A Neural Network Approach to Condition Based Maintenance: Case Study of Airport Ground Transportation Vehicles

Alice E. Smith Department of Industrial and Systems Engineering

207 Dunstan Hall Auburn University

Auburn, AL 36849 USA

David W. Coit Department of Industrial and Systems Engineering

96 Frelinghuysen Road Rutgers University

Piscataway, NJ 08854 USA

Yun-Chia Liang Department of Industrial Engineering

Yuan Ze University Taoyuan, Taiwan

Abstract This paper describes a joint industry/university collaboration to develop a prototype system to provide real time monitoring of an airport ground transportation vehicle with the objectives of improving availability and minimizing field failures by estimating the proper timing for condition-based maintenance. Hardware for the vehicle was designed, developed and tested to monitor door characteristics (voltage and current through the motor that opens and closes the doors and door movement time and position), to quickly predict degraded performance, and to anticipate failures. A combined statistical and neural network approach was implemented. The neural network “learns” the differences among door sets and can be tuned quite easily through this learning. Signals are processed in real time and combined with previous monitoring data to estimate, using the neural network, the condition of the door set relative to maintenance needs. The prototype system was installed on several vehicle door sets at the Pittsburgh International Airport and successfully tested for several months under simulated and actual operating conditions. Preliminary results indicate that improved operational reliability and availability can be achieved.

Keywords: Condition monitoring, Conditional maintenance, Predictive maintenance, Preventive maintenance, Neural network, Transportation

1. Introduction

Ground transportation people mover vehicles are found in every major airport in the world. These

vehicles have stringent availability and safety demands. Although they operate over short, fixed

routes, they are subject to nearly constant use, sometimes under adverse outdoors environments, with

frequent stops and a large volume of passengers. One of the world’s leading producer of these

systems, in cooperation with the University of Pittsburgh, and Auburn University, researched a proof

of concept for utilizing a condition-based maintenance approach for improving the operational

reliability and availability of the vehicle door systems. Door systems were monitored in real time

with respect to their current operational state and predictive models were developed to suggest when

maintenance actions are required. The benefits anticipated using condition-based maintenance over

the current scheduled maintenance approach are: 1) more cost effective maintenance because system

and/or components are maintained only where and when needed and 2) degradation-type failures and

downtime reduced to a minimum because systems and their components can be maintained during an

early phase of degradation, long before failure can occur.

The door system has the greatest need for maintenance at the people mover system sites and is a

major factor impacting the availability of the transit system. Failure in the door system may force the

people mover to be temporarily shut down due to safety concerns. Because of passengers holding the

doors open, numerous open/close cycles and harsh weather conditions, the door’s system components

are subject to a large amount of stress and deterioration. Deterioration in either the electrical DC

motor that controls the operation of the door or the mechanical levers, rollers, tracks, or switches can

cause failures in the subsystem. Currently, a labor-intensive preventive maintenance program is used

to ensure high availability. This scheduled maintenance approach requires experienced personnel to

determine the cause(s) of the problems and can result in components being serviced even though there

is no need for maintenance. In summary, the door system was selected to implement and test the

condition-based maintenance approach because it experiences a large number of degradation failures,

it is fairly readily instrumented for monitoring and it has a very similar design over all models of the

people mover vehicle line.

Although an analytic model was considered and developed early in the project based on physics-

of-failure concepts, it was not practical or effective. The analytic model could not account for all of

the aspects of operation of a real people mover door set. Therefore, an empirical approach was used

to process the signals from the people mover and estimate when maintenance should be performed.

Because of the anticipated nonlinear behavior of the door system, neural networks were chosen as the

primary predictive modeling tool. The appendix shows a flow chart of the monitoring and prediction

system for condition based maintenance.

Predictive maintenance and integrated prognostics involves condition monitoring, fault

detection and prediction of future failures. Degradation models can be used so that reliability

assessments can be made and updated based on observed degradation paths. Prognostics can be

1

defined as the capability to provide early detection and isolation of precursor and/or incipient fault

condition to a component failure condition, and to have the means to manage and predict the

progression of this fault condition to component failure.

Significant research has been conducted on condition monitoring and integrated prognostics.

