IT201800021022A1 - "Procedure for monitoring the quality of a weld, relative welding station and IT product" - Google Patents

Description

DESCRIPTION of the industrial invention entitled:

"Procedure for monitoring the quality of a weld, relative welding station and IT product"

TEXT OF THE DESCRIPTION

Technical field

The embodiments of the present description relate to techniques for monitoring the quality of a weld.

Background

Figure 1 shows a typical welding station of an industrial plant.

In the example considered, welding is used to join two or more metal pieces M1 and M2 along a welding path. For example in Figure 1 two superimposed plates M1 / M2 are shown and the welding should be carried out along a known trajectory in a direction indicated with x.

In particular, in the example considered, the welding is carried out by means of an energy source which comprises a welding head 1, such as for example an electron source for carrying out an electron beam welding (EBW) or a photon source. , typically a laser source. Typically, the source and / or the welding head 1 has associated a control circuit configured to regulate one or more parameters of the source and / or the welding head 1, such as for example the power emitted by the source, the focusing of the beam supplied from the welding head, etc.

Furthermore, the welding station comprises at least one actuator 2 for moving the beam supplied by the welding head 1 along the welding path. For example, this can be achieved by rotating the welding head 1 and / or (as shown in Figure 1) moving the welding head 1 with respect to the pieces M1 and M2 and / or moving at least one axis of the optical chain. For example, in Figure 1 a robotic arm 2 is shown for this purpose. Therefore, typically, also the actuator or actuators 2 have associated a control circuit 20 configured to drive the actuator or actuators 2 to move the beam supplied by the welding head 1 along the welding path.

Therefore the beam of electrons or photons generated by the source and supplied by the welding head 1 strikes the upper piece M1 along the welding path and melts the materials of the pieces M1 and M2 in a welding area SA, thus creating a bead of welding. In general, the metal pieces M1 and M2 can have any shape and it is sufficient that the pieces M1 and M2 are in contact, or have complementary shapes, along the welding path / in the welding area SA. For this purpose, locking / clamping means are typically used which are configured to lock the pieces M1 and M2, in particular in the welding area SA, in such a way as to ensure appropriate contact between the pieces M1 and M2.

In addition, pieces M1 and M2 can also be made with different materials. For example, this is typically the case when batteries are to be made, particularly for electric vehicles. For example, for this purpose, reference can be made to documents US 2015/0207127 A1, US 2017/0341144 A1 or Das, A .; Li, D .; Williams, D .; Greenwood, D., “Joining Technologies for Automotive Battery Systems Manufacturing,” World Electr. Veh. J. 2018, 9, 22.

For example, in this case, soldering can be used to connect a first tab to a busbar and / or to a second tab. For example, this is schematically shown in Figures 2A to 2C. In particular, Figure 2A shows a sectional view, in which a first tab M1, for example in aluminum (Al), is welded to a busbar M2, for example in copper (Cu). Similarly, Figure 2B shows a sectional view, in which a second tab M3, for example made of nickel (Ni) or copper (Cu), is soldered to the busbar M2. Finally, Figure 2C shows a sectional view, in which the first tab M1 is welded to the second tab M3, and possibly also to the busbar M2. For example, such tabs and busbars can have a thickness between 0.3 and 0.8 mm.

Even if modern welding stations meet high quality and stability criteria over time, a verification of the quality of the weld may be required. For example, this is particularly relevant in the context of batteries for the automotive sector, since such batteries must have a uniform electrical resistance between the tabs and the busbars.

The skilled in the art will appreciate that a non-destructive monitoring / control method, so-called Non Destructive Testing, is typically required for continuous monitoring of welding quality. In particular, these non-destructive tests are experimental investigations aimed at identifying and characterizing discontinuities in the weld bead, which are potentially capable of compromising the performance of the final product. The common point of non-destructive testing techniques is therefore their ability not to influence in any way the chemical, physical and functional characteristics of the object analyzed. For example, in this context reference can be made to the UNI EN 473 standard.

Synthesis

The purpose of various embodiments of the present description are therefore new solutions which allow the quality of a weld bead to be monitored.

