Time-delay neural network for the prediction of ... - Semantic Scholar
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 52, NO. 4, AUGUST 2003
1125
Time-Delay Neural Network for the Prediction of Carbonation Tower’s Temperature Dan Shi, Hongjian Zhang, and Liming Yang
Abstract—The carbonation tower is a key reactor to manufacturing synthetic soda ash using the Solvay process. Because of the complexity of the reaction in the tower, it is difficult to control such a nonlinear large-time-delay system with normal measurement instrumentation. To solve this problem, a time-delay neural network (TDNN) is used in the soft measurement model in this paper. A special back-propagation algorithm is developed to train the neural network. Compared with the model based on multilayered perceptron, it is shown that TDNN can describe the system’s dynamic character better and predict much more precisely. The influences of the input variables to the output of the model are analyzed with the online data. Analysis results show this model matches the reaction kinetics and the real operating conditions. Index Terms—Back-propagation (BP) algorithm, carbonation tower, soft measurement, temperature prediction, time-delay neural network (TDNN).
I. INTRODUCTION
T
HE Solvay process is the dominant method used in the world to manufacture synthetic soda ash today. And the carbonation tower is the most important reactor in the Solvay process. Fig. 1 shows its main structure. Because of the complexity of the mechanism and kinetics in carbonation process [1], [2], it is very hard to build an accurate mathematical model to describe such a reactor. It is even more difficult to set up an algorithm used to control the tower online based on the mechanism model. The temperature profile along with the height of the carbonation tower should be carefully controlled to ensure sodium bicarbonate crystals growing to the required size and to maximize soda ash yield. If the crystal size is too small, filtration of the product becomes very difficult and other problems arise at the calcining stage. Because of the nonlinear and large time delay character of the carbonation tower, it is very hard to control its temperature profile by using ordinary control strategy, such as PID which is used most widely in chemical plants. The neural network has been applied successfully for the time-delay compensation of nonlinear processes [3], [4]. Particularly, time-delay neural network (TDNN), originally designed for speech recognition, performs excellent in the modeling of dynamic system with large time delay [5], [6]. In this paper, TDNN is used to predict the temperature of the 12th layer in the carbonation tower, which is a key variable in the temperature profile. Comparing with the model based Manuscript received June 15, 2002; revised January 9, 2003. The authors are with the National Laboratory of Industrial Control Technology, Department of Control Science and Engineering, Zhejiang University, Hangzhou, China. Digital Object Identifier 10.1109/TIM.2003.815985
Fig. 1.
Structure of Carbonation Tower.
on multilayered perceptron (MLP), the prediction of TDNN model is more accurate and robust. Furthermore, based on the TDNN model, the effectiveness factor of each input variable is analyzed. II. SOFT MEASUREMENT MODEL A. Model Structure Fig. 2 shows the main structure of the prediction model based on TDNN. The soft measurement model has 9 input variables and the output is the prediction of the temperature of the 12th layer in the carbonation tower. The time interval of the inputs is 1 min. There are totally 19 variables that can be measured online and used to predict the 12th layer’s temperature. These variables include five temperatures (5th, 7th, 12th, 17th, and 23rd layer) that can be measured in the tower, the temperature and flowrate of reactants (the pre-carbonated ammoniacal brine, calcining gas I and calcining gas II), product (the carbonated ammoniacal brine), chilled water and cooling water and the pressure of
0018-9456/03$17.00 © 2003 IEEE
1126
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 52, NO. 4, AUGUST 2003
Fig. 3. Structure of TDNN.
The last unit of the prediction model transforms the output of to 1, to a TDNN, which is a normalized value ranging from prediction value. Also an FIR filter, whose length is ten, is used to enhance the stability of the model output. B. Training Algorithm For TDNN Fig. 2. Structure of prediction model.
