GB2616336A

GB2616336A - Improvements in and relating to welding and quality control

Info

Publication number: GB2616336A
Application number: GB2219536.6A
Authority: GB
Inventors: Smart Matt; Smart Paul; Swain Ryan; Rooker Tim; Balakrishnan Jeyaganesh; Liu Xuanang
Original assignee: Cavendish Nuclear Ltd
Current assignee: Cavendish Nuclear Ltd
Priority date: 2021-12-22
Filing date: 2022-12-22
Publication date: 2023-09-06
Also published as: GB202219536D0

Abstract

A method of welding may comprise providing welding apparatus and a plurality of sensor types, and defining a first set of welding conditions. Substrate(s) to be welded may be introduced into the apparatus and welding conducted. During the welding process, data may be obtained from the sensors and correlated between at least two elected sensor types. One or more data points from at least one of the sensor types may be synchronised with those from another of the sensor types. Data from the plural sensor types may be compared with reference data and, based on the comparison, determined if the welding quality is acceptable or unacceptable. If the quality is unacceptable, action(s) may be taken. Apparatus for monitoring welding may comprise plural of inputs for data from plural sensor types, one or more processors for correlating and synchronising the data, and one or more outputs for processed data.

Description

IMPROVEMENTS IN AND RELATING TO WELDING AND QUALITY CONTROL

This disclosure concerns improvements in and relating to welding, quality control in welding and to methods of use thereof, based upon the use of multiple sensor types and combinations of data from those.

During welding, a significant number of variables exist which impact upon the formation of the weld and the quality of that weld. Attempts have been made to monitor individual variables during welding to detect deviations in those and potential impaired quality welding and/or defect generation as a result. As a weld is formed and immediately after formation, the weld has various characteristics. Attempts have been made to determine individual characteristics during welding to detect deviations in those and potentially impaired quality welding and/or defect generation as a result.

Amongst the potential aims of the disclosure is to provide a method of welding and verification of quality in which increased benefits are obtained by making greater use of variable monitoring and/or characteristic determining to mitigate against welding defects forming or other quality issues arising. Amongst the potential aims of the disclosure is provide a method of welding and verification of welding conditions which seeks to synchronise and fuse data from multiple sensor types and/or sources so as to obtain more information than the sum of the parts. In addition, potential benefits may be provided from seeking to analyse each of the data types in a more informative manner.

According to a first aspect of the disclosure there is provided a method of welding, the method of welding comprising: a) providing welding apparatus; b) providing a plurality of sensor types; c) defining a first set of welding conditions for the welding method; d) introducing one of more substrates to be welded to the welding apparatus; e) conducting welding of the one or more substrates; f) obtaining data from the plurality of sensor types during welding; g) providing a correlation between the data obtained from at least two elected sensor types in the plurality of sensor types; h) synchronising one or more data points in the data obtained from one of the at least two elected sensor types with one or more data points in the data obtained from another of the at least two elected sensor types; i) comparing the obtained data from the plurality of sensor types with reference data for one or more of the sensor types; j) based upon one or more such comparisons, determining whether the welding is of acceptable quality or unacceptable quality; k) where, if the welding is of unacceptable quality, the method includes then taking one or more actions.

The first aspect of the disclosure may include any of the features or possibilities or options set out in the first and/or second aspect of the disclosure relating to the inspection of a weld.

The method may be arc welding.

The method may provide that the correlation is both a temporal correlation and a positional correlation.

The method may provide that the correlation is a temporal correlation.

The method may provide that the data from the at least two elected sensor types include a time of occurrence for one or more data points within the data. The method may further provide that the correlation is a matching time of occurrence for data points in the data from the at least two elected sensor types.

The method may provide that the data from the at least two elected sensor types include a plurality of times at which a temporal correlation is provided.

The method may provide that the data from the at least two elected sensor types include a time of occurrence for all of the data points within the data. The method may provide that the data from the at least two elected sensor types include a time of occurrence for all of the data points within each range, such as a bandwidth, that the data is sub-divided into.

The method may provide that the time of occurrence is obtained relative to a timestamp introduced into the data from an elected sensor type for each of the two or more elected sensor types providing a correlation. The method may provide that the time of occurrence is obtained directly from a timestamp at that time, for instance as a timestamp is applied at the same time as the time of occurrence. The method may provide that the time of occurrence is obtained by the time elapsed since the timestamp and the time of the data point, for instance where a timestamp occurs and the time of occurrence is calculated from the elapsed time since the timestamp when the time of occurrence is reached.

The timestamp may be applied by a processor, for instance a processor including a clock generator and/or provided with a clock signal.

The method may provide that the correlation is a positional correlation.

The method may provide that the data from the at least two elected sensor types include a plurality of positions at which a positional correlation is provided.

The method may provide that the data from the at least two elected sensor types include a position of occurrence for one or more data points within the data. The method may further provide that the correlation is a matching position of occurrence for data points in the data from the at least two elected sensor types.

The method may provide that the data from the at least two elected sensor types include a position of occurrence for all of the data points within the data. The method may provide that the data from the at least two elected sensor types include a position of occurrence for all of the data points within each range, such as a bandwidth, that the data is sub-divided into.

The method may provide that the position of occurrence is obtained relative to a position stamp introduced into the data from an elected sensor type for each of the two or more elected sensor types providing a correlation. The method may provide that the position of occurrence is obtained directly from a position stamp at that position, for instance as a position stamp is applied at the same position as the position of occurrence. The method may provide that the position of occurrence is obtained by the time elapsed since the position stamp and the time of the data point, for instance where a position stamp occurs and the position of occurrence is calculated from the elapsed time since the position stamp when the position of occurrence is reached and/or from the elapsed distance travelled since the position stamp when the position of occurrence is reached.

The positional correlation may be relative to a position relating to the one or more substrates being welded, for instance relative to a position on the one or one or more substrates. A plurality of positional correlations may be provided relative to the one or more substrates at the same time. The position may have a known relationship relative to a location at which welding is conducted, for instance relative to the start of the weld and/or relative to the end of the weld.

The positional correlation may be relative to a position relating to the welding apparatus, for instance relative to a position on the welding apparatus. A plurality of positional correlations may be provided relative to the welding apparatus at the same time. One or more of the positions on the welding apparatus may be proximate to the welding electrode. The position may have a known relationship relative to a location at which welding is conducted, for instance relative to the start of the weld and/or relative to the end of the weld, potentially with the known relationship being the separation of electrode and substrate.

Both a positional correlation relative to a position relating to the one or more substrates being welded, for instance relative to a position on the one or one or more substrates and a positional correlation relative to a position relating to the welding apparatus, for instance relative to a position on the welding apparatus, may be provided.

The position stamp may be applied by a processor. The processor may receive one or more positional signals. The positional signals may be emitted by one or more position sensors and/or position indicators. The position sensor may be a vision-based sensor, for example a camera. The position indicator may be a movement detection-based position indicator, for example an incremental encoder, for instance providing occurrence of movement and/or direction of movement and/or encoder position. The tracking signal-based position indicator may be a radio frequency identification-based system, potentially with one or more radio frequency identification tags on the substate and/or welding apparatus.

The method may include providing at least one type of correlation between the data obtained from at least three, potentially at least four and possibly at least five elected sensor types in the plurality of sensor types. The same type of correlation may be provided between at least two of the elected sensor types. The same type of correlation may be provided between all of the elected sensor types.

The method may include providing at least two different types of correlation between the data obtained from at least three, potentially at least four and possibly at least five elected sensor types in the plurality of sensor types. The same different types of correlation may be provided between at least two of the elected sensor types. The same different types of correlation may be provided between all of the elected sensor types. The different types may include a temporal correlation and a positional correlation.

The method may include synchronising one or more data points in the data obtained from each of the elected sensor types. The method may include synchronising one or more data points in the data obtained from each of at least two, potentially at least three, possibly at least four, of the elected sensor types using the same correlation or correlations.

