基于学习速率自适应的脉冲神经网络快速学习方(英文)

For accelerating the supervised learning by the SpikeProp algorithm with the temporal coding paradigm in spiking neural networks (SNNs), three learning rate adaptation methods (heuristic rule, delta-delta rule, and delta-bar-delta rule), which are used to speed up training in artificial neural networks, are used to develop the training algorithms for feedforward SNN. The performance of these algorithms is investigated by four experiments: classical XOR (exclusive or) problem, Iris dataset, fault diagnosis in the Tennessee Eastman process, and Poisson trains of discrete spikes. The results demonstrate that all the three learning rate adaptation methods are able to speed up convergence of SNN compared with the original SpikeProp algorithm. Furthermore, if the adaptive learning rate is used in combination with the momentum term, the two modifications will balance each other in a beneficial way to accomplish rapid and steady convergence. In the three learning rate adaptation methods, delta-bar-delta rule performs the best. The delta-bar-delta method with momentum has the fastest convergence rate, the greatest stability of training process, and the maximum accuracy of network learning. The proposed algorithms in this paper are simple and efficient, and consequently valuable for practical applications of SNN.     

基于改进自适应模糊推理系统的半导体制造系统瓶颈设备预测方法(英文)

Semiconductor manufacturing (SM) system is one of the most complicated hybrid processes involved continuously variable dynamical systems and discrete event dynamical systems. The optimization and scheduling of semiconductor fabrication has long been a hot research direction in automation. Bottleneck is the key factor to a SM system, which seriously influences the throughput rate, cycle time, time-delivery rate, etc. Efficient prediction for the bottleneck of a SM system provides the best support for the consequent scheduling. Because categorical data (product types, releasing strategies) and numerical data (work in process, processing time, utilization rate, buffer length, etc.) have significant effect on bottleneck, an improved adaptive network-based fuzzy inference system (ANFIS) was adopted in this study to predict bottleneck since conventional neural network-based methods accommodate only numerical inputs. In this improved ANFIS, the contribution of categorical inputs to firing strength is reflected through a transformation matrix. In order to tackle high-dimensional inputs, reduce the number of fuzzy rules and obtain high prediction accuracy, a fuzzy c-means method combining binary tree linear division method was applied to identify the initial structure of fuzzy inference system. According to the experimental results, the main-bottleneck and sub-bottleneck of SM system can be predicted accurately with the proposed method.     

Artificial neural networks model for the prediction of steady state phenol biodegradation in a pulsed plate bioreactor

BACKGROUND: A recent innovation in fixed film bioreactors is the pulsed plate bioreactor (PPBR) with immobilized cells. The successful development of a theoretical model for this reactor relies on the knowledge of several parameters, which may vary with the process conditions. It may also be a time‐consuming and costly task because of their nonlinear nature. Artificial neural networks (ANN) offer the potential of a generic approach to the modeling of nonlinear systems. RESULTS: A feedforward ANN based model for the prediction of steady state percentage degradation of phenol in a PPBR by immobilized cells of Nocardia hydrocarbonoxydans (NCIM 2386) during continuous biodegradation has been developed to correlate the steady state percentage degradation with the flow rate, influent phenol concentration and vibrational velocity (amplitude × frequency). The model used two hidden layers and 53 parameters (weights and biases). The network model was then compared with a Multiple Regression Analysis (MRA) model, derived from the same training data. Further these two models were used to predict the percentage degradation of phenol for blind test data. CONCLUSIONS: The performance of the ANN model was superior to that of the MRA model and was found to be an efficient data‐driven tool to predict the performance of a PPBR for phenol biodegradation. Copyright © 2008 Society of Chemical Industry     

Flowsheet generation through hierarchical reinforcement learning and graph neural networks

