The objective of this paper is to investigate the performance of a wireless sensor network for force monitoring of cable stays. The network consists of a base station, a data logging and configuration unit that represents the data sink, and several remote measurement nodes attached to the cables and representing the data sources. Each node is equipped with an acceleration sensor, a microprocessor and a radio transceiver. Since the nodes are powered by batteries, the network operates as a multi-hop communication network for reducing power consumption. The accelerations of a cable are acquired with a capacitive MEMS sensor. The measured acceleration time series are processed directly in the nodes. Only the estimated natural frequencies are transmitted to the base station. This increases the node lifetime since data compression needs much less energy than data communication. Since the memory of the data processing unit is small, the natural frequencies are extracted from the acquired samples using a simple ARmodel. The cable force is estimated by minimizing the error between measured and computed natural frequencies that are based on an analytical cable model. It considers axial tension force, bending stiffness and cable sag. The wireless sensor network is validated with laboratory tests performed with the model bridge of Empa. The tests reveal that the natural frequencies could be estimated with an accuracy of approximately 1%. Much more critical is the choice of boundary conditions of the cable model. This affects the force estimation by approximately 10%. Since on bridges no quantitative information on the boundary conditions is available, the sensitivity to boundary conditions may produce significant systematic errors, in particular, when any independent measurements of cable tension forces are available for model calibration. energy consumption is distributed evenly among the nodes. Another feature of the network is its capability of self-organisation. If a single node or parts of the network fail, new routes to the destination are established. Therefore, a wireless sensor network is usually operated as a multihop network. When using digital communication technologies, bandwidth is limited. This limitation becomes more evident if multiple nodes are sharing the available bandwidth. For example, when monitoring the vibrations of a civil structure, the upper bound of the interesting frequency range is about 50 Hz. If the signal is sampled with a resolution of 16 bits, the resulting baud rate amounts to 2 kbit/s. If some administration and configuration overhead is added, the rate increases to about 3 kbit/s. This represents the raw data rate per channel. If multiple channels are transmitting data, this rate increases linearly with each added channel. Even more bandwidth is needed if signals have to be sampled at a higher rate or with a higher resolution. Such high transmission rates require too much power and limit the lifetime of battery powered wireless sensor networks to a few hours. Therefore, data compression at the node level is an essential topic of wireless sensor networks. Network nodes operating for several months or even years from batteries have to perform a dramatic data compression in order to communicate as few data as possible. This increase the node lifetime since compressing data needs less energy than communicating data. As a rule of thumb, the compression rate should be at a level that the communication time can be kept around 1% of the desired network lifetime. If more time is spent in the communication mode, energy resources degrade too fast. 1.2 Cable force monitoring of cable stay bridges A potential application of wireless sensor networks is cable tension force monitoring of stay cable bridges using measured natural frequencies. The first investigations of a vibration-based cable tension evaluation were based on simple taut string theory (Kroneberger-Stanton & Hartsough (1992), Casas (1994)). However, this simple theory may cause significant systematic errors if sag and bending stiffness can not be neglected. More advanced identification methods include the effect of bending stiffness by simple approximation formulas of the natural frequencies (Ladret et al. (2002), Geier et al. (2005)). Since the accuracy of approximation formulas diminish with increasing mode order, its use is restricted to cables with very low bending stiffness. In this paper, this limitation is avoided by working with eigenvalue equations instead of approximation formulas. Furthermore, the natural frequencies are usually estimated by using frequency spectra or output only system identification algorithms. However, these methods are usually too expensive in term of CPU memory to be applicable in wireless sensor networks. This paper investigates natural frequencies estimation using a simple autoregressive model requiring significantly less memory.