Environmentally compatible process control and fault state detection in chemical plants using neural networks

Term: 07 / 1998 – 12 / 2002

Funding: BMBF




In the fine chemical and pharmaceutical industry, complex strongly exothermic reactions are usually carried out in stirred tank reactors (STR). These so-called batch reactors are characterised by nonsteady-state conditions due to the discontinuous operation so that even experienced operators have difficulties to distinguish allowable from undesired process deviations and to identify the cause of process trends. In particular, in general there is no measuring system for chemical on-line analysis in chemical plants to monitor and assess the most important parameters of a complex chemical reaction, the concentration profiles of the reactants, intermediates and products, during the whole batch process in real time. Therefore, the aim of this work was to develop an on-line monitoring system (MoSys) that is able to estimate concentration profiles as well as conversion profiles and the remaining period up to the complete reaction without using expensive on-line analytics.

Model process

To test the monitoring system at industrial scale, hydrogenation of a aromatic nitro compound (R-NO2) to the corresponding aromatic amine (R-NH2) was chosen. The process is carried out in semi-batch mode by dosage controlled feed of hydrogen according the following overall reaction equation:


This hydrogenation is a complex multiphase (heterogeneous) process with consecutive reactions and with a concurrent reaction path according to the Haber’s  reduction scheme. If the concentrations of the intermediates (nitroso and hydroxylamine compound) increase, the slower concurrent reaction path via the azo compound can occur. Due to the accumulation of the intermediates, the hazardous potential of the process is increased because then a strongly exothermic condensation reaction might take place without hydrogen uptake. Moreover, increased intermediate concentrations can lead to problems regarding the space-time yield and the product quality as well. Therefore, the main task of the monitoring system is to supervise accumulations of intermediates during the process.


MoSys is based on a hybrid balance model consisting of coupled heat and mass balances and adaptive model components. The adaptive components consider the difficulty to model heat losses, heat bridges and systematic measuring uncertainties at industrial plants in terms of simple neural networks.



Structure of the on-line monitoring system


However, a process signals set of one batch at normal operating conditions is enough to adapt the monitoring system to the target plant. During the adaptation process, adaptive parameters, introduced as additional coefficients in the balance equations, can be adjusted in off-line operation mode using a neural network approach.

The connection of MoSys to the process control system (PCS) of the chemical plant was the precondition for the on-line ability of the monitoring system. This was done by integration of MoSys into a complex batch information management system (BIMS) consisting of the following components:

  • monitoring system MoSys

  • data management module

  • visualisation and operation system


Data and information flow within the BIMS


The integration of MoSys into BIMS provides the operators an opportunity to access concentration profiles, conversions and reaction times of current or previous batches together with further process signals at the PLS terminals.

Mosys_Umsatzanzeige Mosys_Trendanzeige
On-line display of the conversion rates at the operator stations (screenshot) Trend display of the MoSys results at the operator stations (screenshot)


The on-line ability and the performance of MoSys and BIMS were tested in a multi-purpose stirred tank reactor, located in a chemical plant of the DEGUSSA AG, during several hydrogenation campaigns. The monitoring system was validated by comparison of the MoSys results with accompanying HPLC analyses for the following results:

  • educt concentration,

  • intermediate concentration,

  • product concentration and

  • remaining period up to the complete hydrogenation.


MoSys concentration profiles compared to HPLC analyses during two hydrogenation batches (strongly increased intermediate concentration in batch 2)


Example of a process state classification – MoSys vs. HPLC


Profiles of chemical conversion and remaining period up to the complete hydrogenation

By using MoSys as an integral part of BIMS, the operator of an industrial plant can benefit from the additional process information in the following manner:

  • Adaptive energy & mass balance approach is suitable to estimate concentration courses of educts, products and intermediates without any expensive on-line chemical process analysis.

  • Accuracy of the estimations is sufficient to identify undesired process states at an early stage.

Absolute measuring inaccuracies (RMSE-values): Educt 2.5 Mol%
  Product 3.1 Mol%
  Intermediate 4.6 Mol%
  • The data generated by the monitoring system can be stored and archived batch-wise together with the process signals, laboratory findings as well as plant and substance parameters.

  • By using MoSys, the demand for traceability (e.g. hazardous operating state, batch with poor product quality) of complex batch processes can be fulfilled.

  • Archived data allow to optimise chemical batch processes (e.g. yield, quality) and to repeat them with high product quality.

Further information

Monitoring-System for batch reactors using adaptive heat balances (Poster)

On-line concentration estimation during chemical reactions using adaptive heat/ mass balances  (Presentation)




  • Hessel, G.; Kryk, H.; Schmitt, W.; Seiler, T.; Weiß, F.-P.; Hilpert, R.; Roth, M., Modellbasiertes Online-Zustandserkennungssystem für exotherme chemische Prozesse, Technische Überwachung Bd. 44, Nr. 5 (2003) 43-47

  • Hessel, G.; Kryk, H.; Schmitt, W.; Seiler, T.; Hilpert, R.; Roth, M.; Deerberg, G., Monitoring-System mit adaptiven Wärmebilanzen für Batch-Reaktoren, Chemie Ingenieur Technik 74, 12 (2002) S. 1692-1698

  • G. Hessel, J. Heidrich, R. Hilpert, H. Kryk, M. Roth, W. Schmitt, T. Seiler, F. P. Weiss (2002), Umweltgerechte Prozessführung und Zustandserkennung in Chemieanlagen mit neuronalen Netzen – Teilvorhaben 2: Konzipierung und Erprobung des Zustandserkennungsverfahrens, Wissenschaftlich-Technische Berichte Forschungszentrum Rossendorf, Report FZR-355 (2002) S. 1-137


Degussa AG, Fraunhofer UMSICHT


Dr. H. Kryk