How MRT Track Inspection Works Using Mathematics: Turning Kilometres of Rail into a Continuous Health Map

An MRT track looks simple from a train window.

Two rails continue into the distance. Sleepers or track-support structures repeat beneath them. Fasteners hold geometry in place. The train rolls forward.

But a railway does not experience its track as one long object.

It experiences millions of measurable points.

At one location, alignment may drift slightly. At another, rail surface becomes uneven. A fastener may be missing. A crack can begin below what a passenger could ever see. Gauge can change. A recurring vibration can reveal geometry deterioration. A camera can observe an object that differs from the expected image. A measurement train can produce an enormous sequence of position-linked values while moving through the network.

The challenge is not merely finding something unusual.

The railway must decide:

Is the measurement correct?
Is the observation actually abnormal?
Is the abnormality getting worse?
Is it a harmless irregularity or a meaningful defect?
How uncertain are we?
When should engineers inspect physically?
Where should maintenance be prioritised?

This is why modern track inspection is a mathematics problem as much as a visual inspection problem.

The railway does not merely inspect rail. It converts space into measurements, measurements into evidence, evidence into risk, and risk into maintenance decisions.

Singapore already uses this logic in public operation. LTA says two newer North East Line trains carry Automatic Track Inspection systems using underframe cameras and sensors to monitor rail condition while trains are in service. LTA specifically lists defects such as rail cracks, rail corrugation and missing rail fasteners. The system complements physical inspections and allows more timely maintenance.

SMRT also publicly describes its Rail Vision automatic track-inspection system as allowing continuous monitoring of track conditions during operating hours, alongside daily physical track inspection and periodic rail grinding.

Research is advancing rapidly as well. A 2026 review of railway track-geometry degradation compares prediction methods by accuracy, uncertainty quantification, scalability, interpretability and cost, while recent reviews of machine vision and railway condition monitoring highlight the growing use of cameras, signal processing and machine learning.

This article continues the eduKateSG MRT mathematics series. Begin with How MRT Timing Works Using Mathematics, and see How MRT Predictive Maintenance Works Using Mathematics for the maintenance layer surrounding this inspection problem.

The RFE — Why Does Track Inspection Mathematics Exist?

The weakest possible objective is:

find every imperfection

That would be a poor engineering objective.

No physical track is mathematically perfect.

Measurements contain noise. Rails contain small manufacturing and wear variations. Geometry changes slightly with temperature, load and maintenance. Sensors have finite accuracy. The train itself vibrates. Lighting changes. Dirt appears. Water reflects. Cameras can be partially obscured.

If every deviation becomes an alarm, engineers are buried in false positives.

If thresholds are too permissive, genuine deterioration can remain invisible.

The Reason for Existence is therefore:

measure track condition reliably
+
distinguish signal from noise
+
identify meaningful degradation early
+
locate it accurately
+
measure how it changes over time
+
prioritise justified physical inspection and maintenance
+
prevent avoidable reliability and safety problems

The goal is not maximum alarm generation.

It is maximum useful evidence.

A good track-inspection system does not ask, “Can I detect a difference?” It asks, “Is this difference real, meaningful, worsening and worth acting on?”

Prompt 1 — What Exactly Is Being Measured on an MRT Track?

A railway track has both geometry and material condition.

Geometry tells us where the rails are.

Material condition tells us what physical state they are in.

Track gauge

Gauge is the lateral distance between the rails at the defined measurement points.

Let the measured gauge at position x be:

G(x)

Deviation from a reference value G₀ is:

ΔG(x) = G(x) − G0

The public educational point is not the actual operational threshold.

It is that gauge becomes a spatial function rather than one number for the whole line.

Alignment

Horizontal alignment describes lateral deviation of the rail from its intended path.

Let:

y(x) = lateral rail position

An ideal or reference path is y₀(x).

ealignment(x) = y(x) − y0(x)

Longitudinal level

Vertical geometry can be represented by rail elevation z(x).

elevel(x) = z(x) − z0(x)

Changes over particular wavelengths can influence ride quality and dynamic wheel–rail forces.

Cross level and twist

Let left and right rail heights be z_L(x) and z_R(x).

Cross-level difference is:

C(x) = zL(x) − zR(x)

Twist over a base length L can be represented:

Twist(x;L)
=
[C(x+L) − C(x)]/L

The concept matters because geometry can change over distance, not merely at one isolated point.

