Dynamic sampling rate algorithm (DSRA) implemented in self-adaptive software architecture: a way to reduce the energy consumption of wireless sensors through event-based sampling

In the maintenance industry, condition monitoring can be defined as a technique used to monitor physical variables (e.g., temperature, vibration, and pressure) to draw conclusions about the condition of an asset being monitored. Condition monitoring techniques are used as a data source in condition-based maintenance to provide an early warning to plan and trigger cost-effective maintenance actions (Owen et al. 2009; Al-Najjar 2012). With the recent digitalization trends in industry, wireless sensors are, in particular, gaining a growing interest. This is due to the possibility of installing them in inaccessible locations for wired sensors, such as in embedded systems and rotating machines. This avoids the burden of unreliable electrical connections and wiring expenses (Owen et al. 2009).

Although great success has already been achieved in this area, energy limitation remains a major obstacle for further advances (Alippi and Anastasi 2010; Zhang et al. 2013). For this reason, energy harvesting for wireless sensors received considerable attention from several researchers (Owen et al. 2009; Bogue 2010, 2015). However, harvested energy sources are typically unstable and vary with time, weather, or the season. As such, managing energy consumption wisely is still desirable (Yan et al. 2012; Zhang et al. 2013).

Generally, it is assumed that the energy consumption for sensing is relatively far lower than it is for transmitting and receiving. However, in practice, this is not always the case. There are so-called “hungry sensors” (e.g., a gas sensor) whose energy consumption is relatively high compared to that of data transmitting and receiving (Alippi and Anastasi 2010; Shu et al. 2017). In such applications, efficient sampling management is a very important issue as reducing data sampling means, in many cases, reducing data acquisition, including computations and data transmission.

For wireless sensors, there is usually a trade-off between monitoring quality and energy consumption. This is because with the increase in collected data comes increased monitoring quality and increased energy consumption. However, if only a little data is collected, there is a lack of information, which eventually, will come with the cost of poor monitoring quality and consequent bad decisions (Lu et al. 2017). Therefore, it is important to optimize the sampling with a sufficient rate to catch important information cost-effectively and without excessive energy consumption. One way to achieve sufficient sampling is to adapt the sampling rate dynamically to the events that occur in the physical phenomenon being monitored.

Several studies investigated adaptive sampling. For instance, in Shu et al. (2017), an algorithm for adaptive sampling was developed and tested for water quality. Two key parameters were used to evaluate the algorithm, dissolved oxygen (DO) and turbidity. Normalized mean error was used to evaluate the performance of those predetermined parameters. In comparison with fixed intervals, the study showed that battery life could be increased to 30.66% within 3 months of monitoring. In Alippi and Anastasi (2010), an adaptive sampling algorithm was proposed that estimates the optimal sampling rate online. The proposed algorithm determines the sampling rate by using the fast fourier transform algorithm and compares it with the maximum frequency. The study showed that the algorithm could reduce the number of samples up to 79% compared with the traditional fixed-rate approach. Yan et al. (2012) suggested an adaptive sampling algorithm that adapts the sampling mode to the present energy state. The test was made using a CO₂ sensor, and the results showed that the sensor sampling mode was changing based on the changes in the energy state. More detailed surveys in this domain can be found in Rault et al. (2014) as well as Khan et al. (2015).

In general, an adaptive sampling technique is computational-dependent and, therefore, software plays a significant role in it. For this reason, a reference framework of software architecture is important to facilitate adaptive sampling design, implementation and modeling. Despite its importance, there is a lack of studies on developing a reference framework for adaptive sampling models. From the software engineering field, the architectural approach appears to be suitable for this task. It provides the right level of abstraction and generality to construct a knowledge representation scheme (Oreizy et al. 1998; Garlan et al. 2004; Kramer and Magee 2007; Weyns and Iftikhar 2016; Algabroun 2017). Garlan et al. (2004) proposed a self-adaptive based architecture, named Rainbow, that can be used as a reference for systems that are self-adaptable to uncertainties at the runtime. In this paper, we argue that this reference architecture is suitable for adaptive sampling applications.

The contribution of this paper is twofold: first, proposing a reference architecture, that is, Rainbow, for adaptive sampling algorithms, and second, developing a parametric algorithm, that is, dynamic sampling rate algorithm (DSRA), that dynamically adapts the sampling rate based on the events in the signal. Hence, the aims of this study are as follows: (1) to establish a reference framework for adaptive sampling algorithms, (2) to develop an adaptive sampling algorithm, and (3) to implement and test the proposed reference framework using the developed algorithm.

The remainder of the paper will be as follows: the next section will discuss the self-adaptive software architecture; Sect. 3 will explain the development of the algorithm, which will then be implemented in the proposed software architecture; in Sect. 4 the algorithm will be tested using two datasets; Sect. 5 will discuss the results; and in Sect. 6 conclusions will be drawn.

Self-adaptive software architecture is designed to endow a system with autonomous adaptation to dynamics at runtime (Kephart and Chess 2003; Gil De La Iglesia and Weyns 2015). Typically, this approach contains a managed system and a feedback loop, that is, a managing system. The managed system is concerned with the domain’s functionality, while the managing system is concerned with the adaptation of the managed system (Oreizy et al. 1998; Kramer and Magee 2007).

