ARC WELDING SYSTEM AND ASSESSMENT DEVICE FOR ARC WELDING

Abstract

Arc welding system (1) comprising a robot (4), an arc welding section (41) which is transportably attached to the robot (4) and is configured to perform an arc welding process, an arc welding power source (5) which is configured to supply the arc welding section (41) with a specific welding voltage and a specific welding current, an analysis section (61) which is configured to perform a frequency analysis on a time series waveform of at least one of the welding voltage and the welding current and to generate frequency spectrum data, and an evaluation section (62) which is configured to evaluate a welding condition of the arc welding section (41) on the basis of the frequency spectrum data generated by the analysis section (61), wherein the analysis section (61) is designed to perform a frequency analysis on the time series waveform which has been subjected to convolution using a window function, the assessment section (62) is designed to assess the presence of a welding defect by comparing the analysis frequency spectrum data generated by the analysis section (61) with the normal frequency spectrum data for the case of a normal welding condition, the assessment section (62) is designed to assess the presence of a given error category by comparing the analysis frequency spectrum data in a frequency band of the frequency range corresponding to the error category in question with the normal frequency spectrum data.

Description

Technical field

[0001] Disclosed embodiments relate to an arc welding system and an assessment device for arc welding. General state of the art

[0002] Patent document 1 discloses a method for assessing defects in the welding condition by means of statistical analysis of means, distributions, reference deviations, and the like in welding parameters such as arc voltage, arc current, or the like. JP 2016-26 878 A shows an arc welding system in which a frequency analysis is performed. State-of-the-art documents, patent documents

[0003] Patent document 1: Japanese utility model registration JP 3 197 100 U Brief description of the invention; Problem of the present invention

[0004] However, the time series waveform of welding voltage and welding current is easily influenced by noise from other factors, which is why sufficient detection accuracy of minor waveform changes, as required for fault assessment of the welding condition, cannot be achieved, resulting in low fault assessment accuracy.

[0005] The present invention was made because of this problem, and its objective is to provide a control device and a control method that can increase the suitability of a control device with which the assessment performance of the welding condition can be improved. Means of solving the task

[0006] To achieve the aforementioned objective, an arc welding system according to claim 1 is provided. According to one aspect of the present invention, an arc welding system is used which comprises: a robot, an arc welding section which is transportably attached to the robot and configured to perform an arc welding process, an arc welding power source which is configured to supply the arc welding section with a specific welding voltage and a specific welding current, an analysis section which is configured to perform a frequency analysis on a time series waveform of at least one of the welding voltage and welding current and to generate frequency spectrum data, and an evaluation section which is configured to evaluate the welding condition of the arc welding section based on the frequency spectrum data generated by the analysis section.

[0007] According to a further aspect of the present invention, an evaluation device for arc welding is also used, comprising: an analysis section configured to be detected on a time series waveform of a welding voltage or welding current, which are configured to be from an external arc welding power source, to perform a frequency analysis and to generate frequency spectrum data, and an evaluation section configured to assess, on the basis of learning content from a mechanical learning process, the presence of a fault of a fault category corresponding to a characteristic value exhibited by the frequency spectrum data in at least one of a time domain and a frequency domain. Effect of the invention

[0008] According to the present invention, the performance in assessing the weld condition can be improved. Brief description of the characters

[0009] They show: Fig. 1 an example of a schematic system block structure of a robot arc welding system according to one embodiment; Fig. 2 an example of time series waveform data and frequency spectrum data of a welding current during normal welding; Fig. 3 an example of time series waveform data and frequency spectrum data of a welding voltage during normal welding; Fig. 4 an explanatory view of a procedure for generating frequency spectrum data at the analysis section; Fig. 5 an example of time series waveform data and frequency spectrum data of a welding current in the case of a welding crack defect; Fig. 6 an example of time series waveform data and frequency spectrum data of a welding stress in the case of a welding crack defect; Fig. 7 an example of time series waveform data and frequency spectrum data of a welding current in the event of a gas shortage fault; Fig. 8 an example of time series waveform data and frequency spectrum data of a welding voltage in the event of a gas shortage fault; Fig. 9 an example of a schematic model structure of a neural network of the assessment section when applying deep learning; Fig. 10 a view of an individual data set of the assessment section; and Fig. 11 an example of a schematic model setup of a neural network of the assessment section when applying deep learning, taking into account the welding wire feed rate. Embodiments of the invention