El-Wardany et al. (1996) presented a study on condition monitoring of tool wear and failure in

drilling operation using vibration signature analysis techniques; Szecsi (1999) demonstrated a cutting

tool condition monitoring system for CNC lathes. The system is based on the measurement of the

main DC motor current of the lathes. The system alarms when the cutting tool wear exceeds a pre-

defined value. The monitoring system is trained by genetic algorithm and fuzzy logic based machine

learning techniques; Dimla Jr. et al. (1997) presented a comprehensive review of tool condition

monitoring systems, developed or implemented through application of neural networks; Prickett and

Johns (1999) provided an overview of approaches to the detection of cutting tool wear and breakage

during the milling process.

Lu et al. (2001) developed a technique for predicting individual system performance reliability

in real-time considering multiple failure modes. This technique includes on-line multivariate

monitoring and forecasting of performance measures and conditional performance reliability

estimates. The performance measures are treated as a multivariate time series and a state-space

approach is used to model the multivariate time series. The predicted mean vectors and covariance

matrix of performance measures are used for the assessment of system reliability. This technique

provides a means to forecast and evaluate the performance degradation of an individual system in a

dynamic environment in real-time.

Greitzer et al. (1999) discussed a prototype monitoring and prognostic system for gas turbine

engines. Artificial neural networks, rule-based algorithms, and predictive trend analysis tools were

used to diagnose and predict engine conditions. Roemer and Kacprzynski (2000) described some

novel diagnostic and prognostic technologies for gas turbine engine risk assessment. They presented

an integrated set of turbo machinery condition monitoring, diagnostic and prognostic technologies.

These technologies can be implemented across the entire spectrum of turbo machines from mid-sized

pumps to land-based gas and steam engines as well as aircraft engines. They claimed that

implementation of these technologies offers significant potential for reducing current product life

cycle costs.

2

Neural networks have effectively been used in other applications to predict performance

degradation of operating systems in real-time. Specifically, neural networks have been effective in

anticipating failure for cutting tools (Choudhury et al., 1999, Das et al., 1997, Quan et al., 1998) and

machinery (Chow et al., 1991, Lin and Wang, 1996). For these applications, sensors have been used

to detect vibration and other effects of wear and to input into a trained neural network model.

Additionally, there have been other applications including mining (Sottile and Holloway, 1994),

hydro-electric power plants (Isasi et al., 2000) and component placement for surface mount

technology (May et al., 1998).

2. Door System Description

The door system operates as an open-loop system receiving signals from an automatic train control

system that initiates opening or closing of the motor. The motor turns an operator arm, which runs

along a guidance slot in the door (see Figure 1). The movement of the operator arm pulls or pushes

the door to the open and closed positions. The door is equipped with rollers that ride in a track at the

top of the vehicle and there is a track at the bottom of the vehicle in which the door rides. As the door

goes through its cycle, micro-switches connect and disconnect resistors in the electrical circuitry,

which change the speed of the moving door. The door typically opens in 3 to 3.5 seconds, and closes

in 4 to 4.5 seconds.

Motor

Operator Arm

Top Rollers and Track

Lower Track

Operator Arm Guidance Track

Figure 1. Diagram of the door system.

The available signals from the doors are time, current and voltage of each door leaf as it passes

through a set of five switches. It is easy to calculate energy from the latter two, therefore the signals

used were time and energy. Door position, while of potential value, cannot be readily monitored.

3

However, energy and timing can sufficiently characterize the condition of the door regarding

degradation caused by contaminants in the track, track warping, engine wear out, etc.

3. Data Acquisition

3.1 On board data collection

Because neural networks are data driven models, data under a variety of conditions needed to be

obtained. Although it would have been useful to monitor some operating people movers in the field,

there were practical restrictions that prevented this type of data collection. First, because of the public

nature of airports, any alteration to a vehicle, even in the form of passive monitoring, is difficult to

obtain approval for. Second, monitoring a field site does not allow control over the degradations or

failures that might be experienced. Passive data collection from fielded systems may involve lengthy

data collection intervals without significant degradation. Instead, an experimental set-up was made to

gather data using a test vehicle provided by the industrial partner at their test track.

The key measurable signals collected were the current through the motor, the voltage across the

motor, the time interval of the open and close cycle, and the timing of the micro-switches. Using a

sampling rate of 100 samples per second, all fluctuations in the data were easily detected. All signals

pass through a circuit board that contains voltage dividers that scale down the voltage, and filters for

the current measurements and motor voltage. The circuit board sends the signals to the built-in A/D

converter, which, in turn, sends them to a laptop computer on-board. A software program was

developed to compile the data in ASCII format. The data collection software was designed to collect

data when both the closing voltage and the opening voltage are not equal to zero, i.e., data was only

collected when the door system was in operation. Then, the collected data is used to calculate the

energy and time consumption for the process.

