I read some literature regarding the LANL experiments and the reason of the delay between extreme acoutic signals and the ttf being 0 is the fact that we are predicting \"Time to end of failure\".
\n\nIn the original papers, they tried predicting both the \"Time to start of failure\" and \"Time to end of failure\". Since they got better results in the latter variable, we also have to do the predictions on the same in the competition. The extreme acoustic signal occurs just before the start of failure, and since the failure doesn't end in an instant but rather takes some time to complete, there seems to be a delay.
\n\ntl;dr: We assume that ttf was \"time to start of failure\" since it makes more sense to predict when the failure is going to start but in reality it's \"time to end of failure\" and hence we have this difference in the extreme signal and ttf.
", "messages": [ { "id": "486693", "postDate": "03/09/2019 08:34:42", "content": "I read some literature regarding the LANL experiments and the reason of the delay between extreme acoutic signals and the ttf being 0 is the fact that we are predicting \"Time to end of failure\".
\n\nIn the original papers, they tried predicting both the \"Time to start of failure\" and \"Time to end of failure\". Since they got better results in the latter variable, we also have to do the predictions on the same in the competition. The extreme acoustic signal occurs just before the start of failure, and since the failure doesn't end in an instant but rather takes some time to complete, there seems to be a delay.
\n\ntl;dr: We assume that ttf was \"time to start of failure\" since it makes more sense to predict when the failure is going to start but in reality it's \"time to end of failure\" and hence we have this difference in the extreme signal and ttf.
", "rawMarkdown": "I read some literature regarding the LANL experiments and the reason of the delay between extreme acoutic signals and the ttf being 0 is the fact that we are predicting \"**Time to end of failure**\". \n\nIn the original papers, they tried predicting both the \"Time to *start of failure*\" and \"Time to *end of failure*\". Since they got better results in the latter variable, we also have to do the predictions on the same in the competition. The extreme acoustic signal occurs just before the start of failure, and since the failure doesn't end in an instant but rather takes some time to complete, there seems to be a delay.\n\n**tl;dr:** We assume that ttf was \"time to start of failure\" since it makes more sense to predict when the failure is going to start but in reality it's \"time to end of failure\" and hence we have this difference in the extreme signal and ttf.", "votes": null }, { "id": "486886", "postDate": "03/09/2019 15:54:30", "content": "Abdur has made an excellent comment which is very relevant to what LANL has challenged the competitors to model. It is apparent that many \"micro-quakes\" (my term) occur before the failure occurs. This seems very natural to me based on the way I would expect tectonic plates to move, there are many small movements that occur that produce small acoustic signals. These small movements followed by an earthquake are what I would expect when brittle materials are shifting.
\n\nThere are clearly some general trends in the data which the leaders on the leader board have captured in their results. But I have thought for some time that none of the scores are particularly accurate wrt predicting quakes. Even in the highly controlled mechanical environment of the LANL model the train data shows large amounts of randomness wrt when certain events occur in each time sequence. \nIn my opinion the real challenge of the competition is to look very closely at the details of the acoustic signatures of each test sample and determine what bucket they should go into wrt to TTF. You are not going to come up with a model that is exactly right but if you are reasonably successful in decoding the info buried in each signature you will be close to right.
\n\nThanks for reading these comments and I wonder if other people see the same things in the train data.
\n\nkickback
", "rawMarkdown": "Abdur has made an excellent comment which is very relevant to what LANL has challenged the competitors to model. It is apparent that many \"micro-quakes\" (my term) occur before the failure occurs. This seems very natural to me based on the way I would expect tectonic plates to move, there are many small movements that occur that produce small acoustic signals. These small movements followed by an earthquake are what I would expect when brittle materials are shifting. \n\n There are clearly some general trends in the data which the leaders on the leader board have captured in their results. But I have thought for some time that none of the scores are particularly accurate wrt predicting quakes. Even in the highly controlled mechanical environment of the LANL model the train data shows large amounts of randomness wrt when certain events occur in each time sequence. \n\n In my opinion the real challenge of the competition is to look very closely at the details of the acoustic signatures of each test sample and determine what bucket they should go into wrt to TTF. You are not going to come up with a model that is exactly right but if you are reasonably successful in decoding the info buried in each signature you will be close to right. \n\nThanks for reading these comments and I wonder if other people see the same things in the train data.\n\nkickback", "votes": null }, { "id": "486953", "postDate": "03/09/2019 18:58:05", "content": "That's very insightful, thanks.
", "rawMarkdown": "That's very insightful, thanks.", "votes": null }, { "id": "487034", "postDate": "03/10/2019 01:31:11", "content": "I described my findings on this subject in this post a few days ago. I trained models on the \"start of failure\" so far, but I am planing to do the \"end of failure\" horizon next.
", "rawMarkdown": "I described my findings on this subject in [this post](https://www.kaggle.com/c/LANL-Earthquake-Prediction/discussion/82510) a few days ago. I trained models on the \"start of failure\" so far, but I am planing to do the \"end of failure\" horizon next.", "votes": null }, { "id": "488239", "postDate": "03/12/2019 06:51:21", "content": "Thanks for the insights Abdur. I was assuming that the delay was caused by the time that took the shock wave to reach the piezoelectric sensor (that was just my guess). Is there any particular paper that you would highlight?