[0010] In the following, a first embodiment of the present invention will be described with reference to the figures. <Schematischer Aufbau des Roboterlichtbogenschweißsystems>

[0011] Fig. Figure 1 shows an example of a schematic system block structure of a robotic arc welding system according to one embodiment. The workpiece of the robotic arc welding system is a vehicle frame F, and the system performs an arc welding operation along a specific weld line of the frame F. The robotic arc welding system 1 (arc welding system) consists of Fig. 1 comprises a superior control device 2, a robot controller 3, a robot 4, a welding power source 5, and an evaluation device 6 for arc welding. In the example of the present embodiment, the case is shown where the vehicle frame F forms the workpiece; however, another machine structure can also form the workpiece, and it is also possible to apply it to arc welding operations where the arc welding line is a weld seam or the like between elements of such a workpiece.

[0012] The higher-level control device 2 can, for example, be a universal PC comprising a CPU, ROM, RAM, an operator panel, a display panel (each not shown separately), and the like, and manages the operation of the robot arc welding system 1 as a whole. Specifically, it issues certain work commands to the robot controller 3, feed rate commands to a welding wire feed section of the robot 4 (described later), or power supply commands to a welding power source 5 (described later) so that the latter outputs a specific welding energy. Alternatively, based on evaluation information input by the arc welding evaluation device 6 (described later), it performs a final inspection, interrupts, or resumes the arc welding operations.

[0013] Based on work commands entered by the higher-level control device 2, the robot controller 3 performs a power supply control for electrical drive power to drive shaft motors (not shown) of the robot 4 in order to move a front end of an arc welding section, described later, attached to an arm front end section 4a of the robot 4, along a specific welding line on the frame F.

[0014] In the illustrated example of the present embodiment, the robot 4 is a manipulator arm with six articulating axes (a 6-axis robot). The arc welding section 41 is attached to its front arm end section 4a, and the position and attitude of the arc welding section 41 can be controlled by driving the drive shaft motors of the robot 4 with the electrical drive power supplied by the robot controller 3.

[0015] The arc welding section 41 is a gripping element that, in this example, continuously feeds a welding wire 43 as filler metal along with a shielding gas and performs a short-circuit welding process on the base metal frame F using electrical welding energy supplied by the welding power source 5 described later. The welding wire 43 is continuously advanced by a welding wire feed section 42, also provided on the robot 4. The welding wire feed section 42 advances the welding wire 43 to the arc welding section 41 at a feed rate based on feed rate commands from the higher-level control device 2 and outputs the actual measured feed rate of the welding wire 43 to an evaluation section 62 of the evaluation device 6 for arc welding, which is described later.The robot 4 also has a shielding gas supply section (not shown) which generates shielding gas and supplies it to the arc welding section 41 based on commands from the higher-level control device 2.

[0016] The welding power source 5 (arc welding power source) is a device that converts electrical current from a commercial power source and, based on power supply commands from the higher-level control device 2, generates electrical welding energy and supplies electrical energy between the arc welding section 41 and the frame F. In the illustrated example, a direct current is supplied as the electrical welding energy, with the arc welding section 41 being the anode side and the frame F the cathode side. Furthermore, the welding power source 5 outputs detection values ​​of the actual supplied welding current and the actual welding voltage as welding current detection values ​​and welding voltage detection values ​​to the analysis section 61 of the arc welding evaluation device 6, which is described later.

[0017] The arc welding assessment device 6 is a device that, based on at least one of the welding current and welding voltage measurement values ​​input by the welding power source 5, assesses whether a defect exists in the welding condition of the welding work performed by the arc welding section 41. This arc welding assessment device 6 comprises the analysis section 61 and an assessment section 62.

[0018] Analysis section 61 is a processing section that generates frequency spectrum data by performing a frequency analysis on time-series waveform data in which the welding current or welding voltage measurement values ​​are represented in a time series. The procedure used in analysis section 61 and the content of the frequency spectrum data are described later.

