We are probably working on the problem of an assembly line producing several products in few hundred variants (not 1000+ completely different products). \nIndustry 4.0 - Bosch Rexroth Multi Product Line and few similar movies may be helpful.
", "messages": [ { "id": "133467", "postDate": "09/02/2016 12:31:09", "content": "We are probably working on the problem of an assembly line producing several products in few hundred variants (not 1000+ completely different products). \nIndustry 4.0 - Bosch Rexroth Multi Product Line and few similar movies may be helpful.
", "rawMarkdown": "We are probably working on the problem of an assembly line producing several products in few hundred variants (not 1000+ completely different products). \r\n[Industry 4.0 - Bosch Rexroth Multi Product Line][1] and few similar movies may be helpful. \r\n\r\n\r\n [1]: https://www.youtube.com/watch?v=DrE0FShBfF4", "votes": null }, { "id": "135192", "postDate": "09/12/2016 02:21:32", "content": "This was my thinking, too. I find the number of different flow paths through the stations suspect -- do you think it's possible that sometimes items that are supposed to flow through the same path actually don't get all the same measurements? Or maybe they get all the same measurements except for one or two?
", "rawMarkdown": "This was my thinking, too. I find the number of different flow paths through the stations suspect -- do you think it's possible that sometimes items that are supposed to flow through the same path actually don't get all the same measurements? Or maybe they get all the same measurements except for one or two?", "votes": null }, { "id": "135193", "postDate": "09/12/2016 02:41:27", "content": "You should be able to identify product lines using a cluster algorithm on the measurement data. Some of the data is clearly multi-modal, and some of the modes may reflect differences in part specification. Check out parameters like L0_S0_F6 and L0_S4_F104. On the other hand, parameters like L0_S0_F18 and L3_S30_F3519 seem to be more closely related to tool performance
\n\nThis kernel shows some of this behavior.
", "rawMarkdown": "You should be able to identify product lines using a cluster algorithm on the measurement data. Some of the data is clearly multi-modal, and some of the modes may reflect differences in part specification. Check out parameters like L0_S0_F6 and L0_S4_F104. On the other hand, parameters like L0_S0_F18 and L3_S30_F3519 seem to be more closely related to tool performance\r\n\r\n[This kernel][1] shows some of this behavior.\r\n\r\n\r\n\r\n [1]: https://www.kaggle.com/jeffhebert/bosch-production-line-performance/visualize-station-parameters-over-time/output", "votes": null }, { "id": "135213", "postDate": "09/12/2016 05:58:33", "content": "HI Jeff, thanks for replying.
\n\nI tried training a naive bayes classifier on a set of flow paths that looked "correct" (flow paths that had shown up more than 45 times in the first 10000 records), but after a couple of hours, I decided to kill the job. I guess ~1000 features for ~3000 samples is too expensive.
\n\nInstead, I decided to learn locality sensitive hashing. I think this would enable me to bucket the different flow paths in a way that would keep similar ones close to each other, so differences of one or two stations should be well-tolerated.
\n\nGiven your experience, do you think it's reasonable to suppose that out of a set of measurements for a particular product, some records might just be missing a couple of measurements here or there? Or is it more likely that the records with missing measurements are actually very small batches, maybe even one-off products? For example, in a sample of the first 10000 products, many flow paths show up only once.
", "rawMarkdown": "HI Jeff, thanks for replying. \r\n\r\nI tried training a naive bayes classifier on a set of flow paths that looked \"correct\" (flow paths that had shown up more than 45 times in the first 10000 records), but after a couple of hours, I decided to kill the job. I guess ~1000 features for ~3000 samples is too expensive. \r\n\r\nInstead, I decided to learn locality sensitive hashing. I think this would enable me to bucket the different flow paths in a way that would keep similar ones close to each other, so differences of one or two stations should be well-tolerated. \r\n\r\nGiven your experience, do you think it's reasonable to suppose that out of a set of measurements for a particular product, some records might just be missing a couple of measurements here or there? Or is it more likely that the records with missing measurements are actually very small batches, maybe even one-off products? For example, in a sample of the first 10000 products, many flow paths show up only once.", "votes": null } ], "comments": [ { "id": 135192, "author_name": "fuzzylogic", "author_url": "", "post_date": "09/12/2016 02:21:32", "content": "This was my thinking, too. I find the number of different flow paths through the stations suspect -- do you think it's possible that sometimes items that are supposed to flow through the same path actually don't get all the same measurements? Or maybe they get all the same measurements except for one or two?
