Assessment Of Component Structural Health By Rule Based Classification

Assessment Of Component Structural Health By Rule Based Classification

Paulin Evans Pagalla

Mtech-IIyear, Department of Computer Science and

Engineering, Vignan University, Guntur, India

R. Prathap Kumar

M.Tech CSE: Assistant Professor Department of Computer Science and Engineering, Vignan University Guntur, India

Abstract

Structural health is crucial for overall performance of the entire system. Technical failure of any component can effect its success. Such problems can be avoided well before by assessing the structural health and thus preventing unforeseen catastrophes. Rule Based Classification uses IF THEN rules for the assessing and classifying the data to different class labels. For example; using a flight database, a technician may want to assess the condition of a component in an aircraft with respect to few spatial dimensions (like route temperature, landing conditions, load etc). Depending on the structural health (of component) he might want the aircrafts to be scheduled for either the next flight or for maintenance. In this paper, we brief about the rule based classification, the spatial dimensions and propose appropriate algorithm for classifying the flight data into different class label. Synthetic data sets are used for implementing the system.

Keywords: Rule based classification, spatial databases

1. Introduction

   Spatial database is a large collection of data like maps, satellite imagery, scientific and engineering application data etc. A System comprises of many components; components wear during operations, faulty components can decrease the systems life or even completely fail the system. Technical failures can be avoided before they actually occur by analyzing their health. An airliner database contains flight data and flight maintenance information. Flight data can be categorized as spatial data. It is a collection of both non spatial (ex. altitude, velocity, passenger load, temperature of the fuel lines etc.) and spatial data (commands from ground systems, images, cockpit voice recordings, route information etc.). Flight Maintenance information is a record of all the maintenance operations that are done on the aircraft during routine servicing. The data are analyzed by an analysis team for technical and structural health of the aircraft. Periodical Maintenance of the aircrafts is scheduled as per the manufacturers advice, for instance, for every 1000 hours of flight. Components on an older aircraft experience frequent technical malfunctions than newer aircrafts hence the older components needs to be serviced more often than the manufacturers advice. In this paper, we propose algorithm for assessing the health of the components and assign class labels to the data sets using rule based

   classification of spatial data (flight data). The rest of the paper is structured as follows: Section 2 provides explanation on rule based classification and the Sequential Covering Algorithm. Section 3 describes the application and the results finally; Section 4 concludes the paper with future research directions.

2. Rule Based Classification of Spatial Data

   Given a set D of data, rule based classification will assess the health of the component and assign the data sets to class labels. The health of a component is defined by structural and technical strength evaluated after the assessment. As a motivating example, a technician might want to assess the health of the engine of the aircraft with respect to the temperature exposure on fuel lines during the flight or for an instance he might want to assess the health of the landing system with respect to the vibration during its landing (whether a smooth landing or rough landing), and load on the aircraft. The basic structure of a rule is as follows:

   IF condition THEN conclusion.

   The condition is called the precondition of the rule which is also called as test condition, and the conclusion is the consequence if the condition is satisfied. The precondition usually consists of one or more test conditions that should be met. The rule consequence contains the class prediction or the class label. For ex., consider the rule R1given below, temp is the parameter and threshold is the limit where the parameter is not allowed to exceed i.e., if the temp parameter crosses the given threshold then an alarm set off. If the condition holds, we say that the rule covers the given data set.

   R1: IF temp>threshold THEN alarm.

   Usually a Rule Based Classifier extracts rules by using either by algorithms or by decision trees. As the size of the data sets increases the Decision tree extraction become complex over time and usually contain repetition of branches which have

   to be pruned later. Algorithms like AQ, CN2, and RIPPER are few examples. In our paper we have used a Sequential Covering Algorithms for rule extraction, which extracts rules from the data by one rule at a time.

   Sequential Covering Algorithms:

   The general strategy of the algorithm is to learn one rule at time for a class label. Every time a new rule is learned (created) the tuples that are covered by the rule are removed. The process continues until all tuples are covered or the termination condition is met. Basically we would like the rule to cover all or many tuples for the given class label and none from other classes since there could be many rules for a single class. The outline of the sequential covering algorithm is given below.

