Crack Detection in Welding Process using Acoustic Emission

(1)

Abstract: The process of inspecting welds done in production at Volvo Aero in Trollhättan is time consuming and a lot of this time goes into examining faulty objects. The aim of this thesis is to start development of a system that analyses acoustic emission from cooling welds to determine the quality of the weld. Our aim is to be able to detect cracks in the material and to give information on the cracks using the data gathered by our sensors. To do this we will use methods to locate sound sources and then rate our findings and do some simplifications on the result of our calculations. We will analyze our calculated data to find crack signatures and classify our findings and give alarms if we find cracks that are considered too big for comfort. We will also give insight in to future aspects of our work and look at ways to improve on our proposed methods. We will discuss our systems pros, cons and what things have been taken into consideration during design, and what strategies we propose to handle the results from the system.

Crack Detection in Welding Process

using Acoustic Emission.

Linus Karlsson Lkn05007 Mälardalen University Supervisor: Erik Olsson Contact at Volvo: Patrik Boart Examinator: Peter Funk

(2)

Introduction

In this thesis work we will consider the possibility to detect cracks caused by welding using acoustic emission. To do this we will focus on sounds generated during the time after the welding is done when the material hardens, this is done during the cooling of the material and the duration of the recording will be 5 seconds or more, tough the data provided by Volvo and used for this thesis work covered somewhere between 5 and 15 seconds.

We will be studying one scenario where we have one microphone and some different situations using multiple microphones. Our hope is to be able to determine whether cracks can be found with use of one microphone and determine the reliability of such a system. When using multiple

microphones we will focus on locating sound events that might indicate cracks and also to gather other usable data by use of acoustic emission.

We will examine data, provided by technicians at Volvo Aero, which has been recorded on actual welds using one microphone that was mounted on the piece to be welded. Our purpose is to use this data to determine whether a single microphone is enough to detect cracks in the material. Examining both cracked and non cracked welds we will try to find any information about the quality of the weld. Using two or more microphones the acoustic data obtained will be analyzed and our goal is to locate and pinpoint the sources of individual sounds and then to cluster sounds of similar types from the same region. We will then analyze this data, searching for areas where sounds indicating a crack are frequent. To accomplish this we have designed some different approaches and algorithm going in to different aspects and depths in the problem, trying to find an optimal balance between needed resources, computing time and accuracy. The problem, proposed solutions, consideration and future work will be discussed. To accomplish this we will use at least two microphones set up as pairs of microphones so at least one pair will be used. We will use triangulation between microphones to deduct the source of the sound in question. We will also examine different sampling rate to see where the optimum frequency lies in terms of accuracy versus data management.

The first section will deal with related work and some insight in to why this project came about at Volvo in Trollhättan.

The second section will contain an introduction to the Problem of crack detection, sound data handling and introduce some difficulties with it.

The Section that follows this is where we look at the data from Volvo and report on our work using single microphones. In this section we try to find things that indicate whether crack detection can be done with a single microphone. This section contains two major parts, one dedicated to each data set provided by Volvo.

The section on multiple microphones contains both a description on problems we face when trying to use more than one microphone as well a section on sound preprocessing. This section also contains our approach when trying to use multiple microphones to detect cracks.

In the last chapter we will discuss our findings, our results and what this tells us about using this approach when dealing with the problem off crack detection. We will also contemplate on how to improve accuracy and limit the need for computational power. Lastly we will have a summary of the entire paper which will also hold our conclusions.

(5)

Background and related work

Background

The idea for this project was thought up at Volvo Aero in Trollhättan, Sweden. In their process of creating airplane parts and related products they do a lot off welding and wanted to improve the performance of their welding operation, more precisely they wanted to improve their quality

assurance controls and the methods used to validate that any piece of equipment that was delivered to a customer was according to the requirement specifications.

Their current method of control is time consuming and relies on experts finding faults in the finished products and as time is money it is in any company’s best interests to limit the time spent on all processes within its production work and, by doing so, lower their expenses.

Today the method used by Volvo Aero to determine the quality of their work and more precisely to assure that each finished product is as close to perfect as possible and that it will function at the level specified in the product specification. Today this is done by sending each finished peace to control department where the piece is X-rayed and then experts examine the results of these tests to determine whether a weld is ok or not. As we said this takes time, and since the experts must look at every finished product, even the ones that are not up to specification, a lot of time goes into

examining pieces that later goes straight into the trash bin.

The test done by these experts gives to important pieces of information. It tells us whether a weld is good enough or if it should be discarded, secondly it tells us where in the piece there are any faults and how severe these faults are. This is of course very helpful since it can tell us things about the welding process and also about the work of individual welder. This information will let us see reoccurring faults and help us to alter our processes to avoid these faults in the future. However Volvo wants a system of control that gives as close to the same info as possibly but uses a lot less time, the main focus of the system however will be to discard faulty pieces before they take up expensive expert time.

Volvos idea was to examine the quality of welds using acoustic emission and use some method to detect cracks and also to get as much information about these cracks as possible. One of the main points of using acoustic emission is that it is a form of non-destructive testing same as X-raying and therefore it can be used directly on products that are to be delivered rather than samples from the production line. The main focus of the project and the most important goal is to determine whether cracks have began growth in the object and if so, are these cracks severe enough to cause a major difference in performance. If this method can find cracks that are off the dangerous kind they can be weeded out before going in to X-ray and this is of course the optimal scenario. Another application is that of finding cracks that are minor or in areas where they have less impact, the pieces indicating these cracks will still need X-raying but since the analysis of acoustic emission has already localized an area with suspect acoustic activity the experts will now where to begin their examination in order to find existing severe faults earlier in the examination process.

When Volvo had gotten this far in their planning of the system and the considerations to be made they contacted Mälardalen University and introduced their idea to Erik Olsson and Peter Funk who took an initial look at the project and decided that it would make a good master thesis work for students at advanced level spanning over 20 weeks. The thesis work was to be done by one student

(6)

depending on the people at Volvo for expert knowledge, measurement data and other forms of necessary assistance.

The first focus of the thesis work was to do research in the field of acoustic emission, analysis of sound data, source localization and crack appearance in welds. This research work was ours main focus for roughly three weeks and a summary of the earlier work and related subjects will now follow.

Related work

A lot of research work has been done in the field of analyzing of acoustic emission and we have limited our research to papers concerning either acoustic emission used as a non destructive testing method, acoustic emission in the welding process or work on multiple sound source localization. We have divided this section in two three parts each concerning one aspect of our research.