According to one or more embodiments, one or more of the preceding objects are achieved through a method having the distinctive elements set out specifically in the following claims. The embodiments also relate to a relative welding station, as well as a corresponding corresponding computer product, which can be loaded into the memory of at least one computer and comprising portions of software code for carrying out the steps of the process when the product is executed on a computer. As used herein, a reference to such a computer product is intended to be equivalent to a reference to a computer-readable medium containing instructions for controlling a computer system to coordinate the execution of the procedure. A reference to “at least one computer” obviously intends to highlight the possibility that the present description is implemented in a distributed / modular manner.

The claims are an integral part of the technical teaching of the description provided here.

As explained above, various forms of the present disclosure relate to a method for analyzing the quality of a weld bead in a weld zone.

In various embodiments, the welding seam is generated by means of a continuous welding in which an energy beam supplied by a source with relative welding head follows a welding path, melting the material of at least two metal pieces.

In various embodiments, the welding area is monitored by means of a thermal imaging camera. In particular, the thermal imager provides a series of thermal images / frames in which a certain area corresponds to the welding area. For example, this area can be determined by welding. For example, a rectangular or trapezoidal area can be defined in the thermal image as a region of interest. In the hypothesis of an automatic search for the region of interest, starting from an observation window of fixed and set dimensions, the processing circuit 30 can position the region of interest in the image by maximizing the functional represented by the sum of the temperatures of the pixels in it. included for all frames.

In various embodiments, this area is divided into a plurality of sub-areas and for each sub-area a respective temperature is determined as a function of the values of the pixels comprised in the respective sub-area. For example, the temperature of a given sub-area can be determined by averaging or a weighted average of the pixel values included in the respective sub-area.

During a learning phase, a plurality of welds are carried out, in particular at least for a plurality of examples in which the weld has a sufficient quality and a plurality in which the process is not of quality. In addition, the temperature trend of each sub-area is monitored during each welding.

During a training phase, the temperature trends monitored during the learning phase are processed to train a classifier. For example, for this purpose, an operator can classify the quality of the welds, i.e. the system can receive (from the operator) data that identify the respective weld quality for each weld. Therefore, the classifier is configured to estimate a weld quality as a function of respective temperature trends.

In particular, in various embodiments, a respective cooling curve is extracted from each temperature trend and for each cooling curve parameters are determined that identify the trend of the cooling curve. Therefore, in various embodiments, these parameters are used as input features for the classifier.

For example, in various embodiments, the cooling curve trend is approximated for this purpose by interpolation with a function consisting of a plurality of basic functions, thus selecting a respective set of parameters for each basic function. Therefore, in this case, the parameters of the interpolation can be used as input features for the classifier. For example, in various embodiments, exponential interpolation is used.

Instead, during a normal welding operation phase, the temperature trend of each sub-area can be monitored again during one or more welds and the respective weld quality can be estimated using the previously trained classifier. For example, the classifier can be an artificial neural network. In general, the same or an additional classifier can also be used to estimate a defect class.

In general, the classifier can also receive one or more further input features, such as for example the peak of each temperature trend, the power emitted by the source, the speed of advancement with which the energy beam follows the welding path, etc.

Brief description of the figures

The embodiments of the present description will now be described with reference to the attached drawings, which are provided purely by way of non-limiting example, and in which:

Figure 1 shows an example of a soldering station;

- Figures 2A to 2C show some examples of welds;

Figure 3 shows an embodiment of a soldering station which comprises a thermal imaging camera;

Figure 4 shows an embodiment, in which the welding station of Figure 3 fuses two materials in a welding zone;

- Figure 5 shows an example of the image of the welding area as taken by the thermal imaging camera of Figure 3;

Figure 6 shows an embodiment of a segmentation of the welding zone into different sub-zones;

- Figure 7 shows an example of the temperature trends of the sub-zones of Figure 6;

Figure 8 shows an embodiment for extracting a cooling curve from a respective temperature trend;

Figure 9 shows an embodiment for determining a welding quality as a function of the extracted cooling curves;

Figures 10 and 11 show embodiments of the classifier used in Figure 9; And

- Figure 12 shows an embodiment for the training and use of the classifier of Figure 9.

Detailed Description

In the following description, numerous specific details are given to provide a thorough understanding of the embodiments. The embodiments can be implemented without one or several specific details, or with other processes, components, materials, etc. In other cases, well-known operations, materials or structures are not represented or described in detail so as to avoid making the aspects of the embodiments unclear.

