Investigating the Detection Capability of Acoustic Emission Monitoring to Identify Imperfections Produced by the Metal Active Gas (MAG) Welding Process

Round 1

Reviewer 1 Report

The manuscript "Investigating the Feasibility of Using Acoustic Emission Sensors for Weld Quality Inspection" presents an interesting study investigating the use of acoustic emission sensors as a non-destructive and less time-consuming method for inspecting weld quality. The paper is well-structured, and the methodology used is presented.

Revisions:

1. Provide more information on the extracted data from the sensors and how it was analyzed.
2. The experimentation setup should be real.
3. The authors should provide a detailed description of the experimental design, including the selection of parameters, the setup of the acoustic emission sensors, and the data analysis methods. Additionally, the authors should provide a rationale for selecting the parameters used in the experiment.
4. The discussion section should provide a more detailed interpretation of the results obtained. The authors should discuss the implications of their findings and provide recommendations for future research.
5. Provide a more detailed conclusion summarizing the study's main findings and their implications.
6. Include a discussion on the limitations of the study.
7. Why was the signal represented using STFT? Extend more discussion using recent papers such as "Development of Deep Belief Network for Tool Faults Recognition," "Demagnetization Fault Diagnosis of Permanent Magnet Synchronous Motors Based on Stator Current Signal Processing and Machine Learning Algorithms."

Author Response

From previous Part 2 (Reviewer comments):

We will display the part 2 comments first as the original reviewer comments were made to this paper. Part 1 was refused due to lack of novelty and scope as it stood on its own but nevertheless the authors have addressed these comments too to ensure the paper is a good level to take forward.

Reviewer 1

The manuscript "Investigating the Feasibility of Using Acoustic Emission Sensors for Weld Quality Inspection" presents an interesting study investigating the use of acoustic emission sensors as a non-destructive and less time-consuming method for inspecting weld quality. The paper is well-structured, and the methodology used is presented.

1. Provide more information on the extracted data from the sensors and how it was analyzed.

This was an oversight of the authors and was not included – now the text below has been added to aid greater understanding for the reader.

Line 809 Page 13:

For the GRAS airborne microphone, the system’s software allows the user to action the acquisition with the parameters set for recording as mentioned in the setup.  This is also the same for MISTRAS contact AE sensor, where the synchronisation comes about from the activity of the MAG arc phenomena recorded at the same time where both events are aligned. The USB AE MISTRAS system produced a parameter recorded file where outputs such as amplitude, AE energy, AE counts and time responses were used to discriminate the phenomena at the position of interest.

As the airborne microphone records a constant recording, which is a continuous time series waveform, it can be easily converted to a time-frequency content. The MISTRAS contact AE sensor, however, records events as output parameters and each parameter contains a short burst of waveform data.  However this is not a continuous time series and therefore only time-frequency content can result from the burst waveform data attached to each output parameter.  

2. The experimentation setup should be real.

We agree with your comments and the experimental setup was based on previous works to induce anomalies of interest to prove the sensitivity of the  detection method and the following text produced to convey such observations.

Page 538 Page 13:

Varying metal inserts of increasing complexity in chemistry were deployed to enhance imperfection sensitivity within the deposit.  The chemistry of the three types of inserts used (alloy 718, CMSX4 and EN8 Steel) are provided in table 2.

3. The authors should provide a detailed description of the experimental design, including the selection of parameters, the setup of the acoustic emission sensors, and the data analysis methods. Additionally, the authors should provide a rationale for selecting the parameters used in the experiment (1).

Initially we thought this was covered well and then specifics in regard to the sensor and recording s/w were not present and therefore the following text has been modified and updated:  

Line 622 Page 16:

Sensors deployed were the GRAS 46BE, ¼ inch constant current power CCP airborne microphone (contactless) and MISTRAS (PAC) ISPKWDI wide band microphone (contact). The GRAS airborne microphones were placed a distance of 35 mm to reflect the work of [20] Ladislav et al., (2003). The contact AE sensor was connected 100 mm from the start of the weld line [21] (Cayo and Absi, (2008) to record the phenomena further way.  The setup schematic for welding insert tests can be seen in figure 3. The GRAS microphone used Sinus’s Apollo USB acoustics analyser to record the airborne sound acquisition. The sampling rate used to capture the airborne soundwaves was 200kHz. The MISTRAS sensor however used AE Win for USB where the acquisition system could be run from a laptop USB as a portable solution simulating real time monitoring in the field. The sampling rate used for the contact AE sensor was 2MHz to ensure the full bandwidth of ISPKWDI sensor could be recorded between 100-800kHz. The ISPKWDI sensor is an intrinsically safe AE sensor with low power integral pre-amplifier.  Each recording (both contact and none contact sensors) were triggered when the machine controller starts the welding process and therefore was synchronised to the phenomena of interest.       

4. The discussion section should provide a more detailed interpretation of the results obtained. The authors should discuss the implications of their findings and provide recommendations for future research.

This is good point made by the reviewer and the following has been added:

Line 1104 Page 34:

The STFT analysis of the GRAS airborne acoustic signal is difficult to see the changing amplitude as all signals recorded for the baseline cases have a frequency coverage across the whole sensor bandwidth spectrum. In addition, to make easy comparisons, all the amplitude colour maps were normalised by the same amount. By providing such normalisation can restrict features from being more salient.

The Mistras contact AE sensor produces output parameters that are recorded based on passing a dB noise threshold which means the overall signal is not continuous with time however gives one snap shots of events and each one of those snap shots give small continuous waveforms which can have STFTs applied to it. For future work, a two channel system would be more beneficial where STFTs could be compared where both channels would give two fully continuous signals. In addition, phenomena location would also exist.  The work mentioned here looks at a single AE USB system which is more likely for a test engineer as it’s easier to setup in remote locations.  

