Early detection of signi cant traumatic events, e.g. a terrorist attack or a ship capsizing, is important to ensure that a prompt emergency response can occur. In the modern world telecommunication systems could play a key role in ensuring a successful emergency response by detecting such incidents through signi cant changes in calls and access to the networks. In this paper a methodology is illustrated to detect such incidents immediately (with the delay in the order of milliseconds), by processing semantically annotated streams of data in cellular telecommunication systems. In our methodology, live information about the position and status of phones are encoded as RDF streams. We propose an algorithm that processes streams of RDF annotated telecommunication data to detect abnormality. Our approach is exempli ed in the context of a passenger cruise ship capsizing. However, the approach is readily translatable to other incidents. Our evaluation results show that with a properly chosen window size, such incidents can be detected e ciently and e ectively.
At 21:30 on 1st June, 2015, the cruise ship \Eastern Star" traveling on the
Yangtze River began to list as a consequence of stormy weather. One minute
later, the ship capsized with 458 passengers and crew on board. No distress signal
was sent. It was several hours before the emergency services became aware of the
tragedy and 442 lives were lost [
When a ship capsizes and is incapable of communication by radio, is there
any way to detect the ship capsizing as soon as it happens? In a
telecommunication system, the location and status information of a phone are stored in
the service providers' databases (the speci c name of the databases are visiting
location register, VLR, and home location register, HLR). When a cell phone
is switched o normally (e.g., by pressing power button, or batteries wearing
out), it will register its status as \detached" to the database. However, if a
device is forcefully shut down (e.g., dropped into the water, physically damaged,
or the battery removed), or it enters a blind spot with no signal coverage, it
does not have time to register its status [
This paper is structured as follows. Section 2 presents our illustrative scenario based on the \Eastern Star" shipwreck accident. Section 3 introduces some technical background knowledge of telecommunication networks that are used in our proposal. Section 4 presents the ontology we have adopted in our proposed methodology. Section 5 gives details of the algorithms. The evaluation requirements, results and discussions based on our scenario are highlighted in Section 6. Related work is presented in Section 7 and our conclusions are given in Section 8. 2
We outline a scenario based on the cruise ship \Eastern Star" capsizing on the
Yangtze River, in Jianli, Hubei, China [
In a city with a busy telecommunication datacenter with high volumes of phones (approx. 286,000 phones in the city covering an area of 373.520 km2), we have evaluated the algorithm proposed in this paper against the time taken to detect abnormalities, the accuracy of the detection. In particular, we have investigated the tradeo s between the time taken in generating linked-data streams for the massive amount of data in telecommunication system, and the accuracy of the abnormalities detected.
It is possible that some situations would reduce the accuracy of the approach described in this paper. For example, although the telecommunication service providers are trying very hard to achieve full signal coverage, there are some regions, especially in rural areas, that have no or limited signal coverage. These areas are known as blind zones. If a large number of phones enter a blind zone, then a large number of phone signals are lost at the same spot. Each service provider keeps a list of its blind zones. Such false alarms could be avoided by a simple comparison to the list before an alert is sent. 3 3.1
In the last few years, stream processing has gained a prominent attention in the
Semantic Web community [5{9]. The data in real telecommunication system, e.g.,
system log, SNMP polling, tcpdump, and con guration data, etc., are extremely
dynamic. The granularity of SNMP, Syslog and tcpdump update period is in
terms of a few seconds, one second and microseconds, respectively [
To ground our discussion, it is necessary to consider how the telecommunication
system maintains records of the location of a phone. In telecommunication
networks, phones measure the channel environment and reselect cells every 200ms,
and report its location to the cell tower periodically. This period varies from
every 30 minutes to every hour, depending on the service providers [
Lost silence needs to identify the spatial location where an abnormally high
number of phones lose their signals almost simultaneously. Key to the lost silence
algorithm is a uniformed spatial segmentation among the data. The current
spatial position representation, Universal Transverse Mercator (UTM), identi es
a location by dividing the earth into six zones, and applying a secant transverse
Mercator projection in each zone [
We made a simpli cation with regards to geo-pixel. At the equator, one third decimal degree of longitude and latitude both cover about 100 meters. However, when moving toward the pole, the distance represented by one degree of longitude degeneralize to zero, while the latitude stays almost the same, for the distance resolution of longitude (east-west distance) depends on the latitude. For example, the distance between one degree of longitude worth up to 111.320 km when latitude is 0 , and reduces to only 28.920 km when latitude is 75 . However, most countries with signi cant mobile phone using populations lie between 75 to 75 latitude. Thus, the spatial resolution of geo-pixel varies between 100 m 100 m to 100 m 30 m, which does not have any impact on the algorithm except that making the result more accurate for some countries at high latitudes.
To be able to transform telecommunication system data into RDF, we adopted an ontology that models the raw data of the system. The ontology which is named TOUCAN Ontology (TOCO) is developed for TOUCAN project4. TOCO is proposed to represent the knowledge within a heterogeneous telecommunication system, consisting various technology domains, e.g., mobile network, computer network, optical network, light delity (LiFi), and wireless delity (WiFi), etc. TOCO is composed of one core ontology and 5 sub-ontologies, namely, network resource, service, user, time, location. A diagram of the portion of the TOCO class hierarchy adopted in lost silence is shown in Fig. 1.
