人工神经网络在化学与工程中的应用：新研究与优化

需积分: 19 65 浏览量更新于2024-07-19 1 收藏 9.62MB PDF 举报

"本书《Artificial Neural Networks_New Research-Nova Science(2017)》是一本关于人工神经网络（ANNs）最新研究的专著，涵盖了化学、化工、能源效率优化、微波晶体管模型参数提取、太阳能辐射估算、皮肤防晒滤光剂设计以及在中医远程医疗系统中的应用等多个领域。书中强调了ANNs在多因素同时优化和复杂问题建模上的优势，并探讨了其在化学工程中解决非线性问题的能力。" 人工神经网络（Artificial Neural Networks，ANNs）是一种受生物神经元结构启发的计算模型，能够通过学习和调整连接权重来模拟大脑的处理能力。书中的第一章节讨论了ANNs在过程优化中的应用。传统方法如单变量优化一次只考虑一个因素，而ANNs则可以同时分析所有过程参数对响应的影响，提供更优解。例如，在电弧炉电力消耗优化中，输入变量是熔炼用废钢的化学成分，输出变量是单位电耗。此外，ANNs也被用于提取和吸附过程的建模，通过调整如提取时间、溶剂性质、固液比等参数来优化生物活性化合物的提取效率。第二章节进一步阐述了ANNs在化学和化工领域中的应用。由于这些问题通常涉及复杂的非线性系统，ANNs凭借其高计算能力、数据学习和泛化性能、容错性和上下文信息处理能力，成为处理此类问题的有效工具。通过多种数学函数描述化学过程，ANNs能处理多个变量对目标响应的影响，广泛应用于化学数据分析和复杂问题求解。书中的研究还涉及了ANNs在太阳能辐射预测中的作用，利用神经网络估计太阳辐射量，对于可再生能源的规划和管理具有重要意义。另外，通过多层感知机（Multilayer Perceptron, MLP）模型预测铣削工具的磨损，可以提高机械加工的精度和效率。在微波晶体管的小信号和噪声模型参数提取中，ANNs也展示了其潜力，帮助研究人员无须全面了解物理机制就能进行建模。最后，书中提到了ANNs在深度学习和预测分析方面的应用，特别是在语义中医远程医疗系统中的角色。这表明，神经网络技术不仅可以用于传统的科学计算，还可以与大数据和人工智能相结合，推动医疗健康领域的创新。《Artificial Neural Networks_New Research-Nova Science(2017)》这本书全面介绍了ANNs在不同领域的最新研究成果，展示了这一技术在解决实际问题时的强大能力，并预示着其在未来化学工程和相关科学中将扮演更加重要的角色。

Ivan M. Savic, Dragoljub G. Gajic and Ivana M. Savic-Gajic

medium are commonly used as the input variables, while the amount of adsorbate after

treatment is selected as output variable. Based on the literature review, it can be

concluded that the application of artificial neural networks will surely have an important

role in the modeling and optimization of chemical processes in the future.

Keywords: modeling, optimization, artificial neural network, extraction, adsorption,

chemical processes

ARTIFICIAL NEURAL NETWORKS

Artificial neural networks (ANNs) were found application in the various disciplines with

the aim of modeling the complex problems. They are consisted of the linked simple

processing elements, so called artificial neurons or nodes which provide a parallel processing

of data (Hecht-Nielsen, 1990; Schalkoff, 1997). The basic characteristics of biological

system, which give a special importance to the neural network, are nonlinearity, high

parallelism, robustness, learning, the ability to manage the imprecise and insufficiently clear

information, as well as the ability to generalize (Jain et al., 1996). The artificial models with

those characteristics are desirable because (1) the nonlinearity provides a better fitting the

data, (2) the insensitivity of noise provides precisely prediction in the presence of unreliable

data and errors during the measurement, (3) the high parallelism enables fast data processing,

(4) the learning and adaptability give a possibility to the system to update (edit) all the interior

structure and (5) the generalization provides the application of model on the data that are not

included in the learning process of networks. The main goal of the ANNs based on the

calculation (neural pocessing) is development of the mathematical algorithms that enable to

learn by imitating the information processing and learning as in human brain. The models of

ANNs are empirical in nature, however they can give the correct solutions for precisly and

imprecisely formulated problems and for the phenomena that can only explain experimentaly

(Leondes, 2007).

ARTIFICIAL AND BIOLOGICAL NEURAL NETWORKS

A function of biological neuron is necessary to detail explain, because it is needed for

better understanding of artificial neurons function and analogy between artificial and

biological networks.

BIOLOGICAL NEURON

Human neural system is consisted of the great number of neurons different types and

lenghts depending on their positions in the body (Schalkoff, 1997). The simplified shematic

representation of biological neuron with cleary visible main functional units (dendrites, cell

body and axon) is given in Figure 1.

Applications of Artificial Neural Networks in Chemical Engineering

Figure 1. Scheme of biological neuron.