[0019] Assessment section 62 is a processing section that uses the frequency spectrum data generated by analysis section 61 to determine whether a specific category of weld defects is present. The procedure used in assessment section 62 is described later.

[0020] According to the robotic arc welding system 1 with the setup described above, an automatic arc welding process can also be carried out on welding lines with various three-dimensional locations on a vehicle frame F. By automatically assessing whether a defect exists in the welding process, the arc welding evaluation device 6 enables the higher-level control device 2 to take appropriate measures when a defect occurs. <Merkmale der vorliegenden Ausführungsform>

[0021] For several years, there has been a demand for arc welding systems that perform arc welding processes while, as described above, an arc welding section attached to a robot moves along a weld line (weld seam). However, arc welding requires adherence to high standards in various conditions, such as workpiece positioning (e.g., weld overlap ratio), distance between the workpiece and the filler material (in this example, the welding wire 43), filler material or shielding gas supply, and the workpiece's ambient temperature. Otherwise, the weld quality is severely compromised, making it an extremely demanding process. Therefore, it is crucial to assess the weld quality when performing arc welding with a robot.

[0022] Methods for assessing weld defects have already been developed by performing a statistical analysis of means, distributions, reference deviations, and the like on the time-series waveform of welding voltage or welding current supplied by a welding power source to an arc welding section (i.e., an analysis exclusively in the time domain). However, the time-series waveform of welding voltage and welding current is easily influenced by noise from other factors, which is why it is not possible to achieve sufficient accuracy in detecting minor waveform changes, as is necessary for assessing weld defects. Therefore, the assessment is difficult depending on the type of defect.

[0023] Therefore, in the present embodiment, the analysis section 61, which performs a frequency analysis on the time-series waveform of welding voltage or welding current output by the welding power source 5 and generates frequency spectrum data, and the evaluation section 62, which assesses the welding condition of the arc welding section 41 based on the frequency spectrum data generated by the analysis section 61, are provided. In this way, a phase analysis of the time-series waveform of welding voltage or welding current is also made possible, and in particular, by observing the temporal change of the frequency spectrum in the frequency spectrum data in connection with welding condition defects, noise can be detected and thus a high degree of defect assessment accuracy can be ensured. <Verfahrensweise der Analyse durch den Analyseabschnitt>

[0024] Fig. Figure 2 shows an example of time series waveform data of the welding current during normal welding and frequency spectrum data generated on the basis of this data by analysis section 61, and Fig. Figure 3 shows an example of time-series waveform data of the welding voltage during normal welding and, based on this, frequency spectrum data generated by analysis section 61. The respective in Fig. 2 and Fig. The three time-series waveform data shown below are waveform data obtained by sequentially recording welding current and welding voltage values, acquired at a specific sampling frequency during a stability period after the start of welding operations by the arc welding section 41, in a time series. In the time-series waveform data, the horizontal axis represents time and the vertical axis represents welding current and welding voltage, respectively, and a waveform of an approximately cyclically repeated discharge can be observed.

[0025] Since the discharge phenomenon occurs naturally during arc welding with short-circuit welding, there are, as can be seen from the illustration, strong fluctuations between the discharge cycles and the cycle waveforms in the time series waveform data. Therefore, even with separate frequency analysis, no frequency spectrum data can be obtained between time series waveforms sampled in the same cycle that can be meaningfully compared.

[0026] Therefore, in the present embodiment, the analysis section 61 generates frequency spectrum data based on the Fig. 4. Procedure shown. As shown furthest left in Fig. As shown in Figure 4, when performing a frequency analysis of the time-series waveform data at an arbitrary time t1, the frequency analysis is carried out using an anti-aliasing filter to avoid aliasing errors that can occur during sampling. In other words, the frequency analysis is performed on a waveform where the power with respect to sampling has been sufficiently attenuated in a frequency range at or above the Nyquist frequency.

