METHOD FOR WELD SEAM OPTIMIZATION

Description

[0001] The invention relates to a method for predicting the quality of a manually produced weld during the production of the weld.

[0002] While welding robots can now produce highly predictable welds in many areas, this is not possible in all fields. Manual welding is still necessary, especially for specialized applications. One example is the welding of a submarine's pressure hull. These are typically one-off pieces or very small production runs. Furthermore, the result is safety-critical; a weld failure is unacceptable. Therefore, such processes usually involve a complete inspection of the finished weld, particularly using X-rays. If a defect is found in the weld, it typically has to be cut open and re-welded.If this happens, for example, with a weld connecting two assemblies of a submarine, the submarine is cut open again across its entire cross-section and welded back together, which is an enormous undertaking.

[0003] It would therefore be desirable to detect a welding defect as soon as it occurs, or even better, to receive a warning when a defect is likely to develop. In the first case, the weld seam that needs to be cut is shorter and less working time is lost; in the second case, such subsequent work can ideally be avoided altogether.

[0004] US Patent 2021 / 318673 A1 discloses an in-situ testing method based on a digital data model of the weld.

[0005] A system and a method for monitoring welding information are known from US patent 2015 / 379894 A1.

[0006] The object of the invention is to enable a method for predicting defects in manually produced welds as early as possible.

[0007] This problem is solved by the method with the features specified in claim 1. Advantageous further developments are described in the dependent claims, the following description, and the drawing.

[0008] The method according to the invention serves to predict the quality of a manually produced weld. It therefore concerns a weld that is produced manually by a welder. In contrast to welds produced by robots, for example in the automotive industry, such welds are subject to the human factor and can therefore exhibit variations.

[0009] The procedure includes a learning process a) and a monitoring process b).

[0010] The learning process a) comprises the following steps: a1) Creating a manual weld seam and thereby a1.1) Capturing the welding position, a1.2) Recording at least one welding parameter from the group including welding voltage, welding current, wire feed, shielding gas quantity, wire feed motor current, a2) Defect detection at the weld seam completed in step a1), a3) Using the data from steps a1) and a2) to train a network.

[0011] According to the invention, the network is an LSTM network.

[0012] The training process serves to train a network with data from a welding process and to assign error-free and error-free welding results to the data. For this purpose, data is continuously, selectively, or intermittently recorded and stored during the welding process. Simultaneously, the position of the welding process is also recorded or determined in such a way that the position can be assigned to the respective parameter set. Ideally, the position is determined simultaneously with the data, and both data sets are saved together. Alternatively, a later assignment can also be made.

[0013] Particularly for training purposes, it can be advantageous to intentionally create individual defects. In this case, purely spot welding can be performed. The recording of the welding position in step a1.1) then results directly from the created defect, so that the welding parameters can be assigned to the defect immediately and without an additional recording step.

[0014] The welding position is recorded in step a1.1) relative to the weld seam and only needs to be such that a subsequent correlation between the data recorded in step a1.2) and a defect detected in step a2) is possible. Assuming a constant welding speed, recording the starting position and the time would be sufficient to assign the events of the welding process to the weld seam. However, continuous position recording, which captures the position during the welding process, is also possible. Position recording can be optical (including infrared), ultrasonic, or radar-based.

[0015] If more than one welding parameter is recorded in step a1.2), which is preferred, then all recorded welding parameters are used in step a3).

[0016] Defect detection in step a2) is usually performed separately in a subsequent inspection step. For example, and preferably, the finished weld is examined using X-rays and manually evaluated, and defects are identified and assigned. If defects are detected, the weld usually has to be cut open and re-welded. The aim is precisely to avoid this time-consuming step as much as possible, or at least to reduce its complexity. Naturally, position assignment must also be possible in step a2). With imaging methods, the position is included in the image data. With other methods, the position can also be determined via a starting point and time or motion information. Similarly, position determination during defect detection can be performed in a manner comparable to the position acquisition in step a1.1).

