A Concept of the Real-time Diagnostic System for Prototype Engines

Architecture and Algorithm

Vitaly Promyslov and Stanislav Masolkin

V. A. Trapeznikov Institute of Control Sciences, 65 Profsoyuznaya, Moscow 117997, Russia

Keywords: Modeling, Diagnostic, Real Time, SVM, Software Architecture.

Abstract: The paper summarizes the main ideas and methods used in a software design of the real time diagnostic

system for an advanced engines prototype test bed. The software architecture of the diagnostic systems is

built on a top of the multiprocessor computer system which allows affectively performs various tasks. The

SVM (support vector machine) algorithm is discussed from a point of view its real time implementation.

The simulation results are presented and discussed.

1 INTRODUCTION

Modeling offers a way to automation of the anomaly

diagnosis in the behavior of complex technical

systems during tests. This method transforms

knowledge into a model used for automatic detection

of abnormalities. By the behavior abnormality is

meant any change in the diagnostic information

arriving from the plant and regarded as significant

for the diagnostic system independently of its

causes. Analysis of the causes of abnormal behavior

lies outside the scope of the present paper. For

diagnosis of the abnormal behavior of the pilot

plants such as high-powered engines or experimental

machines, direct transformation of knowledge about

the tested plant into a physical model or expert

conclusions is not necessarily possible. This is due

to complexity and insufficient examination of the

tested plant or classification of information by the

designer for reasons of confidentiality. However,

importance of the timely detection of behavior

abnormalities is evident because it allows one to

minimize the losses of plant destruction and

simplify the post-emergency situation analysis. That

is why the designers of the diagnostic systems of

such plants make emphasis on the diagnosis of

abnormalities by the information-oriented method

which is aimed at constructing the anomaly

detection model relying on the data acquired in the

course of operation, rather than on the expertise. The

information-oriented algorithms for anomaly

detection, which are also known as those of outlier

detection, attempt to identify the portion of data

which differs somehow (is abnormal) for the given

data collection. The readers are referred to

(Chandola et al., 2009); (Markou and Singh,

2003a,); (Markou and Singh, 2003b) for review and

analysis of the abnormality detection algorithms. All

such algorithms need learning data consisting of a

set of the examples of anomalies and a set of

examples of normal (or nominal) data. Using the

learning data, the algorithm trains the model which

is used to verify the hypothesis of nominality and

abnormality of the tested data.

In our opinion, the need for diagnostic system

(DS) to operate at the rate of plant events, that is, in

real time and with the time of response smaller than

the characteristic time of development of the

unfavorable situation, is an essential restriction. For

the engines under examination, this time is about

tens of ms. The resource and time constraints

appreciably restrict the set of realizable algorithms,

as well as increase the role of the technical solutions

used such as system architecture or software and

hardware facilities used to realize the diagnostic

system.

We considered the support vector machines

(Vapnik, 1995) as an algorithm to construct the

information-oriented model. The support vector

algorithm is realized with the use of the free library

(SVM) (Chang and Lin, 1997) that was already used

to solve a similar problem of diagnosis of complex

experimental plants (Iverson et al., 2009),

(Schwabacher et al., 2009), (Jeong et al., 2012),

(Kang et al., 2012).

360

Promyslov V. and Masolkin S..

A Concept of the Real-time Diagnostic System for Prototype Engines - Architecture and Algorithm.

DOI: 10.5220/0004426703600365

In Proceedings of the 10th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2013), pages 360-365

ISBN: 978-989-8565-70-9

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

Since by the time of tests the amount of data for

classification of the abnormal plant modes will be

insufficient, consideration was given to the single-

class algorithm of anomaly detection which learns

using a single data set of which majority or all are

assumed to be nominal. It learns the nominal data

model that may be used to signal anomalies if the

new data do not coincide with the model.

The selected algorithm and diagnostic system

architecture were verified using the simulated data

reflecting the typical forecasted real data obtained

from the plant. Consideration was given to the

system architecture enabling efficient real-time

diagnosis of the engine.

2 SOFTWARE ARCHITECTURE

2.1 Introduction

The proposed diagnostic system architecture

implies division of the algorithms into two classes of

algorithms requiring rigid real-time (RTT) and

calculation algorithms which do not require

immediate response to the inputs (STT). Table 1

presents classification of the diagnostic software

(DSW) and describes functionality of the tasks.