A reference throughout this description to "an embodiment" means that a particular feature, distinctive element or structure described with reference to the embodiment is included in at least one embodiment. Thus, the appearances of the phrases "in one embodiment" in various places throughout this description do not necessarily all refer to the same embodiment. Furthermore, the particular features, distinctive elements or structures can be combined in any suitable way in one or more embodiments.

The references used here are provided simply for convenience and therefore do not define the scope of protection or the scope of the forms of implementation.

In the following Figures 3 to 12, the parts, elements or components which have already been described with reference to Figures 1 and 2 are indicated with the same references used previously in these Figures; the description of these elements described above will not be repeated later in order not to overload the present detailed description.

As mentioned above, the present disclosure provides solutions for monitoring the quality of a weld bead.

Figure 3 shows an embodiment of a welding station in accordance with the present description. Basically, the embodiment is based on the soldering station described with reference to Figures 1 and 2, and the related description applies entirely. In particular, also in this case, the welding station is configured to melt the material of several metal pieces M1 and M2 in a welding zone SA. For this purpose, the soldering station includes:

- an energy source with relative welding head 1, preferably a laser source, controlled by means of a respective control circuit 10; And

- one or more actuators 2, such as for example a robotic arm, controlled by a control circuit 20 in such a way as to move the beam supplied by the welding head 1 along a welding path.

Then, as also shown in Figure 4, the beam supplied by the welding head 1 moves along a welding path SP and heats the material of the overlapping pieces M1 and M2 in a welding zone SA, thus creating a weld bead through a fusion of the materials of the pieces M1 and M2. As shown in Figure 4, in general, the welding path SP does not necessarily start and end at the edges of the pieces M1 and M2. Furthermore, the weld path SP can be of any shape, although a straight path along an x direction is preferable.

In the embodiment considered, the welding station further comprises a thermal imaging camera 3. In particular, in various embodiments, the thermal imaging camera 3 is mounted in a fixed position and positioned in such a way as to frame the welding area SA, i.e. the thermal imaging camera 3 is configured to provide an IMG thermal image which captures the welding zone SA. In general, the thermal imaging camera 3 can also be implemented with a plurality of thermal imaging cameras, in which each thermal imaging camera only captures a part of the welding area SA, i.e. the IMG image can correspond to a panoramic image resulting from the union of the images provided. from a plurality of thermal imaging cameras. The thermal imager (s) then provide a two-dimensional IMG image in two directions x 'and y' (see also Figure 5), in which the value of each pixel identifies a respective temperature.

In the embodiment considered, the IMG thermal image is then processed by a processing circuit 30, such as for example a microprocessor programmed through software code, such as for example a computer. In general, the processing circuit 30 can also be implemented together with the control circuit 10 and / or the control circuit 20 in a single electronic circuit.

For example, Figure 5 schematically shows the IMG thermal image provided by the thermal imaging cameras 3, in which a certain area SA 'corresponds to the welding area SA. Therefore, by welding two pieces M1 and M2, the value of each pixel in the SA ’area identifies the temperature of a respective point of the SA welding area. In general, the zone / area SA 'in the IMG image can be selected manually or the processing circuit 30 can determine the zone SA' automatically. In particular, by carrying out a welding, the pixels in the SA 'area will have higher values, or temperatures, and the processing circuit 30 can therefore detect the SA' area in the IMG image which corresponds to the SA welding area. In particular, in the case where the SP welding path is straight, the SA 'area will typically have a rectangular area or, considering any distortions of the IMG image, trapezoidal.

Therefore, in various embodiments, the processing circuit 30 is configured to compare the value of each pixel of the IMG image (or in a sequence of IMG images) with a reference threshold, selecting the pixels which have a value greater than a threshold and approximate the area where the selected pixels are located with a rectangular or trapezoidal area.

