基于復(fù)雜系統(tǒng)模型的地下采礦無線傳感器網(wǎng)絡(luò)中的優(yōu)化混合神經(jīng)網(wǎng)絡(luò)
發(fā)布時間：2018-11-09 09:50
【摘要】：地下采礦是全球最危險的工作之一,在過去的二十年,在諸如火災(zāi)、落石、洪水、有毒氣體等重大事故中均有重大人員傷亡和巨大的生命和財產(chǎn)損失。生命的傷亡和財產(chǎn)的巨大損失破壞了社會的穩(wěn)定和可持續(xù)發(fā)展。政府、采礦業(yè)、工程師、科學家和計算機專家提出的各種有效的措施正在緩解情況。這些措施包括用以預(yù)測安全和采取救援措施的地下挖掘模型,以便在危險的時候通過監(jiān)測來幫助跟蹤和營救礦工。 然而,大部分用以分析疏散問題和溝通的先進計算機模型通常運算起來非常耗時的。例如,包含額外的平方運算但獲得了全球公認的高斯徑向基模型,隨著網(wǎng)絡(luò)規(guī)模變大,將比其他沒有平方運算的模型更需要時間。本文側(cè)重于優(yōu)化路由或傳輸路徑(R),事故地點被構(gòu)建為純粹的隨機事件,通過建立神經(jīng)網(wǎng)絡(luò)模型對特定巖層信息來估算信息鏈沒有損壞的概率。這樣,當真正的事故發(fā)生時,機器人通過存儲記憶的復(fù)雜關(guān)系的神經(jīng)網(wǎng)絡(luò),運用實時看到的巖層信息就能夠立刻預(yù)測信息鏈沒有損壞概率。如果預(yù)測結(jié)果是積極的,機器人等待接收救援信號；否則,它將進入更深巖層,重復(fù)這個過程。 文章第二部分通過改進已有文獻中S函數(shù)和緊湊R函數(shù)方法,給出了用絕對值運算代替平方運算來減少高斯模型運算負擔的新模型,進而降低計算成本和提高運行效率。 最后,基于這一模型引入一系列混合神經(jīng)網(wǎng)絡(luò)。為了減少學習過程中的錯誤以及分析其他優(yōu)化參數(shù),采用了對混合神經(jīng)網(wǎng)絡(luò)中的轉(zhuǎn)移函數(shù)引入了線性和非線性加權(quán)等措施。 大腦具有由神經(jīng)系統(tǒng)控制的故障識別功能和自愈合機制；本文試圖考慮到兩個神經(jīng)功能共同參導(dǎo)時混合模型有效性,尤其為地下礦難救援行動提出了一個獨特的模型,來進一步減少大腦的工作。 本文研究合理性在于將目前的研究焦點從小世界網(wǎng)絡(luò)的系統(tǒng)分析轉(zhuǎn)移到數(shù)以百萬計節(jié)點的網(wǎng)絡(luò),這將要求計算機具有高處理能力并且要進行相當長的時間運行。因此,對大量的傳感器信息處理的快速計算算法的需要已經(jīng)成為當務(wù)之急,更不用說基本數(shù)據(jù)有可能在事故中可能被毀掉。此外,簡單的模仿人類大腦的神經(jīng)網(wǎng)絡(luò),可以演示以前需要人類專家來參與的快速學習和準確處理分類屬性問題。盡管這些工具顯然無法取代人類專家,但他們?yōu)閱栴}診斷和決策支持提供了依據(jù)。 自適應(yīng)變異粒子群優(yōu)化(A MPSO)f(?)編碼的遺傳算法用于更新基礎(chǔ)轉(zhuǎn)移函數(shù),這使得修正后的函數(shù)可以加快訓(xùn)練過程,提高神經(jīng)網(wǎng)絡(luò)的學習精度。 從基本模型獲得的結(jié)果表明,CRBF模型在各參數(shù)(平均迭代次數(shù)、收斂時間、標準方差誤差和計算時間)表現(xiàn)更優(yōu)；在采用PSO算法訓(xùn)練的模型中,最優(yōu)誤差分別為CRBF模型(0.0111)、ZRBF(0.0140) S函數(shù)模型(0.0157)、高斯算法(0.0120),而采用遺傳算法訓(xùn)練的模型中,CRBF, SBF, ZRBF和GRBF的最優(yōu)誤差值分別是0.012923,0.0126,0.012183,0.12291。相對于目標的最優(yōu)誤差0.01而言,在采用PSO算法和遺傳算法訓(xùn)練負余弦中非線性加權(quán)混合模型中合模型中,最佳的優(yōu)化誤差分別為0.009和0.01109；其次是在采用PSO算法和遺傳算法訓(xùn)練線性混合模型中,最優(yōu)誤差是0.0103,在非線性負余弦加權(quán)g-比例混合模型中,最優(yōu)誤差是0.011。 從結(jié)果可以看出,PSO算法模型被證明是地下礦山、隧道和其他自然(如滑坡)救援行動中的強有力的選擇。遺傳算法模型,特別是SBF訓(xùn)練良好,但在巖石滲透是有困難并且容易出錯,但是在諸如地表采礦、醫(yī)院和建筑物等的疏散行動中非常有效。 本文分為五個章節(jié)。第1章討論了復(fù)雜的自適應(yīng)系統(tǒng)的背景、目標、意義,并提出模型的概念框架。第二章著重介紹一些相關(guān)文獻,包括S函數(shù)和徑向基函數(shù),以及研究這些函數(shù)所用的方法。第三章對路由路徑生成提出假設(shè),并構(gòu)建了基本模型。第四章探討了已提出模型的幾種混合情況,包括線性和非線性加權(quán)混合模型及其分析。第五章對于已提出的遺傳算法的路由路徑進行了分析,并討論了模型在粒子群和遺傳算法中的應(yīng)用趨勢。最后,第六章是總結(jié)和擬進一步開展的研究工作。
[Abstract]:Underground mining is one of the most dangerous work in the world. In the past two decades, there are major casualties and great loss of life and property in major accidents, such as fire, rockfall, flood, and toxic gas. The loss of life and the loss of property damage the stability and sustainable development of society. The various effective measures proposed by the Government, the mining industry, the engineers, scientists and computer experts are being mitigated. These include an underground mining model to predict safety and to take rescue measures in order to help track and rescue miners at risk through monitoring. However, most of the advanced computer models used to analyze evacuation problems and communication are typically operational When, for example, an additional square operation is included, a globally accepted Gaussian radial basis model is obtained, and as the network scale becomes large, there will be a need for a model that does not have a square operation Time. This paper focuses on the optimization of routing or transmission path (R). The accident site is constructed as a pure random event, and the information chain is estimated to be free from damage by establishing a neural network model. The probability is that, when the real accident happens, the robot can predict the information chain without damage immediately by using the neural network of the complex relation