The method may include synchronising at least 10% of the data points in the data obtained from at least one of the elected sensor types with at least 10% of the data points in the data obtained from at least another of the elected sensor types.

The method may provide that the method includes at least one of the plurality of sensor types, potentially at least one of the elected sensor types, are selected from: voltage sensors, current sensors, welding arc sound emission sensors, weld topology sensors, weld imaging sensors and ultrasound imaging sensors.

A voltage sensor may be one of the plurality of sensor types. A voltage sensor may be one of the elected sensor types.

A current sensor may be one of the plurality of sensor types. A current sensor may be one of the elected sensor types.

A welding arc sound emission sensor may be one of the plurality of sensor types. A welding arc sound emission sensor may be one of the elected sensor types. The welding arc sound emissions sensor may be an acoustic sensor, such as a microphone. The welding arc sound emissions sensor may be sensitive to emissions coming from the arc forming the weld and/or from the interaction of the arc with the substrate and/or from the interaction of the arc with the shield gas. The welding arc sound emissions sensor may provide data which informs on characteristics relating to welding speed and/or weld location sidewall arcing and/or weld location sidewall fusion and/or shielding gas flow rate.

A weld topology sensor may be one of the plurality of sensor types. A weld topology sensor may be one of the elected sensor types. The weld topology sensor may be a radiation-based sensor. The weld topology sensor may be an image-based sensor. The weld topology sensor may be a laser sensor. The weld topology sensor may be sensitive to radiation applied to and reflected by the substrate and/or weld. The applied radiation and/or reflected radiation may be focused by the weld topology sensor. The weld topology sensor may inform on characteristics relating to the geometry of the substrate and/or weld.

A weld imaging sensor may be one of the plurality of sensor types. A weld imaging sensor may be one of the elected sensor types. The weld imaging sensor may be a camera-based sensor. The weld imaging sensor may provide data that informs on characteristics relating to weld pool size and/or weld pool geometry and/or weld pool temperature and/or the presence of deposited material on the weld and/or substrate and/or the patterns of deposited material on the weld and/or substrate and/or shape of deposited material on the weld and/or substrate and/or presence of anomalies on the weld and/or substrate.

An ultrasound imaging sensor may be one of the plurality of sensor types. An ultrasound imaging sensor may be one of the elected sensor types.

One or more or all of the sensor types may provide raw data sets. The method may include processing of the raw data sets to give processed data sets A determination for an individual sensor type may consider the raw data and/or processed data. A determination for an individual sensor type may consider the distance that a data point is from a known set of data points, for instance a known set of data points indicative of acceptable welding. The known set of data points may be represented by a single element, for instance the mean of the known set of data points.

The known set of data points may be represented by a Principal Component Analysis, PCA, model for the sensor type and its signal values and/or data values.

The distance may be a Mahalanbis distance.

The distance may be expressed as an outlier score.

The determination may be based upon the distance of the data point relative to a threshold value, for instance a threshold distance. The determination may be based upon whether the data point is within or outside a boundary for the known set of data points.

The determination may be made using a Mahalanbis distance novelty detection model by comparing the data points or signal values to a previously developed Principal Component Analysis, PCA, model for the sensor type and its signal values and/or data values.

The processing of the raw data sets may include applying a noise cancelling algorithm. The raw data sets may be processed to divide them into a series of bandwidths.

The determination may use a model, for instance a PCA model of acceptable performance, to make the determination on subsequent data points and/or signals.

The model and/or the mean of the model may be recalculated as data points and/or data sets determined to be acceptable welding and/or determined to be unacceptable welding are added.

One or more or all of the sensor types may provide raw data sets, with the method including processing of the raw data sets to give processed data sets, the processing making use of a model. The method may include the use of different variants of the model for different parts of the welding process. In particular, a variant of the model may be used for one or more welding passes and a different model may be used for one or more further welding passes within the overall weld generating process. A different variant of the model may be used for each weld pass in the overall weld generating process. One or more further different variants may be used for remedial welding within the weld generating process, for instance where a part of the weld is removed and then rewelded. A separate further different variant of the model may be used for each weld pass or weld bead in the overall welding process that is being subjected to remedial welding.

The model may be an acoustic signal processing model. The model may be a voltage and/or current and/or power, such as output power, processing model.

The model may be comprised of multiple different data sources or sub-models.

The method may provide that the synchronising provides an overall data structure in which data points from the at least two elected sensor types are aligned in time and/or position with one another.

The method may provide that the data points from the at least two elected sensor types are displayed to a user, with the data points align with respect to time of occurrence and/or position of occurrence.

The method may provide that the data from two or more of the elected sensor types and the correlated data from the at least two sensor types are displayed to a user and/or stored.

The method may provide that the determination as to whether the welding is of acceptable quality or unacceptable quality and/or wherein the one or more actions taken are displayed and/or stored.

The method may provide that the method includes providing a correlation between the data obtained from at least four elected sensor types in the plurality of sensor types.

The method may include inspection of the whole weld. The method may include the inspection of a multi-pass weld, for instance before all of the passes have been completed, such as after each pass and before the next pass.

The welding apparatus may be mounted on any autonomous automated deployment system. The welding apparatus may be mounted on a guided rail system or column and boom system. The welding apparatus may be mounted on a robotic arm, for instance a multi-axis arm. The welding apparatus may be an arc welder.

The weld inspection apparatus may be one of the plurality of sensor types. The weld inspection apparatus may be one of the elected sensor types. The weld inspection apparatus may use ultrasound, for instance phased array ultrasound.

The weld inspection apparatus may be mounted on any autonomous automated deployment system. The welding apparatus may be mounted on a guided rail system or column and boom system. The welding apparatus may be mounted on a robotic arm, for instance a multi-axis arm. The weld inspection apparatus may be a phased array ultrasound transducer. The weld inspection apparatus may include an ultrasound emitter and a receiver. The weld inspection apparatus may include a substrate contacting surface. The weld inspection apparatus may be in physical contact with the inspection location. The substrate contacting surface may be in physical contact with the inspection location.

The method may include rolling the weld inspection apparatus, for instance the substrate contact surface thereof, over the surface of the substrate, for instance to one side of, but potentially parallel to the weld.

The weld inspection apparatus may be provided with internal cooling. The method may include the feeding of coolant into the weld inspection apparatus and/or the removal of coolant from the weld inspection apparatus.

The method may further provide that at least one sensor type of the plurality of sensor types, potentially at least one of the elected sensor types, is a part of weld inspection apparatus and the method includes a step of inspecting the weld using the weld inspection apparatus. The method may provide that the method includes a step of inspecting the weld using weld inspection apparatus to determine the absence or presence of a welding defect at a welding location, particularly a continuous series of locations forming the weld. The method may provide that the method includes a step of inspecting the weld using weld inspection apparatus to determine one or more characteristics of a defect at a location. The method may provide that the characteristics include one or more of size, position, defect type, defect shape or defect position relative to the geometry of the weld and/or relative to the length of weld.

The method may include a determination of the absence or presence of a welding defect at a welding location in real-time. The method may include a determination the absence or presence of a welding defect at a welding location within 100 milliseconds, potentially within 50 millisecond and possibly within 20 milliseconds, of the welding apparatus moving away from the location. The method may provide that the method further includes a comparison of one or more of the characteristics with one or more standards, and further includes a determination of whether the weld with the defect at a location meets a weld standard or does not meet the weld standard.

The method may include a determination of whether the weld with the defect at a location meets a weld standard or does not meet the weld standard in real-time. The method may include a determination of whether the weld with the defect at a location meets a weld standard or does not meet the weld standard within 100 milliseconds, potentially within 50 millisecond and possibly within 20 milliseconds, of the welding apparatus moving away from the location.

