I created some features with undergraduate physics:
\n\nElasticity\n```\nF = -kx\na = x.shift(-1) -2*x + x.shift()\nregress a ~ x
\n\n\nPhase (hysteresis) diagram slope \n\nv = x.shift(-1) - x\nregress x ~ v
\nEnergy decay speed\n\nE = x**2\ndE = E.shift(-1) - E\nregress E bin avg ~ dE bin min, at low energy\nor \nregress log(E bin avg) ~ log(-dE bin min)
```\nBecause x is integer, I used various ema to smooth it. Also I calculated the above quantities at various energy levels.
\n\nI think stiffness is related to resonance frequency.\nSo, I also computed the weighted fft frequency of length 1500 slice.
", "messages": [ { "id": "543760", "postDate": "06/04/2019 19:28:29", "content": "I created some features with undergraduate physics:
\n\nElasticity\n```\nF = -kx\na = x.shift(-1) -2*x + x.shift()\nregress a ~ x
\n\n\nPhase (hysteresis) diagram slope \n\nv = x.shift(-1) - x\nregress x ~ v
\nEnergy decay speed\n\nE = x**2\ndE = E.shift(-1) - E\nregress E bin avg ~ dE bin min, at low energy\nor \nregress log(E bin avg) ~ log(-dE bin min)
```\nBecause x is integer, I used various ema to smooth it. Also I calculated the above quantities at various energy levels.
\n\nI think stiffness is related to resonance frequency.\nSo, I also computed the weighted fft frequency of length 1500 slice.
", "rawMarkdown": "I created some features with undergraduate physics:\n\nElasticity\n```\nF = -kx\na = x.shift(-1) -2*x + x.shift()\nregress a ~ x\n\n```\nPhase (hysteresis) diagram slope \n```\nv = x.shift(-1) - x\nregress x ~ v\n\n```\nEnergy decay speed\n```\nE = x**2\ndE = E.shift(-1) - E\nregress E bin avg ~ dE bin min, at low energy\nor \nregress log(E bin avg) ~ log(-dE bin min)\n\n```\nBecause x is integer, I used various ema to smooth it. Also I calculated the above quantities at various energy levels.\n\nI think stiffness is related to resonance frequency.\nSo, I also computed the weighted fft frequency of length 1500 slice.", "votes": null }, { "id": "543765", "postDate": "06/04/2019 19:46:49", "content": "Thanks! Interesting point of view.\nPh.D. in Physics approved! 😉
", "rawMarkdown": "Thanks! Interesting point of view.\nPh.D. in Physics approved! 😉", "votes": null }, { "id": "543820", "postDate": "06/04/2019 21:18:30", "content": "I like these regression type features. Did it helped in your solution?
", "rawMarkdown": "I like these regression type features. Did it helped in your solution?", "votes": null }, { "id": "543966", "postDate": "06/05/2019 01:45:02", "content": "Yes, these are important CV features to me.
\n\nHowever, I was hoping to achieve something like: \nEnergy dissipation per cycle(hence weighted freq) is related to strain\nEnergy = kx*2 instead of x*2\nCritical damping occur when sqrt(k) increase to resonance freq
\n\nI hope features with physical meaning could be helpful to actual physicists.
", "rawMarkdown": "Yes, these are important CV features to me.\n\nHowever, I was hoping to achieve something like: \nEnergy dissipation per cycle(hence weighted freq) is related to strain\nEnergy = kx**2 instead of x**2\nCritical damping occur when sqrt(k) increase to resonance freq\n\nI hope features with physical meaning could be helpful to actual physicists.", "votes": null }, { "id": "547053", "postDate": "06/07/2019 07:43:24", "content": "Also Physics Ph.D. approved :).\nAdditional remarks:\n- FFT based features weren't really useful, the spectrum is practically uncorrelated with ttf\n- in the end, all that matters are the peaks or 'bursts' in the signal. Their size (energy) and especially their temporal distribution. It helped not to think of the data as a continuous signal, but as a collection of relevant bursts separated by periods of noise. Maybe calculating your features for individual bursts would be really great?
", "rawMarkdown": "Also Physics Ph.D. approved :).\nAdditional remarks:\n- FFT based features weren't really useful, the spectrum is practically uncorrelated with ttf\n- in the end, all that matters are the peaks or 'bursts' in the signal. Their size (energy) and especially their temporal distribution. It helped not to think of the data as a continuous signal, but as a collection of relevant bursts separated by periods of noise. Maybe calculating your features for individual bursts would be really great?", "votes": null }, { "id": "548598", "postDate": "06/09/2019 15:58:36", "content": "It's sad that there are too few data or even no data at high energy for regression. But I believe lots of physics can be extracted by analyzing those bursts carefully.