The energy required to move the door between two positions is given by,

TcVcVcIEnergy DC ×××−××= 321 || (1)

where 54

3

ccVc

I sg

×

×= (2)

denotes the current (Vsg is the voltage shunt to ground), V and V represent closing voltage and

opening voltage respectively, and T denotes the time between two samples. c , c , , and c are

conversion factors determined by the circuit board designer.

C D

1 2 3c 4c 5

A data acquisition prototype was developed with two encoder inputs, eight isolated digital inputs,

4

two voltage inputs, one temperature input, and three current sensing inputs. Five of the eight digital

inputs are used to read the door switch contacts to determine the door position in relation to time. The

two voltage inputs measure the voltage on each side of the motor’s armature. The three current

sensing inputs measure the current through the motor by subtracting current through either the open or

close resistor circuit from the total current. The temperature input was designed to compensate for

temperature differences that would cause accuracy of the measuring system to change. For the testing

done with the prototype system, it was found that this was not necessary. Figure 2 shows a typical set

of signals over a single open and close cycle.

3.2 Friction degradation data

Detection of system performance degradation required a detailed understanding of the door

assembly failure modes and effects. As the people mover operates, weather conditions, foreign

substances in the path of the door, passengers holding the door open, etc., cause degradation of the

door’s components. This can result in the following failure modes: door failure to close, worn out

overhead rollers, bent operator arm, and worn out operator arm track. All of these failures increase

the frictional resistance against the door, causing the motor to work harder. Therefore, the effect of

friction on the door is the most important diagnostic parameter for the vehicle door system. The

frictional resistance depends on dirt, damage, wear, and obstruction by foreign materials. It impacts

the door in both directions, and leads to door failure if not maintained correctly.

����

����

��

����������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������

�����������������������������������������������������������������������������������������������������������������������

�������������������������������������������������������������������������������������������������������������������������������������������������������������

���������������������������������������������������������������������������������������������������������������������������������������������������������������

����������������������������������������������������

�����������������������������������������������

����������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������

0

500

1000

1500

2000

2500

3000

3500

1 101 201 301 401 501 601 701

Number of Sample

Volta

ge

Switch5Switch4Switch3Switch2Switch1Opening Voltage��������Closing VoltageCurrent

Figure 2. Typical signals over an open and close cycle from one door set.

To simulate frictional forces on the door, a device was designed, built, and installed onto one of the

5

doors of the test track vehicle (see Figure 3). A metal bar was bolted to the outside of the door. A U-

shaped metal ring was attached to the outside of the vehicle, and a metal bar passed through this ring.

A force meter was used to apply a force to the bar, allowing simulation of frictional forces in both

directions.

Figure 3. Frictional experiment device.

Since the degradation of the door depends on dirt, damage, wear, and obstruction by foreign

materials, the frictional resistance against the door will occur in different ways. Therefore, four

different levels of forces were applied to the frictional device (2, 4, 6, and 8 lbs.). In addition, these

forces were applied at three different places (after the door has traveled approximately 0, 1/3, and 2/3

through its cycle). The combination of these two factors results in twelve different degradation levels.

In order to check the repeatability and variation of data, each individual experiment was run for ten

close-open cycles. Also, ten cycles of normal operation were collected to establish a baseline.

The most extreme case (8 lbs) had been determined to be indicative of nearly immediate failure of

the door assembly. This was designated to correspond to a dimensionless “degradation measure” of

1. In practice, if the door reaches this level, it would be reasonably concluded that a failure occurred.

The objective of this project was to predict degradation such that preventive measures can be taken

prior to failure. Other degradation measures were considered in comparison to the maximum level,

thereby creating a continuous degradation measure ranging from 0 to 1.

4. Data Preprocessing

The physical performance of the door system is monitored continuously and the resulting

sequenced data can be analyzed using time-series modeling. From the experimental data described in

Section 3, simulated data was generated to fill in the gaps between force values. This was done to

create the pattern of continuous degradation from healthy to fully degraded. In order to form the

simulated data, the mean and standard deviation of each set of experimental data were calculated.

