This paper proposes a particle swarm optimization inversion algorithm for avoiding the dependency on initial model and local solution. This algorithm is applied to induction logging inversion on the models of different thickness layers, and yields consistent results with the models in the noise-free case. When noises of 5%, 10% and 20% are added to the models, the results of inversions remain fairly good. Numerical experiment results demonstrate that this particle swarm optimization inversion algorithm has advantages of being independent of initial models, capable of global optimization and anti-noise, and making induction logging data inversion more effective.