We began our study of related work by looking at the field of acoustic emission in general. We read about the theory of using acoustic emission as a method for non destructive testing described by Baifeng JI and Weilian QU[1]. We then went on to read about using acoustic emission to measure crack growth caused by welding and stress corrosion in an article by C.E. Hartbower, W.G. Reuter, C.F. Morais and P.P. Crimmins[2].

We found an article describing how acoustic emission was used to find cracks in welds during their cool down time. This method described by A. N. Ser’eznov, L. N. Stepanova, E. Yu. Lebedev, S. I. Kabanov, V. N. Chaplygin, S. A. Laznenko, K. V. Kanifadin, and I. S. Ramazanov used multiple sound channels and clustering to find events and also gave the location of found events [3].

When we realized the potential of using multiple channels and localization we researched basic methiods for sound source localization and first read about the music algorithm and clustering in an article by E. D. Di Claudio, R. Parisi and G. Orlandi[4]. We then went on to examine methods of using time delays between microphones and difference in amplitudes to locate sound sources and found two interesting articles. One by Brent C. Kirkwood[6] and one by Ming Jaan, Alex C. Kot and Meng H. Er[5].

(7)

Problems and goals

In this section we will take a close look at the basic problem and a lot of the things that we have taken in to consideration. We will examine our different approaches and why they were chosen and also why some things where not taken into account when facing the task. We will start by addressing some main points of the problem and then go into more detail on the points that will be handled in more depth throughout the thesis.

The basic problem we face is to use some sort of acoustic analyze to discern a crack created during a welding process and to locate it if that is possible. This consists both of localizing the sound source and to determine how to determine whether what we see is a crack or not. Another aspect of the problem is to do this with small resources and without taking up large amounts of time. Tradeoffs in accuracy and robustness might be preferable to make the system light, fast and as easy to handle as possible. To accommodate this we have chosen simple and straight forward methods throughout the thesis work to keep the performance requirements low.

The situation that we face is a weld that occurs and cools off in a given area. We will be able to mount our microphones the way we see fit in and around that area, and given that, we can basically choose any microphone array setup. However there is much to gain in having a microphone array that takes up little space and the optimal solution stated by Volvo is for the final crack detector to be as close to fully portable as possible.

Because of the above mentioned problems we also encounter another problem when it comes to detection and localization. The task of finding a crack and its location might be possible by use of two different approaches. We either, first find a crack signature in the sound data and then pin points its location or we map all sounds and sources and the search for cracks.

One of the main goals with our thesis work is to determine whether the single microphone approach is applicable on the problem as specified by Volvo. To determine this we will examine data from both non cracked and cracked welds and try to determine if there are any indicators for cracks or the absence of cracks that can be found with relatively low computational power.

If, after our analysis we find that we have difficulties determining whether a weld is cracked or not from a single microphone we will start work on simple algorithms for sound source localization using simplified methods that will work on the given problem. We will also discuss clustering and sound data handling for the multiple microphone approach.

If these two methods are usable then they each have their own pros and cons, the problem with the first approach is that of miss localization which will lead to insecurities when giving the position for found cracks, and also the problem with determining the size of the crack and if there are multiple signs of crack growth in the object.

The major problem with the second approach is data handling and computer power, though this does not need to become a major problem it will still use up more power than the first approach. The good thing with the first approach is that it gives us fast and easy indication of cracks and that it requires very little computer power to do so. The major advantage with the second approach is the amount of information it could give us, the accuracy and also the fact that we can alter this approach in many beneficial ways, improving our different criteria and methods as we go along. The second approach also offers the ability to use different clustering strategies that allow us to rate different parts of the welding area according to their level of sound generating activity.

(8)

Apart from the task specific obstacles we have discussed so far we also face problems related to recording acoustic emission such as background noise, random noise, microphone precision and also delays caused by other factors than the distance to the source of a sound. As for background noise the normal strategy is to choose threshold values to sort out noise that do not meet required levels of amplitude. When it comes to random noise the problem is a little trickier, since a random noise could have the same signature as any other type of event. Microphone precision is a big factor when deciding the maximum resolution of the system, since we will be relying on differences in amplitude and arrival times between microphones to deduct the source location we can only have as fine a precision as our microphones allow. When considering delays from unknown factors we have two cases one is a constant delay which can easily be taken care of, therefore it is not a big issue, and the other is a random delay which might greatly hinder our work. Preferably we must remove all random delays in the system and this will have to be taken care of before calculations take place.

(9)

Working with the data from Volvo

When using one single microphone we will focus on analyzing sound data based on frequency and by observing the pressure curves generated by the microphones. The major problem here is to

determine what kind of sounds a crack will result in and if these can be perceived and singled out by sound data manipulating and in the long run an autonomic approach.

In this section we use authentic weld data provided by Volvo and our goal is to determine whether a crack can be found using only one microphone mounted on the piece to be welded. This means that we will analyze data from cracked and non cracked welds and compare the results to see if we find any key differences that provide information on crack appearances and growth.

Data was provided by Volvo in two separate sets. The second data set was recorded after our initial analyzes of the first set and therefore the measurement parameters and experimental approach was altered for the second recording.

The first data set

The first data set we got from Volvo contained a total of nine recordings with eight being welds and seven of these included some cool down time. In our analysis our focus will be on the seven data sets that included cool down time.

Title Sensitivity Length in seconds Comments

First cracked weld 50mV/div 13:33 With another setup than the other welds

Second cracked weld 10mV/div 13:39 Third cracked weld 5mV/div 13:48 Fourth cracked weld 5mV/div 14:02

Fifth cracked weld 5mV/div 14:07 With air mounted microphone Non-cracked weld 5mV/div 13:45

Table 1, here we see information about the measurement data provided by Volvo.

This data was recorded on welds made on a test piece set up with a sound sensing device mounted on the piece. All welds where done in a small area of the piece and should therefore be comparable when analyzing sound pressure amplitudes.

(10)

Figure 1, the setup used by the people at Volvo to record data set 1.

In figure 1 we see the setup that was used at Volvo to make the recordings. At the point marked A in the top left half of the figure we see the microphone mounted on the piece. At point B in the right part of the figure is where the welding took place. All welds where made in the polished area and where made from top to bottom given the perspective of the figure.

(11)

Figure 2, the welds after they have had time to cool down.

