RU2827116C1 - Welded joint identification method - Google Patents

Abstract

FIELD: measuring.

SUBSTANCE: invention relates to the field of non-destructive testing of welded joints of main pipelines, namely to the automated analysis of the weld surface in order to personalize the welder. A database of identified personalized welded joints is created, after which technical parameters are measured and the unidentified welded joint is scanned. Unidentified welded joint is personalized by visual measurement control, by correlation method with construction of static dependencies of informative features and their comparison with static dependencies of informative features contained in said database, and/or by constructing matrices of the form "n*k", where n is the width of the matrix, and k is the length of the matrix, containing in its cells sets of functions of informative features for each point of unidentified welded joint with comparison of obtained results with values of formed initial matrices, contained in said database. In case of coincidence of informative features as a result of the above comparisons, a conclusion is made on the possibility of personalization of an unidentified welded joint, and in case of mismatch, I conclude on the absence of the welder in the database.

EFFECT: efficient high-precision method for identification of welder by features of unidentified welded joint with accuracy higher than 90%.

4 cl, 14 dwg, 4 tbl

Description

The invention relates to the field of non-destructive testing of welded joints of main pipelines, namely to automated analysis of the surface of welded joints.

In order to accurately and exclusively interpret the description of the claimed invention, the following definitions have been adopted:

- an identified personalized weld seam is understood to mean a weld seam whose author is known;

- an unidentified weld is a weld whose author is unknown;

- an identified welder is understood to mean a welder whose identity has been established;

- an unidentified welder is a welder whose identity has not been established.

Most countries have national regulations governing welding work performed at particularly hazardous facilities, which include main pipelines. According to such documents, the work must be accompanied by certain precautions. In particular, a mark (brand) of the specialist who performed it must be placed next to the welded joint. If it turns out that the seam is made with a defect, then, visually, using this brand, using the materials of a single database of welders, you can easily determine the full name of the employee, his position, type of activity and category. Thus, increasing the level of responsibility of each individual welder improves the quality and reliability of the welds performed. Identification of the welder in this case does not seem difficult. However, there is a problem of determining the performer of the weld in the absence of a brand at the work site or, for example, in the event of a conflict between different parties, such as the customer and the contractor. The occurrence of this problem on the main pipeline, as a particularly dangerous object, increases the risk of emergency situations, including fires and environmental pollution, due to oil and oil product leaks, as a result of poor-quality and irresponsible performance of work by an unidentified welder. Consequently, when identifying a welder at particularly dangerous objects, the cost of an error in such identification is too high, since it entails personal liability of the welder for possible damage caused.

There are various methods of weld analysis:

- visual and measuring control (checking the conformity of the geometric parameters of welded joints with the requirements of regulatory, technical and design documentation, detection of surface (reaching the surface) and through defects of welded joints such as cracks, undercuts, lack of fusion, unwelded craters, burns, non-metallic inclusions, delaminations, etc. and determination of their location, size and orientation on the surface);

- methods of metallographic analysis (analysis of the metal structure in the heat-affected zone, which can help determine the welding technology and the type of materials used);

- spectral analysis (determination of the elemental composition of a metal by measuring spectral lines).

Each of these methods has its limitations, and the best result is achieved by using a combination of them. All of these methods examine the characteristics of welds and are used mainly to determine the skill level of already identified welders in the course of their work.

If the author of the weld is unknown, then its personalization is possible only in certain cases. For example, if the welder has a unique welding style, which is manifested in his work over a long period of time. It is known that the factors influencing the formation of the weld, and at the same time depending on the welder, include the following: the angle of the electrode, the amplitude of the electrode oscillations, the trajectory of the electrode movement, the welding speed. The greater the amplitude of the electrode oscillations, the greater the width of the weld, the height of the weld depends on the welding speed and the trajectory of the electrode movement. If the welder makes simple oscillatory movements without a long delay on the edges, then the height of the weld will be the greatest in the center and will smoothly decrease towards the edges and the base metal. Also, if you do not make a delay on the edges, but perform loop-shaped movements, the effect will be the same, but the shape of the scales will change. The shape of the scales depends on the welding speed, current supply, etc. With increasing speed, the scales are drawn out, giving them a parabolic shape, since the weld pool will be smaller than when welding at low speeds, it will solidify faster in the tail of the weld. The same welder can repeat defects, such as undercuts, along the same weld or on different welds.

