This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.
The performance of chemical process plant gradually degraded due to deterioration of process equipments and unpermitted deviation of the characteristic variable of the system. Hence, advanced supervision is required for early detection and isolation of faults for highly integrated and complex chemical processes. This work is aimed to perform fault diagnosis of dynamic gas phase polypropylene (PP) production process mainly on the process faults, by using adaptive neuro-fuzzy inference system (ANFIS). The PP production process is simulated using Aspen Dynamic to generate training data for the neuro-fuzzy system to learn the fault and symptom relationships and to classify the faults occurring in the process. Detection and classification of various fault cases, including multiple faults are successfully performed. The simulation results demonstrate the ability of the proposed fault diagnosis system to handle uncertainty in the measurements, adapt and quickly recognize the process upset of different fault severity levels. Also, the advantages such as sensitivity and robustness of the proposed method over conventional multivariate statistic based on the principal component analysis (PCA) are demonstrated. The simulation results clearly illustrates that the proposed method effectively detect and diagnosis faults in a large scale, dynamical process.
Keywords: Fault Diagnosis, UNIPOL, ANFIS, Principal Component Analysis.
Modern chemical processes are highly integrated system with complicated network of material, energy and equipments. Due to performance degradation in the equipments and process itself, fault detection and diagnosis has received considerable attention from industry. Fault detection is the task of determining whether a fault has occurred, whereas fault diagnosis is the task of determining which fault has occurred. The demand for increasingly advanced supervision is required to improve plant reliability, safety and performance. However, the implementation of the fault detection and diagnosis system in real process plants presents several difficulties such as complexity of modern plants design and control systems, lack of accurate process models, measurements corrupted by noise, highly non-linear symptom and fault relationships and unexpected factors. Therefore, more sophisticated methods are required for detecting, identifying and diagnosing faults in real process plant.
One such process where fault detection and diagnosis are important to maintain its production consistency, above-average product quality and safe operation is the gas phase polypropylene production, UNIPOL PP process established by Union Carbide Corp. Although this process is flexible in its operation and capable of producing broad range polymer products to meet various market demands, the frequent operation mode changes for market pressure may lead to unstable steady-state and unsteady high temperature operation. These can lead to loss of product due to agglomeration of the product particles, equipment damages and process breakdown. Therefore, an efficient online monitoring system which can be easily developed to adapt frequently changing operation modes, quickly recognize the new steady state and identify undesired process upset is required.
Generally, fault detection and diagnosis methods can be classified into data-driven, analytical and knowledge-based methods. The proficiency of the approaches depends on the quality and type of available models, and on the quantity and quality of data available. No single approach can be applied to all applications and usually the best scheme employs a combination of several methods . However, the model-based methods require detailed process model to be effective and may not applicable for large scale system. Also, the industrial processes have different operation modes to meet different demand of product mix. The fault detection and diagnosis must be able to adapt these different operation modes to avoid false alarm. Therefore, artificial intelligence which has the ability to handle uncertainty, adapt changing operation modes, quickly recognize the new steady state and identify undesired process upset is applied for fault detection and diagnosis.
Artificial intelligence such as neural networks and their applications have been extensively studied. Jamal et al.  provided an extensive review of the various applications of neural networks for chemical engineering purposes and the comparisons to existing conventional methods are also shown. Hussain  provided an extensive review of the various applications utilizing neural networks for chemical process control, both in simulation and online implementation. Being an excellent tool for pattern recognition and modeling non-linear relationships, neural networks have been widely applied in the field of fault diagnosis where the primary goal was discriminating among the faults. Watanabe et al.  presented a neural network to detect single faults followed by estimation of the fault severity in a heptane-to-toluene aromatization process. Sorsa et al.  investigated different neural network structures in order to obtain a suitable one for fault diagnosis of an exothermic CSTR system. Fan et al.  added a number of functional units into the input layer of a conventional back-propagation neural network to enhance the representation capability of the network. Hussain et al.  illustrated the application of neural networks in process fault diagnosis of acrolein and acrylic acid production plants.
