• Aucun résultat trouvé

3.2 Reliability in IoT

3.2.2 Device Reliability

Typically, an IoT device is a tiny object that contains four main components: (1) a sensing subsystem to sense and collect information from the environment; (2) a processing sub-system to control all device operations; (3) a communication subsub-system to transmit the collected data; (4) a power source to supply the energy needed for all device operations. As we see, the damages in the power source may interrupt the device operations and directly impact the whole IoT system. Unfortunately, the power source is a battery with limited energy capacity. Recharging or replacing such battery is very costly or even impossible, because the devices may be deployed in constrained environments (e.g., under the sewer networks or in the deep forest). In the IoT context, the devices must have a sufficient

3.2. Reliability in IoT 29

lifetime to fulfill the application requirements. Therefore, an energy-efficient mechanism for IoT-device has been received special attention from the research community. There are many proposed approaches, but most of them are inherited from Wireless Sensor Network context. Due to the scope of our thesis, we only review data-driven techniques designed to reduce the number of sampled and transmitted data by the device.

3.2.2.1 Related Work

In the context of Internet of Things, the sensing subsystem may consume more energy than the rest of device elements. This is caused by various factors [163]:

• Power hungry transducers: Many types of sensors use high power resources to per-form sensing tasks, such as multimedia sensors or chemical sensors.

• Power hungry Analog/Digital converter: Some sensors need sensing data from analog to digital format.

• Long acquisition time: Some sensors may even require seconds to performing sensing tasks.

Reducing the energy for communications by selecting low-power communication technology may be not enough. The energy-efficient methods need also decrease the number of data acquisitions (collecting information via sensing subsystem). Thereby, we also reduce the number of communication as well.

The number of sampling could be reduced by exploiting the correlation in the collected data. The authors in [164] use the temporal analysis to build an adaptive sampling for a snow monitoring scenario. They propose an algorithm based on a modifiedCUSUM [165]

that adaptively estimates the current frequency from the trend of historical data. The lim-itation of such approach is too heavy for constrained-resource devices. A similar algorithm is also proposed in [83], which uses a Kalman filter to calculate the sampling rates. Spatial correlation is used in [166] to propose a scheme named “backcasting”. The main idea is that the collecting frequencies of data regions having high variation are higher than the others.

The same idea is exploited in [167]. Another approach named “hierarchical sampling tech-nique” is applied to IoT devices included various sensor types. This method uses the simple sensors with low-energy consumption in normal situations. When abnormal situations are detected, the complex sensors using high-energy is activated to deal with such situation.

This kind of techniques is also called “triggered sampling”. In [168], such technique is used to detect a fire emergency scenario. The environment is monitored by low-cost sensors, such as temperature or C02. When a given area has abnormal states (e.g., the increase in temperature or CO2), high-resolution sensors residing in the area are activated. The similar solutions proposed in [169, 170] to detect the state of structures.

In other way, the data prediction is applied to reduce the number of samples by using machine learning models (e.g., linear model, naive bayes ). The general idea is that instead of using the sensing subsystem to collect the data, we use a well-trained model to predict such data. The data acquisition is only triggered at low frequency to ensure the model is well-updated. The solution in [171] uses a probability density function as a base model to forecast the next values. The authors of [172] following the same way, but they use a Kalman filter as the prediction model. In [173], a Dynamic Probabilistic Model (DPM) is used to implement a probabilistic view of available sampled data. For the time series data, the typical models, such as Moving Average (MA), Auto-Regressive (AR), and an

Auto-Regressive Moving Average (ARMA) are used as prediction models. PAQ [174] uses an AR model aiming to reduce the computation by the devices. The authors in [175] enrich the AR model to deal with inconsistent data and outliers.

3.2.2.2 Open Challenges

• Multiple dimensions: Adaptive sampling is a promised technique to achieve high energy-efficient. However, most of the proposed solutions only optimize the frequency based on a single characterization like in time or space. Thus, we need a comprehensive solution that could combine both time and space to exploit multiple information at the same time.

• Complexity: To effectively reduce power consumption, the reviewed approaches pro-cess a large amount of data. For example, the data prediction approaches require data and device resources to train the prediction model. The trigger sampling ap-proaches need complex methods and historical data to identify the abnormal situa-tions. However, storing and processing these data on the device side is ineffective for constrained-resource devices. They have to distribute such computation on a device network or cloud instead of a single device. This could emerge the issues related to communication (e.g., How to minimize the cost for distributing data). Therefore, we need a method that is light-weigh and easily deploys on constrained devices while ensuring the high energy-efficiency.

• Flexibility: Most of the reviewed algorithm are highly designed for specific purposes.

For example, triggered sampling approach more targets in data accuracy than power-saving while prediction approach is contrary. There is no generic algorithm that could handle user’s interests related to the power-saving degree. Depending on the user-cases, users could define either data accuracy or power-saving be focused.

Part II

Interoperation in IoT