System Integration of an IEEE 802.11 based TDoA Localization System Stefan Schwalowsky, Henning Trsek
Reinhard Exel
Nikolaus Ker¨o
Oregano Systems Institute for Integrated Sensor Systems inIT - Institut Industrial IT Design & Consulting GmbH Austrian Academy of Sciences Ostwestfalen-Lippe University of Applied Sciences 1030 Wien, Austria 2700 Wiener Neustadt, Austria 32657 Lemgo, Germany
[email protected] [email protected] {stefan.schwalowsky,henning.trsek}@hs-owl.de
Abstract—Network entities with synchronized clocks are an enabler for many interesting and innovative applications. Localization of mobile WLAN devices by means of propagation delay measurements and position calculation according to time difference of arrival (TDoA) method is one of the most interesting applications. Those localization techniques are favored over standard signal strength based methods when high accuracy positioning is desired. This paper presents a system that utilizes special synchronized receivers in conjunction with a central unit to locate WLAN devices. The system architecture is described in terms of the integration of all participating entities for both, hardware and software. 1
I. I NTRODUCTION Nowadays wireless networks are ubiquitous in the consumer market and the trend towards even more flexibility and mobility using wireless communication is still unbroken. Wireless communication systems were recently also established in the industrial market, where they are used to connect remote sensors, actuators and other devices with the automation system. Their huge economization potential and a high flexibility are the main reasons for their deployment. Besides providing several benefits, wireless technologies, like IEEE 802.11 Wireless LANs [1], pose also a number of challenges, e.g., the dynamic wireless channel, or their vulnerability. This stems from the vast density of wireless devices that are equipped with several radio technologies interfering with each other. Especially at sites covered only by a few access points (AP) this can degrade availability of crucial network services. Through a localization system numerous services, such as resource allocation, channel planning, or security can be supplemented in order to achieve a quality similar to wired solutions. New innovative applications could be realized with localization systems. Location information of wireless human machine interfaces (HMI), for instance, can be used to grant special rights or deliver information to the personal operating the production machines depending on their actual position. However, most indoor communication systems lack a reliable method to precisely determine the position of certain mobile stations without modification of the user’s mobile device. Localization solutions should seamlessly integrate into existing wireless infrastructures with no or 1 The work presented in this paper is founded by the FIT-IT Project ε-WiFi Embedded Position Determination and Security in Wireless Fidelity Networks. Grant Number 813310
978-1-4244-5977-3/10/$26.00 ©2010 IEEE
minimal impact on standard data transmissions. It should be avoided to generate a significant amount of additional traffic on the wireless medium to prevent an increase of network latency and load on the channel. The remainder of this paper is organized as follows: In section II existing systems and techniques are discussed. Section III explains details of the used localization technique and equipment. The processing of location information in the system is described in section IV. A presentation of results is given in section V. Finally an outlook for future work is given. II. R ELATED W ORK The intensive research in the field of WLAN node localization yields many different approaches to realize accurate localization. The two major methods are based on signal strength and signal propagation time. A. Signal Strength based localization The most common method is the localization by means of the received signal strength (RSS) of several APs in the area of interest. Most of the methods utilize a map created in a training-phase before the actual position determination phase. They are also referred to as manual methods. The offline generated map is used in the consecutive online phase to estimate the position based on current RSS measurements. This can be achieved for instance by nearest neighbor estimation as shown in [2] and [3]. The achievable accuracy is in the range of approximately 2 to 3 m. One decisive drawback of this method is a mandatory re-training phase, once the environment has been subject to major changes, like moved furniture or a rearrangement of an assembly line. Another approach is called automatic method, and omits the training phase by using algorithms for estimating the environment to determine the position of a mobile station (MS). In [4] a system is presented that achieves accuracies around 3 to 4 m. Though, the possible accuracy is not equal to manual methods as they offer more flexibility in case of environmental changes, which happen to occur frequently in industrial environments.