Curvature

For a planar curve y(x), curvature can be approximated mathematically by:

κ(x)
=
y″(x)/[1 + y′(x)²]3/2

For small slopes, this is approximately:

κ(x) ≈ y″(x)

Curvature influences wheel–rail interaction and local wear.

Rail surface condition

Rail condition is not only geometric alignment.

Inspection may look for:

• surface irregularity,
• corrugation,
• cracks,
• wear,
• fastener condition,
• objects or damage near the track,
• and other defects defined by the inspection programme.

LTA publicly lists rail cracks, corrugation and missing fasteners among the defects monitored by the North East Line Automatic Track Inspection system.

SMRT says uneven track surface is a major contributor to track noise and that rail grinding is used to restore a smoother surface.

Dynamic condition

A track can also be inferred through the behaviour it creates in a passing train.

Axle-box or vehicle acceleration a(t) can contain signatures of track irregularity.

Position is related through train speed v:

x(t) = ∫v(t)dt

So a vibration measured in time can be mapped back to a track location.

This is one reason condition monitoring can be performed by ordinary service trains rather than only specialised inspection vehicles.

Track health is not one variable. It is a field of geometry, material condition and dynamic response distributed along kilometres of space.

Prompt 2 — How Does a Moving Train Turn Track into Data?

A track-inspection system needs three things:

measurement
+
location
+
time

If a camera sees a missing fastener but the railway cannot locate it precisely, the observation is much less useful.

If a geometry sensor records a spike but the timestamp is misaligned with train position, engineers may inspect the wrong place.

So every observation should become something like:

observation = (track position x, time t, sensor value s)

Sampling distance

Suppose the train moves at velocity v metres per second and a sensor samples at f samples per second.

Average spatial distance between samples is:

Δx = v/f

If:

v = 20 m/s
f = 1,000 Hz

then:

Δx = 20/1000
   = 0.02 m
   = 2 cm

One sensor stream can therefore produce a measurement every few centimetres at this illustrative rate.

But sampling more frequently does not automatically improve inspection.

It creates:

• more storage,
• more computation,
• more correlated observations,
• and potentially more noise.

Nyquist reasoning

If a signal contains a highest meaningful temporal frequency f_max, the classical Nyquist condition says sampling frequency should exceed:

fs > 2fmax

For a spatial wavelength λ_min, a comparable spatial argument requires enough samples to resolve the feature.

At train speed v, a spatial wavelength λ corresponds to temporal frequency approximately:

f = v/λ

So faster inspection speeds demand higher sampling rates to preserve spatial resolution.

This is a beautiful link between train speed and signal processing.

Aliasing

If sampling is too sparse, a short-wavelength irregularity can appear as a completely different lower-frequency pattern.

This is aliasing.

Inspection mathematics must therefore choose sample rates based on the smallest relevant feature rather than merely the amount of data convenient to store.

Position uncertainty

Suppose estimated track position is x̂ and true position is x.

eposition = x̂ − x

If standard deviation of position error is σ_x, the inspection system should not pretend a defect is known to greater precision than that uncertainty allows.

A practical defect record might therefore include:

estimated location
+
location uncertainty
+
defect probability
+
observation time

Repeated passes improve evidence

One service train passes today.

Another passes tomorrow.

The same location is observed again.

Repeated inspection creates a spatial time series:

H(x,t1), H(x,t2), H(x,t3), ...

This transforms inspection from:

What does the track look like?

into:

How is this exact location changing?

The most valuable inspection is often not the first measurement. It is the second measurement at the same place, because change becomes visible.

Prompt 3 — How Does Mathematics Separate Track Signal from Measurement Noise?

Every sensor measures:

observed value
=
true physical state
+
measurement error

In symbols:

y(x) = h(x) + ε(x)

where h(x) is the underlying track condition and ε(x) is noise.

The first challenge is estimating h(x).

Moving average

A moving average smooths high-frequency random variation:

m(x)
=
(1/n)Σi=0n−1 y(x−iΔx)

But too much smoothing can erase a short, localised defect.

This creates a classic bias–variance trade-off.

large smoothing window
→ low noise
→ risk of hiding local defect

small smoothing window
→ preserve local detail
→ more false fluctuation

The 2026 track-geometry degradation review specifically notes that localised exceedance events can be under-detected when models smooth too aggressively.