The fundamental difference between tasks (H)

and (S) lies only in the priority and planner. They

are realized as a pool of algorithms executed within

an integrated process. Reasoning from the time

requirements and hardware configuration, there may

be more than one task pool (H and S) with different

sets of independent algorithm. (I)–(INI), this part of

DSW is executed at the stage of activation and

performs the functions of determination of

configuration (distribution of the computer

resources) and generation of the necessary number

of tasks (H) and (S). Connection of inputs and

outputs of the algorithm with nonblocking queues.

The inputs and outputs of the algorithm are

connected by the nonblocking queues at the stage of

initialization; the execution cycles of the rigid real-

time algorithms are generated at the stage of

initialization. DAS – digital acquisition system.

The system architecture realized on a

multiprocessor system is shown in Fig. 1 depicting

allocation of the tasks to the multiprocessor system

processors, the affinity interface being used for

binding.

2.2 Description of the Algorithm

of Task Running

The diagnostic system tasks execute the following

stepwise functions:

2.2.1 Task (INI)

1. Inventory of the RTT algorithms.

a) Determination of the number of RTTs - N.

b) Initialization of the internal structures of the

algorithms.

c) Assignment of output identifier (for task (M)).

Table 1: Program Structure of DS

Task Type Function and description

(INI) Initialization STT Responsible for DSW initialization, in particular, binds the algorithms to particular

tasks and supports the real-time requirements (dynamic balancing).

(I) Reception and

multiplexing of the inputs

RTT Supports reception of inputs and their allocation to the input ring buffers

(H) Task of rigid real-time

algorithms

RTT Supports execution of the parts of algorithms that must be executed in a strictly

allocated time. In the case of multiprocessor system, the number of task can be

more than one. Determined is from the results of verification of the system

architecture with regard for the computational complexity of the algorithms and

availability of resources.

(S) Task of soft real-time

algorithms

STT Accumulates all algorithms or their parts which are not rigidly timed.

In the case of multiprocessor system, the number of tasks can be more than one

Determined is from the results of verification of the system architecture with regar

for the computational complexity of the algorithms and availability of resources.

(M) Task of meta-

algorithms and output

RTT Supports generation of the outputs on the basis of the results of operation of the

algorithms of SТT and RTT. This tasks includes also generation and transmission

of the control signals.

(DI) Service diagnosis STT Outputs service and diagnostic information.



Tasks (I) and (М) may be combined in one process executed as a real-time task. Their algorithm which is assumed to be event-triggered is

described in what follows.

AConceptoftheReal-timeDiagnosticSystemforPrototypeEngines-ArchitectureandAlgorithm

361

DAS

I/OIndustrial

Network

RTT

CPU1

CPU

CPU2‐N/H

Internal

realtime

bus



Internal

realtime

bus

STT

CPUM+1/DI,INI

CPU

CPUN‐M

LAN

Figure 1: Architecture and task allocation on the multiprocessor computer. CPU1,2-N,M are individual processors or a set

of processors executing the tasks.

2. Determination of the input points for the soft real-

time algorithms (STT).

a) Determination of the list of connected input

signals for each STT.

b) Assignment of the signal receiver identifier.

3. Determination of the input points for RTT.

a) Determination of the list of connected input

signals for each RTT.

b) Assignment of the signal receiver identifier.

c) Determination of the project deadline for each

RTT.

4. Determination of the input points for the meta-

algorithm.

a) Determination of the list of connected input

signals.

b) Assignment of the signal receiver identifier.

5. Algorithm balancing with respect to tasks (H and

S) and processors.

a) Determination of the number of processor

kernels - Np.

b) Determination of the sequences of test signals

for each RTT.

c) Activation of each RTT on the test sequence of

input signals for determination of the maximal

delay per iteration.

d) Determination of the required processor

resources for each RTT on the basis of the

established maximal delay and project deadline.

e) RTT layout with respect to tasks (H) on the

basis of the required processor resources.

f) Determination of N(H), the number of tasks of

types (H) and N(H) of the sets of the points of

input in RTT for each task .

g) N(H) should not exceed Np-2 to satisfy the

requirements for time characteristics.

6. The process with task (S) is generated, its affinity

to Np - (N(H) + 1) processors is established, the

points of input for all STT are given to it.

7. N(H) processes with tasks (H) are generated,

affinity to one free processor is established for

each, and the corresponding sets of the points of

input to RTT are passed.

