I've been doing some research into laboratory seismological experiment methods in order to clarify how the data for this competition was derived. The experiment for this competition is modeled on that of Johnson et al (2013), a research group that also included Chris Marone who performed much work in earlier, similar experiments. The paper can be read here:
\n\nhttps://agupubs.onlinelibrary.wiley.com/doi/full/10.1002/2013GL057848
\n\nThe equipment they used for their readings were a Brüel and Kjær model 4393 accelerometer and a Brüel and Kjær 2635 charge amplifier, whose product specifications are here:
\n\nhttps://www.bksv.com/media/doc/Bp2043.pdf\nhttps://www.bksv.com/media/doc/bp0099.pdf
\n\nThe amplifier is designed to output a constant gain across its frequency range for measured acceleration, although this does dip towards the higher end. One important note is that the competition data has an extremely high sample rate, capable of measuring frequencies far beyond the range of any available piezoceramic accelerometers I've found online. I'll have to leave it to more qualified users to determine if this is at all significant - perhaps the sampling rate is independant of the accelerometers own frequency range.
\n\nNow, for Johnson et al's work, paragraph 7 in the article above is of particular note. They applied a 34-45kHz bandpass filter and isolated the acoustic waves above a certain threshold, classifying these as Acoustic Events (AE). The presence of these was a strong indicator of 'micro-slips' that preceded a failure event. The equations relating wave speed, frequency and strain in this paragraph may be of use for developing a microslip threshold in your own feature engineering. In this work, microslips were identified by a separate apparatus measuring strain, data which we lack in this competition. We will have to identify them with AE alone; as the researchers noted, the signal-to-noise ratio was much higher for AEs than for microslips so while not every microslip causes an AE, almost every AE was indicative of a microslip.
\n\nAnother finding of interest is that both the number of microslip occurances, and the number of AEs, rose exponentially above the background probability at a TTF of around 10 seconds - unfortunately for our sake, it appears to be random before this stage, but does indicate that the natural log of AE counts can be a useful, linear feature for TTF < 10s. The article has a number of helpful graphs for their results, but I'm unsure about Kaggle's policy on posting images taken from academic papers so I won't repost them here - they can be viewed via the link above.
\n\nTheir findings would explain why the acoustic variance at various quintiles is a useful feature for many of the models so far, as it acts like a less refined means of identifying regions displaying AEs. Finding a threshold for the competition data that identifies microslips more exactly, and can count their number in a 150000-row segment, may prove very useful.
\n\nObviously there will be differences between this experiment and the one used in this competition, but hopefully this will give you some more insight into the data and how it might be used. Best of luck everyone!
\n\nFor additional reading, Chris Marone's seminal 1998 paper 'Laboratory-derived friction laws and their application to seismic faulting' is available to read here: ftp://ftp.gps.caltech.edu/pub/avouac/GE277-2005/Articles/MaroneAREPS1998.pdf
", "messages": [ { "id": "507308", "postDate": "04/04/2019 14:34:51", "content": "I've been doing some research into laboratory seismological experiment methods in order to clarify how the data for this competition was derived. The experiment for this competition is modeled on that of Johnson et al (2013), a research group that also included Chris Marone who performed much work in earlier, similar experiments. The paper can be read here:
\n\nhttps://agupubs.onlinelibrary.wiley.com/doi/full/10.1002/2013GL057848
\n\nThe equipment they used for their readings were a Brüel and Kjær model 4393 accelerometer and a Brüel and Kjær 2635 charge amplifier, whose product specifications are here:
\n\nhttps://www.bksv.com/media/doc/Bp2043.pdf\nhttps://www.bksv.com/media/doc/bp0099.pdf
\n\nThe amplifier is designed to output a constant gain across its frequency range for measured acceleration, although this does dip towards the higher end. One important note is that the competition data has an extremely high sample rate, capable of measuring frequencies far beyond the range of any available piezoceramic accelerometers I've found online. I'll have to leave it to more qualified users to determine if this is at all significant - perhaps the sampling rate is independant of the accelerometers own frequency range.
\n\nNow, for Johnson et al's work, paragraph 7 in the article above is of particular note. They applied a 34-45kHz bandpass filter and isolated the acoustic waves above a certain threshold, classifying these as Acoustic Events (AE). The presence of these was a strong indicator of 'micro-slips' that preceded a failure event. The equations relating wave speed, frequency and strain in this paragraph may be of use for developing a microslip threshold in your own feature engineering. In this work, microslips were identified by a separate apparatus measuring strain, data which we lack in this competition. We will have to identify them with AE alone; as the researchers noted, the signal-to-noise ratio was much higher for AEs than for microslips so while not every microslip causes an AE, almost every AE was indicative of a microslip.