Filtering

Different wavelengths can correspond to different physical phenomena.

A filter can preserve a band of interest and suppress others.

For signal y(t), Fourier transform is:

Y(f) = ∫ y(t)e−i2πft dt

Signal energy in frequency band [f₁,f₂] is proportional to:

Eband = ∫f1f2 |Y(f)|²df

A recurring rail-surface pattern can therefore appear as energy concentrated in certain frequency or wavelength bands.

Kalman-style state estimation

If track condition evolves over time, a state-space model can separate process evolution from measurement noise.

state:
xk = A xk−1 + wk

measurement:
yk = H xk + vk

where:

• x_k is latent physical condition,
• w_k is process uncertainty,
• y_k is observed measurement,
• v_k is sensor noise.

The filter combines a prediction with a measurement according to their uncertainties.

A noisy sensor is trusted less.

A precise sensor is trusted more.

This is a general estimation principle, not a claim about a specific Singapore implementation.

Outliers

A single extreme measurement may come from:

• a real defect,
• sensor glitch,
• dirt or water,
• temporary obstruction,
• positioning error,
• or unusual train motion.

So the system can require spatial or temporal persistence.

For example, an anomaly is stronger when:

same location
+
same type of deviation
+
multiple independent passes
+
consistent growth

all agree.

Noise is not the enemy of inspection. Pretending noise does not exist is.

Prompt 4 — How Does Computer Vision Inspect Rail and Fasteners?

A camera converts track condition into an image.

An image is a matrix.

I(i,j,c)

where i and j identify pixel position and c identifies colour or another channel.

The inspection system must transform this matrix into an engineering statement.

image
→ locate rail region
→ detect objects or surface features
→ classify condition
→ locate anomaly along track
→ assign confidence

Image normalisation

Lighting changes from station to tunnel.

The system may correct brightness or contrast so the same physical object has a more consistent representation.

A simple standardisation is:

I′ = (I − μI)/σI

More sophisticated methods correct illumination, motion blur and perspective.

Segmentation

The model separates important regions:

rail
fastener
sleepers or supports
background
foreign object

For pixel p, segmentation model returns class probability:

P(class = k | pixel p)

Object detection

A fastener detector may output bounding box b and confidence score p:

detection = (b,p,class)

If an expected fastener is absent at a known position, the system can flag the location for verification.

Classification

For image patch x:

P(defect class k | x)

may be estimated by a machine-learning model.

The important engineering challenge is class imbalance.

Healthy track is common.

Serious defects are rare.

A model that simply predicts “healthy” all the time can achieve high overall accuracy while being useless.

Why accuracy can lie

Suppose 100,000 image regions contain:

99,900 healthy
100 genuine defects

A model predicting “healthy” for every image obtains:

accuracy = 99.9%

but detects:

0 of 100 defects

Recall is:

Recall = TP/(TP+FN) = 0

This is why inspection systems need metrics such as:

precision
recall
specificity
F1 score
false alarms per kilometre
missed defects per inspection cycle

Precision and recall

Precision = TP/(TP+FP)

Recall = TP/(TP+FN)

F1 = 2PR/(P+R)

High precision means alarms are usually meaningful.

High recall means genuine defects are usually detected.

The threshold must balance both according to consequence.

A 2025 systematic review of machine-vision railway inspection highlights the broad use of image processing and learning techniques while noting the practical challenge of turning visual detection into reliable real-world inspection.

A camera does not replace an inspector merely because it can classify pixels. It becomes useful when its detections are accurate enough, localised enough and explainable enough to support engineering action.

Prompt 5 — How Does Sensor Fusion Make Track Inspection More Reliable?

One sensor can be wrong.

Two different sensors can fail differently.

That makes sensor fusion valuable.

Suppose a location produces:

• a camera anomaly,
• higher vehicle vibration,
• a geometry deviation,
• and repeated observations across several train passes.

Combined evidence can be stronger than any one measurement.

Bayesian fusion

Let D be the event that a meaningful defect exists.

Prior probability is:

P(D)

Camera evidence is C.

Vibration evidence is V.

Geometry evidence is G.

Posterior becomes:

P(D|C,V,G)
∝
P(C,V,G|D)P(D)

If evidence sources are approximately conditionally independent given D:

P(C,V,G|D)
≈
P(C|D)P(V|D)P(G|D)

Real sensor dependencies must be modelled carefully, but the structure shows how multiple weak clues can combine.

