Heng Yue

Multi-model direct adaptive decoupling control with application to the wind tunnel system.

In this paper, a new multi-model direct adaptive decoupling controller is presented for multivariable processes, which includes multiple fixed optimal controllers, one free-running adaptive controller, and one re-initialized adaptive controller. The fixed controllers provide initial control to the process if its model lies in the corresponding region. For each controller selected, the re-initialized adaptive controller uses the values of this particular controller to improve the adaptation speed. This controller may replace the fixed controller at a later stage according to the switching criterion which is to select the best one among all controllers. A free-running adaptive controller is also added to guarantee the overall system stability. Different from the multiple models adaptive control structure proposed in Narendra, Balakrishnan, and Ciliz [Adaptation and learning using multiple models, switching, and tuning. IEEE Control Syst. Mag. 15, 37-51 (1995)], the method not only is applicable to the multi-input multi-output processes but also identifies the decoupling controller parameters directly, which reduces both the computational burden and the chances of a singular matrix during the process of determining controller parameters. Several examples for a wind tunnel process are given to demonstrate the effectiveness and practicality of the proposed method.

Soft sensor for parameters of mill load based on multi-spectral segments PLS sub-models and on-line adaptive weighted fusion algorithm

The parameters of mill load (ML) not only represent the load of the ball mill, but also determine the grinding production ratio (GPR) of the grinding process. In this paper, a novel soft sensor approach based on multi-spectral segments partial least square (PLS) model and on-line adaptive weighted fusion algorithm is proposed to estimate the ML parameters. At first, frequency spectrums of the shell vibration acceleration signals are obtained. Then the PLS sub-models are constructed with the low, medium and high frequency spectral segments. At last, the PLS sub-models are fused together with a new on-line adaptive weighted fusion algorithm to obtain the final soft sensor models. This soft sensor approach has been successfully applied in a laboratory-scale wet ball mill grinding process.

Soft Sensor Modeling of Ball Mill Load via Principal Component Analysis and Support Vector Machines

In wet ball mill, measurement accuracy of mill load (ML) is very important. It affects production capacity and energy efficiency. A soft sensor method is proposed to estimate the mill load in this paper. Vibration signal of mill shell in time domain is first transformed into power spectral density (PSD) using fast Fourier transform (FFT), such that the relative amplitudes of different frequencies contain mill load information directly. Feature variables at low, medium and high frequency bands are extracted through principal component analysis (PCA), which selects input as a preprocessing procedure to improve the modeling performance. Three support vector machine (SVM) models are built to predict the mill operating parameters. A case study shows that proposed soft sensor method has higher accuracy and better predictive performance than the other normal approaches.

Modeling of operating parameters for wet ball mill by modified GA-KPLS

Load of the ball mill affects the productivity, quality and energy consumption of the grinding process. But sensors are not available for the direct measurement of the key parameters for mill load (ML). A new soft sensor approach based on the shell vibration signals to measure the operating parameters is proposed in this paper. Vibration signal is first transformed into power spectral density (PSD) via fast Fourier transform (FFT), such that the relative amplitudes of different frequencies could contain information about operating parameters. As the spectral curve consists of a set of small peaks, the masses and the central frequencies of the peaks are extracted as the spectral features, then the kernel partial least square (KPLS) is used to built the soft sensor model. The kernel parameters, the input variables of the models including the masses and the central frequencies of the peaks are selected by Genetic algorithm (GA). At last, a new approach for the updating of the KPLS model is proposed. Experimental results show that proposed method has higher accuracy and better predictive performance than the other approaches.

A hybrid intelligent soft-sensor model for dynamic particle size estimation in grinding circuits

The purpose of this paper is to develop an on-line soft-sensor for dynamic estimation of the particle size distribution of hydrocyclones overflow in consecutive grinding process. The hybrid model based soft-sensor is based on the following model structures: 1. a neural net-based dynamic model of state space description for hydrocyclone with a neural net-based model for classifier and a population balance model for ball mill and sump, 2. an ANFIS-based model mainly for abnormal operating conditions, 3. a fuzzy logic coordinator for the final predictive result according to output values of aforementioned models. The fact that the soft-sensor performs well in particle size estimation demonstrates that the proposed hybrid intelligent soft-sensor model is effective for dynamic estimation of particle size distribution.

