This application claims priority of Japanese Patent Application No.: 2018-101718 filed on May 28, 2018, the content of which is incorporated herein by reference.
The present invention relates to a welding state determination device, a welding state determination method, and a program.
JP-A-10-235490 A discloses a method in which a power spectrum, through frequency analysis, of at least one of a welding current, a welding voltage, or a welding arc sound is previously determined in each of a normal state and an abnormal state of the welding, thereby causing a neural network to learn distinction between the normality and abnormality of each power spectrum. In the method, by using the neural network which has leaned the distinction, an actual power spectrum of at least one of the welding current, the welding voltage, or the welding arc sound during welding is evaluated, thereby determining either the normality or abnormality of the power spectrum, and simultaneously determining whether the welding state is an abnormal state or not.
However, the method disclosed in the above-mentioned patent document is difficult to achieve because it needs to prepare lots of data regarding abnormal patterns in advance on various welding conditions through experiments that reproduce the abnormal patterns.
The present invention has been made in view of the above-mentioned problems, and it is a main object of the present invention to provide a welding state determination device, a welding state determination method, and a program which can easily determine the welding state.
To solve the above-mentioned problems, a welding state determination device according to one aspect of the present invention includes: an acquisition unit that acquires a pulse waveform of a pulse current or a pulse voltage supplied to an electrode for pulse arc welding, the pulse waveform including a falling portion, a rising portion, and a flat portion therebetween; a preprocessing unit that shapes the pulse waveform such that the flat portion has a predetermined width; and a determination unit that determines a state of the pulse arc welding based on a difference between the shaped pulse waveform and a normal pattern created based on a plurality of past shaped pulse waveforms.
A welding state determination device according to another aspect of the present invention includes: an acquisition unit that acquires a probability density of a pulse current or a pulse voltage supplied to an electrode for pulse arc welding; and a determination unit that determines a state of the pulse arc welding based on a difference between the probability density and a normal pattern created based on a plurality of past probability densities.
A welding state determination device according to another aspect of the present invention includes: an acquisition unit that acquires a value at a predetermined point of a pulse current or a pulse voltage supplied to an electrode for pulse arc welding; and a determination unit that determines a state of the pulse arc welding based on a difference between the value at the predetermined point and a normal pattern created based on a plurality of past values at the predetermined point.
A welding state determination method according to another aspect of the present invention includes: acquiring a pulse waveform of a pulse current or a pulse voltage supplied to an electrode for pulse arc welding, the pulse waveform including a falling portion, a rising portion, and a flat portion therebetween; shaping the pulse waveform such that the flat portion has a predetermined width; and determining a state of the pulse arc welding based on a difference between the shaped pulse waveform and a normal pattern created based on a plurality of past shaped pulse waveforms.
A welding state determination method according to another aspect of the present invention includes: acquiring a probability density of a pulse current or a pulse voltage supplied to an electrode for pulse arc welding; and determining a state of the pulse arc welding based on a difference between the probability density and a normal pattern created based on a plurality of past probability densities.
A welding state determination method according to another aspect of the present invention includes: acquiring a value at a predetermined point of a pulse current or a pulse voltage supplied to an electrode for pulse arc welding; and determining a state of the pulse arc welding based on a difference between the value at the predetermined point and a normal pattern created based on a plurality of past values at the predetermined point.
A program according to another aspect of the present invention causes a computer to function as: an acquisition unit that acquires a pulse waveform of a pulse current or a pulse voltage supplied to an electrode for pulse arc welding, the pulse waveform including a falling portion, a rising portion, and a flat portion therebetween;
a preprocessing unit that shapes the pulse waveform such that the flat portion has a predetermined width; and a determination unit that determines a state of the pulse arc welding based on a difference between the shaped pulse waveform and a normal pattern created based on a plurality of past shaped pulse waveforms.
A program according to another aspect of the present invention causes a computer to function as: an acquisition unit that acquires a probability density of a pulse current or a pulse voltage supplied to an electrode for pulse arc welding; and a determination unit that determines a state of the pulse arc welding based on a difference between the probability density and a normal pattern created based on a plurality of past probability densities.
A program according to another aspect of the present invention causes a computer to function as: an acquisition unit that acquires a value at a predetermined point of a pulse current or a pulse voltage supplied to an electrode for pulse arc welding; and a determination unit that determines a state of the pulse arc welding based on a difference between the value at the predetermined point and a normal pattern created based on a plurality of past values at the predetermined point.
According to the present invention, the welding state can be easily determined.
