Hui-Xin He，Ning Li，Geng-Feng Zheng，Xu-Zhou Lin，Da-Ren Yu

(1.School of Astronautics，Harbin Institute of Technology，Harbin 150001，China;2.National Institutes for Food and Drug Control，Beijing 100050，China;

3.Fujian Special Equipment Ispection and Research Institute，F(xiàn)uzhou 350008，China)

1 Introduction

With the development of information technology，the amount of data stored grows exponentially in reality.Just by now，humankind has been able to store 2 × 1020optimally compressed bytes［1］. T he data storage is to extract and apply the information from it.It is the main purpose of data mining that relies on the massive data analysis and mining，to find out the useful rules which human can understand and apply［2］.Anomaly Detection isan importanttopic ofdata mining’s research［3－5］.It refers to the problem of finding patterns in data which do not conform to the expected behavior.But the data collected world-wide and drowning the researcher in the deluge of no ending，while anomalous events occur in a relatively infrequent proportion.

A gas turbine is a new kind of power equipment.It has a relatively high thermal effect which has made it be widely applied.But due to its complexity，special supporting conditions and many other nonlinear factors，a gas turbine may be frequent out of order at work.If the failures can not be found or processed，the mechanical equipment＇s reliability will be reduced，even lose the production function，causing great economic losses or casualties.The application of gas turbines＇condition monitoring and faultdiagnosis technology can largely improve the safety and reliability of the unit operation，greatly reduce maintenance and repair costs.Therefore，the research of the gas turbine fault diagnosis methods has an important theoretical significance and high application value［6－7］.

In the method based on knowledge，the traditional expert system is based on the second order deduction system［8］.What＇s more，expert system exploits the existing knowledge only，fails to discover unknown but useful implicit knowledge，and can not give advice to unknown problems.However，the data of gas turbines collected in reality is an incomplete，noise，fuzzy，random big data sample［9］.There does not exist known，determinate fault judging rules，so data mining technology with lower conditional assumption dependence is more suitable.

This article investigates the anomaly detection in a novel situation in which without any history data and knowledge，and construct a method to deal with it by multi-classify detection mechanism. Finally， we classify and analyze the operation and maintenance data of a certain Siemens typhoon unit in recent years.Combining the evaluation with actual information，we verify the validity of this method.

2 Start-Up Process of Gas Turbines and Fault Samples Information

In the start-up process，a gas turbine changes from the cold into hot state and the thermal stress becomes a serious problem.Thermal shock of high temperature parts will influence the operating life;Compressor may have surge problem-high pressure unit＇s surge problem will be more prominent and need to take bleed gas measure.It＇s vital to ensure the success rate of start-up to increase the operation reliability and availability.In the start-up process of the unit，preventing overheating fault is not only necessary to the turbine operation，but also the key to ensure its smooth start-up.In addition，shorter start time meets the unit＇s economy and the timely outside load demand，aswellasa vital measurement index of starting performance.Therefore，it is necessary to analyze the problem and impact factors during the start-up process of the gas turbine.

To different units，start-up failure is not the same，but from the aspect of a gas turbine itself，we generally have the following aspects:

It fails to control the temperature rise rule in startup process.When the injection quantity is controlled improperly，the starting and accelerating process line will be close to or into the surge area，leading the turbine inlet temperature too high and the failure of speed rise.Some gas turbines designed and built by China have the start failure caused by high temperature and"hot pensile"after ignition.Such fault happens more frequently in summer.

There exist failure of fuel system and failure of ignition system.For example，the nozzle wearing，the atomization quality will deteriorate;Air existing in the fuel systems，fuel supply ripple phenomenon will occur;Igniter having carbon deposition，it can＇t form an ignition torch normally.

Part of flow passage of compressor is polluted.This will make compressor＇s efficiency reduce，the characteristic curve go bad，energy consumption during start-up increase or start machine underpowered then boot failure.Partofflow passage ofturbine is deposited.This willlead the turbine resistance increasing.Thus，the operation line during the start-up process of units will get close to the compressor surge border，and even into surge condition and make the unit fail to start."Hot suspension"is a possible fault in a gas turbine ＇s start-up process.It happens after tripping.That is，after the starting engine tripping，the unit halts the rise of speed and has an abnormal voice.If we continue to increase the amount of fuel，T3increases.Instead of rising，however，speed tends to have the downward trends，which eventually lead to the failure of start.The main reason of hot suspension is that the start-up process line is too close to the compressor surge boundary.

After deep analysis of the fault mechanism in start-up process startup，directing at some Siemens typhoon units output start system with an output power of 4900 kW and with a rated speed of 6228 RPM(turn/cent)，we extractand countthe fault information of the year 1998－2009.In this typhoon units start system，main faults include improper control of controller laws(14%)，gas system failure(14%)，fuel system failure(14%)，ignition head failure(11%)，and the percentage of other faults include oil leakage or starting motor drifting of hydraulic system，oil to gas failure， lubricating system fault， joint leakage，adjustable guide vane failure，drain valve failure，nozzle badness，start module damage，etc.is lower.

Thus，we select the main faults in gas turbine start system，including improper control of controller laws，gas system failure，fuel system failure，ignition head failure and so on(as shown in Fig.1).They take a large percent of total failure，fairly representative and have more practical research value.The fault data is analyzed.

Fig.1 Fault of gas turbine(unit:1 time)

3 Anomaly Detection Based on Fusion Detector

The amount of data in real applications is very large.Itwillcause additionalcomputing burden diagnosis the whole data set directly.We put forward the pre-judging method to data set before anomaly detection to screen sample that need diagnosis，and do fault classification afterwards.

Anomaly detection mines and labels abnormal data samples in no-label data set.In the current research，the extensive abnormalform isdefined as:the frequency and probability of the anomaly samples in total sample is extremely low;Anomaly samples and normal sample is completely different.In the industrial environment，especially in the data recorded in long normal operation of the gas turbine work，the possible fault information matches perfectly the conditions［3］.

In the research ofanomaly detection，the classification is used to learn a model(classifier)from a set of labeled data instances(as training data set)and then，using the learned model classify a test instance into one of the classes(testing).Anomaly detection techniques based on classification operate in a two-phase fashion.The training phase learns a classifier using the available labeled training data.The testing phase classifies a test instance as normal or anomalous，using the classifier.

As shown in Fig.2，classification based anomaly detection techniques operate under the following general assumption that a classifier which can distinguish between normal and anomalous classes can be learned in the given feature space.One-class classification based anomaly detection techniques assume that all training instances have only one class label.Such techniques learn a discriminative boundary around the normal instances using a one-class classification algorithm，for example，one-class SVMs，one-class KernelFisher Discriminants.Any test instance which does notfallwithin the learned boundary is declared as anomalous.

Because of the shortcomings of one-class classification used in anomaly detection，we construct a detection method fusion of multiple classifiers.As shown in Fig.3，one-class classification always loses into alarm as false，and higher recall with lower prediction，while the real alarm samples are similar to the normal samples.Tradition method only concerns about the wide range of data separated，but the highvalue anomaly is embedded in the normal，so it is hard to judge them.Every classifier has its applicability，while multi-classify fusion，it excludes individual blind spot.

As construct，this method selects all anomaly samples as part of the training data，and selects some normal data as supplement.The step of building a“weak”biased classify，in which the normal training data is selected as below，the select probability of sample xiin the T+1 turn is:

While MT(xi)is the T turn classify judge xito normal or anomaly，and yiis the real class of xi，and βTas the adjustment of samples to update the weight coefficients，which is calculated as

where εtis on the behalf of the calculated residuals，which is described as

To judge a sample as normal or anomaly， we use majority voting function，as below:

Fig.2 Multi-detector split the normal data effective

Fig.3 Mechanism of multi-detector

4 Experiment

The gas turbine unit equipment will monitor many parameters such as temperature，pressure，vibration quantity，noise，speed as well as the motor speed，power，current and so on in the start-up process.The choice of parameters is based on the following two principles:parameters can character the whole system completely;Redundancy between the parameters is small to analyze the start system＇s fault of a gas turbine.In this experiment，we choose gas generator speed，average temperature of exhaust and the power，the speed，current，and the power of the starter motor as the feature in a whole startup process，and so as a sample data.

A single sample is shown in Table 1 as an example.

Table 1 Single sample

In orderto have unified processing，we standardize the data set.

where x，y ∈ Rn，xmin=min(x)，xmax=max(x).

After standardization，the original data will be neat to［0，1］.

In the distance measure，we use the Euclidean distance

The main selected faults in a gas turbine start system in the experiment include improper control of controller laws A，gas system failure B，fuel system failure C，ignition head failure D.Normal sample identified for N =2000.We experiment anomaly detection process with LOF algorithm［10］， support vectors machines［11］，13-dectors fusion and 20-dectors fusion method，adjust the recall rate up to 0.93，the precision of detection is shown in Fig.4.

From the experiment results，choosing multidetector fusion has great influence on the final analysis results.But this method has a good performance in the recall rate，a high applicability in the capture of abnormaland faultinformation in the exception handling.This will be an ideal possible use in the fault signal process because the abnormal information is grabbed mainly as a preliminary information，followed by further failure analysis.Without abandoning real and possible breakdown，it has a good practical significance in the diagnosis of automation system.Detection accuracy can also achieve good results with appropriate use fusion method.

Fig.4 Experiment Results

5 Conclusions

The gas turbine fault diagnosis has important theoretical significance and high value in practice.In this paper，we study the failure of the gas turbine start system and impact factors systematically，and introduce the anomaly detection method to pre-test operational data based on multi-detector fusion.After classifying and analyzing the operation and maintenance data of a certain actual unit，the evaluation of actual data verifies the validity of this method.

［1］Hilbert M，Lopez P.The world＇s technological capacity to store，communicate，and compute information.Science，2011，332(10):60－65.

［2］Han J，Kamber M，Pei J.Data Mining:Concepts and Techniques. SanFrancisco:MorganKaufmannPub，2011.45－51.

［3］Chandola V，Banerjee A，Kumar V.Anomaly detection a survey.ACM Computing Surveys，2009，41(3):1511 －1558.

［4］Hodge V，Austin J.A survey of outlier detection methodologies.Artificial Intelligence Review，2004，22(2):85－126.

［5］Lazarevic A，Ertoz L，Kumar V，et al.A comparative study of anomaly detection schemes in net-work intrusion detection.Proceedings of SIAM International Conference on Data Mining.Quebec:SIAM，2003.25－36.

［6］LiY G. Performance-analysis-based gas turbine diagnostics:a review.Journal of Power and Energy，2002，216(5):363－377.

［7］Edwards S，Lees A W，F(xiàn)riswel M I.Fault diagnosis of rotating machinery.Shock and Vibration，1998，30(1):4－13.

［8］Depold H R，Gass F D.The application of expert systems and neuralnetworks to gas turbine prognostics and diagnostics.Journal of Engineering for Gas Turbines and Power，1999，121(4):607 －612.

［9］Rajeev V，Niranjan R，Ranjan G.Gas turbine diagnostics using a soft computing approach.Applied Mathematics and Computations，2006，172(2):1342 －1363.

［10］Breunig M M，Kriegel H P，Ng R T，et al.LOF:identifying density-based local outliers.Proceedings of the 2000 ACM SIGMOD International Conference on Management of Data，2000，29(2):93－104.

［11］Steinwart I，Hush D，Scovel C.A classification framework foranomaly detection.JournalofMachine Learning Research，2005，6(1):211－232.