3.1. Problems of Average Current-Based Arc Sensing
Arc sensing defines the left and right weaving areas based on the welding center line to take into account the torch’s weaving motion. It uses the average values of the feedback current measured in each area to estimate the torch’s bias. By comparing the average values of feedback current measured in the left and right areas, the area with the relatively higher average value is considered the area where the torch is biased.
Figure 2 shows the changes in feedback current due to torch bias. Based on this, an offset is applied to shift the center of the torch. This weaving-area-based average current value arc-sensing technique shows effective performance in large environments, such as when the weld seam is straight. However, when the workpiece has significant curvature, causing rapid changes in the distance between the weld target workpiece and the torch, or when a certain curvature and gap exist on the weld workpiece, the reliability of the feedback current can decrease or the response can be delayed, increasing the possibility of seam tracking failure.
We would like to analyze problem instances by observing the actual feedback current values during the welding of workpieces with a certain level of curvature and gap.
Figure 3 shows the weaving-area-specific feedback current values in conditions where there is a 3 mm gap in the target workpiece: (a) when the gap is located at the midpoint of the torch’s weaving, and (b) when seam tracking is delayed and the gap is located at the stay point of the weaving. In condition (a), it can be seen that the feedback current values are collected stably at a certain level. However, in (b), there is a sudden increase in feedback current values followed by a sudden decrease due to the transition to short-circuit control.
During the weaving process, the torch stays at the ends of the left and right weaving areas for a certain cycle before moving to the opposite side. Under normal circumstances, the gap is positioned at the midpoint of the torch’s weaving and passes quickly by without significantly affecting the feedback current values. However, if the offset correction is delayed due to inaccurate and unstable average feedback current values in each area caused by curvature, an instance can occur where the torch stays at the gap point, as shown in
Figure 3c. This results in decreased resistance due to the reduced molten pool, causing the feedback current to increase drastically. As mentioned earlier, these sudden high currents can be a major cause of welding defects. Therefore, control of short-circuit transition is performed by the inverter welder, and the measured feedback current shows an abnormal pattern of rapid increases and decreases. In order to alleviate the complexity of the proposed scheme, we assumed general conditions for weld geometry parameters such as weld angle, bead size, and surface characteristics. Based on this, only two scenarios, such as a workpiece with severe curvature and a workpiece with a gap, are adopted for evaluating the effect of the proposed scheme.
Figure 4 shows the weaving-area-specific average feedback current values measured in the same condition as above, displayed for each weaving cycle. Due to the curvature of the workpiece, the torch gradually gets closer to the left side of the workpiece. Therefore, as the weaving progresses, the average current value in the left area should increase, while the average value in the right area should decrease. However, the actual measured average value in the left area does not show much difference even as the weaving progresses and, instead, a sudden large decrease is observed in the fifth cycle, indicating unstable results. Afterwards, in the 11th, 12th, and 13th cycles, corresponding to the part in
Figure 3b, very unstable average values are observed in the right area. Moreover, although the torch is closer to the left area and an increased feedback current value should be observed, the average value does not change significantly due to the control of short-circuit transition, which has already lowered the current levels. As a result, the average current value in the left area maintains a certain level, while the average current value in the right area shows very unstable values. According to these actual observations, it can be seen that, in conditions where there is a gap in a curved workpiece, if the offset is not corrected promptly even under the normal state, it can lead to an irrecoverable failure instance like this.
Figure 5 shows an instance where seam tracking has failed due to the previous reason. According to the formed bead shape, it can be seen that the torch fails to accurately recognize the V-groove center, which is the welding center line, and continues to be biased in one direction. This problem worsens as the welding progresses. As a result, the torch became biased and the welding ends as abnormal due to contact between the torch and the workpiece. The weaving-area-specific average feedback current values can show unpredictable values due to various environmental factors, such as the sampling cycle, sampling success rate, degree of curvature, gap size, and sensor accuracy. In other words, as in the example above, the existence of a gap in a curved workpiece or in workpieces with severe curvature means that the reliability of the area-specific average current values is reduced. Therefore, this study proposes a technique to correct the torch’s position more quickly and reliably through the detection of outliers in the feedback current values.
3.2. MBSC-Algorithm-Based Seam Tracking Technique for Automation of Welding in Unstructured Environments
This paper proposes the MBSC algorithm, a technique that enables effective seam tracking based on arc sensing, even in conditions where the feedback current changes drastically due to the control of short-circuit transition. Since actual welding works are carried out under challenging environmental conditions, such as in shipyards, the collection cycle of feedback current data for arc sensing can be quite irregular; that is, despite the feedback current data trend time-series characteristics, the time intervals can be somewhat irregular, making it difficult to apply typical time-series processing techniques such as moving averages. Therefore, the application of a clustering-based algorithm is required to effectively process and analyze irregularly collected data during the welding process.
