Detecting the rate of failure to detect the sudden change of variables in the Tennessee Eastman process using neural networks

Abstract

Changing the parameters in an industrial process causes the process to go out of its optimal working point. This change, in turn, will reduce the efficiency of closed-loop controllers that are designed for the optimal operating point of the system. Therefore, it is necessary to first detect and identify these changes as a defect and then correct the system behavior by changing the process parameters or controller structure as needed. Such a system that has the mentioned feature is called a fault tolerant controller.

It is necessary to design such a system effectively, in the first stage, the proper design of the fault detection system is to correctly identify the location, time and magnitude of the fault. The evaluation of the fault detection system is done with the features of detection/non-detection rate and false alarm rate. Utilizing the appropriate power of neural networks for classification for use in this matter has also been considered in recent years.

In this research, a special type of neural network that has a strong generalized error backpropagation training algorithm has been designed and implemented in order to understand the change of process parameters. This type of network analyzes the information collected from the process in a parallel structure, and we will have the aggregated output of the network as a time and place information indicator of the fault occurrence. After detecting the fault, the controller coefficients will be adjusted. 

Keywords: fault detection and identification, dynamic neural networks, Tennessee Eastman process, non-recognition rate.

1-1- Introduction

During the past decades, with the advent of computer control, the monitoring of complex processes has made tremendous progress. Despite all these advances, there is still a very important control task in managing processes manually and it is performed by human factors. This task is the operation of responding to abnormal events in a process, which includes timely tracking of abnormal events, identifying the causes of their occurrence, and then making appropriate control and monitoring decisions and returning the process to a normal, safe, and operational state. This operation is called the management of abnormal events, which is considered a key element in supervisory control. 1-2- Necessity of troubleshooting

Completely trusting only human factors to control abnormal events and emergency affairs is difficult for several reasons. Diagnosis operations include a wide range of malfunctions, including control unit failures, sudden reduction in process unit efficiency, changes in parameters, and so on. It has complexity. For example, in a large industrial process, 1500 process variables may be observed in a few seconds, which contains a lot of information. After all, fault diagnosis operation based on insufficient data or unstable data leads to unpredictable consequences. In such a difficult situation, where the speed of action is essential, any hasty action by the user may lead to accidents. Industrial statistics show that about 70% of industrial accidents were due to human errors. These abnormal events have significant economic, safety and environmental effects. One of the recent important incidents is the explosion at the Al-Ahmadi oil refinery in Kuwait in June 2000, which caused about 100 million dollars in damage [1]. After all, industry statistics have shown that although major disasters and breakdowns in chemical processes may be rare, small incidents are very common and occur every day. Small accidents lead to many work injuries and cost the industry billions of dollars every year. Therefore, it is very important to quickly and correctly reveal and identify the defects of a process from the point of view of economy, performance safety, and human and biological views. Also, the successful detection of a defect in the early stages can increase the rate of system performance improvement and prevent subsequent dangerous events and unnecessary warnings.     

In systems where safety and reliability are considered as the main indicators of system quality, error detection and identification is considered as a way to improve these indicators.Even small accidents due to errors and malfunctions of tools and equipment will result in loss of life and huge financial costs. Identifying the location of the error, with timely repairability, will reduce the direct and indirect costs of operation and increase the safety and reliability of the system. By spreading the error in the system, in addition to expanding the damage level, it becomes much more difficult to identify its primary cause. Therefore, an accurate error detection and identification method can prevent further errors from spreading in the system and their consequences. Neural networks as an intelligent method for modeling the non-linear complexities of the process can greatly help to improve the performance of the troubleshooting system.

The purpose of using error detection and identification methods, applying advanced monitoring, error management, improving reliability and availability, reducing accidents and their consequences, as well as applying optimal maintenance. These capabilities are obtained through continuous monitoring of the process and measurement of parameters and finally the use of suitable algorithms for analysis. 1-4- An overview of the work done Defect detection methods are divided into two main categories: model-based methods and data-based methods. Methods such as parameter estimation and Kalman filter are among the model-based methods. Artificial neural networks is one of the most widely used data-based methods, which is based on the use of real data measured from the process and is independent of mathematical relationships, so it is widely used in complex and non-linear systems. There are different structures for using neural network in troubleshooting:

Backpropagation neural networks[2]

Multiple neural networks[3]

Sequential neural networks[4]

Hybrid neural networks[5]

Modular neural networks[6]

In recent years, dynamic neural networks [7] have received much attention due to their ability to model nonlinear dynamics of processes. For this reason, different articles and works have been done using this network. For example:

data filtering [7,8]

diagnosis and isolation of operating faults in non-linear systems [9]

diagnosis of satellite engine faults [10]

diagnosis of sugar evaporation process faults [11]

jet engine fault diagnosis [12]

In reference [13], a method based on principal component analysis [1] (PCA) called statistical pattern analysis [2] (SPA) is used for troubleshooting the Tennessee Eastman process, in which, instead of T2 statistics, two statistics Dp and Dr, which are soft and optimistic, are used for troubleshooting in the presence of nonlinear dynamics. It has a good answer. In reference [14], the combination of multivariate PCA and an adaptive neural-fuzzy network [3] (Anfis) is used for troubleshooting. In this combination, PCA is first reduced to linear dimension and then an Anfis network is trained for different classes.

In this thesis, a dynamic neural network has been designed and implemented to detect the defective behavior of the Tennessee Eastman process[4]. We will discuss the pros and cons of each method. In the third chapter, two main methods of fault diagnosis, PCA and artificial neural networks[5] have been investigated in order to find the appropriate structure. These solutions have led to the implementation of the proposed structure. In the fourth chapter, the PCA algorithm implementation on the tested process is explained and simulated. In the fifth chapter, the structure of the proposed neural network is presented along with how it is implemented on the Tennessee Eastman process and the results obtained from the simulation. At the end, it has been summarized and presented suggestions.