Weighted fusion

A simpler health score is:

H(x)
=
w1 zgeometry(x)
+ w2 zvibration(x)
+ w3 zvision(x)
+ w4 zultrasonic(x)

where z values are standardised anomaly scores.

The weights reflect confidence and relevance.

A sensor with poor data quality should contribute less.

Let quality q_j range from 0 to 1:

wj,effective = qj wj

Now evidence automatically weakens when sensor health is uncertain.

Spatial agreement

Two sensors may report anomalies at nearby but not identical locations because of positioning error.

Let observations have location estimates x₁ and x₂ with uncertainties σ₁ and σ₂.

A simple consistency condition is:

|x1 − x2|
≤
k√(σ1² + σ2²)

for chosen confidence multiplier k.

This avoids treating tiny localisation differences as contradictory evidence.

Redundant observation

Suppose a defect is detected independently on n passes with single-pass detection probability p.

The probability of at least one detection is:

P(at least one)
=
1 − (1−p)n

If p = 0.7 and there are three independent observations:

P = 1 − 0.3³
  = 0.973

Repeated observation can therefore increase detection confidence substantially.

But independence is an assumption. If all passes use the same faulty sensor or lighting condition, repeated agreement can repeat the same error.

More measurements are powerful only when they contribute genuinely new evidence.

Prompt 6 — How Does Inspection Become a Track Health Map?

A railway needs more than a list of alarms.

It needs a map of condition along the line.

Divide the track into segments of length ΔL.

For segment i:

Si = [xi, xi+ΔL)

Each segment receives features:

gauge deviation
alignment variation
level variation
surface score
fastener score
vibration score
historical defect count
maintenance age

A composite track quality index could be written:

TQI_i
=
wG fG(i)
+ wA fA(i)
+ wL fL(i)
+ wS fS(i)
+ ...

The exact real-world indices and standards vary.

The educational idea is that many measurements can be projected into one comparable health scale.

Standard deviation as geometry quality

For geometry measurements y₁…y_n within a segment:

σsegment
=
√[(1/n)Σ(yi − ȳ)²]

A larger standard deviation means geometry is more variable within that segment.

However, a segment-level average can hide one sharp local defect.

That is why modern research distinguishes segment-level indices from point-wise forecasting and local exceedance-event detection.

Two-layer health representation

A robust public conceptual model uses two layers.

Layer 1:

segment health
= overall condition of a track region

Layer 2:

local event
= isolated anomaly or exceedance at a precise position

This avoids a dangerous averaging problem.

A 100-metre segment can be healthy on average while containing one important 20-centimetre defect.

Spatial clustering

Suppose several nearby anomaly points appear:

x = 10.21 km
10.23 km
10.24 km
10.28 km

A clustering algorithm can treat them as one local condition rather than four unrelated faults.

For DBSCAN-style clustering, points are grouped when enough neighbours fall within radius ε.

This can help identify defect regions while rejecting isolated noise.

Heat maps

A health map can display:

position along line
×
health severity
×
time

Now engineers can see whether:

• a problem is local or widespread,
• one region deteriorates faster than others,
• maintenance restored the signal,
• or a recurring issue returns after intervention.

The railway has changed from a list of inspection reports to a living spatial database.

A track health map is a memory of the railway: where it was healthy, where it changed, what was repaired, and what began changing again.

Prompt 7 — How Does Track Inspection Predict Degradation Rather Than Merely Detect It?

Detection says:

this location is unusual now

Prediction asks:

how will this location evolve?

Let track-health indicator H(x,t) change over time.

At fixed location x₀:

H(t) = H(x0,t)

Linear degradation

H(t) = H0 + kt

If intervention boundary is H_crit:

tcrit
=
(Hcrit − H0)/k

This creates an estimated maintenance horizon.

Nonlinear degradation

Track deterioration may accelerate.

H(t) = a + bt + ct²

or follow exponential-like growth:

H(t) = H0 ekt

The correct model depends on the condition being represented.

Traffic loading

Degradation may correlate better with cumulative train loading than calendar time.

Let cumulative traffic be M(t).

H = H(M)

A heavily trafficked segment can age faster than a lightly used segment over the same number of days.

Maintenance resets the process

Suppose rail grinding or another intervention occurs at time t_m.

