I'm curious if converting acoustics to just the absolute values would improve modelling. What is the relationship of the negative measurements towards the positive ones? Is there any -extra- information in the negatives? \nIf there isn't it would be logical to only use absolutes since it makes for easier and potentially better modelling with decreased multicollinearity. \nAny thoughts on this? - In any case I'm doing a bit of research and might add results here later.
", "messages": [ { "id": "480671", "postDate": "02/28/2019 14:05:24", "content": "I'm curious if converting acoustics to just the absolute values would improve modelling. What is the relationship of the negative measurements towards the positive ones? Is there any -extra- information in the negatives? \nIf there isn't it would be logical to only use absolutes since it makes for easier and potentially better modelling with decreased multicollinearity. \nAny thoughts on this? - In any case I'm doing a bit of research and might add results here later.
", "rawMarkdown": "I'm curious if converting acoustics to just the absolute values would improve modelling. What is the relationship of the negative measurements towards the positive ones? Is there any -extra- information in the negatives? \nIf there isn't it would be logical to only use absolutes since it makes for easier and potentially better modelling with decreased multicollinearity. \nAny thoughts on this? - In any case I'm doing a bit of research and might add results here later.", "votes": null }, { "id": "480832", "postDate": "02/28/2019 18:19:47", "content": "I have those features, the +/- ratio and abs and they improved my score. I have some other also based on positive negative and values. \n I didn't had the chance to work on the problem lately, but one of the features I am planning is to test different thresholds bigger than 0 and to take only the values bigger that threshold. In the beginning I was thinking about what represents the negative values, how they calibrated the device- what is ground 0.
\n\nEdited to remove some misleading information.
", "rawMarkdown": "I have those features, the +/- ratio and abs and they improved my score. I have some other also based on positive negative and values. \n I didn't had the chance to work on the problem lately, but one of the features I am planning is to test different thresholds bigger than 0 and to take only the values bigger that threshold. In the beginning I was thinking about what represents the negative values, how they calibrated the device- what is ground 0. \n\n*Edited to remove some misleading information.*", "votes": null }, { "id": "482340", "postDate": "03/02/2019 19:24:48", "content": "1. info about acoustic emission testing\n2.simple explanations about negatives in sound visualization \nIn the first link they explain how cracks in materials are tested and what part of the signal has any value to the test. They do not use negatives and even filter the positives. The second link gives a general idea how to interpret + and -.
\n\nIf you want to transform everything to positive numbers, I think you should change the zero( not having sound) to some positive value and translate the signal. For example, if we have -6000 the max negative and +7000 max positive, we should add 6000 to each data point and keep the info that the zero( no-sound) border is moved to 6000.
\n\nI keep the set as it is. Beside the whole 150 000 portion I make 2 extra series- one with positive, one with negatives.\nThe graph in posNegDiff.png represents 4212 batches each with 150 000 points in it. Note, that the the cycles are not in the same order like in the training test. With pink is the the last value of ttf for each portion. In red is the difference between the log of the positive's sum and the log of the abs values of the negative sum.
\nampl.png is interesting too. With red is the sum of the max positive value for each portion and the abs value of the max negative. \nWhenever how we choose to keep and work with + and -, I am sure we have to now where the zero is.
@petya5q \nI am playing with positive and negative values as well. What do you mean by +/- ratio? I am computing the ratio of +/- means and this feature it looks interesting on a chart that looks very similar to the posNegDiff.png. It is even more interesting with a rolling window over the means.
", "rawMarkdown": "petya5q \nI am playing with positive and negative values as well. What do you mean by +/- ratio? I am computing the ratio of +/- means and this feature it looks interesting on a chart that looks very similar to the posNegDiff.png. It is even more interesting with a rolling window over the means.", "votes": null }, { "id": "482406", "postDate": "03/02/2019 23:28:35", "content": "Ratios between same type of features extracted from my positive and negative series or positive/all , negative/all etc.. I have means also. Counting the zeroes could bring some info too. I just tried it and it improve slightly my local result.\n I am still learning ML, so I need to research deeper before applying some new technique. Rolling window is on my list, because it seams very logical for that case. Especially when we do not have info about the history of the system. Meanwhile, I created some noobish way to track the trend within each portion and I am sure that there is a potential in that.
", "rawMarkdown": "Ratios between same type of features extracted from my positive and negative series or positive/all , negative/all etc.. I have means also. Counting the zeroes could bring some info too. I just tried it and it improve slightly my local result.\n I am still learning ML, so I need to research deeper before applying some new technique. Rolling window is on my list, because it seams very logical for that case. Especially when we do not have info about the history of the system. Meanwhile, I created some noobish way to track the trend within each portion and I am sure that there is a potential in that.", "votes": null }, { "id": "482579", "postDate": "03/03/2019 09:19:26", "content": "If I understand correctly, our acoustic waveforms should act just like any soundpressure waveform and normally the negative measures should cancel the positives out - but due to the insensitivity of the measuring device (or maybe a bias of the device for the positive wavesides), in our data this is not the case?
", "rawMarkdown": "If I understand correctly, our acoustic waveforms should act just like any soundpressure waveform and normally the negative measures should cancel the positives out - but due to the insensitivity of the measuring device (or maybe a bias of the device for the positive wavesides), in our data this is not the case?", "votes": null } ], "comments": [ { "id": 480832, "author_name": "petya5q", "author_url": "", "post_date": "02/28/2019 18:19:47", "content": "I have those features, the +/- ratio and abs and they improved my score. I have some other also based on positive negative and values. \n I didn't had the chance to work on the problem lately, but one of the features I am planning is to test different thresholds bigger than 0 and to take only the values bigger that threshold. In the beginning I was thinking about what represents the negative values, how they calibrated the device- what is ground 0.