B. Propagation time based localization Current research activities also deal with localization methods which use signal propagation time. These methods can be grouped in Time of Arrival (ToA) and Time Difference of Arrival (TDoA). The ToA methods determine the propagation time between the MS and the base stations (BS). A prerequisite for ToA is that all entities share the same notion of time, they have to be synchronized with an accuracy in the range of single-digit ns in order to achieve localization results of around 1 to 3 m. Such an accuracy is impossible to achieve with commercial-of-the-shelf (COTS) WLAN devices [5], as this would require hardware timestamping mechanisms in all entities and highly stable oscillators. In TDoA localization schemes clock synchronization is only required among BS. Usually, they are connected to a wired backbone, which eases the synchronization between the BS’s. However, the complexity of estimating the position is higher in TDoA localization systems as every TDoA is jointly dependent on two arrival time measurements. This results in a hyperbolic system consisting of a set of non-linear equations. Given that the equation matrix is overdefined (due to more TDoA measurements than required), a minimum mean square error solution (MMSE) solution leads to a high computational complexity. Nonetheless, this approach is taken by systems described in [6], [7]. C. Comparison of both schemes Selecting a suitable localization scheme depends on the individual weights put on the characteristics of the system. One of them is the complexity of a system in terms of hardware and software amendments. RSS based schemes have a definitive advantage in this aspect, because almost every WLAN device provides means to read out the current RSS value. Thus necessary efforts merely remain on the software development for such a system. ToA or TDoA schemes require synchronization of the entities and timestamping of the wireless frames in the range of ns which can only be achieved by means of hardware support. As network standards like IEEE 802.3 do not offer this feature by default, custom implementations for synchronization as well as wireless timestamping are required. The most interesting metric is the achievable accuracy. Propagation based methods and RSS cannot be compared directly as the measurement principle is completely different. In general, RSS methods work only well in the vicinity of the BS as the slope of the free-space path loss decreases with increasing distance, thus making the range estimation inaccurate for large distances. In addition, the RSS may change by 10 dB and more by multipath effects or polarization, consequently rendering the solution unusable. ToA/TDoA methods do not suffer from these problems, they are affected by the bandwidth of the wireless link, the SNR and multipath effects. In total, the RSS methods dwindle as the indoor environment is subject to changes, whereas ToA or TDoA is by far not affected in the same extent as RSS, hence making it to the preferred method for localization applications.
In any case, the decision whether to favor accuracy or minimize complexity is fundamental in any localization system. Table I summarizes the core attributes. TABLE I: Comparison of localization schemes Accuracy
Complexity
Requirements
RSS (automatic)
–
+
Software only
RSS (manual)
◦
◦
previous map creation
ToA
+
––
synchronized BS & MS
TDoA
+
-
synchronized BS
III. T HE TD OA L OCALIZATION S YSTEM The TDoA localization concept used in the ε-WiFi project has similarities to the well-known GPS position determination system with basically the transmitters and receivers swapped. The unmodified MS transmits a frame, which is received by multiple receivers (N + 1 receivers for N dimensions) with given positions. Using the arrival times of N + 1 receivers, N propagation delay differences can be obtained. As the free air propagation speed is known, the TDoA can be converted to range differences, each defining a hyperboloid (3D hyperbola) of possible positions of the MS, whereas the final position can be calculated by intersecting the N hyperboloids. As the propagation speed is close to the speed of light, the measured delays and hence the time differences are in the range of nanoseconds. Therefore, it is of utmost importance to keep timing errors below a few nanoseconds. These errors arise due to loose synchronization between the receivers or by inaccurate estimation of the arrival time of the wireless frames. Although the positional accuracy depends on the intersection angle of the hyperbolas (positional dilution of precision), it can be roughly expected that an inaccuracy of 3 ns causes a localization error of 1 m. A receiver architecture (as described in [8]) was built able to perform timestamping of wireless frames in the subnanosecond range. The approach is to calculate the TDoA based on timestamps, which are taken at the receivers in contrast to traditional approaches [9] operating on a general crosscorrelation scheme. The inherent advantage of this method is that the data sampled at every receiver does not need to be transfered to a central cross-correlation unit over a network. Instead timestamps are taken at a specific position that is present in all frames (e.g., at the end of the WLAN header). The timestamps together with some identification information are forwarded to a central server which calculates the position based on the preceding position and the new measurement. The timestamp approach relieves the server from the burden to perform multiple cross-correlations in real-time. The receiver consists of an analog front-end board called SMiLE board and an Altera Stratix II GX evaluation board. The SMiLE board demodulates, amplifies and finally digitalizes the received data and passes the digital data stream to