5. Provide a more detailed conclusion summarizing the study's main findings and their implications.

Line 1119 page 34:

The use of materials offering a weldability gradient in the form of chemistries that vary significantly in chemistry, welded to an ASTM A36 steel with a fully austenitic welding wire (307Si) promotes a varied set of events through expanding the solidification profile.  Alloy 718 and CMSX-4 have proven to be ideal candidates in inducing favourable imperfections with such a combination of base metal and filler metal chemistries.  These materials promote the formation of topological close-packed phases e.g., the intermetallic and brittle Laves phase that accentuated an enhance crack sensitivity, which in turn improved the ability to test the potential use of remote NDT methods to determine a weld’s integrity.

Through using nickel-base alloy inserts, it has been possible to create additional complexities to promote geometrical changes to instigate gross imperfections in the form porosity.  This combination increased the sensitivity reminiscent of the circular patch weldability test developed by R.D. Stout and published within the welding research council in 1987, which was used to determine the weldability of steels.   This chemistry combination heightens microsegregation due to the rejection of solute elements that coalesce and promote low-temperature liquid films, which decorate grain boundaries and rupture.  Combine this with the change in metal flow, resulting from differential rheological properties and thermal expansion values, further induced microscopical stress regions at the dendritic and grain boundary scale that grew into macro-stressed regions, which resulted in cracking.  This is the result experienced with the ASTM A36 and 307Si + alloy 718 and ASTM A36 and 307Si + CMSX-4 combinations. Such multipart solutions were proven to increase the propensity for hot cracking and cavities, that release sonic waves at specific frequencies. 

It was concluded that there could be a strong correlation between ultrasound omitted during MAG welding and corresponding joint integrity in terms of crack response that could be of significant value to fabricators. Certain frequency ranges and amplitudes are potentially unique for different material chemistries exposed to welding. The frequency rise during and immediately after the process denotes the formation of imperfections, phase transformation or defects in the weld. AE sensors, both airborne and contact can identify such frequencies and amplitudes, which guides automation in terms of acceptable or non-acceptable weld integrity.  AE sensors can be integrated into manufacturing systems for real-time inspection rather than having to undertake timely, expensive, and destructive material analysis to qualify the weld structure.

It was noticed that STFT plots of the AE airborne microphone gave much higher intensities and frequency band utilisation for regions consistent with cracking. The same was observed for the contact sensors where the frequency range was between 250 - 550 kHz with high intensities of 70 dB. Such values were significant indicators of cracking during the cooling phase, circa 80 seconds after welding. This shows that with using high-resolution instrumentation and advanced Digital Signal Processing (DSP) techniques it is possible to obtain useful in-situ and post-machining information to discern cracks and thus the general weld quality. 

For this work to be accepted into the general fabrication mainstream further material analysis with AE sensors is required, which should aim to generalise all welding anomaly conditions in-situ. In addition to this, such anomalies are indicative to certain frequency ranges and amplitudes.  This information could, in future, be fed directly into filter-banks and real-time monitoring control systems as well as advanced material models and machine learning architecture. Further work will look into similar work using thermocouples or IR camera systems processed with Digital Image Correlation. More work will be designed around establishing the different associated welding anomalies and all data will be processed using established and novel machine learning techniques to show the accuracy and acceptance of automated signature inspection.

6. Include a discussion on the limitations of the study.

This is a good comment by the reviewers and we agree, such a discussion will ensure a wider audience will receive and use such works.

Line 664 Page 17:

In terms of limitations for the experiments carried out within this research, the different insert materials stimulates the majority of joining anomalies, specifically that of crack formation and propagation, which affects the weld integrity and therefore important for proof of concept for a welding defect detection system. There are other recognised anomalies indicative of a poor setup however these are for future work.

7. Why was the signal represented using STFT? Extend more discussion using recent papers such as "Development of Deep Belief Network for Tool Faults Recognition," "Demagnetization Fault Diagnosis of Permanent Magnet Synchronous Motors Based on Stator Current Signal Processing and Machine Learning Algorithms."

The authors agree with this comment and the need to highlight why STFTs have been used for the research work. In addition, they have looked at these two references and more to increase the transparency and verification of why STFT was used over other techniques. The following text is what they wish to add to the manuscript:

Line 424 Page 10:

AE signals used to distinguish cracks from porosity would look at short burst high amplitude data vs. shorter decay time and lower amplitude respectively as discussed by Roca et al [32]. These differences were stored within an Artificial Neural Network to give a computer model of the gas-metal arc welding (GMAW) process. This research is perhaps the closest research discussed in the paper where several AE parameters are used to distinguish different crack phenomena.  However no machining learning work has been used within this research work, as the focus is purely on connecting the signal analysis phenomena with the physical material analysis phenomena.

The review made by [32] discussed frequency domain, time-frequency analysis methods such as empirical mode decomposition (EMD) and STFT provide useful information about the type and location of the defect. This is one major reason why STFT is used over other techniques. Further to this work, the research completed by [28] reinforces the use of STFTs to provide demagnetisation fault diagnosis of Permanent Magnet Synchronous Motors (PMSM). The STFT provided the stator current signals for both the fault, and non-fault case where significant harmonics were in more abundance and more salient for the former compared with the latter. STFTs however have not been seen before in fault diagnostics where CWTs, Hilbert – Huang transform (HHTs) and Wigner-Ville distribution (WVD) have been used before. This is not the case for joining technologies where anomalies and none anomalies have been displayed before using STFTs. However specific to the setup in terms of material inserts, and MAG process, the authors of this paper believe this to be a first. Kale et al [33] used a number of machine learning techniques to separate the anomalies which is promising considering future work associated with this research is intended to pursue future ideas, correlating automatically welding defects using AE sensors (both contact and none contact) along with DSP to discriminate between an acceptable and non-acceptable weld. Work presented in a paper by Pietrzakand Wolkiewicz [34] identified that STFT analysis of vibration signals allowed the differentiation between a good machine cutting tool and machine cutting tools with five different faulty conditions. Similar to this work [33] and, [34] used deep belief networks to differentiate between different states. Vibration extracted signals use similar technologies to AE which again reinforces why STFTs should be used in preference to other techniques. 