Next, 200 simulated data points for healthy operation were generated by using the mean and standard

deviation of the experiment data with no friction force applied. 100 data points for the interval

6

between each applied force level were simulated by linearly increasing the mean and standard and

deviation and using a Gaussian random number generator. For instance, if the mean of experiment

with no friction forces is , the standard deviation of the experiment with no friction forces is ,

the mean of the 2 pound experiment is µ , and the standard deviation of the 2 pound experiment is

, the mean and standard deviation of the Gaussian random number generator for each simulated

point between them will be

Nµ Nσ

2

2σ

100Nµ−2

Nµ

+µ=µ (t) and 100

2 NN

σ−σ+σ=σ (t) respectively, where t is time.

A plot of the combined actual and simulated energy data is shown in Figure 4.

In order to reduce the noise in the data and detect the degradation trend, an exponential smoothing

method was used. The exponential smoothing formula is shown below.

O-S-A (One-Step-Ahead) Forecast: (3) 11 −− += ttt GSF

Mean (4) 1 1(1 )( ) (1 )t t t t tS D S G D− −= α + − α + = α + − α tF

1t−Trend (5) 1( ) (1 )t t tG S S G−= β − + −β

where represents the original data and and β denote the smoothing constants. Three different

pairs of ( α ) - (0.1, 0.1), (0.2, 0.2), (0.3, 0.3) - were used in the exponential smoothing model. In

order to compare the effect of the constants, the MAFE (Mean of Absolute Forecast Error) was

calculated. When ( ) is equal to (0.1, 0.1), the results give the smallest MAFE. Therefore, this

pair of values was used for generating the exponentially smoothed data. A plot of an exponentially

smoothed data example is shown in Figure 5.

tD α

β,

βα,

150

200

250

300

350

400

450

1 101 201 301 401 501

solid = opening energy : dashed = closing energy

Figure 4. An example of combining actual and simulated energy data.

7

150

200

250

300

350

400

450

1 101 201 301 401 501

dark = opening energy, light = closing energy

Figure 5. An example of the exponentially smoothed energy data.

5. Reliability Assessment Based on Observed and Projected Degradation Paths

The degradation neural network models are used to assess the conditional reliability of an

individual door assembly and to make forecasts based on projected degradation trends. These are to

be used a basis for proactive predictive maintenance policies.

Often research efforts are focused on reliability prediction based on degradation test data across

an entire population. This information is useful for system designers, who are concerned about

reliability characteristics of products in large volumes. In practice, however, individual users may be

more interested in reliability characteristics of an individual unit rather than the population.

Reliability prediction based on on-line degradation monitoring provides the potential to address this

problem. The degradation process of an individual unit is monitored on-line. Reliability prediction

for the unit can be continuously updated based on new observed degradation measurements.

The degradation neural network model can be applied to individual system reliability prediction

utilizing condition monitoring and integrated prognostics. The problem to be addressed here is, given

the observed degradation path and current degradation measure, the conditional reliability and mean

residual life distribution are required to identify future maintenance activity. This offers the potential

to minimize down-time caused by unscheduled maintenance, and also reduce replacement of

operating units with significant remaining useful life.

Consider z(x(t)) to be the dimensionless degradation output measure as a function of vector inputs,

x(t), observed at time t. Maintenance can be triggered by two types of observations. Consider the

failure threshold to be D, and δ is a degradation measure (δ < D) that is indicative of unacceptable

degradation, but still not a failure. A reactive, but preventive task can be initiated at this observation

8

(z(x(t)) ≥ δ). Then, system failure can be prevented as long as the degradation-level does not

deteriorate to the D level before the first available time for maintenance, i.e., at night between shifts.

A second type of maintenance can be implemented by projecting observed trends to predict a

lower-bound value for time-to-failure. This serves as planning guides for preventive maintenance that

can be updated and revised as more data is collected. The projected failure times are determined

individually for each door assembly based on the observed model inputs and predicted output.

Consider the data collection times (at each open-and-close cycle) as t1, t2, …, tn, model input

parameters as x(t1), x(t2), …, x(tn), and the successive neural network model predictions to be z(x(t1)),

z(x(t2)), …, z(x(tn)). Exponential smoothing time series models are used to predict future degradation.

Then, the models are inverted to predict time-to-failure. A lower-bound on time-to-failure is made by

numerically adjusting the prediction. In practice, the lower-bound should be a statistical lower-bound

at the α-level. However, with the amount of failure data available in this research study, this was not

possible. Instead, a subjective numerical adjustment was made based on maintenance and design

personnel’s experience.