[0027] Furthermore, in the present embodiment, a moving average of the waveform around the sampling time t1 is weighted using a specific window function, and the data cycle is extended. Numerous types of window functions are known for waveform analysis; here, a window function of a type corresponding to the welding conditions during the arc welding process is applied to the time-series waveform data. In the example shown in the present embodiment, the welding conditions are such that the welding method is short-circuit welding and the welding current is 250 A or more, which is why a so-called Gaussian window is applied.As shown in the figure by the bold solid line, this Gaussian window is a window function whose midpoint is the maximum value 1.0 at sampling time t1 and which decreases non-linearly, and by adding (weighting) the Gaussian window value at each time point, harmonic distortion is prevented by performing a cyclic extension of the data and a Fourier transform (time frequency analysis).

[0028] As shown in the second position from the left, a frequency analysis by Fourier transformation is performed on the waveform at time t1, which has undergone cyclic data extension processing, using FFT or a similar method. Furthermore, a superposition and convolution of the spectrum data obtained at each time point is performed with a time shift, thus gradually conducting a frequency analysis for the time domain of the time-series waveform data. By arranging the frequency spectrum data in the time series, as shown on the far right of the figure, frequency spectrum data are obtained as two-dimensional mapping data (two-dimensional pixel sequence data), as shown in the respective figures. Fig. 2 and Fig. 3 shown. Regarding the frequency spectrum data from Fig. 2 and Fig. 3 is the horizontal axis, just as in the corresponding time series waveform data, the time axis (corresponding to the time domain), and the vertical axis is the frequency (corresponding to the frequency domain), where the thickness at the individual points represents the spectrum density.

[0029] Fig. Figures 5 to 8 show examples of time series waveform data and frequency spectrum data obtained using the above acquisition and analysis methods in the case of a weld crack defect or a gas deficiency defect. Fig. Figure 5 shows an example of time series waveform data and frequency spectrum data of the welding current in a welding crack defect where the weld bead breaks during arc welding operations and a hole is created. Fig. Figure 6 also shows an example of time series waveform data and frequency spectrum data of the welding stress in the case of a welding crack defect. Fig. Figure 7 shows an example of time series waveform data and frequency spectrum data of the welding current in the case of a gas shortage fault, where there is insufficient shielding gas during arc welding operations. Fig. Figure 8 also shows an example of time series waveform data and frequency spectrum data of the welding voltage in the event of a gas shortage fault. <Verfahrensweise der Beurteilung durch den Beurteilungsabschnitt>

[0030] When considering the frequency spectrum data from Fig. 2 and Fig. In the three data points corresponding to normal welding, it can be seen that in both frequency ranges, the spectrum density is comparatively high in the band range from 0 to 140 Hz, while in higher band ranges, the spectrum density gradually decreases with increasing frequency. It can also be seen that in both frequency spectrum data types, the frequency spectrum shows no fluctuations in the time domain and is stable.

[0031] If, on the other hand, the frequency spectrum data is taken from Fig. 5 and Fig. 6. Considering the data corresponding to a weld crack defect, it can be seen that the frequency spectrum data fluctuates intermittently in the time domain, and the spectrum density at the time of these fluctuations decreases significantly across the entire frequency range compared to normal welding (see the dotted area in the figures). Thus, the characteristic is that, compared to normal welding, large fluctuations occur in the frequency spectrum data in the time domain when a weld crack defect occurs.

[0032] By detecting this feature in the frequency spectrum data, the evaluation section 62 in the example of the present embodiment judges that a weld crack defect has occurred. That is, frequency spectrum data during normal welding (hereinafter referred to as "normal frequency spectrum data") are stored in advance, and by comparing them in the time domain with the frequency spectrum data generated by the analysis section 61 during arc welding (hereinafter referred to as "analysis frequency spectrum data"), it is judged whether a weld crack defect is present. Specifically, for example, the normal frequency spectrum data and the analysis frequency spectrum data are compared at several points in time in the time domain, and if a large deviation between them is continuously detected for at least a certain period, it can be judged that a weld crack defect has occurred.As a method for comparing the normal frequency spectrum data and the analysis frequency spectrum data at the respective times, for example, the relationship between spectrum density waveforms at a given time is determined across the entire frequency range, or the mean or sum value of the spectrum density across the entire frequency range at a given time is compared with an index value, etc.