Process synthesis experiences a disruptive transformation accelerated by artificial intelligence. We propose a reinforcement learning algorithm for chemical process design based on a state-of-the-art actor-critic logic. Our proposed algorithm represents chemical processes as graphs and uses graph convolutional neural networks to learn from process graphs. In particular, the graph neural networks are implemented within the agent architecture to process the states and make decisions. We implement a hierarchical and hybrid decision-making process to generate flowsheets, where unit operations are placed iteratively as discrete decisions and corresponding design variables are selected as continuous decisions. We demonstrate the potential of our method to design economically viable flowsheets in an illustrative case study comprising equilibrium reactions, azeotropic separation, and recycles. The results show quick learning in discrete, continuous, and hybrid action spaces. The method is predestined to include large action-state spaces and an interface to process simulators in future research.     

基于神经网络的有机朗肯循环过程及循环性能计算方法

有机朗肯循环（ORC）是中低温热能-电能转换中最具前景的技术之一,近年来受到越来越多的关注。工质是ORC实现的载体,由于热源及可选工质的多样性,工质筛选及系统的优化对于提升ORC综合性能非常重要,而物性及过程特性的准确预测是关键。提出了基于神经网络-基团贡献法的ORC系统性能计算方法,建立了涵盖11个基团的基团表,从REFPROP中调用51种工质7958组数据进行神经网络训练,获得了ORC中各个热力过程能量转换和熵差的计算关联式。计算了21种常用工质在1584组工况下的ORC系统性能,并与基于传统方法计算的ORC系统性能参数进行了对比。结果显示预测得到的ORC系统热效率、净输出功和系统?效率与用REFPROP计算得出的结果相比误差分别为1.01%、1.02%和1.61%,相比传统方法,预测精度有显著提高。     

Adaptive neural models for on-line prediction in fermentation

This paper deals with on-line prediction of fermentation variables by neural network techniques. It is shown that the accuracy of the on-line prediction based on a neural model, obtained from an initial learning sequence, decreases when kinetic changes occur during the course of the fermentation. Therefore, sliding window learning schemes are proposed. For a given network structure, the proposed learning procedures progressively refresh the knowledge integrated within an initial neural model. The influence of the length of the learning window, the number of iterations and the initial neural model on the predictive accuracy of adaptive neural models are investigated. Sliding window learning schemes can be also used when fermentation measurements are delayed and/or infrequent.     

Generalized hybrid control synthesis for affine systems using sequential adaptive networks

BACKGROUND: A generalized methodology for the synthesis of a hybrid controller for affine systems using sequential adaptive networks (SAN) is presented. SAN consists of an assembly of neural networks that are ordered in a chronological sequence, with one network assigned to each sampling interval. Using a suitable process model based on oxygen metabolism and an a priori objective function, a hybrid control law is derived that can use online measurements and the states predicted by SAN for computing the desired control action. RESULTS: The performance of the SAN–hybrid controller is tested for simulated fed‐batch production of methionine for three different process conditions. Simulations assume that online measurements of dissolved oxygen (DO) concentration are available. The performance of the SAN–hybrid controller gave an NRMSE of ～10^?4 in the absence of noise, ～10^?3 and ～10^?2 for ± 5% and ± 10% noise in the DO measurement and ～10^?2 for parameter uncertainty when compared with the ideal model prediction. CONCLUSIONS: The observed performance for unmeasured state prediction and control implementation shows that the proposed SAN–hybrid controller can efficiently compute the manipulated variable required to maintain methionine production along the optimized trajectory for different conditions. The test results show that the SAN–hybrid controller can be used for online real‐time implementation in fed‐batch bioprocesses. Copyright © 2009 Society of Chemical Industry     

基于神经网络的pH中和过程非线性预测控制

针对pH中和过程这一化工过程系统中的典型非线性对象特点，应用神经网络建模思想和模型预测控制方法，并结合Hammerstein模型特点，研究pH中和过程非线性系统的两种新型模型预测控制手段，分别建立基于神经网络的非线性预测控制系统整体求解策略和基于Hammerstein模型的两步法预测控制策略，并用MATLAB对其进行仿真。控制仿真结果表明，建立的神经网络预测控制策略和非线性Hammerstein模型预测控制均优于传统PID控制方法，具有良好的设定值跟踪效果和抗干扰控制响应，说明这两种控制策略是非线性过程的有效控制方法。     