Instead, in a currently preferred embodiment, the size of the rectangle (or of the trapezoidal area) is fixed. For example, knowing the size of the welding area SA, the processing circuit can calculate the size of the rectangle from the parameters of the thermal imaging camera 3, eg. the focal length and the distance from the piece M1. Alternatively, the size of the rectangle (or of the trapezoidal area) can be set by an operator. Subsequently, once a welding has been carried out, the processing circuit 30 positions this rectangle (or trapezoidal area) on a plurality of positions in the thermal image IMG and calculates for each position the sum of the values of the pixels that are inside the rectangle (or trapezoidal area) for all frames. Finally, the processing circuit 30 selects the position / area which has the highest sum. Therefore, in this case, the processing circuit chooses as the zone SA 'the zone (with fixed size) that includes the pixels that have the maximum value as a sum, thus realizing a compensation of small displacements of the welding path SP for each welding.

As explained above, in various embodiments, the control circuit 20 is configured to move by means of the actuator or actuators 2 the beam supplied by the welding head 1 along a straight line in an x direction. In this case, the thermal imaging cameras 3 are preferably aligned in such a way that the x 'or direction (as shown in Figure 5) the y' direction of the IMG image corresponds to the x direction. For example, in this way the rectangular area SA 'is also aligned with the pixel matrix of the IMG image.

Alternatively or additionally, the processing circuit 30 can process the thermal image IMG to correct the image acquired by the thermal imager 3, for example rotate the image IMG in such a way as to align the x direction with one of the x 'or y' axes. of the IMG image, compensate for the distortion of the IMG image due to the inclination of the thermal imager 3 relative to the surface of the workpiece M1 and / or the deformation of the IMG image due to the lens of the thermal imager 3. Similar operations are widely known in the context of traditional cameras and can also be applied to images obtained from thermal cameras. For example, as described in US 2018/0082133 A1, knowing the mounting of the chamber, the distortion compensation can be made based on the tilt information.

In the embodiment considered, the processing circuit 30 then processes the pixel values in the area SA '.

In particular, as shown in Figure 6, in various embodiments, the processing circuit 30 divides the area SA 'into a plurality of sub-areas A1..An. For example, considering that the area SA 'has a certain number of pixels N1 in width, for example in the direction x' in Figure 5, each sub-zone A1..An can have a width of N1 pixels and a height of N2 pixels. For example, to increase the accuracy of the analysis, the number of pixels N2 can be chosen between 2 and 20 pixels. On the other hand, to reduce the computation time, the number of sub-zones A1..An can be chosen between 10 and 50, for example on the basis of the length of the welding seam / welding area SA, and the relative number of pixels. N2 can be calculated according to the selected area number A1..An. Therefore, in various embodiments, the number N2 can be selected, for example, between 0.2 · N1 and 2 · N1, preferably between 0.2 · N1 and 0.5 · N1.

Subsequently, the processing circuit 30 processes the values of the pixels in each sub-zone A1..An to associate each sub-zone A1..An with a single instantaneous temperature value Ti. For example, the processing circuit 30 can calculate the temperature value Ti of a given sub-zone Ai by using, for example, the average value or the maximum value of the pixel values in the respective sub-zone Ai. For example, in various embodiments, the processing circuit 30 is configured to calculate the temperature value Ti of a given sub-zone Ai by means of a weighted average which associates to each pixel a weight which varies in the width direction of the area SA '(for example x' in Figure 5), for example by using a lower weight for the pixels at the lateral edges of the respective sub-zone Ai and a higher weight for the central pixels of the respective sub-zone Ai.

Therefore, for each IMG image, each sub-zone Ai will have associated a respective temperature value Ti. Furthermore, as shown schematically in Figure 6, by analyzing a sequence of a plurality of thermal images IMG in time t, or a video, the processing circuit 30 can monitor the trend of the temperature Ti (t) of each sub-zone Ai.

As shown in Figure 7, also considering the time necessary to follow the welding path SP by means of the beam provided by the welding head 1, the processing circuit 30 should therefore analyze a plurality of temperature curves / trends T1..Tn which are out of phase with each other.

As shown in more detail in Figure 8, each Ti (t) temperature curve comprises:

- during a heating phase (between instants t0 and t1 in Figure 8) an increase of the temperature Ti (t) from an ambient temperature Tamb to a maximum temperature Tmax, since the respective area Ai is subjected to the beam supplied by the welding head 1 to carry out the welding, e

- during an immediately following cooling phase (from the instant t1 in Figure 8) a decrease in the temperature Ti (t) from the maximum temperature Tmax towards the ambient temperature Tamb.

