An Online Training Platform For Soft-sensor Based On Embedded Systems

There are great demands and challenges that neural network soft-sensor and support vector machine soft-sensor realized in embedded systems. To solve the soft measurement in complex petrochemical and metallurgy industrial, an online training platform for soft-sensor is designed based on OMAP-L137 (ARM and floating-point DSP dual-core controller). For real-time purpose, by analyzing the related factors, the optimized BP neural networks (BPNN) and support vector machine (SVM) are realized in the dual-core platform that supports online training of soft-sensor model. The testing of datasets demonstrates the excellent high-precision and real-time performance of the instrument.The development of an online training platform for soft-sensor based on embedded system is described in details. The main contributions and innovative points of the project are as follows:(1) Based on Texas Instrument's first dual-core controller for industrial applications, the software and hardware of platform are completed according to modular design concept. With a high degree of flexibility, platform consists of core board, expansion boards, panel board and interface board. The software is divided into low-level functions, application modules function and main function.(2) The optimized BPNN and SVM are realized in the dual-core platform that supports online training of soft sensor mode. For real-time purpose, the stop criterion of training, hidden-layer nodes and learning rate of BPNN, the penalty factors and epsilon-insensitive factor of SVM, gamma of RBF kernel function are tested and analyzed systematically. The appropriate strategy to choose those factors is developed, and the optimized BPNN and SVM can improve real-time performance without losing much precision.(3) Online training platform and strategy for soft-sensor are completed. To integrate the hardware and software of the embedded system, the optimized BPNN and SVM, an online training platform for soft-sensor is developed. A self-adaptive training strategy that combines the automatic and manual settings is also completed.The powerful and easy-to-use instrument is convenient for real-time detection and dynamic display. Unlike offline training instrument, the optimized BPNN and SVM are realized in the dual-core platform that supports online training of soft sensor or pattern recognition. The platform provides a low-cost, portable and real-time solution for the soft measurement of complex petrochemical industrial processing.