gases (calcining gas I and calcining gas II). Based on the reaction mechanism and experiences of the operators, 9 of them are chosen as the model inputs. Especially, the temperature of 12th layer is not one of the inputs because of its too strong effect and correlation with the model output, which is the 3-min prediction of that temperature. The Pretreatment of the input variables includes filtering and normalizing of the input variables that are online measurement values of a carbonation tower. Usually, the online data include much noise, which is very harmful to the stability of the model. Especially for a model based on artificial neural network, a small change of input can lead to large output change. Since the model output is temperature, which does not change very fast, an finite impulse response (FIR) filter is designed to treat the input data and the length of filter’s window is 1 min. Normalizing is another important function for the input pretreatment. For the convenience of artificial neural network design, the input of the and 1. network is normalized to TDNN is a multilayer feedforward network whose hidden neurons and output neurons are replicated across time. Fig. 3 and are input, hidden shows the structure of TDNN. , and are the weight of and output neurons, respectively. the connection between neurons. It is very similar with multilayered perceptron (MLP). It has three layers, including input layer, hidden layer and output layer. Different from MLP, the weights in TDNN are not scalars but vectors. The length of the vector should be decided carefully. If the length of the weight vector is too short, the model would not be able to represent the large-time-delay property of the system. On the other hand, the model can not response the input changes in time if it is too long. In the model developed in this paper, the weights from the input layer to the hidden layer have a length of 17. And the length for the weights from the hidden layer to the output layer is 9. The input layer of TDNN used in this model has nine neurons for the nine input variables, including the flowrate of carbonated ammoniacal brine, the flowrate of pre-carbonated ammoniacal brine, calcining gas I and calcining gas II, the temperature of 5th, 7th, 17th, and 23rd layer of the tower and the quantity of heat removed by the cooling water and chilled water. There are 21 neurons in the hidden layer and one neuron in the output layer.
Because of the similar structure of TDNN and MLP, backpropagation can be applied to train the TDNN. Back-propagation through time [7] is used to train the TDNN of this model. The performance index of the algorithm is (1) is the error between model output and training goal where for each input at discrete time . Then, the chain rule of each error’s gradient can be expressed as (2) (2) where the weight of the neuron in the th layer to the neuron in the ( )th layer; the output of the neuron in the th layer. For the layers :
Others
(3)
is the input of the neuron in the th layer and is a where vector of the weight, whose length is . From (1)–(3), (4) and (5) (seen at the bottom of the next page) can be obtained as the weight modification algorithm [8] used to train the TDNN, where
the coefficient used to adjust the step of BP algo; rithm, usually the input of the neuron in the th layer; the Sigmoid function used in the neuron. C. Model Input Evaluation There are totally 19 variables that can be measured online and used to predict the 12th layer’s temperature. So it is an important issue to evaluate the effectiveness of each measurable variable on the model output, which will be helpful to choose the model is defined inputs. For this purpose, an effectiveness factor
SHI et al.: TIME-DELAY NEURAL NETWORK
1127
Fig. 4.
Comparison between the model output and actual temperature based on TDNN model.
Fig. 5.
Comparison between the model output and actual temperature based on MLP model.
as (6) to show the impact of each input model
on the output of the
(6)
where
is the output of TDNN model with and is the total the input of is the average of , number of inputs. and is the total number of the samples. III. EXPERIMENT RESULTS Using a training set from the historical data of a carbonation tower, the TDNN model is trained to predict the 12th layer’s temperature of the carbonation tower 3 min ahead by the training
algorithm mentioned above. Fig. 4 is the comparison between actual temperature and the TDNN model’s output with inputs from a validation data set. Using the same training set, a three-layer MLP model is trained. Fig. 5 is the comparison between actual temperature and the MLP model’s output. The MLP model has the same number of nodes for input and output layer as the TDNN model, while it has 31 nodes for the hidden layer. Both of the two figures show a sample of about 900 points, i.e., is 15 hours’ data. Fig. 5 shows that the model based on MLP fails to predict the actual value of output. Table I compares the accuracy of two models based on TDNN and MLP with the same input data shown as Figs. 4 and 5. It shows that TDNN performs much better than MLP to model such a nonlinear object with large time delay. of each input is compared in The effectiveness factor Fig. 6. It shows that the flowrate of calcining gas II has the largest impact on the 12th layer’s temperature. It fits well with the real situation. Operators usually adjust the flowrate of the calcining gas II first to control the temperature profile of the carbonation tower. Another important fact can be found from