At the same time, for precise identification of the welder, special examinations and labor-intensive comparative analysis with comparison of the unidentified weld with identified personalized ones are required. At present, such analysis is carried out by the visual-measuring control (VMC) method, the disadvantages of which are high labor intensity, low accuracy and reliability, due to the great influence of the human factor and the lack of a sufficient degree of automation of the process.

A method for tracking and ranking the qualifications of welders is known from the prior art (RU 2763708, published on 30.12.2021), which ensures high accuracy in calculating the welder qualification index due to the fact that when comparing the coordinates of the points of the surface of a weld with a standard, the axis of the spatial image of the standard is aligned with the axis of the spatial image of the weld, while the coordinates of the axis of the 3D digital image of the surface of the weld are calculated by dividing in half the segment included between the coordinates of the abscissas of two profile points with maximum curvature, and the boundaries of the 3D digital image of the surface of the weld are determined by the coordinates obtained by the laser device by linear approximation of the coordinates of the abscissas of the points, which are calculated by dividing in half the segment included between the coordinates of the abscissas of two points of the profiles of the cross sections of the digital image of the weld with maximum curvature of the profile and defined as the boundary points of the weld, and the obtained data on deviations of the cross-sectional areas of the weld from the cross-sectional areas of the standard are transmitted via communication channels to an electronic database for storage and subsequent automated creation of a welder's qualification rating.

Despite the fact that this method implements an algorithm for comparing a standard with a digitalized shape of a weld, the disadvantage of this method is the impossibility of identifying the author of the weld being studied if it is not known in advance, since the method does not have an algorithm that allows for a sufficiently accurate comparison of the impersonal weld being studied with the entire database of already identified welds.

A method for assessing the qualification of a welder is known from the prior art (RU 2569276, IPC G09B 19/24, B23K 9/00, published 20.11.2015), including measuring the geometric dimensions of a weld with a template and comparing them with the dimensions established by regulatory and technical documents, characterized in that, based on the dimensions established by regulatory and technical documents, taking into account the physical and mechanical properties of the molten material of the weld, the shape of the standard, the area of the standard of the weld are calculated, based on the results of measuring the geometric dimensions of the surface of the weld, the shape of the surface of the weld is determined and the absolute deviations of the cross-sectional area of the surface of the weld from the area of the standard are calculated and the welder qualification index is calculated using the formula:

where QW is the welder qualification index;

N=Lw/dLw+1 – the number of measured sections of the weld seam with a measurement step dLw greater than or equal to 0.1 mm;

Lw - weld length;

j - current value of the measured cross-section of the weld;

S ^E - area of the standard;

- the absolute deviation of the area of the j-th cross-section of the weld surface from the area of the standard, which is determined by the formula:

Where - the absolute value of the deviation of the shape of the weld surface in the j-th section from the external and internal shape of the standard, which is determined by the formula:

B - measurement width, B values are in the range from 1.1e _max to 10e _max ,

e _max - maximum width of the welded seam, established by regulatory and technical documents;

- measured values of the coordinates of the surface of the j-th cross-section of the weld with a step dx from 0.1 mm to 5 mm;

Z, x - coordinates in the Z0X coordinate system;

Z ^Emax (x) - a curve that determines the external shape of the weld seam standard;

Z ^Эmin (х) is a curve defining the internal shape of the weld seam standard; the external Z ^Эmax (х) and internal Z ^Эmin (х) shapes of the weld seam standard curves in any spatial position are calculated according to the limit values of the width (e) and height (g) of the weld seam convexity cross-section established by normative and technical documents by solving the integro-differential equation:

where l is the distance from the initial point of the reference curve to the calculation point, measured along the curve, the value of 1 varies from 0 to L;