Health indicator may drop:

H(tm+) < H(tm−)

Then degradation begins again from a new state.

This creates a saw-tooth process:

degrade
→ maintain
→ improve
→ degrade
→ maintain

The inspection system should learn not only how fast the track deteriorates, but how strongly each maintenance action resets condition.

Uncertainty band

A forecast should produce:

expected future health
+
confidence interval

For example:

H(t+30 days) = 0.72
90% interval = 0.58 to 0.91

These are fictional values.

If the upper tail approaches a maintenance boundary, engineers may inspect earlier even though the mean forecast remains acceptable.

This is why the 2026 review of track-geometry prediction evaluates uncertainty quantification as one of the major dimensions of model quality.

Prediction error

If predicted health is Ĥ and measured health later is H:

e = H − Ĥ

Mean absolute error is:

MAE = (1/n)Σ|ei|

Root mean squared error is:

RMSE = √[(1/n)Σei²]

RMSE penalises large forecast errors more strongly.

But a model with good average error can still miss rare local exceedances.

Evaluation should therefore include both:

continuous prediction accuracy
and
event-detection accuracy

The railway does not care whether a model predicts the average track beautifully if it misses the one location that matters.

Prompt 8 — How Does Inspection Decide Where Maintenance Should Go First?

Imagine 500 kilometres of track data and 2,000 anomaly locations.

Engineers cannot treat every observation as equally urgent.

Maintenance prioritisation combines severity, confidence, rate of degradation, consequence and engineering access.

For location i, define:

• s_i = anomaly severity,
• p_i = probability the anomaly is genuine,
• g_i = degradation rate,
• c_i = consequence if condition worsens,
• u_i = uncertainty,
• m_i = maintenance effort.

An illustrative priority score is:

Priorityi
=
pi si ci (1 + αgi)(1 + βui)
/
mi

This is not an operator formula.

It reveals the decision structure.

A mild but rapidly worsening anomaly on a high-consequence section may outrank a larger but stable anomaly elsewhere.

Risk matrix

Risk is often conceptualised as:

Risk = Probability × Consequence

This simple product is useful but incomplete.

Two conditions can have the same expected risk:

0.01 × 1000 = 10
0.50 × 20   = 10

but may deserve different responses because one is rare and severe while the other is common and modest.

Tail risk and safety constraints can therefore override expected-value ranking.

Inspection versus maintenance

An automatic system may recommend:

inspect physically
not
repair automatically

This distinction matters.

Sensor evidence can narrow the search.

Engineering inspection confirms the physical state and selects an authorised intervention.

LTA explicitly describes the NEL Automatic Track Inspection system as complementing existing physical inspections rather than replacing them.

That is a strong public design principle.

automation expands observation
human engineering retains physical verification and authority

Maintenance-window optimisation

Suppose intervention i requires duration d_i during engineering hours.

Binary variable x_ij indicates whether work i is assigned to window j:

xij ∈ {0,1}

A scheduling model can minimise:

expected degradation risk
+
maintenance mobilisation cost
+
service impact

subject to:

available engineering time
staff
machinery
worksite compatibility
safety rules
required service availability

Track inspection therefore connects directly to timetable construction.

The railway must find not only what needs work but when the system can safely do it.

The final output of track inspection is not an image, a waveform or a defect score. It is a justified engineering action placed at a real location and a real time.

A Complete Fictional Track-Inspection Example

Let us inspect a fictional 1-kilometre track section.

This example uses invented values and does not reproduce Singapore MRT inspection limits or maintenance thresholds.

Step 1 — Divide the track into segments

Use 10-metre analysis segments:

1,000/10 = 100 segments

Segment i has:

Si = [10(i−1),10i) metres

Step 2 — Collect geometry

Suppose each segment receives 100 gauge-deviation samples.

For Segment 42:

mean deviation = 0.18 normalised units
standard deviation = 0.31

Healthy peer segments average:

mean σ = 0.12
peer SD of σ = 0.06

Standardised variability score:

zgeometry
= (0.31 − 0.12)/0.06
≈ 3.17

The segment is unusual.

Step 3 — Add camera evidence

The vision system sees an abnormal fastener region with:

P(visual defect) = 0.82

But lighting quality is moderate:

qvision = 0.75

Quality-adjusted evidence is roughly:

0.82 × 0.75
= 0.615

This is not strong enough alone for confident physical diagnosis.