For example, in various embodiments, the processing circuit 30 can start recording the temperatures Ti (t) when the control circuit 10 provides a trigger signal that signals the start of a weld. On the other hand, the duration of the recording of the temperatures Ti (t) can be constant.

In general, a datum indicative of the welding quality is the maximum temperature Tmax reached, since this datum is indicative of the melting of the materials M1 and M2.

However, the inventors observed that, while achieving the same maximum temperature Tmax, the profile of the cooling curve varies according to different welding defects, for example as a result of contamination of the welding area, for example due to the presence of drops of water. water or dust.

Therefore, in various embodiments, the processing circuit 30 is configured to analyze the cooling curve and determine a welding status signal S as a function of the cooling curves, i.e. of the data Ti (t), with t> t1, for all sub-areas A1..An.

For example, in a first embodiment, the processing circuit 30 is configured to record for each sub-area A1..An a respective reference cooling curve during a test phase in which the weld is classified as correct (e.g. S = 1 / OK) and then the processing circuit 30 records during normal operation for each weld carried out a respective cooling curve and classifies the status (e.g. S = 1 / OK or S = 0 / NOK) of each weld by comparing the respective cooling curve recorded with the reference cooling curve. For example, a possible solution for determining the similarity between two data sequences and a respective classification of the similarity is described in the Italian patent application 102017000048962, the content of which is incorporated herein by reference.

Instead, Figure 9 shows a second embodiment. In particular, in the embodiment considered, the processing circuit 30 processes, as described above, at a preprocessing step / block 300 the sequence of images IMG provided by the thermal imager 3 to determine a plurality of temperature curves Ti (t) for the respective area A1..An.

Such temperature curves Ti (t) are supplied to a step / block 302, in which the processing circuit 30 processes the temperature curves Ti (t). In particular, as previously described, the processing circuit 30 is configured to extract the data of the cooling curve, for example by identifying the instant t1 in which the curve Ti (t) reaches a maximum value Tmax, and selecting the data of the Ti (t) curve with t> t1. Subsequently, the processing circuit 30 processes the cooling curve and determines one or more characteristics of the cooling curve, so-called features F. Therefore, the step / block 302 performs a so-called feature extraction.

For example, in various embodiments, a first feature corresponds to the maximum temperature Tmax. Other features F can identify the descent of the cooling curve, for example one or more values indicating the time required until the temperature Ti drops to a certain percentage of the maximum temperature Tmax, for example:

- a first time Δt1 until the temperature Ti drops to 75% of the temperature Tmax;

- a second time Δt2 until the temperature Ti drops to 50% of the temperature Tmax; And

- a third time Δt3 until the temperature Ti drops to 25% of the temperature Tmax.

Instead, in a currently preferred form, the processing circuit 30 performs an interpolation operation to approximate the trend of the cooling curve Ti (t), with t> t1, with a parameterized function. In general, this parameterized function is composed of one or more basic functions, in which each basic function has associated a respective set of parameters. Therefore, by varying the parameters of the basic functions, a combination of parameters a0..am can be chosen which minimizes a cost function. For example, the cost function can correspond to the sum of the absolute differences (Sum of Absolute Difference, SAD) or the mean squared error (MSE) calculated between the trend of the cooling curve and the parameterized function it uses. the chosen parameters.

For example, in various embodiments a polynomial interpolation is used in which the basic functions are represented by polynomials of different degrees and the parameters are the coefficients of the polynomial, for example:

Instead, in a currently preferred embodiment, an exponential interpolation is used, in which the basic functions are exponential functions, for example:

Therefore, at the end of the interpolation, the processing circuit 30 chooses as features F (possibly in addition to the maximum temperature Tmax) the parameters a0..am selected during the interpolation.

In various embodiments, the circuit 30 can also determine other features F. For example, for this purpose, the step / block 302 can receive a first set of data D1 from the control circuit 10 of the source / welding head 1 and / or a second data set D2 from the control circuit 20 of the actuator (s) 2 (see also Figure 3). For example, the data D1 may include the power emitted by the source and / or the focus of the welding head 1. Instead, the data D2 may include the speed of advancement of the beam supplied by the welding head 1 along the welding path SP. However, other sensors can also be used and / or the processing circuit 30 can determine further characteristics as a function of the thermal image IMG provided by the thermal imager 3.