\n\nCompetition aside, computing rolling mean on mfcc and other features can show what is the state of the earth quake. The energy, elasticity, mfcc increase steadily. When they reach resonance state, it seems earthquake can be triggered any time. That's why predicting time since failure is a lot easier. But predicting ttf involves predicting how much time they spent on resonance phase.
", "rawMarkdown": "It's sad that there are too few data or even no data at high energy for regression. But I believe lots of physics can be extracted by analyzing those bursts carefully.\n\nCompetition aside, computing rolling mean on mfcc and other features can show what is the state of the earth quake. The energy, elasticity, mfcc increase steadily. When they reach resonance state, it seems earthquake can be triggered any time. That's why predicting time since failure is a lot easier. But predicting ttf involves predicting how much time they spent on resonance phase.", "votes": null } ], "comments": [ { "id": 543765, "author_name": "stecasasso", "author_url": "", "post_date": "06/04/2019 19:46:49", "content": "Thanks! Interesting point of view.\nPh.D. in Physics approved! 😉
", "votes": null, "replies": [] }, { "id": 543820, "author_name": "titericz", "author_url": "", "post_date": "06/04/2019 21:18:30", "content": "I like these regression type features. Did it helped in your solution?
", "votes": null, "replies": [ { "id": 543966, "author_name": "frankw", "author_url": "", "post_date": "06/05/2019 01:45:02", "content": "Yes, these are important CV features to me.
\n\nHowever, I was hoping to achieve something like: \nEnergy dissipation per cycle(hence weighted freq) is related to strain\nEnergy = kx*2 instead of x*2\nCritical damping occur when sqrt(k) increase to resonance freq
\n\nI hope features with physical meaning could be helpful to actual physicists.
", "votes": null, "replies": [] } ] }, { "id": 547053, "author_name": "friedchips", "author_url": "", "post_date": "06/07/2019 07:43:24", "content": "Also Physics Ph.D. approved :).\nAdditional remarks:\n- FFT based features weren't really useful, the spectrum is practically uncorrelated with ttf\n- in the end, all that matters are the peaks or 'bursts' in the signal. Their size (energy) and especially their temporal distribution. It helped not to think of the data as a continuous signal, but as a collection of relevant bursts separated by periods of noise. Maybe calculating your features for individual bursts would be really great?
", "votes": null, "replies": [ { "id": 548598, "author_name": "frankw", "author_url": "", "post_date": "06/09/2019 15:58:36", "content": "It's sad that there are too few data or even no data at high energy for regression. But I believe lots of physics can be extracted by analyzing those bursts carefully.
\n\nCompetition aside, computing rolling mean on mfcc and other features can show what is the state of the earth quake. The energy, elasticity, mfcc increase steadily. When they reach resonance state, it seems earthquake can be triggered any time. That's why predicting time since failure is a lot easier. But predicting ttf involves predicting how much time they spent on resonance phase.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "543760": "I created some features with undergraduate physics:\n\nElasticity\n```\nF = -kx\na = x.shift(-1) -2*x + x.shift()\nregress a ~ x\n\n```\nPhase (hysteresis) diagram slope \n```\nv = x.shift(-1) - x\nregress x ~ v\n\n```\nEnergy decay speed\n```\nE = x**2\ndE = E.shift(-1) - E\nregress E bin avg ~ dE bin min, at low energy\nor \nregress log(E bin avg) ~ log(-dE bin min)\n\n```\nBecause x is integer, I used various ema to smooth it. Also I calculated the above quantities at various energy levels.\n\nI think stiffness is related to resonance frequency.\nSo, I also computed the weighted fft frequency of length 1500 slice.", "543765": "Thanks! Interesting point of view.\nPh.D. in Physics approved! 😉", "543820": "I like these regression type features. Did it helped in your solution?", "543966": "Yes, these are important CV features to me.\n\nHowever, I was hoping to achieve something like: \nEnergy dissipation per cycle(hence weighted freq) is related to strain\nEnergy = kx**2 instead of x**2\nCritical damping occur when sqrt(k) increase to resonance freq\n\nI hope features with physical meaning could be helpful to actual physicists.", "547053": "Also Physics Ph.D. approved :).\nAdditional remarks:\n- FFT based features weren't really useful, the spectrum is practically uncorrelated with ttf\n- in the end, all that matters are the peaks or 'bursts' in the signal. Their size (energy) and especially their temporal distribution. It helped not to think of the data as a continuous signal, but as a collection of relevant bursts separated by periods of noise. Maybe calculating your features for individual bursts would be really great?", "548598": "It's sad that there are too few data or even no data at high energy for regression. But I believe lots of physics can be extracted by analyzing those bursts carefully.\n\nCompetition aside, computing rolling mean on mfcc and other features can show what is the state of the earth quake. The energy, elasticity, mfcc increase steadily. When they reach resonance state, it seems earthquake can be triggered any time. That's why predicting time since failure is a lot easier. But predicting ttf involves predicting how much time they spent on resonance phase." }, "source": "meta" }