Author Response File: Author Response.pdf

Reviewer 2 Report

In the reviewed paper, the authors are investigating experimentally the influence of welding conditions on the acoustic emission as well as to the airborn-detected signals. The presence of various inserts in the welding area is clearly visible in the acquired signals. It might be judged from the obtained results that simultaneous employment of AE and airborne sensors could enhance damage detection capabilities during the welding process. In general, the paper content is within the scope of the Acoustics journal and could have some scientific merit for the scholars and engineers dealing with welding control. However, certain improvements could be introduced to the paper.

The main issue is that the review of the current research endeavors in the research field is slightly out-of-date. The latest paper mentioned in subsection 2.2 is dated back to the year 2015. Certainly, there exist more recent developments which could be mentioned. For instance, in DOI:10.1109/ACCESS.2019.2935117, time-freqency analysis was applied for the classification of acoustic emission events during the welding, in DOI: 10.1007/s00170-018-3042-2, simultaneous employment of air-coupled and conventional contact-based AE sensors is discussed; a rather comprehensive review of in-situ approaches for welding quality control is provided in DOI: 10.1007/s40194-021-01229-6. These papers are mentioned here just as examples and could encourage the authors to provide more details regarding the novelty of the current paper with respect to them and to many other recent findings in the field of in-situ welding quality control.

Some minor remarks regarding the manuscript text:

1) Lines 236 and 263 - there is a contradiction between the distance from airborne microphone to the investigated sample (35 or 350mm ?)

2) Line 241: "therefore filtration is carried out both through software and hardware means." How exactly was it done?

3) What were the particular models of GRAS airborne microphone and TRAS (PAC) ISPKWDI AE sensor? What was the signal acquisition equipment? What particular equipment was used for welding?

4) For clarity, the length of top and bottom subplots in Figure 2 and further on should coincide so that time history could be immediately matched with its STFT.

5) What were the parameters of the STFT?

6) Line 348: Are the authors sure that kHz should be used here? - the general suggestion - the authors should carefully check all scales and values throughout the paper text and plots.

7) Line 322: "... and frequencies are much higher in ..." - From the presented plots it is not clear to what extent the frequencies are higher.

8) Since with contact-based AE sensors the signals are acquired during the cooling process after the welding is ended, how such testing could be referred to "real time"? (Lines 492-493).

Author Response

Reviewer 2
In the reviewed paper, the authors are investigating experimentally the influence of welding conditions on the acoustic emission as well as to the airborn-detected signals. The presence of various inserts in the welding area is clearly visible in the acquired signals. It might be judged from the obtained results that simultaneous employment of AE and airborne sensors could enhance damage detection capabilities during the welding process. In general, the paper content is within the scope of the Acoustics journal and could have some scientific merit for the scholars and engineers dealing with welding control. However, certain improvements could be introduced to the paper.

The main issue is that the review of the current research endeavors in the research field is slightly out-of-date. The latest paper mentioned in subsection 2.2 is dated back to the year 2015. Certainly, there exist more recent developments which could be mentioned. For instance, in DOI:10.1109/ACCESS.2019.2935117, time-freqency analysis was applied for the classification of acoustic emission events during the welding, in DOI: 10.1007/s00170-018-3042-2, simultaneous employment of air-coupled and conventional contact-based AE sensors is discussed; a rather comprehensive review of in-situ approaches for welding quality control is provided in DOI: 10.1007/s40194-021-01229-6. These papers are mentioned here just as examples and could encourage the authors to provide more details regarding the novelty of the current paper with respect to them and to many other recent findings in the field of in-situ welding quality control.

Line 370 Page 9:

In terms of the specific digital signal processing techniques to show separation between the two sources of extracted acoustic date the following discussions are necessary.

He and Li’s work [28] used time-frequency analysis in the form of continuous wavelet transform (CWT) in order to compress the AE signal and use it as a signature fingerprint to discriminate between a good/bad MIG weld. PCAs are used to determine the most significant compressed-transformed data of the CWT applied AE data. A Support Vector Machine is then used to provide automated classifications between a good/bad weld. PCA used as the CWT AE information can produce computationally heavy data signatures on their own. This is perhaps the closest work completed when compared with the discussed work here, where short time Fourier transform (STFT) is used a method to differentiate such phenomena. STFT was chosen as CWT was considered more computationally expensive and needing techniques such as PCA to display signal differences. STFTs have been used by the authors before in previous work [27] with the setup experiment providing definite phenomena from material interactions.  STFT is perfectly adequate to describe such focused phenomenon of interest, because converting from CWT to the time-frequency domain can result in a loss of substantial information.  

Basantes-Defaz et al, [29] research also discusses very similar work to what is discussed here. However, instead of using a non-contact microphone and a contact acoustic emission sensor (AE), as in this paper’s research [29] looks at using two contact AE sensors as airborne sensors, which is certainly novel, because the use of contact dedicated to non-contact, airborne audio acoustic activity is very unorthodox. Surely using dedicated airborne microphones as in the research presented here is more appropriate and sensitive to airborne change phenomena? Within the work presented in this paper both contact and non-contact microphones have been used, where the former monitors phenomena in the ultrasonic frequency range and the latter, phenomena in the audio frequency range. Nevertheless, the airborne acoustic signature is obtained in [29] and used in a qualitative manner to represent the depth of penetration of the weld where different signatures result in different achieved depths of weld penetration. The two AE sensors work independent of each other as they both have different characteristics; one being specifically used for a low frequency range, and the other, a wide band frequency range. In addition, the dB noise threshold for sensitivity is set differently for each sensor. The third sensor however is configured in a conventional manner on a track to move along at the same rate as the weld torch. The sensor in this case is a contact ultrasonic pulse echo system and based on time of flight measurements, which can discriminate between different material characteristics through a change of boundaries based on different recorded penetration energies.