8. Affinity to one free processor is established for

the current process.

9. Control is passed to task (I).

2.2.2 Task (I)

The task is waiting for input data. After getting

them, the signals are placed into the input ring

buffers. The data readiness signal is sent to all tasks

(H) and (S). Control is transferred to task (M).

2.2.3 Task (H)

The task is waiting for the data readiness signal.

After getting the input data readiness signal, the

task executes RTTs successively fixed to it.

(Individual RTTs place the outputs into the output

ring buffers.)

Upon cycle completion, passage is made to the

mode of waiting for the input data (duration of one

cycle is at most 5 ms.).

2.2.4 Task (S)

After activation individual flows are generated to

execute STT.

Each algorithm operates using its own program

and reads data from the input buffer.

The data readiness signal is processed by each

STT using the internal algorithm. (Individual STTs

place the output signals in the output ring buffers.)

2.2.5 Task (M)

Selection of meta-algorithms: each meta-algorithm

ICINCO2013-10thInternationalConferenceonInformaticsinControl,AutomationandRobotics

362

checks availability of connected signals in the output

ring buffers; if any, the algorithm is executed and,

possibly, the output is generated and then placed in

the output ring buffers.

Transmission of signals kept in the output ring

buffers.

2.2.6 Task (DI)

The task periodically diagnoses DSW with low

priority. The diagnostic results are archived in the

common system database and may be displayed on

the service monitor.

2.3 One-class Support-Vector Machine

(SVM)

The one-class SVMs attempt to describe with the use

of the model the set of normal learning data so as to

enable the resulting model to discriminate between

the normal and abnormal data. It is assumed that the

learning data may contain some anomalies. The one-

class SVM first transforms the learning data from the

initial data space into a larger one or, possibly,

infinite-dimensional space of attributes and then

seeks there a linear model (hyperplane), which

allows actually all normal data to stay on the one

side separately from the learning data deviating from

the norm, if any.

2.3.1 Formulation of the Task

and Description of the Algorithm

Each data object is represented as a vector (point) in

the

-dimensional space (sequence of

numbers).

Each of these points belong only to one class. We

assume that the points are given by

)},).....(,(),,{(

11 nnii

txtxtx

where the label

assumes value 1 or −1 depending

on whether the point

belongs to the class or not.

Each

is a

-dimensional vector.

Needed is to generate the function

TXF : (classifier) assigning the class t to an

arbitrary object

The predictions of SVM rely on the functions like

(;)sign ( )

















xx,x







train i i



 x

, where

Rx

{ 1,1}, 1,...,

tiN





). At that, the weights of

the objects

{}





are nonnegative and upper-

bounded by the constant

0C  . The function

()K x,

also has parameters on which depends

essentially the performance of the resulting

classifiers.

The kernel of basic radial Gaussian functions

returning the distance between examples x

and x

















exp),(



xxK

(1)

was considered as the kernel. The one-class SVMs

require that the user establishes the maximal

number of points in the learning sample that can be

outliers. We take in our experiments that it is

0.0025 because this provides acceptable results

(Schwabacher et al., 2009). The value  of the

SVM kernel is selected using the procedure of

(Runarsson et al., 2003) which takes the least value

such that the assigned part 0.0025 of the learning

points is classified as abnormal.

The results of modeling are considered below,

the data used to verify the algorithm are similar to

the actual data; as a preliminary the algorithm was

tested using the normal (Gaussian) model.

Modeling was carried out on a computer with four

Intel Core 3.07GHz processors.

2.3.2 Simulation

The data for altogether 16 simulated parameters

were considered. A sample of data knowingly

belonging to the normal and abnormal modes was

generated for learning.

Learning sample: data duration 6 s., sampling

step 5 ms.

Figure 2: Location of points on the hyperplane for

example (a).

AConceptoftheReal-timeDiagnosticSystemforPrototypeEngines-ArchitectureandAlgorithm

363

Consideration was given to two examples:

a. Test sample corresponding to the abnormal

mode, data for 0.15 s., sampling step 5 ms.

b. Test sample for the nominal operation mode,

data for 0.2 s., sampling step 5 ms.

2.3.3 Analysis, Resource Consumption

and Speed of the Algorithm

Classification precision for example a. was about 80

%, 2 points of 11 were classified as normal. One can

see on the whole the increase in the distinction of the

tested data from the learning sample in the course of

abnormality development.