The proposed technique is based on the DBSCAN clustering algorithm. DBSCAN performs density-based clustering within multidimensional datasets and defines center points based on a given (epsilon) value and the minimum (minimum number of points to form a cluster), and then connects them to form clusters. This method, when and are appropriately set, can effectively detect outliers in datasets with varying shapes and sizes of different clusters. The proposed technique uses DBSCAN to detect values that are outside the normal range in the weaving-area-specific feedback current data as outliers and compares the number of these outliers for each left and right area. As the distance between the torch and the workpiece narrows, the feedback current values become unstable, and sometimes control of short-circuit transition occurs. These conditions should be detected as outliers through clustering. In other words, the area where outliers are detected more frequently is considered to be closer to the workpiece, and by applying an offset in the opposite direction of that area, highly responsive seam tracking can be performed.
It should be noted that applying the proposed technique to the condition described in
Section 3.1 does not identify and avoid the unstable measurements when the gap is located at the weaving stay point since this phenomenon occurs due to the slow response speed of seam tracking. Therefore, by applying the proposed technique, the response speed for detecting curvature can be increased through control based on outlier detection, which enables the quick avoidance of the example condition itself.
Since the proposed technique is based on DBSCAN, the values of and must be set according to the distribution and characteristics of the data to effectively detect the appropriate outliers. In a structured and stable welding environment, the range of current values is constant and the measurement cycle is stable, so the appropriate value can be determined through experimentation. However, in actual welding environments, the range and distribution of current values change dynamically due to the conditions of the workpiece and the changes in the distance between the torch and the workpiece, and time intervals between the collected current values are irregular due to unstable measurements. Therefore, in order to detect the appropriate outliers in actual welding environments, it is necessary to find an value that can quickly respond to dynamic conditions rather than using a fixed value.
In the proposed technique, the value is newly searched for each weaving cycle that includes the left and right weaving areas once, and this determination process is executed every half cycle of weaving; that is, after deriving the left and right areas of cycle 1, the process next derives them for the right area of cycle 1 and the left area of cycle 2, and then searches for the left and right areas of cycle 2. By dynamically determining using only the current input values in this way, the responsiveness of seam tracking can be enhanced.
Algorithm 1 shows the process of determining
based on the median value for each of the two weaving areas and the procedure for applying MBSC based on that. The MBSC procedure takes the two target areas and the
value as input, calculates the range of current values
for each area, and then applies minmax scaling to normalize all values to be between 0 and 1. After that, appropriate
values,
and
, are calculated for each area. First, in lines 18 and 19, the median current value
m is found and all points with current values within 5% of this median are added to the median points group
. In lines 20 to 26, for each median point, the distance to all points within the area is calculated, and then the average distances to the
th nearest points
are calculated. For example, if
is 5, it calculates the average distance to the fourth most distance point for all points in the median points group. This average value is determined as
for that area. Afterwards, from lines 7 to 11, scaling operations are performed on
based on the range of current values for each area and, finally, clustering is performed using the determined
and
.
Algorithm 1 Determination of using median and MBSC |
Input: |
points as (time, current) pairs of current values measured from left and right weaving areas. |
Minimum number of neighbors required to form a cluster (). |
Output: |
// and with normalization and labeling performed. |
- 1:
procedure MBSC() - 2:
- 3:
- 4:
Perform minmax scaling on to make the values range from to . - 5:
- 6:
- 7:
if then - 8:
- 9:
else - 10:
- 11:
end if - 12:
- 13:
- 14:
return - 15:
end procedure - 16:
- 17:
function CalculateEps() - 18:
- 19:
- 20:
- 21:
for all do - 22:
- 23:
Sort ascending D and let - 24:
- 25:
end for - 26:
▹ Calculate the average of the -th distances - 27:
return - 28:
end function
|
The reason for preferring the median as a standard is that, even when the current values within a certain weaving areas are constant or show large amplitudes up and down, the median is likely to represent the normal current values in the area. Additionally, to maximize clustering formation and outlier detection ability, the minimum
value appropriate for the given
is determined. If
is too large compared to the level of
, there is a high likelihood that outliers will be included in other clusters, and if too small, many outliers may occur. Generally, for 2D data, it is recommended to set the
value to 4 to balance the cluster formation and outlier detection performance. However, in this study, since the main purpose is to detect outliers in the feedback current,
is set to 6 [
32]. The
value is set as the distance to the
th closest point from the median point, which ensures that the median point always belongs to a cluster.
With the value set as somewhat large, the process of determining the value becomes easier due to the influence of outliers, and the value can be greatly affected by the time index of the median. For example, if the point selected as the median is close in time index to an outlier, the likelihood increases that outliers will be included among the points up to . Alternatively, if the median point is located in a region with a sparse time index, an value could be derived that is large enough to include outliers in a cluster. Therefore, to reduce the influence of the median point’s time index, the Euclidean distance to the th nearest point is averaged using additional data within an appropriate range from the median, and this average is used as the value to decrease the possibility of distortion. At this time, when selecting the median point group, setting a range too narrow, around to of the median, does not sufficiently correct the distortion due to the median index. On the other hand, setting too wide a range, around , reduces the outlier detection performance. Hence, a range is set as appropriate. This range can vary depending on the weaving speed or sampling cycle.