This work not only verifies the setup with the work being discussed here, it also discusses that such technologies play a vital role in the detection of welding anomalies. Finally, the work discussed by [29] concentrates specifically on anomaly identification namely burn through.  Further work will look at using machine learning technologies to provide automatic detection. The work presented in this paper also concentrates on anomalies, namely weld quality and crack detection. Future work will also look at automated discrimination through the use of machine learning technologies.

Interrogating Basentes-Defaz et al’s research work [29] revealed that data collected by these AE sensors also provided information on possible superficial discontinuities or defects in the weld metal.  If a superficial indentation or defect is displayed, both the ASL scan and the AE absolute energy plot showed the exact location of this surface discontinuity, which was located at 85 mm from the start of the weld. The sudden signal burst that occurred at the exact location of the superficial indication was determined confirming that where a non-uniform condition or weld defect appears this appears as a sudden surge in the AE absolute energy and therefore a key indicator in determining the signature and potential characteristic for automatic detection. These observations correlate very well with the findings in the work and  displayed in this paper.

There are a few sources such as Madhvacharyula et al [30], , Kanungo et al [31], and Kale et al [33] that have started reporting machine learning techniques applied to detected AE sources that can indicate welding defects. Moreover  Kanungo et al’s work [31] used cluster k-means analysis to show the more significant features from several AE parameters namely, peak amplitude, kurtosis, energy and the number of counts. As there are several dimensions here and cluster k-means analysis requires the most significant features, principal component analysis is used.

AE signals used to distinguish cracks from porosity would look at short burst high amplitude data vs. shorter decay time and lower amplitude respectively as discussed by Roca et al [32]. These differences were stored within an Artificial Neural Network to give a computer model of the gas-metal arc welding (GMAW) process. This research is perhaps the closest research discussed in the paper where several AE parameters are used to distinguish different crack phenomena.  However no machining learning work has been used within this research work, as the focus is purely on connecting the signal analysis phenomena with the physical material analysis phenomena.

Minor Revisions

1. Lines 236 and 263 - there is a contradiction between the distance from airborne microphone to the investigated sample (35 or 350mm ?)

This should be 35mm throughout and the text has been modified – many thanks for spotting this and pointing it out.

2. Line 241: "therefore filtration is carried out both through software and hardware means." How exactly was it done? (2)

This means the hardware filters within MIstras propriety s/w has hardware filters set up which provide a low pass and high pass filtration to remove any unwanted noise. The s/w filters however are also within the Mistras propriety s/w and this is where a pickup noise limit will be set and therefore only record when the signal surpasses this noise level.

Line 600 Page 15:

The hardware filtering on the sensor used a Chebyshev 20kHz high pass filter to remove any unwanted mechanical and audible noise and, a Chebyshev 1000kHz low pass filter was used to remove any unwanted white noise. The software filtering is based around recording past a defined noise limit, which was set at 35dB. It is expected that the contact sensors will give a higher resolution to distinguish finer features than that expected of the airborne sensors.

3. What were the particular models of GRAS airborne microphone and TRAS (PAC) ISPKWDI AE sensor? What was the signal acquisition equipment? What particular equipment was used for welding? (3)

Text modified to reflect the suggested changes. See below:

Line 627 Page 16:

The GRAS microphone used Sinus’s Apollo USB acoustics analyser to record the airborne sound acquisition. The sampling rate used to capture the airborne soundwaves was 200kHz. The MISTRAS sensor however used AE Win for USB where the acquisition system could be run from a laptop USB as a portable solution simulating real time monitoring in the field. The sampling rate used for the contact AE sensor was 2MHz to ensure the full bandwidth of ISPKWDI sensor could be recorded between 100-800kHz. The ISPKWDI sensor is an intrinsically safe AE sensor with low power integral pre-amplifier.  Each recording (both contact and none contact sensors) were triggered when the machine controller starts the welding process and therefore was synchronised to the phenomena of interest.       

4. For clarity, the length of top and bottom subplots in Figure 2 and further on should coincide so that time history could be immediately matched with its STFT.

We have modified Figures in the manuscript to include the suggested changes.

Line 837 Page 23:

Line 848 Page 24:

Line 863 Page 24:

Line 872 Page 24:

5. What were the parameters of the STFT?

Line 820 Page 25:

The matlab function specgram and selected parameters were used to define the STFT for GRAS airborne signal:

specgram(filename,128,204800,kaiser(128,10),87);

Where the sampling rate was 204,800 kHz, 128 was the number of points for block size within the signal, the window allows the FFT segments and in this case, a Kaiser matlab window function is used which has a resolution length of 128 points and shape factor beta value of 10. The parameter at the end gives a number of points overlap to embed the window function into the signal giving a continuous image. Such parameters gave a lot more resolution in the time domain compared to the frequency domain where more clarity was required.

specgram(filename,1024,2000000,Kaiser(1024,10),875);  

The above function and parameter values was used for the AE contact sensors which translated the AE data to burst mode time-frequency AE data. As the AE contact sensors provide much higher pickup bandwidth frequency, the sampling frequency was 2MHz and a resolution time length for the Kaiser window at 1024 points giving better resolution in the time domain than the frequency domain (similar amount was used for window overlap to embed the continuous signal into the image). 

