Forecast Modeling Raw Material Decomposition Rate Of Cement Decomposing Furnace Based On Dynamic Data Dimension Reduction And Semi-supervised Learning

The new type dry-process, which can realize large-scale production, has become the mainstream technology of the cement industry. Pre-decomposition and suspension preheating are regarded as its core technologies. Decomposing furnace is the main equipment of pre-decomposition process. Fuel combustion and raw material decomposition under suspension situation are finished in it, which reduces the heat load of burning zone of kiln, and consequently increases the efficiency of kiln. Decomposition rate of raw material is an important quality indicator in cement clinker production process. It is very significant to stabilize raw material decomposition rate for stability control of rotary kiln and ultimate improvement of product quality.Process variables of cement decomposing furnace can be measured via the online instrumentation under fast sampling rate, while the decomposition rate of raw material cannot be measured online through the existing measurement methods. Current offline detection method is sampling and analysis in manual, which cycle is long and cannot reflect the real production status. Cement decomposing furnace is a typical multiple capacity, high order and large inertia process. It is difficult to establish a precise mechanism model because of the complicated process mechanism. The existing data-driven raw material decomposition rate forecast models are static models based mainly on the slow sampling rate input and output data. The large sampling interval results in the loss of process dynamic information and poor model performance. Therefore, it is necessary to further research ways to forecast model raw material decomposition rate, which is beneficial to achieve artificial decision guidance of decomposing furnace, process security monitoring, automatic inference control, and real-time operating optimization, and which is essential to reduce energy consumption and improve product quality and yield.For above problem of modeling decomposing furnace multirate system, this paper is supported by the national natural science fund general project "product quality parameter prediction modeling based on the fusion of image and process data about kiln". We establish forecast model of raw material decomposition rate based on combination of data and knowledge, considering the dynamic process information. The main work of this paper is as follows:(1) At the present stage, the input feature of raw material decomposition rate forecasting model, which form the static feature set with certain spatial correlation, is selected via all process variables which affect the raw material decomposition rate. In this paper, we consider the dynamic information, which makes each process variables form a time series dynamic feature set with time correlation. For the characteristic of both time relevance and space relevance that the time series dynamic features formed by process variables, this paper proposes a combined dimension reduction method of feature extraction and feature selection, in which, we can reduce the redundancy of each variable’s time series dynamic features through feature extraction and remove redundant feature between space-related features through feature selection.1) For the characteristic of time series dynamic feature, considering the data has potential manifold structure, this paper proposes a feature extraction method-Penalized Preserving Projections (PLPP), based on manifold dimension reduction algorithm Preserving Projections (LPP). For time series dynamic features of each variable, we do feature extraction using PLPP.2) For the spatial correlation between process variables extracted by PLPP, we take a two-stage method to selection features, combining Filter with Wrapper. The minimal redundancy maximal relevance (mRMR) criterion based on mutual information is used to get the candidate feature set in Filter part, and after this we combine sequential forward selection (SFS) and SVR to get a compact feature set. The methods proposed in this paper are used to establish a raw material decomposition rate model with the dynamic information in the actual decomposing furnace process. The experimental results demonstrate the effectiveness of the dimension reduction method.(2) The accuracy of supervised data-driven static model based on slower sampling rate input and output data is low. This paper first introduces semi-supervised learning mechanism of machine learning field to the issue of raw material decomposition rate forecast modeling, considering process dynamic information. So we can fully mine the information of the unlabeled samples related to regression problems and improve the model generalization performance. This idea provides more effective method of data-driven product quality parameter prediction modeling for complex industrial processes. After finishing feature extraction and selection for dynamic input data with the proposed dimensionality reduction method, around the idea of semi-supervised learing, we make use of labeled and unlabeled samples to train regression models, and finally establish forecast model of cement raw material decomposition rate based on semi-supervised ε-LapSVR algorithm, taking into account the compromise between forecast error, the data’s intrinsic geometry structure and complexity of the decision function. The simulation results show that the forecast model of raw material decomposition rate based on semi-supervised ε-LapSVR has better generalized performance.