Study On The Dynamic Soft Sensor Method For The Parameters Of Time-varying Industrial Processes

Due to the limitations of existing detection technology and other conditions, the industrial process can not directly measure many key industrial parameters online. Research on soft sensor technology provides key solutions for online measurement of these parameters, but the most studies focused on non-time-varying processes, the application of these technologies in time-varying industrial processes are many flaws. This paper discusses three kinds of time-varying process parameters for dynamic soft sensor modeling methods.The main works of this research are as follows:As static least squares support vector machine (LSSVM) model is large error in predicting the production process parameters, the paper proposes LSSVM algorithm based on dynamic moving window. The algorithm adds a moving window in the traditional LSSVM, selects the sample to construct the model by using incremental learning algorithm and decrement learning algorithm update strategy. Distillation Column parameters soft sensor experiments show that this method can improve the prediction accuracy and generalization performance of the model.From the perspective of the kernel function, this paper proposes a dynamic multicore squares support vector machine algorithm based on moving window, uses a combination of multi-core instead of a single core, enhanced model’s interpretability, And uses moving window technology, increased the recognition ability and update efficiency of model. Finally, the simulation results of a single variable functions and multivariable functions suggest that this algorithm has better adaptability and forecast results in solving convex optimization problems.Because the past dynamic modeling method to be re-modeled for each sample, and the calculation is very complicated, the paper proposes a dynamic LSSVM algorithm based on just in time(JUST), "the principle of similar inputs produce similar outputs" is adopted in the algorithm, and the algorithm utilizes the similarities and differences between the two time points before and after, rapidly recurrence the new models, greatly simplify the calculation. Through a time-varying function and distillation columns parameters soft sensor simulation experiments is proved that this method can be well adapted to the time-varying process, the model has better generalization performance.