Preferred embodiments of the present invention will be described below with reference to the accompanying drawings. The following respective embodiments are illustrative only to exemplify a method and device for implementing the technical idea of the present invention, and the technical idea of the present invention is not limited to the following. Various modifications can be made to the technical idea of the present invention within a technical range mentioned in the accompanied claims.
The pulse arc welding device 8 includes a welding torch 83 that is supported by a robot arm 81. The welding torch 83 has an electrode 85 for generating an arc and implements arc welding, such as, for example, Metal Inert Gas (MIG) welding or Metal Active Gas (MAG) welding.
The pulse arc welding device 8 implements pulse arc welding with a pulse current and a pulse voltage supplied from the power supply device 9. The power supply device 9 includes an ammeter or a voltmeter and outputs a detected signal of a pulse current or a pulse voltage to the welding state determination device 1.
The welding state determination device 1 is a computer that includes a CPU, a RAM, a ROM, a nonvolatile memory, an input/output interface, and the like. The CPU executes information processing in accordance with a program loaded from the ROM or nonvolatile memory into the RAM. The program may be supplied via an information storage medium, such as an optical disk or a memory card, or for example, may be supplied via a communication network, such as the Internet.
The data acquisition unit 11 is an example of an acquisition unit, the preprocessing unit 13 is an example of a preprocessing unit, the welding state determination unit 15 is an example of a determination unit, and the normal pattern creation unit 17 is an example of a creation unit.
First, the CPU acquires a pulse waveform from a detected signal of a pulse current or a pulse voltage supplied from the power supply device 9 to the pulse arc welding device 8 (S11, process as the data acquisition unit 11). The pulse waveform is cut out in units, each unit including a falling portion, a rising portion, and a flat portion therebetween. The flat portion may be a base portion or a peak portion.
Then, the CPU shapes the pulse waveform such that the flat portion thereof has a predetermined width (S12, process as the preprocessing unit 13), and stores the shaped pulse waveform in the database 2 (S13). The width of the flat portion of the pulse waveform may vary depending on welding conditions, power supply control, and the like. Due to this, the widths of the flat portions of the pulse waveforms are equalized in order to facilitate comparison between the pulse waveforms.
Then, the CPU creates a normal pattern based on the plurality of shaped pulse waveforms stored in the database 2 (S14, process as the normal pattern creation unit 17) and stores the normal pattern in the database 2 (S15). It is not necessary to prepare a lot of abnormal pulse waveforms, the number of which is smaller than that of the normal pulse waveforms, because the normal pattern is created in the present embodiment.
First, the CPU acquires a pulse waveform from a detected signal of the pulse current or the pulse voltage supplied from the power supply device 9 to the pulse arc welding device 8 (S21, process as the data acquisition unit 11). Here, the pulse waveform is cut out in the same unit as that in the normal pattern creation process of S11 shown in
Then, the CPU shapes the pulse waveform such that the flat portion thereof has a predetermined width (S22, process as the preprocessing unit 13). Here, the pulse waveform is shaped such that the flat portion thereof has substantially the same width as the flat portion in the normal pattern creation process of S12 shown in
Then, the CPU reads out the normal pattern stored in the database 2 and calculates a difference between the normal pattern and the pulse waveform shaped in S22 which is an immediately preceding step (S23). Subsequently, the CPU determines the welding state based on the calculated difference (S24, process as the welding state determination unit 15). The present embodiment utilizes the normal pattern and thus can easily determine the welding state.
More specific examples of the normal pattern creation process and the welding state determination process will be described below.
A method of detecting abnormality of a pulse waveform will be described which involves preprocessing a pulse waveform to shape it into a scaled waveform and then conducting pattern matching thereon.
Here, the pulse waveform is proposed to be cut out for each pulse.
In either case, another waveform with a different pulse width from its original waveform is mixed in the original waveform due to the influence of control on a power supply device side regarding the pulse width. This makes it difficult to extract the disturbance from the respective pulse waveforms. For this reason, as shown in
Specifically, after taking a moving average at several points of the pulse waveform, a difference in the value between the adjacent sample points is determined to thereby calculate a gradient there, and subsequently, the process is conducted to widen a portion of the pulse waveform that has an absolute value of the gradient of a predetermined value or less. When widening the portion, a process is performed to conduct linear interpolation between the sample points. Thereafter, only by widening the flat portion with the gentle gradient, the pulse waveforms may not be strictly matched in terms of the number of sample points in the lateral width of the pulse waveform when including the falling portion and rising portion with the steep gradients. Due to this, the process is performed to adjust the entire width of each pulse waveform to approximately 100 points, while the pulse waveform includes the falling portion with the steel gradient, the widened flat portion (base portion) with the gentle gradient, and the rising portion with the steep gradient.