Figure 3: Location of points on the hyperplane for example

(b).

For example b., classification precision was about

50%, 16 of 38 were classified as abnormal. However,

on the whole one can see the correspondence

between the tested data and the learning sample. The

relative low percent of correct classification in

example b. is attributed to an insufficient volume of

the learning data in the example, increased volume of

the learning sample from 6 s. to 1 min. improves

precision up to 70%.

The probability of the false detection of the

abnormal situation may be considerable reduced if

use several consecutive detections of the abnormal

situation as a sign that situation is really changed and

deviated from a nominal behavior. The pitfall of the

such solution will be reducing the reaction time of

the DS.

The algorithm realizing the SVM method

performs two distinct tasks:

1. modeling of calculation and learning,

2. state classification.

The first task belongs to the STT tasks and

required up 200 ms. on the simulated data samples.

The second task requires much less resources and

may be run in real time (calculation required about

1 ms.).

3 CONCLUSIONS

The paper discusses the architecture and

algorithmic aspects of the design the fault diagnosis

tested for prototype engines. The distributed

architecture of the test bed allows affectively

realizing the complex SVM fault diagnoses

algorithm with reasonable time response. The SVM

algorithm demonstrated its practicability for

preliminary diagnosis of abnormalities of the

objects on the test bed. It was possible to diagnose

an abnormality already at the initial stage, which

would enable reduction in the outcomes of the

abnormality the tested object. However, extended

studies on a larger data volume of real data are

required for confident use of the method.

Estimation of the efficiency of the SVM algorithm

for detection of abnormalities as applied to real data

is a challenge because the number of anomalies in

the data usually is not known. One of the

approaches to estimation of algorithm efficiency

lies in estimating it on the artificial data where the

number of anomalies is known.

REFERENCES

Vapnik, V., 1995. The Nature of Statistical Learning

Theory, Springer-Verlag, New-York.

Iverson, D., Martin, R., Schwabacher, M., Spirkovska, L,

Mackey, R., Castle, P., et al., 2009-1909. General

Purpose Data- Driven System Monitoring for Space

Operations, AIAA Infotech@Aerospace Conference,

AIAA, Washington, DC, 2009, AIAA paper.

Schwabacher, M, Nikunj, Oza, Matthews, B., 2009.

Unsupervised Anomaly Detection for Liquid-Fueled

Rocket Propulsion Health Monitoring. Journal of

aerospace computing, information, and

communication. Vol. 6, July.

Jeong, S, Kang, I., Jeong, M. K., Kong, D., 2012. A New

Feature Selection Method for One-Class

Classification Problems, IEEE Transactions on

Systems, Man, Cybernetics, Part C, 42, pp. 1500-

1509.

Kang, I. Jeong M. K., and Kong, D., 2012. A

Differentiated One-Class Classification Method with

Applications to Intrusion Detection, Expert Systems

with Applications, 39, pp. 3899-3905.

Tax, D. M. J., Duin, R. P. W., 1999. Support Vector

Domain Description, Pattern Recognition Letters,

Vol. 20, No. 1113, pp. 1191–1199.

ICINCO2013-10thInternationalConferenceonInformaticsinControl,AutomationandRobotics

364

Chandola, V., Banerjee, A., Kumar, V., 2009. Anomaly

Detection: A Survey, (to appear), ACM Computing

Surveys.

Markou, M., Singh, S., 2003. Novelty Detection: A Review.

Part 1: Statistical Approaches, [Online], Signal

Processing, Vol. 83, pp. 2481–2497. http://dx.doi.org/

10.1016/j.sigpro.2003.07.018 (Accessed: 6 Јune 2013).

Markou, M., Singh, S., 2003. Novelty Detection: A

Review. Part 2, Neural Network Based Approaches,

Signal Processing, Vol. 83, pp. 2499–2521.

Chang, C., Lin, C., 2007. LIBSVM: a Library for Support

Vector Machines, (Online), http://www.csie.ntu.edu.tw/

~cjlin/papers/libsvm.pdf (Accessed: 6 Јune 2013).

Runarsson, R. T., Unnthorsson, R., Johnson, T. M., 2003.

Model Selection in One Class Nu-SVMS using RBF

Kernels, 16-th Conference on Condition Monitoring

and Diagnostic Engineering Management.

AConceptoftheReal-timeDiagnosticSystemforPrototypeEngines-ArchitectureandAlgorithm

365