   Algorithm: Sequential Covering

   Input: D, a data set class labeled tuples

   Att_vals:the set of all attributes and their possible values.

   Output: A set of IF-THEN rules Procedure:

   1. Rule-Set= {}; //initial set of rules learned is empty

   2. For each class c do

   3. repeat

   4. Rule=learn-one-rule (D,Att_vals,c)

   5. remove tuples covered by

      Rule from D;

   6. until terminating condition;

   7. Rule_Set=Rule_Set+Rule;

   8. endfor

   9. return Rule_Set;

      Learn-one-rule procedure finds the best rule for the current class for the given data set. It basically

      adopts Greedy depth -first strategy. Initially the rule is created with an empty test condition for a current class, then each time the procedure encounters new data tuple ( from the training data) it picks only those test conditions that improves the rule quality. Rule Quality is a measure of how accurate the rule classifies the given tuple to the current class. A rule is assessed by its coverage and accuracy

      to specific rule search is done to establish the rules.

      IF<empty condition> THEN Can_fly=yes

      Rule Coverage: Coverage is the number of data tuples a rule covers It can be defined as below:

      =

      where, R is the current rule, n covers is the number of data tuples that are covered by the rule and is the total number of tuples in the database.

      Rule Accuracy: Accuracy is a measure of the correctness of the predicted class label. It can be

      IF load=avg THEN

      Can_fly=yes

      IF landing condition=smooth AND load=less THEN Can_fly=yes

      IF landing condition=smooth

      THEN Can_fly=yes

      IF landing condition=smooth AND

      load=avg THEN Can_fly=yes

      IF load=less THEN

      Can_fly=yes

      IF landing condition=smooth AND load=less THEN Can_fly=yes

      defined as below:

      =

      where, R is the current rule, is the number of tuples that are correctly classified to their class label and is the nmber of tuples that are covered by the rule R.

      The idea is to aiming for high accuracy rather than high coverage.We initially start with an empty test condition for the class label, then we logically conjunct the test conditions to it. A general

      Figure 1. A general to specific rule search

      In our example, the test condition landing condition= smooth best improves the quality of the current (empty) rule R2. We add it as our test condition and the rule which becomes:

      R2: IF landing condition=smooth THEN

      can_fly=yes

      In the next iteration another test condition load=avg best improves the quality of the rule. After logically appending the test condition, the rule becomes:

      R2: IF landing condition=smooth AND load=med

      THEN can_fly=yes

      The rule R2 classify the data to the class label can_fly=yes only when the test conditions load and landing condition are satisfied.

3. Application

   The application has been developed on a Java platform with MySQL as Database Management System. The intended users of the system are administrator, project manager and technician. The Functional Requirements and respective users of the

   application are identified as shown in the table 1. User Management deals with creation of user credentials and registration process. Component Management deals with addition of new component details and updating the status of obsolete components. Parameter Mangement deals with adition, removal parameters into the application database. All the functional requirements are implemented. The Menu is formatted as shown in the Figure 2 based on the requirements outlined in Table 1.

   Table 1 Functional Requirements of the application

   Users	Functional Responsibilities
   System Administrator
   Project Manager
   Analyst
   	Home Page

   • User Management

   • Component Management

   • Parameter Management

   • Allocation of Component s Parameters

   • Setting the bounds for the Parameters

   • Assessment of the Component Health

   • Preparing the Structural Health reports and MIS reports

   Unauthorized

   Registration

   Login Page

   Assess

   Allocate Component

   Component Management

   PM Home

   Admin Home

   Analyst Home

   Parameter Management

   Report

   Logout

   Figure 2. Menu Format

   The GUI of the project manger and the technician are given in the figure3 and figure 4 respectively.