of memory and memory, Probability. If the prediction is positive, the robot waits to receive a rescue signal; otherwise it will go deeper into the formation and repeat this in that second part of the article, by improving the S-function and the compact R-function method in the existing literature, a new model is given to reduce the computational burden of the Gaussian model by using the absolute value operation instead of the square operation, thus the calculation cost and the extraction cost are reduced. High operating efficiency. Finally, introduce a model based on this model In order to reduce the errors in the learning process and to analyze other optimization parameters, the linear sum of the transfer functions in the hybrid neural network is introduced. Nonlinear weighting and other measures. The brain has a fault recognition function and a self-healing mechanism controlled by the nervous system. To further reduce the brain's work. The rationale is to transfer the current research focus from the system analysis of the small world network to the network of millions of nodes, which will require the computer to have high processing performance the need for a large number of fast computing algorithms for sensor information processing has become a priority, not to say, The data is likely to be destroyed in the accident. In addition, a simple imitation of the human brain's neural network can demonstrate the rapid involvement of human experts Speed learning and accurate treatment of the problem of classification attributes. Although these tools are clearly unable to replace human experts, they are Problem diagnosis and decision support provide a basis. Self-adaptation should The genetic algorithm encoded by the variant particle swarm optimization (A MPSO) f (?) is used to update the base transfer function, which makes the modified function add The results from the basic model show that the CRBF model is better in each parameter (average number of iterations, convergence time, standard variance error and calculation time), and the optimal error is the CRBF model (0.0111) and the ZRBF (0. 0111) in the model trained by the PSO algorithm. The optimal error value of CRBF, SBF, ZRBF and GRBF is 0. 012923,0. 0, respectively. The best optimization error is 0. 009 and 0. 01109 in the nonlinear weighted mixed model of the inverse cosine by the PSO and the genetic algorithm, and the second is in production. In the linear mixed model, the optimal error is 0. 0103 and the nonlinear negative cosine weighting is obtained by using the PSO algorithm and the genetic algorithm. In the g-scale hybrid model, the optimal error is 0.011. As can be seen from the results, the PSO algorithm is proved to be an underground mine, a tunnel, Strong choice in road and other natural (such as landslides) rescue operations. The genetic algorithm model, especially the SBF, is well trained, but it is difficult and prone to error in rock penetration, but in the such as surface mining, hospital and The evacuation of buildings and so on is very effective. This article is divided into five chapters. Chapter 1 discusses complex The background, object, meaning of the adaptive system and the conceptual framework of the model are put forward. including an s function and a radial basis function, Methods: The third chapter makes a hypothesis about the route generation, and constructs the basic model. The fourth chapter discusses the proposed model. Several mixed cases, including linear and non-linear weighted hybrid models and their analysis, are presented in the fifth chapter for the proposed genetic algorithm. The analysis of the model in the particle swarm and the genetic algorithm is also discussed.
