Dynamic Soft Sensor Based On Two-dimensional PLPP For Raw Material Decomposition Rate In Cement Process

Dry-process precalciner kiln technology, which moves the preheating and decomposition of material to the suspension preheater and calciner respectively, has become the mainstream of the cement industry. Speed of heat transfer and decomposition in the suspended state is far higher than that in the accumulation state, so the precalciner kiln has fundamentally changed the heat transfer way of the raw material decomposition process and significantly increased thermal efficiency and production efficiency. The decomposition rate of raw material is an important index to measure the quality of products. The key to control calciner and the whole kiln system is ensuring the decomposition rate held steady in the range of the process requirements. Therefore, the accuracy and real-time performance of decomposition rate feedback are crucial.Process variables of raw material pre-decomposition system can be measured via the online instrumentation with fast sampling rate, while the decomposition rate cannot be measured online through the existing detection methods. Current offline detection method is through artificial sampling and chemical analysis. Offline data sampled with slow rate cannot reflect the real-time process condition, which problem could be solved by soft sensor. The precalciner system is complicated and difficult to establish precise mechanism model, while data driven modeling method has been widely studied and applied. Most of the existing data driven soft sensors are static model, whose performance is poor due to the excessive resampling interval that lead to the loss of dynamic information. Multi-rate system identification-based dynamic soft sensor need to acquire informative process dynamic data under well-designed nonlinearity test experiments. Moreover, the implementation of the optimization algorithm for parameter/state estimation would be expensive, and the solution may converge to a local optimum. Therefore, it’s suitable to establish inferential modeling-based dynamic soft sensor driven by multi-rate input and output data, which is beneficial to artificial decision guidance, process security monitoring, automatic inference control and real-time operating optimization.Motivated by above observation, in this paper, we boil down the dynamic soft sensor of raw material decomposition rate to a regression problem of multivariate time series. The input variables of the regression model are composed of temporospatial data, which can be expressed in 2D matrix, is nonlinear and high-dimensional. Combining with the related knowledge of data mining and machine learning, we proposed the support vector regression model based on feature extraction of dynamic data and realized the soft sensor of raw material decomposition rate. The main work is as follows:(1) Multivariate time series data sampled from complex industrial process, such as precalciner system, is nonlinear, high-dimensional and redundancy. Locality preserving projection (LPP), a manifold learning method, is usually used to extract the main features of such data. However, the dynamic data is functional data and of correlation structure in both space and time. Penalized locality preserving projection (PLPP), which proposed by introducing the roughness penalty into LPP, could keep the smooth of functional data. In the case of high-dimensional, strongly correlative input space or inadequate training samples, dimensionality reduction methods based on ID vector, such as LPP and PLPP, usually result in ill-conditioned matrix which leads to the failure of the eigenvalue decomposition. On the other hand, the matrix-to-vector transform may also lead to the loss of useful spatial-temporal structure information embedding in the original multivariate time series data. In order to solve the problem, we introduced the roughness penalty into two-dimensional locality preserving projection (2DLPP) and proposed two-dimensional penalized locality preserving projection (2DPLPP), which could capture 2D correlation structure and the smoothness in temporal dimension simultaneously. Ulterior, we proposed two-directional 2DPLPP ((2D)2PLPP) to reduce the dimensions of temporal and spatial simultaneously and obtain the more effective low-dimensional expression. The novel feature extraction methods can address the problem of ill-conditioned matrix, reduce the eigenvalue decomposition scale and be suitable for small sample learning problem. In this paper, we proposed a novel soft sensor scheme based on the two-directional 2DPLPP and ε-S VR and then realized the accurate prediction of cement raw material decomposition rate. The experimental results of soft sensors based on different feature extraction methods show the effectiveness of the two-directional 2DPLPP.(2) In manifold learning, such as 2DPLPP, similarity matrix is usually based on Euclidean distance in linear space and heat kernel. The Euclidean distance, which regards all forms data as vector, does not exploit the topological structure of matrix (two order tensor) capture the underlying multimode relations. Chord distance on the Grassmann manifolds addresses the issue. It could take into account the priori knowledge about invariance in the input space, and its robustness is much higher than the Euclidean distance due to the capture of the topological structure. Basing on the chord distance, we proposed tensorial factor similarity matrix and improved 2DPLPP algorithm. However, the similarity matrix based on distance measure is depended on the reliability of data and sensitive to the noise. The temporal similarity matrix was proposed in this paper to improve the robustness of the algorithm. This is because the sampling information is treated as priori knowledge and introduced into unsupervised learning. Finally, a variety of hybrid similarity matrices are also proposed. The experimental results show that, both the tensorial factor similarity matrix and the temporal similarity can improve the accuracy of soft sensor.