machine learning for advanced process control1. valida on model machine learning for advanced...
TRANSCRIPT
1. Valida�on model
Machine Learning for Advanced Process ControlYingzhao Lian1,2
1 ABB Corporate Research, Control & Optimization group, Segelhofstrasse 1, 5405 Baden, Switzerland, 2 École Polytechnique Fédérale de Lausanne (EPFL), School of Engineering, Mechanical engineering Section, Station 15, 1015 Lausanne, Switzerland,
References: E. D. Klenske, M. N. Zeilinger, B. Schölkopf, and P. Hennig. IEEE Transactions on Control SystemsTechnology, 24(1):110–121, 2016.C. V. Rao, S. J. Wright, and J. B. Rawlings。 Journal of optimization theory and applications, 99(3):723–757, 1998.
contact: [email protected]
As digitalization thrives in industrial applications, pervasive sensors give us wide access to data. Meanwhile, Machine Learning(ML) is a field of computer science dedicated to extracting knowledge from large data. This thesis focus on implementing ML techniques in advanced process control(APC), especially model identification and integration of machine learning into optimal controller design.
2. Gaussian process(GP) based modelGP is a is a collection of random variables, any finite number of which have consistent joint Gaussian distributions. It is a non-parametric probabilistic model, with which an auto-regressive model is used as an system identification model.
1. RNN based MPC
Neural network is a scalable inter-connected model, including fully-connected networks, recurrent networks(RNN), etc. We proposed a custom recurrent network structure with a special training mechaism, which is efficiently scalable to large system with strong compatibility with MPC.
Introduc�on
2. GP based itera�ve learning control
This thesis dicussed the mechanism of the fully-connected networks. Later, investigated both GP and neural networks based system identification, in particular, a custom structure and training mechanism is proposed. On the basis of the RNN model, MPC is tested, exploiting the possibility of MPC with RNN. At last, an explicit-like MPC and GP based iterative learning control is explored. Later on, we will focus on building rigorous machine learning based scalable control methods.
Conclusions & outlook
Machiner Learning based system iden�fica�on Machine learning integrated controller
Gaussian process and neural networks are used for system identification and tested on a four-tank system. This system is a multiple-input-multipleoutput(MIMO) system and its states, the level of the water in each tank, are coupled. In this thesis, it is assumed that only the tank 1 and tank 2 are observed and no prior knowledge, including the number of the tanks, is presumed before system identification. The pumps of this system are controlled by two PID controllers, whose feedback signals are the level of tank 1 and tank 2.
400 420 440 460 480 500 520 540 560 580 600
time (min)
-3
-2
-1
0
1
2
3
level
SEiso kernel 20-step forward prediction of tank 1
actual output
predictive mean
predictive variance
400 420 440 460 480 500 520 540 560 580 600
time (min)
-3
-2
-1
0
1
2
3
level
SEiso kernel 20-step forward prediction of tank 2
actual output
predictive mean
predictive variance
With this model, a multi-step-forward prediction is achieved by moment match method.
0 500 1000 1500 2000 2500 3000
time (s)
-4
-2
0
2
4
level of ta
nk 1
Response comparison
MPC without GP
desired output
MPC with GP compensation
0 500 1000 1500 2000 2500 3000
time (s)
-2
-1
0
1
2
level of ta
nk 2
MPC without GP
desired output
MPC with GP compensation
GP model is integrated into a linear model to capture non-linear dynamics. It is tesed both on active vibration compensation as well as iterative learning schemes. The following outcome shows the improvement in tracking iterative signals on four-tank system.
C. E. Rasmussen and C. K. Williams. Gaussian processes for machine learning, volume 1. MIT press Cambridge, 2006.R. Pascanu, C. Gulcehre, K. Cho, and Y. Bengio.. arXiv preprint arXiv:1312.6026, 2013.
The model identified with custom RNN structure is integrated into the MPC framework, where interior point method is used to solve the non-linear programming problem. Following outcomes show the performance improvement in tracking and the robustness against unknown noise, which are tested on four-tank system with low level PID controllers.
Lorem ipsum
3. Neural network based model