The method may provide that if the weld meets the weld standard, then a record for the weld is created and stored, potentially with the position of the defect relative to the geometry of the weld and/or relative to the length of weld included. The method may provide that the method further includes that the record includes data from one or more of the plurality of sensor types. The method may provide that if the weld does not meet the weld standard, one or more remedial steps are applied to the weld.

The method may provide that at least one sensor types of the plurality of sensor types, potentially at least one of the elected sensor types, are weld conditions sensors and the method includes a step of inspecting the weld conditions using the weld condition sensors.

The method may provide that the method includes a step of inspecting the weld conditions to determine one or more parameters of the weld as it is formed.

The method may provide that the method includes a comparison of one or more of the parameter with one or more control parameters, and further includes a determination of whether a risk level for weld defects is exceeded.

The method may provide that the method includes the one or more actions being to alter the welding conditions from the first set of welding conditions for the welding method. The method may provide that the method includes the alteration in the welding conditions from the first set of welding conditions is to stop welding and/or alert an operator. The method may provide that the method includes the alteration in the welding conditions from the first set of welding conditions is to change the welding conditions back to the first set of conditions and/or to change the welding conditions to a second set of conditions.

The method of welding may further include: inspection of a weld formed by the method of welding. The method of welding may further comprise: i. providing weld inspection apparatus; ii. elevating the temperature of the one or more substrates above ambient temperature using heating; iii. conducting welding of the one or more substrates at an elevated temperature above ambient temperature using the welding apparatus; iv. inspecting the weld generated using the weld inspection apparatus; v. wherein the inspecting of the weld is provided by the inspection apparatus at an inspection location on the one or more substrates, with the inspection location at an elevated temperature above ambient.

The first aspect of the disclosure may include any of the features or possibilities or options set out elsewhere in the document, including in the other aspects of the disclosure.

According to a second aspect of the disclosure there is provided apparatus for monitoring welding, the apparatus comprising: a) a plurality of inputs for data from a plurality of sensor types; b) one or more processors, wherein a processor from amongst the one or more processors: a. receives the inputs; b. processes the data obtained from at least two elected sensor types amongst the plurality of senor types to apply a correlation between the data obtained from at least two elected sensor types amongst the plurality of senor types; c. processes the data obtained from at least two elected sensor types amongst the plurality of senor types using the correlation to synchronise one or more data points in the data obtained from one of the at least two elected sensor types with one or more data points in the data obtained from another of the at least two elected sensor types; c) one or more outputs for processed data.

A processor may further provide a comparator for receiving and comparing data from one or more of the plurality of sensor types, such as one or more of the elected sensor types, and reference data for one or more of the sensor types.

The comparator may output a determination on whether the welding is of acceptable quality or unacceptable quality based upon the compared data. If the determination is that the welding is of unacceptable quality, the apparatus may further provide a control signal to trigger one or more actions by a control unit.

The apparatus may further comprise a control unit, for instance for receiving a first set of welding conditions for the welding method and/or for receiving a control signal to trigger one or more actions by the control unit, such as a revision to the first set of welding conditions or the termination of welding.

The apparatus may provide that the apparatus includes weld inspection apparatus to determine one or more characteristics of a defect.

The apparatus may provide inputs from at least one of the plurality of sensor types, potentially at least one of the elected sensor types, wherein the sensor types and/or elected sensor types are selected from: voltage sensors, current sensors, welding arc sound emission sensors, weld topology sensors, weld imaging sensors and ultrasound imaging sensors.

The apparatus may provide that the apparatus further provides the comparator includes a first comparator for receiving and comparing one or more of the characteristics with one or more standards, and wherein the first comparator outputs a first determination on whether the weld with the defect meets a weld standard or does not meet the weld standard. The apparatus may provide that the apparatus further provides the comparator includes a second comparator for receiving and comparing one or more of the parameters of the weld as it is formed with one or more control parameters, and wherein the second comparator outputs a second determination on whether a risk level for weld defects is exceeded. The apparatus may provide that if a risk level for weld defects is exceeded that a control signal provided by the apparatus is sent to a controller to trigger one or more actions, and wherein the one or more actions is to alter the welding conditions from the first set of welding conditions for the welding method, for instance, to stop welding and/or alert an operator.

The second aspect of the disclosure may include any of the features or possibilities or options set out elsewhere in the document, including in the other aspects of the disclosure.

According to a third aspect of the disclosure there is provided a method of welding, the method of welding comprising: a) providing welding apparatus; b) providing a plurality of sensor types; c) defining a first set of welding conditions for the welding method; d) introducing one of more substrates to be welded to the welding apparatus; e) conducting welding of the one or more substrates; f) obtaining data from the plurality of sensor types during welding; g) comparing the obtained data from the plurality of sensor types with reference data for one or more of the sensor types; h) based upon one or more such comparisons, determining whether the welding is of acceptable quality or unacceptable quality; i) wherein, the obtained data is obtained by processing the data from the plurality of sensor types, the obtained data for at least one of the plurality of sensor types being processed using a computer model and wherein the method includes the use of different variants of the model for different parts of the welding process.

In particular, a variant of the model may be used for one or more welding passes and a different model may be used for one or more further welding passes within the overall weld generating process. A different variant of the model may be used for each weld pass in the overall weld generating process. One or more further different variants may be used for remedial welding within the weld generating process, for instance where a part of the weld is removed and then rewelded. A separate further different variant of the model may be used for each weld pass or weld bead in the overall welding process that is being subjected to remedial welding.

The model may be comprised of multiple different data sources or sub-models.

The third aspect of the disclosure may include any of the features or possibilities or options set out elsewhere in the document, including in the other aspects of the disclosure.

According to a fourth aspect of the disclosure there is provided apparatus for monitoring welding, the apparatus comprising: a) a plurality of inputs for data from a plurality of sensor types; b) one or more processors, wherein a processor from amongst the one or more processors: a. receives the inputs from at least one of the plurality of sensor types; b. processes the data to provide obtained data for at least one of the plurality of sensor types, the processor having access to a computer model and wherein the method includes the use of different variants of the model for different parts of the welding process; c) one or more outputs for processed data.

The fourth aspect of the disclosure may include any of the features or possibilities or options set out elsewhere in the document, including in the other aspects of the disclosure.

According to a fifth aspect of the disclosure there is provided a method of welding, the method of welding comprising: a) providing welding apparatus; b) providing a plurality of sensor types; c) introducing one of more substrates to be welded to the welding apparatus; d) conducting welding of the one or more substrates; e) obtaining data from the plurality of sensor types during welding; f) comparing the obtained data from the plurality of sensor types with reference data for one or more of the sensor types; g) determining whether the welding is of acceptable quality or unacceptable quality; h) wherein, the determining includes the consideration of combined data from at least two of the plurality of sensor types.

The method may provide that the combined data is formed and then the determination is made by considering the combined data.

The method may provide for a first determining whether the welding is of acceptable quality or unacceptable quality from the obtained data from one or more of the plurality of sensor types compared with reference data from one or more of the sensor types.

The method may provide that the determining whether the welding is of acceptable quality or unacceptable quality using the combined data is a separate second determining to the first determining. The outcome of the second determining may be given precedence over the first determining in establishing whether the welding is of acceptable quality or unacceptable quality.

The combined data-based determination may include a neural network based step. The combined data-based determination may include a machine learning supported decision engine The combined data-based determination may include the use of a model.

The combined data may be used in a separate determination of acceptable and/or unacceptable welding to one or more such determinations made using data from only one senor type in each determination.

The combined data-based determination may provide a determination as to acceptable and/or unacceptable welding when a determination made using data from only one senor type in each determination cannot make a determination and/or cannot make a determination with an acceptable level of associated error and/or makes a different determination on acceptable and/or unacceptable welding.