Some important concepts and relations are presented below. \UserEquipment" - user equipments in mobile communication system, such as phones, tablets, wearable facilities, etc. It has an object property hasStatus with the object of class Status. It is de ned in the network resource subontology. \Point" - describes the location of phone, extended from SpatialThing of the W3C WGS-84 vocabulary5. It is de ned in the location sub-ontology. It has three datatype properties, namely, long, lat, and alt, describing the longitude, latitude, and altitude of the phone. \Status" - the status of UE connectivity in communication system. There are three instances of this class, namely, Attached, Detached, and unReachable, which denotes the phone is connected, detached after deregister to the system, and unreachable without deregister to the system, respectively. It is de ned in the network resource sub-ontology. It is the range of the property hasStatus of UserEquipment.
For example, to describe the fact that \a phone lost signal at location (329.860, 246.792)", the corresponding triples are: net:Phone 1 a net:UserEquipment. pos:Point 1 a pos:Point. net:Phone 1 pos:location pos:Point 1. pos:Point 1 pos:latitude \329.860". pos:Point 1 pos:longitude \246.792". net:unReachable a net:Status. net:Phone 1 net:hasStatus net:unReachable. 5
As the data volume from telecommunication system is extremely large, the
philosophy of divide-and-conquer is adopted in the algorithm design. In the
datacenter of the telecommunication system, for our scenario more than 400 phones
lost their signal at the same location within seconds6. This abnormality could
have been detected if continuous queries were being executed on the streams in
real-time. An alert could have been raised much earlier and help sent to those on
the ship. In the datacenter of the telecommunication system, every time a phone
updates its status, a linked-data event is generated and published as a collection
of triples. As the phones in the city keep updating their location information and
signal status, a linked-data stream is generated. In order to estimate the volume
and velocity of data variation from telecommunication system during accidents,
we referred to the literatures. The mobile phones penetration rate is 96.7% in
China, [
The C-SPARQL query string to detect lost phones is shown in Listing 1.1. One C-SPARQL instance is generated for each geo-pixel. Thus, there are 29,215 C-SPARQL instances running in the experiment. With the massive data volume in telecommunication system and the heterogenous population density, it is unrealistic to execute the query based on the arrival of new triples. Thus we choose C-SPARQL, which supports a time step execution model. The query is evaluated every 5 seconds calculating for each geo-pixel the count of unreachable devices over the last 30 minutes. A result is only returned if the count is above 10. Each stream is denoted as <http://home.eps.hw.ac.uk/ qz1/ ontologies/wirelessnetwork/networkResource.owl/stream> in the query. To avoid the omissions of accidents, we choose the window size the same as the phone's position update period in communication system { 30 minutes. We adopt thread-per-geo-pixel architecture to process telecommunication data and generate a RDF stream in one geo-pixel with one thread. The threads are main
Algorithm 1 Stream Generator: Generate RDF streams for each given geopixels.
// Input are the longitude and latitude of the city, ratio of phone lost signal, thread sleep time. Output is the RDF stream created Require: lat, long, lostRatio, sleepTime Ensure: Stream keepRunning ( 1 while keepRunning do
Insert triple (net:Phone, geo:locateIn, geo:Point ) to Stream Insert triple (geo:Point, geo:latitude, \lat "^^xsd:double ) to Stream Insert triple (geo:Point, geo:longitude, \long "^^xsd:double ) to Stream With a ratio of lostRatio:
Insert triple (net:Phone, net:hasStatus, net:unReunReachableachable) to Stream With a ratio of(1 lostRatio):
Insert triple (net:Phone, net:hasStatus, net:Attached) to Stream
Thread sleep for sleepTime seconds end while Algorithm 2 Thread Pool Scheduler: Schedule the C-SPARQL engine threads. tained by a thread pool. Each thread in the pool has the same execution priority and receives an equal share of CPU time.
The details of the generated streams and thread scheduling process are illustrated in an self-explanatory way in Algorithm 1 and 2 for brevity. Algorithm 1 shows the stream generating process for di erent areas. The lostRatio is adopted to control the ratio of phones that lost signal in the simulated streams. For example, if the lostRatio = 0:1, it means 10% of the phones in this stream will lose signal, and if the lostRatio = 1, it denotes all the phones in this stream will lose signal. The density of phones is determined by the sleepTime. For example, in the rural area Hongcheng Village where the phone density is 6 phones per pixel, as the location update period is 5 seconds (= 5000 milliseconds), in order to simulate the scenario that there are 6 phones signal update per 5 seconds, we have to generate the RDF triples for one phone every 5000 6 833:333 milliseconds. Thus, the sleepTime for rural area is 833 milliseconds. Similarly, the sleepTime for Rongcheng Town is 5000 21 238 milliseconds. 6
We run an extensive and exhaustive evaluation based on the tragedy detection algorithm. The evaluation was carried out on a MacBook Air OS X 10.9.5 with 3MB cache running on Intel Core i5 at 1.5 GHz. The system has 128 GB SSD and 4 GB RAM. 6.1
A small number of phones will randomly disappear from random pixels along the river. At a randomly chosen time, 424 phones will lose their signal simultaneously in the geo-pixel (329.863, 246.792). With a query window of 30 minutes, we experimented with query steps of 5, 20, and 30 seconds respectively. For the three step sizes, we ran in total 3 iterations of the algorithm. Our evaluation for the shipwreck scenario focus on two criteria: the time taken to detect abnormalities and the accuracy of the detection, in terms of the cases of fail-to-report. 6.2
We run the experiment for three times, with the query steps of 5 seconds, 20 seconds, and 30 seconds, respectively. The simulation results are shown in Fig. 3.