Also, the concept of combining fuzzy logic and neural network has grown into a popular research topic [8-12]. This hybrid system possesses both the capability of fuzzy system in handling partial knowledge and the learning capability of neural networks. One of the advantages of having this hybrid system is faster convergence speed with smaller network size as compared to classical neural networks . Calado and Sa da Costa  developed a fault detection and isolation (FDI) system based on a hierarchical structure of several fuzzy-neuro networks to detect actuator faults. Kok et al.  applied a Fuzzy Min-Max (FMM) neural network which integrated with rule extraction algorithm to fault diagnosis tasks of a Circulation Water (CW) system in a power generation plant. Akhlaghi et al. [16-17] applied principal component analysis (PCA) and independent component analysis (ICA) to extract salient features from available input dimension and employed multiple neuro-fuzzy network to detect and diagnose the faults occurred in theoretical distillation column.
Several process fault detection and diagnosis system have been developed for the polypropylene production process. Xiong et al.  developed a multivariate statistical process monitoring based on principal component analysis (PCA) and independent component analysis (ICA) and combined the statistical methods with kernel density estimation (KDE) to monitor the polypropylene catalytic reactor. The studies of knowledge-based process monitoring system using digraph-based system were conducted by Kyung et al.  and Nam et al. . In the digraph-based system, a RCED (Reduced Cause Effect Digraph) was constructed and stored in knowledge base. Then, a PGTT (Pattern Graph Through Time) was generated on the real-time basis during the diagnosis period to represent cause-effect relationships and current state of the process.
Gani et al.  developed a model-based process monitoring system using fault detection filter to detect occurrence of fault in the control actuator and further derived an active fault-tolerant control with switching policy for a gas phase polyethylene reactor. Charles et al.  developed a model based fault-detection and isolation (FDI) scheme using asynchronous measurements. Then, a fault-tolerant control method with a supervisory control component was employed to achieve stability in the presence of actuator failures using control system reconfiguration. A process monitoring system based on self-organizing maps and multiple linear local models was developed by Abonyi et al.  to detect the typical operating regions related to different product grades and predict the product quality of a polyethylene plant.
In this paper, a comprehensive fault diagnosis system for a dynamic gas phase polypropylene production, UNIPOL PP process is developed by using an adaptive neuro-fuzzy inference system (ANFIS). The application of ANFIS for fault diagnosis aims at developing a model that can detect abrupt, incipient as well as single and multiple process fault correctly or at least with a low misclassification rate. The UNIPOL PP process is simulated using Aspen Dynamic to generate process data for the neuro-fuzzy systems to learn process fault and symptom relationships as well as to classify the faults occurring in the process. This neuro-fuzzy system acquires symptom and fault relationships without requiring a priori knowledge of the process, avoiding the difficulty of acquiring process information that is faced by most complex industrial processes. Also, this system has the ability to adapt changing operation modes, quickly recognize the new steady state and identify undesired process upset. Successful detections for various fault cases are obtained and the faults severity levels are simultaneously identified. The effectiveness of the proposed system is compared with the conventional multivariate statistical process monitoring based on the principal component analysis (PCA) results.
This paper is organized as follows. In the next section, the basic of ANFIS and feature selection using ANFIS are introduced. Section 3 explains fault diagnosis using ANFIS. Application of the proposed approach to the UNIPOL PP process and the comparison with multivariate statistical process monitoring using PCA are discussed in Section 4. The last section contains conclusions for this study.
2.1 ADAPTIVE NEURO-FUZZY INFERENCE SYSTEM (ANFIS)
ANFIS is a data-driven based method. A fuzzy inference system which is based on the Sugeno fuzzy model [24-25] is created by fuzzy clustering. The characteristic of non-linear system can be concisely represented by identification of the grouping of the data in a collection of If-Then rules. A typical fuzzy rule in a Sugeno fuzzy model has the format
If x is A and y is B then z = f(x, y), (1)
where A and B are fuzzy sets in the antecedent; z = f(x, y) is a crisp function in the consequent. For a first-order Sugeno fuzzy model, f(x, y) is a first-order polynomial. For a zero-order Sugeno fuzzy model, the output level, f is a constant. To facilitate the learning (or adaptation) of Sugeno fuzzy model, an adaptive networks that can search optimal parameters systematically is incorporated. The resulting network structure, called ANFIS (Adaptive Neuro-Fuzzy Inference System) is shown in Figure 1. A typical ANFIS consists of 5 layers where the function of each layer as described below. The output of the i-th node in l-th layer is denoted as Ol, j.
