Research On Anomaly Detection Models For Time Series Data In Wireless Sensor Networks

Due to the dimensional complexity of the data,the analysis of agricultural production data was quite tricky.The use of Wireless Sensor Networks(WSNs)was steadily increasing to cover various applications and domains.This trend is supported by the technical advancements in the sensor manufacturing process which allow a considerable reduction in the cost and size of these components.However,several challenges are facing the deployment and the good functioning of this type of networks.Indeed,WSNs applications have to deal with the limited energy,memory and processing capacities of sensor nodes as well as the imperfection of the explored data.This dissertation addresses the problem of detecting anomalies in WSNs.Throughout this work,we are providing different solutions that allow us to meet these requirements.First,we have presented the existing research on the treatment of data anomalies.Depending on the field of technology,unexpected patterns are named exceptions,outliers,or flaws.Anomalies were occurrences of data that did not correspond to a well-defined standard behaviour.According to Grubbs,an aberrant or "outlier" phenomenon occurs to deviate significantly from other parts of the sample in which it existed.Many algorithms have been used to detect abnormal or outliers.We implemented some outlier detection techniques that have been proposed with the fundamental concepts to properly handle multidensity and multi-granularity density-based approaches,such as Local Correlation Integral(LOCI)approach,self-organized map(SOM),Support Vector Regression(SVR),Local Outlier Factor(LOF),and Angle Based Outlier Detection(ABOD),discussing their advantages and disadvantages,which helped us to better understand our research on data anomaly detection.To identify trends in the data that do not correspond to the expected behaviour known as failure and anomaly,we proposed and implemented a new approach for detecting outliers based on time series data in WSN.Our approach differed from current model-based solutions,including various time series sensor data(temperature,humidity,CO2,NH3 and NH2S).It associated the autoencoder with the k nearest neighbours denoted KNN using the Mahalanobis distance and the 3-sigma rule.The automatic encoder used to denoise the dataset in order to extract the data.In contrast,KNN used Mahalanobis to calculate the distance between values.The three-sigma was calculated as the threshold to find the abnormal after the prediction process.Our proposed model had advantages,such as the elimination of noise on datasets and less error detection.Moreover,the proposed model provided a technique of handling missing values,in particular,it solved the corruption of the data collection and eliminated the irregularities in the typical data structure.The experimental results confirmed the capacity of the proposed model to eliminate noise thanks to the DEA algorithm and to reduce the error detection rate by 0.02.Finally,we have compared the results between other models and our model approach that associated the autoencoder and the k-nearest neighbours.We performed our solution efficiency using matrix confusion calculation(ROC,MSE)and test precision,accuracy,F-1score,and recall.The results obtained show that the proposed model achieved a detection accuracy of 97.33%,0.76% in AUROC,and 0 false positives.Experiments on real data sets show that the proposed algorithm outperforms other techniques in terms of effectiveness and efficiency.This technique provided an efficient tool to deal with the noisy nature of the WSN environment as well as to detect spikes in the sensory data.