Synonyms

Clustering; Data aggregation; Machine learning; Routing; Wireless sensor Networks

Definition

A Wireless Sensor Network (WSN) consists of spatially distributed autonomous devices using sensors to monitor physical or environmental conditions. A WSN system incorporates a gateway that provides wireless connectivity back to the wired world and distributed nodes. Machine learning (ML) is the science of getting computers to learn and act like humans do and improve their learning over time in autonomous fashion, by feeding those data and information in the form of observations and real-world interactions.

Introduction

Wireless Sensor Networks (WSN) is a vital component in Internet of Things (IoT). The small sized, low powered sensors are capable of monitoring and collecting data from environment. Most of the recent research works have not concentrated to provide a solution for analyzing the potentially huge amount of data generated by these sensor nodes. Thus, there is a need for machine learning (ML) algorithms in WSNs. When volume, velocity, and variety of data generated by WSN is very high, then data analytics tools are needed for data aggregation and clustering. ML tools are used in several applications such as intrusion detection, target tracking, healthcare, home automation, smart city. The main aim of this entry is to provide a basic knowledge in machine learning and its applications in WSN.

Background

Generally, the designers of sensor network symbolize machine learning as a branch of artificial intelligence and it is a collection of algorithms that is capable of creating prediction models. On the other hand, ML experts characterize it as a field, which is having huge amount of patterns and themes useful in sensor network applications (Alsheikh et al. 2014).

Supervised Learning

Supervised learning is merely a formulation of the concept of learning from examples. Supervised learning approach is used to resolve various issues for WSNs such as event detection, objects targeting, localization (Shareef et al. 2008), processing of query, Medium Access Control (MAC), intrusion detection, security, data integrity, and QoS.

Unsupervised Learning

No output vectors or labels are provided in unsupervised learning. Data sets are classified by finding the similarities among them. Unsupervised learning algorithm determines the concealed relationships and it can be used for solving the problems in WSN, where relationship between the variable is complex. These types of algorithms are mainly used for clustering and collection of data in WSN.

Reinforcement Learning (RL)

It allows the agent to learn from the environment by interacting with it. Here, sensor nodes learn to capture the best measurement in order to maximize the advantage. The most famous reinforcement learning algorithm is Q-learning, in which every node tries to extract measurements which are expected to increase the rewards. Sensor nodes regularly update the bonuses it has achieved derived from the actions taken at a given state. Total rewards of future can be computed by the equation:

$$\begin{aligned} & Q\left({s}_t+1,{\alpha}_t+1\right)\\ &\quad=Q\left({s}_t,{\alpha}_t\right) +\gamma \left(\mathrm{r}\left({s}_t,{\alpha}_t\right)-Q\left({s}_t,{\alpha}_t\right)\right)\end{aligned} $$
(1)

(,) indicates the incentive for taking the action \({\mathit{a}}_{\mathit{t}}\) at state \({\mathit{s}}_{\mathit{t}}\) and \(\mathit{\gamma }\) is the rate of learning, which decides how frequently the learning happens (between the values 0 and 1).

Machine Learning for Localization

A sensor node recognizes the position either by using GPS device or by any specific localization methods. One of such method is to collect the knowledge (e.g., connectivity, pair-wise distance measure) on the subject of the whole network into single place, where the composed information is managed in a centralized manner and identify the nodes’ locations using mathematical algorithms.

Security and localization are the two main applications of Support Vector Machine in WSN. The localization method (Tran and Nguyen 2008) for mobile node uses SVM and knowledge about connectivity capabilities. By using RSSI metric (Received signal strength indicator), it detects the node location. Despite the fact that Localized SVM is capable for distributed localization in a speedy manner, it is still susceptible to outliers in training data set.

A localization system based on Self-Organizing Map (SOM) (Paladina and Paone 2007) provides a few artificial intelligence (AI) features to sensor nodes. Without any supervision, SOM can able to learn how to classify. Self-Organizing Map is used on each sensor nodes in order to evaluate the position of the node. Eight anchor nodes spatial coordinates surrounding the unspecified node form the input layer. Output layer gives the special coordinates in 2D space of the unspecified node after the training. Demerit of this method is that nodes should be organized consistently throughout the monitoring area.

An improved localization using Support Vector Regression (SVR) (Wang et al. 2016) discusses about a novel extraction method of training data for localization model. It improves the accuracy of localization method without affecting the hardware cost. The location of the unknown node is found accurately using SVM-based learning with minimum number of anchor nodes (Mary Livinsa and Jayashri 2015). Accuracy is provided by finite size of grid cells. For larger scale networks, SVM-based DV-hop is suitable.

Machine Learning for Clustering and Data Aggregation

It is really difficult to transfer entire data to sink instantly in a large scale energy-constrained sensor network. An effective solution for this problem is to send the data to an aggregator node, which is also called as cluster node. The node aggregates information from all members of the cluster and sends to the sink or base station, thus reduces the energy consumption. Numerous algorithms are actually suggested for the election of Cluster Head to increase the energy efficiency. The machine learning dependent strategies can develop the advantages of clustering and aggregation of data between nodes in WSNs through several procedures.

