Abstract: A localization algorithm based on semi-supervised manifold learning is proposed. Manifold structures hidden in the information of received signal strength can be obtained by the algorithm. It is used to compute a subspace mapping function between the signal space and the physical space by using a small amount of labeled samples and a large amount of unlabeled samples. Existing theories and experiential signal propagation models need not to be known in the algorithm, and localization errors generated by inaccurate models can be avoided. A number of unlabeled samples were used to decrease the difficulty of collecting data and increase the practicality of the algorithm. Real nodes were used to setup the network in metallurgical industry environments. Experimental results in metallurgical enterprises show that a higher accuracy with much less calibration effort is achieved in comparison with RADAR localization systems.