[0017] In step a3), data from the acquisition step are at least partially fed into a network in order to train this network to detect defects.

[0018] For the purposes of this invention, a network is understood to be a machine learning device, in particular a neural network. Data is acquired by sensors on the workpiece or welding fixture and fed to a data processing device, which forms a neural network or a machine learning device that processes the data. Feeding data to the network, whether indicating errors or not, leads to network training and thus to the optimization of the network's error detection capabilities. An LSTM network is particularly suitable as a neural network for this process. LSTM stands for Long Short-Term Memory. LSTM networks have proven advantageous in image recognition and are also suitable for this process, even though the data acquired in step a1.2) is not image data.The preferred LSTM network is a convolutional neural network. The common English term for a convolutional neural network is "convolutional neural network." Therefore, the network is preferably an LSTM network.

[0019] In a preferred embodiment, the LSTM network is an LSTM with a low-pass filter. The behavior of the LSTM network could also be described as a low-pass filter. Transitions take some time but are more stable. Short-term changes within a single measurement are suppressed. This low-pass behavior is due to the structure of an LSTM cell. To reduce data volume, the acquired data can therefore be pre-filtered through a low-pass filter.

[0020] In a preferred embodiment, the LSTM network is an LSTM with wavelet transformation. As an effective data preprocessing method, wavelet analysis provides a time-frequency representation of signals with many different periods in the time domain. It can decompose time-series data into approximate and detailed parts to extract potential information from noisy data. The idea behind the wavelet transformation is to decompose the original sequence into different sub-sequences to obtain detailed information about the multiscale properties of time series. The overarching function of the wavelet transformation is to simultaneously represent information about the time, location, and frequency of a signal. The wavelet transformation is generally divided into continuous wavelet transformation (CWT) and discrete wavelet transformation (DWT).The processed and wavelet-decomposed sub-time series is used as input for the LSTM model to improve the output. Wavelet functions that can be used include, for example, Daubechies wavelets (dbN, where N represents the number of vanishing moments), Meyer wavelets, or Haar wavelets.

[0021] In a preferred embodiment, no labeled data is used. This is known as unsupervised learning. The advantage is that unlabeled data is typically available in significantly larger quantities.

[0022] The process therefore includes a machine learning step. Only in this way is it possible to later retrieve this knowledge in real time during the welding process.

[0023] The monitoring process b) comprises the following steps: b1) Creating a manual weld seam and thereby b1.1) Recording at least one welding parameter from the group including welding voltage, welding current, wire feed, shielding gas quantity, wire feed motor current b2) Using the network trained in step a3) and the welding parameters recorded in step b1.1) to predict whether a defect in the weld seam is to be expected at the welded area.

[0024] Preferably, the welding parameters recorded in step a1.2) and step b1.1) are identical. Particularly preferably, more than one welding parameter is recorded. The aim is to infer from fluctuations, for example in the welding current, especially in combination with the wire feed, a change in the amount of material introduced, which in turn could ultimately lead to a defect. The approach is that fluctuations that have already led to a defect can lead to another defect. However, since it is a manual process, it is subject to fluctuations anyway, which complicates a simple evaluation of only the welding parameters.

[0025] The network trained or optimized in step a) is used by feeding the welding parameters acquired during the weld creation process in step b) into the network. This input of welding parameters can occur during weld creation or only after welding is complete. The network evaluates this data to determine whether a defect is generated and outputs this evaluation as a result or prediction. Whether a defect actually exists can only be confirmed in a subsequent analysis process. For example, a fluctuation in the welding current can indicate a change in the distance of the welding machine. Therefore, if the welding machine is inadvertently moved away from the workpiece, the weld may become too weak, resulting in a defect.The basic idea of ​​the invention is to detect such a tendency, for example, an unintentional movement away from the workpiece, by means of a change, for example, in the welding current. Similarly, material variations in the workpiece can alter, for example, its conductivity or thermal conductivity, which also leads to a detectable change, for example, in the welding current. The crucial aspect is to recognize, through training, the fluctuations that are unavoidable, such as the digital noise in the measured values, and that have no influence on the result, and to detect only those events that do affect the quality of the weld.