L - length of the reference curve of the weld seam;

dl - differentiation step along the reference curve;

dλ - integration step along the reference curve;

τ(l) is the angle of inclination of the tangent of the reference curve to the horizon at the calculation point;

κ ₀ - curvature of the reference curve of the weld seam at the initial point, determined during the solution of the equation;

a _K - physical and mechanical property of the molten material of the weld - capillary constant of the molten metal of the weld;

δ - angle of longitudinal inclination of the weld seam;

ϕ - angle of transverse inclination of the weld seam;

Z(x) is the equation describing the reference curves of the weld seam, written in parametric form {Z(l), х(l)} and has the form:

the area of the standard is calculated using the formula:

Despite the fact that this method allows us to evaluate the qualifications of a welder based on the deviation of the weld surface from the standard, its disadvantage is the impossibility of identifying the author of the weld being examined if he is not known in advance, since the method does not have an algorithm that allows us to compare the impersonal weld being examined with sufficient accuracy with the entire database of already identified personalized welds.

The technical problem that the claimed invention is aimed at solving is the creation of an effective, high-precision method for identifying a weld to establish the identity of a welder based on the characteristics of an unidentified weld with an accuracy of over 90%, the reliability of which is ensured by an automated three-stage process of analyzing and comparing welds using elements of machine learning and a method of correlative comparison of informative features of welds.

The technical result of implementing the claimed method consists in increasing the reliability of identification of a weld with subsequent personalization of the welder based on the characteristics of the unidentified weld.

The technical result is achieved in that in the method for identifying a weld seam, a database of identified personalized weld seams is created by measuring technical parameters and laser scanning weld samples to obtain two-dimensional and three-dimensional images, dividing each identified personalized weld seam into sections no less than 70 mm long along the longitudinal scanning coordinate, using a section passing through the center of the weld seam, with the allocation of informative features in each section of each identified personalized weld seam, constructing static dependencies of informative features and forming an initial matrix of the type "n * k", where n is the width of the matrix, and k is the length of the matrix containing in its cells sets of functions of informative features for each point of the identified personalized weld seam, followed by placing the created database of identified personalized weld seams in the computer memory unit, after which the technical parameters are measured and the unidentified weld seam is laser scanned, then the unidentified weld seam is personalized by visual and measuring control using comparing the measured technical parameters of the unidentified weld with the corresponding technical parameters of the identified personalized welds contained in the said database, after which the most significant informative features of the unidentified weld are selected by means of a correlation method with the construction of static dependencies of the informative features and their comparison with the static dependencies of the informative features contained in the said database, and/or by constructing matrices of the type "n*k", where n is the width of the matrix, and k is the length of the matrix containing in its cells sets of functions of informative features for each point of the unidentified weld with the implementation of a comparison of the obtained results with the values of the generated initial matrices contained in the said database in the event of a match of the informative features as a result of the above comparisons, a conclusion is made on the possibility of personalizing the unidentified weld.

The technical result is also achieved by using the width and height of the weld seam, the width of the weld seam scales, the distance between the scales, the differences in the heights of the scales, the depressions between the scales, the craters of the locks, the places of re-ignition and extinguishing of the welding arc when changing the electrode as technical parameters.

The technical result is also achieved by measuring technical parameters using a caliper, ruler, thickness gauge, micrometer, templates, feeler gauges, angle gauges with vernier, micrometer and indicator bore gauges.

The technical result is also achieved by the fact that the average value of the width of the cross-section profile, the average value of the area under the curve of the weld profile, the average value of the center of mass of the profile section, the average value of the position of the maximum of the approximating polynomial, the average value of the coefficients of the approximating polynomial, the average value of the residual of the approximating polynomial, the standard deviation of the width of the cross-section profile, the standard deviation of the area under the curve of the weld profile, the standard deviation of the center of mass of the profile section, the standard deviation of the position of the maximum of the approximating polynomial, the standard deviation of the coefficients of the approximating polynomial, the standard deviation of the residual of the approximating polynomial, spectral readings of the power density of spatial frequencies, the tangent of the angle of inclination of the spatial oscillations of the weld profile are used as informative features.

