TW200630819A - Method of using intelligent theory to design heat dissipation module and device thereof - Google Patents

Abstract

The method of the present invention is provided to design a heat dissipation module in a finite three-dimensional space, which is most suitable for the finite three-dimensional space to achieve lowering the operating temperature of heat generating devices. In the design of the invention, eleven attribute vectors of the heat dissipation module are used as eleven input vectors of a back propagation neural network to respectively correspond to eleven output vectors and then sequentially perform learning and order-descending process, detecting process and re-learning process corresponding to the eleven input vectors respectively. In the learning and order-descending process, K-L expansion method is employed to convert the attribute vectors of designed parameters of the heat dissipation module onto the orthogonal main axes for preventing the attribute vectors from interfering with each other, and determine the minimum number of main axis vectors required for maintaining the estimation precision, thereby reducing the estimation complexity of the neural network. Further, the neural network uses the known input values and output values of the training samples (that is, the attribute vectors of the training samples and the corresponding design rules of the heat dissipation module in the learning sample database) to adjust the weight of each node so as to obtain a minimum error between the output value of the neural network and the actual output value of the sample, which is used as a target function to optimize the bonding value of each node thereby increasing the estimation precision of neural network. When the learning and order-descending process is completed, the weight of each node is fixed to facilitate estimation in the detecting process. In the detecting process, the attribute vectors of the sample under detection are processed by K-L expansion method for performing main axis conversion and order descending, and the order-descended axes are used as input vectors to perform heat dissipation module design and evaluation via the neural network. If there are erroneous samples in the evaluation process (wherein the erroneous sample represents a sample with an error actually evaluated by the neural network which is larger than a tolerance value), the erroneous samples are stored in the learning sample database to facilitate obtaining data for re-learning. In the re-learning process, with the erroneous samples added into the learning sample database, the K-L expansion method re-adjusts the orientations of the main axes and the neural network adjusts the weight of each node until not encountering erroneous determination for samples similar to the aforementioned erroneous samples in the subsequent detecting process, thereby increasing the estimation precision for the method and device using intelligent theory to design a heat dissipation module in a finite space, so as to design a heat dissipation module with a heat dissipating efficiency suitable for this finite space and thus decrease the working temperature of the heat generating device in the finite three-dimensional space.