Consider the following prediction conceptual model, based on exponential smoothing, for

predicted degradation at failure time, t (t > tn). t represents the lower-bound failure time prediction

and ε is an adjustment factor.

ˆ

1 2 11

1 2 1

ˆ( ) ( ( ), ( ),..., ( ), ( ))ˆ ( ( ), ( ),..., ( ), ( ))

n n

n n

z t f t t t t D

t f D t t t t−

−−

= =

= −

x x x x

x x x x ε (6)

The failure time predictions for each system are then used as a guide for planning and scheduling

preventive maintenance. Determination of δ and ε are not based only reliability behavior. These need

to be determined and customized by the airport operator based on their propensity for risk,

reliability/availability quantitative requirements, and the cost trade-off between scheduled and

unscheduled maintenance.

6. Neural Network Model Development for Condition Monitoring

6.1 Choice of neural network paradigm

Artificial neural networks began in the 1940’s, with the intention of emulating the strength and the

processing power of the human brain. A neural network is a parallel distributed processor and it

acquires knowledge through iterative “learning.” This acquired knowledge is stored in its connection

weights, which transcend from an initial random state to a fixed state through the learning process. A

typical neural network consists of several layers of interconnected processing units. Because of its

9

theoretical property of universal approximation (Funahashi, 1989, Hornik et al., 1989), a

backpropagation network was chosen as the primary modeling vehicle (Figure 6). This is a fully

connected, multi-layer network that consists of input units, hidden units, and output units (Rumelhart

et al., 1986). The learning (also known as training) process involves three stages: 1) the feedforward

of the input training pattern, 2) the calculation and backpropagation of the associated error, and 3) the

adjustment of the weights according to an error message (usually squared error). The

backpropagation learning algorithm is the most well known training algorithm and adjusts the

connection weights according to the gradient descent method where the squared error is minimized in

the direction of greatest improvement.

.

.

.

.

.

.

.

.

.

InputLayer

FirstHiddenLayer

SecondHiddenLayer

OutputLayer

11y

1x

2x

3x

nx

21y)( 11yf )( 21yf

1z )( 1zf

)( 2zf

)( 3zf

2z

3z

11ny22 ny)(

11nyf )(22nyf

[ ])(1)()(,)exp(1

1)( yfyfyf

yyf −=′

−+=binary sigmoid

transfer function:

Figure 6. A typical backpropagation neural network.

A serious drawback of the backpropagation paradigm is the necessity to a priori select the

architecture, i.e., number of hidden layers and hidden nodes. Two other paradigms, which reduce

dependence on this selection, were also tested. These are the cascade correlation network (Figure 7),

which builds its own architecture incrementally, also using an error feedback algorithm to minimize

squared error, and the radial basis function network (Figure 8), which uses basis functions (hyperbolic

10

tangent in this case) to localize inputs prior to using the error feedback minimization algorithm to

determine output layout weights.

Inputs Outputs

Middle NeurodesAdjoined

1x

2x

3x

4x

1y

2y

)( 1yf

)( 2yf

1z

2z

3z

)( 1zf

)( 2zf

)( 3zf

[ ][ ])(1)(1)(,)2exp(1

)2exp(1)( yfyfyf

y

yyf −+=′

−+

−−=

hyperbolic tangenttransfer function:

Figure 7. A typical cascade correlation neural network.

6.2 Variable selection

Artificial neural networks are trained using observations collected from the system under

investigation. Once trained, the network recognizes patterns similar to those it was trained on and

classifies new patterns accordingly. The development of the neural network-based condition

monitoring system requires training data for classifying the relative condition of the door system.

Each training pattern must be associated with a level of degradation. Two approaches to selection of

input variables were used in the neural network:

1. four input variables (closing energy, opening energy, closing time, and opening time)

2. 24 input variables (closing energy, opening energy, closing time, and opening time over the

five-switch locations, creating six physical longitudinal door sections (Figure 9).

11

.

.

.

.

.

.

.

.

.

Input Layer Output LayerHidden Layer

1x

2x

1−nx

nx

1

1y

Ny

)( 1yf

)( Nyf

z)( zf

[ ][ ])(1)(1)(,)2exp(1)2exp(1

)( yfyfyfyy

yf −+=′−+

−−=

hyperbolic tangenttransfer function:

Figure 8. A typical radial basis function neural network.