   Figure 3 GUI for allocation of component

   Figure 4 GUI for Component Management

   Data acquisition will be done by sensors placed on the aircrafts components. The number and type of sensors to be deployed depends on the component being assessed. For ex, for assessing the health of a rudder, the parameters of interest would be wind velocity, stress on the wing, etc. Load on the

   aircraft may not be a crucial parameter for such scenario. Data thus recorded are used for assessment post flight. For convenience, we have assumed data are clean and free from inconsistencies. Upper and Lower bounds for the parameters are obtained as per the standards given by the components manufacturers. The data are randomly picked up from the database. Each of the picked data is compared with the bounds and any values that are outside of the range are reported. All the Data that falls between the lower and upper bounds are considered safe where as those that fall outside of the range are considered as unsafe. The farther the deviation from the safe zone the more abnormal the data are. The notify zone, alarm zone and auto zone are the unsafe zones.Data are assessed by plotting and counting their occurances. Each occurances in the unsafe zones are individually counted and based on their count the component is given a health status. The landing component of the aircraft has been assessed by considering three parameters, namely vibration experienced during the landing, climatic temperature and the load on the aircraft in the application. Half of the data from the database is used as training data sets and the rest as the test data. Rule Quality is determined by the rule accuracy and Rule Coverage as explained in the section 2.

   Consider the sample training data sets from the flight database. The first column in the table uniquely identifies each row in the table. The second, third and fourth columns stores the summarized data for the parameters temp, vibration and load on a landing component. The last column Class:can_fly stores the prediction for the class label can_fly. A total of 18 data sets are represented in the table 2. Consider the rule R2 from the section 2. It covers two tuples (RID 11 and 17) of the18 tuples. It can correctly classify both tuples. Therefore, Rule coverage and accuracy for the rule R2 are calculated.

   coverage(R2) = 2/18=11.11%

   where = 2 and =18

   accuracy(R2) = 2/2=100%

   where =2 and =2

   Table 2 Sample training data sets.

   RID	Route Temperat ure	Landing Conditi on	Load	Class: can_fly
   1	high	rough	heavy	no
   2	high	rough	average	no
   3	high	rough	light	no
   4	high	smooth	heavy	no
   5	high	smooth	average	no
   6	high	smooth	light	yes
   7	normal	rough	heavy	no
   8	normal	rough	average	no
   9	normal	rough	light	yes
   10	normal	smooth	heavy	yes
   11	normal	smooth	average	yes
   12	normal	smooth	light	yes
   13	low	rough	heavy	no
   14	low	rough	average	no
   15	low	rough	light	no
   16	low	smooth	heavy	no
   17	low	smooth	average	yes
   18	low	smooth	light	yes

   There is no known solution (application) that assesses the post flight structural health of a component using Rule based Classification. The application developed has achieved all functional requirements listed in Table 1. Rule Quality is determined by Rule Accuracy rather than Rule Coverage the application has achieved 100% Rule Accuracy as explained above.

4. Conclusion

In this paper, we have studied Rule based classification of Spatial Data. Rule based Classification uses decision trees and sequential covering algorithms for rule extraction. We have used sequntial covering algorithms to extract rules from the training data sets and classified the test data sets. Rule Extraction using decision trees becomes complex when the database is large. Branch repetition and pruning has to be performed. We can reduce the time and space complexities by using sequential covering algorithm. The application was developed by considering the landing system in the aircraft. The various parameters that are considered to assess the health of the landing system are three, namely vibration experienced by the landing gear

while landing, load on the aircraft, and the temperature exposure. In future we will extend this study by considering multiple dimensions, keeping in view the overall performance of the application.

REFERENCES

1. UML for Java programmers, Martin, Low Price Edition

2. RAPID JAVA application development using JBUILDER#, Y.Daniel Liang, Prentice Hall

3. JSP2.0 The Complete Reference, Hanna, TMH Edition

4. Software Requirements & Specifications a lexicon of practice, principles & prejudices, Micheal Jackson, Addison-Wesley

5. The Web Programming Desktop Reference 6 in 1, Micheal Afergan, Rick Darnell, Brian Farear, Russ Jacobs, David Medinets, Robert Muller, Micheal O Foghtu, PHI