When the fault detection or fault diagnosis problem is considered as a binary or multiple class classification problem, there are some challenges, i.e., highly dimensional input variables, high correlation among some input variables, overlap among the input variable spaces of different fault classes and invisible distribution of fault classes. To tackle the above problems, a novel chemical process monitoring and fault diagnosis approach, which integrates canonical variate analysis (CVA) with multiple self-organizing map (multi SOM), is proposed. CVA is employed to extract fault classification feature information as much as possible, to reduce dimension and to eliminate correlation via CVA features. Based on CVA features, multi SOM, whose structure is similar as a tree structure, is employed to distinguish all fault classes clearly. The output plane of the root SOM is obtained based on the CVA features of all fault classes. According to the root plane, each mixing region is distinguished and the output plane of one second layer SOM is further employed to partition fault classes within the mixing region. In this way, each mixing region in father plane is further partitioned by one his son plane until there is no mixing region on the output planes of all leaf SOMs. A case study on the Tennessee Eastman process benchmark shows the effectiveness and feasibility of the proposed fault diagnosis and process monitoring approach.