6. Line 348: Are the authors sure that kHz should be used here? - the general suggestion - the authors should carefully check all scales and values throughout the paper text and plots.

The scales have been checked and are correct. Where the STFT plots display the continuous waveform of the airborne microphone showing a bandwidth upto 100 kHz. The STFT for the burst plots is based on continuous data but only continuous waveforms for each parameter recorded as a burst signal – this bandwidth is from 100 kHz to 1000 kHz. 

7. Line 322: "... and frequencies are much higher in ..." - From the presented plots it is not clear to what extent the frequencies are higher.

Yes, thank you for your comments and we agree as the comments within paper are confusing and so have been changed:

Line 856 page 24:

The remainder of the emitted higher frequencies display the continued weld over the insert material in comparison with Figure 8, the baseline case. The displayed amplitude intensities are much higher in the Alloy 718 insert case displayed in Figure 9, the baseline case. Such results give credence to detecting welding anomalies in-situ with airborne sensors. The differences in emitted frequencies and noise levels can be attributed to the changes in arc noise relating to abnormalities during the solidification terminus.

8. Since with contact-based AE sensors the signals are acquired during the cooling process after the welding is ended, how such testing could be referred to "real time"? (Lines 492-493).

Yes, this comment is confusing and so the following text has been changed to the following:

Line 1052 page 33:

Contact AE was found to give more significant results in the identification of defects within the weld in real time and during its cooling phase.

Author Response File: Author Response.pdf

Reviewer 3 Report

Most of the suggestions related to Part 1 (Materials Analysis Perspective) can be written here. Some of the suggestions will be repeated and some are new.

In general, combining these two manuscripts into one would make for a great research.

The study addressed examination the sensitivities of metal active gas welding imperfections through monitoring acoustic emission. The topic is relevant to the research field. The originality of the research is currently hidden so that it is not clearly stated what the scientific contribution is compared to previously published manuscripts. The conclusions are written in accordance with the presented evidence and arguments, but could be supplemented. References are appropriate but too few. More recent references are also missing. Tables and figures are clear.

1. In this research (Part 2 - Signal Analysis Perspective) there are identical sentences as in the previous research (Part 1 - Materials Analysis Perspective). It is not even possible to cite, since currently none of the manuscripts have been published. I think that the combination of these two manuscripts was the best solution for the authors of the article.

2. Technical suggestion. References are not adequately cited in the text. Please see Instructions for Authors (In the text, reference numbers should be placed in square brackets [ ], and placed before the punctuation; for example [1], [1–3] or [1,3]. For embedded citations in the text with pagination, use both parentheses and brackets to indicate the reference number and page numbers; for example [5] (p. 10). or [6] (pp. 101–105).) Also, the references format is not adequate (References section).

3. The authors used 18 references. The references are relevant, but are too few. The section pertaining to review of previous contributions in the field must be expanded with other references, especially the most recent ones.

4. The abstract is not well written. It needs to be rewritten. The abstract must be presented in a clear way: problem, objective, idea, description of idea, methods, results, quantitative comparison of results with significant findings, conclusions.

5. The authors should clarify the originality and the scientific contributions given by the manuscript. This is expected above all in the Introduction, but also in other parts of the paper. The authors should clarify the originality and the scientific contributions given by the manuscript. Manuscripts must explain the significant advances provided in approaches and understanding compared to previous literature, and/or demonstrate convincingly potential in new applications. A highlight of your hypothesis, new concepts and innovations.

6. The state-of-the-art comparisons for the proposed work are missing in this paper. Then do a critical analysis of previous research. State explicitly the shortcomings of previous research. What is positive in previous research and what is negative. Based on that, you explicitly define the goal of the research and the scientific hypothesis.

7. The shortcoming of this research is that potential errors and measurement uncertainty were not analysed and discussed.

8. The obtained results could be additionally discussed and, if possible, compared with the results of previous research.

9. In the conclusions, emphasize the scientific contribution, the innovativeness of the research, as well as the possibilities of practical application.

Author Response

The study addressed examination the sensitivities of metal active gas welding imperfections through monitoring acoustic emission. The topic is relevant to the research field. The originality of the research is currently hidden so that it is not clearly stated what the scientific contribution is compared to previously published manuscripts. The conclusions are written in accordance with the presented evidence and arguments, but could be supplemented. References are appropriate but too few. More recent references are also missing. Tables and figures are clear.

1. In this research (Part 2 - Signal Analysis Perspective) there are identical sentences as in the previous research (Part 1 - Materials Analysis Perspective). It is not even possible to cite, since currently none of the manuscripts have been published. I think that the combination of these two manuscripts was the best solution for the authors of the article.

The authors agree with these comments and hence why the corrections have taken longer than expected where there was a need for authors to get together and state what is necessary and what is not and then merge the two papers together in a consistent manner.

2. Technical suggestion. References are not adequately cited in the text. Please see Instructions for Authors (In the text, reference numbers should be placed in square brackets [ ], and placed before the punctuation; for example [1], [1–3] or [1,3]. For embedded citations in the text with pagination, use both parentheses and brackets to indicate the reference number and page numbers; for example [5] (p. 10). or [6] (pp. 101–105).) Also, the references format is not adequate (References section). (9 – do later)

The authors have made the necessary changes.

3. The authors used 18 references. The references are relevant, but are too few. The section pertaining to review of previous contributions in the field must be expanded with other references, especially the most recent ones.

The authors agree with this statement and the following references and text have been added to the paper:

               Line 370 page 9:

In terms of the specific digital signal processing techniques to show separation between the two sources of extracted acoustic date the following discussions are necessary.

