Final year project · FAST NUCES Karachi
Low-cost sensor nodes read a building continuously. A cloud risk engine turns what they read into one health index and five plainly worded states, each with an action and a timeframe. An augmented reality view puts every reading back on the wall it came from.
8 measurements · 6 modules · 3 roles in one app
Why we are building this
The standard for multi-storey buildings in Pakistan is periodic manual inspection. An engineer visits, looks at the building, and files a report. Between two visits, which can be years apart, nobody is watching it at all.
The building in Lyari had been declared dangerous by the Sindh Building Control Authority in 2023 and had been served evacuation notices, yet people were still living in it on the morning it fell. Nobody could tell a building with years of life left from one with days, because nothing was measuring either of them. Sources.
What existing work misses
Low-cost boards have already been validated against commercial instruments for vibration, tilt and crack measurement. What does not exist is a deployable system that carries those readings all the way to the people who are supposed to act on them.
Existing systems output accelerometer traces, frequency spectra and strain values. A building manager is not a structural engineer and cannot act on a spreadsheet. The data has to become a status, an action and a timeframe.
Most published systems watch a single quantity, usually vibration or cracks. Buildings fail through several slow processes acting together, and those only make sense when they are read together over months.
No low-cost system connects structural readings to the management, the maintenance staff and the residents. Lyari is exactly this failure: the information existed at the authority level and never reached the building.
How it works
Nothing in the chain is exotic. Cheap boards take the readings, one small computer in the building collects them, and the work of turning numbers into advice happens in the cloud where the history lives.
01
Low-power nodes are fixed to columns, walls, risers and tanks. Each one wakes on a schedule, takes its reading, sends it over the local network and goes back to deep sleep, which is what makes months of battery life possible.
ESP32, ESP32-CAM, LiDAR02
One gateway per building collects all node traffic, drops corrupt packets, timestamps every reading, and buffers locally whenever the internet goes down. When the link returns it uploads the backlog.
Raspberry Pi03
Readings are validated, stored against each sensor's learned baseline, then run through threshold rules, trend fitting and anomaly detection at the same time. The three scores fuse into one health index and five states.
Time series database04
A deep learning model outlines the crack in each camera image, post-processing measures its width in pixels, and the laser distance to that surface converts the figure into millimetres.
Segmentation network05
The camera and laser nodes together give a measured model of the room, so every reading can be anchored to the surface it came from and read back through the phone camera on site.
Ordinary smartphone06
One application with three role-based views for management, the site team and residents, with critical alerts pushed straight to whoever is on call.
Web and mobileWhat it senses
Each sensor answers a different question, and none of them mean much on their own. Read together and over months, they show whether a building is settling into a new normal or drifting away from one. The camera and the laser sensor come first, because everything else is easier to place once the building has been seen properly.
ESP32-CAM node
A camera fixed in front of a column, a joint or a crack, photographing the same spot on a schedule so today's image can be compared with last month's.
Phase 1LiDAR time of flight node
Millimetre distances to walls, columns and ceilings. Room geometry that can be measured again later to see what has moved.
Phase 1MEMS accelerometer node
How the building moves and how quickly it settles. A shift in natural frequency is a well established early damage sign and shows up long before anything is visible.
Phase 2MEMS inclinometer node
Differential settlement tracked to a hundredth of a degree, following low-cost inclinometer designs already validated against commercial instruments.
Phase 2Camera and LiDAR fused
The model extracts the crack from the image, and the laser distance supplies the physical scale that converts pixels into millimetres.
Phase 2Handheld survey
Pulse speed through concrete, which indicates internal voids and material quality. Survey readings are entered into the same database.
OptionalLevel and flow node
Sump levels, tank levels and riser flow, so a blockage is caught while it is still just a blockage.
OptionalCurrent clamp node
Current clamp readings per distribution board. An unusual draw usually means a failing motor or a fault worth a look.