Fig. 1. Typical multi-input-single-output (MISO) structure of ANFIS.
Each node in this layer generates membership degrees through a node function. Example of node function is generalized bell membership function:
where x is the input to node i, ai, bi, and ci is the parameters that change the shapes of the membership function. The parameters in this layer are referred to as the premise parameters.
Each node in this layer generates firing strength of a rule via multiplication:
The i-th node calculates the ratio of the i-th rule's firing strength to the total of all firing strengths:
The i-th node calculates the contribution of the i-th rule towards the overall output with following node function:
where pi, qi and ri are the consequent parameters.
The single node in this layer computes the overall output as the summation of contribution from each rule:
The training process of ANFIS for searching of optimum parameters can be assisted by hybrid learning method which combines gradient descent and the least squares method. The Fuzzy Logic Toolbox from Matlab is utilized to generate the ANFIS model in this research. More detailed coverage of ANFIS can be found in [24-25].
2.1.1 Implementation of Fault Diagnosis using ANFIS
Usually, the symptom and fault associations of industrial processes are known. By training the ANFIS with the symptom and fault relationships, the symptoms of a fault can be used to pinpoint the root cause of the abnormal state. This eliminates the requirement of detailed process mathematical model for large scale systems which is expensive to obtain. Figure 2 shows typical structure of ANFIS for the fault diagnosis. The measurement variables obtained from the process are normalized and assigned as the inputs of ANFIS while the output works as fault indicators. Each ANFIS is uniquely trained to classify and diagnose specific fault. Since ANFIS itself only suitable for single output system, multiple ANFIS models are constructed to diagnose the faults occurred in the process. When the inputs are associated with a particular fault, only the respective ANFIS will be activated and indicates the fault has occurred.
Fig. 2. Structure of ANFIS for process fault diagnosis.
In a complex chemical plant, the numbers of measured variables can easily be up to hundreds or thousands. Monitoring all the measured variables is highly redundant and it also increases the computational burden of the system. Statistical dimension reduction methods such as principal component analysis (PCA) and independent component analysis (ICA) have been used for feature extraction [16-17] in order to improve the recognition process. However, the statistical techniques linearly decorrelate the original measurement data with certain assumptions on the data distribution or correlations which may not be appropriate in practice. Also, the higher order statistical techniques can be computationally quite complicated and unreliable.
Jang  has proposed a heuristic way to quickly determine the potential inputs and select the inputs with highest priority for ANFIS modeling. The proposed method was based on the assumption that the ANFIS model with the smallest RMSE after one epoch of training has a greater potential of achieving a lower RMSE when given more epochs of training. The input selection method is summarized as follows:
Step 1: Given n candidate inputs, a subset of k inputs are selected as the inputs to ANFIS.
Step 2: For each different combination of the k inputs, one AFNIS model is constructed.
Step 3: Each ANFIS model is trained for single epoch and the training error is recorded.
Step 4: Step 3 is repeated for other ANFIS models with different combination of k inputs.
Step 5: The ANFIS model with the smallest training error is selected for further training.
The feature selection method is straight forward and computationally effective as fast training can be achieved by the least squares method. The model produces lowest training error consists of the most influential inputs for the recognition process will be selected for further training to fine-tune the parameters.
2.2 PRINCIPAL COMPONENT ANALYSIS (PCA)
Numerous applications of principal component analysis (PCA) have been reported to extract structure from multidimensional chemical process data and modeling the multivariate process for monitoring purpose [1, 27]. PCA, as an optimal dimensionality reduction technique, determines the most accurate lower dimensional representation of the data and captures the data directions that have the most variance. PCA determines a set of loading vectors by solving the singular value decomposition (SVD) or the covariance matrix. The loading vector corresponding to the larger singular value described the linear trends in the data while the remaining loading vector corresponding to the smaller singular value describes the random noise. Detailed descriptions of PCA and selection of loading vectors are covered in [1, 27].
