Mode identification method based on model generalization and probability density difference for process data

The multimode problem exists widely in the production process, resulting in a great difference in the process data distribution, which causes a mismatch between the modeled data and the real-time data distribution, ultimately affecting the prediction accuracy. In order to discern the changes in data modes accurately during operations and achieve the explicit division of the modes, this paper proposes a process data mode division and identification method based on model generalization and least-squares density-difference. Firstly, a multidimensional process variable prediction model is established; and forms the deviation information matrix based on the deviation of the prediction results from the real process data to reflect the generalization ability. Then the deviation information matrix is cut using sliding windows, and the difference in the distribution of the deviation information matrix between windows is measured by the least-squares probability density difference estimation, where the mode is considered to have changed when the difference is greater than a given threshold. The numerical simulation experiments and Tennessee Eastman simulation process experiments proved the feasibility and effectiveness of the proposed method.