Francisco Souza

Adaptive fuzzy identification and predictive control for industrial processes

This paper proposes a method for adaptive identification and control for industrial applications. The learning of a T-S fuzzy model is performed from input/output data to approximate unknown nonlinear processes by a hierarchical genetic algorithm (HGA). The HGA approach is composed by five hierarchical levels where the following parameters of the T-S fuzzy system are learned: input variables and their respective time delays, antecedent fuzzy sets, consequent parameters, and fuzzy rules. In order to reduce the computational cost and increase the algorithms performance an initialization method is applied on HGA. To deal with nonlinear plants and time-varying processes, the T-S fuzzy model is adapted online to maintain the quality of the identification/control. The identification methodology is proposed for two application problems: (1) the design of data-driven soft sensors, and (2) the learning of a model for the Generalized predictive control (GPC) algorithm. The integration of the proposed adaptive identification method with the GPC results in an effective adaptive predictive fuzzy control methodology. To validate and demonstrate the performance and effectiveness of the proposed methodologies, they are applied on identification of a model for the estimation of the flour concentration in the effluent of a real-world wastewater treatment system; and on control of a simulated continuous stirred tank reactor (CSTR) and on a real experimental setup composed of two coupled DC motors. The results are presented, showing that the developed evolving T-S fuzzy model can identify the nonlinear systems satisfactorily and it can be used successfully as a prediction model of the process for the GPC controller.

Genetic fuzzy system for data-driven soft sensors design

This paper proposes a new method for soft sensors (SS) design for industrial applications based on a Takagi-Sugeno (T-S) fuzzy model. The learning of the T-S model is performed from input/output data to approximate unknown nonlinear processes by a coevolationary genetic algorithm (GA). The proposed method is an automatic tool for SS design since it does not require any prior knowledge concerning the structure (e.g. the number of rules) and the database (e.g. antecedent fuzzy sets) of the T-S fuzzy model, and concerning the selection of the adequate input variables and their respective time delays for the prediction setting. The GA approach is composed by five hierarchical levels and has the global goal of maximizing the prediction accuracy. The first level consists in the selection of the set of input variables and respective delays for the T-S fuzzy model. The second level considers the encoding of the membership functions. The individual rules are defined at the third level, the population of the set of rules is treated in fourth level, and a population of fuzzy systems is handled at the fifth level. To validate and demonstrate the performance and effectiveness of the proposed algorithm, it is applied on two prediction problems. The first is the Box-Jenkins benchmark problem, and the second is the estimation of the flour concentration in the effluent of a real-world wastewater treatment system. Simulation results are presented showing that the developed evolving T-S fuzzy model can identify the nonlinear systems satisfactorily with appropriate input variables and delay selection and a reasonable number of rules. The proposed methodology is able to design all the parts of the T-S fuzzy prediction model. Moreover, presented comparison results indicate that the proposed method outperforms other previously proposed methods for the design of prediction models, including methods previously proposed for the design of T-S models.

Learning of a single-hidden layer feedforward neural network using an optimized extreme learning machine

This paper proposes a learning framework for single-hidden layer feedforward neural networks (SLFN) called optimized extreme learning machine (O-ELM). In O-ELM, the structure and the parameters of the SLFN are determined using an optimization method. The output weights, like in the batch ELM, are obtained by a least squares algorithm, but using Tikhonovs regularization in order to improve the SLFN performance in the presence of noisy data. The optimization method is used to the set of input variables, the hidden-layer configuration and bias, the input weights and Tikhonovs regularization factor. The proposed framework has been tested with three optimization methods (genetic algorithms, simulated annealing, and differential evolution) over 16 benchmark problems available in public repositories.

Online Mixture of Univariate Linear Regression Models for Adaptive Soft Sensors

This paper proposes a mixture of univariate linear regression models (MULRM) to be applied in time-varying scenarios, and its application to soft sensor problems. Offline and online solutions of MULRM will be obtained using the Expectation-Maximization Algorithm. A forgetting factor will be introduced in the online solution to discount the information of already learned data, so that it can be applied in time varying settings. The solution of the proposed method allows its online and recursive application in any regression problem, without the necessity of storing any past value of data. The recursive solution of the MULRM will then be applied in two time-varying real-world prediction problems. The proposed method is compared with four state of art algorithms. In all the experiments, the proposed method always exhibits the best prediction performance.