The claimed invention is illustrated by drawings, where:

Fig. 1 shows an example of the data acquisition process using laser illumination technology.

Fig. 2 shows an example of the result of scanning a three-dimensional image.

Fig. 3 shows an example of a cross-section of a weld profile before transformation.

Fig. 4 shows an example of a cross-section of a weld profile after transformation.

Fig. 5 shows an example of approximation of the cross-section of the weld profile.

Fig. 6 shows the stages of determining the central longitudinal section.

Fig. 7 shows an example of the profile of the central longitudinal section of a weld seam.

Fig. 8 shows an example of a spatial frequency power spectrum.

Fig. 9 shows the graphs of the relationship between the selected informative features using the mean width of the weld cross-section “Mean_Width”.

Fig. 10 shows the graphs of the relationship between the selected informative features without using the feature of the average width of the weld cross-section “Mean_Width”.

Fig. 11 shows graphs of the dependence of the accuracy of different trained models on the number of selected informative features.

Fig. 12 shows graphs of the dependence of the accuracy of different trained models on the number of selected informative features.

Fig. 13 shows comparative graphs of the dependence of reliability on the number of features.

Fig. 14 shows comparative graphs of the dependence of reliability on the number of features.

The claimed method is an automated process of comparing the measured technical parameters and informative features of an unidentified weld with a database of identified personalized welds, using a sequential combination of three methods: visual-measuring control (VMC), by constructing matrices and a correlation method, with subsequent assignment of an unidentified weld to an identified welder. The method is implemented as follows:

Initially, a database of identified personalized welds is created by measuring technical parameters and laser scanning weld samples to obtain two-dimensional and three-dimensional images, constructing static dependencies of informative features and forming an initial matrix of the "n*k" type, where n is the matrix width and k is the matrix length. The initial matrix of the "n*k" type contains in its cells sets of functions of informative features for each point of the identified personalized weld. Laser scanning is performed as follows: each identified personalized weld is scanned (Fig. 1, Fig. 2), the data from the scanning device, representing the height of the scanned seam at each of its points, are loaded into the computer. The results of scanning the profile of each identified personalized weld are divided into sections no less than 70 mm long along the longitudinal coordinate of scanning and informative features are identified in each section of each identified weld, and are supplemented with informative features of the transverse and longitudinal section and a two-dimensional spatial spectrogram of each identified weld. The division of sections is used to increase the number of elements and testing samples of informative features for the purpose of further accurate construction of a stable classifier of the trained model and effective evaluation of the work.

In order to obtain informative features of the cross-section of each identified personalized weld, when scanning the weld, in order to eliminate the influence of the weld preparation form factor and the geometric factors of the relative position of the scanner and the part, the cross-section of the weld preparation is rotated, the origin of the scanner coordinate system is moved, and the weld cross-section profile is scaled along the Y axis of the scanner so that the weld boundaries are located at points with coordinates (y=-1; z=0) and (y=1; z=0). The Z coordinate axis of the scanner is rotated by 180° (Fig. 3 and Fig. 4).

For each section of the identified weld seam, the geometric shape of the cross-sectional profile is determined using approximation by a 4th degree polynomial according to formula 1:

Next, for each section of the identified weld, informative features of the cross-section are determined:

a) the width of the cross-sectional profile W before approximation using formula 2:

where y _o , y _n are the coordinates of the initial and final points of the profile along the Y axis;

b) area A under the weld profile curve according to formula 3:

where z _t is the height of the weld profile points,

Δу - step of profile points along the cross-section of the weld seam profile,

c) the center of mass CM of the weld seam cross-section profile according to formula 4:

d) the position of the maximum of the approximating polynomial of the cross-section of the weld;

d) coefficients of the approximating polynomial of the cross-section of the weld seam a ₀ , a ₁ , a ₂ , a ₃ , a ₄ ;

e) the discrepancy obtained when calculating the approximating polynomial of the cross-section of the weld.

