Also, neuro controller, the reverse trained neural network of neuro emulator, is designed for the speed control system of a diesel engine. The selective neuro controller is proposed for the purpose of improving the neuro controller's performance and by combining a PI controller with the proposed controller, the efficiency of this combination speed control system of a diesel engine is determined. Due to the non-linearity of many parameters of the diesel engine system, the control parameters must be adjusted to be applicable to real system in the whole area through many experiments.
It is difficult to find the right parameters that control the diesel engine system satisfactorily in all situations. In this paper, even if the control plant is nonlinear and the parameters of the plant system are not clear, in order to forcefully control the speed of the diesel engine used to drive the generator, it is proposed to design the engine speed control system with oil using neural network. In Chapter 3, an emulator is designed to model the forward dynamics of a diesel engine by various back-propagation algorithms.
The neuro emulator is trained using training data collected from a real diesel engine system under different back propagation algorithms.
Neural Networks
If a network has only input and output layers, it is called a single-layer network. On the other hand, if the network contains one or more hidden layers in addition to the input and output layers, it is called a multilayer network. NNs can be classified as feedforward or recurrent networks based on the direction of flow of the input signals.
In a NN whose input flows in only one direction from input to output, the NN is called a feedforward network. If, on the other hand, the network has a feedback loop, it is called a recurrent network.
Learning of Neural Networks
The network structure and node interconnection are very important in determining the performance of the network. In this particular learning model, the learning rule adjusts the network weights based on the difference between the desired output and the network output, usually in such a way that the difference is minimized. Unsupervised learning is required in cases where the network is only provided with an input signal.
In this learning mode, the weights are adjusted by a learning rule based solely on the input signals and the current network outputs. The network's weights are adjusted to develop an input and output behavior that maximizes the chance of receiving a reward and minimizes that of receiving a penalty. Backpropagation is a supervised learning technique based on the gradient descent method, which minimizes a square error criterion measured at the output layer by changing network weights.
Learning can be performed by following comparisons to reduce errors for the output layer. There are often cases of falling into local minima when a system is learned with simple backpropagation and the error cannot be reduced. The former tends to reduce the possibility of the sum-squared error staying in local minima with high errors, and the latter tends to train faster than just backpropagation algorithm.
Gradient descent is a very simple search technique where parameters such as weights and biases are shifted in the opposite direction to the gradient of the error. Applying momentum changes this only slightly by making the changes proportional to the running average of the gradient. The Levenberg-Marquardt algorithm modifies parameters according to rules and is more powerful than gradient descent.
Initialization of Neural Networks
This method re-adjusts the weights between the input layer and the hidden layer before training and results in reduced training time in some applications, including the XOR operation [5]. As mentioned above, NNs include parameters such as learning rate and momentum chosen to optimize learning performance. In order to obtain a LV to emulate a diesel engine system, the LV input-output ratio must be designed considering the characteristics of the diesel engine system.
For this, a model of the diesel engine system represented by a block diagram is considered in this chapter. Then, NNs are trained using the above algorithms and compared in terms of training efficiency. The trained network is used for the speed control system as the neuro-emulator for diesel engines in the next chapter.
The combustion system and rotation can be modeled as a first-order system and shown in Fig. This engine system used for generator maintains 1800 [rpm] for four pole diesel generator. So, at a high speed of 1800 [rpm], the combustion system taken into account with the dead time obtained from.
In this study, the diesel engine system is assumed to be a third-order system considering the dead time.
To determine the correct number of hidden nodes, the same structure as Fig. As expected, if the number of hidden nodes is small, the network error tends not to converge at large error. However, if the number of hidden nodes is too large, the convergence tends to be improved, but the number of epochs for training is large.
Based on the trial-and-error experiment[5], when the node number is chosen from 9 to 12, this network meets the convergence speed. The output layer consists of only one node corresponding to the estimated motor output value. Tangent sigmoid function is used for activation function of input neurons and hidden neurons and linear function for output layer[2].
Data Collection
Data acquisition is performed from a diesel engine speed control system consisting of four parts; a digital governor, an actuator, an MPU and a computer. When MPU approaches close to the flywheel teeth, it generates pulse related to rpm and this rpm data gives feedback to the digital governor. According to the difference between reference rpm and feedback rpm, appropriate control inputs are generated.
This control input signal causes an actuator to operate and the amount of fuel injected into the engine is adjusted according to the control input. And, the reference, rpm and control input are transmitted to the PC and saved in the form of text files using RS232 serial communication.[4].
In the case of applying BPM and BPX, the mean square error did not reach the error target despite 4000 epochs, but the BPLM algorithm needed 4000 epochs to converge to the error target. To construct the neurocontrol scheme, a neurocontroller is designed using training data.
Neuro Controller Design
This architecture is similar to the plant identification scheme above, but the input signals of the networks are different from the plant modeling case. Since this inverse identification is taken for control purposes, it must generate the control signal in relation to the output signal. Inversely trained networks will generate suitable outputs and will be able to control the speed of the diesel engine instead of the controllers that have been used conventionally.
The following graph compares the actual control input with the control input generated by inversely trained neural networks, the so-called neurocontroller in the case of using confirmed data.
However, after the system is stimulated by a perturbation, it takes a long time to return to the reference.
As shown in Figure 4.10, the response approaches the reference level faster than the control scheme using only the neurocontroller. A neural emulator and a neurocontroller for a diesel engine speed control system are proposed in the thesis. To improve the control efficiency, a selective neurocontroller training method is proposed, and it is proven that the neurocontroller trained by this method is more effective in controlling the diesel engine speed in the presence of an existing disturbance.
For a quick response in the case of an existing disturbance, a combined control system is proposed, which is combined with a neuro controller and a conventional PI controller. The simulated results show that the combined control system with neuro controller and conventional PI controller effectively controls the speed of the diesel engine driven generator in case of existing disturbance. In the future, a study should be done on the development of a dedicated controller to implement this combined neurocontrol system and real diesel engine application.
지난 2년간의 대학원 생활을 되돌아보면 많은 추억이 떠오릅니다. 하지만 이 소중한 추억을 저 혼자 책임지는 것이 아니기 때문에, 참여해주신 많은 분들께 말로라도 감사의 마음을 전하고 싶습니다. 제가 부담 없이 다가갈 수 있도록 항상 친절하게 많은 도움을 주시는 천행춘 원장님께 감사의 말씀을 전하고 싶습니다.
그리고 대학원 생활 2년 동안 누구보다 많은 시간을 저와 함께 해주시고, 건전한 조언도 해주시고, 지칠 때마다 격려의 말씀으로 옆에서 함께 해주신 영일님께도 감사하다는 말씀 전하고 싶습니다. 그리고 1년 동안 선배로서, 친구로서 도움과 격려를 해준 인호에게도 감사 인사를 전하고 싶습니다. 2년 동안 연구실에서 동고동락했던 복산과의 추억도 소중히 간직하고 싶습니다.