In addition to nucleus that contains the information about the hereditary properties, a cell

body has plasma with the molecules for production of the necessary material (Jain et al.,

1996). Dendrites receive signals from the neighboring neurons, so in this way the signals

reach to the cell body. The total area of dendrites of one typical neuron which could receive

signal is approximately 0.25 mm

(Zupan and Gasteiger, 1993). Axon contains the branches

and receives signals from a cell body. The signals are transferred through the synapse

(microscopic interspace) to the dendrites of neighboring neurons. Schematic illustration of

signal transfer between two neurons through the synapse is presented in Figure 2. The

impulse in the form of electric signal is transferred inside of dendrite and cell body toward the

presynaptic membrane. After arriving in the membrane, a neurotransmitter (chemical) is

released from the vesicles in the amount proportional to the strength of the input signal. The

neurotransmitter is released inside of synaptic interspace toward the post-membrane and

optionally in the dendrites of neighboring neurons. It forces them to generate a new electrical

signal depending on the threshold of the receiving neuron. The generated signal passes

through the neighboring neuron in an identical way. The quantity of signal, which passes

through the receiving neuron, depends on the intensity of source signal of each neurons, their

synaptic strength and threshold of the receiving neuron. The transfer of multiple signals

simultaneously is possible, because a neuron has a great number of dendrites/synapses. This

simplified mechanism of signal transfer has caused the early development of neuro-

computing and construction of the units of ANNs (Vedmedenko, 2007).

ANALOGY BETWEEN ARTIFICIAL AND BIOLOGICAL NEURONS

Analogy between artificial and biological neurons is that the connections among the

nodes represent the axons and dendrites, the weights represent the synapses, and threshold

approximates an activity in the body (Jain et al., 1996). In Figure 3, n biological neurons with

different signal intensities x and synaptic strength w, which enter in the neuron with the

Applications of Artificial Neural Networks in Chemical Engineering





















bxw

(1)

where 1 represents “on”, and 0 “off” (Figure 3) during solving the classification problems. A

positive weight (w

> 1) increases the neuron, while a negative weight reduces ξ and inhibits

the neural activity. The system is consisted of artificial neuron and inputs as shown in Figure

3. It is called a perceptron which establishes a mapping between input activity (stimulus) and

output signal. In Equation 1, the neural threshold can be considered as an additional input

node whose value is always 1, and its weight is equal to b. In this case, the summarizing is

performed from 0 to n (in Equation 1), and a network signal ξ is compared to 0.

PERCEPTRONS

Perceptron (Figure 3) could be trained on the set data using a special rule of learning

(Lingireddy and Brion, 2005). The weight coefficients of perceptron including threshold are

being changed proportionally the differences between the target response (Y) and perceptron

solution (y) for the each data. Using the special techniques of finding, the learning process

tends to find the set of weights which correspond to the global minimum. Rosenblatt derived

the perceptron rule which helps to find the optimal weight vector in a finite number of

iterations, despite the initial value of the weights in 1962. This rule could be precisely used

only for linearly separable classes (Hecht-Nielsen, 1990). In this case, a linear hyper-plane

separating one class of objects on one side, and the second class on the other side. The

linearly and non-linearly separable classes are presented in Figure 4.

In order to solve the non-linearly separable problems, the additional layers of neurons are

inserted between the input layer (which contains the input nodes) and output neuron, whereby

the architecture of multilayer perceptron (MLP) is obtained as presented in Figure 4 (Hecht-

Nielsen, 1990). How this layer could not interact with the environment, it is called a hidden

layer. The nodes of this layer are called the hidden nodes. The additional hidden layers

increase the ability of perceptron to solve a non-linearly classification problem. Using a

similar dynamic of neuron, the information passes through the hidden layer and then reaches

to the output layer. Learning of MLP is not such direct as in the case of the simple perceptron.

Backpropagation neural network is a type of the trained MLP (Harris, 1994). However, the

learning process is an upgrade of the algorithm of simple perceptron during which the weights

are being adjusted in the hidden layers (Hecht-Nielsen, 1990).

The activation function of neurons in the hidden layer has a role to provide the learning

of neural network. The neurons of hidden layers do not have a higher ability than ordinary

perceptron network in the case of linear problems. The non-linear activation functions are

often used for the neurons in the output layer, what influences that the network with multiple

layers to be a particularly powerful. All non-linear functions are often used. The sigmoid

functions (logistic, arctangent or Gaussian function) are commonly used at the

剩余242页未读，继续阅读

WindStand

粉丝: 35
资源: 367

人工神经网络在化学与工程中的应用：新研究与优化

AIS.rar_ais_ais--m1731.pdf_immune artificial_免疫_免疫算法

Artificial-Neural-Networks-in-NVIDIA-Blackwell-Chip.pptx

Spiking neural networks.rar_Neural networks_SNN_Spiking-Neural_s

AI_Roadmap_Exec_Summary-FINAL-.pdf

books_[.md_&_.ipynb]_Series_of_Artificial_Intell_AI-Notes.zip

_Introduction_to_Artificial_Intelligence_-_2019_Fa_AI.zip

Mastering Machine Learning with scikit-learn -2017.7.24

Advanced-Certification-in-Data-Science-and-Artificial-Intelligence.pdf

PyTorch Recipes_ A Problem-Solution Approach.pdf

Artificial Neural Networks New Research 无水印原版pdf

最新资源