The invention provides an assembly system and method based on a class feature knowledge base, which collects force sensing data corresponding to random assembly actions of a manipulator, pretreats thedata, and extracts force sensing data features by using a principal component analysis algorithm. Support vector machine (SVM) is used to classify the samples after feature extraction, to judge whether the assembly action corresponding to the current contact state is successful or not, and the failure data be multi-classified using the support vector machine- Extreme learning machine, optimizes the model parameters, builds the mapping model from force-sensing features to contact state features, and constructs the knowledge base of assembly contact state features. Based on the above-established knowledge base of assembly contact state classification, the current contact state is retrieved in real time, the manipulator is controlled to make corresponding posture adjustment according to theretrieval results, and the current new contact state is retrieved again, and the posture adjustment is continued until the assembly operation is successful.