One output variable, the level of degradation, was a continuous value between 0 and 1, where 0 is

perfectly healthy and 1 is fully degraded. All data was exponentially smoothed as described in the

preceding section.

Because neural networks can be sensitive to number of trainable weights (architecture) and

learning rate (Geman et al., 1992), different networks were constructed using different combinations

of numbers of hidden neurons and learning rates. Learning rates of 0.1 and 0.3 were tested for all

networks with little difference between the two. For the backpropagation four input network, hidden

neurons of 5, 7 and 10 were tried (in a single hidden layer) and for the 24 input case, 15, 20, and 25

hidden neurons were used. For all radial basis function networks 15, 20, and 25 hidden neurons were

tested for all cascade correlation networks an upper bound of 50 hidden neurons was set. Table 1

shows results of the four input case. While the radial basis function network achieved the lowest

error on the training set, its ability to generalize was weak, as evidenced by the large testing error rate.

The R2 values of all networks were high, however, indicating that neural network predictive modeling

is effective for this application. Table 2 shows the results of 24 input case. Here, the

12

backpropagation network was superior for both training and testing. This network is also

significantly more accurate than the four input case. Therefore, it can be concluded that dividing the

energy and time into physical increments along the door track is helpful in estimating door condition.

Figure 10 shows typical predicted (both training and testing) versus actuals over three degradation

cycles using the backpropagation, 24 input trained neural network.

Table 1. Results of the different neural network paradigms for the four input case.

Network Training RMS

Training R2

Testing RMS

Testing R2

Number of hidden neurons in the network

Backpropagation .101 .968 .073 .985 5 Cascade Correlation .050 .978 .071 .955 16 Radial Basis Function .035 .989 .100 .917 15

RMS is root mean squared error.

R2 is the coefficient of determination.

Table 2. Results of the different neural network paradigms for the 24 input case.

Network Training

RMS

Testing

RMS

Number of

hiddBackpropagation 0.00674 0.04419 25 Cascade Correlation 0.04144 0.08430 50* Radial Basis Function 0.04662 0.10998 20

RMS is root mean squared error.

50 was set as an upper bound.

OpenClose

Figure. 9. Segmentation of track into six segments by the five switches.

6.3 Software platform

The data preprocessing was implemented using Borland C++. The backpropagation neural

13

network was developed in the BrainMaker ProfessionalTM (four input case) and the NeuralWorksTM

(24 input case) software packages. In both packages, all networks are complied C code after training.

-0.1

0.1

0.3

0.5

0.7

0.9

1.1

1 62 123

184

245

306

367

428

489

550

611

672

733

794

855

916

977

1038

1099

1160

1221

1282

1343

1404

1465

1526

1587

1648

1709

1770

sample points

Degr

adat

ion

TARGETTRAINTEST

Figure 10. Actual (target), training and testing data for a backpropagation network with 25 hidden

neurons and a learning rate of 0.3 over three degradation cycles.

7. Field Test at Pittsburgh International Airport Site

The next phase of the project was to install the prototype system at the Pittsburgh International

Airport Site for a three-month period to determine the first round of refinements that would have to be

made. In order to facilitate an easy installation of the prototype system into a vehicle at the Pittsburgh

Airport site, harnesses were made in the shop and labeled to assure quick and accurate wiring when

installed on a vehicle. This installation was installed during third shift when traffic would be a

minimum. Eight units were installed in the landside car on the south train.1 Installation not only

included installing the prototype units, it also included testing all the doors on the vehicle to assure

that safety or operation was not compromised.

Figure 11 shows the prototype installed into the control circuit of one of the door leaves of the

Pittsburgh International Airport vehicle. Two harnesses were required for installation – one for the

digital signals, communication, and voltage signals and a second harness of a thicker gauge wire for

the current connections. If any failure occurred within the prototype data acquisition unit itself, it

could quickly be disconnected from the system by retracting the 104 pin connector at one end and

1 At the Pittsburgh airport, there are two “must ride” vehicles, one on the south side and one on the north side. Both move back and forth from the landside terminal to a single transport terminal.

14

removing and connecting together the four wires of the current sensing harness.

A laptop computer was installed using Velcro to hold it in place and was put in the end of the

vehicle where the technicians keep their toolbox. The laptop computer could be quickly lifted up to

view if necessary without disconnecting any of the wiring. The same circuit that is used to power the

utility outlets was used to power the computer. Data was removed from the laptop once every other

day when the vehicle came in for routine maintenance. Compiled data was sent to the industrial

partner and Auburn University for analysis and monitoring.