\n\nEdited to remove some misleading information.
", "votes": null, "replies": [] }, { "id": 482340, "author_name": "petya5q", "author_url": "", "post_date": "03/02/2019 19:24:48", "content": "1. info about acoustic emission testing\n2.simple explanations about negatives in sound visualization \nIn the first link they explain how cracks in materials are tested and what part of the signal has any value to the test. They do not use negatives and even filter the positives. The second link gives a general idea how to interpret + and -.
\n\nIf you want to transform everything to positive numbers, I think you should change the zero( not having sound) to some positive value and translate the signal. For example, if we have -6000 the max negative and +7000 max positive, we should add 6000 to each data point and keep the info that the zero( no-sound) border is moved to 6000.
\n\nI keep the set as it is. Beside the whole 150 000 portion I make 2 extra series- one with positive, one with negatives.\nThe graph in posNegDiff.png represents 4212 batches each with 150 000 points in it. Note, that the the cycles are not in the same order like in the training test. With pink is the the last value of ttf for each portion. In red is the difference between the log of the positive's sum and the log of the abs values of the negative sum.
\nampl.png is interesting too. With red is the sum of the max positive value for each portion and the abs value of the max negative. \nWhenever how we choose to keep and work with + and -, I am sure we have to now where the zero is.
If I understand correctly, our acoustic waveforms should act just like any soundpressure waveform and normally the negative measures should cancel the positives out - but due to the insensitivity of the measuring device (or maybe a bias of the device for the positive wavesides), in our data this is not the case?
", "votes": null, "replies": [] } ] }, { "id": 482393, "author_name": "calabist", "author_url": "", "post_date": "03/02/2019 22:40:28", "content": "@petya5q \nI am playing with positive and negative values as well. What do you mean by +/- ratio? I am computing the ratio of +/- means and this feature it looks interesting on a chart that looks very similar to the posNegDiff.png. It is even more interesting with a rolling window over the means.
", "votes": null, "replies": [ { "id": 482406, "author_name": "petya5q", "author_url": "", "post_date": "03/02/2019 23:28:35", "content": "Ratios between same type of features extracted from my positive and negative series or positive/all , negative/all etc.. I have means also. Counting the zeroes could bring some info too. I just tried it and it improve slightly my local result.\n I am still learning ML, so I need to research deeper before applying some new technique. Rolling window is on my list, because it seams very logical for that case. Especially when we do not have info about the history of the system. Meanwhile, I created some noobish way to track the trend within each portion and I am sure that there is a potential in that.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "480671": "I'm curious if converting acoustics to just the absolute values would improve modelling. What is the relationship of the negative measurements towards the positive ones? Is there any -extra- information in the negatives? \nIf there isn't it would be logical to only use absolutes since it makes for easier and potentially better modelling with decreased multicollinearity. \nAny thoughts on this? - In any case I'm doing a bit of research and might add results here later.", "480832": "I have those features, the +/- ratio and abs and they improved my score. I have some other also based on positive negative and values. \n I didn't had the chance to work on the problem lately, but one of the features I am planning is to test different thresholds bigger than 0 and to take only the values bigger that threshold. In the beginning I was thinking about what represents the negative values, how they calibrated the device- what is ground 0. \n\n*Edited to remove some misleading information.*", "482340": "[1. info about acoustic emission testing](http://www.nde-ed.org/EducationResources/CommunityCollege/Other%20Methods/AE/AE_Signal%20Features.htm)\n[2.simple explanations about negatives in sound visualization ](https://forums.stevehoffman.tv/threads/why-do-audio-waveforms-have-a-negative-side.261141/#post-6896180)\nIn the first link they explain how cracks in materials are tested and what part of the signal has any value to the test. They do not use negatives and even filter the positives. The second link gives a general idea how to interpret + and -. \n\n\nIf you want to transform everything to positive numbers, I think you should change the zero( not having sound) to some positive value and translate the signal. For example, if we have -6000 the max negative and +7000 max positive, we should add 6000 to each data point and keep the info that the zero( no-sound) border is moved to 6000. \n\nI keep the set as it is. Beside the whole 150 000 portion I make 2 extra series- one with positive, one with negatives.\nThe graph in posNegDiff.png represents 4212 batches each with 150 000 points in it. Note, that the the cycles are not in the same order like in the training test. With pink is the the last value of ttf for each portion. In red is the difference between the log of the positive's sum and the log of the abs values of the negative sum. \nampl.png is interesting too. With red is the sum of the max positive value for each portion and the abs value of the max negative. \nWhenever how we choose to keep and work with + and -, I am sure we have to now where the zero is.", "482393": "petya5q \nI am playing with positive and negative values as well. What do you mean by +/- ratio? I am computing the ratio of +/- means and this feature it looks interesting on a chart that looks very similar to the posNegDiff.png. It is even more interesting with a rolling window over the means.", "482406": "Ratios between same type of features extracted from my positive and negative series or positive/all , negative/all etc.. I have means also. Counting the zeroes could bring some info too. I just tried it and it improve slightly my local result.\n I am still learning ML, so I need to research deeper before applying some new technique. Rolling window is on my list, because it seams very logical for that case. Especially when we do not have info about the history of the system. Meanwhile, I created some noobish way to track the trend within each portion and I am sure that there is a potential in that.", "482579": "If I understand correctly, our acoustic waveforms should act just like any soundpressure waveform and normally the negative measures should cancel the positives out - but due to the insensitivity of the measuring device (or maybe a bias of the device for the positive wavesides), in our data this is not the case?" }, "source": "meta" }