DE102021002682A1 - Procedure for quality control during welding - Google Patents

Abstract

The invention relates to a method for quality monitoring during welding. According to the invention, faulty welding processes are reproduced and process data of process variables are recorded, - the recorded process data are used to train models (MD) from the field of artificial intelligence to detect poor weld quality, and - the trained Models (MD) are used within a graphical user interface while welding processes are being carried out to identify the poor quality of the weld seam.

Description

The invention relates to a method for quality monitoring during welding according to the features of the preamble of claim 1.

From the prior art, as in DE 103 51 579 A1 described, a method for checking weld seams during a welding process and a device for performing this method are known. Part of the radiation intensity that is generated by the arc is detected by a photosensitive detector after it has passed through an intensity reducer. The signal from the detector is mathematically processed by a digital computer unit and a digital pattern is generated therefrom, which is compared with a reference pattern stored in the memory and a deviation is reported via a signal transmitter.

The invention is based on the object of specifying a method for quality monitoring during welding which is improved over the prior art.

The object is achieved according to the invention by a method for quality monitoring during welding with the features of claim 1.

Advantageous refinements of the invention are the subject matter of the subclaims.

• - fehlerhafte Schweißprozesse reproduziert und dabei Prozessdaten von Prozessgrößen, insbesondere Schweißstrom, Schweißspannung und während des Schweißprozesses emittierter Schall, aufgezeichnet,
• - die aufgezeichneten Prozessdaten zum Trainieren von Modellen aus dem Bereich künstlicher Intelligenz zur Erkennung einer mangelnden Schweißnahtqualität verwendet, und
• - die trainierten Modelle innerhalb einer graphischen Benutzeroberfläche während einer Durchführung von Schweißprozessen zur Erkennung der mangelnden Schweißnahtqualität verwendet.

In a method according to the invention for quality monitoring during welding, in particular during gas-shielded metal arc welding (MIG / MAG welding)

• - Faulty welding processes are reproduced and process data of process variables, in particular welding current, welding voltage and sound emitted during the welding process, are recorded,
• - the recorded process data are used to train models from the field of artificial intelligence to identify poor weld seam quality, and
• the trained models are used within a graphical user interface while welding processes are being carried out to identify the poor quality of the weld seam.

In particular, the recorded process variables, d. H. the process data, preprocessed and derived from the preprocessed signals of the process variables, in particular statistical parameters, which are used to train the models for machine learning, consisting of data preparation and model creation. The preprocessing step includes, in particular, a transformation of the recorded process variable into a frequency and / or frequency range. Advantageously, various characteristics are calculated from the transformed, which for example provide information on the course, the structure or the distribution. In particular, statistical parameters are derived from these characteristics, which are used as input variables for training the machine learning models. The respective resulting model is able to divide welds into good welds (OK = weld is OK) and bad welds (not OK = weld is not OK) based on the process variables already mentioned above, welding current, welding voltage and sound. H. the trained models are used while the welding processes are being carried out to evaluate the respective weld as a good weld or a bad weld.

So far, the evaluation of the weld seam quality in gas-shielded metal arc welding has mostly been done by automated test systems for larger quantities. These work on the basis of optical test procedures and are therefore prone to errors. Furthermore, their use is associated with high costs, since space has to be reserved for test stations and the test can only be carried out afterwards. Process-related irregularities can therefore only be recognized and remedied at a later point in time and lead to possible rejects until then. This problem is solved by the invention in that, as described, faulty welding processes and thus the welding defects are reproduced, in particular reproduced under real conditions in the laboratory and the process data are recorded, the models from the field of artificial intelligence are trained to detect poor weld seam quality and the trained models within a graphical user interface (GUI) in the productive environment (inline / inprocess), d. H. during the implementation of welding processes, in particular used to detect the poor quality of the weld seam.