Figure 11. Prototype system installed on a vehicle at the Pittsburgh International Airport.

Figure 12 shows some of the data from the field trial. There are three door sets which, although

all initially healthy, exhibit different operating characteristics. It is evident that inherent differences

exist among the signal (energy and time) traces of the vehicles’ door set. This motivates the need to

customize the predictive algorithm to each door set. This could be accomplished by gathering data

passively during a specified period when the system is first installed in a vehicle, assuming the door

was healthy when installed. Once the norms of each door set are known, the algorithm could be

adjusted by a positive or negative constant to correctly predict condition for that door set.

15

222223N =

Door Leaf

632

Mea

n Va

lue

of T

otal

Ope

n En

ergy

140

120

100

80

60

40

20

12/22/9912/20/99

12/18/99

222223N =

Door Leaf

632

Mea

n Va

lue

of T

otal

Clo

se T

ime

3.2

3.1

3.0

2.9

2.8

2.7

02/27/0001/28/00

01/11/00

Figure 12. Box plots of open energy (left) and close time (right) of three door leaves during the trial.

The field trial data showed that energy is a better indicator of door health than time. The data also

showed that time and energy were not well correlated for individual door sets. This is contrary to

expectations. This probably occurs because open and close times are too affected by riders in the

vehicle while energy used to open and close the doors primarily relates to the condition of the track

along which the door moves. Figure 13 shows two of the door leaves with their energy tracings over

about 2 ½ months (over various days during the field trial). These plots are by date (x axis) with the

observations during each day summarized as a box plot (y axis). Of course, there are different

number of open and close cycles each day. This door leave shows degradation from the beginning of

the trial through the end, although the increase in energy is non linear. Note how the open energy and

close energy of this door leave are correlated.

Total Close Energy at Door Leaf 6

Date

227

225

222

217

215

209

207

205

131

128

126

120

118

115

111

101

1230

1228

1226

1222

1220

1218

300

200

100

0

Total Open Energy at Door Leaf 6

Date

227

225

222

217

215

209

207

205

131

128

126

120

118

115

111

101

1230

1228

1226

1222

1220

1218

160

140

120

100

80

60

Figure 13. Open energy (left) and close energy (right) of a door leaf exhibiting degradation over time.

Similarly, in a door leaf not exhibiting degradation, the energy traces are clear, as in Figure 14.

16

Total Close Energy at Door Leaf 3

Date

227

225

222

217

215

209

207

205

131

128

126

120

118

115

111

101

1230

1228

1226

1222

1220

1218

46

44

42

40

38

36

Figure 14. Stable door leaf over time.

8. Concluding Discussion

Having reviewed the data captured on the test vehicle and the data captured at the Pittsburgh

Airport Site, a case can be made for condition-based maintenance on the door systems. The following

conclusions from the field test have already been made:

• The idea of predictive maintenance for degradation-type behavior is technically sound.

• It is possible to design, build and implement a predictive maintenance system for people mover

doors without major design changes and without significant investment.

• Vehicles do exhibit a wide variety of “normal” behavior and therefore predictive models will need

to be customized to individual systems, in this case, individual door sets.

• Significantly more data from door sets in some known state of degradation is needed to build

effective predictive models.

• Analytic modeling does not appear to be practical and therefore traditional model based

approaches will probably fail.

• Empirical modeling does appear to be practical and effective; a supervised neural network

approach, such as backpropagation, preceded by exponential smoothing works well.

• Much of the time and cost of this project was spent on developing the necessary hardware to

collect the time, current and voltage signals and on unfruitful attempts at analytic modeling and

collecting door position data. The project is now poised, if continued, to move forward in an

efficient and expedient manner.

17

In addition, a U.S. provisional patent have been filed and several Asian and European patents have

been granted for this system.

18

Appendix. Flowchart of Condition Based Maintenance System.

Continuous Data

Collection by Boards

Transfer Files to Decimal Format &

Store in Laptop

Neural Network

A

Degree of Degradation

Necessary Maintenance Recommended

If Value > 0.5 (or other determining value)

A

Others

Historical Data Current Data

Trend Analysis Generate Input Data by

Calculation, Exponential Smoothing

Inputs (energy and time consumption) are generated by calculation, exponential smoothing

1 output (degree of degradation) is generated by using a Neural Network (backpropagation, radial basis function, or cascade correlation network).