An example of approximation of the cross-section of the profile is shown in Fig. 5, with the polynomial coefficients: a ₀ = 2.518; a ₁ = - 0.925; a ₂ = - 2.656; a ₃ = 0.901; a ₄ = - 0.090; and the residual R = 0.007.

Then, the mean value and standard deviation of each of the specified informative features are determined in each section of the identified personalized weld seam, which are subsequently used for classification in the trained model.

To determine informative features in all sections of the longitudinal section of each identified personalized weld, a section passing through the center of the weld is used. To determine the central longitudinal section of each section of the weld, the previously determined coordinates along the Y axis of the weld boundaries are approximated by a linear relationship. The points lying on the line passing through the middle between the lines approximating the boundaries of the weld section are selected as the central section. Fig. 6 shows the stages of determining the central longitudinal section.

For each section of the longitudinal section, the power spectrum of spatial frequencies is determined, characterizing the periodic fluctuations in the height of the weld profile (Fig. 7, 8).

The first 20 spatial frequencies of the spectrum (excluding the constant component) with height fluctuations with a period of at least 3 mm are used as informative features characterizing the longitudinal section of each identified personalized weld seam.

The tangent of the angle of inclination of the spatial oscillations of the weld height profile, determined by the position of the frequency peak on the two-dimensional spectrogram, is used as an informative feature extracted from the two-dimensional spatial spectrogram of each identified weld. The use of a two-dimensional spatial spectrogram allows, in addition to the informative features of the longitudinal and transverse sections, to characterize the slope of the spatial oscillations of the weld height profile.

All informative features extracted from two-dimensional and three-dimensional images of identified welds are given in Table 1.

Next, the most significant informative features for each identified personalized weld are selected for use as a training set in the trained models.

At this stage, to eliminate the influence of different scales of informative features, they are normalized. For this, μ _j is calculated - the average value of the j-th feature and σ _j - the standard deviation of the j-th feature for all sections for each identified personalized weld. Then, each feature is normalized using formula 5:

Where - normalized value of the j-th feature of the i-th sample;

μ _j is the average value of the j-th feature for all samples;

σ _j is the standard deviation of the j-th feature across all samples.

The selection of the most significant informative features, and, consequently, the reduction of their number by choosing those that most fully characterize the features of the identified personalized weld, provides the following advantages:

- simplifies the trained model to improve interpretability;

- reduces the dimensionality of the data processing process, which reduces the model training time and the required computing power for the classifier to operate;

- increases the generality of the classifier;

- allows to avoid the possible “curse of dimensionality”;

- improves the generalizing ability of the trained model and eliminates its overfitting.

The "curse of dimensionality" refers to the exponential growth of the number of required experimental data depending on the dimensionality of the space. With a limited number of identified personalized weld samples, a large number of informative features will lead to low generalization ability. Generalization ability is associated with the ability to work not only on the training sample of already existing identified personalized welds, but also on new samples that have not been previously used in training. The problem of reducing the generalization ability of the trained model is associated with the danger of overfitting or overtraining, when an attempt is made to describe specific data more accurately than the noise level in the data and the error of the trained model itself allow in principle. In this case, the trained model will be based not only on real differences in the data characteristic of different classes, but also on errors present in the experimental data. For example, the problem of selecting informative features was solved in the Python 3.8 programming language using the "SelectKBest" class from the "scikit-learn" library, which implements one-dimensional feature selection ("univariate feature selection").

The selection of the most significant informative features for further training of the models is performed by constructing the distribution of informative features in each section of each identified personalized weld. To select the most significant informative features, a criterion based on mutual information is used, which allows estimating the relationship between random values of informative features. The required number of informative features is determined by enumerating options in a certain range. In this way, those informative features are selected that have the highest repeatability (Fig. 9). Among the selected most informative features, there is the average width of the weld cross-section "Mean_Width". This feature can be compared not with the individual handwriting of the welder, but with the features of the welded blanks, therefore, for the sake of example, two options were further considered: with and without using the "Mean_Width" feature (Fig. 10).

