Designing a decision support system for accident prevention at process plants using artificial neural networks

A.V. Chikurov, M.Kh. Khusniyarov, D.S. Matveev,

A.V. Midyutov, I.R. Saifutdinov

Ufa State Petroleum Technological University, Russia

Designing a decision support system for accident prevention at process plants using artificial neural networks

Ensuring process safety at refineries, chemical and petrochemical facilities is an issue of high priority due to the large number of failures and accidents resulting in many occupational injuries, illnesses and financial losses. On one hand the high rate of accidents is caused by physical and moral wearing of equipment and human errors. On the other hand it is influenced by the desire to achieve better performances, to intensify processes, which leads to unpredictable deviations and errors in equipment performance.

In order to achieve an acceptable level of safety it is necessary to use advanced methods for optimal decision making in abnormal situations. These methods, often enhanced by artificial intelligence, have been implemented in the form of integrated computing environments for complex decision making called decision support systems (DSS). DSS can provide good solutions to ensure better safety in the process industries by helping operators to make decisions in abnormal situations.

A common approach to build a DSS for accident prevention is to use a knowledge based expert system, which can offer an insight into problem solving in process plants, but has its own limitations such as tedious nature of knowledge acquisition, the inability of the system to learn or dynamically improve its performance and the unpredictability of the system outside its domain of expertise. Not possessing these limitations artificial neural network (ANN) can provide a better solution to detecting faults and making decisions because of its effectiveness in representing input-output data, making predictions, classifying data and recognizing patterns [1,2].

In this paper, the application of ANN to detect failures and make decisions in a gas fractionation plant is presented.

Case study description

The gas fractionation plant consists of three distillation columns connected in series. The feed to the plant is liquefied petroleum gas (LPG), which contains hydrocarbons from propane to butanes. In the depropanization column light products (methane and ethane) are recovered in the overhead. The bottom stream is fed to the second column where the mixture of butanes is separated from the natural gasoline. The butanes then enter the debutanization column where i-butane is recovered leaving the n-butane. All the distillation columns operate at high pressure conditions. Due to high operating temperatures, kerosene is used as a heating medium in the reboilers. The column is equipped with a pump-around system. Liquid collected on the trapout tray is drawn out from a side-draw stream. The stream is split into two – a reflux stream and a stream that goes through an external cooler. The cooled stream is the product distillate. The pump-around system provides a means for the vapor in the column to be condensed through direct contact with the cooled liquid from above.

UniSim Design process simulator is used to simulate this process plant. The simulation of the plant can provide data for normal and faulty operation needed for training the ANN and also for testing and validating the DSS. It provides an integrated steady-state and dynamic simulation capabilities. Figures 1 and 2 show the main environment and column environment of the depropanazer column in UniSim Design.

The measurement variables to be used as input for training and testing the ANN are column top stage temperature, column bottom stage temperature, column top stage pressure, reflux mass flowrate, feed mass flowrate, bottom mass flowrate and distillate mass flowrate. The dynamic measurement patterns used to train the network are obtained from plant model simulation.

Process Fault Detection

In order to get normal and faulty data of the gas fractionation plant, different kinds of process conditions are simulated using UniSim Design. To facilitate the analysis of this data a graphical user interface (GUI) is developed using GUIDE toolbox within MATLAB. The GUI can help operator to select different measurement variable data to be compared with the normal ones in dynamic state. The following faults were considered: leak in the feed flow stream, deviations of the feed composition and sensor faults. In this paper only the leakage fault is considered in details.

Figure 1 – Main environment

Figure 2 – Column environment

The leak is simulated by inserting a splitter in the feeding pipeline of the column. For the leak line, the percentage of leakage is controlled by a control valve as shown in Figure 1. Table 1 shows the selected fault conditions and logical representations that represent each of the defined faults. For example, a 5 % decrease in feed flowrate (A) was introduced and simulated; in figure 3 and 4 we can see that the column pressure and top column temperature respectively increased compared to the normal condition.

Table 1 – Selected fault for depropanization column

Fault	Symbol
5 % decrease in feed flowrate	A
10 % decrease in feed flowrate	B
15 % decrease in feed flowrate	C

Pattern Recognition

The development of ANN model for decision support system was carried out using neural network toolbox available within MATLAB. A data set of 100 data points (20 seconds time frame) for each fault and its logical representation were used as input and output of the ANN respectively. The input data was normalized to scale from 0 and 1 before training to ensure the efficiency of the network.

To correctly recognize the fault pattern, interlayer feed forward network was selected as the connection scheme because it is the least complicated one and very straightforward to implement. The network consists of three layers, in which tansig (hyperbolic tangent sigmoid transfer function) is chosen as the transfer function in the hidden layer, and purelin (linear transfer function) is chosen for the output layer.

Network training was implemented using Levenberg-Marquardt Backpropagation (trainlm). The network was cross validated at every batch training and thus the cross-validation errors of the network were monitored throughout the training. Network weights and biases were selected based on the minimum cross-validation error achieved in the training.

Figure 3 – Comparison between A,B,C and normal condition on TS1 - top stage pressure measurement data (kPa)

Figure 5 – Comparison between A,B,C and normal condition on TS1 - top stage temperature measurement data (^oC)

The number of hidden layers and number of hidden nodes are the two important factors that ensure correct classification of faults. The neural network with one hidden layer was used in the simulation. The number of hidden nodes required to perform accurate classification was based on the lowest value of mean square error (MSE).

After the network topology had been selected, the ANN was subjected to novel input data. To assist user in conducting training and validating input, a GUI was developed using GUIDE toolbox within MATLAB. The GUI includes the fault definition and its representation graphs showing the classification results, training and validation command buttons.

To illustrate the network classification capability, faulty conditions were introduced into the depropanization column and the data obtained from the simulation was then subjected into to the network to be classified. For example, when 7% decrease was introduced in feed flowrate (between A and B) the ANN successfully detected the fault. Similar result was obtained when 12% decrease in feed flowrate (between B and C) was introduced into the simulation. Figure 6 shows the results obtained from the GUI when 7 and 12% leak occur. Similar results were obtained when the network was subjected to other fault patterns. The network successfully detected all the faults introduced into the plant.

Figure 6 - Comparison between A, C and fault condition on TS1 – top stage temperature measurement data (^oC)

Conclusion

A fault detection unit as a part of the decision support system for the gas fractionating plant based on an artificial neural network was presented. The proposed scheme for leak detection consists of two stages: process estimator and fault classifier. On the first stage, process plant is simulated in order to get normal and faulty data. On the second stage, the data obtained from the first stage, is analyzed through ANN for compliance with the faults, which have previously been used to train the network. The further development of this study will include detection of other faults and building the framework of the decision support system for accident prevention at process plants.

Bibliography

1 M.R. Othman, M.W. Ali and M.Z. Kamsah “Process Fault Detection and Diagnosis Using Boolean Representation on Fatty Acid Fractionation Column”, International Conference on Chemical and Bioprocess Engineering (ICCBPE) 2003. Sabah, Malaysia.

2 A. Ahmad and M. K. A. Hamid “Pipeline Leak Detection System in a Palm Oil Fractionation Plant Using Artificial Neural Network”, International Conference on Chemical and Bioprocess Engineering (ICCBPE) 2003. Sabah, Malaysia.