrum characteristic diagram showing an abnormal state with no gas in each case.

As shown in FIG. 5, based on this characteristic diagram, the teacher data is shown in FIG.
1 = 0, No2 = 0, No3 = 0, No4 = 0, NO5 = 1, If the torch target position is abnormal, No1 = 0, No2 = 0, No3 = 0, No4 = 1, No5 =
0, No1 = 0, No2 = 0, No3 = 1, No if the torch height is abnormal
4 = 0, No5 = 0, No1 = 0, No if the groove is abnormally wet
2 = 1, No3 = 0, No4 = 0, No5 = 0, If the shield gas is abnormally low, No1 = 1, No2 = 0, No3 = 0, No4 = 0, No5 = 0
, And give. The neural network which can evaluate the welding condition appropriately by constructing the threshold value and the connection coefficient of each neuron satisfying this condition by learning is constructed.

Table 4 shows the results of evaluating 18 normal-state power spectra using this neural network.
In all of # 1 to # 18, the output terminal (neuron) No5 is almost 1, and No1 to No4 are 0 (zero), so that the accuracy rate is 100%. Similarly, Table 5 shows the case where the torch aiming position is abnormal, Table 6 shows the case where the torch height is abnormal, Table 7 shows the case where the groove is abnormally wet, and Table 8 shows the case where the shielding gas is not discharged. The evaluation results of each network are shown.

[0040]

[Table 4]

[0041]

[Table 5]

[0042]

[Table 6]

[0043]

[Table 7]

[0044]

[Table 8]

Since the output terminal No. 4 in Table 5 is almost 1 and the other output terminals are almost 0, the accuracy rate is 100%. Table 6
Since the output terminal No. 3 is almost 1 and the other output terminals are almost 0, the accuracy rate is 100%. The value of output terminal No2 in Table 7 is # 2
Since 0.5 and # 4 are 0.2, they could not be evaluated properly. Therefore, the accuracy rate is 16/18 = about 90%. Output terminal No. in Table 8
1 is almost 1 and the other output terminals are almost 0, so the accuracy rate is 100%.

That is, it was confirmed that the determination of normal / abnormal was 100% correct, and the determination of the type of abnormality was extremely high.

[0047]

As described above, according to the present invention,
By analyzing the welding current, welding voltage, and welding arc sound during welding, and evaluating the power spectrum with a neural network, it is possible to determine whether the welding condition is normal or abnormal, and even to determine the type of abnormality. Therefore, it is possible to accurately evaluate the welding state of the electric welding machine with a high probability, and it is possible to greatly contribute to improvement in welding quality.

Further, compared to the conventional method of collating with a database, the determination method of the present invention using a neural network is less affected by fluctuations in welding conditions and the like, and there is a condition change beyond the degree of freedom of the neural network. Even though they can learn and respond quickly, they are highly adaptive.

[Brief description of the drawings]

FIG. 1 is a schematic diagram showing a configuration of an embodiment of an electric welding machine of the present invention.

FIG. 2 is a block diagram showing a configuration of a welding abnormality detection device used in the present invention.

FIG. 3 is a block diagram showing a configuration of a welding control device used in the present invention.

FIG. 4 is a schematic diagram showing an example of a neural network structure used in the present invention.

FIG. 5 is a schematic diagram showing another example of a neural network structure used in the present invention.

FIG. 6A is a plan view showing the shape of a work used in the experiment,
(b) It is a side view.

7A and 7B are characteristic diagrams showing a normal state and a abnormal state of a welding current waveform, respectively.

8 (a) is a characteristic diagram showing a normal state and FIG. 8 (b) is a characteristic diagram showing an abnormal state.

FIG. 9A shows a normal state of the welding arc sound waveform.
(b) is a characteristic diagram showing an abnormal state.

10A and 10B are characteristic diagrams showing a power spectrum of a welding current in a normal state and an abnormal state, respectively.

11 (a) and 11 (b) are characteristic diagrams showing a welding voltage power spectrum showing a normal state and an abnormal state, respectively.

12A and 12B are characteristic diagrams illustrating a power spectrum of a welding arc sound showing a normal state, and FIG. 12B showing an abnormal state.

FIG. 13 (a) shows a normal state of the power spectrum of the welding state, (b) shows an abnormal state of aiming the torch, (c) shows an abnormal state of the torch height, (d) shows an abnormal state of groove wetness, (e) is a characteristic diagram showing an abnormal state without gas.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Work 1a, 1b Test material 2 Welding electrode 3 Welding power supply 4 Earth cable 5 Power cable 6 Acoustic sensor 7 Welding abnormality detector 8, 9, 10 Signal line 11 Welding current waveform sampler 12 Welding voltage waveform sampler 13 Welding Arc acoustic waveform sampler 14 Frequency analyzer 15 Neural network 16 Input layer 17 Middle layer 18 Output layer 19 Display device 20 Backing material 21 Welding control device 22 Data memory 23 Signal transmitter 24 CPU 25 Operation setting device 26 Electric welding machine ( Electric welding robot, etc.)

Claims (6)

[Claims] 1. A method for evaluating a welding state when welding a workpiece using an electric welding machine, wherein at least one of a welding current, a welding voltage, and a welding arc sound in a normal state and an abnormal state of the welding in advance. One is subjected to frequency analysis to obtain a power spectrum, and the difference between normal and abnormal is learned by a neural network. Using the learned neural network, welding current, welding voltage and welding arc sound during welding are used. A method for evaluating a welding state of an electric welding machine, comprising: determining whether a welding state is normal or abnormal by evaluating at least one power spectrum. 2. A method for evaluating a welding state when a workpiece is welded using an electric welding machine, comprising a welding current, a welding voltage and a welding voltage in a normal state of welding and a plurality of abnormal states specified in advance. At least one of the welding arc sounds is frequency-analyzed to determine a power spectrum, and a neural network is used to learn the discrimination between some normal and specified abnormalities. Using the learned neural network, welding during welding is performed. Welding of an electric welding machine characterized by evaluating at least one power spectrum of a current, a welding voltage and a welding arc sound to determine whether a welding state is normal or one of several specified abnormalities. State evaluation method. 3. The method for evaluating a welding state of an electric welding machine according to claim 1, wherein said power spectrum is obtained by an autoregressive analysis method. 4. At least one of a welding current, a welding voltage and a welding arc sound when welding in an electric welding machine.
Means for frequency analysis of the two to obtain a power spectrum;
Neural network means, which has learned in advance the distinction between a normal state and an abnormal state of welding based on the input from the means for obtaining the power spectrum, and a welding state from the at least one power spectrum using the output of the neural network means. A determination unit for determining whether the state corresponds to a normal state or an abnormal state. 5. At least one of a welding current, a welding voltage, and a welding arc sound when welding in an electric welding machine.
Means for frequency analysis of the two to obtain a power spectrum;
Neural network means, which has learned in advance the distinction between the normal state of welding and some specified abnormal states based on the input from the means for obtaining the power spectrum, and the output of the neural network means, Determining means for judging from the power spectrum whether the welding state corresponds to a normal state or any of a plurality of specified abnormal states. An evaluation apparatus for evaluating a welding state of an electric welding machine. 6. The means for determining the power spectrum,
The apparatus for evaluating the welding state of an electric welding machine according to claim 4 or 5, wherein said means is a means for obtaining by an autoregressive analysis method.