I would appreciate verification. I believe the PS3 in House 4 is on L2 in the training data and L1 in the testing data (at least in the Testing_09_12_1347433201.mat datafile). The power signature of the PS3 is clearly seen in this file on L1 at time 1347489976.43 (or 2012/9/12 22:46:16 GMT).
\nFrankly, this timestamp is outside the 15 time periods covered in the Submission file, but it is very troubling to me that the PS3 was apparently moved between the training and the testing data collection. It makes me wonder if other appliances might have moved, too. This is particularly important when testing for spectral features because presumably the location in the house's electrical circuit affects the resultant HF noise spectrum.
\nIn my analysis methods I have assumed that the presence of an appliance on either L1 or L2 would be constant. Obviously, in real life this is not necessarily so (portable appliances are often plugged in to different outlets all over a house), but it often works out this way (most appliances basically never move). If we need to account for portable appliances, we would also need to have training data for each appliance on each possible outlet.
\nI ran across this when trying to test my classification method that uses spectral features. The PS3's HF noise contribution in the testing data is vastly different from the training examples (which are all very similar). I'm not sure why this is observed, but I'd also appreciate suggested explanations.
", "messages": [ { "id": "30628", "postDate": "09/10/2013 04:38:34", "content": "I would appreciate verification. I believe the PS3 in House 4 is on L2 in the training data and L1 in the testing data (at least in the Testing_09_12_1347433201.mat datafile). The power signature of the PS3 is clearly seen in this file on L1 at time 1347489976.43 (or 2012/9/12 22:46:16 GMT).
\nFrankly, this timestamp is outside the 15 time periods covered in the Submission file, but it is very troubling to me that the PS3 was apparently moved between the training and the testing data collection. It makes me wonder if other appliances might have moved, too. This is particularly important when testing for spectral features because presumably the location in the house's electrical circuit affects the resultant HF noise spectrum.
\nIn my analysis methods I have assumed that the presence of an appliance on either L1 or L2 would be constant. Obviously, in real life this is not necessarily so (portable appliances are often plugged in to different outlets all over a house), but it often works out this way (most appliances basically never move). If we need to account for portable appliances, we would also need to have training data for each appliance on each possible outlet.
\nI ran across this when trying to test my classification method that uses spectral features. The PS3's HF noise contribution in the testing data is vastly different from the training examples (which are all very similar). I'm not sure why this is observed, but I'd also appreciate suggested explanations.
", "rawMarkdown": "", "votes": null }, { "id": "30631", "postDate": "09/10/2013 05:12:57", "content": "Hello Chris,
\n\n
We are looking into this and will get back to you.
\n\n
Regards,
\nJinesh Jain
", "rawMarkdown": "", "votes": null }, { "id": "30755", "postDate": "09/11/2013 20:28:01", "content": "Chris,
\nIn a real house, there is definitely a possibility that portable appliances may have been moved between training and testing days. This is clearly a problem that needs to be solved within this bigger appliance disaggregation problem. The algorithm is expected to be able to handle this scenario.
\nThat said, I don't know what you mean by "time period". (I am guessing you already know this but) The timestamp (1347489976) you are talking about is outside the scripted times and thus outside the submission file timestamps for that day (1347486360). So this will in no way impact the score. Hope this helps.
\nRegards,
Jinesh Jain
Hi Jinesh,
\nThanks for the response. What I mean by "time period" is that the submission file's timestamps (the minute-long periods for which we are to provide predictions) are grouped into 15 six-to-eight hour long blocks that each reside within one of the testing datafiles. I'm sorry if this was confusing. (By the way, the submission file does not include any timestamps covered by the last datafile for H4.)
\nI agree, the instance of the PS3 that I found was in the datafile but outside the requested timestamps of the submission file. This case won't affect the scoring. However, if the PS3 was active during the times included in the submission file, then it will be on the "wrong" line.
\nIf the algorithm needs to account for portable appliances moving between lines, then we will need training data of those portable appliances on both lines. While I acknowledge that the power signature should look the same regardless of line (this is how I spotted the PS3 by eye), the HF spectral noise will be significantly different. In the referenced literature it is claimed that two installations of the same model of CFL can be distinguished by their spectral noise features because their location in the house (and on the wiring circuitry) affects the spectral features. If we are to learn the spectral features for the PS3 on only L2, we can only safely apply that training information to L2.
