The joining of a 6-mm thickness Al 6061 to Stainless steel 304 has been performed by solid state welding. A selection method of optimum friction welding condition using neural networks is proposed. The data used for a...The joining of a 6-mm thickness Al 6061 to Stainless steel 304 has been performed by solid state welding. A selection method of optimum friction welding condition using neural networks is proposed. The data used for analyses are the friction stir welding condition, the input parameters of the model consist of welding speed and tool rotation speed. The outputs of the ANN (Artificial Neural Network)model includes resulting parameters, namely, maximum reached temperature,and heating rate for both aluminum alloy 6061 and stainless steel 304 during friction stir welding process.The results of analysis suggest that the proposed method is an effective one to select an optimum welding condition.Good performance of the ANN model was achieved. The combined influence of welding speed and tool rotation speed on the maximum reached temperature and heating rate for both aluminum alloy 6061and stainless steel 304 friction stir welding was simulated. A comparison was made between the output of the ANN program and finite element model. The calculated results were in good agreement with that of finite element model.展开更多
Fast simulated annealing is implemented into the learning process of neural network to replace the traditional back-propagation algorithm. The new procedure exhibits performance fast in learning and accurate in predic...Fast simulated annealing is implemented into the learning process of neural network to replace the traditional back-propagation algorithm. The new procedure exhibits performance fast in learning and accurate in prediction compared to the traditional neural networks. Two numerical data sets were used to illustrate its use in chemistry.展开更多
In this paper,an approach is developed to optimize the quality of the training samples in the conventional Artificial Neural Network(ANN)by incorporating expert knowledge in the means of constructing expert-rule sampl...In this paper,an approach is developed to optimize the quality of the training samples in the conventional Artificial Neural Network(ANN)by incorporating expert knowledge in the means of constructing expert-rule samples from rules in an expert system,and through training by using these samples,an ANN based on expert-knowledge is further developed.The method is introduced into the field of quantitative identification of potential seismic sources on the basis of the rules in an expert system.Then it is applied to the quantitative identification of the potential seismic sources in Beijing and its adjacent area.The result indicates that the expert rule based on ANN method can well incorporate and represent the expert knowledge in the rules in an expert system,and the quality of the samples and the efficiency of training and the accuracy of the result are optimized.展开更多
A designed visual geometry group(VGG)-based convolutional neural network(CNN)model with small computational cost and high accuracy is utilized to monitor pulse amplitude modulation-based intensity modulation and direc...A designed visual geometry group(VGG)-based convolutional neural network(CNN)model with small computational cost and high accuracy is utilized to monitor pulse amplitude modulation-based intensity modulation and direct detection channel performance using eye diagram measurements.Experimental results show that the proposed technique can achieve a high accuracy in jointly monitoring modulation format,probabilistic shaping,roll-off factor,baud rate,optical signal-to-noise ratio,and chromatic dispersion.The designed VGG-based CNN model outperforms the other four traditional machine-learning methods in different scenarios.Furthermore,the multitask learning model combined with MobileNet CNN is designed to improve the flexibility of the network.Compared with the designed VGG-based CNN,the MobileNet-based MTL does not need to train all the classes,and it can simultaneously monitor single parameter or multiple parameters without sacrificing accuracy,indicating great potential in various monitoring scenarios.展开更多
Multivariate analysis and filtering techniques are widely applied to simultaneous and/or selective determination of multicomponent systems. Many methods among them are based on the principle of linear addition, while ...Multivariate analysis and filtering techniques are widely applied to simultaneous and/or selective determination of multicomponent systems. Many methods among them are based on the principle of linear addition, while this principle does not always hold due to various physical and chemical factors. Using quite a different way, neural network (NN) based on a given learning rule, such as back propagation (BP) model, needs neither knowing nor using any form of input/output relationship. Particularly, NN can resolve various problems such as those with casual relation, those with fuzzy backgrounds, and those with uncertain inferential processes. NN was used by us to investigate quantitative struc-展开更多
文摘The joining of a 6-mm thickness Al 6061 to Stainless steel 304 has been performed by solid state welding. A selection method of optimum friction welding condition using neural networks is proposed. The data used for analyses are the friction stir welding condition, the input parameters of the model consist of welding speed and tool rotation speed. The outputs of the ANN (Artificial Neural Network)model includes resulting parameters, namely, maximum reached temperature,and heating rate for both aluminum alloy 6061 and stainless steel 304 during friction stir welding process.The results of analysis suggest that the proposed method is an effective one to select an optimum welding condition.Good performance of the ANN model was achieved. The combined influence of welding speed and tool rotation speed on the maximum reached temperature and heating rate for both aluminum alloy 6061and stainless steel 304 friction stir welding was simulated. A comparison was made between the output of the ANN program and finite element model. The calculated results were in good agreement with that of finite element model.
文摘Fast simulated annealing is implemented into the learning process of neural network to replace the traditional back-propagation algorithm. The new procedure exhibits performance fast in learning and accurate in prediction compared to the traditional neural networks. Two numerical data sets were used to illustrate its use in chemistry.
文摘In this paper,an approach is developed to optimize the quality of the training samples in the conventional Artificial Neural Network(ANN)by incorporating expert knowledge in the means of constructing expert-rule samples from rules in an expert system,and through training by using these samples,an ANN based on expert-knowledge is further developed.The method is introduced into the field of quantitative identification of potential seismic sources on the basis of the rules in an expert system.Then it is applied to the quantitative identification of the potential seismic sources in Beijing and its adjacent area.The result indicates that the expert rule based on ANN method can well incorporate and represent the expert knowledge in the rules in an expert system,and the quality of the samples and the efficiency of training and the accuracy of the result are optimized.
基金supported by the National Key Research and Development Program of China (Grant No.2019YFB1803700)the Key Technologies Research and Development Program of Tianjin (Grant No.20YFZCGX00440).
文摘A designed visual geometry group(VGG)-based convolutional neural network(CNN)model with small computational cost and high accuracy is utilized to monitor pulse amplitude modulation-based intensity modulation and direct detection channel performance using eye diagram measurements.Experimental results show that the proposed technique can achieve a high accuracy in jointly monitoring modulation format,probabilistic shaping,roll-off factor,baud rate,optical signal-to-noise ratio,and chromatic dispersion.The designed VGG-based CNN model outperforms the other four traditional machine-learning methods in different scenarios.Furthermore,the multitask learning model combined with MobileNet CNN is designed to improve the flexibility of the network.Compared with the designed VGG-based CNN,the MobileNet-based MTL does not need to train all the classes,and it can simultaneously monitor single parameter or multiple parameters without sacrificing accuracy,indicating great potential in various monitoring scenarios.
基金Project supported by the Japanese Ministry of Education,CultureScience(Monbusho),the Ministry of Mechanical Industry of China(MMIC),the State Education Commission of China(SECC)the National Natural Science Foundation of China(NSFC).
文摘Multivariate analysis and filtering techniques are widely applied to simultaneous and/or selective determination of multicomponent systems. Many methods among them are based on the principle of linear addition, while this principle does not always hold due to various physical and chemical factors. Using quite a different way, neural network (NN) based on a given learning rule, such as back propagation (BP) model, needs neither knowing nor using any form of input/output relationship. Particularly, NN can resolve various problems such as those with casual relation, those with fuzzy backgrounds, and those with uncertain inferential processes. NN was used by us to investigate quantitative struc-
