Paweł D. Domański

Non-Gaussian properties of the real industrial control error in SISO loops

The paper presents results of the observations on behavior and properties of the control quality that is met in real industrial practice. The analysis is based on data from several hundreds of control loops operating in different process industries located in several sites all over the world. Practice shows that theoretical assumption about Gaussian properties is hardly met. The author suggest novel approach to the loop analysis and the assessment of process control quality based on the fractal approach. Alternative tools based on the R/S plots, Hurst index and fat-tail probabilistic distributions seems to be valid extension to the existing Gaussian perspective.

Optimization of Coal Mill Using an MPC Type Controller

The article presents the question of optimization of a ventilation coal mill on the basis of a predictive optimizing controller with a receding horizon, which is an extension of the standard linear MPC (Model Predictive Control) type controllers. The controller has been realized in a digital version operating with a certain sampling period dependent upon the process dynamics. All calculations of the control rules are performed in one cycle which enables the controller to operate in the running mode. On the basis of a right optimization procedure the controller regulates the correction of settings, which are introduced to classic control structures in a fuzzy control system. The non-linear process model, implemented in the controller, is based on the basis of fuzzy neural networks. This structure enables to design, learn and tune NARMAX type models (Nonlinear Auto Regressive Moving Average with auXiliary input) [1]. The process model uses fuzzy rules, where fuzzy rules figure on the side, which helps to avoid sharp switching between them. The consequences of the rules take the form of differential equations of the linear ARX type models. The use of neural networks ensures a fast and efficient implementation and effective learning and tuning. The problem of control is based in on a periodically performed optimization of the performance index, defined on the basis of the assumed project goals. The aim of the controller operation is to eliminate undesired events occurring during mill operation. Such events are: instability of temperature value of air-dust mix after the mill, excessive fluctuation of air temperature before the mill and positioning of primary and secondary air dampers outside the control range. These goals are realized through appropriate control of the primary air damper and revolving speed of the mill. The implementation carried out of the described controller in a digital automatic control system on 8 ventilation mills of a 360 MW brown coal fired boiler. This article presents the results obtained and a carried out analysis.Copyright

Fractal measures in control performance assessment

In case of numerous loops control engineers need fast and accurate method finding out loops with the poorest control quality. There exits several measures however, industrial practice reveals their deficiencies. Some of them require introduction of external disturbance into the process (like the step test) that in many cases is considered as a risk or threat. The other group of automatic approaches uses regular process data and is based on the loop Gaussian properties foundation. Analysis of real process data shows that this assumption holds only in a few cases. Both facts initiated research for alternative measures grouped into statistical non-Gaussian indexes and fractal ones based on the properties of R/S plot and Hurst index. The analysis is performed on both simulation and real industrial data revealing interesting properties and proving validity of selected approach.

On-line control loop assessment with non-Gaussian statistical and fractal measures

This paper present results of on-line control loop performance assessment using non-Gaussian statistical and fractal measures. Research shows importance of loop quality indexes that are not biased with Gaussian assumption about signal characteristics. Industrial data show frequent fat-tail properties and thus relevant indexes are proposed, like non-Gaussian statistical factors or persistence fractal measures. There are frequent reports presenting results calculated off-line. The paper extends the analysis to the on-line performance monitoring with non-Gaussian statistical and fractal measures. Results are evaluated through simulations for known scenarios and real industrial variables. Results show detection potential of considered approach.

Gradient Approach to Curve Fitting with Incomplete Data

The paper deals with the problem of nonlinear curve fitting in situation of incomplete data. Research was motivated by the industrial identification of Hammerstein models. It was noticed that for the model robustness and quality the fitness of the static nonlinear element is much more crucial then efficiency of dynamic operation of linear part. Industrial data are incomplete, i.e. they do not cover the whole process domain. It is due to the operation around selected steady states, close loop operation, extensive manual model use, etc. In case of multi-regional approach we often get regions with no data. Proposed methodology is addressing that issue. Included results are for both simulation and real industrial case.

Memory-Based Prediction of District Heating Temperature Using GPGPU

The paper presents application of the memory-based prediction to the problem of the return water temperature prognosis in a district heating network. CHP (Combined Heating Plant) problem is defined as well as the algorithm based on the memory of the historical process realizations together with its novel, parallel implementation using CUDA on GPGPU. The use of the calculation extensive methods from one side enables to get good and reliable predictions, but in opposite the prognosis evaluation is done at high cost. An alternative application of the massively parallel version of the Memory-based time series prediction algorithm has been implemented and tested. The paper shows very good and promising improvement in comparison to the common applications. The algorithm is tested on the real process data.

Statistical measures for proportional–integral–derivative control quality: Simulations and industrial data

This article focuses on investigation of statistical approaches to the task of control performance assessment. Different statistical measures with Gaussian and non-Gaussian probabilistic distributions are taken into consideration. Analysis starts with the observations for simulated proportional–integral–derivative control error histograms followed by its statistical investigation using selected probabilistic distribution functions. Simulation experiments are followed by the analysis of control data originating from real industrial loops. Shadowing effect of long-tail control error histograms is identified, as it may significantly disable proper loop quality assessment. Results show that non-Gaussian approach with Cauchy or α-stable distributions seems to be reasonable assessment alternative in case of disturbances existing in industrial processes.

Estimation of control improvement benefit with α–stable distribution

The paper discusses the subject of estimation of potential financial benefits achievable with the rehabilitation (modification or tuning) of the control system. This issue appears almost in any process improvement initiative giving arguments for control upgrades. The subject exists in literature for several years with well established the same limit approach. The procedure is based on the assumption of Gaussian properties of the considered variable reflected in its histogram. Review of industrial data shows frequent situations when process variables are of different character featuring long tails. Such properties are well described by α–stable distributions. This paper presents extension of the method on such general probability density functions family. The analysis is illustrated with the simulation and industrial data examples.

Reproduction of equipment wear characteristics with kernel regression

Predictive maintenance task is of crucial role for any plant equipment supervision and scheduling of service activities. For this purpose it should be known what is current aging status of any equipment. Presented approach assumes that we know the nominal (starting) element curve and a damage one as well. It is also assumed that the aging course progresses according to some good practice aging Lorentz attrition (wear) curve. Kernel Regression algorithm is used to perform curve adaptation and then enabled to identify current element status and varying aging curve. The approach is tested on SISO and two-dimensional examples proving its ability to reproduce equipment aging characteristics.

Evaluation of High Efficiency Operating Rules for Grate-Fired Boilers

Synthesis of control rules (extraction of knowledge) from historical data may be used for optimal control of the pulverized coal boiler, however it appears to be quite a difficult task. The best effects, allowing fast design and reliable tuning, might be obtained through comprehensive approach with use of all available sources of information. Presented methodology covers one aspect: knowledge extraction through analysis of process operational historical data. Proposed methodology is specifically developed for the task: synthesis of the control rules and conditions for improvement of the combustion process efficiency based on the historical. The approach is stepwise. It starts from the historical data acquisition, followed by the efficiency calculations, data processing. The methodology ends with evaluation of the control rules that may suggest boiler operators optimal operation. This approach copes with several practical problems met during daily boiler operation. Proposed algorithm uses original combination of heuristics, polynomial approximation and data clustering to find patterns in scattered data allowing to assign operational rules to meet the goals. The approach is validated on the real data from the industrial grate boiler.