The combined data-based determination may be used for data points and/or series of data which are close to a threshold, close to a threshold distance and/or close to a distance relative to an indication of acceptable and/or unacceptable welding. The combined data-based determination may not be used where the data points and/or a series of data points are far from the threshold, threshold distance or distance.

The combined data may be from one or more sensor types relating to welding conditions applied during the welding. The combined data may be from one or more sensor types relating to welding observations, conducted within less than 1 second of welding, for instance less than 1/10th of a second of welding, at a location. The combined data may not include data from one or more sensor types relating to NDT of the weld and/or relating to welding observations conducted more than 1 minute after the welding at a location.

The neural network may be trained by the provision of labelled data to the neural network. The labelled data may be data points and/or series of data points and an indication of whether they are consistent with acceptable welding and/or unacceptable welding. The labelled data may be data points and/or series of data points and an indication of whether they are consistent with acceptable welding and/or unacceptable welding with respect to a particular characteristic which can give acceptable or unacceptable welding.

The neural network may be trained by the provision of supervised learning. The neutral network may be provided with supervised learning through operator based inputs, such as labels, for instance whether the data point and/or series of data points are indicative of acceptable and/or unacceptable welding.

The training using labelled data and/or supervised learning may be provided during a calibration stage and/or during production welding.

The neural network may be trained from a library of existing data, particularly labelled data. The library may be fed to a neural network classifier.

The labelled data and/or supervised learning and/or library may provide data from multiple sensor types.

The second manner, the library, may be used as the starting labelled data set. The conduct of the first manner, calibration or test runs on the actual welding system with operator calls, can be used as an alternative from the outset. The first manner can be used to add to the second manner data set.

The neural network may revise a definition used in the determination of acceptable and/or unacceptable welding with the passage of time and/or learning.

The neural network may be trained by unsupervised learning. The unsupervised learning may be carried out during a calibration stage.

The neural network may provide clustering or grouping based processing. The neural network may look for patterns in the data, for instance of similarities and/or of anomalies in the data.

The neural network may be trained in a first manner only or may be trained in a first manner and then in a second manner or may be trained in a first manner and a second manner in parallel.

The fifth aspect of the disclosure may include any of the features or possibilities or options set out elsewhere in the document, including in the other aspects of the disclosure.

According to a sixth aspect of the disclosure there is provided apparatus for monitoring welding, the apparatus comprising: a) a plurality of inputs for data from a plurality of sensor types; b) one or more processors, wherein a processor from amongst the one or more processors: a. receives the inputs from at least one of the plurality of sensor types; b. compares the obtained data from the plurality of sensor types with reference data for one or more of the sensor types; c. determines whether the welding is of acceptable quality or unacceptable quality for an output determination; the apparatus further comprising: c) a processor from amongst the one or more processors: a. forms combined data from the data from two or more of the plurality of sensor types to provide combined data; b. determines a second determination using the combined data, the second determination being whether the welding is of acceptable quality or unacceptable quality; d) one or more outputs for the output determination and the output of the second determination.

The sixth aspect of the disclosure may include any of the features or possibilities or options set out elsewhere in the document, including in the other aspects of the disclosure.

Various embodiments of the disclosure will now be described, by way of example only, and with reference to the accompanying drawings, in which: Figure 1 is a schematic illustration of an adaptive control capability for welding processes conducted according to the disclosure; Figure 2a is an image of an ultrasound probe detected defect; Figure 2b is an image of the ultrasound probe used in the detection of Figure 2a; Figure 3 is a plot of outlier score obtained from acoustic signals v data point for a variety of different welding conditions; Figure 4a is a perspective view of a profile sensing device relative to a substrate and weld; Figure 4b is a schematic illustration of the sequence and geometry of the formation of an overall weld through a series of weld passes; Figure 5 is a series of camera images of a welding location during welding; Figure 6 is a plot of arc voltage and a plot of Gaussian amplitude x Gaussian centre v time for a welding process; Figure 7 is an illustration of the combined data types displayed to a user; Figure 8 is an illustration of a second level of processing applied to data from a plurality of sensor types.

In is known to conduct non-destructive testing (or examination), NDT, of welds after they have formed and cooled to see if any defects are present. Remedial action can then be taken on any defects found. Such NDT is however, conducted in the prior art a long time after the weld is formed and so the information on defects is provided a long time after the event and so is too late to prevent the defect formation or to mitigate further defect formation.

During welding, a significant number of variables exist which impact upon the formation of the weld and the quality of that weld. These may include: wire feed speed, voltage, current, welding speed, welding device to welding location separation, shield gas, shield gas flow rate, substrate shape/profile/configuration/dimensions, pre-heating temperature for the substrate, welding groove shape/profile/configuration/dimensions, weld or weld pass shape/profile/configuration/dimensions.

Attempts have been made to monitor individual variables during welding to detect deviations in those and potential impaired quality welding and/or defect generation as a result.

As a weld is formed and immediately after formation, the weld has various characteristics. These may include: acoustic emissions from the welding location and/or arc, the topology of the weld formed, the visual characteristics of the weld as it is formed and/or after formation.

Attempts have been made to determine individual characteristics during welding to detect deviations in those and potentially impaired quality welding and/or defect generation as a result.

Overall Process The present disclosure seeks to obtain increased benefits in welding by making greater use of variable monitoring and/or characteristic determining to mitigate against welding defects forming or other quality issues arising. The disclosure also seeks to synchronise and fuse data from multiple sources so as to obtain more information than the sum of the parts. In addition, the disclosure seeks to analyse each of the data types in a more informative manner.

The disclosure also reduces the risks of defect formation through improved initial weld quality. This is achieved by an enhanced range of data types relating to the welding that are collected, analysed and employed as the welding progresses.

The disclosure makes use of data derived from multiple sensors and/or sensor types processed in real time with machine learning / artificial intelligence algorithms being used to derive near real time decision making for weld process control. Thus. the decisions benefit from the involvement of the multiple different sensor types.

In the exemplified disclosure, four different sensor types are used to collect data from use in combination and to provide adaptive control capability for the welding and hence mitigate the chance of defect formation. However, the methodology is suitable for handling greater numbers of sensor types or data types and is also suitable for use with different sensor types to those exemplified.

Referring to Figure 1, a schematic illustration of the adaptive control capability is provided.

The Welding Process is where a series of variables which influence and control the welding are defined and controlled. These may include: wire feed speed, voltage, current, welding speed, welding device to welding location separation, shield gas, shield gas flow rate, substrate shape/profile/configuration/dimensions, pre-heating temperature for the substrate, welding groove shape/profile/configuration/dimensions, weld or weld pass shape/profile/configuration/dimensions.

Within the Welding Process, on the left-hand side of Figure 1, an NDT step [Post Weld Hot NDT] is conducted using a high temperature compatible roller type ultrasound probe [described in further detail below] and this step provides the ultrasound data [NDT Indications] which are fed to the overall assessment of Weld Qualification which provides the end of the process.

As well as that monitoring of the weld produced and prompting any remedial action on that, the Welding Process leads to a parallel series of steps commencing with Process Monitoring that seek to prevent or minimize the extent of defects arising.

In the Process Monitoring step, data on the applied voltage and current with time for the welding device is measured and collected. Sensing of voltage and current is used in this embodiment to consider correct weld formation, or issues in correct weld formation.

The data from three other sensor types are feed into the data stream at this point to form the Sensor Data. These are Laser, Vision and Acoustic data streams from appropriate sensors of that type described in more detail below. Sensing of weld profile using laser scanning, visual assessment of the weld and acoustic sensing of the sounds originating from the weld are also used in this embodiment to consider correct weld formation, or issues in correct weld formation.

