Inference System Based Fault Detection Computer Science Essay

Published Date: 02 Nov 2017

Adaptive Neuro Fuzzy Inference System based Fault Detection and Isolation Scheme for Pneumatic Process Control Valve

K. Prabakaran, T. Uma Mageshwari, and A. SugunaAbstract— As modern process industries become more complex, the importance to detect and identify the faulty operation of pneumatic process control valves is increasing rapidly. The fault diagnosis in the control valve is not an easy process due to inherent nonlinearity. The prior detection of faults leads to avoiding the system shutdown, breakdown, raw material damage and etc. This paper proposes a new integrated diagnostic system for pneumatic control valve fault diagnosis by means of a neuro fuzzy approach. The particular values of five measurable quantities from the valve are depend on the commonly occurring faults such as Incorrect supply pressure, Diaphragm leakage and Actuator vent blockage. The correlations between these parameters from the fault values for each operating condition are recognized by an Adaptive Neuro Fuzzy Inference System (ANFIS). The parameter consideration is done through the committee of Development and Application of Methods for Actuator Diagnosis in Industrial Control Systems (DAMADICS). The simulation results using MATLab prove that Adaptive Neuro Fuzzy Inference System has the ability to detect and identify various magnitudes of the faults and can isolate multiple faults.

Index Terms— ANFIS, BP, DAMADICS, MATLab

INTRODUCTION

A common element in modem process industries is the pneumatic process control valve, which is used to control the flow of a liquid, gas or slurry. Failures in these valves due to some abnormality operating conditions give rise to disturbances in the system process. The result is process output deviate from required output and some time unscheduled process shut down. The increasing complexity of process industries and the needed to reduce the overall

production costs leads to development of suitable techniques for detecting and assigning causes to valve failures [1].

A number of techniques for fault detection and identification (FDI) have been developed and can be applied to process control valves. In general, FDI techniques follow some measureable parameters related to the performance of the system. When the parameters deviate from their original values, it is assumed that a fault occurred in the valve. If the dependent parameters are carefully selected then that is enough to identify each fault. The design of an effective FDI system requires: (i) a method for obtaining fault dependent parameters related to the system performance, and (ii) a decision process that identifies the specific operating condition related to a particular set of dependent parameters [1].

There are a number of papers that propose different techniques for the fault detection. Beard (1971) & Jones (1973) proposed Beard-Jones Fault Detection Filter based on Observer-based fault detection scheme [2]. Clark & Fosth & Walton, (1975) developed Luenberger Observers based on residual generation scheme [3]. Rank (1987) & Isermann (1991) & Basseville and Nikiforov (1993) proposed Classical fault diagnosis but this method is theoretical only [4]. Cordier et al., (2000) & (de Kleer and Kurien, 2003) implemented Model-Based Diagnosis (MBD) but exact model of complex system is difficult [5].

To build highly efficient, timely and accurate fault diagnosis systems which ensure production safety has become focus research in control field. Modern methods to solve FDI problems in systems with inherent dynamics can be classified into three broad categories. The first being a model based techniques. But developing mathematical models for complex systems or for nonlinear systems is difficult, so that model-based scheme has some disadvantage. Further more even make a model; the model should be validated by means of experimental method. This method is not suitable for practical complex system. The third class method is to use neuro fuzzy approach as fault classifiers to solve FDI problems. This paper investigates an Adaptive Neuro Fuzzy Inference System to diagnose faults in the process control valve using MATLab [6].

Pneumatic Control Valve

The most important final control element (FCE) in the process automation industries is the pneumatic control valve. The pneumatic control valve adjusts the manipulated variable such as flow of gas, steam, water, or chemical agents, to keep the desired value of set point which affected from the load disturbance.

A pneumatic servo-actuated industrial control valve, which is used as test bed of the fault detection approach proposed in this paper .The internal structure of pneumatic control valve, is shown in Fig. 1.

Fig. 1. Internal structure of pneumatic control valve.

Actuator Components

The pneumatic control valve includes three parts: spring-and diaphragm pneumatic servo-motor, positioner, and valve, as shown in Fig. 1.