", "votes": null, "replies": [] }, { "id": 135193, "author_name": "jeffhebert", "author_url": "", "post_date": "09/12/2016 02:41:27", "content": "You should be able to identify product lines using a cluster algorithm on the measurement data. Some of the data is clearly multi-modal, and some of the modes may reflect differences in part specification. Check out parameters like L0_S0_F6 and L0_S4_F104. On the other hand, parameters like L0_S0_F18 and L3_S30_F3519 seem to be more closely related to tool performance
\n\nThis kernel shows some of this behavior.
", "votes": null, "replies": [] }, { "id": 135213, "author_name": "fuzzylogic", "author_url": "", "post_date": "09/12/2016 05:58:33", "content": "HI Jeff, thanks for replying.
\n\nI tried training a naive bayes classifier on a set of flow paths that looked "correct" (flow paths that had shown up more than 45 times in the first 10000 records), but after a couple of hours, I decided to kill the job. I guess ~1000 features for ~3000 samples is too expensive.
\n\nInstead, I decided to learn locality sensitive hashing. I think this would enable me to bucket the different flow paths in a way that would keep similar ones close to each other, so differences of one or two stations should be well-tolerated.
\n\nGiven your experience, do you think it's reasonable to suppose that out of a set of measurements for a particular product, some records might just be missing a couple of measurements here or there? Or is it more likely that the records with missing measurements are actually very small batches, maybe even one-off products? For example, in a sample of the first 10000 products, many flow paths show up only once.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "133467": "We are probably working on the problem of an assembly line producing several products in few hundred variants (not 1000+ completely different products). \r\n[Industry 4.0 - Bosch Rexroth Multi Product Line][1] and few similar movies may be helpful. \r\n\r\n\r\n [1]: https://www.youtube.com/watch?v=DrE0FShBfF4", "135192": "This was my thinking, too. I find the number of different flow paths through the stations suspect -- do you think it's possible that sometimes items that are supposed to flow through the same path actually don't get all the same measurements? Or maybe they get all the same measurements except for one or two?", "135193": "You should be able to identify product lines using a cluster algorithm on the measurement data. Some of the data is clearly multi-modal, and some of the modes may reflect differences in part specification. Check out parameters like L0_S0_F6 and L0_S4_F104. On the other hand, parameters like L0_S0_F18 and L3_S30_F3519 seem to be more closely related to tool performance\r\n\r\n[This kernel][1] shows some of this behavior.\r\n\r\n\r\n\r\n [1]: https://www.kaggle.com/jeffhebert/bosch-production-line-performance/visualize-station-parameters-over-time/output", "135213": "HI Jeff, thanks for replying. \r\n\r\nI tried training a naive bayes classifier on a set of flow paths that looked \"correct\" (flow paths that had shown up more than 45 times in the first 10000 records), but after a couple of hours, I decided to kill the job. I guess ~1000 features for ~3000 samples is too expensive. \r\n\r\nInstead, I decided to learn locality sensitive hashing. I think this would enable me to bucket the different flow paths in a way that would keep similar ones close to each other, so differences of one or two stations should be well-tolerated. \r\n\r\nGiven your experience, do you think it's reasonable to suppose that out of a set of measurements for a particular product, some records might just be missing a couple of measurements here or there? Or is it more likely that the records with missing measurements are actually very small batches, maybe even one-off products? For example, in a sample of the first 10000 products, many flow paths show up only once." }, "source": "meta" }