Research On Automatic Cluster System Of SOM Network With Sensitive Parameter

Funding by financial supports from NSFC (Grant No 60970073), automatic cluster algorithm of SOM for multi-dimensional data was mainly researched. Automatic cluster for multi-dimensional data was the main method of optimizing dynamic flow soft sensor.The methord was mainly used to optimize training samples. The methord challenged one of the topics of dynamic flow soft sensor. Aiming at real-time and accuracy was needed by dynamic flow soft sensor, improving convergence rate and automatic cluster quality of automatic cluster algorithm was researched. The issue was very important in theory and reality significance for dynamic flow soft sensor.Firstly, the SOM algorithm with sensitive parameter was proposed. Through the introduction of sensitive parameter, weight adjustment formula was improved. The convergence rate was improved and training accuracy was ensured.Secondly, aiming at the low automatic cluster accuracy because of noise pollution, a new threshold function is constructed in weighted average method. Then, the noise impact on automatic cluster accuracy which appears in the character of Gaussian distribution was reduced, and the noise was removed by wavelet threshold method.Thirdly, aiming at the low automatic cluster accuracy because of not meticulous submerge stop parameters.For higher automatic cluster qulity, an expanded range Sigmoid function was introduced to improve the setting of submerge stop parameter. The separation of local maxima was prevented.Finally, the whole automatic cluster system was realized by MATLAB R2007.The experiment result indicated that the real-time and accuracy of automatic cluster algorithm was improved by the whole automatic cluster system. The improved method proposed in this paper obtains better automatic cluster effects and beneficially explores optimizing dynamic flow soft sensor's training samples. The proposed algorithms have great potential in other field.