[0026] The inventive method now makes it possible in situ to detect situations in which a defect occurs or situations that lead to a defect, and thus to prevent the formation of a defect. This makes it possible to at least partially reopen and remake the weld at an earlier stage, resulting in significantly less wasted work.

[0027] In a further embodiment of the invention, the learning process a) additionally includes the step: a1.3) Capturing at least one hand parameter of the welder, wherein the hand parameter is selected from the group including position and acceleration.

[0028] In step a3), the data from step a1.3) is also used. Thus, the network, and in particular the LSTM network, is also trained on these parameters.

[0029] The monitoring process also includes the following step: b1.2) Capturing at least one hand parameter of the welder, wherein the hand parameter is selected from the group including position and acceleration.

[0030] In step b2), the data from step b1.2) is also used.

[0031] Recording a manual parameter allows for significant improvement, as it captures and considers not only the machine-related values ​​but also the movement of the welder or welding equipment. This position can refer to the orientation in space, for example, in relation to a horizontal or vertical alignment, the weld seam, or at least one reference point. This is particularly advantageous because fluctuations in manual parameters can cause fluctuations in welding parameters, either compensating for or amplifying them.

[0032] The sensor for recording hand parameters can be integrated into the welding machine. Alternatively or additionally, it can be positioned directly on the welder's hand, for example, in the form of a glove. Multiple sensors can also be used, for example, on the welding machine and on the hand, to improve orientation tracking. An external sensor can also be used, for example, on the welding machine's power supply unit.

[0033] In a further embodiment of the invention, the position is captured as an additional hand parameter. For this purpose, a further additional capture device is preferably used. For example, the welder's hand (e.g., by means of a glove) and / or the welding machine can have markings for a motion capture method.

[0034] In a further embodiment of the invention, the welding machine has a distance sensor which detects the distance from the welding machine to the workpiece to be welded. The distance is recorded as a welding parameter in steps a1.2) and b1.1).

[0035] In a further embodiment of the invention, the learning process a) additionally includes the step: a1.4) Identifying the person of the welder or an identifier uniquely assigned to the person of the welder.

[0036] Recording only one identifier uniquely assigned to the welder enables data collection and use while avoiding the performance recording of an employee and thus avoiding legal problems.

[0037] The individual can be a crucial influencing factor. For example, there are right-handed and left-handed people. Therefore, it is to be expected that hand movements will differ (and, to a first approximation, be mirror images). It is also to be expected that, especially with curved components, problematic and thus defect-prone areas will be symmetrically aligned with each other. It is therefore helpful to consider such personal data and, for example, to take into account the welder's handedness.

[0038] In step a3), the data from step a1.4) are additionally used to train the network, in particular the LSTM network.

[0039] The monitoring process also includes the following step: b1.3) Identifying the person of the welder or an identifier uniquely assigned to the person of the welder.

[0040] In step b2), the data from step b1.3) is also used.

[0041] It may be possible to generate a separate dataset for each person recorded in step a1.4) for training the network, in particular the LSTM network. In step b2, the dataset of the person recorded in b1.3) can be accessed and used.

[0042] In a further embodiment of the invention, defect detection in step a2) is performed manually. This means that the defects are assigned in a supervised manner; it is therefore a supervised learning process. Simultaneously, another neural network could also be trained to automate defect detection, or at least the detection of suspected regions that may exhibit defects, or at least to minimize the time required for an expert to perform defect detection.

[0043] In a further embodiment of the invention, the result from step b1.2) is visually transmitted to the welder in real time. For example, it can be displayed on the helmet or protective visor. Since a welder always wears appropriate eye protection, this can be used very efficiently as a display. To minimize distraction for the welder, a preferred embodiment provides for the visual transmission of the result in the form of a color scale, for example, a traffic light (green = OK, yellow = Caution, red = Defect). The color scale can also display the risk level in greater detail using a smooth color gradient and is significantly easier to cognitively grasp than a numerical representation.It may also be provided that the visual representation of increased risks for defects is displayed for a slightly longer period than the event itself, in order to extend the perception period and thus reduce the risk of overlooking the event.