the FPGA board. The FPGA decodes the frame completely and takes a timestamp which is forwarded via Ethernet to the Ranging Engine. It has been shown that the SMiLE architecture is able to timestamp frames with an accuracy of below 1 ns [10]. Besides timestamping, the synchronization is of similar importance for the operation of TDoA; the boards have to be synchronized to a few nanoseconds. Even though the-stateof-the-art clock synchronization protocols, like IEEE 1588 in conjunction with hardware timestamping can provide an accuracy in the range of a few nanoseconds, this accuracy is hardly sufficient for this application. Maintaining such an accuracy with IEEE 1588 and individual oscillators in the receivers requires high speed digital logic and expensive oscillators. Still as some Ethernet PHYs have a non deterministic clock lock-in behavior, the transmission line becomes asymmetric which results in a synchronization offset rendering TDoA unusable [11]. Hence the synchronization is based on a combined approach: adjusting clock rates by distributing a single frequency over the network (as proposed by synchronous Ethernet) and secondly by determining the clock offsets at the SMiLE boards. The frequency distribution is maintained via Ethernet clocking: The SMiLE boards are connected to a common switch and use the recovered Ethernet clock as their local clock source. As every board is incrementing the local counter with the same rate, the clock offset between two boards remain constant. Since the geometry of the system and the signal propagation speed are known, the propagation delays and delay differences from the transmitter with known position (e.g., the AP) to the SMiLE boards can be calculated. Assuming that the wireless frames from the AP can be timestamped by every board, the time offsets can be determined. These offsets are stored in the server and used to correct the timestamps creating a virtual common timebase. IV. S YSTEM I NTEGRATION The integration of different existing components to the overall system is an important aspect, because the achievable accuracy also depends on their interaction. The system integration is divided in several steps which are discussed in this section. A. Timestamp processing The first topic of the system integration is the timestamp extraction. The SMiLE receivers generate timestamps at the exact reception time of a WLAN frame. The timestamp is compiled of the essential WLAN frame information, the counter value and the RSS information. In our approach we use a central PC running a software named the Ranging Engine, which performs the task of timestamp aggregation and processing. In the architecture it is intended, that all SMiLE receivers are connected to the Ranging Engine over a single Ethernet switch, as it is a requirement for the syntonisation of the receivers. All information from the set of SMiLE receivers is gathered at the Ranging Engine, so the first processing stage
RCV 1
RCV 2
RCV n
Fig. 1: Timestamp compilation process
acts as a information dispatcher, which evaluates the WLAN frame information to relate a timestamp to a MS based on the source IP address. The timestamps together with identification information are put into N queues, one for each SMiLE receiver. As it is not guaranteed that the timestamps arrive in the same order as they have been timestamped, due to varying delays in the Ethernet switch, a queue storing a few timestamps is required to find all matching timestamps. If sufficient timestamps are in the queue, the algorithm tries to find a matching tuple, whereas the first timestamp serves as reference for the compilation of a tuple. The Ranging Engine memorizes the value of the last processed timestamp, if there is an orphan timestamp in a queue, that relates to a WLAN frame received in the past, it is abolished immediately. If the necessary amount of timestamps for a tuple could not be compiled, the timestamps are also dropped and the next reference is fetched from one of the timestamp queues. The identification of the timestamps belonging to a WLAN frame is solved by comparison of the frame check sum (FCS) available in all WLAN frames. Possible hash collision problems are eliminated with the additional consideration of the receive time and the source MAC address [12]. A complete tuple of up-to-date timestamps is passed to the localization algorithm and the calculated position is stored within a MS representing object. This process is depicted in figure 1. B. Enhancement measures In order to enhance the achievable accuracy, a suitable filtering has to be applied. It is either possible to filter the raw TDoA or the calculated positions. Due to the implementation of a decoupled localization algorithm, it is also possible to apply filters at both stages in the processing chain. Figure 2 shows the filter application at different stages. One important fact for the filtering process is, that it has to consider a variable timestamp interval as the traffic generated
GUI
Data from Receivers
Ranging Engine SMiLE Rcv
Timstamp gathering
Visualization & Interface
Position calculation
TS Filter
