![\"Sample](\"https://storage.googleapis.com/kaggle-forum-message-attachments/481846/11481/Screenshot%20from%202019-03-01%2014-51-42.png\")
I am working on figuring out exact time of the earthquake event in the acoustic data. In my understanding the quake should create a loud(er) and long(er) sound event that should stand out in the waveform of the acoustic data.
\n\nIt turns out that these loud events happen on average 0.315 seconds before the time_to_failure reaches minimum for a given laboratory quake experiment. Given that 150000 samples chunk covers 0.0375 seconds, the quake seems to happen about 100 chunks before the moment time_to_failure runs to the minimum (which is, notable, not zero).
\n\nHow we know that the quake happened?
\n\nAs we know there are 16 earthquake experiment data segments in the training data. These segments can be identified by finding the indices of minimal values of time_to_failure values. However within span of the each experiment data segment there is more than one acoustic event that exceeds 5 standard deviations of the mean acoustic signal value while lasting for at least 1000 samples. The difference between these events and the final \"quake\" event is that after the final event the fault becomes quiet because the tension is relieved. This fact justifies inclusion of the trailing 100 chunks because they would carry evidence for the release of pressure. This 100 chunks period perhaps is the time resolution of the device measuring sheer stress mentioned in the Additional Info post:\n
\n\n\n\n\nTime to failure is based on a measure of fault strength (shear stress, not part of the data for the competition). When a labquake occurs this stress drops unambiguously.
\n
The image below is a waveform of the laboratory earthquake. It is the quake #1. The quake is the dark blob is in the middle of the image. You can see a lot of variability leading to the quake on the left of it and less of it to the right of the quake. \n
The attached mp3 file is the sound of the quake #5 with acoustic event at 0.320 sec. before time_to_failure reaches minimum.
\n\nWhat should be considered the time of failure?
\n\nIt would make sense to use beginning of the associated acoustic event as time of failure. This way the quake itself would not be used for feature creation hence avoiding information leakage.
\n\nPlease let me know in the comments if you would like me to release a kernel illustrating these findings.
", "messages": [ { "id": 481846, "postDate": "2019-03-01T22:59:43.837Z", "content": "I am working on figuring out exact time of the earthquake event in the acoustic data. In my understanding the quake should create a loud(er) and long(er) sound event that should stand out in the waveform of the acoustic data.
\n\nIt turns out that these loud events happen on average 0.315 seconds before the time_to_failure reaches minimum for a given laboratory quake experiment. Given that 150000 samples chunk covers 0.0375 seconds, the quake seems to happen about 100 chunks before the moment time_to_failure runs to the minimum (which is, notable, not zero).
\n\nHow we know that the quake happened?
\n\nAs we know there are 16 earthquake experiment data segments in the training data. These segments can be identified by finding the indices of minimal values of time_to_failure values. However within span of the each experiment data segment there is more than one acoustic event that exceeds 5 standard deviations of the mean acoustic signal value while lasting for at least 1000 samples. The difference between these events and the final \"quake\" event is that after the final event the fault becomes quiet because the tension is relieved. This fact justifies inclusion of the trailing 100 chunks because they would carry evidence for the release of pressure. This 100 chunks period perhaps is the time resolution of the device measuring sheer stress mentioned in the Additional Info post:\n
\n\n\n\n\nTime to failure is based on a measure of fault strength (shear stress, not part of the data for the competition). When a labquake occurs this stress drops unambiguously.
\n
The image below is a waveform of the laboratory earthquake. It is the quake #1. The quake is the dark blob is in the middle of the image. You can see a lot of variability leading to the quake on the left of it and less of it to the right of the quake. \n
The attached mp3 file is the sound of the quake #5 with acoustic event at 0.320 sec. before time_to_failure reaches minimum.
\n\nWhat should be considered the time of failure?
\n\nIt would make sense to use beginning of the associated acoustic event as time of failure. This way the quake itself would not be used for feature creation hence avoiding information leakage.