\n\nAnother finding of interest is that both the number of microslip occurances, and the number of AEs, rose exponentially above the background probability at a TTF of around 10 seconds - unfortunately for our sake, it appears to be random before this stage, but does indicate that the natural log of AE counts can be a useful, linear feature for TTF < 10s. The article has a number of helpful graphs for their results, but I'm unsure about Kaggle's policy on posting images taken from academic papers so I won't repost them here - they can be viewed via the link above.
\n\nTheir findings would explain why the acoustic variance at various quintiles is a useful feature for many of the models so far, as it acts like a less refined means of identifying regions displaying AEs. Finding a threshold for the competition data that identifies microslips more exactly, and can count their number in a 150000-row segment, may prove very useful.
\n\nObviously there will be differences between this experiment and the one used in this competition, but hopefully this will give you some more insight into the data and how it might be used. Best of luck everyone!
\n\nFor additional reading, Chris Marone's seminal 1998 paper 'Laboratory-derived friction laws and their application to seismic faulting' is available to read here: ftp://ftp.gps.caltech.edu/pub/avouac/GE277-2005/Articles/MaroneAREPS1998.pdf
", "rawMarkdown": "I've been doing some research into laboratory seismological experiment methods in order to clarify how the data for this competition was derived. The experiment for this competition is modeled on that of Johnson et al (2013), a research group that also included Chris Marone who performed much work in earlier, similar experiments. The paper can be read here:\n\nhttps://agupubs.onlinelibrary.wiley.com/doi/full/10.1002/2013GL057848\n\nThe equipment they used for their readings were a Brüel and Kjær model 4393 accelerometer and a Brüel and Kjær 2635 charge amplifier, whose product specifications are here:\n\nhttps://www.bksv.com/media/doc/Bp2043.pdf\nhttps://www.bksv.com/media/doc/bp0099.pdf\n\nThe amplifier is designed to output a constant gain across its frequency range for measured acceleration, although this does dip towards the higher end. One important note is that the competition data has an extremely high sample rate, capable of measuring frequencies far beyond the range of any available piezoceramic accelerometers I've found online. I'll have to leave it to more qualified users to determine if this is at all significant - perhaps the sampling rate is independant of the accelerometers own frequency range.\n\nNow, for Johnson et al's work, paragraph 7 in the article above is of particular note. They applied a 34-45kHz bandpass filter and isolated the acoustic waves above a certain threshold, classifying these as Acoustic Events (AE). The presence of these was a strong indicator of 'micro-slips' that preceded a failure event. The equations relating wave speed, frequency and strain in this paragraph may be of use for developing a microslip threshold in your own feature engineering. In this work, microslips were identified by a separate apparatus measuring strain, data which we lack in this competition. We will have to identify them with AE alone; as the researchers noted, the signal-to-noise ratio was much higher for AEs than for microslips so while not every microslip causes an AE, almost every AE was indicative of a microslip.\n\nAnother finding of interest is that both the number of microslip occurances, and the number of AEs, rose exponentially above the background probability at a TTF of around 10 seconds - unfortunately for our sake, it appears to be random before this stage, but does indicate that the natural log of AE counts can be a useful, linear feature for TTF < 10s. The article has a number of helpful graphs for their results, but I'm unsure about Kaggle's policy on posting images taken from academic papers so I won't repost them here - they can be viewed via the link above.\n\nTheir findings would explain why the acoustic variance at various quintiles is a useful feature for many of the models so far, as it acts like a less refined means of identifying regions displaying AEs. Finding a threshold for the competition data that identifies microslips more exactly, and can count their number in a 150000-row segment, may prove very useful.\n\nObviously there will be differences between this experiment and the one used in this competition, but hopefully this will give you some more insight into the data and how it might be used. Best of luck everyone! \n\nFor additional reading, Chris Marone's seminal 1998 paper *'Laboratory-derived friction laws and their application to seismic faulting'* is available to read here: ftp://ftp.gps.caltech.edu/pub/avouac/GE277-2005/Articles/MaroneAREPS1998.pdf", "votes": null }, { "id": "507382", "postDate": "04/04/2019 16:00:39", "content": "Great write up! Perhaps it could be interesting to try a similar band-pass filter in a preprocessing step.
\n\nAs for the effect of the competition data sampling rate, and thus measurable frequencies, being far beyond the frequency range of the piezoelectric accelerometer, the main issue would be inaccurate response of transducer. If you refer to the high-frequency response of the B&K type 4393 transducer you linked, for example, it is evident that at 30 kHz and above there are some very significant fluctuations from the true values (approx +5 dB @ 30 kHz, +30 dB @ 50 kHz, -15 dB @ 100 kHz).