This eliminates the need for manual inspection and automated test systems that have already been used. An inline / in-process test is made possible. In particular, an early detection of a drifting quality of the weld seam is made possible, whereby appropriate corrective measures are made possible. In addition, parameterization can be omitted.

The solution described is particularly also suitable for modern process control variants, in particular according to DVS 0973, which are based on high dynamics in the process have not yet been monitored.

The method is used, for example, in the production of a battery housing for a vehicle, in particular for a motor vehicle.

Embodiments of the invention are explained in more detail below with reference to drawings.

• 1 schematisch Schweißstrom, Schweißspannung und Schall über die Zeit,
• 2 schematisch einen Modellerstellungsteil eines Verfahrens zur Qualitätsüberwachung beim Schweißen,
• 3 schematisch ein Originalsignal,
• 4 schematisch Nulldurchgänge,
• 5 schematisch einen Spektralschwerpunkt,
• 6 schematisch einen spektralen Roll-Off,
• 7 schematisch eine spektrale Bandbreite,
• 8 schematisch eine Datenerfassung mittels einer Vorrichtung, und
• 9 schematisch eine Bewertung eines Schweißprozesses.

Show:

• 1 schematically welding current, welding voltage and sound over time,
• 2 schematically a model creation part of a method for quality monitoring during welding,
• 3 schematically an original signal,
• 4th schematic zero crossings,
• 5 schematically a spectral center of gravity,
• 6th schematically a spectral roll-off,
• 7th schematically a spectral bandwidth,
• 8th schematically a data acquisition by means of a device, and
• 9 schematically an evaluation of a welding process.

Corresponding parts are provided with the same reference symbols in all figures.

Based on 1 until 9 The following describes a method for quality monitoring during welding, in particular for gas-shielded metal arc welding (MIG / MAG welding). In summary, faulty welding processes are reproduced in this method and process data of process variables, in particular welding current, are reproduced in the process I. , Welding voltage U and sound emitted during the welding process S. , recorded, the recorded process data for training models MD from the field of artificial intelligence used to detect poor weld seam quality, and the trained models MD used within a graphical user interface while welding processes are being carried out to identify the poor quality of the weld seam. In particular, input signals of the process variables, ie the process data, are preprocessed, and statistical parameters are derived from the preprocessed signals, which are used to train the models MD can be used for machine learning, consisting of data preparation and model creation. The respective resulting model MD is able to perform welds using the process variables already mentioned above, welding current I. , Welding voltage U and sound S. to be divided into good welds (OK = weld is OK) and bad welds (NOK = weld is not OK), ie the trained models MD are used during the welding process to evaluate the respective weld as a good weld or a bad weld.

The background and the process sequence of this procedure are described in detail below.

In gas-shielded metal arc welding, digitally controlled welding power sources are advantageously used, with which innovative process variants can be implemented. If irregularities occur during welding, these are reflected in the process variables. Process variables considered so far are the welding voltage U and the welding current I. . Another signal, ie another process variable in which or in which irregularities become noticeable, is the sound emitted during the welding process S. . It is known that an experienced welder can make statements about the quality of the weld based on acoustics. The three signals, ie process variables, welding current, are thus combined I. , Welding voltage U and sound S. affected by irregularities in the welding process. 1 shows this as an example for a CMT pulse welding process based on the three process variables welding current I. , Welding voltage U and sound S. over time t, whereby the arrows shown indicate areas in which irregularities in the welding process can be recognized.

The basic approach of the method described here is with the help of the three signals mentioned, ie the process variables welding current I. , Welding voltage U and sound S. , and using machine learning ML , ie machine learning, to draw conclusions about the quality of the weld. As in 2 shown, input signals, ie recorded process data, of the process variables welding current run through for this purpose in a modeling part of the method I. , Welding voltage U and sound S. initially signal preprocessing in a first step MS1. From these preprocessed signals, statistical parameters, that is to say parameters, are then derived, that is to say extracted, in a second step MS2. These are used for machine learning ML used, which consists of a data preparation, ie a data preparation, in a third step MS3 , and the model creation in a fourth step MS4. The resulting model MD is, as already mentioned, able to perform welds on the basis of the values of the welding current I. , Welding voltage U and sound S. to be divided into good welds (OK) and bad welds (not ok).

