The invention discloses a dual-core embedded platform-based system for online training a support vector machine soft-sensing model, which mainly comprises a microcontroller, a memory, a human-computer interaction interface, a communication interface, a signal input and output interface and the like. An embedded platform-based method for online training the support vector machine soft-sensing model of the invention can maximally meet the requirements on real-time performance and intelligence because related parameters can be automatically set according to data characteristics and manual parameter setting is simultaneously supported. The system has a plurality of functions of being portable, supporting dynamic display and intelligent analysis and the like, realizes complex operations through the simple and direct-viewing human-computer interaction interface, overcomes the defects of difficulty in the realization of miniaturization, difficulty in the determination of related training parameters, the complex operations and the like, particularly the successful transplantation of the online training of a support vector machine algorithm, and provides a scheme which is low in cost, portability and high in the real-time performance for soft-sensing in complex processes in petrochemical technology and the like.