Step 4 — Add dynamic evidence

Two service trains produce elevated vibration at almost the same location.

Pass 1: zvibration = 2.4
Pass 2: zvibration = 2.7

Position estimates differ by 0.8 metres, inside the combined localisation uncertainty of the two passes.

The evidence is spatially consistent.

Step 5 — Build a composite evidence score

Use an illustrative score:

H
=
0.40 zgeometry
+ 0.30 zvision
+ 0.30 zvibration

Let z_vision after calibration equal 2.1 and average z_vibration = 2.55.

H
= 0.40(3.17)
+ 0.30(2.10)
+ 0.30(2.55)

= 1.268 + 0.630 + 0.765
= 2.663

The segment ranks well above typical background observations.

Step 6 — Compare across time

Historical composite scores:

Month 1: H = 0.8
Month 2: H = 1.1
Month 3: H = 1.5
Month 4: H = 2.0
Month 5: H = 2.66

The anomaly is not merely large.

It is growing.

A least-squares trend could estimate slope k.

Suppose:

H(t) ≈ 0.28 + 0.45t

where t is months from the first observation.

The positive slope provides additional evidence for physical inspection.

Step 7 — Prioritise inspection

Suppose the system estimates:

probability anomaly is genuine = 0.88
relative consequence score = 4/5
degradation urgency = 3/5
inspection effort = 1 unit

Illustrative priority:

Priority
= 0.88 × 4 × 3 / 1
= 10.56

Another anomaly might be visually stronger but stable and in a lower-consequence context.

The growing multi-sensor anomaly can therefore rank first.

Step 8 — Physical inspection returns

Engineers inspect the location.

Suppose they confirm a developing physical condition requiring maintenance.

The maintenance is completed.

Next inspection pass shows:

zgeometry falls
zvibration falls
vision anomaly disappears

The repair response confirms that the previous multi-sensor signature was physically meaningful.

measurement
→ anomaly
→ fusion
→ trend
→ prioritisation
→ physical inspection
→ maintenance
→ signal improvement
→ labelled evidence

That is the full track-inspection loop.

The Track-Inspection Deletion Tests

Remove precise location

The system knows something is wrong but cannot efficiently find it.

Remove measurement uncertainty

Sensor noise becomes false physical certainty.

Remove repeated passes

One-off anomalies cannot be separated as effectively from persistent degradation.

Remove local-event detection

Segment averages can hide a short but important defect.

Remove segment-level trends

The system finds individual defects but misses slow widespread deterioration.

Remove sensor fusion

Each sensor must make the decision alone.

Remove physical inspection

Statistical evidence becomes automatic diagnosis without engineering verification.

Remove maintenance feedback

The system never learns whether the anomaly disappeared after intervention.

Remove inspection of the sensors themselves

A drifting sensor can slowly rewrite the railway’s health map incorrectly.

A track-inspection system has to inspect two things at once: the railway and the reliability of its own eyes.

The Track-Inspection Paradoxes

Paradox 1 — More data can hide a defect

If millions of healthy observations dominate the analysis, rare defects can disappear statistically.

Paradox 2 — Smoother data can be less safe

Excessive smoothing can erase short local anomalies.

Paradox 3 — Better detection can initially create more maintenance work

The system begins discovering issues that previously remained invisible.

Paradox 4 — A highly accurate vision model can still be poor inspection

If it cannot localise defects, estimate uncertainty or work under changing tunnel conditions, its laboratory accuracy has little operational value.

Paradox 5 — The quietest track is not necessarily the healthiest track

Noise is one signal among many. Some defects may not create obvious audible symptoms.

Paradox 6 — A good average track-quality score can hide a dangerous local point

Aggregates protect against noise but can conceal extremes.

Paradox 7 — More frequent inspection can produce fewer disruptions

Continuous monitoring may increase the number of observations and maintenance interventions while reducing emergency events.

Paradox 8 — Automation makes human inspection more valuable

When automation narrows millions of track points to a small number of credible locations, engineers can spend scarce physical-inspection time where it matters most.

How Track-Inspection Mathematics Grows from School to Research

Primary Mathematics

• distance and scale,
• measurement,
• averages,
• tables and graphs.

Secondary Mathematics and Science

• standard deviation,
• rates of change,
• coordinate geometry,
• waves and frequency,
• probability,
• classification.