[0044] In a further embodiment of the invention, the method additionally comprises a prediction process c). The prediction process c) comprises the following steps: c1) Using the network trained in step a3) to predict weld quality for a prediction period. c2) Feedback to the welder if a defect is expected within the forecast period.

[0045] The aim is therefore to detect deviations that can immediately precede a defect. For example, a fluctuation in the welding voltage could lead to a change in the welding current, which the welder perceives as a change in heat generation during welding and may instinctively counteract. If the fluctuation in the welding voltage then subsides, the effect can be amplified by the welder's correction. Such effects are particularly well-suited for machine learning, making prediction possible, even if usually only for a very short prediction period. Therefore, prediction process c) is preferably carried out for a prediction period of up to 2 seconds. It has been shown that a time limitation positively influences the prediction quality. Logically, the prediction process is executed within monitoring process b).

[0046] In a further embodiment of the invention, the welding parameters welding voltage, welding current, wire feed rate, shielding gas quantity, and wire feed motor current are recorded in steps a1.2) and b1.1). Recording as many welding parameters as possible enables the most precise possible recognition of comparable situations. All of these parameters have proven to be useful and complementary.

[0047] In a further embodiment of the invention, the welder's hand parameters, position and acceleration, are recorded in step a1.3) and in step b1.2). Here too, recording both parameters is advantageous in order to improve the data basis and thus facilitate the most precise possible recognition of comparable situations.

[0048] In a further embodiment of the invention, the data are acquired in steps a1.2) and b1.1) as well as in the optional steps a1.3) and b1.2) within an interval of 5 ms to 100 ms, preferably within an interval of 15 ms to 50 ms. This has proven to be a time window in which significant data are acquired without processing unnecessary data clutter (for example, due to statistical noise at extremely short time intervals).

[0049] In a further embodiment of the invention, defect detection at the finished weld seam in step a2) is carried out using X-rays. An X-ray examination of the weld seam is thus performed, and the result is evaluated, particularly by a person skilled in the art, and defects are located.

[0050] In a further embodiment of the invention, the monitoring process b) additionally includes the following step: b1.4) Capturing the welding position.

[0051] Analogous to the welding position recording in step a1.1), the welding position must also be recorded in such a way that a subsequent correlation between the data recorded in step b1.4) and a defect detected in the inspection process d) is possible. Assuming a perfectly constant welding speed, recording the starting position and the time would be sufficient. Position detection can be performed optically (including infrared), by ultrasound, or radar.

[0052] This allows for a later comparison between the data and detected defects.

[0053] In a further embodiment of the invention, the monitoring process b) is followed by a verification process d). Quality control of the finished weld is typically performed as standard practice for products where failure would be unacceptable, for example, the pressure hull of a submarine. The verification process d) comprises the following steps: d1) Defect detection on the weld seam completed in step b1).

[0054] Following this standard product quality check, the network undergoes ongoing training, including an additional training step: d2) Using the data from steps b1) and d1) to retrain the network trained in step a3), in particular the LSTM network.

[0055] In a further embodiment of the invention, additional training data is used in step a3). This additional training data is generated from the data from steps a1) and a2). Preferably, this additional training data is generated only for defects. For example, and preferably, the additional training data is generated from the original data by applying a method selected from the group comprising imprinting noise, shifting at least one series of measurements, and applying a window function to the original data. For example, the originally measured welding current data is taken, and a new data set is generated in which noise is added to the welding current. Shifting at least one series of measurements means that the numerical value is changed uniformly, i.e., for example, all values ​​of the welding current are halved.The von Hann window has proven to be a particularly suitable window function. The von Hann window is based on a superposition of three spectrally shifted silicon functions to achieve stronger suppression of sidelobes compared to the rectangular window with only one silicon function in the spectrum. The disadvantage is a reduction in frequency resolution.