Figure 2 shows us the welds made in area B from a perpendicular angle. The two welds placed to the right are the welds that are not cracked. These welds are somewhat closer to the microphone tough we hope that this will not affect our measurements too much. One of these non cracked welds where made without recording cool down time and has therefore been omitted from analysis. When considering the non cracked weld there is however one thing that we must take in

consideration and that is that the welding apparatus was no turned off directly as was done with all other welds but rather it was toned out over a period of about ten seconds. The impact of this on our data is not fully determined, this happened due to a communications problem between us and the testing team, since our focus lies in the seconds after welding apparatus is off and technicians had a higher focus on time when it was on. We hope that this does not make our data invalid, and we will take steps to avoid this in future measurements.

After the first cracked weld was completed the piece that was welded on and also the sensor was rearranged. Also, as we can see, throughout the testing the sensitivity of the equipment was altered with the first and second cracked weld having different sensitivities from the other four welds. In all the recordings the initial seconds consist of recording while the welder is on, it is then turned off and we have data on the initial seconds of the material cooling down. We will show our work on

examining the welds using the third cracked weld and the cracked weld, beginning with the non-cracked weld.

(12)

Examining the non-cracked weld

The first step we took when examining the sound data was naturally to look at a plot of the sound pressure over time.

Figure 3, the non cracked weld shown as sound pressure over time.

Above in figure 3 we see the data plot and we can notice the change around index 7 on the X-axis, this is where the welding apparatus was turned off. Since we will want to take closer look on the part of the data that was gathered after the weld was turned off we cut that data out and worked on it exclusively. This leaves us with a new array consisting of indexes 7000000 to 10000000 of the original data array.

(13)

In figure 4 we first noticed five clear and visible audio spikes in the later part of the sound and also a higher level of noise in the earlier parts of the recording. Our first step in trying to get more

information from the data was to run it through a FFT analysis, which will let us see how the

amplitudes of spikes and noise are divided over frequency. First we simply looked at all the data over all our frequencies.

Figure 5, the first FFT plot. Sound amplitude, shown over time and frequency

In this figure the right-hand axis indexed from 0 to 300 shows frequency with our full range of 200 KHz split into 256 equal parts indexed from lowest to highest. Sharing origo with this axis is the time axis indexed from 0 to 6000 where each million indexes in our sound data is represented as 2000 points in this graph. The last axis indexed from 0 to 0.01 is the amplitude axis showing the amplitude of the sounds we are seeing placed at their correct time and frequency.

When studying this graph we find some interesting points to note, firsthand we see that almost all of the noise occurs at the low end of the frequency register and therefore should be fairly easy to remove or at least to omit from our analytic methods. We can also note that the five spikes seen in the first data plot are not shown in this graph, which tells us that they are of a higher frequency than 200 KHz.

(14)

Figure 6, the second FFT plot. The data with the lowest frequency omitted.

We removed the lowest frequencies and redid the plot and ended up with the graph seen in figure 6. We found a second smaller section of noise placed just above the first one in frequency. Now that the resolution on the amplitude axis is a lot higher we could also see some low amplitude spikes in the higher ends of the frequency spectrum.

Figure 7, the third FFT plot. This graph shows data from the 246 highest groups of frequencies.

We redid the FFT plot omitting enough of the low frequencies to filter away the noise and the results can be seen in figure 7. To arrive at this plot we removed 10 of the 256 groups of frequencies thus omitting roughly four percent of the sound data. If we cross reference this data plot to the amplitude plot of the whole data in figure 4 we can see that the spikes seen in the higher frequency spectrum of figure 7 occur at the same time as the spikes seen in figure 4.

(15)

Studying the spike in figure 7 we see that they grow in amplitude towards the higher end of the frequency spectrum and that they do so continuously suggesting that it is not a result of filtering where we would have seen a much sharper start to our spikes. The spikes observed in this graph also differ from the ones observed in figure 8 when it comes to amplitude and are a lot smaller than those observed in the other plot. To better determine the characteristics of the spikes, and also the spikes amplitude contra noise amplitude, we singled out the last and biggest spike and redid the initial plot and the FFT yet again and this time plotting indexes 9850000 to 9860000.

Figure 8, indexes 9850000 to 0860000 of the data shown as amplitude over time.

(16)

In figures 8 and 9 we see a smaller portion of the data represented in three different ways. First in figure 8 we see the data one again as sound pressure amplitude over time, in this graph we can see the spikes from figure 8 but this time the resolution of the axis has changed and we also see that the start of the spike is really fast whilst it ends by fading out.

Studying figure 9 we see the spike and how it is fading out and how that spreads over the frequency slots. We that the highest amplitude in the spike is found at the high end of the observed frequency slots, this leads us to assume that an even higher peak can be found at higher frequencies. Since we still cannot see peaks to match those of figure 4 we will have to keep searching or do other

measurements.

In the close up in figure 9 we can also see some effects from the noise climbing its way up in

frequency. The affect is fading and it is fixed in frequency which will make it easier to suppress if that turns out to be necessary.

Assuming that the peaks fades away in a similar fashion to that seen in figure 4 we will not examine that in the FFT graph. We will instead study how the spike grows over frequency slots. To better be able to do this we will first isolate the highest of the peaks, showing only that peaks behavior over the frequency slots. Focusing on indexes 9852500 to 9853524, frequency sots above 150 and examining the peaks as amplitude over frequency we arrive at figure 10.

(17)

From the way the observed amplitudes in figure 10 behave we assume that the biggest amplitude can be found at higher frequencies. This leads us to speculate whether the amplitude spikes seen in figure 4 are located at an even higher frequency than the ones covered by our measurements. From the figure we also see that the relationship between frequency and amplitude is varying and it seems that there are peaks at higher and higher frequencies but also smaller peaks at frequencies in

between. From the figure it is not possible to approximate a frequency at which we will find our sought after amplitude peaks. We do believe however since we see amplitudes of varying height over the frequencies, but still all growing with the frequency that, there might be waves carried by each other. To determine if this is the case we made a high-resolution close up of the peak as amplitude over time, similar to figure 4.

Figure 11, a close up of the spike seen as sound pressure amplitude over time.

From figure 11 we can clearly determine that there are multiple frequencies collaborating to produce the spikes we have observed with only single peaks being responsible for the really high amplitudes. This explains why they were omitted from the FTT which only focuses on the frequencies which can be fully covered by a sample rate of 500 KSS, those frequencies that lie below 250 KHz.