This approach has several advantages (Algabroun 2017):

Several frameworks have been proposed for self-adaptive software architectures (Oreizy et al. 1998; Kephart and Chess 2003; Garlan et al. 2004; Kramer and Magee 2007). Some studies, such as IBM’s MAPE-K (monitor-analyze-plan-execute-knowledge) (Kephart and Chess 2003), considered developing a framework that guides the self-adaptation mechanism. Kramer and Magee (2007) considered three different hierarchical layers, which named component control, change management, and goal management. Others, such as Oreizy et al. (1998) and Garlan et al. (2004), provided a framework that enables assessment of the adaptation decision and allows runtime configuration. All of these frameworks aimed to provide leeway for the engineer to architect a software system. As such, it is nearly possible to reach rather a similar objective using any of them (Weyns et al. 2012). Despite this, the Rainbow framework appears to be the most suitable for the problem addressed in this paper, as it contains the necessary elements for conducting an adaptive sampling operation. These elements are collect data, detect changes in the condition, evaluate changes, decide the adaptation regime (i.e., the sampling rate), and execute the regime. Therefore, in this study we considered Rainbow architecture (Garlan et al. 2004). The next section will describe the Rainbow architecture in more detail.

Sufficient sampling is very important, as once the obtained signal is distorted from insufficient sampling, it will be impossible to retrieve the original signal anymore. In a sampled signal, the pattern between the sampling points is unknown. From a practical point of view, the signal is usually assumed to have a smooth transition between these points (Cohen 2019).

The algorithm developed in this section is based on a realistic assumption to capture the signal changes; that is, we need more data when a change in the monitored phenomenon occurs. So, the sampling rate should be dependent on the change rate in the monitored signal. This motivates the need to consider the slope between two consecutive points. The more slope there is, the more change that occurred and the more following-up there should be; consequently, the higher sampling rate should be implemented. Based on this assumption, this algorithm is dealing with two main parts, that is, identifying the slope and determining the sampling rate.

Two datasets were created and used as examples to test the DSRA algorithm. The first example is intended to visualize the behavior of the algorithm; the second example is intended to show the approximation that could be achieved in a more realistic dataset.

Based on the algorithm implementation and the results of the examples, the Rainbow software architecture provided a knowledge scheme for adaptive sampling to follow and use. It follows systematic steps to perform adaptive sampling that are: data collection, detection of a change in the signal being monitored, evaluation of the change significance, selection of the adaptation strategy, and eventually, execution of the selected strategy. There was no conflict or illogical process in the flow of the Rainbow software architecture to be reported.

Figure 2 shows that the algorithm sampling increases when the slope increases. When the slope value is 0, the total number of samples is ~ 7. When the slope value is 1, the total number of samples is 9, and when the slope is 2, the total number of samples is 28. This is when E and S are 5 and 2, respectively.

In Fig. 3, when there is a redundancy in the signal (i.e., the amplitude value of the signal is stable, and there is no variation in its value), fewer samples were collected. When there were variations in the test signal, higher sampling was exhibited to capture the changes in the signal. The reconstructed signal from DSRA appears to have a pattern close to that of the test signal with MAPE = 0.326% for 300 points. This is when E and S are 12 and 15, respectively. It is crucial to assign a proper value for E and S to achieve efficient sampling with a good approximation to the signal being monitored. The optimization strategies and techniques for tuning E and S are out of the scope of this study and will be the subject of future work.

This approach might not be suitable for highly fluctuating signals (e.g., high noise signals, vibration signals), and in this case, a further process for this signal is required (e.g., averaging, CUSUM). However, further research is still required to investigate its performance on such signals.

In order to overcome the energy limitation in wireless sensor, it is important to optimize sensor sampling to catch only important information without excessive samplings that cause unnecessary energy consumption. A way to achieve this is by adapting the sampling rate to the events that occur in the signal being monitored.

As software plays an important role in the techniques of adaptive sampling, a reference framework for software architecture is important in order to facilitate their design, modeling, and implementation.

The software reference framework for adaptive sampling should allow one to address the following three key questions: is there a change in the signal, is it considerable change, and what adaptation strategy should be conducted?

The Rainbow software architecture appears to have a mechanism that answers those questions, that is, after collecting data, Model manager component detects the change in the signal, Constraint evaluator component determines the relevance of the change, and Adaptation engine component selects the suitable adaptation strategies that are then executed by Adaptation executor component. It is believed that the proposed architecture has the general steps of adaptive sampling and, as such, can be generalized and used as a reference software architecture for a self-adaptive sampling algorithm.

An algorithm for adaptive sampling, namely DSRA, was developed in this study and implemented in Rainbow architecture. As proof of concept, this algorithm was implemented in two examples, and the results prove that the algorithm has adaptive sampling based on the changes in the signal. However, this algorithm might not be suitable for signals with high fluctuation, in such a case, further signal processing, such as averaging and CUSUM, might be necessary. The applicability of DSRA on such a signal still needs to be investigated. Future work could include investigating optimization strategies for E and S parameters and techniques to include multiple sensor signals in the algorithm, as well as examining the Rainbow software architecture with other sampling algorithms.