\n\nPlease let me know in the comments if you would like me to release a kernel illustrating these findings.
", "rawMarkdown": "I am working on figuring out exact time of the earthquake event in the acoustic data. In my understanding the quake should create a loud(er) and long(er) sound event that should stand out in the waveform of the acoustic data. \n\nIt turns out that these loud events happen on average 0.315 seconds before the time_to_failure reaches minimum for a given laboratory quake experiment. Given that 150000 samples chunk covers 0.0375 seconds, the quake seems to happen about 100 chunks before the moment time_to_failure runs to the minimum (which is, notable, not zero). \n\n**How we know that the quake happened?**\n\nAs we know there are 16 earthquake experiment data segments in the training data. These segments can be identified by finding the indices of minimal values of time_to_failure values. However within span of the each experiment data segment there is more than one acoustic event that exceeds 5 standard deviations of the mean acoustic signal value while lasting for at least 1000 samples. The difference between these events and the final \"quake\" event is that after the final event the fault becomes quiet because the tension is relieved. This fact justifies inclusion of the trailing 100 chunks because they would carry evidence for the release of pressure. This 100 chunks period perhaps is the time resolution of the device measuring sheer stress mentioned in the [Additional Info post:\n](https://www.kaggle.com/c/LANL-Earthquake-Prediction/discussion/77526)\n> Time to failure is based on a measure of fault strength (shear stress, not part of the data for the competition). When a labquake occurs this stress drops unambiguously.\n\nThe image below is a waveform of the laboratory earthquake. It is the quake #1. The quake is the dark blob is in the middle of the image. You can see a lot of variability leading to the quake on the left of it and less of it to the right of the quake. \n\n\nThe attached mp3 file is the sound of the quake #5 with acoustic event at 0.320 sec. before time_to_failure reaches minimum.\n\n**What should be considered the time of failure?**\n\nIt would make sense to use beginning of the associated acoustic event as time of failure. This way the quake itself would not be used for feature creation hence avoiding information leakage. \n\n\n*Please let me know in the comments if you would like me to release a kernel illustrating these findings.*", "votes": 6 }, { "id": 486893, "postDate": "2019-03-09T16:03:45.900Z", "content": "I have come to pretty much the same conclusions as sergeman after looking at the training data a lot. I posted some comments on this topic here. https://www.kaggle.com/c/LANL-Earthquake-Prediction/discussion/83377
", "rawMarkdown": "I have come to pretty much the same conclusions as sergeman after looking at the training data a lot. I posted some comments on this topic here. https://www.kaggle.com/c/LANL-Earthquake-Prediction/discussion/83377", "votes": 2 }, { "id": 482391, "postDate": "2019-03-02T22:29:46.543Z", "content": "0.315 is the value of thee time_to_failure at the moment when the last acoustic spike. it varies in 0.311 to 0.320 range for all 16 lab quakes in the data. I assume this is time in seconds
", "rawMarkdown": "0.315 is the value of thee time_to_failure at the moment when the last acoustic spike. it varies in 0.311 to 0.320 range for all 16 lab quakes in the data. I assume this is time in seconds", "replies": [ { "id": 482400, "postDate": "2019-03-02T23:07:05.760Z", "content": "Thank you, of course it is! I was calculating nanoseconds yesterday all day and didn't switch the scale in my mind.
", "rawMarkdown": "Thank you, of course it is! I was calculating nanoseconds yesterday all day and didn't switch the scale in my mind." } ] }, { "id": 481975, "postDate": "2019-03-02T06:01:40.020Z", "content": "@petyaAngelova
\n\nTo answer your question, the huge spikes in acoustic data happen many times in each of the 16 lab cycles appearing at random times except the last one that appears at 0.315 seconds before the end of the cycle.
\n\nI changed wording in a post to clarify your question and removed reference to the \"final\" quake that caused the confusion.