Junior College

• calculus and curvature,
• Fourier analysis concepts,
• probability distributions,
• Bayesian reasoning,
• regression and optimisation.

University and Research

• signal processing,
• state estimation,
• computer vision,
• machine learning,
• sensor fusion,
• spatial statistics,
• reliability engineering,
• track-degradation modelling,
• maintenance optimisation.

The railway again forces these subjects to work together.

A camera pixel becomes a probability.

A probability becomes a point on a map.

The point becomes a trend.

The trend becomes a maintenance priority.

The maintenance priority becomes a real engineering job during a finite maintenance window.

The World Return — When the Physical Track Answers the Inspection Model

An automatic system flags a location.

Engineers inspect it.

The physical railway returns one of several answers.

True positive

model flags condition
→ physical inspection confirms it

The observation becomes labelled evidence.

False positive

model flags condition
→ inspection finds no relevant defect

The system asks whether the error came from:

• camera lighting,
• sensor calibration,
• position mismatch,
• model threshold,
• or an inspection gap.

False negative

physical inspection finds defect
→ automatic system did not flag it

This is especially valuable evidence because it reveals a blind spot.

Was the feature too small?

Was sampling insufficient?

Was the sensor looking at the wrong modality?

Was the model trained on too few examples?

Was smoothing too aggressive?

Maintenance response

After maintenance, the same location is measured again.

If health indicator improves:

Hafter < Hbefore

the intervention has produced measurable physical return.

If it does not:

either the maintenance did not address the source, or the health indicator was not measuring what engineers thought it was measuring.

The complete loop is:

inspect automatically
→ locate anomaly
→ estimate confidence
→ inspect physically
→ maintain if justified
→ inspect again
→ compare before and after
→ update model

LTA’s description of Automatic Track Inspection complementing physical inspection is therefore exactly the right systems architecture.

Automation expands how much of the track can be observed.

Physical engineering closes the evidence loop.

The final judge of a track-inspection algorithm is not its dashboard. It is whether the physical rail behaves the way the model said it did.

RFE Return — What Does Good Track Inspection Owe the Passenger?

The passenger should rarely need to think about rail geometry.

That is the point.

Good inspection should produce an almost invisible public result:

smooth ride
+
reliable service
+
fewer avoidable defects reaching operations
+
better planned maintenance
+
less emergency intervention
+
continued safe use of the infrastructure

But it also owes restraint.

A public inspection system should not expose sensitive local vulnerability maps, operational thresholds or detailed failure conditions.

Its public educational layer should explain:

• what kinds of measurements exist,
• why uncertainty matters,
• how automatic inspection complements people,
• and why repeated evidence enables timely maintenance.

The RFE is therefore not:

find every tiny difference in the track

It is:

detect the differences that matter early enough, accurately enough and responsibly enough that engineers can act before the railway’s condition becomes the passenger’s problem.

Conclusion — The Rail Is Becoming a Measured Landscape

For much of railway history, inspection meant people walking, looking, measuring and listening.

That work remains essential.

But modern MRT systems can now add another layer.

Trains themselves can become inspection vehicles.

Every journey can produce another observation of the infrastructure beneath it.

A camera sees a fastener.

A geometry sensor measures rail position.

An accelerometer feels the vehicle response.

A positioning system places the evidence on the map.

A statistical model asks whether it is unusual.

A time-series model asks whether it is worsening.

A fusion model asks whether other sensors agree.

A maintenance model asks whether it should be inspected now or watched longer.

An engineer returns to the physical rail and decides what is actually there.

track
→ measurement
→ spatial coordinate
→ noise filtering
→ anomaly
→ multi-sensor evidence
→ health map
→ degradation forecast
→ inspection priority
→ physical verification
→ maintenance
→ new measurement

Singapore’s North East Line Automatic Track Inspection systems and SMRT’s Rail Vision represent this shift already: the railway is increasingly able to observe itself during normal operation rather than relying only on isolated inspection moments.

The mathematics makes the difference between one image and a health history.

Between one vibration and a spatial trend.

Between an unusual pixel and an engineering priority.

The passenger sees two steel rails.

The inspection system sees millions of samples.

The statistician sees distributions.

The computer-vision model sees objects and surfaces.

The maintenance planner sees a map of changing risk.

The engineer sees the real rail.

