An advanced coke temperature control system was developed. Final coke temperature measured when coke cake is pushed out is controlled in order to stabilize coke quality. It has two main control functions. (1) The target flue temperature is determined by using the heat transfer model of the coke chamber in order to attain the target final coke temperature under the charged blended coal properties and the planned coking time. (2) The input heat is manipulated by the optimal regulator based on the autoregressive model in order to attain the target flue temperature. This system was applied to a coke oven. As a result, the fluctuation of the coke temperature was decreased and the coke oven heat consumption was remarkably decreased.
Eaton (1962) has solved theoretical problems in computing time-optimal control of linear systems. But in the practical point of view, his method has two main drawbacks with respect to both the convergent rate and the accuracy of the solution. In this paper, we discuss a new indirect method for the terminal control problems, taking into account of the fact that such an approximate control is practically required as making the terminal state error norm smaller than some acceptable value. And combining the method with Eaton's method, a new effective algorithm of finding time optimal control is proposed. Also the integrating method to prevent the accumulation of round off error at the evaluation of the cost function is proposed. The accuracy and the efficiency of the method are shown by using several numerical examples.
When a continuous-time system is sampled by use of a zero order hold, all stable poles are transformed into the unit circle. However, there is no simple relation between the zeros of a continuous-time system and its sampled version. This paper has derived sufficient conditions which guarantee that a continous-time system with zeros in the left half plane is transformed to a sampled system with zeros inside the unit circle for all sampling periods. The criteria shown in this paper are represented by coefficients of the partial fraction expansion of a continuous-time transfer function.
Recently, many applications of neural networks (NN) to commercial products have been reported. We have already considered an application of NN to bill money recognition machines. We have also confirmed that the recognition ability using NN is better than the conventional method based on trial and error. However, the size of the NN may become an obstacle to produce commercial recognition machines. We propose a random mask method which is capable of condensing input pixels in a simple way to solve this problem. Using the proposed method, we show that various kinds of bill moneys such as Japanese yens, Korean wons, and American dollars can be classified by a conventional recognition machine and 32-bit computer even if they are mixed in a random way.