【學位授予單位】：江蘇大學
【學位級別】：博士
【學位授予年份】：2014
【分類號】：TN929.5;TP212.9
本文編號：2320066
[Abstract]:Underground mining is one of the most dangerous work in the world. In the past two decades, there are major casualties and great loss of life and property in major accidents, such as fire, rockfall, flood, and toxic gas. The loss of life and the loss of property damage the stability and sustainable development of society. The various effective measures proposed by the Government, the mining industry, the engineers, scientists and computer experts are being mitigated. These include an underground mining model to predict safety and to take rescue measures in order to help track and rescue miners at risk through monitoring. However, most of the advanced computer models used to analyze evacuation problems and communication are typically operational When, for example, an additional square operation is included, a globally accepted Gaussian radial basis model is obtained, and as the network scale becomes large, there will be a need for a model that does not have a square operation Time. This paper focuses on the optimization of routing or transmission path (R). The accident site is constructed as a pure random event, and the information chain is estimated to be free from damage by establishing a neural network model. The probability is that, when the real accident happens, the robot can predict the information chain without damage immediately by using the neural network of the complex relation of memory and memory, Probability. If the prediction is positive, the robot waits to receive a rescue signal; otherwise it will go deeper into the formation and repeat this in that second part of the article, by improving the S-function and the compact R-function method in the existing literature, a new model is given to reduce the computational burden of the Gaussian model by using the absolute value operation instead of the square operation, thus the calculation cost and the extraction cost are reduced. High operating efficiency. Finally, introduce a model based on this model In order to reduce the errors in the learning process and to analyze other optimization parameters, the linear sum of the transfer functions in the hybrid neural network is introduced. Nonlinear weighting and other measures. The brain has a fault recognition function and a self-healing mechanism controlled by the nervous system. To further reduce the brain's work. The rationale is to transfer the current research focus from the system analysis of the small world network to the network of millions of nodes, which will require the computer to have high processing performance the need for a large number of fast computing algorithms for sensor information processing has become a priority, not to say, The data is likely to be destroyed in the accident. In addition, a simple imitation of the human brain's neural network can demonstrate the rapid involvement of human experts Speed learning and accurate treatment of the problem of classification attributes. Although these tools are clearly unable to replace human experts, they are Problem diagnosis and decision support provide a basis. Self-adaptation should The genetic algorithm encoded by the variant particle swarm optimization (A MPSO) f (?) is used to update the base transfer function, which makes the modified function add The results from the basic model show that the CRBF model is better in each parameter (average number of iterations, convergence time, standard variance error and calculation time), and the optimal error is the CRBF model (0.0111) and the ZRBF (0. 0111) in the model trained by the PSO algorithm. The optimal error value of CRBF, SBF, ZRBF and GRBF is 0. 012923,0. 0, respectively. The best optimization error is 0. 009 and 0. 01109 in the nonlinear weighted mixed model of the inverse cosine by the PSO and the genetic algorithm, and the second is in production. In the linear mixed model, the optimal error is 0. 0103 and the nonlinear negative cosine weighting is obtained by using the PSO algorithm and the genetic algorithm. In the g-scale hybrid model, the optimal error is 0.011. As can be seen from the results, the PSO algorithm is proved to be an underground mine, a tunnel, Strong choice in road and other natural (such as landslides) rescue operations. The genetic algorithm model, especially the SBF, is well trained, but it is difficult and prone to error in rock penetration, but in the such as surface mining, hospital and The evacuation of buildings and so on is very effective. This article is divided into five chapters. Chapter 1 discusses complex The background, object, meaning of the adaptive system and the conceptual framework of the model are put forward. including an s function and a radial basis function, Methods: The third chapter makes a hypothesis about the route generation, and constructs the basic model. The fourth chapter discusses the proposed model. Several mixed cases, including linear and non-linear weighted hybrid models and their analysis, are presented in the fifth chapter for the proposed genetic algorithm. The analysis of the model in the particle swarm and the genetic algorithm is also discussed.
【學位授予單位】：江蘇大學
【學位級別】：博士
【學位授予年份】：2014
【分類號】：TN929.5;TP212.9
【參考文獻】
相關(guān)期刊論文 前1條
1 彭程;潘玉民;;粒子群優(yōu)化的RBF瓦斯涌出量預(yù)測[J];中國安全生產(chǎn)科學技術(shù);2011年11期
，本文編號：2320066
本文鏈接：http://sikaile.net/kejilunwen/wltx/2320066.html
最近更新
教材專著
