In this paper, an optimized extreme learning machine by adaptive particle swarm optimization (APSO- ELM) is proposed to be used in regression problems. In conventional extreme learning machine (ELM), random selection of some parameters causes the network has a number of sub-optimal parameters. This reduces the accuracy of the model, especially in working with noisy data. For solving this problem, adaptive particle swarm optimization (APSO) is used for optimizing Gaussian kernel parameters of the ELM. The proposed APSO-ELM model can be used as a soft sensor. To evaluate the performance of this soft sensor, it is used in an industrial process. Simulation results show that this model is more accurate and more robust than primal ELM model in working with noisy data