Next, an example will be described in which abnormal data is detected by conducting principal component analysis using only the normal pulse waveforms or a large number of pulse waveforms, most of which are normal pulse waveforms. In normal welding, it is considered that most of pulse waveforms become normal pulse waveforms while extremely small parts of the pulse waveforms become the abnormal waveforms. The following calculation can be applied without any problem when most of the pulse waveforms are the normal pulse waveforms.
When there are N shaped pulse waveforms (x1, x2, . . . , xN), one pulse x1=[x11, x21, . . . , xp1]T is a p-dimensional vector and, specifically, a vector of about 100 dimensions. That is, when X=[x1, x2, . . . , xN] is represented by a matrix, the following equation 1 is given.
Here, the mean μ and standard deviation σ for each row of the matrix are calculated. Each of these calculation results is also a p-dimensional vector. Then, normalization that involves subtracting the mean μ and dividing by the standard deviation σ is performed on each of the N pulse waveforms x1, x2, . . . , xN so as to obtain the mean of 0 and the standard deviation of 1.
Thereafter, principal component analysis is conducted to calculate a reconstruction error, thereby obtaining the degree of abnormality of the pulse waveform. Specifically, the degree of abnormality α1(x′) of each shaped pulse waveform x′, which is a target for calculation of the degree of abnormality, can be calculated by the following equation 3 on the assumption that u1, u2, . . . , um are m vectors with the first to m-th highest variances among the obtained principal component vectors as represented by the following equation 2, and Ip is the unit matrix of p rows and p columns, by using x{tilde over ( )}(i.e., x with a wavy sign added) obtained by subtraction of the mean μ and then division by the standard deviation σ.
U
m=[u1,u2 . . . ,um] [Equation 2]
α1(x′)={tilde over (x)}T[lM−UmUmT]{tilde over (x)} [Equation 3]
The calculation result of the degree of abnormality of the pulse waveform will be shown below.
The calculation of the principal component vector by the principal component analysis is an example of the normal pattern creation and corresponds to the normal pattern creation unit 17 and step S14. The calculation of the reconstruction error, that is, the calculation of the degree of abnormality is an example of the welding state determination and corresponds to the welding state determination unit 15 and step S24.
In the way mentioned above, the degrees of abnormality of the abnormal pulse waveforms are calculated, and some of the pulse waveforms with the higher degrees of abnormality are extracted, thus making it possible to detect the abnormality of welding. Although in the above-mentioned example, the normalization process is performed, the calculation of the degree of the abnormality is possible without performing any normalization process.
Even if the pulse waveform is normal, the width or shape of the pulse could vary depending on the set current, the set voltage, and the power source control. However, in the present embodiment, the pulse waveform is scaled and shaped, thus enabling improvement of the detection accuracy of abnormality. That is, even under various welding conditions, the abnormality of welding can be detected because the pulse waveform is scaled and shaped. In addition, since the pulse waveform is scaled and shaped, the abnormality of welding can be detected even under the influence of macro changes in average current, average voltage, or the like, which vary depending on the relational position of a welding torch relative to a workpiece in terms of the height or the lateral position of the welding torch.
Although the principal component analysis can detect the abnormality of welding in the above-mentioned embodiment, the difference from the normal pattern can be clarified by shaping the pulse waveform in the embodiment, and hence it is considered that the abnormality of welding can be detected by various methods as well as the principal component analysis.
For example, only the normal pulse waveforms or a large number of pulse waveforms, most of which are normal pulse waveforms, are shaped. Then, the shaped pulse waveforms are averaged to produce an average pulse waveform, which is referred to as the normal pattern. Subsequently, by calculating a distance between the average pulse waveform and a pulse waveform corresponding to a welding state, which is a target to be determined, the degree of abnormality of the welding state can be calculated. Even though a small number of abnormal pulse waveforms are contained, the influence of these abnormal pulse waveforms can be suppressed by averaging the pulse waveforms. Alternatively, the degree of abnormality of the target pule waveform can be calculated by calculating a Mahalanobis distance between a large number of shaped pulse waveforms and a pulse waveform corresponding to a welding state, which is a target to be determined.
Next, a method of detecting abnormality by performing clustering with k-means after shaping the pulse waveforms will be described.
That is, the pulse waveform assigned to the lower level cluster is regarded as the abnormal one, so that the abnormality of welding can be detected. In addition, one-class support vector machine can also be used, and it is needless to say that supervised learning, such as an ordinary support vector machine or a decision tree, can also be applied in the presence of a relatively large amount of abnormal data.
