Light-weight Online Predictive Data Aggregation for Wireless Sensor Networks Jeremiah D. Deng
Yue Zhang
Department of Information Science University of Otago Dunedin, New Zealand
Department of Information Science University of Otago Dunedin, New Zealand
[email protected]
[email protected]
ABSTRACT
and challenging to both the data mining and networking aspects of WSN-based applications. In particular, data transmission itself consumes significant amounts of energy in sensor nodes, which normally operate on limited power supply from batteries subject to physical constraints (such as size and weight). While there have been proposals on WSN deployment using solar charging [15] or wireless power transfer [16], these have their own constraints such as the availability of mobile charging stations, or good weather. So far, energy consumption reduction in WSNs remains an important research topic. From the perspective of networking protocol stacks such as TCP/IP, energy consumption obviously relates to multiple layers. It is clearly not just limited to operational costs in medium access control (MAC) contention, routing, or TCP retransmission. In fact, the redundancy within the sensory data can be exploited by machine learning techniques, which can contribute to significant transmission reduction. There has been some work on so called compressive sensing, employing distributed prediction on sensory data [2]. Some of these approaches, also termed “data aggregation”, however require large historical data storage or employ algorithms of large complexity [8][9]. In this paper, we argue that the algorithmic overhead in sensory data prediction needs to be considered and propose a light-weight scheme that utilizes a shift-register based implementation and employs a very short data history for effective online data stream prediction. A number of simple prediction algorithms are evaluated using a real-world dataset against several performance metrics. The results are very promising, reducing data transmission down to less than 10% for temperature monitoring. The remainder of the paper is organized as follows. In Section 2 we briefly review some related work and justify our proposed approach. The predictive sensing scheme is then presented in Section 3, along with a brief introduction to the prediction algorithms employed, and the performance indices to be used in assessment. The experiment results are presented in Section 4. Finally we conclude the paper with a short discussion on future research.
Wireless Sensor Networks (WSNs) have found many practical applications in recent years. Apart from both the vast new opportunities and challenges raised by the availability of large amounts of sensory data, energy conservation remains a challenging research topic that demands intelligent solutions. Various data aggregation techniques have been proposed in the literature, but the optimal tradeoff between algorithm complexity and prediction ability remains elusive. In this paper we concentrate on employing a few light-weight time series estimation algorithms for online predictive sensing. A number of performance metrics are proposed and employed to examine the effectiveness of the scheme using real-world datasets.
Categories and Subject Descriptors C.2.1 [Network Architecture and Design]: Wireless communication; I.5.4 [Pattern Recognition]: Applications | Waveform analysis
Keywords wireless sensor networks, data aggregation, time series prediction
1.
INTRODUCTION
Wireless Sensor Networks (WSNs) are becoming more and more important and widely utilized in many civil and military applications, such as environment monitoring, surveillance, healthcare, construction and transportations. Paired with the advances on sensing techniques in many domains, observation data obtained from a variety of phenomena can now be effectively collected and transmitted by radio. The increasing amount of sensory data that are being gathered and transmitted, even potentially in real time from the physical world, is at the same time enabling
2.
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. MLSDA ’13, December 3, 2013, Dunedin, New Zealand Copyright 2013 ACM 978-1-4503-2513-4 ...$15.00.