Variable and delay selection using neural networks and mutual information for data-driven soft sensors

This paper proposes a new method for input variable and delay selection (IVDS) for Soft Sensors (SS) design. The IVDS algorithm is composed by the following steps: (1) Time delay selection; (2) Identification and exclusion of redundant variables; (3) Best variables subset selection. The IVDS algorithm proposed in this work performs the delay and variable selection through two distinct methods, mutual information (MI) is applied to delay selection and for variable selection a multilayer perceptron (MLP) based approach is performed. It is shown in the case studies that the application of the delay selection before applying the variable selection increases the generalization of the MLP-model. The algorithm uses the relative variance tracking precision (RV TP) criterion and the mean square error (MSE) to evaluate the precision of soft sensor. Simulation results are presented showing the effectiveness of the method.

Fault detection and replacement of a temperature sensor in a cement rotary kiln

This paper proposes a method for fault detection and replacement of the sensor responsible by the measurement of the burning zone temperature in a rotary cement kiln. The control of the burning zone temperature is crucial for the control of kiln temperature and therefore for the control of the cement quality, pollutant emissions, and consumed energy. However the flying dust within the kiln can block the pyrometer sensor, causing faults in the temperature sensor. Exploring the analytical redundancy that usually exist in industrial processes, the proposed methodology uses a neural network trained using an online sequential extreme learning machine to online construct a model to estimate the burning zone temperature. Using the error between the measured and estimated temperatures, faults in the measurement can be detected and thus the replacement of the measured temperatures by the estimated output is made.

Adaptive fuzzy generalized predictive control based on Discrete-Time T-S fuzzy model

The paper presents an adaptive fuzzy predictive control based on discrete-time Takagi-Sugeno (T-S) fuzzy model. The proposed controller is based on Generalized predictive control (GPC) algorithm, and a discrete-time T-S fuzzy model is employed to approximate the unknown nonlinear process. To provide a better accuracy in identification of unknown parameters of the model, it is proposed an on-line adaptive law which ensures that the tracking error remains bounded. The stability of closed-loop control system is proved/studied via the Lyapunov stability theory. To validate the theoretical developments and to demonstrate the performance of the proposed control is simulated as nonlinear system a laboratory-scale liquid-level process. The simulation results show that the proposed method has a good performance and disturbance rejection capacity in industrial process.

Design and application of Soft Sensor using Ensemble Methods

Industries are faced with the choice of suitable process control policies to improve costs, quality and raw material consumption. In the paper pulp industry, it is important to estimate quickly the Chemical Oxygen Demand (COD), a parameter that is highly correlated to product quality. Soft Sensors (SSs) have been established as alternative to hardware sensors and laboratory measurements for monitoring and control purposes. However, in real setups it is often difficult to get sufficient data for SS development. This work proposes Ensemble Methods (EMs) as a way to improve the SS performance for small datasets. EMs use a set of models to obtain better prediction. Their success is usually attributed to the diversity. Bootstrap and noise injection are used to produce diverse models. Several combinations of EMs are compared. The SS is successfully applied to estimate COD in a pulp process.

Neo-fuzzy neuron model for seasonal rainfall forecast: A case study of Ceara's eight homogenous regions

The knowledge about the seasonal rainfall in some Brazilian regions is essential for agriculture and the adequate management of water resources. For this purpose, linear and nonlinear models are commonly used for seasonal rainfall prediction, while some of them are based on Artificial Neural Networks, demonstrating great potential as shown in literature. According to this tendency, this work presents a rainfall seasonal forecast model based on a neuro-fuzzy technique called Neo-Fuzzy Neuron Model. Improved performance by using this approach has been obtained in terms of reduced root mean square error RMSE and increased correlation between predicted and real output when compared with dynamic downscaling model using the Regional Spectral Model. Experimental results show the effectiveness of the proposed method in predictions regarding the first four trimesters from year 2002 up to the current one.

Seasonal rainfall forecast using a Neo-Fuzzy Neuron Model

Knowledge about the seasonal rainfall for some regions of Brazil is essential, due to the dependency of agriculture and for a correct management of water resources. For this, linear and nonlinear models are commonly used for seasonal rainfall prediction, some of them are based on Artificial Neural Networks, which have proved to have a great potential for this purpose. Following this idea, this work presents a seasonal rainfall forecast model based on a neuro-fuzzy technique, called Neo-Fuzzy Neuron Model, that showed a better performance, in terms of root mean square error and correlation coefficient between predicted and real output, when compared with a dynamic downscaling model using the Regional Spectral Model. The experimental results show the effectiveness of the proposed method in predicting the first four trimesters from 2002 up to the current year.