Elsevier

Expert Systems with Applications

Volume 40, Issue 12, 15 September 2013, Pages 4760-4769
Expert Systems with Applications

A new approach based on computer vision and non-linear Kalman filtering to monitor the nebulization quality of oil flames

https://doi.org/10.1016/j.eswa.2013.02.008Get rights and content

Abstract

The nebulization quality of oil flames, an important characteristic exhibited by combustion processes of petroleum refinery furnaces, is mostly affected by variations on the values of the vapor flow rate (VFR). Expressive visual changes in the flame patterns and decay of the combustion efficiency are observed when the process is tuned by diminishing the VFR. Such behavior is supported by experimental evidence showing that too low values of VFR and solid particulate material rate increase are strongly correlated. Given the economical importance of keeping this parameter under control, a laboratorial vertical furnace was devised with the purpose of carrying out experiments to prototype a computer vision system capable of estimating VFR values through the examination of test characteristic vectors based on geometric properties of the grey level histogram of instantaneous flame images. Firstly, a training set composed of feature vectors from all the images collected during experiments with a priori known VFR values are properly organized and an algorithm is applied to this data in order to generate a fuzzy measurement vector whose components represent membership degrees to the ‘high nebulization quality’ fuzzy set. Fuzzy classification vectors from images with unknown a priori VFR values are, then, assumed to be state-vectors in a random-walk model, and a non-linear Tikhonov regularized Kalman filter is applied to estimate the state and the corresponding nebulization quality. The successful validation of the output data, even based on small training data sets, indicates that the proposed approach could be applied to synthesize a real-time algorithm for evaluating the nebulization quality of combustion processes in petroleum refinery furnaces that use oil flames as the heating source.

Highlights

► This paper presents a computer vision and Kalman filtering approach to estimate nebulization quality of oil flames. ► Grabbed images from CCD cameras are used to build a set of fuzzy classification rules. ► A state vector composed of grey-level histogram geometrical features of grabbed images is modeled as a random-walk. ► A non-linear Kalman filter estimates the state in order to infer flame nebulization quality. ► Statistically-proven accuracy of the estimates is achieved.

Introduction

Automatic diagnosis of technical defects in industrial plants has been a long date desire of the maintenance engineering community. Such an aim, however, can not be achieved unless the knowledge to detect an anomalous behaviour and to infer its most likely causes be properly represented and coded in order to emulate the ability of the expert responsible for the maintenance task. Various architectures of expert systems to deal with that class of problems have been proposed so far, encompassing rule-based (Qian, Li, Jiang, & Wen, 2003), Bayesian networks (Heo, Changa, Choib, Choic, & Jee, 2005), fuzzy rule-based (Azadeh, Ebrahimipour, & Bavar, 2010), neural networks (Calisto, Martins, & Afgan, 2008) and neural-fuzzy networks (Sahu, Padhee, & Mahapatra, 2011).

Automatically identifying faults in petroleum plants, however, imposes to the expert system developer some very strict demands. The project of a control system capable of optimizing the energetic efficiency of petroleum refinery furnaces in order to reduce the emission rates of pollutants such as CO, NOx and particulate material, requires the setting up of a large network of heterogeneous sensors (thermocouples, flow meters, air–fuel ratio gauges, opacity meters, pressure sensors etc.) dedicated to measure the main variables of the process and to give feedback to the controller. In the last two decades, however, video CCD (Charged Coupled Device) cameras and frame grabbers have been incorporated to this measurement apparatus, since image sequences of flames captured by a near infra-red sensitive CCD and properly analyzed by suitable computer vision methods may provide a large quantity of useful information to the controller.