2.2.1 Fault Detection using PCA
A general and yet simple approach for detecting abnormal status is by using the T2 and Q control chart for the loading vectors retained in the PCA model. The T2 statistic measures the variations in score space which represent the normal operations. This approach detects most of the faults that produced large mean shift in the measurement variables (Kano et al. 2001). By retaining loading vectors corresponding to a, the number of largest singular value, the T2 statistic (Chiang et al., 2001) can be computed by
where is the testing set, P includes the loading vectors associated with the a largest singular values (a < m), Σa contains the first a rows and columns of Σ, and x is an observation vector of dimension m. The threshold of T2 statistic (Himmelblau, 1978; Chiang et al., 2001) can be determined by using the probability distribution,
where Fα (a ,n-a) is the upper 100α% critical point of the F-distribution with a and n - a degrees of freedom. Any value of T2 statistic exceeding the threshold value indicates faulty operating conditions. The portion of the measurement space corresponding to the lowest m - a singular values can be monitored by using the Q statistics (Chiang et al., 2001),
where r is the residual vector. The Q statistic measures the amount of variation not captured by the PCA model. The threshold for the Q statistic can be computed from its approximate distribution,
where , and cα is the normal deviate corresponding to the upper (1 - α) percentile.
The proficiencies of the PCA statistics are investigated based on the misclassification rates for faults of the observation data. The misclassification rate for each fault class is defined as:
where is the original testing set and is the data set of PCA transformation.
3.1 SIMULATION OF THE UNIPOL PP PROCESS
Fig. 3. The UNIPOL PP process.
Figure 3 shows the UNIPOL PP process. The raw materials including propane (inert), propylene and hydrogen will be sent to the purification column prior to the reactor. The impurities in raw material such as water, carbon monoxide and carbon dioxide are removed. Next, the purified raw materials are transferred to the fluidized bed reactor where gaseous propylene is contacted with solid catalysts to initiate the polymerization process. The yield of PP for one pass is very low. For that reason, large portion of the reactants in vapor forms are recycled through the recirculation/cooling loop to improve the yield of PP. The polymer in granular forms is withdrawn from the reactor through the resin discharged system and the unreacted components are separated from the products in the separators. Inert gas (nitrogen) blanketing is used to remove remaining reactants in the products and transfer the products to additive and packaging system. A portion of hydrocarbon is purged to the flare stack to keep the inert from accumulating in the process.
Components feed rate and operating conditions of UNIPOL process.
Flow Rate (kg/hr)
Titanium Tetrachloride (Catalyst)
Polymer phase volume
In this study, the UNIPOL PP process is simulated and controlled using Aspen Dynamic. The fluidized bed reactor is modeled using the CSTR reactor model with two phases: a vapor phase and a polymer phase. The POLYPCSF (PCSAFT) thermodynamic model is used to relate the gas phase monomer, hydrogen and etc. composition to their concentration in the polymer phase. The Ziegler-Natta based kinetic model is used to describe the polymerization reactions in the polymer phase. Both models are build-in model in the simulation software. Each data sample is collected every 6 minutes for a duration of 40 simulation hours. The reactor and separators are equipped with controllers to regulate the temperature, pressure and level in the reactor and separators. The process conditions are as listed in Table I.
3.2 FAULT DIAGNOSIS USING ANFIS
This section describes fault diagnosis of the UNIPOL PP process using the proposed fault diagnosis system. In this study, four different process faults scenarios are considered as follows:
Fault 1: Partial blockage in catalyst supply line, leading to a decrease in the catalyst supply rate, FCAT.
Fault 2: Heat exchanger in recycle stream fouled, leading to a decrease of the heat transfer area, UA.
Fault 3: Performance of compressor in recycle stream degraded, leading to a decrease in the compressor discharge pressure, PVAP-B.
Fault 4: Performance of propylene supply pump degraded, leading to a decrease in propylene feed pressure, PC3FEED.
The possible process faults and fault severity are obtained through abrupt fault simulation by decreasing the value of the parameters FCAT, UA, PVAP-B and PC3FEED. All possible process faults considered are examined for three different severity levels: A fault corresponds to 5% decrease of the parameter is slight, a fault at corresponding to 10% decrease is medium and a fault corresponding to 15% decrease is severe. For severe fault case, the corresponding emergency measures will be taken once the fault is detected and identified. A fault which severe than that will be handled in the same way. Thus, further discrimination is not necessary.