The proposed technique applies DBSCAN based on the appropriately determined value to detect outliers in each of the left and right weaving areas, and compares the number of outliers in each area to apply an offset to the torch based on this comparison. However, since data normalization is performed for each area, there arises a problem where the scale of the base data from which each is derived as different. Therefore, to scale , it is necessary to adjust the value using the range of feedback current value changes in each area before performing DBSCAN. For example, a smaller range of current change means that the data are clustered near the normalized median value, and, in this case, it is desirable to detect fewer outliers compared to other areas. Accordingly, the value for the area with a smaller range of current change is increased proportionally compared to the area with a larger range, and then outliers are detected.
Algorithm 2 shows the process of performing seam tracking based on the outliers detected in each weaving area by the MBSC algorithm. Since the frequency of outliers is high in areas where the instability of the feedback current is relatively high, the position of the torch can be corrected and the weld seam can be accurately tracked by applying an offset in the opposite direction of the area. This approach enhances efficiency by enabling quick seam tracking even in conditions where the existing arc-sensing technique, which is based on the average current of each area, cannot effectively detect the torch’s bias due to the control of short-circuit transitions. In addition, if the frequency of outliers is similar in both areas, the existing arc-sensing technique, which performs seam tracking using the average feedback current, is applied to adjust the position of the torch more frequently. The difference in outlier frequency and the offset in MBSC-based seam tracking should be appropriately set according to the welding environment, and this approach improves the precision of seam tracking.
Algorithm 2 MBSC-Based ArcSensing |
Input: |
labeled current values including clustering result from MBSC. |
current location of the welding torch. |
offset for welding torch movement per iteration. |
threshold for outlier ratio to decide movement. |
Output: |
// updated location of the welding torch. |
- 1:
procedure MBSC-BasedArcSensing() - 2:
- 3:
- 4:
if and then - 5:
▹ Move torch left - 6:
else if and then - 7:
▹ Move torch right - 8:
end if - 9:
return - 10:
end procedure
|
3.3. Application of the MBSC Algorithm Based on Actual Data
In this section, we explain in detail how the proposed algorithm works using actually measured feedback current data. The data used are as shown in
Figure 6, which shows an instance where a left bias occurs in conditions where a gap exists between curved workpieces and the weaving stay point is just before reaching the gap. If this bias is not detected and the weaving stay point reaches the gap, the feedback current pattern becomes like that shown in
Figure 3b, making it difficult to normalize the welding operation with arc sensing alone. In conclusion, it is difficult for arc sensing based on weaving area average values to quickly respond to such bias in challenging and unstructured environments, leading to the failure of automatic welding. On the other hand, the proposed technique can effectively detect outliers even in these highly changing patterns, increasing the probability of successfully performing seam tracking.
Figure 7 shows the process of determining
for the left and right weaving area current data of the second weaving cycle seen in
Figure 6. After normalizing the current data for each area, the median point based on the current values is found. At this time, the current value of the median point of each area can be considered to represent the current value of that area. (For intuitive confirmation, the leftmost graph shows the raw input values before normalization, but, in reality, the median point is found after normalization.) It can be observed that the median point of the right area is biased to the right. If we try to calculate
only at this point, the result may be distorted by the time index. Therefore, data within
of the current value of the median point are selected as the median points group. Afterwards, the Euclidean distance to the
th closest point for each point in this group is calculated, and this is averaged to determine the
value. For the left area in the figure, the distance to the
th closest point from the median point is 0.281, but the arithmetic mean of the median points group is 0.302, so this value is determined as
for that area. The arithmetic mean of the distances to the
th closest point for each point in the median points group of the right area is 0.233, so this value is determined as
.
Next, it is necessary to match the reference scales of calculated in the two weaving areas. The range of current values in the left area, , is , and in the right area is . Since the range of current values in the right area is relatively small, it can be said that the current values in the right area tend to be more stable. Therefore, considering the scale of the left area, should be scaled up in the right area to detect outliers. Thus, the for the right area is finally determined as .
Figure 8 shows the results of performing DBSCAN for each weaving area using the
derived from the previously performed
determination process. The number of groups is ignored, and four outliers were detected in the left area and no outliers were detected in the right area. This means that more unstable feedback current values were measured in the left area compared to the right area, suggesting that the torch is closer to the workpiece in the left area when weaving. Therefore, by applying the proposed technique, an offset can be applied to move away from the left area based on these results. On the other hand, using average-value-based arc sensing on the same dataset would not be able to respond quickly to such unstable current value patterns.
The MBSC results for different weaving areas of
Figure 6 are shown in
Figure 9 and
Table 1. In
Figure 9, it can be seen that parts of the sudden increases and decreases in current values (left area of the second cycle) as well as the entirety (left areas of the first and third cycles) are detected as outliers. In
Table 1, it can be seen that outliers are appropriately detected even when determining
every half cycle. The proposed technique has thus been demonstrated to effectively respond to conditions with large-curvature workpieces or gaps in certain levels of curvature, such as unstructured environments where unstable current values are observed. In the next section, we will analyze in detail the applicability and performance of the proposed technique through experiments targeting various real unstructured environments.