  • ML algorithm is used to identify and remove the nonfunctional nodes from routing schemes and compress data at CH using dimensionality reduction method.

  • Machine learning algorithms are executed for electing the Cluster Head efficiently, where election of suitable cluster head will considerably decrease the usage of energy and augment the network’s lifespan.

The CH election problem can be solved by the application of Decision Tree algorithm (Ahmed et al. 2010) (DT). By considering prominent features as battery, distance, and mobility, DT is electing suitable cluster head. By estimating these features, DT can provide a proficient method for identifying link reliability in Wireless Sensor Networks. Role-Free Clustering (Forster and Murphy 2010) is a clustering method with Q-learning for WSNs. The ability of the cluster head is evaluated by Q-learning algorithm with active network parameters.

A distributed spectral clustering method to group sensor nodes on their location is proposed (Muniraju et al. 2017) to avoid data congestion. This is the combination of two algorithms, distributed eigen vector computation and distributed K-Means clustering. Eigen vector of graph Laplacian is calculated by distributed power iteration technique. Clustering on eigen vector is done by distributed K-Means algorithm. In order to establish the network topology, only the location information of sensor nodes is used and this information is not exchanged in the network. In (Park et al. 2013), energy efficiency is maximized by finding cluster head with k means algorithm.

The Self-Organizing Map (SOM) is unsupervised algorithm method that learns to map from high to low dimensional space. In Cluster-based self-Organizing Data Aggregation (CODA) architecture (Lee and Chung 2006), the nodes classify the aggregated data j∗ as the winning neuron, which is having a weight W(t) nearest to the value of input vector X(t). With CODA scheme, quality of data is increased and energy consumption and network traffic are reduced.

Principal Component Analysis (PCA) algorithm is used for dimensionality reduction and famous in the field of data compression. Principle components are a set of orthogonal variables. Main aim is to select the significant information from data concerning the principal components. Both data compression and dimensionality reduction are multivariate methods. This can reduce the amount of data transmission between the nodes in WSN. By selecting only the momentous principal components and tossing out the lower order components, it solves the big data problem into small data. For improving the process of data aggregation, the combination of following two significant algorithms is used along with Principal Component Analysis (PCA) in WSNs (Morell et al. 2016).

  • Compressive Sensing (CS): The conventional scheme of “sample then compress” is recently replaced by CS. Compressive Sensing uses sparsity property of signal.

  • Expectation-Maximization (EM): This algorithm comprises of two stages, an expectation stage and maximization stage. While performing its expectation (E) stage, EM evaluates the cost function by saving the current expectation of parameters. In the maximization (M) stage, it re-computes parameters that can minimize the estimation error.

Adaptive Learning Vector Quantization (ALVQ) is used to extract compressed model of information from the nodes accurately (Lin et al. 2009). ALVQ uses the LVQ learning algorithm with previous training data for predicting the code-book. This method reduces the needed bandwidth while transmitting data and improves the accuracy of correct reading restoration from the compressed information.

Machine Learning for Routing

Considering features such as memory and computational requirements, communication costs, wireless ad-hoc nature, restricted energy, mobility and topology changes, different ML algorithms are used for energy-efficient routing in WSNs.

In fuzzy logic-based routing method (Arabi 2010), the entire network is grouped into different clusters and the election of suitable cluster head is based on fuzzy variables. The routing in WSNs depends on the combination of hybrid routing methods and fuzzy logic for improved energy savings and improved network life span. Energy and Delay Efficient Routing protocol for sensor network (EDEAR) (Sharma and Shukla 2012) is adaptive routing method (Kumar and Kumar 2010) using Reinforcement Learning (RL). It finds the best path with minimum energy consumption and end to end delay. RL renews routing tables and thus considers all the dynamic parameters that characterize the traffic. The adaptation of routing depends on the varying traffic conditions and hence minimizes data transfer time.

Energy-aware QoS Routing algorithm using Reinforcement Learning (EQR-RL) (Jafarzadeh and Moghaddam 2014) is able to balance QoS requirements and improve network life time. Using a random load balancing algorithm, next hop neighbor is chosen. QoS parameters such as number of hops, latency, and geographical distance are considered. Dynamics of the network is learned using Q-Learning technique and routing decision is made accordingly; however, the network state information is not maintained. Sensor Intelligence Routing (SIR) using Self-Organizing Map (Barbancho et al. 2008) detects optimal routing path. The combination of SOM and Dijkstra’s algorithm model offers QoS guarantees like throughput, latency, duty cycle, and packet error rate.