RELATED WORK
Communication overhead has been identified as a major issue for data transmission in WSNs. Data transmission can be directly between the sensing nodes and the base station (sink), which normally demands a large amount of energy per transmission because of the long transmission distance. Self-organized approaches such as using clusters [7] can
35
help to reduce the transmission overhead by allowing sensor nodes to form into clusters where transmission responsibility is shared. Spatial redundancy within the sensory data gathered from a cluster can be also utilized, so data can be aggregated before the transmission. Naturally, the temporal redundancy of the sensory data stream also exists and this can be exploited for transmission reduction as well. A simple, straightforward approach would be to reduce the sampling rate and simply transmit selectively. For tasks that require continuous monitoring and reporting, e.g., room temperature or electricity voltages, however, it becomes necessary to employ time series prediction methods on sensor nodes so as to avoid costly transmissions whenever the nodes or the sink can predict the next sensory data value accurately. A vast range of time series modeling and prediction methods can be used for this purpose, ranging from a lazy approach of copying the last stream entry [13][1], to heavyweight algorithms such as support vector machines [8]. In between, the Auto Regressive Integrated Moving Average (ARIMA) estimator has been used in a number of studies [10][9]. ARIMA is typically pre-built, but can be made adaptive and online as in [9]. However, due to its inherent offline nature, it requires a considerable of historical data to train the model, which also takes time. As shown in [9], schemes employing ARIMA often require a large historical database, with the window sizes ranged between 15 and 60, for instance. On the other hand, adaptive algorithms that do not require prior knowledge or a long sequence of historical data have been proposed for predictive sensing in WSNs. In [13], impressive performance is achieved by an adaptive scheme based on the Least Mean Squares (LMS) algorithm, for which both an online mode and an offline mode are proposed. Even the lazy estimator, as shown in [4], can be employed for a dual prediction scheme along with auto-regressive models. In [17], an interesting approach is proposed for implementing both temporal and spatial sensing data correlations. Its core prediction algorithm is, however, quite simple, employing a moving average estimator with fixed weights. Other studies look at aggregating sensor data streams both temporally and spatially, and exploit the similarity between data streams from different sensors so that similar streams get aggressively aggregated [5]. These would assume some kind of dense deployment of sensor nodes [3] so as to gain the redundancy, and require close collaboration between neighbor sensor nodes. For a review of the stateof-the-art stream clustering algorithms, see [14]. Our approach differs from previous studies in that we consider only light-weight implementations that require little memory and computational overheads, and we are interested in investigating online predictive sensing schemes only, as we believe that is what real-world environmental data streams normally demand. To assess how well these sensing schemes can deal with real-world data streams, we look at the predictive accuracy as well as the sensing accuracy when predictive sensing schemes are employed.
3.
predictive data aggregation is shown in Figure 1. At the core of the process is a light-weight estimator that uses a short-length shift register to provide data input for online prediction. The prediction outcome will be compared with the true value of the next measurement, which generates feedback into the shift register. Therefore the shift register behaves like a “sliding window” on the predictive sensing output sequence, which provides the historical information for prediction. It differs from traditional prediction schemes, as the sliding window operates not on the original data stream, but on the predictive sequence with corrections. Note there is a tradeoff to be met between prediction ability and computing complexity. The more accurate the prediction can be, the fewer transmissions are needed; on the other hand, better prediction demands more complicated algorithms, potentially adding to energy consumption. As wireless sensor nodes are normally subject to stringent energy constraints, the choice of the prediction algorithms is limited. Since sensor memory is also limited, the algorithms have to operate in an online mode using a short sliding window. Here we consider a few simple approaches that are promising for solving this problem. Let us denote the sensor data stream as x = (x[1], x[2], · · · , x[n], · · · , x[N ]). The task is to use a sliding window of size M , X = (x[n − M − 1], · · · , x[n − 1], x[n]) to predict the next measurement x[n + 1]: x0 [n + 1] = pred(X),
where pred(.) denotes an estimator, and the sensor data aggregation operates depending on whether the prediction is within an error threshold θ: ( x0 [n + 1], if |x0 [n + 1] − x[n + 1]| < θ d[n + 1] = (2) x[n + 1], otherwise. Here d[n] denotes the predictive reading. The idea is that if the prediction is reasonably accurate, other nodes or the base station do not need to receive the source node, but instead use the prediction as the “reading”. For this reason, both the source and the sink should be synchronized and use the same sliding window X for prediction, which now consists of historical values of sequence d[n] rather than x[n]. Bad predictions will be synchronized and the sensing sequence is updated as sort of a correction, bringing d[n + 1] back to the groundtruth x[n + 1]. Threshold θ will be a decisive factor for limiting transmissions. This can be application dependent and set a priori by users. Fortunately, with the help of prediction algorithms, the actual error is usually less than the given threshold, as shown in experiment results later. Since the predictions can be “corrected” by the actual measurement at the source node, the accuracy of the sensing scheme is not simply reflected by the prediction errors. This correction, however, comes with a price as it has transmission involved. Therefore we proposed the following three indices to assess the predictive sensing scheme under different estimation algorithms: 1. Prediction error, i.e., the average prediction error against the groundtruth across all rounds: PN −M 0 n=1 |x[n] − x [n]| . (3) PredErr = N −M
PREDICTIVE DATA AGGREGATION
3.1
(1)
Overview
The diagram of the proposed sensing scheme using
36
Shift Register
x[.]
D
......
D
d[n]
d[n-1]
D d[n-M+1]
Estimator x’[n+1]
d=x
N
|x-x’|