Spring-and-diaphragm pneumatic servomotor

Spring-and-diaphragm pneumatic servomotor is a compressible (air) fluid powered device in which the fluid acts upon the flexible diaphragm, to provide linear motion of the servomotor stem.

Positioner

Eliminate the control-valve-stem miss-positions produced by the external or internal sources such as friction, pressure unbalance, and hydrodynamic forces.

Valve

To prevent, allow and/or limit the flow of fluids through control systems. Changing the state of the control valve is accomplished by a servomotor [7].

Internal Parameters of Actuator

S -Pneumatic servo-motor

V -Control valve

P -Positioner

ZC -Position P Controller (internal loop)

E/P -Electro-Pneumatic Transmitter

Additional external components

V1 -Cut-Off Valve

V2 -Cut-Off Valve

V3 -By-Pass Valve

PSP -Positioner Supply Pressure

PT -Pressure Transmitter

FT -Volume Flow Rate Transmitter

TT -Temperature Transmitter

Set of basic measured physical values

CV -External (Flow or Level) Controller Output

F -Flow Sensor Measurement

P1 -Valve Input Pressure

P2 -Valve Output Pressure

T1 -Liquid Temperature

X -Rod Displacement

Control Valve Faults

DAMADICS committee is concerning on the development of pneumatic control valve fault detection and isolation (FDI) algorithms [6].The main goal of DAMADICS benchmarks is the creation of well defined, repeatable single actuator faults. For this purpose the set of actuator faults was identified [8].

DAMADICS predefined the 19 types of faults which are going to be occurring in the pneumatic valve during the process [9]. The faults of control valve are classified into four following groups: Control valve faults; Pneumatic servo-motor faults; Positioner faults; General faults/external faults. Mostly, single actuator faults are observed in industrial practice whereas multiple faults rarely occur. When fault is occurring, dependent parameters will be vary from the normal condition. So these parameters are sufficient to characterize the changes in the performance of the actuator due to the occurrence of the faults under investigation [6].

Fault Considered for Diagnosis

In real time application several faults may occur in pneumatic control valve. Three commonly occurring faults are

– Incorrect supply pressure

– Actuator vent blockage

– Diaphragm leakage

These faults are going to be diagnosed by the Neural Network methods.

Dependent Parameters Considered for Diagnosis

The following five parameters are considered to identify the above three faults which are approved by the DAMADICS [8].

– Valve Input Pressure (kPa)

– Valve Output Pressure (kPa)

– Flow Sensor Measurement (m3/h)

– Rod Displacement (%)

– External (Flow or Level) Controller Output (%)

Adaptive Neuro Fuzzy Inference System

ANFIS is an adaptive system that constructs a set of fuzzy ‘IF-THEN’ rules with proper membership functions to create the stipulated input-output pairs. Here, the membership functions are tuned to the input-output data by neural network [10].The adaptive neuro-fuzzy systems may be used either for fault identification (fault detection) or for fault classification (fault isolation) purposes. The following subsection explains the adaptive neuro-fuzzy systems used for detecting the parameters of Takagi-Sugeno fuzzy models, which may be used for fault detection [10] and a neuro-fuzzy structure used for fault isolation [12].

Takagi-Sugeno type Neuro-Fuzzy ARMA Model for Fault Detection

Takagi-Sugeno model has as consequence of the fuzzy rules ARMA (Auto Regressive Moving Average) models [10] of higher order as shown in Eq. (1).

(1)

where i=1,…,r, r is the number of rules, x=(x1, x2, …, xk) is the input vector, and x(t-j), y(t-j), j=1,…,n1 or n2, represent the past values for the inputs and output of the system. If the two sums in the consequent of the rule given in Eq. (1) are missing, we obtain the well-known form of a Takagi- Sugeno model of order zero.

In order to design a Takagi-Sugeno model, the following three sets of parameters need to be identified using the available input-output data measurements [13]:

The actual input variables (x1… xk) composing the antecedent of the rule.

Ai1,…,Aik – the membership functions of the fuzzy sets in the rule antecedent

ci, pi, si – the parameters in the consequence of the rule.