A method of hybrid intelligent optimal setting control for flotation process

In the flotation process, the concentrate grade and the tailing grade are crucial technical indices which reflect the product quality and efficiency. Such technical indices have close relationship with the reagent feeding, the air flowrate and the pulp level. There are strong nonlinearity and uncertainties in dynamic behavior, which can hardly be described using any accurate mathematical model. The technical indices which can not be measured online continuously vary with the variation of the slurry density, the slurry flowrate, the particle size and the ore grade, etc. Therefore conventional control methods are incapable of keeping the actual indices values within the target ranges. In this paper, a hybrid intelligent optimal setting method comprised of a pre-setting model based CBR and a feedback compensator model based RBR is proposed innovatively. The hybrid intelligent optimal setting control system updates adaptively the setpoints of control loops of the flotation reagent feeding, the air flowrate and flotation level as long as the boundary condition change so as to control the concentrate grade and the tailing grade within their respective target ranges. This method has been successfully applied to a flotation process in China, and significant application effect has been achieved. The successful application in a flotation process indicates that the proposed method has an extensive prospect in the application domain of the optimal control for the technical indices of the complex industrial process.

Supervisory Control for Rotary Kiln Temperature Based on Reinforcement Learning

The burning zone temperature in rotary kiln process is a vitally important controlled variable, on which the sinter quality mainly relies. Boundary conditions such as components of raw material slurry often change during kiln operation, but related offline analysis data delay to reach or even are unknown to the human operator. This causes unsatisfactory performance of the burning zone temperature controller and subsequent unstable production quality. To deal with this problem, a Q-learning-based supervisory control approach for burning zone temperature is proposed, in which the signals of human intervention are regarded as the reinforcement learning signals, so that the set point of burning zone temperature can be duly adjusted to adapt the fluctuations of the boundary conditions. This supervisory control system has been developed in DCS and successfully applied in an alumina rotary kiln. Satisfactory results have shown that the adaptability and performances of the control system have been improved effectively, and remarkable benefit has been obtained.

Multiple Models Neural Network Decoupling Controller for a Nonlinear System

For a discrete-time nonlinear MIMO system, a multiple models neural network decoupling controller is designed in this paper. At each equilibrium point, the system is expanded into a linear and nonlinear term. These two terms are identified using two neural networkss, which compose one system model. Then, all models, which are got at all equilibrium points, compose the multiple models set. At each instant, the best model is chosen as the system model according to the switching index. To design the controller accordingly, the nonlinear term and the interactions of the best model is viewed as measurable disturbance and eliminated by the use of the feedforward strategy. The simulation example shows that the better system response can be got even when the system is changed around these equilibrium points.

Learning to link human objects in videos and advertisements with clothes retrieval

In this paper, we present a new method for human object-level video advertising. A framework that aims to embed content-relevant ads within a video stream is investigated in this context. In particular, to support content-relevant advertising, we employ the discriminatively trained part based model to detect human objects in a video and then select the ads that are related to the detected human objects. For human clothing advertising, we design a deep Convolutional Neural Network (CNN) using face features to recognize human genders in a video stream. Human parts alignment is then implemented to extract human part features that are used for clothes retrieval. Our novel framework is examined in various types of videos. Experimental results demonstrate the effectiveness of the proposed method for human object-level video advertising.

Spectral Kernel Principal Component Selection Based on Empirical Mode Decomposition and Genetic Algorithm for Modeling Parameters of Ball Mill Load

Parameters of ball mill load (ML) affects production capacity and energy consumption of the grinding process, which have stronger correlation with shell vibration spectrum. A novel spectral features extraction and selection approach combined with empirical mode decomposition(EMD), power spectral density(PSD), kernel principal component analysis(KPCA), genetic algorithms(GA) and partial least square(PLS) was proposed in this paper. At first, shell vibration signals were decomposed into a number of intrinsic mode functions (IMFs) based on the EMD. Secondly, the PSD of each IMF was obtained. At last, the mainly spectral KPCs extracted from the PSD were integrated together as the candidate features set. GA was used to optimize spectral KPCs as the selected features subset, which was used to construct ML parameters soft sensor models based on PLS algorithm. The experimental result shows that the proposed approach has higher accuracy and better predictive performance than other normal approaches.