[0056] In a further embodiment of the invention, the network trained in step a3), in particular the LSTM network, is optimized to avoid identifying non-existent defects as defects; that is, it is about avoiding the detection of a supposed defect that does not actually exist. Such methods always require optimization that balances falsely missed defects (false acceptance, weld incorrectly deemed acceptable) against falsely detected but non-existent defects (false rejection, weld incorrectly rejected as defective). For the problem at hand, the method according to the invention is intended to detect defects early so that they can be corrected with less effort. If a defect is not detected, it will be noticed during final inspection, and nothing is lost compared to not using the method according to the invention.However, if a defect were falsely detected where none exists, this would result in additional effort. Therefore, the process must be optimized to minimize the false rejection rate.

[0057] The method according to the invention is explained in more detail below with reference to an embodiment shown in the drawing. Fig. 1 Flowchart

[0058] In Fig. 1 A flowchart of the method according to the invention is shown in an exemplary embodiment.

[0059] The process is roughly divided into two steps: the learning process a) and the monitoring process b).

[0060] During the training process, a welder manually creates a weld seam (10). The current position is recorded (11). The welding parameters (12) welding voltage, welding current, wire feed rate, shielding gas flow rate, wire feed motor current, the manual parameters (13) position and acceleration, and the welder's identity or an identifier uniquely assigned to the welder (14) are also recorded. While the current position, the welding parameters (12) welding voltage, welding current, wire feed rate, shielding gas flow rate, wire feed motor current, and the manual parameters (13) position and acceleration are continuously recorded, the welder's identity or an identifier uniquely assigned to the welder (14) can be recorded once, for example, through a registration process, such as before starting work.

[0061] In a second step, defects (20) are detected on the finished weld. This involves, in particular, X-ray imaging (21) of the weld, which is then evaluated (22). Defects are identified accordingly.

[0062] The data from the weld generation 10 and the defect detection 20 are combined to train an LSTM network 30. If additional training data is to be generated, for example, white noise can be added to the data acquired during the generation of a manual weld 10 to increase the number of training data sets for defects.

[0063] The LSTM network trained in this way is then used in monitoring process b). Here, a weld is manually created (40). The current position (41), the welding parameters (42) welding voltage, welding current, wire feed rate, shielding gas quantity, wire feed motor current, the manual parameters (43) position and acceleration, and the welder's identity or a unique identifier (44) are recorded. Using the currently recorded data, a real-time prediction (50) of the expected weld quality is then generated. The result can be output to a device (51), such as a display in the welder's protective shield. The output can be, for example, a red-yellow-green color code or simply a warning message indicating when an error is expected (i.e., only red).

[0064] To continuously improve the system, the data acquired during the weld quality inspection and defect detection process 60, together with the data acquired during the manual weld creation process 40, are used again to train the LSTM network 30. For example, and in particular, defect detection 60 again includes X-ray acquisition 61 and subsequent manual evaluation 62 of the acquired X-ray data. Reference sign

[0065] a) Learning process b) Monitoring process 10 Creating a manual weld 11 Recording the welding position 12 Recording the welding parameters 13 Recording the hand parameters 14 Recording the person 20 Defect detection 21 X-ray acquisition 22 Evaluation 30 Training the network 40 Creating a manual weld 41 Recording the welding position 42 Recording the welding parameters 43 Recording the hand parameters 44 Recording the person 50 Prediction 51 Output 60 Defect detection 61 X-ray acquisition 62 Evaluation 70 Generating additional training data