The number and the membership functions of the fuzzy sets Fts, t=1… rs, associated with each input variable xs, s=1,…, k, must be determined before building the neural network. The space associated with each variable can be empirically partitioned into fuzzy sets by analyzing the way the system operates. This can be a very difficult task when dealing with complex systems. Other techniques that can be employed are clustering and genetic algorithms. The fuzzy sets in the antecedent of the rules for input s, s=1,…, k, are elements of the set { Fts | t=1… rs }.

The first set of parameters (actual inputs used in the antecedent) represents a subset of all inputs of the system and it can be determined using the heuristic search algorithm [13]. The method is concerned with making two choices. The first choice represents the choice of the variables that will appear in the antecedent of the rules. Each variable has associated with itself a fuzzy partition on its space. The second choice represents the number of fuzzy sets in the partition.

The third set of parameters is identified using training algorithms for neuro-fuzzy systems for Takagi-Sugeno model implementation. These systems put the set of fuzzy rules of the model under the form of a neural network as shown in Fig. 2 [11].The parameters are identified during the training of the neuro-fuzzy network. The ARMA model in the consequence of a fuzzy rule is implemented by a sub network as shown in Fig. 3

Fig. 2. Neuro-fuzzy network for Takagi-Sugeno fuzzy model implementation.

Fig. 3. The sub network corresponding to i-th neuron in the 4th layer.

Mamdani- type Neuro-Fuzzy Hierarchical Structures for Fault Isolation

A hierarchical architecture of Mamdani-type neuro-fuzzy structures (also called as fuzzy-neural networks (FNNs)) for fault isolation purposes is shown Fig. 4. The structure aims to correctly classify input symptoms corresponding to both abrupt and incipient faults (single or multiple), using only abrupt faults symptoms and normal state symptoms during the training phase. The symptoms are generated by selecting from residuals, and their combinations, those signals that provide the best distinction between different operating states of the system [11].

The hierarchical structure has the three levels shown in Fig. 4. The first-order differences for all available measurements are used as symptoms. The lower level consists of one FNN that receives as input the considered symptoms. The output of this FNN determines which of the FNNs on the medium level will be activated. That is, if the i-th component of the output has a value close to 1, then the i-th FNN on the medium level will be activated.

Fig. 4. A hierarchical structure of neuro-fuzzy networks.

The number of the FNNs on the medium level is equal to the number of faults considered. Each one of them is also fed with all symptoms considered. The upper level is used to perform an OR operation on the outputs of the activated FNNs on the medium level. The components of the outputs considered for the OR operation must have a value close to 1.

The ANFIS training routine for Sugeno-type Fuzzy Inference System (MEX only) is shown in Eq. (2) which is used as basic function for FIS creation.

(2)

Results

The real time data measured under normal and abnormal condition of pneumatic control using data acquisition is given to ANFIS. Totally 600 data are collected under various operating conditions including no fault condition. The TABLE I shows the output of the ANFIS.

TABLE I

Results of Adaptive Neuro Fuzzy Interference System

Parameters

ANFIS Output

No. of training data

1500

No. of checking data

0

Training error

0.000569417

Designated epoch

1000

Computational Time

0.1605minutes

The designated target epoch was 1000 and the ANFIS training completed at epoch 1000 within 0.1605 minutes with minimum number of fuzzy rules. The training error was plotted against the number of epoch as shown in Fig. 5.The Training error was minimized at 1000 epoch with 0.004916507 values.

Fig. 5. Training error verses number of epoch plot

The ANFIS structure created for training was shown in the Fig. 6.The structure shows that only 6 fuzzy rules were created to training the fuzzy interference system.

Fig. 6. Structure of created ANFIS for training.

From the analysis of above plot and ANFIS output the created fuzzy interference system has the perfect ability to diagnosis control valve faults.

Conclusion

In this paper an ANFIS based scheme for detection and isolation of pneumatic control valve faults was proposed. The specific values of five parameters were observed to depend upon the particular type of fault. For each operating condition, the dependent parameters changed its state which is recognized by a fuzzy interference system with the goal of successfully detecting and isolating the faults. The simulation results proved that the trained fuzzy interference system has the capability to detect and identify the various magnitudes of the faults with better performance.