In order to enhance the real-time performance, adaptiveness and robustness of industrial robot assembly systems, drawing on the thoughts of people's perception of learning style,a flexible assembly method based on contact state perception development is presented. The position and force/torque of the robot end effector are used to describe the assembly contact state. Combined with the support vector data description and the improved extreme learning machine for the perception development of the contact state, a self-renewing and growing empirical knowledge base is formed to predict the assembly action of the robot and complete the flexible assembly task. In order to verify the effectiveness of the method, an experiment is carried out by taking the clamping assembly process of the small circuit breaker as an example. The experimental results show that the self-renewal of the assembly experience knowledge base is realized by the contact state perception development, and the flexible assembly of the robot is completed, which verify the feasibility and effectiveness of the method.