He and Li’s work [28] used time-frequency analysis in the form of continuous wavelet transform (CWT) in order to compress the AE signal and use it as a signature fingerprint to discriminate between a good/bad MIG weld. PCAs are used to determine the most significant compressed-transformed data of the CWT applied AE data. A Support Vector Machine is then used to provide automated classifications between a good/bad weld. PCA used as the CWT AE information can produce computationally heavy data signatures on their own. This is perhaps the closest work completed when compared with the discussed work here, where short time Fourier transform (STFT) is used a method to differentiate such phenomena. STFT was chosen as CWT was considered more computationally expensive and needing techniques such as PCA to display signal differences. STFTs have been used by the authors before in previous work [27] with the setup experiment providing definite phenomena from material interactions.  STFT is perfectly adequate to describe such focused phenomenon of interest, because converting from CWT to the time-frequency domain can result in a loss of substantial information.  

Basantes-Defaz et al, [29] research also discusses very similar work to what is discussed here. However, instead of using a non-contact microphone and a contact acoustic emission sensor (AE), as in this paper’s research [29] looks at using two contact AE sensors as airborne sensors, which is certainly novel, because the use of contact dedicated to non-contact, airborne audio acoustic activity is very unorthodox. Surely using dedicated airborne microphones as in the research presented here is more appropriate and sensitive to airborne change phenomena? Within the work presented in this paper both contact and non-contact microphones have been used, where the former monitors phenomena in the ultrasonic frequency range and the latter, phenomena in the audio frequency range. Nevertheless, the airborne acoustic signature is obtained in [29] and used in a qualitative manner to represent the depth of penetration of the weld where different signatures result in different achieved depths of weld penetration. The two AE sensors work independent of each other as they both have different characteristics; one being specifically used for a low frequency range, and the other, a wide band frequency range. In addition, the dB noise threshold for sensitivity is set differently for each sensor. The third sensor however is configured in a conventional manner on a track to move along at the same rate as the weld torch. The sensor in this case is a contact ultrasonic pulse echo system and based on time of flight measurements, which can discriminate between different material characteristics through a change of boundaries based on different recorded penetration energies.

This work not only verifies the setup with the work being discussed here, it also discusses that such technologies play a vital role in the detection of welding anomalies. Finally, the work discussed by [29] concentrates specifically on anomaly identification namely burn through.  Further work will look at using machine learning technologies to provide automatic detection. The work presented in this paper also concentrates on anomalies, namely weld quality and crack detection. Future work will also look at automated discrimination through the use of machine learning technologies.

Interrogating Basentes-Defaz et al’s research work [29] revealed that data collected by these AE sensors also provided information on possible superficial discontinuities or defects in the weld metal.  If a superficial indentation or defect is displayed, both the ASL scan and the AE absolute energy plot showed the exact location of this surface discontinuity, which was located at 85 mm from the start of the weld. The sudden signal burst that occurred at the exact location of the superficial indication was determined confirming that where a non-uniform condition or weld defect appears this appears as a sudden surge in the AE absolute energy and therefore a key indicator in determining the signature and potential characteristic for automatic detection. These observations correlate very well with the findings in the work and  displayed in this paper.

There are a few sources such as Madhvacharyula et al [30], , Kanungo et al [31], and Kale et al [33] that have started reporting machine learning techniques applied to detected AE sources that can indicate welding defects. Moreover  Kanungo et al’s work [31] used cluster k-means analysis to show the more significant features from several AE parameters namely, peak amplitude, kurtosis, energy and the number of counts. As there are several dimensions here and cluster k-means analysis requires the most significant features, principal component analysis is used.

AE signals used to distinguish cracks from porosity would look at short burst high amplitude data vs. shorter decay time and lower amplitude respectively as discussed by Roca et al [32]. These differences were stored within an Artificial Neural Network to give a computer model of the gas-metal arc welding (GMAW) process. This research is perhaps the closest research discussed in the paper where several AE parameters are used to distinguish different crack phenomena.  However no machining learning work has been used within this research work, as the focus is purely on connecting the signal analysis phenomena with the physical material analysis phenomena.

The review made by [32] discussed frequency domain, time-frequency analysis methods such as empirical mode decomposition (EMD) and STFT provide useful information about the type and location of the defect. This is one major reason why STFT is used over other techniques. Further to this work, the research completed by [28] reinforces the use of STFTs to provide demagnetisation fault diagnosis of Permanent Magnet Synchronous Motors (PMSM). The STFT provided the stator current signals for both the fault, and non-fault case where significant harmonics were in more abundance and more salient for the former compared with the latter. STFTs however have not been seen before in fault diagnostics where CWTs, Hilbert – Huang transform (HHTs) and Wigner-Ville distribution (WVD) have been used before. This is not the case for joining technologies where anomalies and none anomalies have been displayed before using STFTs. However specific to the setup in terms of material inserts, and MAG process, the authors of this paper believe this to be a first. Kale et al [33] used a number of machine learning techniques to separate the anomalies which is promising considering future work associated with this research is intended to pursue future ideas, correlating automatically welding defects using AE sensors (both contact and none contact) along with DSP to discriminate between an acceptable and non-acceptable weld. Work presented in a paper by Pietrzakand Wolkiewicz [34] identified that STFT analysis of vibration signals allowed the differentiation between a good machine cutting tool and machine cutting tools with five different faulty conditions. Similar to this work [33] and, [34] used deep belief networks to differentiate between different states. Vibration extracted signals use similar technologies to AE which again reinforces why STFTs should be used in preference to other techniques. 

All the above reinforces the use of STFTs over other transforms. Using STFTs with the specific setup has not been reported before in previous published work. Furthermore, the raw AE parameters being recorded to distinguish welding defects, AE rise time has not been used before. The reason for such a parameter being used was down to it being less noise prone as a source of error when obtaining other parameters such as ASL scan and amplitude energy. Another aspect to the novelty of using this parameter is based on the recording after the welding event to note anomalies during material settlement during the cooling phase and this also as not been communicated in previous literature. 