", "rawMarkdown": "Thanks for the insights Abdur. I was assuming that the delay was caused by the time that took the shock wave to reach the piezoelectric sensor (that was just my guess). Is there any particular paper that you would highlight?", "votes": null }, { "id": "488329", "postDate": "03/12/2019 09:50:22", "content": "Here's the link: https://doi.org/10.1038/s41561-018-0272-8
", "rawMarkdown": "Here's the link: https://doi.org/10.1038/s41561-018-0272-8", "votes": null }, { "id": "488492", "postDate": "03/12/2019 14:44:25", "content": "Thanks!
", "rawMarkdown": "Thanks!", "votes": null } ], "comments": [ { "id": 486886, "author_name": "bkosar1640", "author_url": "", "post_date": "03/09/2019 15:54:30", "content": "Abdur has made an excellent comment which is very relevant to what LANL has challenged the competitors to model. It is apparent that many \"micro-quakes\" (my term) occur before the failure occurs. This seems very natural to me based on the way I would expect tectonic plates to move, there are many small movements that occur that produce small acoustic signals. These small movements followed by an earthquake are what I would expect when brittle materials are shifting.
\n\nThere are clearly some general trends in the data which the leaders on the leader board have captured in their results. But I have thought for some time that none of the scores are particularly accurate wrt predicting quakes. Even in the highly controlled mechanical environment of the LANL model the train data shows large amounts of randomness wrt when certain events occur in each time sequence. \nIn my opinion the real challenge of the competition is to look very closely at the details of the acoustic signatures of each test sample and determine what bucket they should go into wrt to TTF. You are not going to come up with a model that is exactly right but if you are reasonably successful in decoding the info buried in each signature you will be close to right.
\n\nThanks for reading these comments and I wonder if other people see the same things in the train data.
\n\nkickback
", "votes": null, "replies": [] }, { "id": 486953, "author_name": "mateiionita", "author_url": "", "post_date": "03/09/2019 18:58:05", "content": "That's very insightful, thanks.
", "votes": null, "replies": [] }, { "id": 487034, "author_name": "calabist", "author_url": "", "post_date": "03/10/2019 01:31:11", "content": "I described my findings on this subject in this post a few days ago. I trained models on the \"start of failure\" so far, but I am planing to do the \"end of failure\" horizon next.
", "votes": null, "replies": [] }, { "id": 488239, "author_name": "ricarddelgado", "author_url": "", "post_date": "03/12/2019 06:51:21", "content": "Thanks for the insights Abdur. I was assuming that the delay was caused by the time that took the shock wave to reach the piezoelectric sensor (that was just my guess). Is there any particular paper that you would highlight?
", "votes": null, "replies": [ { "id": 488329, "author_name": "abdurrafae", "author_url": "", "post_date": "03/12/2019 09:50:22", "content": "Here's the link: https://doi.org/10.1038/s41561-018-0272-8
", "votes": null, "replies": [] }, { "id": 488492, "author_name": "ricarddelgado", "author_url": "", "post_date": "03/12/2019 14:44:25", "content": "Thanks!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "486693": "I read some literature regarding the LANL experiments and the reason of the delay between extreme acoutic signals and the ttf being 0 is the fact that we are predicting \"**Time to end of failure**\". \n\nIn the original papers, they tried predicting both the \"Time to *start of failure*\" and \"Time to *end of failure*\". Since they got better results in the latter variable, we also have to do the predictions on the same in the competition. The extreme acoustic signal occurs just before the start of failure, and since the failure doesn't end in an instant but rather takes some time to complete, there seems to be a delay.\n\n**tl;dr:** We assume that ttf was \"time to start of failure\" since it makes more sense to predict when the failure is going to start but in reality it's \"time to end of failure\" and hence we have this difference in the extreme signal and ttf.", "486886": "Abdur has made an excellent comment which is very relevant to what LANL has challenged the competitors to model. It is apparent that many \"micro-quakes\" (my term) occur before the failure occurs. This seems very natural to me based on the way I would expect tectonic plates to move, there are many small movements that occur that produce small acoustic signals. These small movements followed by an earthquake are what I would expect when brittle materials are shifting. \n\n There are clearly some general trends in the data which the leaders on the leader board have captured in their results. But I have thought for some time that none of the scores are particularly accurate wrt predicting quakes. Even in the highly controlled mechanical environment of the LANL model the train data shows large amounts of randomness wrt when certain events occur in each time sequence. \n\n In my opinion the real challenge of the competition is to look very closely at the details of the acoustic signatures of each test sample and determine what bucket they should go into wrt to TTF. You are not going to come up with a model that is exactly right but if you are reasonably successful in decoding the info buried in each signature you will be close to right. \n\nThanks for reading these comments and I wonder if other people see the same things in the train data.\n\nkickback", "486953": "That's very insightful, thanks.", "487034": "I described my findings on this subject in [this post](https://www.kaggle.com/c/LANL-Earthquake-Prediction/discussion/82510) a few days ago. I trained models on the \"start of failure\" so far, but I am planing to do the \"end of failure\" horizon next.", "488239": "Thanks for the insights Abdur. I was assuming that the delay was caused by the time that took the shock wave to reach the piezoelectric sensor (that was just my guess). Is there any particular paper that you would highlight?", "488329": "Here's the link: https://doi.org/10.1038/s41561-018-0272-8", "488492": "Thanks!" }, "source": "meta" }