", "rawMarkdown": "@petyaAngelova\n\nTo answer your question, the huge spikes in acoustic data happen many times in each of the 16 lab cycles appearing at random times except the last one that appears at 0.315 seconds **before** the end of the cycle.\n\nI changed wording in a post to clarify your question and removed reference to the \"final\" quake that caused the confusion.", "replies": [ { "id": 482016, "postDate": "2019-03-02T07:44:52.793Z", "content": "I see your point. It is a good idea and it's worth a try, although I think those precursor tremors should not be completely excluded from the training process. Those emitted acoustic waves mean some drop in the the accumulated stress and this is valuable information. I will read with interest any further results in that direction.\nAlso interesting is that we do not know how big is the actual quake- the failure. The group says in their papers, that they use pre-quake data that has been ignored as noise before that research. Maybe exactly those tremors are holding meaningful information. I do not have some great CV and LB score, but using features generated with those peaks in mind, improved the numbers.
\n\nSomething small puzzles me yet. Maybe it is not seconds or the number is not correct, or I am missing something?
\n\n\n", "rawMarkdown": "I see your point. It is a good idea and it's worth a try, although I think those precursor tremors should not be completely excluded from the training process. Those emitted acoustic waves mean some drop in the the accumulated stress and this is valuable information. I will read with interest any further results in that direction.\nAlso interesting is that we do not know how big is the actual quake- the failure. The group says in their papers, that they use pre-quake data that has been ignored as noise before that research. Maybe exactly those tremors are holding meaningful information. I do not have some great CV and LB score, but using features generated with those peaks in mind, improved the numbers. \n\nSomething small puzzles me yet. Maybe it is not seconds or the number is not correct, or I am missing something?\n > 0.315 seconds \n\n" } ] }, { "id": 481946, "postDate": "2019-03-02T04:54:16.883Z", "content": "0.315 seconds
\n
Just a question to check if I understand you correctly.\n \"Final quake\" is it the huge spike in the acoustic data in each cycle or the actual system failure after the end of the cycle?
", "rawMarkdown": "Just a question to check if I understand you correctly.\n \"Final quake\" is it the huge spike in the acoustic data in each cycle or the actual system failure after the end of the cycle?" }, { "id": 481939, "postDate": "2019-03-02T04:38:25.737Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 486893, "author_name": "kickback", "author_url": "", "post_date": "2019-03-09T16:03:45.900000", "content": "I have come to pretty much the same conclusions as sergeman after looking at the training data a lot. I posted some comments on this topic here. https://www.kaggle.com/c/LANL-Earthquake-Prediction/discussion/83377
", "votes": 2, "replies": [] }, { "id": 482391, "author_name": "sergeman", "author_url": "", "post_date": "2019-03-02T22:29:46.543000", "content": "0.315 is the value of thee time_to_failure at the moment when the last acoustic spike. it varies in 0.311 to 0.320 range for all 16 lab quakes in the data. I assume this is time in seconds
", "votes": 0, "replies": [ { "id": 482400, "author_name": "petyaAngelova", "author_url": "", "post_date": "2019-03-02T23:07:05.760000", "content": "Thank you, of course it is! I was calculating nanoseconds yesterday all day and didn't switch the scale in my mind.
", "votes": 0, "replies": [] } ] }, { "id": 481975, "author_name": "sergeman", "author_url": "", "post_date": "2019-03-02T06:01:40.020000", "content": "@petyaAngelova
\n\nTo answer your question, the huge spikes in acoustic data happen many times in each of the 16 lab cycles appearing at random times except the last one that appears at 0.315 seconds before the end of the cycle.
\n\nI changed wording in a post to clarify your question and removed reference to the \"final\" quake that caused the confusion.