A comparative study of a CRT colorimetric prediction model by neural networks and the models by conventional method

In order to produce desired colors on CRT screens, much work has been done on the problem of the CRT colorimetric prediction. However, it would take great pains to overcome the troubles such as the constant channel chromaticity, the gun or channel independence, and the screen background effect, etc., with the conventional prediction methods such as PLCC and PLVC models, etc. To solve such problems, we propose a completely different CRT colorimetric prediction model by using a set of Artificial Neural Networks (ANN), where a set of back‐propagation (BP) neural networks is used to perform a nonlinear conversion between RGB values and XYZ values. By comparing some typical conventional CRT colorimetric prediction models with our neural‐networks‐based model theoretically, the article indicates that our new model can overcome the troubles faced by the conventional models, and by experiment the article shows that our new model can yield a satisfactory prediction result. © 1999 John Wiley & Sons, Inc. Col Res Appl, 24, 45–51, 1999     

基于TA-ConvBiLSTM的化工过程关键工艺参数预测

化工过程中,掌握关键工艺参数的变化趋势对于消除潜在波动、维持工况稳定作用巨大。然而,传统的浅层静态模型很难对非线性和动态性显著的复杂序列数据进行精准预测。针对上述难题,提出一种深度预测模型TA-ConvBiLSTM,将卷积神经网络(convolutional neural networks, CNN)和双向长短时记忆网络(bi-directional long short term memory, BiLSTM)集成到统一框架内,使其不仅能在每个时间步上自动挖掘高维变量间的隐含关联,更能横跨所有时间步自适应提取有用的深层时序特征。此外,引入时间注意力(temporal attention, TA)机制,为反映目标变化规律的重要信息增加权重,避免其因输入序列过长、深层特征太多而被掩盖。所提出方法的有效性在国内某延迟焦化装置炉管温度预测的案例中得到验证。     

Sequential adaptive networks: An ensemble of neural networks for feed forward control of L-methionine production

Intracellular methionine synthesis is strictly regulated and apparently, results in highly nonlinear concentration-time profiles observed in fed-batch production of this amino acid. For controlling methionine concentration along a predefined trajectory, a control strategy was developed using a sequential adaptive network (SAN) in conjunction with a mechanistic feedforward control law. SAN is an assembly of chronologically ordered networks, with one sub network assigned to each sampling interval, so that feature memory is distributed. Data for training SAN was obtained using a model whose parameters were calculated from experimental data. A range of different operating regimes was simulated using the model to create process scenarios for evaluating the performance of the SAN-feedforward controller (SAN-FFC). The adaptation of the weights of SAN is driven by the error between predicted and measured values of dissolved oxygen concentration at each sampling interval. Under simulated conditions, the feedforward control law uses the values of state variables predicted by SAN and measured values to determine a control action that is in tune with process evolution. The SAN-feedforward controller is robust and exhibits stable tracking of the methionine concentration trajectory in the presence of measurement noise and parametric uncertainty. It is perceived that the online implementation of SAN-FFC for a general bioprocess is practicable.     

Nonlinear system identification and model reduction using artificial neural networks

We present a technique for nonlinear system identification and model reduction using artificial neural networks (ANNs). The ANN is used to model plant input–output data, with the states of the model being represented by the outputs of an intermediate hidden layer of the ANN. Model reduction is achieved by applying a singular value decomposition (SVD)-based technique to the weight matrices of the ANN. The sequence of state values is used to convert the model to a form that is useful for state and parameter estimation. Examples of chemical systems (batch and continuous reactors and distillation columns) are presented to demonstrate the performance of the ANN-based system identification and model reduction technique.