Although the base portion on the lower side of the pulse waveform is focused on in the embodiment mentioned above, the present invention is not limited thereto. For example, the pulse waveform from the rising portion to a next rising portion thereof, that is, the entire pulse waveform including both the peak portions and the base portion thereof may be preprocessed.
Regarding the scaling for adjusting the levels of the peak portions and the base portion in the manner mentioned above, a scaled value a{circumflex over ( )} (a with {circumflex over ( )} thereon) is determined by the following equation 4:
where μp is an average current value at the peak portion of the pulse waveform, μB is an average current value thereof at the base portion, and a is a current value at each time.
In the above-mentioned embodiment, the preprocessing is performed on the pulse current, but may be on both the pulse current and the pulse voltage.
The left half part of the diagram shows preprocessed current values of about 100-dimensional vectors, and the right half part thereof shows preprocessed voltage values of about 100-dimensional vectors, which result in about 200-dimensional vectors by simply placing both parts side by side laterally. The voltage value varies due to the influence of weaving and the like. However, variations in the voltage value of such an extent is considered not to be problematic, because the normalization that involves subtracting the mean μ and dividing by the standard deviation σ is performed on the pulse waveforms so as to exhibit a mean of 0 and a standard deviation of 1 in calculating the degree of abnormality of the pulse waveform.
In this way, the preprocessing can be applied not only to the base portion of the pulse waveform, but also to the peak portion thereof. In addition, the preprocessing can be applied not only to the pulse current, but also to a pulse voltage. Here, the pulse voltage may be simply shaped in the same manner as the pulse current mentioned above. However, as the voltage value significantly varies, sample points used for a process of matching the widths of flat portions of the pulse voltage waveforms may be the sample points used for the process of matching the widths of the flat portions of the pulse current waveforms, each sample point having a small absolute value of the gradient of the current value. In this case, which sample point counted from the front is used to shape the pulse current is remembered, and by using such a remembered sample point, the process of matching the widths of the flat portions of the pulse voltage waveforms may be performed.
Next, a method of detecting the abnormality will be described which involves extracting an output pattern of a current value by probability density estimation and conducting pattern matching. The probability density estimation is expressed by, for example, equation 5 below.
where n is a sample size, K is the kernel smoothing function, and h is a bandwidth.
It is found that by performing preprocessing of the probability density estimation every weaving, the influence by changes in the current value due to the weaving is equalized, thereby emphasizing the difference between the normal pulse current and the abnormal pulse current.
Then, an example will be described in which the principal component analysis is performed on the pulse waveforms using the results of the probability density estimation, thereby detecting abnormal data.
The calculation result of the degree of abnormality of the pulse waveform will be shown below.
As mentioned above, the degrees of abnormality of these pulse currents are calculated, and some of the pulse currents with the higher degrees of abnormality are extracted, thus making it possible to detect the abnormality of welding. The calculation of the principal component vector by the principal component analysis is an example of a process executed by a creation unit, and the calculation of a reconstruction error, i.e., the calculation of the degree of abnormality is an example of a process executed by a determination unit.
Since the probability density is calculated in a cycle including one weaving, a set of changes in the current or voltage during one weaving can be obtained even when the relative position between the welding torch and a workpiece changes due to the weaving. In this way, it is considered that the probability density pattern of the pulse waveform can be stably acquired. That is, as the pulse waveform is scaled and shaped, the abnormality of welding can be detected even in a situation where the relative position between the welding torch and the workpiece changes due to the weaving.
Next, a method of detecting the abnormality will be described which involves extracting output patterns of current values by setting the same corresponding points of the base portions of the respective pulses as a sample point and then by conducting pattern matching, in a set of repeated pulse waveforms of the pulse current.
Here, the point near the center of the base portion is proposed to be sampled.
Here, the degree of abnormality of a newly obtained sample point x′ is represented by the following equation 6:
where x1, x2, . . . , xN are sample points, μ is the mean thereof, and σ is the standard deviation thereof.
The calculation result of the degree of abnormality will be shown.
As mentioned above, by using one point for each pulse, the abnormality of welding can be detected while suppressing the process time. The extraction of the sample point is an example of the process executed by a creation unit, and the calculation of the degree of abnormality is an example of the process executed by the determination unit. Although only one point is extracted from each pulse in the embodiment, a plurality of points may be extracted therefrom. The embodiment is not limited to the pulse current and may be applied to a pulse voltage.
Number | Date | Country | Kind |
---|---|---|---|
2018-101718 | May 2018 | JP | national |