Position Filter
WLAN AP
Other clients
MS Existing Infrastructure
Fig. 2: Filter application in the processing chain
by the MS is not regular. Estimation algorithms like the Kalman filter are usually expressed in state space equations. The matrix representation of the involving equations simplifies the integration of a varying sample-time interval. The Kalman filter estimates the next state of the mathematical model by aggregating noisy measurement data and predictions of the process. The positions calculated by the localization algorithm are the measurements for the estimation process. Since we track moving WLAN MS’s, the process model is based on the mathematical model of accelerated linear motion, 1 (1) ~s = ~s0 + ~v · t + ~a · t2 2 Where ~s, ~s0 = [x, y], ~v = [vx , vy ] and ~a = [ax , ay ]. In our case the time is considered as interval since the last timestamp reception ∆t = tk − tk−1 and s0 as the estimated MS position at tk−1 . When transforming (1) into a state space representation, ~xk = A · ~xk−1 (2) the resulting matrices form the initial point for T the Kalman equations. ~xk = x y vx vy ax ay denotes the state space vector at the time k. A is the state transition matrix containing the time interval ∆t. For the Kalman filter process additional parameters have to be defined that describe the process noise covariance Q and the measurement noise covariance R. These covariance matrices can be adopted to react on modified environment conditions, for instance higher motion of the MS or degraded receiving conditions. The application of the Kalman algorithm as timestamp filter, right before the localization stage, emerges differently. The state vector is reduced to the set of TDoA values; xk = ∆t12 ∆t23 ∆t34 . The state transition matrix A forms the identity matrix. This parameterization of the Kalman filter forms an estimator for constant values, which is in our case applicable to stationary WLAN devices, like APs. The estimated values are valuable for the time offset determination in order to create the virtual common timebase mentioned in section III. C. System Deployment The proposed localization system is composed of the SMiLE receivers, the central Ranging Engine and visualization clients. The receivers are connected to an Ethernet switch passing the timestamps to the Ranging Engine. The Ranging
Ethernet
Fig. 3: System deployment
Engine itself can be located anywhere in the network, as it does not require synchronization with the SMiLE receivers. The localization system is intended to act as a passive unit within the existing network and only reacts on direct requests. When connecting the Ranging Engine is connected to the same Ethernet segment as the SMiLE receivers, the rest of the network, wireless as well as wired, is completely unaffected by the presence of localization system. An exemplary setup is depicted in figure 3. The MS, implied by the user with a mobile WLAN device, has unaffected access to the services offered by the existing network. Meanwhile the wireless traffic is recognized by the SMiLE receivers, which forward the timing information to the Ranging Engine. The obtained position information can be polled by devices in the network, either to grant access to services or to just display location information. This architecture features some advantages worth to be mentioned. Firstly, the central Ranging Engine keeps the processing work away from the distributed receivers, thus enabling a small low-cost receiver design. A common PC is sufficient for the Ranging Engine process and works best to keep track of multiple connected wireless clients without lacking processing power. Moreover, due to the passive (receive only) behavior of the SMiLE receivers no disturbing wireless traffic is injected, hence preserving the maximal throughput for user data exchange. The additional wiring of the SMiLE receivers to an Ethernet switch is the only extra installation overhead to be afforded. However, this is the only way to achieve high accuracy and to spare the existing infrastructure from extra load. The most important characteristic is, that the system does not require any modification on the wireless devices, neither the WLAN infrastructure (APs) nor any MS. V. R ESULTS System integration itself cannot be quantified in terms of measurements, for this reason we present results of the system. A. Timestamping results Figure 4 shows a histogram of 100 000 TDoA measurements based on the receive timestamps of two BS’s in direct line
5330mm
Rcv1
wsill
Pillar
Sideboard
3450mm
Node 1
Windowsill
Rcv2
5320mm
Windo
2500mm
Rcv3
Rcv4
Fig. 4: Delay histogram of 100 000 free-air measurements
Bookshelf Height = 2300mm
7740mm
B. Localization results The high synchronization accuracy of the receivers is the key for good localization results. A test performed in a conference room showed the maximal possible accuracy of the system. The dimensions of the conference room are shown in figure 5. The SMiLE receivers are placed at the corners of the room at a height of about 1.5 m. As device under test an AP is placed in the center of the room and configured
Fig. 5: Conference room used as test environment
1
1 0.9
3.5
0.8
3
0.7
2.52
0.6
2
0.5 0.4
1.5
0.3 1
0.2
4
0.5 03 0
1
2
3
4
5
6
0.1
Deviation from exact position [m]
Position determination out of 6706 timestamps 4
Y Coordinate [m]