Examining the third cracked weld

As with the non cracked weld our first task when examining the recording of the cracked weld was to look at the sound data as sound pressure amplitude over time.

(18)

Figure 12, Sound data from the cracked weld shown as pressure over time.

The graph in figure 12 shows us that there are not as many spikes in this data as in the data from the non cracked weld. There are some clear spikes around index 5000000 in the graph which we hope will give us some information. There is also some activity around index 5000000 and 8000000 in the graph which we want to take a closer look at. We will begin with examining the relation between spikes and background noises through a FFT focused on the part of the recording that contains the spike and the other sound activity, indexes 4500000 to 8000000, showing all frequencies.

Figure 13, amplitude shown over time and frequency for indexes 4500000 to 8000000.

We see that when looking at frequencies which we can fully cover given our sampling rate the spike is not to be seen against the noise. We will omit the lowest frequencies to get a view of the sound activity as amplitude plotted over time and frequency.

(19)

Figure 14, showing a closer look at the activity found between indexes 4500000 and 8000000. By studying figure 14 we see that the activity can be described as two clear spikes beginning somewhere in the top 50 frequency slots and also noise and seemingly random activity at roughly the same frequency. The activity that we observe is of rather low amplitude and does not account for the spikes seen in the plot off sound pressure amplitude over time. This tells us that the spikes appear at a higher frequency and the same goes for the noise seen in the figure.

We will now single out the highest of the spikes observed in the sound pressure over time graph to examine it in more detail, focusing on indexes 5105000 to 5115000.

Figure 15, indexes 5105000 to 5115000 of the cracked weld shown as sound pressure amplitude over time. We see in figure 15 that the event begins sharply and then fades out which I consistent with a spike and we also see things to indicate that it is an event that acts throughout different frequencies. We

(20)

will examine it in more detail with the use of yet another FFT focused on the same indexes as figure 15.

Figure 16, the spike between indexes 5105000 and 5115000 shown as amplitude over time and frequency. Figure 16 shows us how the spike behaves through the various frequency slots. We see that it grows in amplitude towards the higher end of the frequency slots and also that that the amplitudes fade independently throughout the frequency slots. We will conclude our study off the data from the cracked weld by examining how steeply the amplitude grow over frequency, this is done, as in the non cracked case, by narrowing down our view on the spike itself and examine only the two indexes that make up the spike. We focus our last modified FFT on the cracked data on indexes 5108400 to 5109424.

(21)

Figure 17, Shows how the amplitude grows with frequency in the spike between indexes 5108400 and 5109424. In figure 17 we see that there can be many estimates for the growth of amplitude over frequency and that it is hard to make a clear calculation of the frequency at which we might find the high amplitudes found in the figures showing the data as plain pressure amplitude over time. Comparing the two welds

In this section we will make a comparison between our two recorded sound data and try to reason our way around similarities and differences in the two sound data recordings. We will begin by simply showing the data as sound pressure amplitude over time.

Figure 18, a comparison between the two welding data. The non cracked weld is seen in its entirety to the right and the cracked weld is shown to the left.

Observing figure 18 we see a lot off differences between the two sets of sound event data. The first and most obvious difference is the difference in focus between the two recordings. Both are of similar length but in the case of the non cracked weld recording roughly 70% is recorded whilst the welding apparatus is still turned on compared to around 35% in the case of the cracked weld. The transmission from the part where the welding apparatus is on to the part where it is turned off also looks a lot different between the two recordings. In the data from the non cracked weld we see

(22)

what looks like an immediate turning of off the apparatus, the transition from on to off is done in one single step and in a very small timeframe. Whereas in the case of the cracked weld the apparatus is turned off successively through several steps and also the turning of takes up more time than in the recording from the non cracked weld. It is not clear if and how this affects our data but the issue has been noted and will be taken into account when conducting the recordings for our second data set. Another thing that we can see clearly from the figure is the difference in behaviors of the spikes between the two recordings. There are clear spikes in both cases but there are big differences between them both in number, frequency and amplitude. (Notice that frequency in this sense corresponds to the number of spikes and their relative timing rather than the frequency of the sounds that generate the spikes). In the recording from the non cracked weld we see a number of high to medium amplitude spikes with short time in between beginning roughly two seconds after the apparatus is turned off. In the recording from the cracked weld we see one medium amplitude spike and some small spikes spaced out throughout the recording with the highest spike at about 3 seconds after the apparatus is turned off.

In both recordings we see some noises that differ from the base level of noise though it appears at different timing in the recordings and also varies in length between them.

These differences are consistent throughout all the data with all recordings from cracked welds showing medium to low amplitude spikes and not as many occurrences as in the case of the non cracked weld.

Note worthy similarities between our recordings on this stage is that the sound pressure amplitude recorded whilst the apparatus is on and also the noise when it has been turned off is at close to the same level and also that it acts the same which tells us that the experiments have been conducted with roughly the same environmental parameters.

Figure 19, Showing the two recording as amplitude over time and frequency.

Taking a closer look on the data as amplitude over time and frequency as shown in figure 19 we can also note and discuss some differences and similarities. The first and most obvious difference is that of amplitude between the two recordings. As when the data is shown as sound pressure over time the amplitudes observed in the data from the non cracked weld have much higher amplitude than the data taken from the recording of the cracked weld. This affects the resolution and reference values of the axis set on the plot by Matlab and gives the impression that the cracked weld contains a

(23)

lot more noise spread out over the frequencies. This however in not true it is merely the higher resolution in the graph showing that data that gives the illusion of a higher noise.

We note that the noise seen in the pressure over time graphs described as deviating from the background noises can be observed at the high end of the frequency slot in both graphs and is visible in the non cracked weld data even though the resolution is lower than in the cracked weld data. This noise is seen prior to the spikes in the data from the non cracked weld and directly after the spikes in the data recorded on the cracked weld.

Furthermore we can also see that the spikes, apart from the difference in amplitude, are similar in both data sets. Particularly their growth and the behavior of their respective amplitude as the frequency increases are very similar. We will take a closer look at the spikes to conclude our comparison of the two data sets.

Figure 20, a close up FFT made on the largest spikes from both data sets, the one from the non cracked weld is shown to the left and the data from the cracked weld is to the right.

Examining the two graphs in figure 20 we see a lot of similarities between the two spikes. We see that their behavior regarding both time and frequency is very much alike and this gives strength to our theories that claim that the events in themselves are similar. The only significant difference is yet again the amplitude and a small variation to the fading out of the sound pressure at high frequencies, though this can be explained by the difference in resolution.

