Carlos, Without seeing your code or the data, it is hard to identify the problem. One possibility is that NuPIC is not handling the simultaneous tracks properly.
How long is the time window where you have data? --------- Matt Taylor OS Community Flag-Bearer Numenta On Mon, Feb 22, 2016 at 12:25 PM, carlos arenas <[email protected]> wrote: > Hi, I've divided the tracks in North and South heading. I think that this > can make it easier for the system to figure out the patterns. However, it > hasn’t shown any great improvement. > > I have achieved some improvement increasing the scale. It is logical, > because if a higher number of positions are taken as the same or as closer > it will identify a higher number of routes as similar. My problem right now > is that it still gives back all the positions as anomalies unless I select a > very big scale (like 1000) which, I think, would reduce the anomaly > detection capacity. > > I would like to know if is there any other model parameter that I can modify > in order to make the system converge more quickly. > > 2016-02-19 19:48 GMT+01:00 carlos arenas <[email protected]>: >> >> I've already been introducing it separated by ships and reseting the >> sequence after every ship. >> >> I am using data from an API service of marinetraffic.com. I'm sorry but >> it isn't free, however, if you are using it for academic purpouses they >> sometimes give it for free. >> >> My code is basically a modification of Geospatial Tracking, with a >> different preprocessing and a final representation on Google Earth with a >> KML file. >> >> 2016-02-19 17:04 GMT+01:00 Matthew Taylor <[email protected]>: >>> >>> How have you been feeding in data to the one model you've created? If >>> you just send all the position data into the model chronologically, >>> that would be very confusing. The data needs to be broken up into >>> "tracks", each track representing the path one ship took through the >>> area. This is going to take some pre-processing, and I'm not sure how >>> you'll apply it to live data, but I have some ideas. >>> >>> In the meantime, even if you are going to use just one model, I >>> suggest you make sure you are doing something like this... First, for >>> the time period you are evaluating, break all the positional data into >>> "tracks" that contain the positional points of each ship over time. >>> Then, starting with the first chronological track in the area, feed >>> the points of the track into the model. When you get to the end of the >>> track, reset the sequence. Then go to the next track and do the same >>> thing. This means you won't be processing the points chronologically, >>> but as long as you reset the sequence in between the tracks I think >>> that is ok. I would love to see if it works. >>> >>> By the way, is this data free? Can you share your code? >>> >>> Thanks, >>> >>> >>> --------- >>> Matt Taylor >>> OS Community Flag-Bearer >>> Numenta >>> >>> >>> On Fri, Feb 19, 2016 at 1:58 AM, carlos arenas <[email protected]> >>> wrote: >>> > Hi Matt, >>> > >>> > >>> > >>> > I have added two pictures. The first one is the zone I'm evaluating. >>> > The red >>> > areas delimit the sea lines. The two lines at the right are North-way, >>> > and >>> > the two at the left South-way. In the second picture I’ve added all the >>> > ship >>> > positions I had recorded yesterday morning. >>> > >>> > My main idea was to create only one model for all de area introducing >>> > each >>> > ship track as a new one. I have also tried to filter the data inflow to >>> > only >>> > ships going north. My results still are very poor, but I think that I >>> > haven’t introduced enough data yet. >>> > >>> > One of the things I’m testing is if the system is able to stablish some >>> > patterns, even though there are ships going north and south at the same >>> > time. >>> > >>> > 2016-02-18 23:55 GMT+01:00 Matthew Taylor <[email protected]>: >>> >> >>> >> Carlos, I hope this helps. Sometimes I just need to try to explain >>> >> something with graphics :) >>> >> >>> >> https://youtu.be/aednR5_xgZ8 >>> >> >>> >> https://realtimeboard.com/app/board/iXjVOf4zo2g=/ >>> >> --------- >>> >> Matt Taylor >>> >> OS Community Flag-Bearer >>> >> Numenta >>> >> >>> >> >>> >> On Thu, Feb 18, 2016 at 2:24 PM, carlos arenas >>> >> <[email protected]> >>> >> wrote: >>> >> > Not in an automatic way for every ship >>> >> > >>> >> > 2016-02-18 23:21 GMT+01:00 Matthew Taylor <[email protected]>: >>> >> >> >>> >> >> Is there any way you can get the port of origin and destinations >>> >> >> from >>> >> >> each >>> >> >> ship? >>> >> >> --------- >>> >> >> Matt Taylor >>> >> >> OS Community Flag-Bearer >>> >> >> Numenta >>> >> >> >>> >> >> >>> >> >> On Thu, Feb 18, 2016 at 1:56 PM, carlos arenas >>> >> >> <[email protected]> >>> >> >> wrote: >>> >> >> > Good evening Matt, >>> >> >> > >>> >> >> > I’m analyzing the traffic in a Traffic Separation Scheme (TSS), >>> >> >> > specifically >>> >> >> > the Finisterre TSS, which is like a sea highway. This is the big >>> >> >> > area >>> >> >> > I >>> >> >> > was >>> >> >> > talking about. For tracking the ships I’m using AIS data, which >>> >> >> > includes >>> >> >> > the >>> >> >> > ship MMSI, latitude, longitude, speed, course, status and >>> >> >> > timestamp, >>> >> >> > from >>> >> >> > the marinetraffic.com’s API for AIS Data. I tried to modify some >>> >> >> > of >>> >> >> > the >>> >> >> > model parameters and it worked better. However, I did it kind of >>> >> >> > trial >>> >> >> > and >>> >> >> > error. So far I have introduced the data of only one day. I >>> >> >> > suppose >>> >> >> > that >>> >> >> > as >>> >> >> > the time goes by it will improve. >>> >> >> > >>> >> >> > I would get as an anomaly any ship that goes out of the defined >>> >> >> > lane, >>> >> >> > that >>> >> >> > sails with the inverse course or even that has a too oblique >>> >> >> > course. >>> >> >> > >>> >> >> > Thank you for answering. >>> >> >> > >>> >> >> > >>> >> >> > >>> >> >> > Best regards, >>> >> >> > >>> >> >> > Carlos >>> >> >> > >>> >> >> > >>> >> >> > 2016-02-18 22:03 GMT+01:00 Matthew Taylor <[email protected]>: >>> >> >> >> >>> >> >> >> Hi Carlos. I can probably help you, but I have some questions >>> >> >> >> first... >>> >> >> >> >>> >> >> >> What kind of event would classify in your mind as an anomaly in >>> >> >> >> this >>> >> >> >> situation? >>> >> >> >> >>> >> >> >> Are there maritime "highways" that ships stick to? Or are they >>> >> >> >> all >>> >> >> >> over the place? >>> >> >> >> --------- >>> >> >> >> Matt Taylor >>> >> >> >> OS Community Flag-Bearer >>> >> >> >> Numenta >>> >> >> >> >>> >> >> >> >>> >> >> >> On Thu, Feb 18, 2016 at 11:13 AM, carlos arenas >>> >> >> >> <[email protected]> wrote: >>> >> >> >> > Good evening, >>> >> >> >> > >>> >> >> >> > For my degree final project I'm adapting Geospatial Tracking >>> >> >> >> > for >>> >> >> >> > anomaly >>> >> >> >> > detection in maritime routes. I´m covering around 4555 square >>> >> >> >> > Km >>> >> >> >> > and >>> >> >> >> > an >>> >> >> >> > inflow of 5000 ship positions every day. So far the sistem is >>> >> >> >> > struggling >>> >> >> >> > to >>> >> >> >> > detect the patterns. I would be grateful if you can give any >>> >> >> >> > advice >>> >> >> >> > about >>> >> >> >> > how to modify the sistem in order to make it able to process >>> >> >> >> > this >>> >> >> >> > data >>> >> >> >> > size. >>> >> >> >> > I'm studing Mechanical Eingeniering, so I'm not very deep in >>> >> >> >> > mathematics >>> >> >> >> > nor >>> >> >> >> > programmation. however I've been working hard in learning how >>> >> >> >> > NuPIC >>> >> >> >> > works. >>> >> >> >> > I'm very grateful in advance for your help. >>> >> >> >> > Best regards, >>> >> >> >> > Carlos >>> >> >> >> >>> >> >> > >>> >> >> >>> >> > >>> >> >>> > >>> >> >