of sight (LOS) to the MS. It should be noted that the histogram not only includes the timestamping errors of the BS’s but also the effect of the synchronization system. In total the system is able to generate timestamps with a standard deviation of about 400 ps. In case the BS’s are positioned 34 m apart, i.e., reducing the SNR significantly, the standard deviation roughly doubles. Compared to Ethernet hardware timestamping for IEEE 1588 the timestamps are very accurate due to the fact that the timestamping is done on the physical layer and not above [8]. In terms of accuracy, the results are comparable to crosscorrelation or super-resolution algorithms as described in [9] and [13]. However, these algorithms operate on the sampled baseband data, which imposes a significant network load on the backbone. As the WLAN signal has a bandwidth of 20 MHz, the data has to be sampled with at least 40 MS/s; with 8 bit per sample every BS produces 40 MB/s. Even if only selected frames are sent to the correlation unit (e.g., 20 %), this may overload a 100 Mbit/s Ethernet line. In addition, finding the correlation peak in presence of carrier and symbol frequency offsets results is a computational extensive task. The timestamp approach inherited from clock synchronization protocols requires to transmit just a single timestamp per WLAN frame, e.g., by a UDP packet. For the example above, if there are 500 Frames/s, the backbone would be loaded with only 32 kB/s/BS and not with 4 MB/s/BS. Currently not implemented, it would be possible to pack multiple timestamps into a single packet increasing the efficiency of the timestamp transmission. The only minor drawback of the timestamp method is that it is bound to the current physical and MAC layer of IEEE 802.11 and is not open to future amendments, whereas cross-correlations methods work independently of the modulation and data formats.
0
X Coordinate [m]
Fig. 6: Conference room localization results
to send out management messages with an interval of 10 ms, which are received by the SMiLE boards and processed by the Ranging Engine accordingly. The scenario is absolutely static, i.e. without MS movement. Also the filter stages are inactive to visualize pure results from the localization stage. In this test the mean deviation from the exact determined position accounted up to 5 cm, the standard deviation is 2 cm. The result is shown in figure 6. An interesting observation is made when people entered the room and shadowed the LOS between one SMiLE receiver and the DUT. As expected, the interference downgrades the results of the TDoA measurements shown in Fig. 7a, but also the recorded RSS values displayed in figure 7b are heavily affected. This fortifies the initial assessment analysis about RSS based localization. It can be observed that the TDoA values of the receiver pairs still having LOS conditions are only exposed to minor influences. The corresponding RSS values behave in an almost identical way. As the two parameters are strongly correlated, this offers the opportunity to include the RSS to detect possible non-LOS conditions. Hence, the receivers contributing to position determination can be dynamically selected based upon RSS changes.
70
60
RCV2
t
D12
40
D13 D14
measured attenuation
Time difference [ns]
3
RCV4
t
20 0 −20 −40 −60
50
RCV1
40 30 20 10
−80 −100 0
RCV
60
t
20
40
60
80
100
120
140
160
Time [s]
(a) Time difference tD1x for each receiver with respect to Rcv1
0 0
20
40
60
80
100
120
140
160
Time [s]
(b) Attenuation over time
Fig. 7: SMiLE receiver measurement results in case of NLOS
VI. C ONCLUSION AND O UTLOOK Indoor localization of WLAN devices without deploying additional localization infrastructure is nearly impossible. The key is to design a system offering high accuracy combined with a simple architecture that does not negatively affect the network infrastructure and can be easily integrated. The system proposed in this paper utilizes the well-known TDoA localization scheme, which implies heavy synchronization requirements for the receiver clocks. This could be achieved using the common clock of an Ethernet switch. The central Ranging Engine calculates positions and provides this information to eligible clients. The architecture of the system is designed in a manner, that integration into an existing IEEE 802.11 network infrastructure is uncomplicated and without modifications on any BS or MS. Future work comprises the integration of IEEE 1588 to ease the offset determination. Also improved localization algorithms and enhancement methodologies are under investigation, especially the incorporation of adaptive Kalman filters [14] and non-linear Kalman filtering [15] is of major interest. The key to localization improvement is the application of a redundant number of receivers to attenuate influences of multipath propagation [16]. To exploit an optimal result of redundant measurements the fusion of RSS information with TDoA measurements is pursued. The investigations show that the RSS value can serve as indicator for changing environmental conditions and it might be utilized to improve the position determination. R EFERENCES [1] IEEE, Std. 802.11-2007 for Information technology-Telecommunications and information exchange between systems-Local and metropolitan area networks-Specific requirements - Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, IEEE Std., June 2007. [2] P. Bahl and V. Padmanabhan, “RADAR: an in-building RF-based user location and tracking system,” in INFOCOM 2000. Nineteenth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE, vol. 2, 2000, pp. 775 –784 vol.2.
