In view of the problem that objective function is difficult to calculate online

a process-data-based online economic performance grading assessment method is proposed. Autoregressive projection to latent structure algorithm (AR-PLS) is used to decompose input data matrix. Then

offline models of different performance grades are established in the subspace related to output latent variable

and thus the unrelated-output variation is eliminated. Afterwards

a similarity-grid model is designed using strategy of “calibration zoning

then comparing the similarity of adjacent grades”. The method can divide process performance into steady performance grade state and transition state. Performance variations caused by factors excluded in offline model can be identified to enrich the offline database further. When the evaluation result is nonoptimal

the cause of performance variation can be diagnosed by the variable contribution. Finally

ethylene cracking process data test shows the method can help to detect performance deviation in time and accurately.