The process sequence includes the collection of the Sensor Data, through the use of Laser. Vision and Acoustic data streams. In addition, the Process Monitoring step collects data on the applied voltage and the current for the welding device. The Post Hot Weld NDT also provides the NDT Indications; a fifth data type. These form the full data stream that is considered. Each of those sensor types and the data type they provide are now considered. The different sensor types are: * Phased Array Ultrasound sensor type -measuring weld quality; * Acoustic sensor type -deriving weld process performance from the sound of the welding arc; * Laser sensor type -monitoring the topology of the weld; * Visual monitoring type -replacing the operator eyesight in visualisation of the weld pool; * Voltage and current monitoring type -measuring the power to maintain the welding process in its optimum state.

Ultrasound sensor type This sensor type is concerned with the results from the weld produced, whereas the other sensor types are more focused on the weld as it is produced. Thus, this sensor type provides information on the need for remedial action, whereas the later sensor types described below can be used to mitigate the generation of defects in the first place.

The ultrasound sensor type includes a transducer which provides a 5Mhz 64 element phased array that is mounted to generate 550 ultrasound waves into the substrate. A 0.5mm pitch and 10mm elevation can be used. An angled beam is beneficial in being able to inspect the weld fully from a laterally spaced location. Frequently, that laterally spaced location will be more amenable to good contact between the probe and the object, that at the location where welding is occurring. For instance, in multi-pass welds, until the weld is completed, there will be a significant depression which will interfere with good contact and ultrasound propagation into the object. This is an issue with 00 or low angle-based approaches.

This type of transducer and an associated ultrasound conveying block configuration can be used to provide a sectorial scan beam defined by the upper extremity beam [angled away from the perpendicular to the transducer face] and the lower extremity beam [near perpendicular to the transducer face] emitted.

In terms of the performance sought for the probe on high temperature performance, the probes are capable of inspecting for prolonged period substrates that are at P:i300°C.

The coupling is dry but still achieves the necessary levels of ultrasound propagation through the interface into and back from the abject.

The high-temperature polymer used in the coupling component is able to withstand prolonged contact with objects at such temperatures and still propagate the ultrasound to and from the interface successfully.

The coolant and hence the coolant filled gaps between elements of the sensor type are able also to effectively propagate the ultrasound waves to and from the conveying block.

Optimal propagation properties for the conveying block are provided as the block is exposed to near ambient temperatures only and so there is no need to choose high temperature resistant materials which have lesser ultrasound propagation properties.

The sensor type also includes an integrated surface temperature measurement and a coolant temperature measurement sensor type.

Welding location conditions and impact Arc welding uses a power supply to generate sufficient voltage difference between the electrode of the welding device and the substrate to be welded to produce an arc. A current results. The arc heats the substrate to a molten state [potentially with consumption of the electrode]. On cooling, the molten metal solidifies to join two substrates together.

The speed with which the welding device moves relative to the substrate has an impact upon the quality of the weld through impacting upon the extent of melting, shape of the melt pool and the like.

The welding location is usually protected by a shield gas to prevent oxygen, water or water vapour in the atmosphere reaching the welding location. Inert or semi-inert gases are typically used for the shield gases, examples include argon and helium. Flow rates for the shielding gas impact upon its ability to perform its role.

For the quality of the weld to be high, careful control of many operating variables is necessary within the welding process. These variables can be considered indirectly, as exemplified in the section below. Although not directly sensed in the exemplifying embodiment, further sensor types could measure speed of welding device movement, shield gas flow, shield gas flow rate and add those sensor types and their data sets to the processing.

Acoustic sensor type The acoustic sensor type collects high frequency audio signals that arise as the welding occurs. These come from the arc forming the weld and the interaction of the arc with the substrate and the interaction of the arc with the shield gas. The audio signals detected have been established to be sensitive to a number of important variables within the welding process. Referring to Figure 3, this shows examples for the values for outlier score for different welding characteristics.

The outlier score is obtained by a mathematical approach which considers how far a particular data value is from known set of data values, which have been classified as acceptable values for the characteristic and/or sensor type data being assessed.

One such approach, used in the disclosure, is the use of a Mahalanbis distance novelty detection model by comparing the incoming audio signal values to the previously developed Principal Component Analysis, PCA, model for the sensor type and its signal values and/or data values.

The signal processing includes taking the audio signal and applying a noise cancelling algorithm to the raw data. The audio signal is further processed by the use of a short-time Fourier transform to convert the raw time series of data into the frequency domain. Statistical features are then extracted from each of a series of bandwidths spanning the frequency range of interest. In the example, the bandwidth used is 39.1kHz and that yields 312 features describing the acoustic signal in a given instance of the signal. The total feature set [312 features for each signal instance] is then optimised by the removal of redundant features and then standardised to improve the robustness of the features with smaller standard deviations.

In the initial establishment of the PCA model, the signals and hence the remaining feature set arising from the processing above is established to be for acceptable performance of the welding with respect to that variable, in this case acoustic signals. Hence, the remaining feature set can be finally reduced using a PCA to give the model and the principal components that define that model in respect of acceptable performance of the welding.

The PCA model of acceptable performance can then be used to consider subsequent signals. Those subsequent signals are subjected to the same noise cancelling and other steps defined above. For each feature, a feature value position relative to the distribution of feature values for established acceptable weld performance and hence feature values is determined. A mean is calculated for the distribution and the feature value distance is measured relative to the mean. The distance provides the outlier quantification displayed in Figure 3 and so establishes variations in distance which are still consistent with acceptable welding conditions and establishes variations in distance which are more exceptional and so indicative of impaired welding performance.

The approach is beneficial in giving a unitless, scale-invariant quantification that takes account of the correlation of the feature values within the distribution. The mean can be recalculated as acceptable feature values are added to the distribution and/pr the PCA model or can be based upon a fixed set of pre-existing acceptable feature values used in a PCA model.

Referring to the example results obtained using this processing and displayed in Figure 3, the first set of data points are illustrative of good welding operating parameters occurring. These parameters where independently verified as applying. As can be seen these give a well clustered set of data points A, relative to the log scale outlier score axis.

The second set of data points are illustrative of too high a welding speed; that is the welding device and the substrate are moving too fast relative to one another. Two different welding speed deviations were demonstrated and these give two well clustered sets of data points B1 and E32, with few outliers.

The third set of data points are illustrative of sidewall arcing during welding; that is the arc is short circuiting to the side wall rather than to the desired welding location within the welding groove Once again, these conditions give a well clustered set of data points C, with few outliers.

The fourth set of data points are illustrative of sidewall fusion; that is the arc is melting the sidewall and causing fusion there rather than in the welding groove. The data points are well clustered as data points D, with few outliers.

The fifth set of data points are illustrative of too high a flow rate for the shielding gas; this can cause porosity issues which are undesirable. As with the other sets of data points, this set E is also well defined with few outliers. A similar position can be detected with too low a flow rate for the shielding gas.

Each of the undesirable welding conditions mentioned, when present, gives a far higher outlier value or feature value than when good welding conditions are present. As a result, a threshold Thac"stIc [a selected value for the Mahalabis distance] can be set and can be used to distinguish between acoustic sensor type derived data indicating good welding conditions or indicating impaired welding conditions. Thus, the acoustic data type has the ability to warn the operate, trigger a stop to welding or the like if the threshold Th""stl is breached or remains breached for a given number of data points. Real-time acoustic signal-based identification of defect generation is thus provided. This analysis can be provided continuously and can be provided for each pass of a multi-pass weld approach.

Significantly, the algorithm used is more complex that just the use of a single previously developed Principal Component Analysis, PCA, model for the sensor type and its signal values and/or data values. As seen in Figure 4b, in multi-pass welding, the weld passes gradually fill the weld groove. This means that the depth of the weld groove, the shape of the weld groove and the volume filled change with each pass. All of these and potentially other changes between passes impact upon the acoustic signals that are emitted and are detected. Thus, the approach uses a separate model for each pass so as to make the most accurate assessment of the observed data against the expected data for that pass.