Example Application Scenarios

Regression-Based Adaptive Incremental Algorithm for Health Abnormality Prediction (RBAIL)

This method employs a regression based incremental algorithm for providing the learning features (Srinivasan et al. 2013). The incremental learning system is wirelessly connected to the patient and receives the stream of input parameters from patient for a fixed time of interval. RBAIL algorithm performs regression on the important health parameters for predicting the indiscretion of patient. System uses history of the patient to check whether previous anomalies were there in order to get the updates and feedbacks. Main features here are aggregation, learning, and prediction. Correctness of the parameter is verified during aggregation. If previous data and current input data are valid and parameter value is greater than a threshold value, then abnormality is detected. If the difference of current and predicted value becomes greater than a threshold, then doctor provides feedback to correct the learning algorithm.

Object Detection and Tracking in Wireless Sensor Networks

Prediction logic is used to predict the exact location of the sink node using current location (Vaidehi et al. 2011). The estimated position is sent to Cluster Head to wake up the node which is in sleep mode. The combination of sleep wake scheduling, clustering, tracking, prediction logic, and shortest path routing minimize the energy consumption in sensor networks. Sink nodes awake cluster head that helps to reach target. Further complex events processing engine is used for detecting abnormal events in a multisensor scenario (Bhargavi and Vaidehi 2013; Gao et al. 2012).

Semantic Intrusion Detection System by Means of Pattern Matching and State Transition Analysis (SIDS)

Semantics Intrusion Detection System (SIDS) (Sri Ganseh et al. 2011) combines pattern matching, state transition, and data mining for increasing the accuracy of intrusion detection. Multiple sensors are deployed in the sensor area. The events generated by sensors are correlated in time spatial domain. The outputs from the sensors are symbolized as patterns and states. When the patterns generated by sensors violate the rule, it is detected as an intrusion. Semantics rules are developed using Another Tool for Language Recognition (ANTLR).

Online Incremental Learning Algorithm for Anomaly Detection and Prediction in Health Care

An Online Incremental Learning Algorithm (OILA) is proposed (Kirthana and Bhargavi 2014) for processing the data in online. It uses the combination of regression and feedback mechanism in order to decrease the prediction error and hence improves accuracy. The vital health parameters are received from the body sensors of a patient. Online Incremental Algorithm evaluates several parameters based on the received data and checks whether any anomalies are found. An alert is set to the doctor, if any anomalies are detected. Regression-based method is used to predict next instance. Prediction of each patient is personalized according to his/her health parameters. This algorithm calculates overall trend by longvalue and recent trend by shortvalue in health parameters. The parameters maxthresh and minthresh captures maximum and minimum threshold value of tolerance. Difference between maxthresh and minthresh is captured by a parameter diffthresh. Patient sensitivity range can be defined through sensitivity range parameter by the doctor. History factor is a parameter that defines number of times a patient affected to abnormalities. After reading every new instance, these parameters are updated, error is adjusted, and according to that prediction is made. The algorithm predicts the abnormality using updated parameters and triggers alert.

A Genetic Approach for Personalized Healthcare

Genetic Algorithm-based Personalized Healthcare System (GAPHS) (Vaidehi et al. 2015), uses a sensor integrated wearable chest strap for the non-invasive monitoring of physiological parameters and body parameters. Wrist wear wireless Blood Pressure (BP) sensor is used for monitoring blood pressure. A fingertip wearable oxygen saturation level (SPO2) sensor is used to detect blood oxygen saturation level. The abnormality levels of the vital parameters are classified into very low (VL), low (L), medium (M), high (H), and very high (VH) and encoded into a 5-bit representation to determine the severity level of the patient. Using fitting function, the best chromosome that represents the personalized vital parameter of the patient is obtained. The proposed GAPHS provides an intelligent, personalized, and efficient healthcare system to serve the needy patient in right time by the doctor.

Dynamic Higher Level Learning Radial Basis Function for Healthcare Application

Traditional Radial Basis Function (RBF) has issues with using complete training set and large number of neurons. Due to these issues, computation time and complexity are increased. Dynamic Higher Level Learning RBF (DHLRBF) (Chandraskar et al. 2014) is applied to health parameters to find normal and abnormal category. The DHLRBF uses both cognitive and higher level learning component for effective classification with less complexity.

Sensor Based Decision Making Inference System for Remote Health Monitoring

Most of the existing methods have difficulty to differentiate between original and fall like patterns. Intelligent Modeling technique, Adaptive Neuro-Fuzzy Inference System (ANFIS) (Dhivya Poorani et al. 2012) is used for detecting the fall automatically with higher accuracy and less complexity. The data received form 3 axis accelerometer is categorized into five states (sit, stand, walk, lie, and fall) using ANFIS model. Mean, median, and standard deviation are selected for training the neural network. When the state is detected as fall, it examines ECG and heart rate of patient to check the abnormal condition and raise alarm.

Future Direction

The new revolution, Internet of Things is rapidly gathering momentum driven by advances in Sensor Networks and cloud technologies. The cloud has to provide security for sensed data. The storage for sensed data needs to be minimized by compression schemes. Multisensor data fusion is used for situation, object and threat refinement. The complex events sensed by sensor network can be analyzed with data analytics tools to find the trends, patterns, and gain new insights and knowledge for variety of applications.

Cross-References