4. The abstract is not well written. It needs to be rewritten. The abstract must be presented in a clear way: problem, objective, idea, description of idea, methods, results, quantitative comparison of results with significant findings, conclusions.

This abstract has now been merged with the first papers abstract and improved.

Abstract. Welding inspection is a critical process that can be severely time-consuming resulting in productivity delays, especially when destructive or invasive processes are required. This paper defines the novel approach to investigate the physical correlation between common imperfections found in arc welding and the propensity to determine these through the identification of signatures using acoustic emission sensors.  Through a set of experiments engineered to induce prominent imperfections (cracks and other anomalies) using a popular welding process and the use of AE technology (both airborne and contact) it provides confirmation that the verification of physical anomalies can indeed be identified through variations in obtained noise frequency signatures. This in-situ information provides signals during and after solidification to inform operators of the deposit/HAZ integrity to support advanced warning of unwanted anomalies and whether the weld/fabrication process should be halted to undertake rework before completing the fabrication. Experimentation was carried out based on an acceptable set of parameters where extracted data from the sensors were recorded, analysed, and compared with the resultant microstructure. This would allow signal phenomena to be captured and catalogued for future use in referencing against known anomalies.  

5. The authors should clarify the originality and the scientific contributions given by the manuscript. This is expected above all in the Introduction, but also in other parts of the paper. The authors should clarify the originality and the scientific contributions given by the manuscript. Manuscripts must explain the significant advances provided in approaches and understanding compared to previous literature, and/or demonstrate convincingly potential in new applications. A highlight of your hypothesis, new concepts and innovations.

The two paragraphs below are summary discussions within the literature review section where the novelty and scientific contributions are given.

Line 354 page 9:

The discussed investigations reinforce the applicability of using both airborne and contact methods of AE for the determination of weld surface and volumetric integrity both in-situ or off-line.

As there are many thermal and kinematic events occurring during the weld cycle it is important to deliberately force or seed the correct defect for detection. Out of all the deliberate defect tests it is thought the use of using different material inserts promotes cracks more readily than other techniques. The experimentation of the presented work will follow the ideas presented by [9]. Such tests will investigate alternative, but more compliant insert materials to closely obtain more expansive realistic data sets where both time and frequency domains will show distinguishing phenomena. By using advanced PAC wideband AE sensors it is possible to acquire a higher resolution recording, displaying the suitability for such technologies which was not displayed by [9]. Other authors also concluded that it was not possible to detect ductile fracture mechanisms as well as flaws that were stressed sufficiently to detect initial crack growth; such sensor enhancements should now ensure this is possible where both elastic and plastic material phenomena is identified during scratch tests [27]. The setup of using an AE contact sensor to distinguish crack initiation / propagations with such a setup using different insertable inserts, has not been seen before in literature.  

6. The state-of-the-art comparisons for the proposed work are missing in this paper. Then do a critical analysis of previous research. State explicitly the shortcomings of previous research. What is positive in previous research and what is negative. Based on that, you explicitly define the goal of the research and the scientific hypothesis.

The authors agree with these comments and the following section has been added to the discussion of results section:

Line 1070, page 35:

Research by authors Ser'eznov et al, (2009) who have used AE contact sensors were inferior to the AE contact sensors used here. Comparisons in the total count of AE signals for the same time of 350 seconds were 1400 counts, but those recorded for CMSX4 cracking evaluation produced 11000 counts.

It can be concluded from such results that the AE contact sensors used within this study are far more superior, providing greater resolution and sensitivity.  Therefore, it’s now possible to detect finer anomalies and the possible onset of cracks resulting from poor joining practices and preparation.

In terms of patterns however, both pieces of work display high amplitudes for the onset and continuation of crack activity. For crack detection, the authors’ were not aware at the time of completing these experiments that researchers have studied such phenomena using both contact and airborne AE sensors.

From work carried out by [20] Grad et al, (2003) an AE contact sensor was used as a microphone to monitor and understand the quality of gas-metal arc welds. Here the recording of the arc process provided a view to quality from the welding processIf however, an electromagnetic acoustic transducer (EMAT) AE sensor is used instead, then this is perhaps more appropriate to record airborne signatures, but it should be noted that it is still detecting elastic waves albeit via a non-contact method.

Finally, we use STFTs where we can such as when transforming a continuous waveform from the time-series to time-frequency domain. The reasons for using STFT in favour of other transform techniques is based on it being less computationally demanding as well as the amount of the information is not massive requiring a high performance computer. In addition, AE parameters are used to give a rich summary picture of the AE events and parameters such as rise-time, which has, not be used in literature before for this type of work. Therefore, with a robust picture of the phenomena taking place, there is a good confidence to verifying the findings within the paper. This is another novel aspect to work carried out here.   

The STFT analysis of the GRAS airborne acoustic signal is difficult to see the changing amplitude as all signals recorded for the baseline cases have a frequency coverage across the whole sensor bandwidth spectrum. In addition, to make easy comparisons, all the amplitude colour maps were normalised by the same amount. By providing such normalisation can restrict features from being more salient.

The Mistras contact AE sensor produces output parameters that are recorded based on passing a dB noise threshold which means the overall signal is not continuous with time however gives one snap shots of events and each one of those snap shots give small continuous waveforms which can have STFTs applied to it. For future work, a two channel system would be more beneficial where STFTs could be compared where both channels would give two fully continuous signals. In addition, phenomena location would also exist.  The work mentioned here looks at a single AE USB system which is more likely for a test engineer as it’s easier to setup in remote locations.  

7. The shortcoming of this research is that potential errors and measurement uncertainty were not analysed and discussed.

