In order to enhance the optimization performance of PSO and cooperative PSO algorithms, an improved cooperative PSO algorithm (ICPSO) is proposed and a novel cooperative strategy is provided. During evolving process, the particles are divided into several sub swarms, each sub-swarm can make use of the experience of its own sub swarm and the whole particle swarm effectively. Each sub-swarm can search in its own domain adequately, which can avoid missing optimization direction. At the same time it can take advantage of the best solution found by the whole swarm periodically. The diversity of particles can be maintained by dividing the particles into several sub-swarms, thus can restrain local optimum phenomena. Experiments study on several classical and complex functions show that the improved algorithm outperforms basis PSO in robustness, converging speed and precision, global searching ability. Finally, ICPSO is applied to construct neural network fault diagnosis model for rotary machinery, then fault diagnosis system is designed. Research results show that the diagnosis system has the properties of high diagnosis accuracy and favorable stability