(4) (5)
1128
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 52, NO. 4, AUGUST 2003
TABLE I COMPARISON BETWEEN THE ACCURACY MODELS BASED ON TDNN AND MLP
OF
this model, the effectiveness factor is defined to analyze each variable’s impact on the model output. Such analysis shows that the model based on TDNN not only can predict the actual temperature, but also fits the kinetics of the carbonation tower’s operation. With this model, a better control strategy can be developed to ensure the carbonating process is more stable and then to improve the yield. REFERENCES [1] P. V. Danckwerts and K. M. Mcneil, “The absorption of carbon dioxide into aqueous amine solution and the effects of catalysis,” Trans. Inst. Chem. Engr., vol. 45, pp. 32–47, 1967. [2] X. Li, “Modeling and simulation of carbonation process of ammoniacal brine. II. reactor/crystallizer model of carbonation columns,” Chem. Reaction Eng. Technol., vol. 15, no. 4, pp. 357–363, 1999. [3] G. W. Ng and P. A. Cook, “Real-time control of systems with unknown and varying time-delays, using neural networks,” Eng. Applicat. Artif. Intell., vol. 11, pp. 401–409, 1998. [4] Y. Tan and A. V. Cauwenberghe, “Neural-network-based d-step-ahead predictors for nonlinear systems with time delay,” Eng. Applicat. Artif. Intell., vol. 12, pp. 21–35, 1999. [5] Y. Yamashita, “Time delay neural networks for the classification of flow regimes,” Comput. Chem. Eng., vol. 21, Suppl., pp. S367–S371, 1997. [6] J. Zhu, J. Zurcher, M. Rao, and M. Q.-H. Meng, “An on-line wastewater quality predication system based on a time-delay neural network,” Eng. Applicat. Artif. Intell., vol. 11, pp. 747–758, 1998. [7] P. J. Werbos, “Backpropagation through time: What it does and how to do it,” Proc. IEEE, vol. 78, pp. 1550–1560, Oct. 1990. [8] S. Hu, The Application Technology of Artificial Neural Network: National Univ. Defence Technology, Changsha, China, 1985, pp. 185–188.
Dan Shi received the B.Sc. degree in automation instrumentation from the Control Science and Engineering Department, Zhejang University, Hangzhou, China. He is currently pursuing the Ph.D. degree at the Chemical Engineering Department, University of California, Los Angeles. His current research interests are in the control of high velocity oxygen fuel thermal spray and the modeling and control of the coating formation. Fig. 6. Comparison of the effectiveness factor .
Fig. 6 is that the heat removed by the cooling water and chilled water has very small impact on the 12th layer’s temperature. It reveals another reality that the cooling ability for that tower is insufficient. 1) Flowrate of calcining gas II. 2) Temperature of the 23rd layer. 3) Flowrate of calcining gas I. 4) Temperature of the fifth layer. 5) Temperature of the seventh layer. 6) Temperature of the 17th layer. 7) Flowrate of carbonated ammoniacal brine. 8) Heat removed by the chilled water. 9) Heat removed by the cooling water. IV. CONCLUSION Based on the above analysis, it can be concluded that the prediction model presented in this paper can predict the 12th layer’s temperature of the carbonation tower accurately. Because of the serious nonlinear character and large time delay of the carbonation tower, the proposed model based on TDNN is more accurate than MLP or other linear regression method. Based on
Hongjian Zhang was born in July 1961. He received the B.Sc. and M.Sc. degrees from Zhejiang University, Hangzhou, China, in 1982 and 1985, respectively, and the Ph. D. degree from University of Shanghai for Science and Technology, Shanghai, China, in 1988. From 1992 to 1994, he was a Visiting Scholar with the Department of Chemical Engineering, Katholieke Universiteit, Leuven, Belgium. Currently, he is a Professor with the Deparment of Control Science and Engineering, Zhejiang University. He is also the permanent staff of the National Laboratory of Industrial Control Technology. His research interests include measurement theory and techniques for process parameters, signal processing, and its algorithms. He is currently Vice Dean of the Information College of Zhejiang University and Chairman of the Control Science and Engineering Department. Dr. Zhang is a council member of China Metrology Society (CMS) and Automation Society of Zhejiang Province and the member of standing committee of Process Measurement and Control Committee of the China Instrument Society.
Liming Yang received the B.Sc. degree in 1983 from the Department of Physics, Yunan University, China. She is a Senior Engineer with the Department of Control Science and Engineering, Zhejiang University, Hangzhou, China. From 1984 to 1987, she was a Research Assistant with the Science Academy of China, working on signal measurement and instrumentation. Since 1987, she has been an Engineer and a Senior Engineer with Zhejiang University, working on instrumentation and process control and teaching a course in process automation and instrumentation. She has undertaken more than ten projects on chemical and biochemical process control and has published 20 papers.