[0033] When considering the frequency spectrum data from Fig. 7 and Fig. 8, which correspond to a gas shortage error, it can be seen that in both frequency ranges within a specific frequency band (in the example shown at 10) 2(Hz) the spectrum density is high compared to normal welding (see the dotted area in the figures). Thus, it can be observed that, compared to normal welding, large fluctuations occur in the frequency spectrum data when a gas shortage fault occurs.

[0034] By detecting this feature in the frequency spectrum data, the evaluation section 62 in the example of the present embodiment judges that a gas shortage fault has occurred. By comparing the normal frequency spectrum data and the analysis frequency spectrum data in the frequency domain, it is thus determined whether a gas shortage fault exists. Specifically, for example, the normal frequency spectrum data and the analysis frequency spectrum data are compared in several frequency bands of the frequency domain, and if a large deviation between them is continuously detected for at least a certain period of time, it can be judged that a gas shortage fault has occurred.As a procedure for comparing the normal frequency spectrum data and the analysis frequency spectrum data in a specific frequency band, for example, the relationship between spectrum density waveforms is determined, restricted to that frequency band at that time, or the mean or sum value of the spectrum density in that frequency band at that time is compared with an index value, etc.

[0035] Although not shown, it is intended that for welding defect categories other than weld crack defects or gas deficiency defects, a feature of the analysis frequency spectrum data will also be evident in the time domain or frequency domain, and that this feature will manifest itself in different temporal fluctuation patterns or different frequency bands. Assessment section 62 records these features according to their temporal fluctuation pattern or frequency band, thus enabling a separate defect assessment for each welding defect category. <Wirkung der vorliegenden Ausführungsform>

[0036] As described above, the robotic arc welding system 1 of the present embodiment comprises the analysis section 61, which performs a frequency analysis on the time-series waveform of the welding voltage or welding current output by the welding power source 5 and generates frequency spectrum data, and the evaluation section 62, which assesses the welding condition of the arc welding section 41 based on the frequency spectrum data generated by the analysis section 61. In this way, a phase analysis of the time-series waveform of the welding voltage or welding current is also possible, and in particular, by observing the temporal changes in the frequency spectrum in the frequency spectrum data in relation to welding condition defects, noise can be detected, thus ensuring high defect assessment accuracy. In this way, the performance in assessing the welding condition can be improved.

[0037] In examining the technique, the inventors discovered that by performing a frequency analysis on the time-series waveforms, on which analysis section 61 had performed a convolution using the window function, the characteristics of waveform fluctuations, which should be captured during the various assessments during arc welding, become particularly evident in the frequency spectrum data. That is to say, it was found that a method previously used only for waveform analysis of human voiceprints can also be advantageously applied to the frequency analysis of the time-series waveforms of welding voltage or welding current during arc welding.In this way, the accuracy of the error assessment for the time series waveforms recorded during the arc welding process can be further increased in the present embodiment by performing the frequency analysis not only on the basis of a single arc discharge cycle, but also by performing the frequency analysis under smoothing by the moving average weighted by the window function around its time range.

[0038] In particular, the present embodiment uses a window function adapted to the welding conditions. Welding conditions, such as the welding technique (short-circuit welding, pulse welding, etc.) or the welding voltage and current, strongly influence the large-droplet material transfer during arc welding, i.e., the waveform of the time-series waveforms recorded during the arc welding process, and thus also the relationship between successive arc discharge cycles in the time domain. Therefore, applying a window function that is appropriately selected according to the welding conditions means that a good balance between frequency resolution and dynamic range can be achieved in the frequency analysis, thereby increasing the accuracy of the weld condition assessment under a wide variety of welding conditions.

[0039] The present embodiment was described using the example of welding conditions such that the welding technique is short-circuit welding and the welding current is at least 250 A. Therefore, a Gaussian window is used as a suitable window function, but there is no limitation in this respect. The accuracy of the error assessment can also be increased by, for example, using a window function (not shown) suitable for welding conditions where the welding technique is pulse welding and there are differences in the welding current and welding voltage corresponding to the change in the large-droplet material transfer.