The separate models could be obtained by a neural network approach to training based upon pre-existing passes or could be learned as the passes are made during operation with increasing number of each pass improving the model for that given pass in the sequence of passes.

The use of a separate model for each pass also extends to the use of separate models for repasses, for instance as a part of remedial work on the weld. The weld groove shape and hence the acoustics in such repass cases are significantly different from the ordinary passes.

Loser sensor type The next sensor type employed is a visual one and this seeks to evaluate the geometry of the weld created.

Figure 4a illustrates a section of substrate 2 with a weld 20 formed on it. The weld 20 is linear in this example, but other weld track can be considered in the same way. A visual sensor type device 22 includes a casing 24 within which is provided a light source 26 which can illuminate the substrate 2 and the weld 20 across an illuminating width 28. The device 22 has an operating range 30 within which accurate imaging is possible. Light returns to the device 22 where a receiver 32 focuses the light onto a senor matrix 34 and signals are generated. In this example a 2D laser profile scanner is used, by other types can be substituted.

The laser as the light source 26 is used to establish various details of the weld and surrounds, including the weld profile, groove profile remaining, weld bead width and any material deposited outside of the weld groove. A plane perpendicular to the substrate surface beside the weld groove and perpendicular to the longitudinal axis of the weld groove may be used for the inspection. In addition, the weld bead profile along the weld groove may be considered.

Figure 4b is an illustration of a typical welding groove during the course of a sequence of welding passes. Each weld pass adds a weld to the welds already present in the groove, in a predetermined sequence [as numbered] to build up the overall weld. As can be seen, the weld passes contribute a predictable geometry for the weld pass itself and a predictable change in the geometry of the weld groove, if welding is proceeding correctly.

The signals from the device 22 can be used to form a profile image across the weld track with each position along the weld track. That actual profile can be compared with an expected profile and deviation noted. The deviation can be compared with a threshold ThPrane. Thus, the profile data has the ability to warn the operate, trigger a stop to welding or the like if the threshold ThPmfile is breached or remains breached for a given number of data points. Real-time profile signal-based identification of defect generation is thus provided. This analysis can be provided continuously and can be provided for each pass of a multi-pass weld approach. Effective geometric verification is provided.

Visual sensor type In the next sensing area, a high dynamic range camera is used to obtain images of the welding location, including the location where a weld has not yet been formed, the location where the weld is being formed and the location where the weld is solidifying and then cooling further.

Figure 5 shows a series of images of this type collected from different weld locations. Processing of these using combinations of artificial intelligence and conventional machine vision tools, the system can detect the changes in the images that correlate to abnormal welding conditions or the generated visual anomalies caused by defects. This may be achieved using the processing of single images, or a combination of multiple images from prior data, both from recent images, and historical parts and passes.

One or more variables may be considered in this area, for instance, the weld pool size [width; trailing length, leading length], weld pool geomehy [elliptical, teardrop or others]; weld pool temperature can be considered, along with analysis on the patterns, shapes of deposited material and visible anomalies. A profile of known judgement parameters for different types of defects may be used to compare the live outputs from each pent of the algorithm against known-good values. These values may be a combination of, presence or absence of a specific visual feature, a numerical band or a threshold value, or a classification. Should the image meet be deemed to exceed these parameters, the situation occurring for one or more or all of those can he compared with the desired position for one or more or all of those and a deviation can be used again to trigger a warning or stopping of the welding.

Voltage and current sensor type The voltage applied will impact upon the formation of the arc and the current within the arc. That in turn impacts the power and hence the rate of melting of the substrate [and if consumed the electrode]. These are important variables for the quality of the weld. These are variables which impact the weld pool size for instance.

The welding voltage also needs to be automatically and continuously adjusted to reflect the separation between the welding device and the substrate being welded. This is based upon the known and fixed position of the substrate and the variable but known X-Y-Z position of the robotic arm carrying the welding device, and hence the separation of the two.

The sensing system for the voltage and current monitoring is orders of magnitude faster than that incorporated in pre-existing automatic voltage control approaches. The power system is able to handle 500A and scale up to 1000A or more, whilst still providing voltage monitoring and current monitoring at the nano-second level between measurements. Hence, very detailed information of the voltage and current are obtained and even short-lived variations can be taken into account.

Input voltage and current, and hence input power, tends to be detected rather than output power to avoid the sensing/detecting itself disrupting the output power performance.

The data and hence the approach used for its processing is similar to that set out above for the acoustic sensor type.

Referring to Figure 6, the arc voltage is plotted against time as plot V. In this example, the welding process was shut down after 30 seconds and so the voltage returns to zero.

Also plotted in Figure 6 is the value for Gaussian amplitude x Gaussian centre versus time. As can be seen, in the initial time period, the first 2 or 3 seconds, the value for this plot is higher than an acceptance threshold and so there would be concerns at the quality of the weld. After that initial 2-3 seconds, and certainly after 8 seconds onwards, the plot is appreciably below the threshold and quality welding is being indicated with respect to this variable set.

In addition to these variables, Figure 6 also includes indications of when low argon as the shield gas is present. The indication shows the occurrence of low argon by being a plot point and also shows the extent to which the argon is missing through the vertical position of the plot point.

Overall Process -Continued Returning again to Figure 1, having established the specifics by which the sensor types operate and their data is considered, the process and the data within it goes through a Data Conditioning step. The Data Conditioning is provided based upon Inputs provided through a User Interface and/or based upon Historic Data obtained from Storage. The Storage may contain data from earlier in this weld's conduct and/or data from a large number of previous welds conducted by the system and/or data from other systems [for instance from calibration processes or the like], all of which can contribute to the Historic Data and hence to the Data Conditioning.

In the Live Analysis step two different levels of processing may be implemented.

The first level of processing provides for the synchronization of data types from different sensor types. This synchronization can accommodate not only the different data types from the different sensor types exemplified above by means of the different sensor types, but also any number of other sensor types introduced and used to measure key characteristics relating directly and/or indirectly to the welding system and the welds arising.

A Programmable Logic Controller, PLC, is used to temporally synchronize the different data types by the provision of a master timestamp at the initiation of data collection from the sensor types within the system. The PLC also continually checks that each sensor of each sensor type within the system is continuing to provide data and that the data provided is collected at a consistent rate. Periodic further master time stamps may be applied during the data collection process as it progresses.

The master timestamp means that the data collected from the microphones acting as the sensor type for the acoustic analysis, input power analysis to measure the draw during welding, area scan cameras to provide the vision system and laser profile scanner which provides the 3D profile sensor type can all being aligned to represent data from the same time and hence the same position within the weld.

Significantly, the PLC also receives data from an incremental encoder which provides position data. The encoder can be mounted on the substrate being welded and/or on the welding device and indicates the physical position of that encoder at that time. The encoder triggers data collection and provide consistent correlation of sensor data to the weld position on the component. Again, the data from the encoder has the same master timestamp applied. This means that the actual position is known for the established same position applicable at the same time in the alignment of the synchronized data. The data is synchronized temporally and spatially.

This processing enables the sensor data that is collected at different locations on the component to be post processed into one cohesive data structure, with the raw sensor data, the post processed information and the output of the analysis to be correlated to the physical position of the weld. This data can then be displayed to the operator within the user display, allowing them to make adjustment to the welding process based on the output of the system.

The second level of processing is implemented by the application of Data Reduction and/or Machine Learning. This second level of processing takes this correlated data and then feeds it to a secondary layer of processing and analysis (by use of machine learning or other appropriate analysis method) to correlate any number of features from both the raw data and the output of the sensor type level analysis to welding defects. In effect, a machine learning supported decision engine is employed.

