Optimized Hybrid Neural Network Based On Complex System Model For Wireless Sensor Network In Underground Mining Environment

Underground mining is one of the most dangerous jobs around the globe. Within the past two decades, there has been heavy casualties and tremendous loss of lives and property as a result of accidents occasioned by fires, rock falls, flooding, and poisonous gasses among others. Rigorous measures by governments, mining industries, engineers, scientists and computer experts are being made to curtail the situation. Efforts in this direction include underground mining models for predicting accidents and rescue measures as well as monitoring mining operations to help track and rescue miners in times of dangerHowever, most advanced computer models for the analysis of evacuation problems and communication are generally computationally expensive. For example, the Gaussian radial basis model which has gained global acceptance involves additional square operation which increases its processing time especially when the network size becomes large. As a result it compares unfavorable with those models which do not use the square operation in their architecture.This work focuses on optimizing the capabilities of a routing or transmission path (R), where an incident location is modeled as a pure random event. Using the model the probability that the communication chain through particular rock layers to the ground is not broken can be calculated. The neural network is then trained to memorize the complicated relationships, such that when real accident happens, the neural network resident in the robot is used to predict the probability based on the rock layer it sees instantly. If the result is positive, the robot waits to receive a rescue signal. Otherwise, it moves deeper to the next layer and repeats the procedure.The study further proposes a model that focuses on removing the computational burden of the Gaussian model by paralyzing the extra it’s power operation and generating sigmoid and compact radial basis functions literally but novel, to reduce cost and increase processing efficiency and best utilization of resource. Finally, a number of hybrid neural networks based on the proposed model are introduced. These hybrids introduce a linear and nonlinear updating for the transfer functions with the intention of reducing the overall error in the learning process, as well as investigating other optimization parameters. The brain has the capability of fault identification and self healing mechanisms which are controlled by the neuro-system. This work attempts to advance further the understanding of the work of the brain than the current state of knowledge in this area shows, especially where the two neural functions are joined together, and proposes a unique model for rescue operation.The justification of this work is that the focus of current research is moving from system analysis of small-world networks to millions of nodes. This will demand computers with high capacity to process, and considerable time to run. Therefore the need for fast computation algorithm to deal with these large numbers of sensors has become imperative, notwithstanding the fact that base stations can be destroyed in times of accident. Furthermore, the simple imitations of the human brain (called neural network), demonstrate fast and accurate learning and classification properties in problems that otherwise require human experts. Although such tools cannot obviously replace human experts, they are used as diagnostic tools and supporting evidence in decision making.This work employs the adaptive mutation particle swarm optimization (AMPSO) and coded genetic algorithm are to update the basis transfer functions such that the modified function could speed up the training process and improve the learning accuracy of the neural networks.The results obtained from the basic models indicate that CRBF has superior performance in terms of contributions made by the various parameters (mean iteration, convergent time, standard variance, error and computational time) and with optimized errors of CRBF (0.0111) and sigmoid (0.0157) as compared to ZRBF (0.0140) and Gaussian (0.0120) for the PSO trained model, while optimised errors for genetic model are0.012923,0.0126,0.012183,0.12291for CRBF, SBF, ZRBF and GRBF, respectively. The nonlinear weighted hybrid of negative cosine emerges the best with an optimized error of0.009and0.01109for PSO and genetic models respectively, as against a target error of0.01among all the hybrids. This is followed by the linear hybrid of67%splenium with an error of0.0103and g-ratio hybrid with nonlinear weight of negative cosine (0.011).From the results, PSO proves to be a competitive alternative to underground mine, tunnel and other natural (i.e. landslides) rescue operations. The model with the genetic algorithm, especially SBF trains well but has difficulty in rock penetration coupled with high error. However, it can be very effective in evacuation operations in facilities such as surface mining, hospitals and buildings.This dissertation is organized into six chapters. Chapter1discusses significant background, objectives, significance and conceptual framework of the proposed models. Chapter2highlights some related literature including sigmoid and radial basis functions as well as the methods used for the study. The preliminary considerations, assumptions in generating the proposed routing path, and basic models are discussed in Chapter3, while Chapter4examines several linear and nonlinear weighted hybrids of the proposed model. In Chapter5the proposed routing path with genetic algorithm is analyzed. The trends of the models are discussed for both the particle swarm and genetic algorithm. Finally, Chapter6provides summary of the findings and conclusions.