Enhanced non-parametric sequence learning scheme for internet of things sensory data in cloud infrastructure

Abstract

The Internet of Things (IoT) Cloud is an emerging technology that enables machine-to-machine, human-to-machine and human-to-human interaction through the Internet. IoT sensor devices tend to generate sensory data known for their dynamic and heterogeneous nature. Hence, it makes it elusive to be managed by the sensor devices due to their limited computation power and storage space. However, the Cloud Infrastructure as a Service (IaaS) leverages the limitations of the IoT devices by making its computation power and storage resources available to execute IoT sensory data. In IoT-Cloud IaaS, resource allocation is the process of distributing optimal resources to execute data request tasks that comprise data filtering operations. Recently, machine learning, non-heuristics, multi-objective and hybrid algorithms have been applied for efficient resource allocation to execute IoT sensory data filtering request tasks in IoT-enabled Cloud IaaS. However, the filtering task is still prone to some challenges. These challenges include global search entrapment of event and error outlier detection as the dimension of the dataset increases in size, the inability of missing data recovery for effective redundant data elimination and local search entrapment that leads to unbalanced workloads on available resources required for task execution. In this thesis, the enhancement of Non-Parametric Sequence Learning (NPSL), Perceptually Important Point (PIP) and Efficient Energy Resource Ranking- Virtual Machine Selection (ERVS) algorithms were proposed. The Non-Parametric Sequence-based Agglomerative Gaussian Mixture Model (NPSAGMM) technique was initially utilized to improve the detection of event and error outliers in the global space as the dimension of the dataset increases in size. Then, Perceptually Important Points K-means-enabled Cosine and Manhattan (PIP-KCM) technique was employed to recover missing data to improve the elimination of duplicate sensed data records. Finally, an Efficient Resource Balance Ranking- based Glow-warm Swarm Optimization (ERBV-GSO) technique was used to resolve the local search entrapment for near-optimal solutions and to reduce workload imbalance on available resources for task execution in the IoT-Cloud IaaS platform. Experiments were carried out using the NetworkX simulator and the results of N-PSAGMM, PIP-KCM and ERBV-GSO techniques with N-PSL, PIP, ERVS and Resource Fragmentation Aware (RF-Aware) algorithms were compared. The experimental results showed that the proposed NPSAGMM, PIP-KCM, and ERBV-GSO techniques produced a tremendous performance improvement rate based on 3.602%/6.74% Precision, 9.724%/8.77% Recall, 5.350%/4.42% Area under Curve for the detection of event and error outliers. Furthermore, the results indicated an improvement rate of 94.273% F1-score, 0.143 Reduction Ratio, and with minimum 0.149% Root Mean Squared Error for redundant data elimination as well as the minimum number of 608 Virtual Machine migrations, 47.62% Resource Utilization and 41.13% load balancing degree for the allocation of desired resources deployed to execute sensory data filtering tasks respectively. Therefore, the proposed techniques have proven to be effective for improving the load balancing of allocating the desired resources to execute efficient outlier (Event and Error) detection and eliminate redundant data records in the IoT-based Cloud IaaS Infrastructure.