[0040] In the present embodiment, the evaluation section 62 assesses the presence of a welding defect, in particular by comparing the analysis frequency spectrum data generated by the analysis section 61 with the normal frequency spectrum data for a normal welding condition. This enables a defect assessment of the welding condition through a relatively simple comparative processing.

[0041] In the present embodiment, the evaluation section 62 also assesses, in particular, the presence of a respective fault category by comparing the analysis frequency spectrum data in a frequency band of the frequency range corresponding to the fault category in question (gas shortage fault, etc.) with the normal frequency spectrum data. This enables an assessment of the presence of a fault with a clear distinction between the several fault categories, whose respective fault characteristics are only clearly apparent in the frequency range of the frequency spectrum data, while avoiding unnecessary comparative processing in frequency bands that are not related to the fault assessment.

[0042] In the present embodiment, the evaluation section 62 assesses, in particular, the presence of a defect of a specific defect category (weld crack defects, etc.) by comparing the analysis frequency spectrum data in the time domain with the normal frequency spectrum data. This enables the assessment of the presence of a defect for defect categories where the defect characteristic is clearly evident exclusively in the time domain of the frequency spectrum data.

[0043] Furthermore, in the embodiment described above, a distinction is made between welding current and welding voltage, and a welding defect assessment is performed for one of them (or for both) by comparison with the frequency spectrum data; however, there is no limitation in this respect. Analysis section 61 can also perform a frequency analysis of the time series waveform (not shown) for one of the impedances (voltage / current) and welding energy (voltage × current) determined from the welding voltage and welding current. In this way, assessment section 62 can perform the assessment processing simultaneously based on both the welding voltage and the welding current. <abwandlungsbeispiele>

[0044] The embodiment described above can be modified in various ways within its scope and technical concept. <Abwandlungsbeispiel 1: Schweißfehlerbeurteilung durch mechanisches Lernen>

[0045] In the embodiment described above, the evaluation section 62 assesses the presence of a welding defect by comparing the normal frequency spectrum data and the analysis frequency spectrum data. The evaluation section 62 can also perform the welding defect assessment using mechanical learning in an example that is not, in itself, according to the invention.

[0046] The assessment procedure of assessment section 62 can be applied to a wide variety of mechanical learning methods, with the following describing, as an example, the application of deep learning to a mechanical learning algorithm. Fig. Figure 9 shows an example of a schematic model setup of a neural network for assessment section 62 using deep learning. Fig. 9 is a neural network of the assessment section 62 designed such that, with regard to analysis frequency spectrum data which are entered into the assessment section 62 as two-dimensional mapping data, the presence of a welding defect, which can be assessed on the basis of a feature that emerges in the analysis frequency spectrum data, and the corresponding category are output in the form of assessment information.

[0047] In the example shown, the analysis frequency spectrum data is inputted to an input node, while the output is a binary value. Only one of the three output nodes provides a "true" value for the welding states: normal welding, weld cracking, and gas deficiency (a so-called clustered output). When inputting the analysis frequency spectrum data, the spectrum density (light and dark in the figure) at each pixel is entered into the input node as two-dimensional mapping data, maintaining the relationship between the time and frequency axes. This evaluation processing is based on learning content from a learning phase of evaluation section 62 in a mechanical learning process. That is, the neural network of evaluation section 62 learns feature sizes that indicate a relationship between the spectrum density scheme and the welding state in the time or frequency domain of the analysis frequency spectrum data.

[0048] For this mechanical learning process of the assessment section 62, the assessment device for arc welding 6 is equipped in the form of software (or hardware) with a multi-layer neural network designed as described above. Following this, learning into the assessment section 62 takes place using a large number of assessment section learning datasets stored in a database (not shown) as so-called "supervised learning." The assessment section learning datasets used here, for example, are as described in Fig. Figure 10 shows how frequency spectrum data (which in the example shown are based on welding current) and a corresponding welding state from an actual arc welding process are related to each other, thereby creating a single-element data set. These single-element data sets are created in large numbers for the individual welding states for a multitude of welding conditions and stored in the database.