For example, in various embodiments, by analyzing the thermal image IMG, the processing circuit 30 can determine in step 300 also dimensional parameters of the welding keyhole, as described for example in document US 2010/0086003 A1, the content of which is incorporated here for reference.

For example, in various embodiments these features can include, for each IMG image during the rewelding phase, i.e. between t0 and t1, the dimension in the x 'and / or y' directions of the keyhole and / or a parameter that identifies the heat distribution in the keyhole.

Additionally or alternatively, the processing circuit 30 can determine the spectral characteristics of each IMG image, for example by means of a Fast Fourier Transform (FFT), and choose a given number of frequencies which have the maximum values.

Additionally or alternatively, the processing circuit 30 can process the last purchased IMG image. In particular, the inventors observed that in this case all the pixels should have substantially the same value, since the pieces M1 and M2 have cooled down. However, when using pieces M1 and M2 that have different materials, you may notice pixels that have substantially different values (i.e. higher or lower) than the average of the pixels. In particular, these pixels correspond to splashes of the material of the lower piece M2 which have deposited on the surface of the piece M1. In particular, such sketches may be visible, as different materials also have different emissivity. Therefore, in various embodiments, the processing circuit 30 can determine such pixels which appear warmer or colder, for example by comparing the value of each pixel with a threshold, calculated for example as a function of all the pixels of the IMG image or only as a function of a certain number of pixels surrounding the respective pixel. Therefore, an additional feature could be the number of the hottest or coldest pixels.

Therefore, in general, the block / step 302 provides a plurality of features F, in which at least a part of the features F identify the trend of the cooling curves Ti (t), with t> t1, of the sub-areas A1 .. An.

In various embodiments, these features F are then provided to a pitch / block 304 configured to classify the state S of the weld as a function of the features F. In particular, in various embodiments, the classifier of the pitch / block 304 is implemented with a machine learning method.

In particular, as shown in Figure 12, after a starting step 1000, the processing circuit 30 monitors a plurality of welds at a learning step 1002. In particular, for this purpose a plurality of welds are carried out with different welding conditions. For example, for this purpose it can be:

- the power emitted by the source and / or the focus of the welding head 1 changed by means of the control circuit 10, and / or

- changed by means of the control circuit 20 the speed of advancement of the beam supplied by the welding head 1 along the welding path SP, and / or

- contaminated the welding area SA, for example with splashes of water and / or dust, and / or

- changed the tightening / blocking between pieces M1 and M2, for example by varying the clamping force.

Subsequently, an operator can check the quality of the weld. For example, the operator can carry out mechanical tests (for example mechanical resistance tests of the connection between the pieces M1 and M2) and / or electrical tests (for example measurements of the electrical resistance between the two pieces M1 and M2), and the operator can classify the quality of the weld as sufficient (eg S = 1) or not sufficient (eg S = 0). In general, one or more of the tests used in this phase can also be destructive, for example the mechanical tests can include a tensile test in which the tensile force is increased until the connection between the pieces M1 and M2 is broken.

Therefore, the data acquired in step 1002 represent a training dataset that includes experimental data both for conditions in which the weld has sufficient quality and for conditions in which the weld has insufficient quality.

Therefore, during a training phase 1004, the processing circuit 30 can extract the features F at least from the cooling curves of the areas A1..An (see also the description of Figure 9) and train the classifier 304 using the features F as data of the classifier 304 and the state of the weld S as the output of the classifier 304. In general, different classifiers of the category supervised learning, or supervised machine learning, such as artificial neural networks or support vector machines can be used.

For example, in various embodiments, an artificial neural network is used, such as for example a feed-forward network. For example, in various embodiments, this network comprises an input layer which comprises a number of input nodes equal to the number of features F. Furthermore, the network comprises a certain number of hidden layers, so-called hidden layers. For example, in various embodiments, the number of hidden layers is between 2 and 5, preferably 3, and the number of nodes / neurons of each hidden layer is chosen between 1.5 and 3 times the number of features F.

Therefore, at the end of step 1004 the classifier 304 is able to estimate the quality of a weld as a function of a set of features F extracted at least from the trend of the cooling curves for areas A1..An.