Claims

1. Method for predicting the quality of a manually-produced weld seam, wherein the method comprises a training process a) and a monitoring process b), wherein the training process a) comprises the following steps: a1)creating a manual weld seam and in so doinga1.1)identifying the welding position (11),a1.2)acquiring at least one welding parameter (12) from the group comprising welding voltage, welding current, wire feed rate, amount of protective gas, wire rotary motor current,a2)defect detection (20) at the weld seam produced in step a1),a3)using data from step a1) and a2) for training a network (30), wherein the network is an LSTM network, wherein the monitoring process b) comprises the following steps: b1)creating a manual weld seam (40) and in so doingb1.1)acquiring at least one welding parameter (42) from the group comprising welding voltage, welding current, wire feed rate, amount of protective gas, wire rotary motor current,b2)using the network trained in step a3) and the welding parameters acquired in step b1.1) for the prediction (50) as to whether a defect in the weld seam is likely to occur at the particular point where welding has just been carried out.

2. Method according to Claim 1, characterized in that the training process a) comprises in addition the step: a1.3)acquiring at least one hand parameter (13) of the welder, wherein the hand parameter is selected from the group comprising position and acceleration, wherein the data from step a1.3) is used in addition in step a3), wherein the monitoring process comprises in addition the step: b1.2)acquiring at least one hand parameter (43) of the welder, wherein the hand parameter is selected from the group comprising position and acceleration, wherein the data from step b1.2) is used in addition in step b2).

3. Method according to one of the preceding claims, characterised in that the training process a) comprises in addition the step: a1.4)acquiring the welder's details or an identifier (14) uniquely assigned to the particular welder, wherein the data from step a1.4) is used in addition in step a3), wherein the monitoring process comprises in addition the step: b1.3)acquiring the welder's details or an identifier (44) uniquely assigned to the particular welder, wherein the data from step b1.3) is used in addition in step b2).

4. Method according to one of the preceding claims, characterised in that defect detection (20) is performed manually in step a2).

5. Method according to one of the preceding claims, characterised in that the result from step b1.2) is visually transmitted to the welder in real time.

6. Method according to Claim 5, characterised in that the result is visually transmitted in the form of a colour scale.

7. Method according to one of the preceding claims, characterised in that the method comprises in addition a prediction process c), wherein the prediction process c) comprises the following steps: c1)using the network trained in step a3) to predict weld seam quality,c2)feeding back to the welder if a defect is to be expected.

8. Method according to Claim 7, characterised in that the prediction process c) is performed for a time period of up to 2 s.

9. Method according to one of the preceding claims, characterised in that the welding parameters comprising welding voltage, welding current, wire feed rate, amount of protective gas and wire rotary motor current are acquired in step a1.2) and in step b1.1).

10. Method according to Claim 2 or a claim that refers back to Claim 2, characterised in that the hand parameters of the welder comprising position and acceleration are acquired in step a1.3) and in step b1.2).

11. Method according to one of the preceding steps, characterised in that data is acquired in the steps a1.2) and b1.1) and in the optional steps a1.3) and b1.2) in an interval of 5 ms to 100 ms, preferably in an interval of 15 ms to 50 ms.

12. Method according to one of the preceding claims, characterised in that defect detection (20) is performed on the finished weld seam in step a2) with the aid of X-rays.

13. Method according to one of the preceding claims, characterised in that the monitoring process b) comprises in addition the following step: b1.4)identifying the welding position (12).

14. Method according to Claim 13, characterised in that an inspection process d) follows the monitoring process b), wherein the inspection process d) comprises the following steps: d1)defect detection (40) at the weld seam produced in step b1),d2)using the data from steps b1) and d1) to retrain the network (30) trained in step a3).

15. Method according to one of the preceding claims, characterised in that additional training data is used in step a3), wherein the additional training data is generated from the data from step a1) and a2).

16. Method according to Claim 15, characterised in that the additional training data is only generated for defects.

17. Method according to one of claims 15 to 16, characterised in that the additional training data is generated (70) by a method selected from the group comprising noise injection, shifting at least one series of measured values, applying a window function.

18. Method according to one of the preceding claims, characterised in that the network trained in step a3) is optimised to avoid identifying a non-existent defect as a defect.