Kauko Leiviskä

The advantages by the use of neural networks in modelling the fluidized bed granulation process

Abstract The use of artificial neural networks (ANNs) in modelling a fluidized bed granulation process is reported. The granules were made in a fully instrumented laboratory-scale granulator (Glatt WSG 5, Glatt GmbH Germany). The independent input variables were inlet air temperature atomizing air pressure and binder solution amount. The input variables varied in three levels. The responses used were mean granule size and granule friability. Neural computing was carried out using a commercial NeuDesk software (Neural Computer Sciences U.K.) in a 486 microcomputer with a specific accelerator card NeuSprint (Neural Computer Sciences U.K.). In total, 36 different ANN models were tested. The results were also compared with a statistical method (multilinear stepwise regression analysis). The results showed clearly that the best networks were able to predict the experimental responses more accurately than the multilinear stepwise regression analysis. On the other hand it also became evident that several different structures should be trained with different training end points to generate a proper model.

Knowledge acquisition for decision support systems on an electronic assembly line

Increasing global competition has made many manufacturing companies recognize that competitive manufacturing in terms of low cost and high quality is crucial for success. Real-time process control and production optimization are, however, extremely challenging areas because manufacturing processes are getting ever more complex and involve many different parameters. This is a major problem when building decision support systems especially in electronics manufacturing. Although problem-solving is a knowledge intensive activity undertaken by people on the production floor, it is quite common to have large databases and run blindly feature extraction and data mining methods. Performance of these methods could, however, be drastically increased when combined with knowledge or expertise of the process. This paper describes how defect-related knowledge on an electronic assembly line can be integrated in the decision making process at an operational and organizational level. It focuses in particular on the efficient acquisition of shallow knowledge concerning everyday human interventions on the production lines, as well as on the factory-wide sharing of the resulting information for an improved defect management. Software with dedicated interfaces has been developed using a knowledge representation that supports portability and flexibility of the system. Semi-automatic knowledge acquisition from the production floor and generation of comprehensive reports for the quality department resulted in an improvement of the usability, usage, and usefulness of the decision support system.

Large-Scale Complex Systems

Large-scale complex systems (LSS) have traditionally been characterized by large numbers of variables, structure of interconnected subsystems, and other features that complicate the control models such as nonlinearities, time delays, and uncertainties. The decomposition of LSS into smaller, more manageable subsystems allowed for implementing effective decentralization and coordination mechanisms. The last decade revealed new characteristic features of LSS such as the networked structure, enhanced geographical distribution and increased cooperation of subsystems, evolutionary development, and higher risk sensitivity. This chapter aims to present a balanced review of several traditional well-established methods and new approaches together with typical applications. First the hierarchical systems approach is described and the transition from coordinated control to collaborative schemes is highlighted. Three subclasses of methods that are widely utilized in LSS – decentralized control, simulation-based, and artificial-intelligence-based schemes – are then reviewed. Several basic aspects of decision support systems (DSS) that are meant to enable effective cooperation between man and machine and among the humans in charge with LSS management and control are briefly exposed. The chapter concludes by presenting several technology trends in LSS.

Advanced Control of a Rotary Dryer

Abstract Two kinds of intelligent, hybrid control systems for a rotary dryer are presented. The main controlled variable is the output moisture of solids and the main manipulated variable is the input temperature of drying air which correlates to the fuel flow. The main disturbances of the process are the input moisture of solids and the feed flow. The one discussed control system includes a fuzzy logic controller (FLC) and a PI-controller and the other a neural network controller and a PI-controller. In both cases the intelligent controller determines the set point value to a PI controller. The control results have been examined both with simulations and with pilot plant experiments.

Influence of granulation and compression process variables on flow rate of granules and on tablet properties, with special reference to weight variation

Abstract The influence of three independent fluidized bed granulation process variables (inlet air temperature, atomizing air pressure and binder solution amount) on the flow rate of granules and on the tablet properties was studied using the 3 3 factorial design. The dependence of the flow rate of granules, the tablet friability and the disintegration time of tablets on the fluidized bed granulation process variables was explained using the multilinear stepwise regression analysis. The flow time of the granules was affected by all three factors. The friability and the disintegration time of tablets were affected by the binder solution amount and the atomizing air pressure. The influence of inlet air temperature on the tablet responses was insignificant.