And if all of them work properly, the passenger sees almost nothing at all.

The best track defect is the one that becomes a maintenance job while it is still only a pattern in the data.

Key Equations

ΔG(x) = G(x) − G0
Gauge deviation

ealignment(x) = y(x) − y0(x)
Alignment deviation

C(x) = zL(x) − zR(x)
Cross level

Twist(x;L) = [C(x+L) − C(x)]/L
Track twist over a base length

κ(x) = y″(x)/[1+y′(x)²]3/2
Planar curvature

x(t) = ∫v(t)dt
Map time-domain measurements to position

Δx = v/f
Spatial sample spacing

fs > 2fmax
Nyquist sampling condition

f = v/λ
Spatial wavelength to temporal frequency

y(x) = h(x) + ε(x)
Measurement = physical state + noise

Y(f) = ∫y(t)e−i2πftdt
Fourier transform

Eband = ∫|Y(f)|²df
Spectral energy in a band

xk = Axk−1 + wk
yk = Hxk + vk
Generic state-space model

Precision = TP/(TP+FP)

Recall = TP/(TP+FN)

F1 = 2PR/(P+R)
Classification performance

P(D|C,V,G) ∝ P(C,V,G|D)P(D)
Multi-sensor Bayesian evidence

H(x) = Σwj zj(x)
Composite health score

wj,effective = qjwj
Sensor-quality weighting

|x1−x2| ≤ k√(σ1²+σ2²)
Illustrative spatial-consistency condition

P(at least one detection) = 1 − (1−p)n
Repeated observation

TQIi = Σwjfj(i)
Conceptual segment health index

σsegment = √[(1/n)Σ(yi−ȳ)²]
Segment geometry variability

H(t) = H0 + kt
Simple degradation model

MAE = (1/n)Σ|ei|

RMSE = √[(1/n)Σei²]
Forecast error measures

Risk = Probability × Consequence
Basic risk structure

Priorityi = pi si ci(1+αgi)(1+βui)/mi
Illustrative inspection-priority score

Reader-safety note: All fictional thresholds, health scores, sensor rates, defect examples and maintenance rankings in this article are educational abstractions. They do not reproduce Singapore MRT inspection limits, engineering tolerances, defect-location data, security-sensitive infrastructure maps or authorised maintenance procedures. Actual railway inspection uses validated asset-specific standards, calibrated equipment, confidential engineering data and professional judgement.

Continue the MRT Mathematics Series

• How MRT Timing Works Using Mathematics — the system-wide pillar.
• How MRT Braking Works Using Mathematics — stopping distance and regenerative braking.
• How MRT Headway Works Using Mathematics — train spacing and throughput.
• How MRT Station Dwell Time Works Using Mathematics — queues, doors and platform flow.
• How MRT Delays Propagate and Recover Using Mathematics — instability, regulation and recovery.
• How an MRT Timetable Is Built Using Mathematics — demand, train cycles and robustness.
• How MRT Energy Use Is Optimised Using Mathematics — traction, coasting and regenerative energy.
• How MRT Passenger Capacity Is Calculated Using Mathematics — vehicle, station and delivered capacity.
• How MRT Routes and Transfers Are Optimised Using Mathematics — graph theory, transfer cost and route choice.
• How MRT Network Resilience Is Measured Using Mathematics — redundancy, cascading overload and recovery.
• How MRT Predictive Maintenance Works Using Mathematics — anomaly detection, failure probability and maintenance timing.

The next natural leg is How MRT Wheel–Rail Contact Works Using Mathematics: contact geometry, forces, friction, wheel conicity, curves, wear, vibration and why two pieces of steel can guide a train through an entire city with remarkably little rolling resistance.

Sources and Further Reading

• Land Transport Authority — North East Line Automatic Track Inspection, cameras and sensors
• SMRT — Rail Vision, daily track inspection and rail grinding
• Land Transport Authority — new Circle Line trains and condition monitoring
• Results in Engineering — Railway track geometry degradation: review of prediction methods, 2026
• International Journal of Transportation Science and Technology — Machine-vision railway defect inspection review
• Journal of Industrial Information Integration — Review of AI in railway-track monitoring, inspection and maintenance
• Neural Computing and Applications — Time-series data mining for railway wheel and track monitoring
• Railway Engineering Science — Recent advances in wayside railway condition monitoring, 2026