Conclusion

We conclude that, as far as we can see in this data, the same type off sound generating events occur in both cracked and non cracked welds and that we will have to achieve measurements that give us a larger number of spikes to work with from both kinds of welds, preferably the new data will have the same consistency in amplitude as we have so far observed in behavior.

We will also want new data with a higher sampling rate so that we can observe the whole sound events. Preferably we want to include frequencies where the spikes are no longer active so that we can watch their behaviors over all the frequencies that they inhabit.

So far we see no difference between cracked and non cracked welds and we lean towards the idea that we need to categorize number of sound events and use some sort of localization techniques to determine whether cracks are present or not, work on the second data set will help to confirm or deny this hypothesis.

(24)

The Second Data Set

The second data set we received from Volvo contained a total of six recordings and five of these where recorded during welding. A similar equipment setup as for the first data set was used apart from the recording in itself which was done with other parameters.

For this data set they used a sample rate of two million samples per second which is four times the sampling rate used for the first dataset. All recordings where done with the same microphone sensitivity and again at roughly the same distance. The sensitivity was set to 20 mV per div.

One of the welds in the set was made with a long shutdown time for the welding apparatus and all the other used a short shutdown time. Some differences occurs even among the welds said to have the same shutdown sequence and this is yet to be explained by Volvo, since this occurs at the beginning of the recording we have not given it to much weight with our focus still lying on the later part of the recording.

Title Sensitivity Length in seconds Comments

Non cracked weld 20mV/div 13:29 Longer turning off of welding apparatus First cracked weld 20mV/div 13:33

No visible cracks 20mV/div 13:36 Second cracked weld 20mV/div 14:41 Third cracked weld 20mV/div 14:44 Fourth cracked weld 20mV/div 13:48

Table 2, this table shows the welds from data set 2.

At first glance we cannot see any clear spikes in the data from the two welds without visible cracks but still see clear spikes in the cracked welds. Since all the cracked welds are very similar to one another we will use one of these to represent them all and we will also examine the non cracked weld and the weld without visible cracks, beginning with the non cracked weld. For all welds we divided the plotting of the data into two parts, one showing the time when the welding apparatus was on and the turning of off the apparatus, the other parts shows the interesting data gathered from the initial cooling of the material.

Non cracked weld

We began by examining the data that was recorded during the non cracked weld. This weld was done with a longer tone out of the welding apparatus. We plotted the data in two parts with the first containing the turning of off the apparatus and the other focusing on the interesting part where the material starts to cool down. The two parts of the data is shown with different values for the Y-axis to get a better understanding of both the noise produced by the welding apparatus and the sound data recorded during cooling.

(25)

Figure 21, the data from when the welding apparatus was on and the turning of off the apparatus.

Figure 22, this data follows after the last figure and shows what was recorded during the initial cool down. The first thing to notice is the lack of visible spikes, which were seen in all data from data set 1. Failure to notice any spikes in this non cracked weld might give us useful information. We do however see some noise in the beginning of the second half of the data as well as in the end of the first part. To get more information from this data we will use an FFT focus on the ending of the observed noise, indexes 2 to 6 million, without the first 50 frequency slots, in figure 19.

(26)

Figure 23, an FFT of indexes 2 to 6 million in the second half of the data.

From figure 23 we learn that there are no clear spikes in the figure however the noise observed earlier is visible in the beginning of the frequency field of approximately 1 MHz, with the first 20% omitted to remove background noise that puts the sounds seen in the picture at around 200 KHz. In the figure we can also observe some sound activity higher up in the frequency field, at around 600 KHz we see something similar to the bigger sound events but at much lower amplitude. This activity is at a high enough frequency to have been missed last time and we will search for similar

phenomena in all our data from the second data set.

The lack of visible spikes in this data and also the noise seen and its amplitude and behavior will be noted as we go on to examination of the other two data sets chosen for inspection.

Weld without visible cracks

We had the same approach with the weld that had no visible cracks; we first looked at the two halves of the recording and tried to find points of interest to investigate further, once again they are shown with different Y-axis for a clearer view and a better understanding.

(27)

Figure 24, the first part of the data from the weld with no visible cracks.

Figure 25, the second part of the weld with no visible cracks.

There are few distinguishing features in both halves of the recording though some noises and one small spike can be seen in the second half of the data. We will focus on the part with the spike in the second half and also the beginning of the noise seen in the beginning of the second half. First we will look at the beginning of the noise focusing on the data between index 8000000 in the first halt and 2000000 in the second half, shown as a FFT.

(28)

Figure 26, a FFT showing the beginning of the noise observed in figures 24 and 25.

We can see in figure 26 that the noise is taking place in the frequencies where we have earlier observed spikes and noises that were not deemed a part of the background noise. We can also see that the nose gain in amplitude at the same point in time as in figure 25. We also note that the rise in amplitude can be observed on the high amplitude reflection as well and that it seems to be of around the same magnitude as the rise in the low frequency data.

We will not look at the only resemblance of spike that we have observed in the data from this weld, found around index 8 million in the second half of the data. The spike can be observed at this index in figure 25 and we will now show it in a FFT showing indexes 7000000 to 9000000 of the second half of the data.

(29)

Figure 27, an FFT of the small spike seen in figure 25.

We can clearly see the spike emerging from the noise and roughly at the same frequency. We can also notice that the highest amplitude of the spike is located at a slightly higher frequency than the top amplitude of the noise. Also visible at a close up of the figure is a reflection of the spike in the reflection of the noise at the higher frequency of about 600 KHz.

This spike that was no seen in the data from the non cracked weld might give us information

regarding non seen cracks in the weld. However from the first data set we did not observe spikes to be exclusive to cracked welds.

We will however not the behavior of this crack and note at where in the data set it took place and move on to the final examination in the second data set.

Cracked weld

We choose to use the third cracked weld for our examination of the cracked welds in the second data set. At first glance we saw that it was only in the cracked welds that we observed clear spikes and we will try to see if we can match their behavior to the spikes earlier observed in data set 1.

(30)

Figure 28, the first half of the data from the cracked weld.

Figure 28 shows us the first half off the data from the cracked weld. We can note that the part that shows the welding apparatus being turned off has a different look than in the other recording from the same data set.

Figure 29, the second part of the recording from the cracked weld.

In this figure we see a lot of spikes and also a rise in noise level as observed in the earlier recordings. We will narrow down our examination on the last 2000000 indexes of the data set and focus on the rise in noise amplitude and the largest of the spikes. We will begin by examining it all with a FFT.