Continuous direct tablet compression: effects of impeller rotation rate, total feed rate and drug content on the tablet properties and drug release

Context: Continuous processing is becoming popular in the pharmaceutical industry for its cost and quality advantages. Objective: This study evaluated the mechanical properties, uniformity of dosage units and drug release from the tablets prepared by continuous direct compression process. Materials and methods: The tablet formulations consisted of acetaminophen (3–30% (w/w)) pre-blended with 0.25% (w/w) colloidal silicon dioxide, microcrystalline cellulose (69–96% (w/w)) and magnesium stearate (1% (w/w)). The continuous tableting line consisted of three loss-in-weight feeders and a convective continuous mixer and a rotary tablet press. The process continued for 8 min and steady state was reached within 5 min. The effects of acetaminophen content, impeller rotation rate (39–254 rpm) and total feed rate (15 and 20 kg/h) on tablet properties were examined. Results and discussion: All the tablets complied with the friability requirements of European Pharmacopoeia and rapidly released acetaminophen. However, the relative standard deviation of acetaminophen content (10% (w/w)) increased with an increase in impeller rotation rate at a constant total feed rate (20 kg/h). A compression force of 12 kN tended to result in greater tablet hardness and subsequently a slower initial acetaminophen release from tablets when compared with those made with the compression force of about 8 kN. Conclusions: In conclusion, tablets could be successfully prepared by a continuous direct compression process and process conditions affected to some extent tablet properties.

Effect of neural network topology and training end point in modelling the fluidized bed granulation process

Abstract The effect of the topology and the training end point of artificial neural networks (ANN) in the modelling of a fluidized bed granulation process is presented. The neural network topologies were designed on the basis of an earlier study (Murtoniemi et al., Int. J. Pharm. , 108 (1994) 155–163). In the first part of this study, the networks contained only one hidden layer in which the number of neurons was either 10, 15, 20 or 25. The training end points with all four networks ranged from 0.15 to 0.07, with a step length of 0.01. In the second part, the training end point was fixed to be 0.12, while the number of neurons in the hidden layer varied from 10 to 25. The main purpose of this study was to find a suitable ANN in regard to the generalization ability and to compare the results to those calculated on the basis of multilinear stepwise regression analysis. The results showed that the number of hidden layer neurons did not affect the generalization ability of the networks and a proper generalization ability was achieved with rather simple networks. The training end point, however, had a significant effect on the generalization ability and it also affects the number of iteration epochs needed. In complicated systems this probably will affect remarkably the time required for the training.

Calibration and validation of a modified ASM1 using long‐term simulation of a full‐scale pulp mill wastewater treatment plant

A mathematical model modified from the well established Activated Sludge Model no. 1 was used for modelling a full‐scale wastewater treatment plant (WWTP) in a bleached kraft pulp mill. Effluents from the pulp and paper industry are typically nutrient deficient, which was considered in the model. The wastewater characterization and model calibration were based on respirometric batch experiments with sludge and wastewater sampled from the WWTP. The model performance was validated in a long‐term simulation using routinely measured process data from the WWTP as the model inputs. The simulation results proved useful in evaluating nutrient dosage strategies at the WWTP and in troubleshooting poor treatment plant performance. However, in order to achieve a completely accurate description of nitrogen removal, more complex phenomena would have to be included in the model. Even though the simulated period was long compared to the brief measurement campaign used in the model calibration, the model was able to describe the treatment plant’s behaviour. The calibrated model can be expected to stay valid for a long time, which allows the use of deterministic modelling in practical applications at pulp and paper WWTPs.

Real-coded genetic algorithms and nonlinear parameter identification

In this study, real-coded genetic algorithms are used in the parameter identification of the macroscopic Chemostat model. The Chemostat model utilized in this work is nonlinear having two distinct operating areas. Thus, the model is identified separately for both operating areas. The process simulator is used to generate data for the parameter identification. The optimizations with genetic algorithms are repeated with 200 different initial populations to guarantee the validity of the results. The parameter identification with genetic algorithms performs well giving accurate results.

Sensor Validation And Outlier Detection Using Fuzzy Limits

In a continuous industrial process, the accuracy and reliability of process and analytical measurements create the basis for control system performance and ultimately for product uniformity. Validation of measured values is the key and a prerequisite to guarantee reliable measurements for process control. This application introduces the use of standard deviation and density function-based absolute limits. Limits are used to cut off outliers and weigh the reliability of the on-line measurement against more reliable, but seldom made, laboratory analysis. Absolute limits are accomplished with constant or adaptively updating fuzzy limits. The adaptive fuzzy limits are recursively updated in real time when a new measured value and reference analysis become available.