1125
Time-Delay Neural Network for the Prediction of Carbonation Tower’s Temperature Dan Shi, Hongjian Zhang, and Liming Yang
Abstract—The carbonation tower is a key reactor to manufacturing synthetic soda ash using the Solvay process. Because of the complexity of the reaction in the tower, it is difficult to control such a nonlinear large-time-delay system with normal measurement instrumentation. To solve this problem, a time-delay neural network (TDNN) is used in the soft measurement model in this paper. A special back-propagation algorithm is developed to train the neural network. Compared with the model based on multilayered perceptron, it is shown that TDNN can describe the system’s dynamic character better and predict much more precisely. The influences of the input variables to the output of the model are analyzed with the online data. Analysis results show this model matches the reaction kinetics and the real operating conditions. Index Terms—Back-propagation (BP) algorithm, carbonation tower, soft measurement, temperature prediction, time-delay neural network (TDNN).
I. INTRODUCTION
T
HE Solvay process is the dominant method used in the world to manufacture synthetic soda ash today. And the carbonation tower is the most important reactor in the Solvay process. Fig. 1 shows its main structure. Because of the complexity of the mechanism and kinetics in carbonation process [1], [2], it is very hard to build an accurate mathematical model to describe such a reactor. It is even more difficult to set up an algorithm used to control the tower online based on the mechanism model. The temperature profile along with the height of the carbonation tower should be carefully controlled to ensure sodium bicarbonate crystals growing to the required size and to maximize soda ash yield. If the crystal size is too small, filtration of the product becomes very difficult and other problems arise at the calcining stage. Because of the nonlinear and large time delay character of the carbonation tower, it is very hard to control its temperature profile by using ordinary control strategy, such as PID which is used most widely in chemical plants. The neural network has been applied successfully for the time-delay compensation of nonlinear processes [3], [4]. Particularly, time-delay neural network (TDNN), originally designed for speech recognition, performs excellent in the modeling of dynamic system with large time delay [5], [6]. In this paper, TDNN is used to predict the temperature of the 12th layer in the carbonation tower, which is a key variable in the temperature profile. Comparing with the model based Manuscript received June 15, 2002; revised January 9, 2003. The authors are with the National Laboratory of Industrial Control Technology, Department of Control Science and Engineering, Zhejiang University, Hangzhou, China. Digital Object Identifier 10.1109/TIM.2003.815985
Fig. 1.
Structure of Carbonation Tower.
on multilayered perceptron (MLP), the prediction of TDNN model is more accurate and robust. Furthermore, based on the TDNN model, the effectiveness factor of each input variable is analyzed. II. SOFT MEASUREMENT MODEL A. Model Structure Fig. 2 shows the main structure of the prediction model based on TDNN. The soft measurement model has 9 input variables and the output is the prediction of the temperature of the 12th layer in the carbonation tower. The time interval of the inputs is 1 min. There are totally 19 variables that can be measured online and used to predict the 12th layer’s temperature. These variables include five temperatures (5th, 7th, 12th, 17th, and 23rd layer) that can be measured in the tower, the temperature and flowrate of reactants (the pre-carbonated ammoniacal brine, calcining gas I and calcining gas II), product (the carbonated ammoniacal brine), chilled water and cooling water and the pressure of
0018-9456/03$17.00 © 2003 IEEE
1126
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 52, NO. 4, AUGUST 2003
Fig. 3. Structure of TDNN.
The last unit of the prediction model transforms the output of to 1, to a TDNN, which is a normalized value ranging from prediction value. Also an FIR filter, whose length is ten, is used to enhance the stability of the model output. B. Training Algorithm For TDNN Fig. 2. Structure of prediction model.