The Fig. 3(a) and 3(b) are the simulation result with the query step of 5 seconds, the Fig. 3(c) and 3(d) are with the query step of 20 seconds, and the Fig. 3(e) and 3(f) 30 seconds. In the Fig. 3(a) 3(c), and 3(e) on the left, the horizontal and vertical coordinates indicate the latitude and longitude of the geographical area respectively. The total number of lost phones detected at each geo-pixel are shown as circles on the sub gures. The number of lost phones are denoted as the radius and colour of the circles, e.g., the smallest circle in black denotes only one lost phone at that location, the second smallest circle in orange denotes two lost phones, and the largest circle in red is the total 442 lost phones detected where the shipwreck took place. The Fig. 3(b), 3(d), and 3(f) on the right show the number of lost phones detected in the geo-pixel (329.863, 246.792), where the shipwreck took place, at each query step, from the beginning of the experiment. To shade light on the detailed detection process, we show the number of lost phones below 10 as well, although no alert will be send for that. As shown in the Fig. 3(b), 3(d), and 3(f), there are sudden jumps of the number of lost phones in each gures, at the time when the shipwreck happens. As the number of lost phones raised to above 10, the lost silence will send alert of that abnormally rise together with the information of the geo-pixel. As the query step grow from 5 seconds (as shown in Fig. 3(b)) to 30 seconds (as shown in Fig. 3(f)), the number of lost phones increase faster, and the slopes of the dots become steeper.
A pressure test is also carried out to simulate an extreme scenario in which 10 ships capsize in di erent locations on the river simultaneously, as shown in Fig. 4. Each ship carries more than 300 passengers. The duration of the capsizing of one ship is about 3 minutes. Thus, in the 10 geo-pixels along the river, there is a massive phone loss incident with the rate of about 300 3 = 100 phones per minute. As shown in the Fig. 4(a), the abnormality of a large number of lost phones in all of the 10 geo-pixels are detected, and alerts are sent successfully. 30.030.0 32.0 34.0 34.034.0 35.0
Latitude 329.864
Latitude (c) 246.786 329.859 329.861 329.863 329.865 329.867 329.869
Latitude (e) 40000
60000
Time (ms) (f) 20000 80000 100000 Fig. 3. The total number of lost phones in the shipwreck area and the query results in the geo-pixel (329.863, 246.792), where the shipwreck took place, at each query step for the three experiments with the query steps of 5 seconds, 20 seconds, and 30 seconds, respectively. Fig. a) and b) correspond to the experiment with the query step of 5 seconds, Fig. c) and d) correspond to 20 seconds, and Fig. e) and f) 30 seconds. Fig. a), c), and e) illustrate the total number of phones lost signal at the water area of city Jianli at each experiment. The number of phones lost signal scales with the colour and radius of the circle, e.g., the brighter colour and larger radius of a circle denotes a larger number of lost phones. Fig. b), d), and f) show the number of detected lost phones in the geo-pixel (329.863, 246.792), at each query step from the beginning of the query. 239.800 239.798 239.796 itendo 223399..779924 u g L 239.790 239.788 239.786 .885223 .985223 .680223 .186223 .268223 .386223Latit.486223ude.586223 .668223 .786223 .826823 .982263 .082723 (a)
The results at each query step for an random geo-pixel is shown in the Fig. 4(b). A sharp increase of the lost phones can be spotted, with the rate about 100 phones per minute. The result demonstrates the e ciency and practicality of the lost silence. 7
The utility of data from telecommunication system in geography and social
science to improve urban planning has been increasingly investigated [4, 15{18].
In [
Scholte and Rozenkrane [
In this paper we have shown that the system data inside cellular networks can be used to detect accidents with large number of phones lost signal at the same location simultaneously, in our scenario a ship capsizing incident.
In lost silence, heterogenous user data from all the vendors are concentrated and organized with Geo-Pixel, which is proposed as a geography unit. Continuous RDF processing is adopted to query the data streams in real time. We perform the simulation based on a real life shipwreck incident in China 2015, and discussed the results in di erent scenarios.
This lost silence service adopts the ontology developed and built for the TOUCAN project. Current ontologies of telecommunication networks generally do not provide information and knowledge inside the system. In our future work, we will intent to extend the lost silence with machine learning approaches and real data from cellular networks.
This research was supported by the EPSRC TOUCAN project (Grant No. EP/L020009/1).