\n\nI have some experience correcting this same problem on pressure microphones used in the free field utilizing the dimensions of the microphone diaphragm and essentially estimating the diffraction about the diaphragm, but while it strikes me that this would be a similar issue based on the range of frequencies affected, I'm not sure how much of that would really apply to an accelerometer. One would of course need to know which model of accelerometer it was too. To my knowledge correcting these types of responses in accelerometers remains relatively unexplored, so most likely it would be a project/thesis in and of itself.
\n\nThank you for the detailed reply, evidently you have more expertise in this field than me!
\n\nThe dimensionality of the accelerometer, and other specific details, have probably been withheld for a reason - the researchers will be looking for models that can generalise as much as possible without being tailored too much to their equipment. Hopefully we can still use these old papers for new ideas when we get stuck. The bandpass filter in particular may prove very helpful if the correct range can be identified.
\n\nAnother issue is that the minimum frequency that can be reliably detected in a 150000-row time window is about 6kHz. If the distinctive AE frequency lies in the ~30kHz range then there's a limit on how many could possibly be recorded for one segment - if the segment is analysed in stages then that pushes us to higher and higher frequencies... we may inevitably be forced to examine the less informative parts of the spectrum.
", "rawMarkdown": "Thank you for the detailed reply, evidently you have more expertise in this field than me!\n\nThe dimensionality of the accelerometer, and other specific details, have probably been withheld for a reason - the researchers will be looking for models that can generalise as much as possible without being tailored too much to their equipment. Hopefully we can still use these old papers for new ideas when we get stuck. The bandpass filter in particular may prove very helpful if the correct range can be identified.\n\nAnother issue is that the minimum frequency that can be reliably detected in a 150000-row time window is about 6kHz. If the distinctive AE frequency lies in the ~30kHz range then there's a limit on how many could possibly be recorded for one segment - if the segment is analysed in stages then that pushes us to higher and higher frequencies... we may inevitably be forced to examine the less informative parts of the spectrum.", "votes": null }, { "id": "519756", "postDate": "04/19/2019 15:30:19", "content": "Thanks for sharing.
", "rawMarkdown": "Thanks for sharing.", "votes": null } ], "comments": [ { "id": 507382, "author_name": "ochiwankenobi", "author_url": "", "post_date": "04/04/2019 16:00:39", "content": "Great write up! Perhaps it could be interesting to try a similar band-pass filter in a preprocessing step.
\n\nAs for the effect of the competition data sampling rate, and thus measurable frequencies, being far beyond the frequency range of the piezoelectric accelerometer, the main issue would be inaccurate response of transducer. If you refer to the high-frequency response of the B&K type 4393 transducer you linked, for example, it is evident that at 30 kHz and above there are some very significant fluctuations from the true values (approx +5 dB @ 30 kHz, +30 dB @ 50 kHz, -15 dB @ 100 kHz).
\n\nI have some experience correcting this same problem on pressure microphones used in the free field utilizing the dimensions of the microphone diaphragm and essentially estimating the diffraction about the diaphragm, but while it strikes me that this would be a similar issue based on the range of frequencies affected, I'm not sure how much of that would really apply to an accelerometer. One would of course need to know which model of accelerometer it was too. To my knowledge correcting these types of responses in accelerometers remains relatively unexplored, so most likely it would be a project/thesis in and of itself.
\n\nThank you for the detailed reply, evidently you have more expertise in this field than me!
\n\nThe dimensionality of the accelerometer, and other specific details, have probably been withheld for a reason - the researchers will be looking for models that can generalise as much as possible without being tailored too much to their equipment. Hopefully we can still use these old papers for new ideas when we get stuck. The bandpass filter in particular may prove very helpful if the correct range can be identified.
\n\nAnother issue is that the minimum frequency that can be reliably detected in a 150000-row time window is about 6kHz. If the distinctive AE frequency lies in the ~30kHz range then there's a limit on how many could possibly be recorded for one segment - if the segment is analysed in stages then that pushes us to higher and higher frequencies... we may inevitably be forced to examine the less informative parts of the spectrum.