The novelty of the solution described lies in the combination of the observed welding current signals I. , Welding voltage U and sound S. as well as in signal preprocessing. Other approaches consider the signals essentially only in their original form. In digitized form, this corresponds to a time series of data points. Statistical key figures are derived from these time series, which are used as input for machine learning ML be used.

The signal preprocessing described here is an intermediate step and transforms the input signals, ie the recorded process data of the process variables welding current I. , Welding voltage U and sound S. , in different dimensions. Various transformations are used for this. These have their origin in audio analysis. However, there are also current and voltage signals like sound S. are mapped by a time series, they can also be transformed. The target dimensions of the transformations are time domain, frequency domain and cross-reference domain.

the 3 until 7th provide an overview of the associated transformations. As described, the target dimensions in signal preprocessing are the time domain, frequency domain and "cross-border domain". The time domain includes the original signal, as in 3 shown, and zero crossings as in 4th shown. The frequency range includes a spectral centroid, as in 5 shown a spectral roll-off, as in 6th shown, and a spectral bandwidth as in 7th shown. The limit range comprises a Mel-Frequency-Cepstrum-Coefficient (MFC), which cannot be represented in a schematic black-and-white representation.

The original signal is the original input data in its raw state.

The spectral center of gravity expresses the frequency at which the energy of a spectrum is centered. So it shows the “center of gravity” of a frequency spectrum. It is calculated as follows: $${\mathrm{C.}}_r=\frac{{\displaystyle \sum {}_{k=1}^{\frac{N}{2}}f\left[k\right]\left|X_r\left[k\right]\right|}}{{\displaystyle \sum {}_{k=1}^{\frac{\mathrm{<figure-callout\; id="2"\; label="N"\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">N</figure-callout>}}{2}}\left|X_r\left[k\right]\right|}}$$

• C_r der Spektralschwerpunkt,
• f[k] die Frequenz des k-ten bin,
• r die aktuelle Frame-Nummer,
• [k] die Kurzzeit-Fouriertransformation des Frame,
• N der limitierende Wert von k,
• k der bin,
• X_r[k] der Spektralwert des k-ten bin.

Where:

• C _{r is} the spectral center of gravity,
• f [k] the frequency of the k-th bin,
• r the current frame number,
• [k] the short-term Fourier transform of the frame,
• N is the limiting value of k,
• k who am
• X _r [k] the spectral value of the k-th bin.

The spectral roll-off resembles the shape of the signal in that it shows at what point high frequencies approach zero. For this purpose, he specifies the frequency that forms the upper limit of the 85% concentration in the size distribution of the frequency spectrum. It is calculated as follows: $${\displaystyle \sum _{k=1}^{\mathrm{F.}}\left|X_r\left[k\right]\right|}=0.85{\displaystyle \sum _{k=1}^{\frac{\mathrm{<figure-callout\; id="2"\; label="N"\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">N</figure-callout>}}{2}}\left|X_r\left[k\right]\right|}$$

• r die aktuelle Frame-Nummer,
• [k] die Kurzzeit-Fouriertransformation des Frame,
• F der limitierende Wert von k,
• N der limitierende Wert von k,
• k der bin,
• X_r[k] der Spektralwert des k-ten bin.

Where:

• r the current frame number,
• [k] the short-term Fourier transform of the frame,
• F is the limiting value of k,
• N is the limiting value of k,
• k who am
• X _r [k] the spectral value of the k-th bin.

