Neural Networks In Control Systems

Authors: Atharva Karwande, Tejas Kolhe, Rushikesh Magar, Soham Kamble, Aboli Joshi.

The necessary scheme of a Neural Network is to define interconnected units, also called as “neuron”, where the connections have weights.  The networks have some input features which when passing through number of neurons, compute one or more outputs. There are three fundamental blocks of a Neural Network. 

1. Input Layer
2. Hidden Layers
3. Output Layers

Fig 1. Different Components in a Neural Network

With the advent of time and computational capabilities of a computer, ANN has time and again proved to be fruitful in prediction and control of systems. Various industries find their roots of working in the correct implementation of optimized neural networks. One the fields in which ANNs are being used exhaustively is Control.

There are usually two steps involved while implementing a Neural Network for control system:

1. System Identification
2. Control Design

System Identification is the process of estimation or inference of the relationship between the input and outputs. This is usually same for all the types of Neural Network Architecture. But the second stage i.e. Control Design varies drastically. System Identification deals with developing a neural network model for the plant we want to control. Design Control stage uses the neural network developed to design a controller that controls the plant.

There are three major Neural Network Architectures used:

1. Model Predictive Control.
2. NARMA-L2 Control.
3. Model Reference Control. 

Let us discuss each one in detail:

1. Model Predictive Control : The neural networks for Model Predictive Control can be based on linear as well as non-linear model predictive controller. We will be looking into the non linear part of this.

• System Identification : As discussed above, the first stage on any control system neural network is the system identification. This part involves training the neural network to represent the forward dynamics of the plant to be controlled. The error between the predicted response and the actual response acts a neural network training signal. The following represents the process:

Fig 2. Block diagram of predictive model control ^[1]

• Predictive Control : In the model predictive control, model predicts the plant responses over a time horizon.  This method is called as receding horizon technique. The training requires a batch algorithm for feed forward network. This is really fast and is optimized. The neural network works to optimize the cost function and then calculates the optimal input to the plant. This controller requires an online optimization algorithm which can be computationally expensive as after every lop the input is re optimized.

2. NARMA-L2 Control : It is also regarded as feedback linearization control. It is referred to as so when the plant model has a particular form. It is referred to as NARMA L2 when the plant control can be approximated by the same form. The central design of this type of controller is to alter non linear system dynamics into linear system dynamics by removing the non linearity.

• System Identification : The first step as stated is to develop a model using the previous data of the plant. The NARMA-L2 approximation is given by,

y(k + d) = h[y(k), y(k – 1), ¼, y(k – n + 1), u(k), u(k – 1), ¼, u(k – m + 1)]

• NARMA-L2 Controller : The subsequent control input is determined to force the plant output to follow a reference signal. The plant model neural network is trained with static back propagation. The controller is basically rearrangement of the plant model system, and requires minimal online computation which is beneficial. And the same can be done in a single neural network.

3. Model Reference Controller : The third type of Neural Network is Model Reference Control. This type of architecture uses two nets.

1. Plant Model Network.
2. Controller Network.

The plant model is identified using plant model neural network. Then secondly the controller neural network is trained for the plant model to follow reference model output. The architecture block diagram is as follows:

Fig 3. Block diagram of Model Reference Control Architecture^[1]

The Model reference control uses online computation. But in this case it is minimal. Also, it needs to train two separate neural networks. Training the controller can be computationally expensive because it requires the dynamic back propagation.

With the passage of time and solid research works, we can come up with even better models that can be implemented flawlessly in the domain of Control Systems.

REFERENCES

[1] Martin T. Hagan, Howard B. Demuth And Orlando De Jesús  : “An Introduction To The Use of Neural Networks in Control Systems”.