", "votes": null, "replies": [] } ] }, { "id": 519756, "author_name": "cpmpml", "author_url": "", "post_date": "04/19/2019 15:30:19", "content": "Thanks for sharing.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "507308": "I've been doing some research into laboratory seismological experiment methods in order to clarify how the data for this competition was derived. The experiment for this competition is modeled on that of Johnson et al (2013), a research group that also included Chris Marone who performed much work in earlier, similar experiments. The paper can be read here:\n\nhttps://agupubs.onlinelibrary.wiley.com/doi/full/10.1002/2013GL057848\n\nThe equipment they used for their readings were a Brüel and Kjær model 4393 accelerometer and a Brüel and Kjær 2635 charge amplifier, whose product specifications are here:\n\nhttps://www.bksv.com/media/doc/Bp2043.pdf\nhttps://www.bksv.com/media/doc/bp0099.pdf\n\nThe amplifier is designed to output a constant gain across its frequency range for measured acceleration, although this does dip towards the higher end. One important note is that the competition data has an extremely high sample rate, capable of measuring frequencies far beyond the range of any available piezoceramic accelerometers I've found online. I'll have to leave it to more qualified users to determine if this is at all significant - perhaps the sampling rate is independant of the accelerometers own frequency range.\n\nNow, for Johnson et al's work, paragraph 7 in the article above is of particular note. They applied a 34-45kHz bandpass filter and isolated the acoustic waves above a certain threshold, classifying these as Acoustic Events (AE). The presence of these was a strong indicator of 'micro-slips' that preceded a failure event. The equations relating wave speed, frequency and strain in this paragraph may be of use for developing a microslip threshold in your own feature engineering. In this work, microslips were identified by a separate apparatus measuring strain, data which we lack in this competition. We will have to identify them with AE alone; as the researchers noted, the signal-to-noise ratio was much higher for AEs than for microslips so while not every microslip causes an AE, almost every AE was indicative of a microslip.\n\nAnother finding of interest is that both the number of microslip occurances, and the number of AEs, rose exponentially above the background probability at a TTF of around 10 seconds - unfortunately for our sake, it appears to be random before this stage, but does indicate that the natural log of AE counts can be a useful, linear feature for TTF < 10s. The article has a number of helpful graphs for their results, but I'm unsure about Kaggle's policy on posting images taken from academic papers so I won't repost them here - they can be viewed via the link above.\n\nTheir findings would explain why the acoustic variance at various quintiles is a useful feature for many of the models so far, as it acts like a less refined means of identifying regions displaying AEs. Finding a threshold for the competition data that identifies microslips more exactly, and can count their number in a 150000-row segment, may prove very useful.\n\nObviously there will be differences between this experiment and the one used in this competition, but hopefully this will give you some more insight into the data and how it might be used. Best of luck everyone! \n\nFor additional reading, Chris Marone's seminal 1998 paper *'Laboratory-derived friction laws and their application to seismic faulting'* is available to read here: ftp://ftp.gps.caltech.edu/pub/avouac/GE277-2005/Articles/MaroneAREPS1998.pdf", "507382": "Great write up! Perhaps it could be interesting to try a similar band-pass filter in a preprocessing step.\n\nAs for the effect of the competition data sampling rate, and thus measurable frequencies, being far beyond the frequency range of the piezoelectric accelerometer, the main issue would be inaccurate response of transducer. If you refer to the high-frequency response of the B&K type 4393 transducer you linked, for example, it is evident that at 30 kHz and above there are some very significant fluctuations from the true values (approx +5 dB @ 30 kHz, +30 dB @ 50 kHz, -15 dB @ 100 kHz).\n\nI have some experience correcting this same problem on pressure microphones used in the free field utilizing the dimensions of the microphone diaphragm and essentially estimating the diffraction about the diaphragm, but while it strikes me that this would be a similar issue based on the range of frequencies affected, I'm not sure how much of that would really apply to an accelerometer. One would of course need to know which model of accelerometer it was too. To my knowledge correcting these types of responses in accelerometers remains relatively unexplored, so most likely it would be a project/thesis in and of itself.\n\n* If one could correct the diffraction response of the transducer in the frequency domain, then one could theoretically transfer back into the time domain to obtain the true waveform (or something close to it).", "510227": "Thank you for the detailed reply, evidently you have more expertise in this field than me!\n\nThe dimensionality of the accelerometer, and other specific details, have probably been withheld for a reason - the researchers will be looking for models that can generalise as much as possible without being tailored too much to their equipment. Hopefully we can still use these old papers for new ideas when we get stuck. The bandpass filter in particular may prove very helpful if the correct range can be identified.\n\nAnother issue is that the minimum frequency that can be reliably detected in a 150000-row time window is about 6kHz. If the distinctive AE frequency lies in the ~30kHz range then there's a limit on how many could possibly be recorded for one segment - if the segment is analysed in stages then that pushes us to higher and higher frequencies... we may inevitably be forced to examine the less informative parts of the spectrum.", "519756": "Thanks for sharing." }, "source": "meta" }