Therefore, once the step 1004 has been completed, the welding station can be used during a normal operation step 1006 in which the quality of the weld is to be estimated without further checks by an operator.

Therefore, in step 1006, the processing circuit 30 monitors again the trend of the cooling curves (see also the description of the step / block 300), determines the features F (see also the description of the step / block 302) and uses the classifier trained to estimate the state / quality of the weld S as a function of the features F (see also the description of step / block 304).

In general, an operator can still carry out further tests to verify the quality of the weld as described with reference to step 1002. For example, this can be useful during the initial stage of setting up a new welding process in order to verify the estimate of the classifier 304 and / or to periodically monitor the results of the estimate, for example to obtain additional data that has not previously been taken into account.

Therefore, as schematically shown in Figure 12, if the operator determines at a verification step 1008 that the result of the classifier is correct (output "Y" of the verification step 1008) the process can continue to step 1006.

On the other hand, in the event that the operator determines at the verification step 1008 that the result of the classifier is wrong (output "N" of the verification step 1008), the operator can store the data of the weld performed and the respective correct quality in the dataset and can start step 1004 to train the classifier again.

Therefore, in various embodiments, the data acquired during normal operation 1006 can be used itself as a training data-set. For example, for this purpose, the processing circuit 30 can be configured, for example by means of a suitable computer program, to store the training dataset directly in the processing circuit 30 and also directly manage the training phase 1004, thus allowing a new classifier training when the training dataset changes.

In various embodiments, the classifier 304 can be configured to provide not only a binary result S, i.e. indicate sufficient or insufficient quality, but can also provide indication C on the type of the defect. For example, for this purpose, the operator can also store information on a type of defect detected in step 1002 in the training dataset. For example, such defects can correspond to the different welding conditions used in step 1002, for example insufficient tightening, impurities / contamination of the pieces M1 / M2, a loss of power of the source, etc.

For example, this is schematically shown in Figure 10. In particular, in the embodiment considered, the classifier 304 comprises a first classifier 306 configured to estimate the state S of the weld which can therefore be corrected or defective. The output of the classifier 306 may still correspond to a continuous value, for example in the range between 0 and 1, which indicates the confidence of the estimate. The classifier 306 can then determine the state S according to the value provided, for example assigning a first value (e.g. S = 1 / OK) if the output value is greater than a first threshold (e.g. 0 , 8), or a second value (eg S = 1 / OK) if the output value is lower than a second threshold (eg 0.2).

In addition or alternatively, the classifier 304 comprises a second classifier 308 configured to estimate the defect class C of the weld which can therefore have multiple values. For example, this is schematically shown in Figure 11, where the values of two features F1 and F2 are mapped to four classes C1..C4. In general, the number of dimensions to be considered corresponds to the number of features F taken into consideration.

For example, in various embodiments, the classifier 308 comprises for each class C a respective output which provides a continuous value of indicative of the distance of the point represented by the combination of the current values of the features F with respect to each class C, i.e. each cluster, for example in the range between 0 and 1. For example, in this case, the classifier 308 can choose the class C which has associated the higher value, possibly limiting the choice only to clusters for which the respective distance is less than a maximum value.

Therefore, during step 1002, the operator can determine not only the state S of the weld, but possibly also the type of defect C. In general, since it is a machine learning approach, the classifier 308 is therefore able to adopt number of classes of defects C that the operator wants to consider, also allowing the addition of new types of defects that emerge only during normal operation 1006 of the welding station (see also the description of Figure 12). For example, during the step 1006 a situation may arise in which the quality of the weld becomes insufficient because the lens of the welding head 1 gets dirty, while this problem was not taken into consideration during the step 1002.

Naturally, without prejudice to the basic principles of the invention, the construction details and the embodiments can vary widely with respect to what has been described and illustrated here purely by way of example, without thereby departing from the scope of the present invention. as defined by the upcoming claims.