[0049] In the learning phase of evaluation section 62 from the example of the present embodiment, combined learning data is used, the input data of which are the frequency spectrum data and the output data of the welding states. A so-called backpropagation process is performed, in which a weighting coefficient of the edges connecting the individual nodes is adjusted such that a relationship between the input layer and the output layer of the neural network of evaluation section 62 is established. In addition to such backpropagation, various known learning methods can also be applied simultaneously, such as an autoencoder, restricted Boltzmann machine, dropout, noise addition, and sparse regularization, to increase the processing accuracy. The learning phase of evaluation section 62 corresponds to the mechanical learning process of the claims.

[0050] As mentioned above, in addition to the deep learning algorithm shown in section 62, other pattern matching algorithms (not shown), such as a support vector machine or a Bayesian network, can also be applied to the algorithm for assessing the welding condition. In this case, the basic structure is the same: the welding condition, which forms the basis for generating the analysis frequency spectrum data, is assessed and output as assessment information. <Wirkung des ersten Abwandlungsbeispiels>

[0051] As described above, in the robot arc welding system 1 of the present modification example, assessment section 62 evaluates the welding condition based on learning content from a mechanical learning process (deep learning and other mechanical learning). Thus, assessment section 62 does not rely on an analysis procedure based on a human-created mathematical model, but instead enables an assessment of the presence of a defect based on defect characteristic parameters obtained from the frequency spectrum data during the mechanical learning process, thereby increasing the accuracy of the defect assessment.

[0052] In this modified example, the evaluation section 62 learns, in particular, a correspondence between the feature parameters of the time domain and the frequency domain of the frequency spectrum data and the presence of an error of a given error category during the mechanical learning process. In this way, the frequency spectrum data generated by the analysis section 61, as two-dimensional pattern data where the time domain and the frequency domain form orthogonal axes, can be subjected to pattern analysis. The evaluation section 62 can then identify feature parameters of errors that appear in the two-dimensional pattern and perform a highly accurate assessment of the presence of an error for the various error categories. <Abwandlungsbeispiel 2: Beurteilung von Schweißfehlem unter zusätzlicher Berücksichtigung der Schweißdrahtvorschubgeschwindigkeit>

[0053] The fluctuation in the feed rate of the welding wire 43, which is the welding filler material, strongly affects the large-droplet material transfer during arc welding, which also significantly alters the waveform of the time series waveform data. Therefore, the assessment section 62 can evaluate the weld condition taking into account the feed rate recorded by a welding wire feed section 42. Fig. Figure 11 shows an example of a schematic model setup of a neural network of assessment section 62 when using deep learning, taking into account the welding wire feed rate.

[0054] In Fig. 11 is a neural network of the assessment section 62 designed such that, with regard to the analysis frequency spectrum data and the welding wire feed rate entered into the assessment section 62, the presence of a welding defect and the corresponding category are output in the form of assessment information.

[0055] In the example shown, the analysis frequency spectrum data and the actual measured welding wire feed rate are inputted, while the output is a binary value. Only one of four output nodes provides a "true" value for the welding conditions: normal welding, weld cracking, gas deficiency, and welding wire feed rate error (a so-called clustered output). The welding wire feed rate input is an average value of the feed rate, acquired in the same time series as the frequency spectrum data. Although not shown, the welding wire feed rate can also be input using time series waveform data acquired in the same time series as the frequency spectrum data.This neural network in assessment section 62 learns feature parameters that indicate the relationship between the analyzed frequency spectrum data and the welding wire feed rate on the one hand, and the welding states on the other. During the learning phase, the training can be performed using the same backpropagation processing or similar methods, whereby, although not shown, training datasets are used in which the frequency spectrum data, the welding wire feed rate, and the welding state are related to each other during actual arc welding operations. <Wirkung des zweiten Abwandlungsbeispiels>

[0056] As described above, in the robotic arc welding system 1 of the present modification example, the arc welding section 41 includes the welding wire feed section 42, which feeds the welding wire 43, while the evaluation section 62 also assesses the weld condition based on the feed rate of the welding wire 43 through the welding wire feed section 42. By assessing the weld condition based on the feed rate of the welding wire 43 through the welding wire feed section 42, it is possible to determine the defect category and its cause. The supply condition of the shielding gas can also be taken into account when assessing the weld condition.