\nI guess we have to do the best we can with the data we have. It is just important for us to know these sorts of facts about the data collection (such as appliances can be portable) because they shape the assumptions we build into our classification schemes.
\nRegards,
\nChris
", "rawMarkdown": "", "votes": null }, { "id": "30776", "postDate": "09/12/2013 06:04:26", "content": "I have found at least one appliance that switched phase within the submission time periods. I think that my score gets worse when I add it to my submissions but I am not sure (it could be something else in those submissions that is causing the worse score). I'll be glad to send more details to Belkin after Luis, Michael and I get an answer about the back end solution's labeling of appliances mentioned in the other thread (about the missing dishwasher).
\nChris, your argument about the location in the house applies for portable appliances moving to a different location in the house even if they happen to stay on the same phase. We should still be able to tell the difference between a CFL and something else.
\nIf there are two CFL's in the house with identical characteristics and they are switched with each other, we should be able to identify the location of the one that was turned on even if we don't know exactly which one it is.
", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 30631, "author_name": "jinesh0", "author_url": "", "post_date": "09/10/2013 05:12:57", "content": "Hello Chris,
\n\n
We are looking into this and will get back to you.
\n\n
Regards,
\nJinesh Jain
", "votes": null, "replies": [] }, { "id": 30755, "author_name": "jinesh0", "author_url": "", "post_date": "09/11/2013 20:28:01", "content": "Chris,
\nIn a real house, there is definitely a possibility that portable appliances may have been moved between training and testing days. This is clearly a problem that needs to be solved within this bigger appliance disaggregation problem. The algorithm is expected to be able to handle this scenario.
\nThat said, I don't know what you mean by "time period". (I am guessing you already know this but) The timestamp (1347489976) you are talking about is outside the scripted times and thus outside the submission file timestamps for that day (1347486360). So this will in no way impact the score. Hope this helps.
\nRegards,
Jinesh Jain
Hi Jinesh,
\nThanks for the response. What I mean by "time period" is that the submission file's timestamps (the minute-long periods for which we are to provide predictions) are grouped into 15 six-to-eight hour long blocks that each reside within one of the testing datafiles. I'm sorry if this was confusing. (By the way, the submission file does not include any timestamps covered by the last datafile for H4.)
\nI agree, the instance of the PS3 that I found was in the datafile but outside the requested timestamps of the submission file. This case won't affect the scoring. However, if the PS3 was active during the times included in the submission file, then it will be on the "wrong" line.
\nIf the algorithm needs to account for portable appliances moving between lines, then we will need training data of those portable appliances on both lines. While I acknowledge that the power signature should look the same regardless of line (this is how I spotted the PS3 by eye), the HF spectral noise will be significantly different. In the referenced literature it is claimed that two installations of the same model of CFL can be distinguished by their spectral noise features because their location in the house (and on the wiring circuitry) affects the spectral features. If we are to learn the spectral features for the PS3 on only L2, we can only safely apply that training information to L2.
\nI guess we have to do the best we can with the data we have. It is just important for us to know these sorts of facts about the data collection (such as appliances can be portable) because they shape the assumptions we build into our classification schemes.
\nRegards,
\nChris
", "votes": null, "replies": [] }, { "id": 30776, "author_name": "noamtene", "author_url": "", "post_date": "09/12/2013 06:04:26", "content": "I have found at least one appliance that switched phase within the submission time periods. I think that my score gets worse when I add it to my submissions but I am not sure (it could be something else in those submissions that is causing the worse score). I'll be glad to send more details to Belkin after Luis, Michael and I get an answer about the back end solution's labeling of appliances mentioned in the other thread (about the missing dishwasher).
\nChris, your argument about the location in the house applies for portable appliances moving to a different location in the house even if they happen to stay on the same phase. We should still be able to tell the difference between a CFL and something else.
\nIf there are two CFL's in the house with identical characteristics and they are switched with each other, we should be able to identify the location of the one that was turned on even if we don't know exactly which one it is.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "30628": "", "30631": "", "30755": "", "30759": "", "30776": "" }, "source": "meta" }