(31)

Figure 30, a FFT of indexes 18 million to 20 million of the cracked weld data.

As we look at figure 30 we see the spikes clearly at around 200-300 KHz which is where we hoped to find them, we also see some low amplitude activity around the same frequencies that grows towards the end of the recording. This matches the rise in noise amplitude that we observed in figure 29. If we look at the higher amplitudes we again see a reflection of the activity around 200-300 KHz appearing at lower amplitude and with a frequency of about 600-700 KHz.

We can also note that the spikes appear within the same areas where we also note a higher noise level. It might indicate that something is indeed taking place in the material during these times of higher noises and that the spikes give us information about this process.

We will take a closer look on the largest of the spikes to better study its behaviors over frequency and also its companion at higher frequency. First we will view it in a plain graph as amplitude over time.

(32)

Figure 31, the largest spike from the cracked weld shown as amplitude over time.

When we examine the spike more closely as in figure 31 we see that I fades out in stages, the fist max point stands alone but after it comes 5 points that reach about the same height and then I fades again to a number of points in the next category of amplitude. I we look at the lower spikes, the minimum points we also see roughly the same behavior, though the time of their lowest point does not match that of the top point.

Figure 32, a FFT of the largest spike in the data from the cracked weld.

Now that we can see the spike in more detail in figure 32 we can study its behavior more closely. We see that the event peaks at around 200 KHz which means that we were close to seeing the peak in the first data set. The over tone at 600KHZ is a reflection of the first sound and was of course omitted

(33)

earlier. Now that we have identified a so clear a spike in the data and also noted that it lies within a period of higher noises we are starting to assume that a correlation between spikes and a rise in the noise level, only measuring noise around 100-200 KHz might actually indicate an event taking place in the material.

Comparisons

The first thing we want to look at in our comparison is the part of the data where the welding apparatus was turned off. This is also a point on which the three cracked welds look alike and therefore we will again look at the first half of the three examined data sets.

Figure 33, the turning of off the apparatus as it was recorded in the three examined welds.

As we can see from the graphs in figure 33 the welding apparatus seems to have been turned off in a different manner between the cracked and non cracked welds. We also note that the noise produced by the welding apparatus looks a bit different in the data from the non cracked weld as compared to the two others; this might have something to do with the manner of turning off the apparatus. Volvo specified that a longer tone out was used for the non cracked weld and that must be what we see in the recording.

We will assume that the difference between the welds will not affect our work but we will still advise to remove these differences from future measurements.

We will also look at the rise in noises at around 150 KHz shown as pressure over time in the three welds. Since we have already concluded that the noise appears at those frequencies we will now just compare it between the welds.

Figure 34, the rise in noise shown in all three welds.

In figure 34 we see the data from all three welds beginning with the non cracked weld followed by the weld without visible cracks and the lastly data from the cracked weld.

(34)

In the first data, from the non cracked weld, we see the whole rise in amplitude come and go. We see that it spans over roughly 10 000 000 indexes which means roughly 5 seconds. We can also see that the activity is a smooth one without any strange occurrences apart from the rise in noise.

In the data without visible cracks the noise begins in the second part and we see 9-10 000 000 indexes off it; we do not see the end of the activity. During the observed part of the sound activity it looks smooth except for a few smaller spikes which could mean that something is happening inside the material for example the beginning of a small crack.

In the data from the cracked weld we see the last 6 000 000 indexes, about 3 seconds, and see a rise in the noise level even if it is not as distinct as in the other two recordings. We do however not that the spikes seen appear in conjunction with the rise in noise beginning roughly at the same time as we can note a difference in noise levels and then gaining in amplitude as the noise rises. This is

interesting and leads us to believe that activity in the material leads to height in noise level and that distinct events leads to different events that can be observed in the recordings.

Conclusion

In the second data set we saw no spikes in the non cracked weld but clear and obvious spike from the cracked weld data and also a high level of similarity between all cracked welds. When we looked at the data from the weld without visible cracks we see something that might indicate some minor activity taking place in the material. Based on the second data set alone we could reach the conclusion that cracks are indicated by spikes that appear around 200 KHz and are mirrored at roughly 600 KHz.

Result

After we have examined both data sets provided by Volvo we are a bit split as to how to read the results. On one hand we have the second data set and parts of the first data set that indicates that the spikes observed at around 200 KHz do indicate cracks or related phenomena in the material. On the other hand we have the non cracked welds in the first data set and also some of the cracked welds that contradict this conclusion.

Of course some of these might be problems with equipment or other noises that destroy the reliability of our data but this cannot be decided upon with only the given data.

What we can decide on however is that sound data gathered in this way has a high potential of being valuable and that we should continue with recording using the same sampling frequency as in the second data set. We can also note that we want to strive to eliminate any and all differences in procedure between measurements on cracked and non cracked welds so that future data will be even more reliable.

Conclusion

From the observed data it is hard to draw a good conclusion. When we consider the data from the first set we can observe that the spikes in the non cracked weld, in figure 3, that are of the highest amplitude observed for spikes do not follow the behavior suggested by other data in that they do not occur in conjunction with the noise generating event. The difference in amplitude relative other sounds in that data set also stands alone from other data sets and this leads to suspicion of this data being off rather than other observations. Since the observed spikes are so similar to the spikes vied in

(35)

other data this might be because other welds are cooling off in the vicinity of the non cracked weld or just a result of the difference in equipment sensitivity.

The conclusion we reach is that we spikes that lie within the field of heightened noise with both spikes and noise have frequencies at around 200 KHz might very well indicate crack growth in the material. The next step in measurements will be to conduct larger scale experiments with more cases in each one and also to keep allowing time in between welds for the material too properly cool off and therefore avoid any eventual interference.

(36)

Proposing solutions for localization and analysis

When considering solutions to our problem we have taken a lot of things into consideration and we have looked at a lot off possible answers. We have compared many different ideas in theory and come up with a few different approaches to go on and test in real life situations. After our own initial tests, which will be done on mundane data with acoustic data that lies within the hearable spectrum, we will go on to apply the real test data on the proposed solutions that produce the results we are looking for.

In this section we will focus on our work on design solutions for the microphone array design, the algorithms used for sound localization, sound data handling and also on our work on recognizing cracks in the output from the first two algorithms. The work put into these three areas will be presented in that order, first array design then source localization followed by data handling and lastly analyzing the data to find cracks. We will begin by taking a closer look at the some of the problems we face.