", "votes": 0, "replies": [ { "id": 482016, "author_name": "petyaAngelova", "author_url": "", "post_date": "2019-03-02T07:44:52.793000", "content": "I see your point. It is a good idea and it's worth a try, although I think those precursor tremors should not be completely excluded from the training process. Those emitted acoustic waves mean some drop in the the accumulated stress and this is valuable information. I will read with interest any further results in that direction.\nAlso interesting is that we do not know how big is the actual quake- the failure. The group says in their papers, that they use pre-quake data that has been ignored as noise before that research. Maybe exactly those tremors are holding meaningful information. I do not have some great CV and LB score, but using features generated with those peaks in mind, improved the numbers.
\n\nSomething small puzzles me yet. Maybe it is not seconds or the number is not correct, or I am missing something?
\n\n\n", "votes": 0, "replies": [] } ] }, { "id": 481946, "author_name": "petyaAngelova", "author_url": "", "post_date": "2019-03-02T04:54:16.883000", "content": "0.315 seconds
\n
Just a question to check if I understand you correctly.\n \"Final quake\" is it the huge spike in the acoustic data in each cycle or the actual system failure after the end of the cycle?
", "votes": 0, "replies": [] }, { "id": 481939, "author_name": "", "author_url": "", "post_date": "2019-03-02T04:38:25.737000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "481846": "I am working on figuring out exact time of the earthquake event in the acoustic data. In my understanding the quake should create a loud(er) and long(er) sound event that should stand out in the waveform of the acoustic data. \n\nIt turns out that these loud events happen on average 0.315 seconds before the time_to_failure reaches minimum for a given laboratory quake experiment. Given that 150000 samples chunk covers 0.0375 seconds, the quake seems to happen about 100 chunks before the moment time_to_failure runs to the minimum (which is, notable, not zero). \n\n**How we know that the quake happened?**\n\nAs we know there are 16 earthquake experiment data segments in the training data. These segments can be identified by finding the indices of minimal values of time_to_failure values. However within span of the each experiment data segment there is more than one acoustic event that exceeds 5 standard deviations of the mean acoustic signal value while lasting for at least 1000 samples. The difference between these events and the final \"quake\" event is that after the final event the fault becomes quiet because the tension is relieved. This fact justifies inclusion of the trailing 100 chunks because they would carry evidence for the release of pressure. This 100 chunks period perhaps is the time resolution of the device measuring sheer stress mentioned in the [Additional Info post:\n](https://www.kaggle.com/c/LANL-Earthquake-Prediction/discussion/77526)\n> Time to failure is based on a measure of fault strength (shear stress, not part of the data for the competition). When a labquake occurs this stress drops unambiguously.\n\nThe image below is a waveform of the laboratory earthquake. It is the quake #1. The quake is the dark blob is in the middle of the image. You can see a lot of variability leading to the quake on the left of it and less of it to the right of the quake. \n\n\nThe attached mp3 file is the sound of the quake #5 with acoustic event at 0.320 sec. before time_to_failure reaches minimum.\n\n**What should be considered the time of failure?**\n\nIt would make sense to use beginning of the associated acoustic event as time of failure. This way the quake itself would not be used for feature creation hence avoiding information leakage. \n\n\n*Please let me know in the comments if you would like me to release a kernel illustrating these findings.*", "486893": "I have come to pretty much the same conclusions as sergeman after looking at the training data a lot. I posted some comments on this topic here. https://www.kaggle.com/c/LANL-Earthquake-Prediction/discussion/83377", "482391": "0.315 is the value of thee time_to_failure at the moment when the last acoustic spike. it varies in 0.311 to 0.320 range for all 16 lab quakes in the data. I assume this is time in seconds", "481975": "@petyaAngelova\n\nTo answer your question, the huge spikes in acoustic data happen many times in each of the 16 lab cycles appearing at random times except the last one that appears at 0.315 seconds **before** the end of the cycle.\n\nI changed wording in a post to clarify your question and removed reference to the \"final\" quake that caused the confusion.", "481946": "Just a question to check if I understand you correctly.\n \"Final quake\" is it the huge spike in the acoustic data in each cycle or the actual system failure after the end of the cycle?", "481939": "" } }