[3] C. Rohrig and F. Kunemund, “Mobile robot localization using WLAN signal strengths,” in Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications, 2007. IDAACS 2007. 4th IEEE Workshop on, 6-8 2007, pp. 704 –709. [4] S. Ivanov, E. Nett, and S. Schemmer, “Automatic WLAN localization for industrial automation,” in Factory Communication Systems, 2008. WFCS 2008. IEEE International Workshop on, 21-23 2008, pp. 93 –96. [5] A. Mahmood and G. Gaderer, “Timestamping for IEEE 1588 based clock synchronization in Wireless LAN,” in Precision Clock Synchronization for Measurement, Control and Communication, 2009. ISPCS 2009. International Symposium on, 12-16 2009, pp. 1 –6. [6] C. Rohrig and M. Muller, “Localization of sensor nodes in a wireless sensor network using the nanoLOC TRX transceiver,” in Vehicular Technology Conference, 2009. VTC Spring 2009. IEEE 69th, april 2009, pp. 1 –5. [7] R. Yamasaki, A. Ogino, T. Tamaki, T. Uta, N. Matsuzawa, and T. Kato, “TDoA location system for IEEE 802.11b WLAN,” in Wireless Communications and Networking Conference, 2005 IEEE, vol. 4, 13-17 2005, pp. 2338 – 2343 Vol. 4. [8] R. Exel, G. Gaderer, and P. Loschmidt, “Localisation of wireless LAN nodes using accurate TDoA measurements,” in WCNC 2010: IEEE Wireless Communication & Networking Conference, 2010. [9] J. Jemai and T. Kurner, “An advanced IEEE 802.11b channel sounder based on PN correlation and high resolution algorithms,” in Antennas and Propagation, 2007. EuCAP 2007. The Second European Conference on, 11-16 2007, pp. 1 –6. [10] R. Exel, J. Mad, G. Gaderer, and P. Loschmidt, “A Novel, High-Precision Timestamping Platform for Wireless Networks,” in ETFA’09: The 14th International Conference on Emerging Technologies and Factory Automation, Sep 2009, pp. 1–8. [11] R. Exel and G. Gaderer, “Boundaries of Ethernet Layer 2 Hardware Timestamping,” G. Cena and F. Siminot-Lion, Eds., May 2008, pp. 255– 258. [12] M. Molina, S. Niccolini, and N. G. Duffield, “A comparative experimental study of hash functions applied to packet sampling,” in ITC-19, 2005. [13] F. Zhao, W. Yao, C. Logothetis, and Y. Song, “Comparison of superresolution algorithms for TOA estimation in indoor IEEE 802.11 Wireless LANs,” in Wireless Communications, Networking and Mobile Computing, 2006. WiCOM 2006.International Conference on, 22-24 2006, pp. 1 –5. [14] C. Hide, T. Moore, and M. Smith, “Adaptive kalman filtering algorithms for integrating GPS and low cost INS,” in Position Location and Navigation Symposium, 2004. PLANS 2004, 26-29 2004, pp. 227 – 233. [15] S. Julier and J. Uhlmann, “A new extension of the kalman filter to nonlinear systems,” in Int. Symp. Aerospace/Defense Sensing, Simul. and Controls, Orlando, FL, 1997. [16] Y. Chan and K. Ho, “A simple and efficient estimator for hyperbolic location,” IEEE Transactions on Signal Processing, vol. 42, no. 8, pp. 1905 –1915, aug 1994.