Within the method, a general assessment is being made as to whether, where a defect is suspected of a location, that defect is within acceptable characteristics, such as size, or exceeds those acceptable characteristics and is thus a defect that needs noting or remedial action. When NDT type sensing is conducted, then a direct measurement of the defect is made by virtue of the imaging conducted. This reports directly on the size and potentially other characteristics of the defect. However, where the consideration of defects is based upon the welding conditions occurring during the welding, the potential defects are being considered indirectly; the question "are these conditions likely to lead to a defect?" is being considered. The second level of processing offers to make improved determinations on acceptable and/or unacceptable welding in this context.

Referring to Figure 8, two different sensor types are being considered, Type A on the left-hand side and Type B on the righthand side. The two sensor types could be any of the sensor types mentioned herein and/or other forms of sensor type which provide data on the weld's conduct or weld outcome.

Referring to the Type A sensor type, there will be a series of data points within area 800 which are firmly established as indicative of acceptable welding. These may be established from test runs verified by other sensing and/or NDT or may be modelled cases. There will also be data points from welding, such as data point 802, which are clearly outside of that acceptable area and also a long way on the unacceptable side of a threshold 804 that may be applied in making a first determination on whether a data point is indicative of acceptable or unacceptable welding. Other data points, such as data point 806, are above the threshold and also deemed unacceptable with that definition of the threshold 804. The difficult to interpret instances are data points, such as data points 808 and 810, which are outside the area 800 but under the threshold 804. A call based upon the single sensor would result in these being deemed acceptable welding as they are below the threshold. Gains are to be made by the second level of processing which considers the data points [802, 806, 808, 810] across more than one sensor type to reach the full determination. Referring to the righthand side and the Type B sensor type, the data point 802 again lies within the established acceptable welding area 800. Similarly, the data point 806 is well beyond the threshold 804 and once again from this sensor type alone is indicative of unacceptable welding.

Turning to the data points 808 and 810, again, both are outside of area 800, but under threshold 804. Under a single sensor type approach, these edge cases, would again be deemed acceptable welding. However, the second level of processing obtains additional information by considering the position across the multiple sensor types.

Two approaches to the consideration of the position across multiple sensor types can be employed using neural networks. These two approaches can be used as alternatives to one another or one may be use in parallel or used in series with the other.

In the first approach, labelled data is provided and supervised learning is involved. The labelled data can come from either or both of two sources. Firstly, particularly during an early stage of the processing, for instance during calibration or during early production welding runs, the labelled data may be obtained from experimental results. So, continuing with the Figure 8 example, the position for both points 808 and 810 may be flagged to an operator for an operator-based assessment of acceptable or unacceptable. The operator is provided not just with the data from a single sensor type to consider and make a call on, but rather is provided with flagged data from multiple different sensor types on that data [data points or a sequence of data points] and so a more nuanced position from which to make a determination call. The outcome is used to label the data and is thus available within the pool of data that the neural network learns from. The human knowledge and interpretation are fed to the neural network by the user's determination and so the supervision is provided. The second manner for providing the labelled data is to make use of a library of existing data. Again, this data is labelled as to the determination and has the necessary transfer of operator knowledge. The library would be fed to a neural network classifier to establish the processed position for the library data set. The library would cover data from multiple sensor types and thus allow determinations of the second level of processing type; the data point position across multiple sensor type inspection and results. A classifier score could then be determined for any data point relative to that classifier library data and in turn a probable error for a data point [or series thereof] to be examined and a determination to be made on could be quantified. Classifiers can use Bayes theorem-based approaches for the classification and the error quantification.

The second manner, the library, may be used as the starting labelled data set. The conduct of the first manner, calibration or test runs on the actual welding system with operator calls, can be used as an alternative from the outset. However, the first manner can be used to add to the second manner data set and so advance the neural network learning from a more general welding system position to a position more tailored to that particular welding system.

Returning to Figure 8, the operator might determine data point 808 to be sufficiently close to the acceptable area 800 to be deemed acceptable but might deem the data point 810 so close to the threshold across the multiple sensor types that it is not acceptable. That might lead to the neural network providing a minor adjustment to the acceptable area 800 and/or the threshold 804 [value or format]. Over time, repeated determinations of this type could lead to a more pronounced and optimised revision of the bounds of the acceptable area 800 and/or of the threshold 804. For instance, the acceptable area 800 might be expanded and/or the threshold might be tightened. The same results as data points 808 and 810 at a later more advanced learning stage for the process might be called acceptable and unacceptable respectively because of where they sit, within the revised acceptable area for data point 808 and over the revised threshold for data point 810.

An example of such an analysis in a real-world scenario might be that the left-hand side relates to a visual image sensor type and one or more images in succession suggest that there is an issue with side wall proximity for the welding. Consideration of an acoustic sensor type as the right-hand side could also suggest a side wall proximity issue and thereby verify an overall determination that the welding was not acceptable.

Whilst the above examples refer to the position from a sensor type and other sensor types giving a determination of acceptable or not acceptable, the determination may be more detailed than that and give a determination of the nature of the problem with the welding conditions. Thus, the data from two sensor types may inform on the nature of the problem, when the data from one sensor type alone might just suggest a problem.

An issue to note, is that the importation of historical data from other welding situations and environments is not a strong starting place for a library from which to judge acceptable welding performance. This is because other welding operations and other welding environments have a wide range of other variables which could have affected the data from those welding operations. For instance, in the context of the visual imaging type sensor, the lighting and lighting angle in that welding environment, the nature of the substrate, the welding torch angle and separation and the like can all influence the data and so that data would not be consistent with that to be looked for in another environment with different lighting, for example.

As an alternative, or additional to this first approach using labelled data and supervised learning, it is possible to reduce or avoid the need for library type data and/or supervised learning, by making use of the ability of neural networks to conduct clustering or grouping based processing [looking for similarities and/or anomalies in the data]. Various techniques, such as K-means clustering exist which enable centroids for clusters to be established and degrees of certainty with distance established around those. Multiple other clustering approaches apply. These could be sued to establish the acceptable area 800 and/or threshold 804 position and modify it with increased data and learning, without the need for a large library of labelled data to start with.

As mentioned previously, the first approach and second approach can be used in combination with one another, as well as being alternatives. Thus, the first approach could be used to start the neural network off on its learning and then the second approach could take over after the first approach has advanced the learning. It is also possible for the neural network to be learning from both the first approach and the second approach at the same time so as to maximise the data fed, particularly where the library is being extended from ongoing welding in other welding systems besides the particularly welding system under consideration. It would also be possible to pool the learning from similarly configured and operated welding systems.

Overtime, particularly when the method is used in production version conducting large amounts of welding, then the amount of data, the precision of the assessments and the complexity of different cases for the data which can be successfully assessed increases.

Any and all of the outcomes from the Live Analysis of the data and/or Data Reduction and/or Machine Learning, once obtained can have copies fed to the Storage for future use or to contribute to the knowledge of the system and/or similar systems.

A key step in the Analysis Results is the data value or position compared with one or more thresholds set for the data type, Thresholding. Examples of approaches on thresholding are provided in the different sections on the specific sensor types exemplified, but are broadly applicable to each sensor type and the data type it generates. The position relative to a threshold may be deemed indicative of a defect, Defect Indicators, and hence be a key part of the Weld Quality Results which in turn are displayed to the user via User Display, which also receives and displays the Sensor Data received.

Figure 7 is an example of a User Display. The User Display can give the real time output of all of the sensor type analysis allowing them to review the welding process as it occurs and adjust the process if the output of the analysis flags any welds of poor quality. In addition, they will be able to review the overall weld data after each pass and interrogate the data to discern any areas that require rernediation or adjustment before or during the following welding to occur.

The Weld Quality Results step can, on encountering an unacceptable weld, stop the welding process and/or provide an alert to an operator, for instance via User Display.

Based upon the User Display position, the operator or even the system itself may adjust one or more variable parameters used to control and conduct the Welding Process in real time.