The spectral bandwidth is the frequency interval in which the spectral concentration is at least half of the maximum value. Thus it is "the power transfer function around the center frequency". For the solution described here, the second, third, and fourth orders turned out to be the most relevant. The spectral bandwidth is calculated as follows: $${\mathrm{B.}}_p={\left({\displaystyle \sum {}_{k=1}^{\mathrm{F.}}\ast {\left(X_r\left[k\right]f\left[k\right]-{\mathrm{C.}}_r\right)}^p}\right)}^{\frac{1}{p}}$$

• B_p die spektrale Bandbreite,
• r die aktuelle Frame-Nummer,
• [k] die Kurzzeit-Fouriertransformation des Frame,
• F der limitierende Wert von k,
• k der bin,
• p die Ordnung,
• f[k] die Frequenz des k-ten bin,
• C_r der Spektralschwerpunkt,
• X_r[k] der Spektralwert des k-ten bin.

Where:

• B _{p is} the spectral bandwidth,
• r the current frame number,
• [k] the short-term Fourier transform of the frame,
• F is the limiting value of k,
• k who am
• p the order,
• f [k] the frequency of the k-th bin,
• C _{r is} the spectral center of gravity,
• X _r [k] the spectral value of the k-th bin.

The zero crossings (zero crossing rate) define how often and / or at what point x and / or at what time t the signal changes sign. This is a measure of the signal smoothness and is calculated as follows: $$Z\mathrm{C.}\mathrm{R.}={\displaystyle \sum _{n=1}^{N-1}\left|sGn\left[x\left(n+1\right)\right]-sGn\left[x\left(n\right)\right]\right|/2\left(N-1\right)}$$

• ZCR die Nulldurchgangsrate,
• N der limitierende Wert von k,
• n der bin,
• x(n) ein diskretes Audiosignal, wobei für sgn[x(n)] gilt:
  • 1 (x(n)≥0
  • -1 (x(n)≤0.

Where:

• ZCR is the zero crossing rate,
• N is the limiting value of k,
• n who am
• x (n) is a discrete audio signal, where for sgn [x (n)]:
  • 1 (x (n) ≥0
  • -1 (x (n) ≤0.

The Mel Frequency Cepstrum Coefficient (MFC) shows the logarithmic energy spectrum for certain frequency bands. The bands are not linear and are based on the Mel scale. This scale replicates the human hearing system. For the solution described here, fifteen bands were zoned, the twelve lowest of which were used. The Mel Frequency Cepstrum Coefficient (MFC) is calculated as follows: $$c\left(n\right)={\displaystyle \sum _{m=0}^{\mathrm{M.}-1}lO{G}_{10}\left(s\left(m\right)\right)cOs\left(\frac{\pi n\left(m-0.5\right)}{\mathrm{M.}}\right)}$$

• c(n) der Mel-Frequenz-Cepstrum-Koeffizient,
• n die Bandnummer,
• M die Gesamtzahl der dreieckigen Mel-Gewichtungsfilter,
• s(m) das Mel-Spektrum,
• m eine Laufvariable von 0 bis M-1.

Where:

• c (n) is the mel frequency cepstrum coefficient,
• n the volume number,
• M is the total number of triangular Mel weighting filters,
• s (m) the Mel spectrum,
• m is a run variable from 0 to M-1.

The signals generated in this way are then used to extract statistical key figures. The method of signal preprocessing thus makes it possible to significantly increase the number of signals available. As a result, the number of key figures extracted is also increased. This stands for the subsequent machine learning ML new features, ie new features, are ready that provide additional information about the division into OK and NOK, ie into good welds and bad welds. This can be the resulting models MD improve with regard to their accuracy, but also with regard to their generalizability.

8th shows schematically a data acquisition by means of a corresponding device. The device includes process sensors 1 , a data acquisition module 2 , a measurement data calculator 3 , a database 4th for recording the process variables welding current I. , Welding voltage U and sound S. , a robot 5 for carrying out the welding process on a component 6th and a welding unit 7th .