The measurement data from the UNIPOL PP process consists of eight input variables: (1) TCYCGAS, (2) PCYCGAS, (3) PRFEED, (4) FRFEED, (5) FVAP-A, (6) PVAP-A, (7) TREACTOR and (8) FPROD. In order to enhance the fault diagnosis and ease the computational burden, the most influential inputs are selected by using the feature selection described in Section 2. Here, the four most relevant inputs will be selected as the inputs to ANFIS. A total of ANFIS models can be constructed and trained with single epoch. The best ANFIS model which produces lowest training error will be selected for further training. To avoid particular inputs dominating the input space, the selected inputs are normalized as:
where and σi denote the normal operation values and standard deviation of corresponding variables respectively, r denotes the user defined scaling factor such that the normalized variables, xi* can map within the desired input range of (-5, 5).
Four ANFIS models were constructed to classify the normal condition and four different fault scenarios with each comprising of three different types of fault severity. Each ANFIS model is uniquely trained to discriminate one fault type. The four most influential features selected are fed as the inputs to ANFIS and further trained. The training data and testing data, each consists of 200 observations with the first 50 observations in the normal condition and the rest 150 in the abnormal one. The training of the ANFIS continues until the mean square error less than 0.001. The ANFIS parameters are adjusted to maintain the size of the network while not upsetting its predictions and generalization capabilities. At fault free case, all ANFIS will take value of 0.1. If the inputs are associated with ith fault, only the output of ith ANFIS will take value encoding different severity levels of the corresponding ith fault: 0.7 for low, 0.8 for medium and 0.9 for severe, while others will take value of 0.1. A threshold, σ = 0.6 is employed to work as detection limit. If the output value less than 0.6, it is considered as fault free.
4 RESULTS AND DISCUSSION
4.1 SINGLE FAULT CASE STUDIES
This section demonstrates fault diagnosis using ANFIS for selected faults. The fault patterns and diagnosis results associated with Fault 1are shown in Figure 4. The fault patterns are corrupted by measurement noise with standard deviation of 3%. At normal condition, the ANFIS outputs take values less than 0.6. As fault is occurred, the first ANFIS is activated and take value of 0.7 to indicate that the catalyst supply line is slightly blocked. Since the fault is not removed from the process, the ANFIS classifier follows the fault effects constantly. Other ANFIS outputs take value around 0.1 as the classifiers recognize the fault patterns are only related to Fault 1.
Fig. 4. Normalized inputs and ANFIS outputs in the case of partial blockage in catalyst supply line.
The fault patterns and fault diagnosis results for incipient Fault 2 are shown in Figure 5. Also, the measured variables consist of measurement noise. At normal condition, all the ANFIS outputs take values of 0.1. As Fault 2 is occurred at 50-th sample, the second ANFIS is triggered to pinpoint that the heat exchanger is fouled and the fouling is medium (output » 0.8). The ANFIS classifiers follow the fault effects until the variables reach their new steady state. This shows that the ANFIS classifiers can correctly pinpoint the root cause of the process malfunction and the corresponding fault severity level. The diagnosis results for single fault cases are summarized in Table II. It is observed that the ANFIS correctly detects all the faults considered and identifies the corresponding fault levels as well. Small deviation of the prediction values and targets is caused by the error of the ANFIS model.
Fig. 5. Normalized inputs and ANFIS outputs in the case of recycle stream heat exchanger fouling.
Diagnosis results of ANFIS for single fault cases.
Fault No. /
ANFIS Output Value
1 / 1
1 / 2
1 / 3
2 / 1
2 / 2
2 / 3
3 / 1
3 / 2
3 / 3
4 / 1
4 / 2
4 / 3
Note: Encoding value of fault levels: Normal = 0.1, Level 1 = 0.7, Level 2 = 0.8 and Level 3 = 0.9.
Multivariate statistical process monitoring of the UNIPOL PP process using principal component analysis (PCA) is compared with the proposed method. The statistical fault detection results for Fault 2 are shown in Figure 6. The statistical fault detection is based on T2 and Q statistics with 99% confidence limit. Three loading vectors were retained which represent a total variance of 92.7%. The abnormal status is detected at 60-th sample. However, the detection turns out to be ineffective (T2 << Tα2 and Q << Qα) after 120-th sample as the process reaches its new steady state. The fault occurred in this case produces small effects on the process behavior which is difficult to be detected using statistical method because the statistics are slow in detecting small changes in mean or data which is statistical stationary.