When doing this thesis work we had some problems with measurements at Volvo and therefore we were unable to obtain data recorded with multiple microphones, we also had problems with determining what kind of equipment would be used. This leads to cut downs in some of the work in this theoretical section. Mostly in the section of clustering and data handling for Volvo data, also we have been unable to determine what filters and sound data algorithms we need to apply to the data before applying our own methods for sound preprocessing.

The problems we face when using multiple microphones

One main problem when using multiple microphones to pinpoint a sound event source is that we must first match sound events between our microphones. Matching sounds is a task that can run in to a lot of problems, such as sound events taking place at times close to each other or sounds that are low enough in amplitude to be represented very differently in the two microphones.

When we examine the acoustic patterns that are emitted from the cooling weld we will discover sounds varying in frequency and this might help us determine the nature of the events that

generated the sound that we are looking at. This fact will also be useful in figuring out which sound event corresponds to which between the different microphones. A problem we face here is that sounds can arrive in different sequences at different microphones and cause confusion as to which measured event corresponds to a given sound generating event. This can be helped by avoiding certain microphone array designs and also by examining the frequencies of measured sounds to better identify its source. Some risk of this phenomena occurring will still be in the system but it will be much reduced and we will implement some different tactics to handle it.

As we have just discussed, problems in array design for the placement of the microphones in conjunction with an inability to discern different sound sources given only the data in one

microphone can lead to problems in source localization. A problem is that two sounds generated at roughly the same time might arrive at the different microphone in another order than the order of appearance. This will lead to our algorithms, using the wrong data, producing false positions for both the sources.

(37)

Figure 35: This figure is used to illustrate the problem sound switching order between microphones In figure 35 we see a setup that uses two microphones set at opposing sides of the welding area, marked as grey in the figure. The two microphones M1 and M2 Will both be recording all sound generated in the target area. For the case of our example we assume that two similar sounds arise at points P1 and P2 at about the same time. This will be measured in both microphones as soon as the sound waves arrive. Given that both sounds are closer two one of the microphones than the other this will lead to the sounds from a Point arriving at different times in the two microphones. We will call the sound generated by events P1 S1 and sounds from P2 S2.

Figure 9: This figure will help us see how the sounds generated at P1 and P2 are represented in the measurement data

from the microphones M1 and M2.

In figure 36 we see the result of measurements for both our microphones M1 and M2 presented in the form off a wave with amplitude on the Y-axis and the time of arrival at the X-axis. As we can see in the figure a situation has occurred where the two sounds has arrived in different order to the microphones and there by introduced the problem of identifying which sound corresponds to which particular event. In the section on proposed solutions we will discuss ways to limit the effects of this problem.

(38)

Microphone arrays

We started out with a lot of theoretical work in considering different approaches to designing our microphone array. We arrived at the problem, described in the problem section, of sounds arriving in different order to different microphones and began work with trying to figure out how to design an array where this is less likely to happen.

Designs using two microphones

We began however with the basic concept of two microphones each positioned at opposite sides of the welding area; it was by considering this design that we realized the problem of arrival time. When considering this area the only thing we need to keep track of is the distance D between the microphones.

Figure 37: our first attempt at array design which led to our discovery of the problem of sounds changing their order in different microphones

The array in figure 37 was discarded due to the problem associated with it that we have described and we started thinking on what changes could be made to this design. Still working with only two microphones and still wanting to keep them at opposing sides of the weld we came up with the idea of using an offset. Using an offset and a more limited weld area would let us guarantee the distance to one of the microphones would always be greater than the distance to the other microphone, using this design we would lessen the risk of arrival time problems. For this design we will have to know both the distance between microphones D and the offset do which is the difference in length from the center of the welding area. We will also have to limit the width of the welding are to insure that the distance will always be greater to one of the microphones.

(39)

Figure 38: In this figure we can see how a design could look when we use a limited weld area and a microphone off-set. The last two microphone solution we will look as is also the one we choose for our algorithm testing, in this method we use two microphones positioned at the same side of the weld area. This will allow us to keep a wider weld area but we will still profit from keeping it as small as possible. What we do in this approach is to use the distance D between our two microphones as a reference and with that we can go on to positioning algorithms. One problem with this design is choosing the size of D, we want to keep D as large as possible to give us a bigger accuracy when measuring distance from the microphones to each other and to sound sources, at the same time we want D to be as small as possible to be able to measure small enough time changes to be able to estimate the frequency of the sound. This minimum length of the distance D will vary depending on how high frequencies we want to monitor but will be equal to the wave length of the examined sound.

Figure 10: Here we see a two microphone design that has both microphones on the same side as the weld. In this design there is theoretically the least risk of running in to the arrival time problem and also since our microphones are close to each other the effects of random noises and acoustic activity from outside of the weld area should be roughly the same in each microphone and therefore easier to handle.

When considering designs using only two microphones we have chosen to proceed with this design using two microphones on the same side of the welding area. With choosing this design for our two microphone testing we hope to minimize the effects of the arrival time problem and also limit the effects of background noises and random interrupting acoustic emission. One thing to take into

(40)

account however is that when using microphones that have longer distances between each other we can filter some background and random noise by comparing between microphones, this takes up fairly large amount of computing power however and we much prefer to just treat these as we would other sounds and later omit sounds that are calculated to have originated outside of the weld area. We also look at the fact that this is the most portable solution and also one that is fairly easy to set up. We see this as the optimal way of configuring two microphones in terms of accuracy, noise handling and also in sorting for arrival times.

Three or more microphones

If we have the opportunity to use more than two microphones our strategy will be to build many pairs of microphones, and also in a case of for example 3 microphones, use them as three pairs. When setting up our microphone arrays we will also consider having different pairs of microphones working from various angels relative to the welding are. And when we have microphones set up to one side of the field as preferred when using two microphones we can also try different approaches off pairing. For example we can pair microphones from different sides of the weld are and use that data in conjunction with data obtained from a single side to get as much info as we possibly can. The big win when using microphones from two side of the weld is that the distance between them is greater and therefore the points for which we can decide their origin will be located with higher precision. The drawback is as we have discussed that sounds can be wrongly matched and sounds may end up in false positions.

Sound preprocessing

When we take a first look at recorded sound we will see that most often it comes with a few problems that must be taken care off. Such as the sound curve being of centre, small noises, peaks and fluctuations due to equipment and environmental factors.