In the process, “OK” (OK) welds, ie good welds, and “not OK” (not OK) welds, ie bad welds, are advantageously carried out first. The bad welds are of poor quality. This can be deliberately provoked, for example by contamination of a joint or the selection of unsuitable welding parameters. The three signals, ie the welding parameters welding voltage U , Welding current I. and sound S. , are recorded in order to generate a database. This means that these three signals are available for each weld.

Each of these signals then goes through various transformations, as described above. As already mentioned, the transformations come from the field of audio analysis. However, since the current and voltage data, like sound data, are also time series, the transformations can also be used here. This multiplies the number of signals available per weld.

• - Arithmetisches Mittel:

$$am=\frac{1}{n}{\displaystyle \sum {}_{i=1}^n{x}_i}$$

All signals obtained in this way are also one-dimensional time series. It is thus possible to extract statistical key figures from it. The key figures used are listed below:

• - Arithmetic mean:

$$am=\frac{1}{n}{\displaystyle \sum {}_{i=1}^n{x}_i}$$

• am das arithmetische Mittel,
• n die Anzahl an Datenpunkten,
• i Laufvariable,
• x der Wert eines Datenpunkts.

• - Varianz:

$${\mathrm{<figure-callout\; id="2"\; label="\sigma "\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">\sigma </figure-callout>}}^2=\frac{{\displaystyle \sum {}_i^n{\left(x_i-\overline{x}\right)}^2}}{n}$$

Where:

• am the arithmetic mean,
• n is the number of data points,
• i run variable,
• x is the value of a data point.

• - variance:

$${\mathrm{<figure-callout\; id="2"\; label="\sigma "\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">\sigma </figure-callout>}}^2=\frac{{\displaystyle \sum {}_i^n{\left(x_i-\overline{x}\right)}^2}}{n}$$

• σ² die Varianz,
• n die Anzahl an Datenpunkten ,
• i Laufvariable,
• x der Wert eines Datenpunkts.

• - Median: Dies ist ein zentraler Wert einer sortierten Datenreihe.

• - Summe:

$$summe={\displaystyle \sum {}_i^n{x}_i}$$

Where:

• σ ^{2 is} the variance,
• n is the number of data points,
• i run variable,
• x is the value of a data point.

• - Median: This is a central value of a sorted data series.

• - Total:

$$summe={\displaystyle \sum {}_i^n{x}_i}$$

• n die Anzahl an Datenpunkten,
• i Laufvariable,
• x der Wert eines Datenpunkts.

• - Minimum: Dies ist der niedrigste Wert einer Datenreihe.

• - Maximum: Dies ist der höchste Wert einer Datenreihe.

• - Schiefe:

$${\mathrm{<figure-callout\; id="1"\; label="g"\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">g</figure-callout>}}_1=\frac{m_3}{{\mathrm{<figure-callout\; id="1"\; label="m"\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">m</figure-callout>}}_1^{3/2}},wobei{m}_k={\displaystyle \sum {}_i^n{\left(x_i-\overline{x}\right)}^k}$$

Where:

• n is the number of data points,
• i run variable,
• x is the value of a data point.

• - Minimum: This is the lowest value in a data series.

• - Maximum: This is the highest value in a data series.

• - skewness:

$${\mathrm{<figure-callout\; id="1"\; label="G"\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">G</figure-callout>}}_1=\frac{m_3}{{\mathrm{<figure-callout\; id="1"\; label="m"\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">m</figure-callout>}}_1^{3/2}},wObei{m}_k={\displaystyle \sum {}_i^n{\left(x_i-\overline{x}\right)}^k}$$

• g₁ die Schiefe,
• m₃ der Kumulant dritter Ordnung,
• m₂ der Kumulant zweiter Ordnung,
• m_k der Kumulant k-ter Ordnung,
• n die Anzahl an Datenpunkten,
• i Laufvariable,
• x der Wert eines Datenpunkts.