Collected data is used to calculate energy and time consumption during each open-close cycle.

Collected data include motor current, voltage, and switch signals in hexadecimal format.

Start

19

REFERENCES

1. S. K. Choudhury, V. K. Jain, Ch. V. V. Rama Rao, “On-line monitoring of tool wear in turning using a neural network,” International Journal of Machine Tools & Manufacture, 1999 v. 39, pp 489-504.

2. M. Chow, P. Magnum, and S. O. Yee, “A neural network approach to real-time condition monitoring of induction motors,” IEEE Transactions on Industrial Electronics, 1991 v. 38(6), pp 448-453.

3. S. Das, P. P Bandyopadhyay, A. B. Chattopadhyay, “Neural-networks-based tool wear monitoring in turning medium carbon steel using a coated carbide tool,” Journal of Materials Processing Technology, 1997 v. 63, pp 187-192.

4. D. E. Dimla Jr., P. M. Lister, N. J. Leighton, “Neural Network solutions to the tool condition monitoring problem in metal cutting – A critical review of methods”, International Journal of Machine Tools & Manufacture, 1997 v. 37, n. 9, pp 1219-1241.

5. T. I. El-Wardany, D. Gao, M. A. Elbestawi, “Tool condition monitoring in drilling using vibration signature analysis”, International Journal of Machine Tools & Manufacture, 1996 v. 36, n. 6, pp 687-711.

6. K. Funahashi, “On the approximate realization of continuous mappings by neural networks,” Neural Networks, 1989 v. 2, pp 183-192.

7. S. Geman, E. Bienenstock, R. Doursat, “Neural networks and the bias/variance dilemma,” Neural Computation, 1992 v. 4, pp 1-58.

8. F. L. Greitzer, L. J. Kangas, K. M. Terrones, M. A. Maynard, B. W. Wilson, R. A. Pawlowski, D. R. Sisk, N. B. Brown, “Gas turbine engine health monitoring and prognostics”, presented at the International Society of Logistics (SOLE) 1999 Symposium, Las Vegas, Nevada, August 30 - September 2, 1999.

9. K. Hornik, M. Stinchcombe, H. White, “Multilayer feedforward networks are universal approximators,” Neural Networks, 1989 v. 2, pp 359-366.

10. P. Isasi, A. Berlanga, A. Sanchis, J. M Molina, “Hydroelectric power plant management relying on neural networks and expert system integration,” Engineering Applications of Artificial Intelligence, 2000 v. 13, pp 357-369.

11. C. C. Lin, H.-P. Wang, “Performance analysis of rotating machinery using enhanced cerebellar articulation controller (E-MAC) neural networks,” Computers in Engineering, 1996 v. 30(2), pp 227-242.

12. S. Lu, H. Lu, W. J. Kolarik, “Multivariate performance reliability prediction in real-time,” Reliability Engineering and System Safety, 2001 v.72, pp 39-45.

13. G. S. May, H. C. Forbes, N. Hussain, A. Louis-Charles, “Modeling component placement errors in surface mount technology using neural networks,” IEEE Transactions On Components, Packaging and Manufacturing Technology-Part C, 1998 v. 2(1), pp 66-70.

14. Y. Quan, M. Zhou, Z. Luo, “On-line robust identification of tool-wear via multi-sensor neural-network fusion,” Engineering Applications of Artificial Intelligence, 1998 v. 11, pp 717-722.

15. M. J. Roemer, G. J. Kacprzynski, “Advanced diagnostics and prognostics for gas turbine engine

20

risk assessment”, IGTI/ASME Turbo Expo, Munich, Germany, May 2000.

16. D. E. Rumelhart, G. E. Hinton, R. J. Williams, “Learning internal representations by error propagation,” Parallel Distributed Processing: Explorations in the Microstructure of Cognition. Vol. 1: Foundations (D. E. Rumelhart and J. L. McClelland, editors), Cambridge, MA: MIT Press, 1986, pp 318-362.

17. J. Sottile, Jr., L. E. Holloway, “An overview of fault monitoring and diagnosis in mining equipment,” IEEE Transactions on Industry Applications, 1994 v. 30(5), pp 1326-1332.

18. T. Szecsi, “A DC motor based cutting tool condition monitoring system”, Journal of Materials Processing Technology, 1999 v. 92-93, pp 350-354.

21