To solve these issues we have constructed some tools for pre processing of sound data that will be applied to data just before it goes to other computations, if for instance we want to use filters to single out or remove certain frequencies that will be done before applying the methods shown in this section. To describe our methods and show their functionality we have applied the step by step to a test data, this data was created to be easily presented and a good example.

(41)

Figure 40, we see the sound data directly after recording

In figure 40 we see the sound data plotted directly after recording and we will now discuss what needs to be corrected and also how we have chosen to do it. As we can see the data in the graph is not centered at zero, this gives us some problems with further calculations and there for we will try to align the data around zero

Centering data around zero

To get the data centered we must find out the distance from the data base line, the line that represents no sound events, to zero and then adjust all points in the data array with this distance. This is done by counting how many reorientations there are of each possible distance to the center and then chose the most common one. To do this we will first chose a resolution of our analysis based on how many decimals we want to use. In the chosen example we have chosen to round to the second decimal, since it gives a good enough accuracy. We now round all numbers and then

determine how many there are of each one; wherever we find the most numbers we have our base line.

(42)

Figure 41, this is the frequency with which each distance is appearing.

In figure 41 we see a graph representing the number of findings for each distance to zero in the sound data, the peak value of 28 means the distance in 0.28 from our min value which is -0.35. This gives us the distance -0.07 in our example. If we subtract that from each element en the sound data array we will receive an array where the values have been normalized around zero.

Figure 42, the sound data has now been normalized around zero

Now that our sound data is normalized around zero we can begin working on sorting out background noise and smaller spikes so that we can identify and analyze important sound events.

(43)

Singling out sound events

Now that the values all revolve around zero we may calculate a Root mean squared value for the data and use that to set up appropriate threshold values.

∑

Now we will calculate a value P relative to the Root mean squared value for each data point and then compare that value to an appropriate threshold.

We will use the threshold value to sort out as much background noise as possible, for testing algorithms we will set the threshold manually to a value that takes away an appropriate amount of background noise. In our example P was set to 3.5. If we apply this as a upper and lower limit to the data and remove points that fall within these limits we will get a data output that holds sound data that will be considered sound events, and also a few stray sounds that will be removed in later stages.

Figure 43, our sound data after background sounds are removed.

Now, the next thing to do is to decide which of the sounds we see corresponds to a given sound event and also if any of these sounds can be omitted as single peaks. We will demand of any sound that is still left in the data at this point to have other sounds within 500 indexes in the array, which means roughly 0.01 seconds. Sound peaks that occur within that time frame are considered as part of the same sound event and peaks which fail to fall within 0.01 seconds of at least one other peak are omitted.

(44)

Figure 44, showing our sound data with an indicator on what constitutes as one sound event.

Here we see the result off our sound event localization technique. The positive flank in the green line marks the beginning of the sound event and the negative flank marks the end. These two points are what we need to locate the sound when we want to use it for various calculations.

Finding a max point to use in calculations

The next thing to do is to try to approximate a max point to use for later calculations. The problem we face here is the phenomena explained as waves in waves. That is, sound waves of various frequencies that lies within each other and thereby makes our approximations more complex. We might also get a good approximation by using the max point within the timeslot though this will require that our sampling rate is big enough.

When approximate the max value we will also look at the negative values since max points might be found amongst them as well. Therefore we will set all values within our timeslot to positive values by use of the root square method. And also we will do our calculations in a new array and shed the old indexes.

(45)

Figure 45, the sound event shown in figure 27 after all values have been made positive.

We will now try to find points suitable for curve fitting that lie within the singled out time slot. To do this we will begin with the first maximum point in the field and from that point search for the next maximum point that has higher amplitude. We will continue this search for all maximum points higher than the last max point until we reach the highest point in the timeslot. When the highest point has been reach we will redo the same thing starting at the end of the time slot and work ourselves back to the maximum point. All points that are not found to be a local maximum higher than the last local maximum will be set to 0.

(46)

Our idea at this point is to try to fit a polynomial to this data and to fit it as tight as possible around the maximum point. Since the length of the data should have no affect on the max point and the fading out part of the data is as irrelevant we have focused all our effort on the maximum points and 3-5 points to every side of it.

After experimenting and calculating we descried to use a second degree polynomial focused on the maximum point and its two nearest neighbors as our approximation.

Of course the quality of our approximation is a key issue for our future calculations and work can be done in this section for years to come. We have chosen a direct and fast approach so that we can move on for now and examine the whole problem.

Figure 47, our second degree polynomial seen in red against the sound event data shown in blue.

The data from our approximation is stored and we now have a time for the sound event and also the approximated maximum amplitude that can be seen in the event.

The final work done on sound preprocessing was to put in a safeguard against single sound peaks breaching our threshold. To handle these kinds of situations we set a simple select statement that forbids events with less than three points above the threshold from entering the approximation calculation.

(47)

Figure 48, an example of a case where some of the found sounds will be omitted from the approximation calculation. In this sections final figure, figure 48, we see an example of the above explained phenomena. The last sound event shown will be omitted on account of being too small for calculations.

Sound source localization

In this part we will look at the task of localizing sound sources using difference in amplitude and arrival time, we will assume that sound activity is arriving in order at both microphones; if this is untrue in some of our cases we will take care of this later by use of clustering to eliminate stray sounds.

We will first look at the easiest available case where we have two microphones set up at one side of the weld area.

Using two microphones

When using two microphones set up at opposite sides of the weld area sound source localization is a process of calculating the distances from the two microphones to the source. This is done by

comparing the sound data in each microphone and treating these as hypotenuses in two triangles that share one side, the y value.

(48)

Figure 49: Two microphones used to create triangles for positioning.

Since we will now the difference in arrival time and amplitude we know both the relation between the two Hypotenuses and their difference in length.

Figure 50: In this figure we see the readout from the two microphones

With the use of this information we can began calculations to determine the position of the source of the analyzed sound. First we will look at the known constants and correlation between variables. We will name the variables for the distance from our microphones to the point P on the X-axis X1 and X2. And the distance on the Y-axis, which is the same for both microphones will simply be called Y. Our output value for X will be the same as X1. When looking at our two X values we see from the figure that:

We also know from earlier reasoning that we can determine the relation in length between the hypotenuses as the inverse of the relation between measured amplitudes.

The difference in length Q between our two hypotenuses is given by multiplying the speed of sound with their difference in arrival times, and knowing that H2 will always be longer that H1 gives us: