QG,(((,QWHUQDWLRQDO&RQIHUHQFHRQ&RPSXWDWLRQDO,QWHOOLJHQFHDQG$SSOLFDWLRQV

Prediction and Analysis of Aircraft Failure Rate Based on SARIMA Model

Yanming Yang, Haiyan Zheng and Ruili Zhang

Qingdao Campus
Naval Aeronautical University
Qingdao 266041, China
e-mail: [email protected]

Abstract—A large number of aviation equipment maintenance model can be used when the time series is stationary, but the
data exhibit seasonal behavior, such as aircraft failure rate. ARIMA model does not have that limitation [1]. The time
Consequently, seasonal forecasting problems are of series generally present ascending or descending trends.
considerable importance in aviation maintenance support. Such series are non-stationary. Non-stationary series can be
Aircraft failure rate is an important parameter of aviation modeled by allowing differencing the data series to
equipment RMS (Reliability-Maintainability-Supportability). stationary. The letter “I” (integrated) in ARIMA indicated
It is indispensable to scientifically predict the aircraft failure that the modeling time series has been transformed into a
rate and to make scientific decisions on aviation maintenance stationary time series. The time series often contain cyclic
to improve maintenance support capability. This paper
features (seasonal effects), to which seasonal differencing is
proposes a seasonal ARIMA (SARIMA) model to solve the
often used to remove the seasonal effects. These kinds of
problem of aircraft failure rate forecasting. Then the
mathematic model and modeling process of the SARIMA are models is known as SARIMA models. Thus, ARIMA
introduced in detail. The application of SARIMA model in models have two general forms: non-seasonal ARIMA and
forecasting the aircraft failure rate is analyzed by examples. multiplicative seasonal ARIMA (SARIMA) [2].
SARIMA (0, 1, 1) (0, 1, 1)12 model was selected as the most A. ARIMA Model (Nonseasonal ARIMA Model)
suitable model to forecast of aircraft failure rate. And the
forecasting results were analyzed and compared. The results Non-seasonal ARIMA models are generally denoted
demonstrate that the SARIMA model is feasible and effective ARIMA( p, d , q) [3] , where
for the prediction of aircraft failure rate. p is the order (number of time lags) of the autoregressive
part.
Keywords- SARIMA model; aircraft failure rate; prediction; d is the order of the differencing (the number of times the
time series analysis
data have had past values subtracted, rarely should >2 be
needed).
I. INTRODUCTION q is the order of the moving-average process.
In the management of aviation maintenance has more Given a dependent time series ^ X t :1 d t d n` ,
prediction problem, the technical state of the equipment, mathematically the ARIMA model is written as
aviation equipment readiness, spare parts consumption, the
aircraft failure rate to make scientific prediction, to improve
aviation maintenance work predictability, countermeasures T ( B)
’d X t H   
and scientific, and improve aviation equipment readiness and I ( B) t
wartime utilization level plays a very important role. The
aircraft failure rate is essential parameter for aviation
equipment maintenance engineering. Often, aircraft failure where
rate data display behavior that is seasonal. Seasonality is B is the backshift operator:
defined to be the tendency of time-series data to exhibit
behavior that repeats itself every S periods. The term season BX t X t 1   
is used to represent the period of time before behavior begins
to repeat itself. S is therefore the season length in periods. I ( B) is the autoregressive operator, represented as a
The SARIMA (Seasonal Autoregressive Integrated Moving
polynomial in the back shift operator:
Average) method is an effective method to solve this kind of
problem. In this paper, we use SARIMA model to forecast
the aircraft failure rate. I ( B) 1  I1B ˜˜˜  I p B p   
II. THE ARIMA MODELS
T ( B) is the moving-average operator, represented as a
The ARIMA model is one of the most popular models for
polynomial in the back shift operator:
time series forecasting. This model has been originated from
the autoregressive model (AR), the moving average model
(MA) and the combinations of the AR and MA. The ARMA  T ( B) 1  T1B ˜˜˜  Tq Bq   

978-1-5386-2030-4/17/$31.00 ©2017 IEEE 

 Ht is the independent disturbance, also called the random time series has a certain trend, need differential processing of
the original series; if the time series is heteroscedastic, you
error. It assumed to be independently and identically must first log according to the logarithmic transformation.
distributed with a mean of zero and a constant variance of
V 2 (white noise). The roots of I ( X ) 0 and T ( X ) 0 B. Model Identification
should all lie outside the unit circle. It was suggested by Box The model identification is to determine the order of p, q,
and Jenkins [4] that at least 50 or preferably 100 P and Q of the corresponding seasonal ARIMA model. The
observations should be used for the SARIMA model [5]. model identification is performed by calculating the
autocorrelation function (ACF) and the partial
’d describes differencing operation to data series to autocorrelation function (PACF) of the time series after
make the data series stationary, and d is the number of preprocessing. In general, we can estimate the possible
differencing. values of model order p, q, P and Q by observing the
B. SARIMA Model (Seasonal ARIMA Model) correlation graph, and then determine the most appropriate
model order by Akakine information criterion (AIC), Bayes
Seasonal ARIMA models are usually denoted information criterion (BIC) and so on [8][9]. The criterion of
SARIMA( p, d , q) u ( P, D, Q) S [6], where model order is shown in Table 1.
P is the order of the seasonal autoregressive (SAR) part.
D is the order of the seasonal differencing (rarely should TABLE I. THE ARMA MODEL ORDER CRITERION
D > 1 be needed).
ACF PACF Model
Q is the order of the seasonal moving-average (SMA)
process. Tailing p order truncation AR(p)
S is the length of the seasonal cycle. q order truncation Tailing MA(q)
Given a dependent time series ^ X t :1 d t d n` , Tailing Tailing ARMA(p, q)
mathematically the SARIMA model is written as
For the SARIMA model, besides using the low order
ARMA model to extract the short-term correlation effects,
T ( B)4( B ) S
the ARMA model with periodic steps is also used to extract
’ d ’ SD X t H   
I ( B )) ( B S ) t the seasonal effects.
C. Model Fitting
where After determining the order of the model, it is necessary
)( B S ) is the seasonal autoregressive operator, to estimate the parameters in the model. Methods used in this
represented as a polynomial in the back shift operator: step is the least square method, in the actual application, can
be obtained through the statistical software fitting value.
)( B S ) 1  )1B S ˜˜˜  ) P B PS    D. Model Diagnosis
In order to ensure the validity of the model, it is
necessary to test the residual white noise of the model. The
4( B S ) is the seasonal moving-average operator, residual LB statistic approximates the chi-square distribution
represented as a polynomial in the back shift operator: of the degree of freedom m. If the P value of the statistic is
greater than the significance level, it can be considered that
4( B S ) 1  41B S ˜˜˜  4Q BQS    the residual sequence is a random sequence, which indicates
that the information has been fully extracted by the model.
Otherwise, some information is not extracted and the model
III. SARIMA MODELING PROCESS needs to be improved. In order to ensure the simplicity of the
model, we must test the significance of the parameters and
To identify the correct model for prediction and eliminate the parameters which are not significantly zero.
explanation, the SARIMA modeling process consists of four
steps: stationarity test, model identification, model fitting
and model diagnosis [7]. IV. APPLICATION EXAMPLE ANALYSIS
A. Stationarity Test A. Problem Description
To test stationarity of the time series, that is, to determine In the period from 2005 to 2016, the failure rate of a
the size of d and D, the most intuitive identification method certain type of aircraft is shown in Table 2. Taking into
is the autocorrelation graph. If the autocorrelation coefficient account the seasonal influence of the aircraft failure rate, the
rapidly approaching zero, namely the self correlation SARIMA model is used to predict the aircraft failure rate of
coefficient is truncated, the time series is stationary. After the 12 months in 2017.
difference, we need to check whether the sequence has been
stabilized by ADF (Augmented Dickey-Fuller) test. If the


 TABLE II. AIRCRAFT FAILURE RATE STATISTICS DATA IN 2005~2016

Month 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
Jan 23.37 17.1 21.09 16.53 9.69 12.54 13.68 9.69 9.12 11.97 10.26 10.26
Feb 25.65 19.38 17.67 13.68 11.4 14.25 10.83 18.24 13.11 16.53 10.83 14.25
Mar 31.35 27.93 14.25 26.79 19.38 14.82 11.97 15.39 14.25 15.39 21.09 14.82
Apr 21.66 25.65 22.8 29.07 22.8 18.81 25.65 17.1 17.67 23.94 25.08 10.26
May 27.36 22.8 23.37 22.8 24.51 16.53 18.81 19.38 19.95 22.23 21.66 21.09
June 31.92 32.49 26.22 42.75 28.5 24.51 19.38 21.66 25.65 23.37 31.92 21.09
July 35.91 35.91 25.08 43.89 31.35 23.94 23.37 28.5 32.49 26.22 32.49 27.93
Aug 33.63 40.47 23.94 35.91 28.5 23.94 32.49 27.36 28.5 33.06 29.07 29.07
Sep 49.59 45.6 29.07 30.21 30.78 22.23 27.36 27.93 26.22 25.08 31.92 21.09
Oct 30.21 29.64 26.22 32.49 21.66 22.23 28.5 19.95 27.36 34.77 27.36 30.78
Nov 32.49 28.5 25.08 27.36 13.68 14.25 15.96 14.25 11.97 19.95 14.25 17.1
Dec 32.49 26.79 22.8 15.39 11.4 12.54 16.53 13.68 7.98 10.83 8.55 10.26

For the lag 12, draw its ACF and PACF plots as shown in
B. Problem Analysis Figure 3 and Figure 4. As can be seen from Figure 3 and
The time series plot of aircraft failure rate and the Figure 4, there is a convex peak not only at 12 and 24, but
difference plot of lag 12 are plotted respectively, as shown in also at 1 and 13. Therefore, it cannot be regarded as a simple
Figure 1 and Figure 2. The Figure 1 shows that there are cycle change. ACF has a truncation at 1, PACF has a
obvious seasonal features (12 months period), and the whole trailing at 1,13, ... According to the criterion of model order
time series has a non-stationary trend with decreasing in Table 1, it can be seen that the difference in lag 12 should
averages. As can be seen from Figure 2, there is no obvious belong to the MA (1) model. Taking into account these
non-stationary trend. situations, this is a simple seasonal model SARIMA (0,1,1)
 and a model MA (1), that is, SARIMA(0,1,1) u (0,1,1)12 .

]OZN YOMTOLOIGTIKROSOZYLUXZNKG[ZUIUXXKRGZOUTY
 


,GOR[XK8GZK

'[ZUIUXXKRGZOUT







 




 


            
      
:OSK
2GM
Figure 1. Time series plot of aircraft failure rate in 2005~2016.
:OSK9KXOKY6RUZUL*OLLKXKTIK Figure 3. The autocorrelation function for difference 12.


]OZN YOMTOLOIGTIKROSOZYLUXZNKVGXZOGRG[ZUIUXXKRGZOUTY



6GXZOGR'[ZUIUXXKRGZOUT




*OLLKXKTIK


 











                 

:OSK 2GM

Figure 2. The difference time series plot of lag 12. Figure 4. The partial autocorrelation function for difference 12.


C. Predictive Results TABLE IV. THE FORECAST VALUES AND RANGES OF AIRCRAFT
FAILURE RATE OF EACH MONTH IN 2017
According to the above analysis, in this example we
employ the model SARIMA(0,1,1) u (0,1,1)12 to forecast the Forecast 95% Normal Confidence Interval
Month
Values Lower Upper
aircraft failure rate in 2017. Figure 5 shows the comparison
between the actual and the fitted values of the aircraft failure Jan 9.373186 0.502873 18.2435
rate. Figure 6 shows the comparison between the actual and Feb 11.71845 2.580081 20.85682
the forecasted values of the aircraft failure rate. Mar 14.7899 5.391117 24.18869

<GXOGHRK Apr 17.60787 7.95569 27.26004
'IZ[GR
,OZY
May 18.41667 8.517592 28.31576


June 23.38289 13.24291 33.52287

,GOR[XK8GZK

 July 26.97771 16.60243 37.353

Aug 27.12389 16.51852 37.72926

Sep 25.51119 14.68062 36.34176
Oct 25.1807 14.12952 36.23188

Nov 14.83935 3.571881 26.10683
 Dec 10.22169 -1.258 21.70138
          
:OSK
Figure 5. The fitted values of aircraft failure rate. D. Predictive Effect Analysis
Analyze the forecast results, first observe whether the

]OZNLUXKIGYZYGTJZNKOX IUTLOJKTIKROSOZY residual ACF and PACF are almost all zero. According to
Figure 7 and Figure 8, the residual ACF and PACF are

almost all falling into the boundary, and it can be considered
that the white noise state has been achieved, so the model has
,GOR[XK8GZK


better forecast effect.

 ]OZN YOMTOLOIGTIKROSOZYLUXZNKG[ZUIUXXKRGZOUTY




'[ZUIUXXKRGZOUT






             

:OSK

Figure 6. The forecast result of aircraft failure rate.


The parameter estimation results are shown in Table 3, 

and the final calculation model can be obtained according to 

the parameter estimation results:           

2GM
Figure 7. The ACF of residuals for aircraft failure rate.
’’12 X t H t  0.8296H t 12  0.7523(H t 1  0.8296H t 13 )   

]OZN YOMTOLOIGTIKROSOZYLUXZNKVGXZOGRG[ZUIUXXKRGZOUTY

TABLE III. THE PARAMETER ESTIMATION RESULTS
6GXZOGR'[ZUIUXXKRGZOUT




Type Coef SE Coef T P

MA 0.7523 0.0564 13.33 0.000 

SMA 0.8296 0.0617 13.45 0.000 



The forecast values and ranges of aircraft failure rate of 

each month in 2017 with SARIMA(0,1,1) u (0,1,1)12 are




presented in Table 4.           

2GM
Figure 8. The PACF of residuals for aircraft failure rate.


E. Residual Plots Analysis effects. This plot helps you to check the assumption that the
Use to examine the goodness of model fit in regression residuals are uncorrelated with each other.
and ANOVA (Analysis of Variance). Examining residual
plots helps you determine if the ordinary least squares V. CONCLUSIONS
assumptions are being met. If these assumptions are satisfied, The SARIMA model is used when the data exhibits both
then ordinary least squares regression will produce unbiased trend and seasonality. It will adapt more quickly to genuine
coefficient estimates with the minimum variance [10]. The changes in the time series, but it might also overreact to freak
residual plots for aircraft failure rate with multiplicative and time series value. The application example shows that the
additive seasonal models are showed in Figure 9. SARIMA model can make full use of historical data, and can
accurately forecast the aircraft failure rate by depicting the
periodic fluctuation. Moreover, the SARIMA model is
simple and effective, and it can predict the aircraft failure
rate in one cycle at a time. It is an effective short-term and
medium-term prediction method. At the same time, the
SARIMA model proposed in this paper not only applies to
the prediction of aircraft failure rate, but also to other
equipment indexes or parameters with time series
characteristics, such as consumption of air material, flight
safety accident rate and aviation equipment readiness rate,
which provides a scientific method and means for equipment
support forecast.
REFERENCES
[1] Narimani, Reza, and A. Narimani. "A New Hybrid Model for
Improvement of ARIMA by DEA." Decision Science Letters
1.2(2012):59-68.
[2] Han, Ping, et al. "Drought forecasting based on the remote sensing
data using ARIMA models." Mathematical & Computer Modelling
An International Journal 51.11(2010):1398-1403.
Figure 9. Residual plots for aircraft failure rate. [3] Babu, C. Narendra, and B. E. Reddy. "Selected Indian stock
predictions using a hybrid ARIMA-GARCH model." International
Conference on Advances in Electronics, Computers and
Communications IEEE, 2015:1-6.
1) Normal Probability Plot of residuals: The points in [4] G.P. Box, G.M. Jenkins, Time Series Analysis: Forecasting and
this plot should generally form a straight line if the residuals Control, Holden-Day, New Jersey, 1976.
are normally distributed. If the points on the plot depart [5] Tseng, Fang Mei, H. C. Yu, and G. H. Tzeng. "Combining neural
network model with seasonal time series ARIMA model."
from a straight line, the normality assumption may be Technological Forecasting & Social Change 69.1(2002):71-87.
invalid. [6] Hyndman, Rob J; Athanasopoulos, George. "8.9 Seasonal ARIMA
2) Histogram of the Residuals: An exploratory tool to models". Forecasting: principles and practice. oTexts. Retrieved 19
show general characteristics of the residuals including May 2015.
typical values, spread, and shape. A long tail on one side [7] Bhutto, Abdul Waheed, et al. "Forecasting the consumption of
gasoline in transport sector in pakistan based on ARIMA model."
may indicate a skewed distribution. If one or two bars are Environmental Progress & Sustainable Energy (2017).
far from the others, those points may be outliers. [8] Jarrett, Jeffrey E, and E. Kyper. "ARIMA Modeling With
3) Residuals Versus Fitted Values: This plot should Intervention to Forecast and Analyze Chinese Stock Prices."
show a random pattern of residuals on both sides of 0. If a International Journal of Engineering Business Management
3.3(2011):53-58.
point lies far from the majority of points, it may be an
[9] Babu, C. Narendra, and B. E. Reddy. "Selected Indian stock
outlier. There should not be any recognizable patterns in the predictions using a hybrid ARIMA-GARCH model." International
residual plot. For instance, if the spread of residual values Conference on Advances in Electronics, Computers and
tend to increase as the fitted values increase, then this may Communications IEEE, 2015:1-6.
[10] Yan-ming Yang, Hui Yu and Zhi Sun. "Aircraft failure rate
violate the constant variance assumption. forecasting method based on Holt-Winters seasonal model. " IEEE
4) Residuals Versus Order of Data: This is a plot of all 2nd International Conference on Cloud Computing and Big Data
residuals in the order that the data was collected and can be Analysis (ICCCBDA), 2017.
used to find non-random error, especially of time-related