• - Kurtosis:

$${\mathrm{<figure-callout\; id="2"\; label="\gamma "\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">\gamma </figure-callout>}}_2=\frac{{\mathrm{<figure-callout\; id="4"\; label="\mu "\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">\mu </figure-callout>}}_4}{{\mathrm{<figure-callout\; id="4"\; label="\sigma "\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">\sigma </figure-callout>}}^4}-\mathrm{3,}wobei{\mu }_4=\frac{{\displaystyle \sum {}_i^n{\left(x_i-\overline{x}\right)}^4}}{n}$$

Where:

• g ₁ the skewness,
• m _{3 is} the third order cumulant,
• m _{2 is} the second order cumulant,
• m _{k is} the cumulant of the kth order,
• n is the number of data points,
• i run variable,
• x is the value of a data point.

• - Kurtosis:

$${\mathrm{<figure-callout\; id="2"\; label="\gamma "\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">\gamma </figure-callout>}}_2=\frac{{\mu }_{\mathrm{4th}}}{{\sigma }^{\mathrm{4th}}}-\mathrm{3,}wObei{\mu }_{\mathrm{4th}}=\frac{{\displaystyle \sum {}_i^n{\left(x_i-\overline{x}\right)}^{\mathrm{4th}}}}{n}$$

• γ₂ die Kurtosis,
• µ₄ das vierte zentrale Moment,
• σ⁴ das Quadrat der Varianz,
• n die Anzahl an Datenpunkten,
• i Laufvariable,
• x der Wert eines Datenpunkts.

• - Root Mean Square:

$$rms=\sqrt{\frac{{\displaystyle \sum {}_i^n{x}_i^2}}{n}}$$

Where:

• γ ₂ the kurtosis,
• µ ₄ the fourth central moment,
• σ ^{4 is} the square of the variance,
• n is the number of data points,
• i run variable,
• x is the value of a data point.

• - Root Mean Square:

$$rms=\sqrt{\frac{{\displaystyle \sum {}_i^n{x}_i^2}}{n}}$$

• rms = Root Mean Square,
• n die Anzahl an Datenpunkten,
• i Laufvariable,
• x der Wert eines Datenpunkts.

• - erstes Quantil:

$${\mathrm{<figure-callout\; id="1"\; label="q"\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">q</figure-callout>}}_1=\frac{n+1}{4}\text{ in einer sortierten Datenreihe}$$

Where:

• rms = Root Mean Square,
• n is the number of data points,
• i run variable,
• x is the value of a data point.

• - first quantile:

$${\mathrm{<figure-callout\; id="1"\; label="q"\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">q</figure-callout>}}_1=\frac{n+1}{\mathrm{4th}}\text{in a sorted data series}$$

• q₁ das erste Quantil,
• n die Anzahl an Datenpunkten.

• - drittes Quantil:

$${\mathrm{<figure-callout\; id="3"\; label="q"\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">q</figure-callout>}}_3=\frac{n+1}{4}\times 3\text{ in einer sortierten Datenreihe}$$

Where:

• q _{1 is} the first quantile,
• n is the number of data points.

• - third quantile:

$${\mathrm{<figure-callout\; id="3"\; label="q"\; filenames="DE102021002682A1\_0018.png"\; state="\{\{state\}\}">q</figure-callout>}}_3=\frac{n+1}{\mathrm{4th}}\times 3\text{in a sorted data series}$$

• q₃ das dritte Quantil,
• n die Anzahl an Datenpunkten.

• - Entropie:

$$e=-{\displaystyle \sum {}_i^n\left(\left|x_i\right|\times log\left|x_i\right|\right)}$$

Where:

• q _{3 is} the third quantile,
• n is the number of data points.

• - entropy:

$$e=-{\displaystyle \sum {}_i^n\left(\left|x_i\right|\times lOG\left|x_i\right|\right)}$$

• e die Entropie,
• n die Anzahl an Datenpunkten,
• i Laufvariable,
• x der Wert eines Datenpunkts.