The Weld Quality Results feed through to the Weld Qualification provided for the weld overall. Where a defect is identified by the NDT, then the size and position of the defect are cross-checked with appropriate standards to ensure that the defect is acceptable at that level. If not, remedial action is taken. If acceptable, then a 3D lifetime record of the data is kept within the Weld Qualification so as to be available through the life of the weld and into any necessary decommissioning support.

Processing and storage of data In addition to the consideration of each data type individually, there are further advantages to be gained from considering the data types as a combined data stream. To enable this, each of the data sets is time stamped and so that allows all of the data sets to be time synchronised by synchronising the time stamps.

This means that all of the data sets can be combined and then saved as a single data file. This also provides for reuse at a subsequent time.

Claims

CLAIMS1. A method of welding, the method of welding comprising: a) providing welding apparatus; b) providing a plurality of sensor types; c) defining a first set of welding conditions for the welding method; d) introducing one of more substrates to be welded to the welding apparatus; e) conducting welding of the one or more substrates; f) obtaining data from the plurality of sensor types during welding; g) providing a correlation between the data obtained from at least two elected sensor types in the plurality of sensor types; h) synchronising one or more data points in the data obtained from one of the at least two elected sensor types with one or more data points in the data obtained from another of the at least two elected sensor types; i) comparing the obtained data from the plurality of sensor types with reference data for one or more of the sensor types; j) based upon one or more such comparisons, determining whether the welding is of acceptable quality or unacceptable quality; k) where, if the welding is of unacceptable quality, the method includes then taking one or more actions.
2. A method according to claim 1, wherein the correlation is a temporal correlation and a positional correlation.
3. A method according to claim 1, wherein the correlation is a temporal correlation.
4 A method according to any preceding claim, wherein the data from the at least two elected sensor types include a time of occurrence for one or more data points within the data, and wherein the correlation is a matching time of occurrence for data points in the data from the at least two elected sensor types.
A method according to claim 4, wherein the time of occurrence is obtained relative to a timestamp introduced into the data from an elected sensor type for each of the two or more elected sensor types providing a correlation.
6. A method according to claim 5, wherein the time of occurrence is obtained directly from a timestamp at that time.
7. A method according to claim 5 or claim 6, wherein the time of occurrence is obtained by the time elapsed since the timestamp and the time of the data point.
8. A method according to claim 1, wherein the correlation is a positional correlation.
9. A method according to any preceding claim, wherein the data from the at least two elected sensor types include a position of occurrence for one or more data points within the data, and wherein the correlation is a matching position of occurrence for data points in the data from the at least two elected sensor types.
10. A method according to claim 4, wherein the position of occurrence is obtained relative to a position stamp introduced into the data from an elected sensor type for each of the two or more elected sensor types providing a correlation.
11. A method according to claim 5, wherein the position of occurrence is obtained directly from a position stamp at that position.
12. A method according to claim 5 or claim 6, wherein the position of occurrence is obtained by the time elapsed since the position stamp and the time of the data point.
13. A method according to any preceding claim, wherein the synchronising provides an overall data structure in which data points from the at least two elected sensor types are aligned in time and/or position with one another.
14. A method according to any preceding claim, wherein data points from the at least two elected sensor types are displayed to a user, with the data points align with respect to time of occurrence and/or position of occurrence.
15. A method according to any preceding claim, wherein the data from two or more of the elected sensor types and the correlated data from the at least two elected sensor types are displayed to a user and/or stored.
16. A method according to claim 15, wherein the determination as to whether the welding is of acceptable quality or unacceptable quality and/or wherein the one or more actions taken are displayed and/or stored.
17. A method according to any preceding claim, wherein the method includes providing a correlation between the data obtained from at least four elected sensor types in the plurality of elected sensor types.
18. A method according to any preceding claim, wherein at least one elected sensor type of the plurality of elected sensor types is a part of weld inspection apparatus and the method includes a step of inspecting the weld using the weld inspection apparatus.
19. A method according to claim 18, wherein the method includes a step of inspecting the weld using weld inspection apparatus to determine one or more characteristics of a defect.
20. A method according to claim 19, wherein the characteristics include one or more of size, position, defect type, defect shape or defect position relative to the geometry of the weld and/or relative to the length of weld with reference to the geometry of the weld and/or relative to the length of weld.
21. A method according to claim 19 or claim 20, wherein the method further includes a comparison of one or more of the characteristics with one or more standards, and further includes a determination of whether the weld with the defect meets a weld standard or does not meet the weld standard.
22. A method according to claim 21, wherein if the weld meets the weld standard, then a record for the weld is created and stored.
23. A method according to claim 22, wherein the method further includes that the record includes data from one or more of the plurality of sensor types.
24. A method according to claim 21, wherein if the weld does not meet the weld standard, one or more remedial steps are applied to the weld.
25. A method according to any preceding claim, wherein the at least two elected sensor types in the plurality of elected sensor types are weld conditions sensors and the method includes a step of inspecting the weld conditions using the weld condition sensors.
26. A method according to claim 25, wherein the method includes a step of inspecting the weld conditions to determine one or more parameters of the weld as it is formed.
27. A method according to claim 26, wherein the method further includes a comparison of one or more of the parameter with one or more control parameters, and further includes a determination of whether a risk level for weld defects is exceeded.
28. A method according to any preceding claim, wherein the method further includes the one or more actions being to alter the welding conditions from the first set of welding conditions for the welding method.
29. A method according to claim 28, where the alteration in the welding conditions from the first set of welding conditions is to stop welding and/or alert an operator.
A method according to claim 28 or claim 29, wherein the alteration in the welding conditions from the first set of welding conditions is to change the welding conditions back to the first set of conditions and/or to change the welding conditions to a second set of conditions.
31 A method according to any preceding claim, wherein at least one of the at least two elected sensor types are selected from: voltage sensors, current sensors, welding arc sound emission sensors, weld topology sensors, weld imaging sensors and ultrasound imaging sensors.
32. Apparatus for monitoring welding, the apparatus comprising: a) a plurality of inputs for data from a plurality of sensor types; b) one or more processors, wherein a processor from amongst the one or more processors: a. receives the inputs; b. processes the data obtained from at least two elected sensor types amongst the plurality of senor types to apply a correlation between the data obtained from at least two elected sensor types amongst the plurality of senor types; c. processes the data obtained from at least two elected sensor types amongst the plurality of senor types using the correlation to synchronise one or more data points in the data obtained from one of the at least two elected sensor types with one or more data points in the data obtained from another of the at least two elected sensor types; c) one or more outputs for processed data.
33 Apparatus according to claim 32 or any claim depending thereon, wherein the processor is adapted to provide a temporal correlation, the output data from the processor including temporal information in the data from the one of the at least two elected sensors and temporal data in the data from the another of the at least two elected sensor types in the data.
34 Apparatus according to claim 32 or any claim depending thereon, wherein the processor is adapted to provide a positional correlation, the output data from the processor including positional information in the data from the one of the at least two elected sensors and positional data in the data from the another of the at least two elected sensor types in the data.
Apparatus according to claim 32 or any claim depending thereon, wherein the processor is adapted to provide in the data from the at least two elected sensor types, a time of occurrence for one or more data points within the data, and wherein the correlation is a matching time of occurrence for data points in the data from the at least two elected sensor types.
36. Apparatus according to claim 32 or any claim depending thereon, wherein the processor is adapted to provide in the data from the at least two elected sensor types, a position of occurrence for one or more data points within the data, and wherein the correlation is a matching position of occurrence for data points in the data from the at least two elected sensor types.
37. Apparatus according to any of claims 32 to 36, wherein inputs are from at least one of the plurality of sensor types, potentially at least one of the elected sensor types, wherein the sensor types and/or elected sensor types are selected from: voltage sensors, current sensors, welding arc sound emission sensors, weld topology sensors, weld imaging sensors and ultrasound imaging sensors.