Correlations between the brightness, spectral and geometric properties of flame images and the corresponding variables of the combustion process have been reported by several authors, who developed different methods to build characteristic vectors and use them to estimate a subset of the state variables that characterize a combustion point of operation. Among those methods, it is worthwhile citing neural networks (Santos-Victor, Costeira, Tomé, & Sencieiro, 1993), fuzzy rules based on triangular membership functions (Tuntrakoon & Kuntanapreda, 2003), hot spots identification through the application of thresholding and logical operators to a collection of sequentially grabbed images of flames (Bertucco, Fichera, Nunnari, & Pagano, 2000), spectral analysis of the hottest zones of flames obtained from images selected through a segmentation process (Baldini, Campadelli, & Lanzarotti, 2000), investigation of correlations between flame image measured parameters (area and centroid of the luminous region, ignition point position and spread angle) and physical data (particle size of the pulverized coals tested and mass flow rate of the primary air) and used them to build a characteristic vector of the combustion process (Yan et al., 2002, Yan et al., 2004, and Lu, Gilabert, & Yan, 2005), and self-organizing feature maps associated to cepstral analysis (Hernández & Ballester, 2008).

Particularly, a combustion diagnosis system developed by Wójcik and Kotyra (2009), based on the analysis of image flames captured at a frequency of 25 frames/s by an ordinary monochromatic video camera equipped with a fiberscope, uses a computer vision algorithm that generates, for each image of the temporal series, a characteristic vector whose components are shape parameters of the image flames, like area, perimeter, centroid coordinates and Fourier descriptors of their contours. Analyzing the temporal values of the shape parameters associated to oil flames exhibiting previously air–fuel ratios and different instability characteristics, the authors stress that there is a clear correlation among the parameters calculated and the phenomena examined, but do not present a method to automatically discriminate the flames according to the physical properties concerning the focused events. Similarly, using a set of four spatial luminous parameters (mean value, standard deviation, kurtosis and skewness of the spatio-temporal brightness image distribution) and one temporal spectral parameter (flicker frequency) extracted from flame images grabbed by a high-speed CCD video camera, the researchers in (González-Cencerrado, Peña, & Gil, 2012) investigated the relative influence of the above-mentioned parameters on combustion processes with different a priori known air-to-fuel ratios and combustion camera mean temperatures. Aided by multivariate regression methods, important correlations between the temporal evolution of both the image features and the respective combustion process variables could be identified; however, as admitted by the authors, an image feature based method to automatically characterize the combustion process should require a more thorough investigation.

In a research project developed at São Paulo State Institute of Technology (IPT) (Fleury, 2006), the authors of the present article proposed various computer vision algorithms to extract features of instantaneous and average flame images, in order to generate crisp decision rules that could be used to diagnose several kinds of abnormalities of the combustion process, encompassing: flame extinction, lack of symmetry, instability, high or low excess air and low nebulization quality. Despite the good agreement between the decisions issued by the application of those rules to image test sets and the known a priori physical conditions concerning the capture of such images, three drawbacks of this diagnostic system must be pointed out: firstly, it required the calculation of average images and the application of heterogeneous computer vision methods to generate the parameters used by the majority of the diagnostic decision rules, what imposed a limitation to the computational performance; secondly, only two states – either strict normality or abnormality of the process – could be diagnosed, although the decisions that can be made by a human expert on the combustion process are not so strict; finally, history of measurements were completely ignored, for the diagnostics were proposed on the basis of present measured values only.

In a further work (Fleury, Trigo, & Martins, 2010), the same authors focused the particular problem of identifying flames exhibiting nebulization patterns of low quality. As the experimental evidence indicate, such flames emerge from combustion mixtures with low values of vapor flow rate (i.e., the quotient between the nebulization vapor flow and the fuel oil flow). After capturing images of flames from combustion processes with varied levels of VFRs, a dedicated diagnostic system was developed to classify them according to the nebulization patterns observed. Compared with the results reported in Fleury (2006), the performance of this new diagnostic system was much improved, due to the following changes: (i) feature vectors are based only on few properties of instantaneous images, which permit to apply simpler computer vision algorithms; (ii) fuzzy linguistic variables are used in the classification process, making the diagnostics more realistic; (iii) predictions are obtained through a stochastic minimum variance least squares estimator, giving rise to more reliable classifications.