Where:

• e is the entropy,
• n is the number of data points,
• i run variable,
• x is the value of a data point.

The method of data processing explained last makes it possible to extract a large number of key figures from three signals. These are then used as features, ie as characteristics, for the downstream machine learning ML used.

The input data are first prepared before they are fed into an algorithm. There are three steps to this preparation. The first step is to scale the data with a standard scaler. This distributes the values of the individual features on a uniform scale in order to avoid that some features show significantly higher absolute values than others, as this can worsen the results of the machine learning algorithms. The formula is: $$z_i=\frac{\left(x_i-\overline{x}\right)}{\sigma }$$

• z_i der standardisierte Wert
• x_i der zu standardisierende Wert
• x der Mittelwert
• σ die Standardabweichung.

Where:

• z _{i is} the standardized value
• x _i the value to be standardized
• x the mean
• σ is the standard deviation.

In order to avoid so-called overfitting, not all, but only the most meaningful features are fed into the algorithm. These are determined using a correlation analysis. the 50 Features that correlate most strongly with the target value (OK or not OK) remain. In order to further reduce the number of features, a linear discriminant analysis (LDA) is then carried out. This compresses the information of features through algorithmic processing into a smaller number of features with minimal loss of information.

• - Gradient Boosting Trees
• - Random Forest
• - Kernel Support Vector Machine
• - Multilayer Perceptron

The features obtained in this way are used in a classification algorithm of feature-based machine learning ML fed. This learns welds based on the welding current data I. , Welding voltage U and sound S. to be divided into OK, ie good welds, and not ok, ie bad welds. The following algorithms are used, for example:

• - Gradient boosting trees
• - Random Forest
• - Kernel Support Vector Machine
• - Multilayer perceptron

9 shows a flow of evaluation of a welding process. The process variables welding current I. , Welding voltage U and sound S. are recorded. In a first evaluation step BS1, the signal preprocessing of these signals, ie the recorded process data of these process variables, the extraction of the statistical key figures, that is, parameters, from these preprocessed signals and the data preparation, ie the data preparation, of these key figures. The key figures prepared in this way are then transferred to the trained model in a second evaluation step BS2 MD supplied as input data, which in a third evaluation step BS3 decides whether the welding process is OK or not OK, ie OK, ie a good weld, or not OK, ie a bad weld. The fourth evaluation step BS4 is the output of the corresponding result, ie OK or NOK, ie there is a good weld or a bad weld.

QUOTES INCLUDED IN THE DESCRIPTION

This list of the documents listed by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent literature cited

• DE 10351579 A1 [0002]

Claims (8)

Method for quality monitoring during welding, characterized in that - faulty welding processes are reproduced and process data of process variables are recorded, - the recorded process data are used to train models (MD) from the field of artificial intelligence to detect poor weld seam quality, and - the trained models (MD) can be used within a graphical user interface while welding processes are being carried out to identify the poor quality of the weld seam. Procedure according to Claim 1 , characterized in that the quality monitoring is carried out during metal arc welding. Method according to one of the preceding claims, characterized in that process data of the process variables welding current (I), welding voltage (U) and sound (S) emitted during the welding process are recorded. Method according to one of the preceding claims, characterized in that the recorded process variables are first preprocessed and then parameters are derived from the preprocessed signals which are used for machine learning, consisting of data preparation and model creation. Procedure according to Claim 4 , characterized in that the preprocessing step includes a transformation of the recorded process variable into a frequency and / or frequency range. Procedure according to Claim 5 , characterized in that various characteristics are calculated from the transform, which for example provide information on the course, the structure or the distribution. Procedure according to Claim 6 , characterized in that statistical parameters are derived from these characteristics, which are used as input variables for training the machine learning models. Use of a method according to one of the preceding claims in the production of a battery housing for a vehicle.

Images