Next, several training models are created, for training which different image recognition methods are used. The selected informative features of identified personalized welds serve as the initial technical data for training the models, and the full names of the identified welders serve as the result. Models are trained with different parameters and different numbers of informative features selected at the previous stage. Table 2 shows the image recognition methods used for training the models.

Next, the most accurate training model and its parameters are determined, as well as the optimal number and specific set of informative features. To do this, the accuracy of each training model is calculated using formula 6, and the most accurate one is left for further implementation of the proposed method.

where A is the accuracy;

P - total number of identified welds.

Fig. 11 and 12 show examples of the dependence of the accuracy of different trained models on the number of selected informative features. The accuracy of the trained models is plotted along the ordinate axis, and the number of informative features is plotted along the abscissa axis. The graph in Fig. 12 is constructed for the case when the optimal set of informative features was selected from all available informative features, and when constructing the graph in Fig. 13, the weld cross-section width feature "Mean_Width" was removed from the original features before determining the optimal set. The graphs in Fig. 13 and 14 are provided for comparison; they show the accuracy of the trained models on the training data set as an example. On the training set, with a large number of informative features, almost all trained models show an ideal result (classification accuracy of 100%).

In Fig. 11-14, identical designations are adopted: 1 - k-nearest neighbors method, k=1; 2 - k-nearest neighbors method, k=3; 3 - support vector method C=10, gamma=0.1; 4 - support vector method C=10, gamma=0.5; 5 - support vector method C=10, gamma=0.7; 6 - random forest method.

From the graphs in Fig. 12 and 13 it is evident that the exclusion of the “Mean_Width” feature reduces the classification accuracy, but even in this case it is possible to achieve a classification accuracy of 0.78 (support vector machine, C=10, gamma=0.1, set of 7 informative features). When using the entire set of features for determination, the classification accuracy increases and reaches 0.85.

As an example, Table 3 provides a detailed report of the statistical characteristics of the optimal training model (support vector machine, C=10, gamma=0.7; set of 6 informative features) when using the "Mean_Width" informative feature set. The optimal set of informative features: Mean_Width, Mean_CntrMass, Mean_p2Coef, Mean_p4Coef, Mean_Residue, FFT_tg.

Table 4, as an example, provides a detailed report of the statistical characteristics of the optimal training model (support vector machine, C=10, gamma=0.1, set of 7 informative features) without using the "Mean_Width" feature in the set of informative features. The optimal set of informative features: Mean_CntrMass, Mean_MaxPos, Mean_p2Coef, Mean_p3Coef, Mean_p4Coef, Mean_Residue, FFT_tg.

From tables 3 and 4 it is evident that the individual characteristics of the welded seam (“handwriting”) of welder 1 are most easily identified, and the greatest number of errors occur when identifying welded seams made by welder 5 - up to 50%, in the case where the “Mean_Width” feature is not used.

Next, the performance of the optimal training model is checked. The performance of training models cannot be checked on the same sample on which the training was conducted, so, for example, the original sample was randomly divided into two parts: in our case, 67% of the samples were used to train the models, and 33% were used to check the quality of the model.

Together with the correlation method or separately from it, a matrix of the form “n*k” is constructed, where n is the width of the matrix and k is the length of the matrix as follows:

- form an initial matrix of the form “n*k”, the cells of which contain values, for example, of the height for each point of the identified weld seam relative to the laser scanning device.

- if necessary, the matrix is reduced by an amount sufficient to compress the data to a size that will make it possible to further form a set of unique identification features of the identified personalized weld. The reduction is performed by successive averaging of adjacent matrix cells by rows and columns. In this case, the cells will take the form of formula 7. The operation is performed several times by rows and columns. As a result of the reduction, a secondary matrix of the form (n/p)⋅(k/p) is obtained, where p is the reduction coefficient.

- form a linear matrix of the third order, consisting of functions of formula 8, describing, to the selected degree of approximation, the fluctuations in the heights of the identified weld seam.