Fig. 6. Fault detection using PCA statistics in the case of recycle stream heat exchanger fouling.
Compared with the proposed method, the misclassification rate for each fault based on PCA statistics are summarized in Table III. The results show that the proposed method can be used to monitor the UNIPOL PP process more effectively than statistical PCA method. Also, the proposed method produces lower missed detection and false alarm rates compared to PCA statistics. This is because the PCA model assumes the data are collected from a steady state process containing only linear correlations between the variables. For the fault with small disturbances or dynamic characteristics, the statistical fault detection method is less likely to produce good results. Although PCA method based on T2 statistic produces satisfactory fault detection results, this statistic does not provide information corresponding to fault severity level.
Comparison of misclassification rates for single fault cases.
4.3 MULTIPLE FAULT CASE STUDIES USING ANFIS
Multiple faults occurring within the same time window are likely to happen for many industrial processes. Fault diagnosis of multiple faults presents more challenging compared to single fault case. This section will discuss fault diagnosis using ANFIS for selected multiple faults scenarios for the UNIPOL PP process. Figure 7 shows the fault patterns and diagnosis result for Fault 2 (recycle stream heat exchanger fouled) and Fault 4 (performance of propylene supply pump degraded). Similar to single fault cases, all faults are introduced at the 5-th simulation hour (50-th sample). Correct detection and classification are again observed. The second and the forth ANFIS take value of approximately 0.9 and 0.7 respectively, indicating that Fault 2 is severe and Fault 4 is slight. Although the process slowly shifts to its new steady state, the ANFIS classifiers still consistently recognize the process upsets and identify the severities for the two different faults.
Fig. 7. Normalized inputs and ANFIS outputs for recycle stream heat exchanger fouling (Fault 2) and performance of the propylene supply pump degraded (Fault 4).
Another multiple faults case is studied. At around 50-th sample, Fault 1 (partial blockage of catalyst supply line) and recycle Fault 3 (stream compressor degraded) are introduced into the process. Figure 8 shows the fault patterns and diagnosis results. As soon as faults are introduced into the process, the first and the third ANFIS take value of approximately 0.9 and 0.7 respectively, indicating that Fault 1 is severe and Fault 3 is slight. The alarms for Fault 2 and Fault 4 are not triggered (output < 0.6) as the ANFIS classifiers recognize that the faults patterns are not belong to the fault classes.
Fig. 8. Normalized inputs and ANFIS outputs for partial blockage of catalyst supply line and performance of the recycle stream compressor degraded.
Comparing the single and multiple fault case studies, it is observed that ANFIS can produce superior fault detection and diagnosis over PCA statistics. Not only does it possess fast detection rate, ANFIS also produces lower and consistent misclassification rates over the entire testing data as it can consistently follows the fault effect until the system reaches new its new steady state. Given appropriate and sufficient training, ANFIS able to detect and identify incipient faults which produce small changes in the measurement variables. The training of ANFIS for multiple simultaneous fault cases, however, are slower compared to that of single fault cases only as the number of the fault combination increases exponentially with the number of faults. Also, the task of diagnosing multiple faults is much more challenging and the proficiencies of the ANFIS classifiers depend on the nature of the combination of the faults. Lower classification accuracy can be expected due to increasingly complexity of the fault patterns.
A method for online fault diagnosis of the UNIPOL process using ANFIS was presented. Occurrence of single fault and multiple simultaneous faults in the UNIPOL process were successfully detected, followed by the diagnosis of the faults and identification of the different fault level. The advantages of ANFIS over PCA method, in terms of promptness and sensitivity, were illustrated by detection of incipient fault which producing almost statistically unchanged measurements. This indicated ANFIS based monitoring is efficient for process that requires rapid fault detection. This ANFIS based fault diagnosis system able to handle measurement noise, adapt changing operation modes, and rapidly recognize the new steady state to identify undesired process upset. Also, it eliminated the need of detailed process mathematical model which were expensive to obtain for large scale systems. Thus, this approach may contribute to minimize the process downtime and production losses by supporting and improving the troubleshooting, decision making and maintenance tasks. Further works and studies involve extracting salient futures from large dimensional measurement data so that a plant-wide fault diagnosis system can be developed.