[0057] When terms such as "vertical", "parallel", "planar", or similar are used in the description above, they are not to be understood in a strict sense. That is to say, "vertical", "parallel", and "planar" are to be understood as "essentially vertical", "essentially parallel", and "essentially planar", such that tolerances and deviations in design and manufacturing are permissible.

[0058] When the above description refers to dimensions and sizes of the external appearance, shapes, positions, and the like as "same," "identical," "different," or similar, this is not to be understood in a strict sense. That is to say, "same," "identical," and "different" allow for tolerances and deviations in design and manufacturing and mean "essentially the same," "essentially identical," and "essentially different," respectively.

[0059] In addition to the above, the procedures of the above embodiments and various modification examples can also be combined. Although not illustrated in detail, the above embodiments and modification examples can be subjected to various changes within their scope. Explanation of reference symbols 1 Robotic arc welding system (arc welding system) 2. Higher-level control device 3 robot controllers 4 robots 5 Welding power source (arc welding power source) 6 Assessment device for arc welding 41 Arc welding section 42 Welding wire feed section 43 Welding wire (welding filler) 61 Analysis Section 62 Assessment section< / abwandlungsbeispiele>

Claims

[1] Arc welding system (1) comprising a robot (4), an arc welding section (41) which is transportably attached to the robot (4) and is configured to perform an arc welding process, an arc welding power source (5) which is configured to supply the arc welding section (41) with a specific welding voltage and a specific welding current, an analysis section (61) which is configured to perform a frequency analysis on a time series waveform of at least one of the welding voltage and the welding current and to generate frequency spectrum data, and an evaluation section (62) which is configured to evaluate a welding condition of the arc welding section (41) on the basis of the frequency spectrum data generated by the analysis section (61), wherein the analysis section (61) is designed to perform a frequency analysis on the time series waveform which has been subjected to convolution using a window function, the assessment section (62) is designed to assess the presence of a welding defect by comparing the analysis frequency spectrum data generated by the analysis section (61) with the normal frequency spectrum data for the case of a normal welding condition, the assessment section (62) is designed to assess the presence of a given error category by comparing the analysis frequency spectrum data in a frequency band of the frequency range corresponding to the error category in question with the normal frequency spectrum data. [2] Arc welding system (1) according to claim 1, wherein a window function is used which corresponds to the welding conditions. [3] Arc welding system (1) according to claim 1 or 2, characterized by , that the assessment section (62) is designed to assess the presence of a fault of a particular fault category by being designed to compare the analysis frequency spectrum data in the time domain with the normal frequency spectrum data. [4] Arc welding system (1) according to any one of claims 1 to 3, characterized by , that the analysis section (61) is designed to perform a frequency analysis of the time series waveform for an impedance and welding energy determined from the welding voltage and welding current. [5] Arc welding system (1) according to any one of claims 1 to 4, characterized by, that the arc welding section (41) has a welding filler feed section (42) which is configured to supply a welding filler (43), wherein the evaluation section (62) is configured to evaluate the welding condition on the basis of the feed rate of the welding filler through the welding filler feed section (42). [6] Assessment device for arc welding comprising an analysis section (61) configured to be detected on a time series waveform of a welding voltage or welding current, which are configured to be from an external arc welding power source (5), to perform a frequency analysis and to generate frequency spectrum data, and an assessment section (62) configured to assess, on the basis of learning content from a mechanical learning process, the presence of a defect of a defect category corresponding to a characteristic value exhibited by the frequency spectrum data in at least one of a time domain and a frequency domain, wherein the analysis section (61) is designed to perform a frequency analysis on the time series waveform which has been subjected to convolution using a window function, The assessment section (62) is designed to assess the presence of a welding defect by comparing the analysis frequency spectrum data generated by the analysis section (61) with the normal frequency spectrum data for the case of a normal welding condition; the assessment section (62) is designed to assess the presence of a respective defect category by comparing the analysis frequency spectrum data in a frequency band of the frequency range corresponding to the defect category in question with the normal frequency spectrum data.

Images