The sets of oscillations according to formula 8 will be unique for each identified weld, however, such sets will contain a system of functions of the form of formula 9, describing unique sets of curves, characteristic personally of each identified welder. The general set of such curves will constitute the features of the weld, characterizing the identified welder, the so-called "handwriting" of the welder.

- form matrices of type 9 by applying the Discrete Fourier Transform (hereinafter referred to as DFT), using functions according to formula 10.

where ω are the cyclic frequency components of the functions.

- supplement the third-order matrices with three columns containing sets of functions reconstructed from curves that constitute the oscillations of the weld seam boundaries in width and the oscillations of the weld seam width. After supplementation, the matrix will take the form of formula 11.

- add sets of curves obtained from identified personalized welds and obtain a matrix according to formula (12).

- save the correspondence numbers of personalized data (full name) of identified welders and sets of Fourier curves in a separate linear matrix of the form of formula (13):

- extract from the obtained matrices the data contained in them on the characteristics of the welded seams ("handwriting") of the welder. To do this, supplement the matrix according to formula 12 with correlation coefficients r _k and obtain a matrix according to formula 14:

- obtaining correlation coefficients in the following way. Compare sets of curves in accordance with formula 9 using the direct correlation method using formula 15, add correlation coefficients sequentially first for identified personalized welds performed by one welder r _k , then (1-r _k ) for identified welds performed by different welders. The final matrix will look like formula 14, however, the coefficients r _k will have different values depending on the iteration number. The iterative introduction of correlation coefficients using formula 15 is repeated until the correlation with identified personalized welds performed by different welders is equal to zero or tends to zero at the level of the selected error. The selected error is the necessary basis for further calculation of the degree of reliability of data processing.

After creating a database with identified personalized welds, the welder is identified using the VIC method. VIC is performed for all welds, which consists of measuring the following technical parameters for each weld: width, height of the weld, width of the weld scales, distance between scales, differences in scale heights, and depressions between scales. For circumferential welds, additional inspection and measurement of lock craters, places of re-ignition and extinguishing of the welding arc when changing the electrode are performed. The following metrological means are used for these measurements:

- calipers, metal and wooden rulers, thickness gauge, micrometer - for measuring linear informative features;

- templates, feeler gauges, angle gauges with vernier, micrometer and indicator bore gauges - for measuring linear and angular informative features.

The obtained technical parameters of each weld are compared with the above-mentioned database in order to find similarities with any technical parameters of the identified personalized weld. If similarities are found, the weld is personalized.

After the comparison above, the weld seam is personalized with further determination of the welder's personal data using the correlation method. Using a laser scanning device connected to a computer with a loaded database of identified personalized weld seams, unidentified weld seams that need to be personalized are scanned.

To identify the author of a weld, the distribution of informative features of such an unidentified weld is compared with the database of identified personalized welds using the created and verified accurate learning model. The accuracy indicator of the match of an unidentified weld with identified ones is determined. If the accuracy indicator for an unidentified weld is more than 60%, such identification is considered preliminary reliable. If the accuracy indicator is less than 60%, then a conclusion is made that the database of identified personalized welds does not contain data on the welder and it is not possible to establish his identity based on the features of the unidentified weld under study.

After the comparison performed above, the weld seam is identified by the correlation method or instead of it, with subsequent determination of the welder's personal data using the matrix construction. Using a laser scanning device connected to a computer with a loaded database of identified personalized weld seams, unidentified weld seams that need to be personalized are scanned. The obtained data of the sets of functions of informative features for each point of the unidentified weld seam are compared with the values of the formed initial matrices contained in the said database. In the event of a discrepancy between them of 50% or less, the welder's identification by the weld seam features is considered successful. In the event of a discrepancy in accuracy indicators of more than 50%, a conclusion is made about the impossibility of accurately determining the author of the weld seam.

For the final identification of the welder, the accuracy indicators obtained at all three stages of the proposed method are compared.

This method is applicable for external butt welds, in particular on main pipelines made by manual arc welding.