The authors dreaded this question as it is very involved but necessary! Joking aside, the authors totally agree with this comment and it should not be omitted and instead, addressed as much as possible. That said, this really needs to be thought out at the beginning of the measurement campaign to warrant more measurements per test case and other factors (measurement algorithms etc such as that for STFT) – this will be looked at in much greater depth for future work. The following was carried out to give the reader some idea of uncertainty with the current measurement campaign.

Line 909 Page 27:

Table 6 Average of Rise Time measurements for each insert state (contact AE sensor)

Table 7 Average of STFT measurements (airborne AE sensor)

Table 6 provides rise time measurement data from the AE contact sensor to give an idea of differences between each insert state. From the table, it is possible to see the standard deviation differences of the rise time measurements are all within a similar percentage giving some confidence of pattern trends. The max rise time shows higher differences for alloy 718 and CMSX4 which is expected considering the differences in chemistry where EN8 and BL are very similar to the plate chemistry and therefore the differences are less. Only the rise time parameter was discussed here as this was not seen in research before and showed good demarcation from other parameters. The data from Table 7 is based around the STFT energy utilization and again, the maximum bandwidth and high amplitude energy utilization is found with the inserts with the very different chemistries when compared with chemistries that are similar to the plate material. Such differences show a good case for using machine-learning techniques and providing automated classification/prediction where all the data would be used in sensor fusion approach. To apply full uncertainty considerations to measurements needs to be applied at the beginning of measurement campaign and this will be applied to future work.  

8. The obtained results could be additionally discussed and, if possible, compared with the results of previous research. (11)

This is a difficult task to carry out based on the current amounts of information in the arena. In addition, the technologies now are far superior than compared with the technologies used for previous works and are therefore a mismatch for comparison. Patterns and bandwidth frequency of anomalies found can however be compared. The following text has been added to a Discussion of Results section within the paper.

It is always important where possible to align current research with previous research to give more confidence with the obtained results. That said, with the amount of data available to date, this is very much a difficult task if not an impossible task. In addition, comparing early works with works carried out in this paper, the sensors used within our work are so much more sensitive to change as well as having much greater resolution and therefore, such comparisons are difficult to be made where higher quantities are achieved with the current setup. Nevertheless, the patterns and bandwidth frequency for anomalies do show good correlation and similar trends.  

9. In the conclusions, emphasize the scientific contribution, the innovativeness of the research, as well as the possibilities of practical application.

The authors agree and thank the reviewers for these comments. They have changed the conclusions as following:

Line 1119 page 34:

The use of materials offering a weldability gradient in the form of chemistries that vary significantly in chemistry, welded to an ASTM A36 steel with a fully austenitic welding wire (307Si) promotes a varied set of events through expanding the solidification profile.  Alloy 718 and CMSX-4 have proven to be ideal candidates in inducing favourable imperfections with such a combination of base metal and filler metal chemistries.  These materials promote the formation of topological close-packed phases e.g., the intermetallic and brittle Laves phase that accentuated an enhance crack sensitivity, which in turn improved the ability to test the potential use of remote NDT methods to determine a weld’s integrity.

Through using nickel-base alloy inserts, it has been possible to create additional complexities to promote geometrical changes to instigate gross imperfections in the form porosity.  This combination increased the sensitivity reminiscent of the circular patch weldability test developed by R.D. Stout and published within the welding research council in 1987, which was used to determine the weldability of steels.   This chemistry combination heightens microsegregation due to the rejection of solute elements that coalesce and promote low-temperature liquid films, which decorate grain boundaries and rupture.  Combine this with the change in metal flow, resulting from differential rheological properties and thermal expansion values, further induced microscopical stress regions at the dendritic and grain boundary scale that grew into macro-stressed regions, which resulted in cracking.  This is the result experienced with the ASTM A36 and 307Si + alloy 718 and ASTM A36 and 307Si + CMSX-4 combinations. Such multipart solutions were proven to increase the propensity for hot cracking and cavities, that release sonic waves at specific frequencies. 

It was concluded that there could be a strong correlation between ultrasound omitted during MAG welding and corresponding joint integrity in terms of crack response that could be of significant value to fabricators. Certain frequency ranges and amplitudes are potentially unique for different material chemistries exposed to welding. The frequency rise during and immediately after the process denotes the formation of imperfections, phase transformation or defects in the weld. AE sensors, both airborne and contact can identify such frequencies and amplitudes, which guides automation in terms of acceptable or non-acceptable weld integrity.  AE sensors can be integrated into manufacturing systems for real-time inspection rather than having to undertake timely, expensive, and destructive material analysis to qualify the weld structure.

It was noticed that STFT plots of the AE airborne microphone gave much higher intensities and frequency band utilisation for regions consistent with cracking. The same was observed for the contact sensors where the frequency range was between 250 - 550 kHz with high intensities of 70 dB. Such values were significant indicators of cracking during the cooling phase, circa 80 seconds after welding. This shows that with using high-resolution instrumentation and advanced Digital Signal Processing (DSP) techniques it is possible to obtain useful in-situ and post-machining information to discern cracks and thus the general weld quality. 

For this work to be accepted into the general fabrication mainstream further material analysis with AE sensors is required, which should aim to generalise all welding anomaly conditions in-situ. In addition to this, such anomalies are indicative to certain frequency ranges and amplitudes.  This information could, in future, be fed directly into filter-banks and real-time monitoring control systems as well as advanced material models and machine learning architecture. Further work will look into similar work using thermocouples or IR camera systems processed with Digital Image Correlation. More work will be designed around establishing the different associated welding anomalies and all data will be processed using established and novel machine learning techniques to show the accuracy and acceptance of automated signature inspection.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Accept.

Reviewer 2 Report

The authors have carefully revised the paper and adequately responded to the comments. The manuscript could be accepted.

Reviewer 3 Report

The manuscript has been significantly improved and corrected. I recommend accepting the manuscript in its current form.