gases (calcining gas I and calcining gas II). Based on the reaction mechanism and experiences of the operators, 9 of them are chosen as the model inputs. Especially, the temperature of 12th layer is not one of the inputs because of its too strong effect and correlation with the model output, which is the 3-min prediction of that temperature. The Pretreatment of the input variables includes filtering and normalizing of the input variables that are online measurement values of a carbonation tower. Usually, the online data include much noise, which is very harmful to the stability of the model. Especially for a model based on artificial neural network, a small change of input can lead to large output change. Since the model output is temperature, which does not change very fast, an finite impulse response (FIR) filter is designed to treat the input data and the length of filter’s window is 1 min. Normalizing is another important function for the input pretreatment. For the convenience of artificial neural network design, the input of the and 1. network is normalized to TDNN is a multilayer feedforward network whose hidden neurons and output neurons are replicated across time. Fig. 3 and are input, hidden shows the structure of TDNN. , and are the weight of and output neurons, respectively. the connection between neurons. It is very similar with multilayered perceptron (MLP). It has three layers, including input layer, hidden layer and output layer. Different from MLP, the weights in TDNN are not scalars but vectors. The length of the vector should be decided carefully. If the length of the weight vector is too short, the model would not be able to represent the large-time-delay property of the system. On the other hand, the model can not response the input changes in time if it is too long. In the model developed in this paper, the weights from the input layer to the hidden layer have a length of 17. And the length for the weights from the hidden layer to the output layer is 9. The input layer of TDNN used in this model has nine neurons for the nine input variables, including the flowrate of carbonated ammoniacal brine, the flowrate of pre-carbonated ammoniacal brine, calcining gas I and calcining gas II, the temperature of 5th, 7th, 17th, and 23rd layer of the tower and the quantity of heat removed by the cooling water and chilled water. There are 21 neurons in the hidden layer and one neuron in the output layer.
Because of the similar structure of TDNN and MLP, backpropagation can be applied to train the TDNN. Back-propagation through time [7] is used to train the TDNN of this model. The performance index of the algorithm is (1) is the error between model output and training goal where for each input at discrete time . Then, the chain rule of each error’s gradient can be expressed as (2) (2) where the weight of the neuron in the th layer to the neuron in the ( )th layer; the output of the neuron in the th layer. For the layers :
Others
(3)
is the input of the neuron in the th layer and is a where vector of the weight, whose length is . From (1)–(3), (4) and (5) (seen at the bottom of the next page) can be obtained as the weight modification algorithm [8] used to train the TDNN, where
the coefficient used to adjust the step of BP algo; rithm, usually the input of the neuron in the th layer; the Sigmoid function used in the neuron. C. Model Input Evaluation There are totally 19 variables that can be measured online and used to predict the 12th layer’s temperature. So it is an important issue to evaluate the effectiveness of each measurable variable on the model output, which will be helpful to choose the model is defined inputs. For this purpose, an effectiveness factor
SHI et al.: TIME-DELAY NEURAL NETWORK
1127
Fig. 4.
Comparison between the model output and actual temperature based on TDNN model.
Fig. 5.
Comparison between the model output and actual temperature based on MLP model.
as (6) to show the impact of each input model
on the output of the
(6)
where
is the output of TDNN model with and is the total the input of is the average of , number of inputs. and is the total number of the samples. III. EXPERIMENT RESULTS Using a training set from the historical data of a carbonation tower, the TDNN model is trained to predict the 12th layer’s temperature of the carbonation tower 3 min ahead by the training
algorithm mentioned above. Fig. 4 is the comparison between actual temperature and the TDNN model’s output with inputs from a validation data set. Using the same training set, a three-layer MLP model is trained. Fig. 5 is the comparison between actual temperature and the MLP model’s output. The MLP model has the same number of nodes for input and output layer as the TDNN model, while it has 31 nodes for the hidden layer. Both of the two figures show a sample of about 900 points, i.e., is 15 hours’ data. Fig. 5 shows that the model based on MLP fails to predict the actual value of output. Table I compares the accuracy of two models based on TDNN and MLP with the same input data shown as Figs. 4 and 5. It shows that TDNN performs much better than MLP to model such a nonlinear object with large time delay. of each input is compared in The effectiveness factor Fig. 6. It shows that the flowrate of calcining gas II has the largest impact on the 12th layer’s temperature. It fits well with the real situation. Operators usually adjust the flowrate of the calcining gas II first to control the temperature profile of the carbonation tower. Another important fact can be found from