Claims (11)

CLAIMS 1. Process of analyzing the quality of a welding seam in a welding area (SA), said welding seam being generated by a continuous welding in which an energy beam supplied by a source with relative welding head (1) follows a welding path (SP) by melting the material of at least two or more metal pieces (M1, M2), the process comprising the steps of: - monitoring said welding area (SA) by means of a thermal imaging camera (3), in which said thermal imaging camera (3) provides a series of thermal images (IMG), and in which a determined area (SA ') in said thermal image (IMG) corresponds to said welding zone (SA); - dividing (300) said area (SA ') into a plurality of sub-areas (A1..An) and determining for each sub-area (Ai) a respective temperature (Ti) as a function of the values of the pixels included in the respective sub-area -area (Ai); - during a learning phase (1002) in which a plurality of welds are carried out both with sufficient quality and with insufficient quality, monitor by means of said thermal imaging camera (3) the temperature trend (Ti (t)) of each sub-area ( Ai) during each weld; - during a training phase (1004), elaborating the temperature trends (Ti (t)) monitored during said learning phase to train a classifier (304) configured to estimate a welding quality as a function of respective temperature trends (Ti (t)), in which said elaboration of the temperature trends (Ti (t)) includes: - extract (302) from each temperature trend (Ti (t)) a respective cooling curve and determine for each cooling curve a plurality of parameters (F) that identify the trend of the cooling curve; And - using said parameters (F) as input features for said classifier (304); - during a normal operating phase of the weld (1008), monitor (300) by means of said thermal imaging camera (3) the temperature trend (Ti (t)) of each sub-area (Ai) during a welding and estimate by means of said classifier (304) the respective weld quality (S, C). 2. Process according to claim 1, wherein said determining for each cooling curve a plurality of parameters (F) which identify the cooling curve trend comprises approximating by interpolation the cooling curve trend with a function consisting of a plurality of basic functions, selecting in this way a set of interpolation parameters, and using said set of interpolation parameters as input features for said classifier (304). Method according to claim 2, wherein said interpolation is an exponential interpolation. Method according to one of the preceding claims, wherein said determining for each sub-area (Ai) a respective temperature (Ti) comprises determining the temperature (Ti) by means of the average or a weighted average of the values of the pixels comprised in the respective sub-area ( To the). 5. Process according to one of the preceding claims, comprising determining said area (SA ') in said thermal image (IMG) through the steps of: - carry out a welding; - defining a rectangular or trapezoidal area of interest in said thermal image (IMG); - positioning said area of interest in a plurality of positions in such a way as to maximize the sum of the pixel values in said thermal image (IMG) for a plurality of frames. Method according to one of the preceding claims, wherein said classifier (304) comprises at least one artificial neural network (306, 308). Method according to one of the preceding claims, using (304) in addition to said parameters (F) one or more further input features for said classifier (304), in which said one or more further features are selected from: - the maximum temperature (Tmax) of each temperature trend (Ti (t)); - the power emitted by said source; - the speed of advancement with which said energy beam follows said welding path (SP); - one or more dimensional data of the keyhole produced during welding; and / or - at the end of the welding, the number of pixels in said thermal image (IMG) which has a substantially different value from the average value of the pixels in said thermal image (IMG). Process according to one of the preceding claims, comprising, following a normal operation step of the welding (1008): - check the quality of the welding; - comparing (1008) the quality of the weld estimated by said classifier (304) with the quality of the verified weld, and - in the event that the quality of the weld estimated by said classifier (304) does not correspond to the quality of the verified weld, train (1004) again said classifier (304) using the temperature trend (Ti (t)) of each sub- area (Ai) monitored both during said learning phase (1002) and during said normal welding operation (1008). 9. Process according to one of the preceding claims, comprising: - during said learning step (1002), classifying each weld which has an insufficient quality in a defect class (C) of a plurality of classes; And - during said training phase (1004), training a classifier (308) configured to estimate a defect class as a function of said temperature trends (Ti (t)). 10. Welding system, including: - a source with relative welding head (1) configured to supply an energy beam; - one or more actuators (2) configured to move said energy beam produced by said welding head (1) along a welding path (SP) in such a way as to melt the material of at least two metal pieces (M1, M2), - a thermal imaging camera (3); and - a processing circuit (30) operatively connected to said thermal imaging camera (3) and configured to implement the method according to one of the preceding claims. Computer product that can be loaded into a memory of at least one processor and comprises portions of software code for implementing the steps of the method according to one of claims 1 to 9.