The present research intends to validate the nebulization quality classification method described in Fleury et al. (2010) by applying it to sequences of images of flames whose values of VFRs are not known a priori, thus asserting the approach as an effective tool for combustion diagnostic. A Tikhonov-regularized version of the Kalman filter is used to estimate the state, a vector of properties from grabbed images. It must be emphasized that Sections 2 Experimental apparatus and data collection, 3 Image flame analysis of this paper, which describe the currently used experimental set-up and the procedure for obtaining the image feature vectors, contain extended versions of the material previously presented by Fleury et al. (2010).

Section snippets

Experimental apparatus and data collection

As illustrated in Fig. 1, the furnace used in the experiments at IPT is a vertical one, with the burner settled at the bottom and the gases exhaustion at the top. Having a total height of 4.0 m, it is subdivided in 12 independent water cooled blocks and can process number 1 fuel oil (number 1 fuel oils are distilled oils, i.e., they have low viscosity and are free of sediments and inorganic ash) at a maximum flow rate of 80 kg/h. The burner has two (primary and secondary) air entrances for

Image flame analysis

Although combustion process characterization could be made on the basis of a large number of image feature properties, encompassing geometry, luminance and spectral aspects of the flame image, it was established that, to attend real-time performance requirements, only the simplest and fastest algorithms should be applied. Considering that the shape of image grey-level histograms changed for flames with different vapor flow rates (Figs. 4(a) and (b), 5(a) and (b)), ten geometric properties of

Estimation problem

The literature (Balbi, Santoni, & Dupuy, 1999, and Mandel et al., 2008) reports some attempts to model the dynamics of flame propagation through discretization of reaction–diffusion partial differential equations in one or two dimensions by finite differences and to estimate the state, the temperature distribution and the remaining amount of fuel, using the Kalman filter. Mandel et al. (2008), for example, generate synthetic ensembles for the Kalman filter from the numerical solution of the

Nebulization quality estimation

It is common, for simplicity, when there is a lack of knowledge on uncertainty of process and observation models, to admit both noise covariance matrices diagonal. In this work, however, noise covariance matrices shall not be assumed diagonal, once it is possible to compute them from the available data.

The process covariance matrix Q used in Eq. (10) is obtained in the following manner: According to the dynamical model from Eqs. (1), (2), state and measurement are the same for each training set

Results and discussion

Since the benchmark for the validation of the method proposed here is Fleury et al. (2010), it is convenient to recall the main results there presented, as summarized on Table 2 in which the fourth column is the point E[x] in the solution space to be sought by the Tikhonov-regularized Kalman filter algorithm (forced solution).

Actual convergence of the estimation process was asserted by the inspection of the observation residuals (Fleury, 1985, and Jazwinski, 2007), the difference between the

Conclusion

In this work, a method for the determination of nebulization quality for oil flames in industrial processes was improved and validated. The approach consisted in using images grabbed from CCD cameras to build a fuzzy classification set of rules, and in employing the resulting characteristic vector in a state-space model of flame dynamics.

A non-linear Tikhonov regularized Kalman filter was able to provide estimates of the state from data of training sets whose nebulization pattern quality was

Acknowledgements

The authors wish to thank the Conselho Nacional de Pesquisa e Desenvolvimento Tecnológico – CNPq (Brazilian National Council of Technological Research and Development), who supported this work as part of the projects 484260/2011-1, ‘Caracterização do início do processo de instabilização de chamas em fornos de refino’, and 471115/2004-5, ‘Sistema de monitoramento em tempo real de chamas em forno de refino’. First author acknowledges also CNPq for funding part of the work.

References (25)

  • Baldini, G., Campadelli, P., & Lanzarotti, R. (2000). Combustion analysis by image processing of premixed flames. In...
  • Bertucco, L., Fichera, A., Nunnari, G., & Pagano, A. (2000). A cellular neural networks approach to flame image...
  • Cited by (0)

    View full text