OptionalChosen from low-cost parts already validated in the literature
| Device | Role in the system | Phase |
|---|---|---|
| ESP32 DevKit boards | Core microcontroller for all sensor nodes | 1 |
| ESP32-CAM modules | Scheduled imaging of columns, joints and cracks | 1 |
| LiDAR ToF sensors, VL53L1X or TF-Luna class | Millimetre room geometry and crack width scale | 1 |
| Raspberry Pi 4 or 5, with storage | Per-building site gateway with local buffering | 1 |
| MEMS accelerometers, ADXL345 or ADXL355 class | Vibration and natural frequency monitoring | 2 |
| MEMS inclinometers and IMU modules | Tilt and differential settlement tracking | 2 |
| Batteries, enclosures, mounting hardware | Long-term unattended node deployment | 1 to 2 |
| Ultrasonic pulse velocity tester, shared | Handheld concrete quality surveys | Optional |
| Water level and flow sensors | Sump, tank and riser monitoring | Optional |
| Non-invasive current clamp sensors | Distribution board load monitoring | Optional |
Augmented reality · Phase 1
The Phase 1 camera and laser nodes together produce a measured model of the room, which means every reading in the database can be anchored to the physical spot it came from.
A technician walks up to a column, holds up the phone, and the crack in front of them carries its own history: how wide it was in March, how wide it is now, and what the last inspector wrote down.
Crack detection model
Crack width is the only reading in the suite that cannot be taken directly off a sensor. We follow the encoder and decoder segmentation approach, because pixel-level output is what a width measurement needs.
01
A 224 by 224 patch cut from the ESP32-CAM image.
02
Convolutional blocks of 64, 128 and 256 filters downsample the patch, with a 512-filter bottleneck at the lowest resolution.
03
The decoder upsamples back to full resolution, and a 1 by 1 convolution with a sigmoid produces a binary crack mask.
04
Post-processing extracts the crack skeleton and measures the perpendicular width, then the LiDAR distance to that surface supplies the scale.
A hairline crack can be one or two pixels wide, so skip connections carry fine spatial detail from each encoder block across to its matching decoder block. Without them the crack disappears in the downsampling.
Training starts on public concrete crack datasets and is then fine-tuned on images from our own instrumented column, where the camera position, lighting and wall surface are fixed and known.
Management view
Building managers are not structural engineers, and they should not have to be. Every reading resolves to one of five states, and every state comes with something to do and a timeframe to do it in.
Building health index
82out of 100
Down 4 points since the March survey
If the north facade keeps moving at the current rate, this reaches 74 by June. That is enough notice to budget for the work instead of reacting to it.
Offline is grey, not red. A sensor that has lost its connection is a maintenance job, and colouring it like a structural risk teaches people to ignore red.
The app
The same building looks different depending on who is asking. Everyone signs into the same system and sees the part of it that is theirs.
Owners and building managers
Technicians, guards and cleaners
The people who actually live there
The long view
The final year project is the first slice of something larger. These are the things we are building towards, so that the decisions we make this year still hold up later.
The first deployment watches one instrumented structure. The same gateway and the same database should carry a second building, then a street of them, without anything being rebuilt.
A crack measured every day for a year tells you when to act. The system should say what is likely to need money in six months, while there is still time to plan for it.
Structure, drainage, power and people all sit in the same building, so they should sit in the same app. Splitting them across four systems is how things get missed.
Nobody managing a building wants a spreadsheet of accelerometer output. They want to know what is fine, what needs watching, and who to call today.
Roadmap
A single year is not long enough to build all of this well. Rather than promise the whole vision by June, the scope is split into what we have committed to, what we plan to reach, and what we will attempt if the first three phases land on time.
Literature survey of low-cost IoT structural health monitoring, vision-based crack measurement and AR inspection. The sensing suite is finalised and Phase 1 hardware is procured.
ESP32-CAM, LiDAR and gateway in handCamera and LiDAR nodes running on one instrumented column, the gateway collecting on site, and the AR proof of concept pinning readings to their physical location.
Demo: point a phone at a column and read itVibration, tilt and crack width sensing added on the same gateway, the first trained version of the crack segmentation model, and cloud ingestion with threshold rules and the alert path wired into the management dashboard.
Demo: an alert fires from sensor to screenHealth index and trend projections computed over the collected history, risk scoring per subsystem, the crack model fine-tuned on our own site images, and the mobile app with all three role views.
Demo: a generated risk report and a service requestEnd-to-end testing of the complete pipeline on the instrumented structure, then final reporting.
Demo: the whole chain running unattendedUltrasonic pulse velocity surveys, drainage and power monitoring, and support for more than one building.
Not guaranteed. Depends on procurement and time.Proposal submitted 25 Aug 2026 · currently in Month 1
Team
Final year project at FAST NUCES Karachi, supervised by Dr. Farrukh Shahid.
23K-0927
23K-0705
23I-0124
Supervisor
Project title
AIoT-Based Autonomous Structural Health Monitoring and Integrated Building Management Platform Using Low-Cost Sensor Fusion and Augmented Reality