(4) (5)
1128
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 52, NO. 4, AUGUST 2003
TABLE I COMPARISON BETWEEN THE ACCURACY MODELS BASED ON TDNN AND MLP
OF
this model, the effectiveness factor is defined to analyze each variable’s impact on the model output. Such analysis shows that the model based on TDNN not only can predict the actual temperature, but also fits the kinetics of the carbonation tower’s operation. With this model, a better control strategy can be developed to ensure the carbonating process is more stable and then to improve the yield. REFERENCES [1] P. V. Danckwerts and K. M. Mcneil, “The absorption of carbon dioxide into aqueous amine solution and the effects of catalysis,” Trans. Inst. Chem. Engr., vol. 45, pp. 32–47, 1967. [2] X. Li, “Modeling and simulation of carbonation process of ammoniacal brine. II. reactor/crystallizer model of carbonation columns,” Chem. Reaction Eng. Technol., vol. 15, no. 4, pp. 357–363, 1999. [3] G. W. Ng and P. A. Cook, “Real-time control of systems with unknown and varying time-delays, using neural networks,” Eng. Applicat. Artif. Intell., vol. 11, pp. 401–409, 1998. [4] Y. Tan and A. V. Cauwenberghe, “Neural-network-based d-step-ahead predictors for nonlinear systems with time delay,” Eng. Applicat. Artif. Intell., vol. 12, pp. 21–35, 1999. [5] Y. Yamashita, “Time delay neural networks for the classification of flow regimes,” Comput. Chem. Eng., vol. 21, Suppl., pp. S367–S371, 1997. [6] J. Zhu, J. Zurcher, M. Rao, and M. Q.-H. Meng, “An on-line wastewater quality predication system based on a time-delay neural network,” Eng. Applicat. Artif. Intell., vol. 11, pp. 747–758, 1998. [7] P. J. Werbos, “Backpropagation through time: What it does and how to do it,” Proc. IEEE, vol. 78, pp. 1550–1560, Oct. 1990. [8] S. Hu, The Application Technology of Artificial Neural Network: National Univ. Defence Technology, Changsha, China, 1985, pp. 185–188.
Dan Shi received the B.Sc. degree in automation instrumentation from the Control Science and Engineering Department, Zhejang University, Hangzhou, China. He is currently pursuing the Ph.D. degree at the Chemical Engineering Department, University of California, Los Angeles. His current research interests are in the control of high velocity oxygen fuel thermal spray and the modeling and control of the coating formation. Fig. 6. Comparison of the effectiveness factor .
Fig. 6 is that the heat removed by the cooling water and chilled water has very small impact on the 12th layer’s temperature. It reveals another reality that the cooling ability for that tower is insufficient. 1) Flowrate of calcining gas II. 2) Temperature of the 23rd layer. 3) Flowrate of calcining gas I. 4) Temperature of the fifth layer. 5) Temperature of the seventh layer. 6) Temperature of the 17th layer. 7) Flowrate of carbonated ammoniacal brine. 8) Heat removed by the chilled water. 9) Heat removed by the cooling water. IV. CONCLUSION Based on the above analysis, it can be concluded that the prediction model presented in this paper can predict the 12th layer’s temperature of the carbonation tower accurately. Because of the serious nonlinear character and large time delay of the carbonation tower, the proposed model based on TDNN is more accurate than MLP or other linear regression method. Based on
Hongjian Zhang was born in July 1961. He received the B.Sc. and M.Sc. degrees from Zhejiang University, Hangzhou, China, in 1982 and 1985, respectively, and the Ph. D. degree from University of Shanghai for Science and Technology, Shanghai, China, in 1988. From 1992 to 1994, he was a Visiting Scholar with the Department of Chemical Engineering, Katholieke Universiteit, Leuven, Belgium. Currently, he is a Professor with the Deparment of Control Science and Engineering, Zhejiang University. He is also the permanent staff of the National Laboratory of Industrial Control Technology. His research interests include measurement theory and techniques for process parameters, signal processing, and its algorithms. He is currently Vice Dean of the Information College of Zhejiang University and Chairman of the Control Science and Engineering Department. Dr. Zhang is a council member of China Metrology Society (CMS) and Automation Society of Zhejiang Province and the member of standing committee of Process Measurement and Control Committee of the China Instrument Society.
Liming Yang received the B.Sc. degree in 1983 from the Department of Physics, Yunan University, China. She is a Senior Engineer with the Department of Control Science and Engineering, Zhejiang University, Hangzhou, China. From 1984 to 1987, she was a Research Assistant with the Science Academy of China, working on signal measurement and instrumentation. Since 1987, she has been an Engineer and a Senior Engineer with Zhejiang University, working on instrumentation and process control and teaching a course in process automation and instrumentation. She has undertaken more than ten projects on chemical and biochemical process control and has published 20 papers.