Based on the results of the development of the claimed method, it was established that the following factors influence the possibility of determining informative features and further personalization of the welder based on the characteristics of the weld:

- the presence of mechanical impact on the weld seams being examined (cuts, dents, grinding of the surface of the beads in order to bring the geometric parameters to standard values, etc.);

- the presence of corrosion damage on the welded seams under study;

- the presence of rolling defects during pipeline production;

- presence of assembly defects;

- differences in the types of welds analyzed (pipe/plate/insert).

- number of welded seams;

- dimensions, length of the seam;

- reflectivity of the weld surface (the presence of glare leads to missing data recording);

- complex geometric shape of the weld (for example, pipe-to-pipe cutting), requiring the development of specialized devices to ensure that an image of the weld is obtained without distortion.

The influence of all the above factors is minimized in the claimed method, since it implements the ability and algorithm for selecting the most informative features of welds for analyzing, data processing is performed taking into account the leveling of all possible errors in collecting such data. Thus, the claimed method ensures not only accurate and fast identification of the welder by the features of an unidentified weld, but also the reliability of such a process due to the maximum exclusion of the human subjective factor, the ability to use multiple software tools to implement the claimed method and a high-quality database with a sufficient amount of weld data. In addition, the claimed method does not require the use of expensive equipment and materials and can be implemented in a short time.

Claims (4)

1. A method for identifying a weld seam, characterized in that a database of identified personalized weld seams is created by measuring technical parameters and laser scanning weld samples to obtain two-dimensional and three-dimensional images, dividing each identified personalized weld seam into sections no less than 70 mm long along the longitudinal scanning coordinate, using a section passing through the center of the weld seam, with the allocation of informative features in each section of each identified personalized weld seam, constructing static dependencies of informative features using a trained model and forming an initial matrix of the form "n * k", where n is the width of the matrix, and k is the length of the matrix containing in its cells sets of functions of informative features of the identified personalized weld seam, with subsequent placement of the created database of identified personalized weld seams in a computer memory unit, after which the technical parameters are measured and the unidentified weld seam is laser scanned, then the unidentified weld seam is personalized by visual and measuring control by comparing the measured technical parameters of the unidentified weld with the corresponding technical parameters of the identified personalized welds contained in the said database, after which the most significant informative features of the unidentified weld are selected by means of a correlation method with the construction of static dependencies of the informative features and their comparison using a trained model with the static dependencies of the informative features contained in the said database, and/or by constructing matrices of the form "n * k", where n is the width of the matrix, and k is the length of the matrix containing in its cells sets of functions of informative features for each point of the unidentified weld, with the implementation of a comparison of the obtained results with the values of the generated initial matrices contained in the said database, wherein in the event of a match between the informative features as a result of the above comparisons, a conclusion is made about the possibility of personalization of the unidentified weld, and in the event of a mismatch between the informative features, a conclusion is made that the identity of the welder of the unidentified weld seam is not installed. 2. The method according to paragraph 1, characterized in that the technical parameters used are the width, height of the weld seam, width of the scales of the weld seam, distance between the scales, differences in the heights of the scales, depressions between the scales, lock craters, places of re-ignition and extinguishing of the welding arc when changing the electrode. 3. The method according to paragraph 1 or 2, characterized in that the technical parameters are measured using a caliper, ruler, thickness gauge, micrometer, templates, feeler gauges, protractors with vernier, micrometer and indicator bore gauges. 4. The method according to claim 1, characterized in that the following are used as informative features: the average value of the width of the cross-section profile, the average value of the area under the curve of the weld profile, the average value of the center of mass of the profile section, the average value of the position of the maximum of the approximating polynomial, the average value of the coefficients of the approximating polynomial, the average value of the residual of the approximating polynomial, the standard deviation of the width of the cross-section profile, the standard deviation of the area under the curve of the weld profile, the standard deviation of the center of mass of the profile section, the standard deviation of the position of the maximum of the approximating polynomial, the standard deviation of the coefficients of the approximating polynomial, the standard deviation of the residual of the approximating polynomial, spectral readings of the power density of spatial frequencies, the tangent of the angle of inclination of the spatial oscillations of the weld profile.