Smart Water Quality Sensor Prediction Project

ISEF Category: Embedded Systems

Subcategory: Sensors  ·  Difficulty: Advanced  ·  Setup: School Lab  ·  Time: Full Year

The Hook

Cheap water sensors often disagree with each other. That is not a bug, it is the whole problem. If you can teach a tiny controller to read those messy signals together, you can turn shaky numbers into a useful safety estimate. That skill matters in places where lab testing is too slow or too expensive.

What Is It?

This project asks a simple question with a tricky answer, can cheap sensors work together well enough to estimate drinking-water safety? You measure things like pH, electrical conductivity (EC), and turbidity. pH tells you how acidic or basic the water is. EC gives a clue about dissolved ions. Turbidity measures how cloudy the water is.

Think of each sensor like one witness at a noisy scene. Each one sees part of the story, but none of them sees everything. A multitask neural network is a model that learns from several outputs at once. Instead of treating each sensor separately, it looks for patterns across all of them and adjusts for cross-sensitivities, which are cases where one sensor responds to more than one water property.

The embedded system part means the model runs on a microcontroller or similar small device, not just on a laptop. That makes the project stronger, because you are testing both the sensing hardware and the edge AI logic that has to work in real time.

Why This Is a Good Topic

This is a strong science fair topic because you can test it with real measurements, compare it against known water samples, and quantify how much the model improves predictions. It connects to drinking water safety, low-cost monitoring, and rural or emergency settings. You can learn sensor calibration, data cleaning, feature selection, and model evaluation, all without needing a full research lab.

Research Questions

• How does combining pH, EC, and turbidity improve water safety prediction compared with using one sensor alone?
• What is the effect of calibration set size on the model's prediction accuracy for unknown water samples?
• Does adding a multitask neural network reduce error from cross-sensitivity more than a simple regression model?
• To what extent does sensor drift over repeated trials change the final safety classification?
• Which sensor combination gives the best tradeoff between cost, accuracy, and computation on a microcontroller?
• What is the effect of water source type, such as tap water, pond water, or filtered water, on model generalization?
• To what extent does training on mixed-sample blends improve performance on new water samples?

Basic Materials

• Microcontroller board with analog inputs, such as Arduino Nano or ESP32.
• Capacitive or analog pH sensor module.
• Electrical conductivity sensor module.
• Turbidity sensor module.
• Breadboard and jumper wires.
• USB cable for programming and power.
• Digital multimeter.
• Clean sample cups or beakers.
• Distilled water for baseline samples.
• Source water samples with known differences, such as tap water and filtered water.
• Notebook or spreadsheet for recording measurements.
• Safety gloves and eye protection.

Advanced Materials

• Microcontroller with enough memory for on-device inference, such as ESP32-S3.
• Higher-quality pH probe and interface circuit.
• Lab-grade conductivity probe and meter for comparison data.
• Turbidity reference standards or calibrated nephelometric setup.
• Reference instruments or access to a water testing lab for ground-truth labels.
• Temperature sensor for compensation.
• SD card module or data logger.
• Stable power supply.
• Shielding or enclosure for sensor electronics.
• Reference software for model training and conversion to embedded format.

Software & Tools

• Python: Cleans sensor data, trains models, and compares prediction methods.
• Google Colab: Lets you train and test models in a browser without local setup.
• Arduino IDE: Programs the microcontroller and reads live sensor data.
• TensorFlow Lite for Microcontrollers: Helps run a small neural network on embedded hardware.
• ImageJ: Can help if you use turbidity images or color-based calibration in a side experiment.

Experiment Steps

1. Define your prediction target, such as a safety class, a water quality score, or a set of measured reference values.
2. Choose the sensor inputs you will compare, and decide whether you will test single-sensor, dual-sensor, or full-fusion models.
3. Plan a reference measurement strategy so you can label each water sample with trusted ground truth data.
4. Build a calibration set that covers multiple water conditions, then reserve separate samples for testing generalization.
5. Select a model structure that can run on a microcontroller, and decide how you will measure accuracy, drift, and latency.
6. Design controls that check whether the model really handles cross-sensitivity better than a simpler baseline.

Common Pitfalls

• Using only one water source for training, which makes the model fail on new samples.
• Skipping a trusted reference method, which leaves you with sensor readings but no real ground truth.
• Mixing up sensor drift with model error, which hides whether the hardware or the algorithm caused the mistake.
• Training a model that looks accurate on the same samples it already saw, which inflates performance.
• Ignoring temperature and probe conditioning, which can shift pH and conductivity readings enough to ruin comparisons.

What Makes This Competitive

A stronger project will compare more than one modeling approach and report where each one breaks down. You can make it stand out by testing the model on water sources it never saw during training, then analyzing which sensor combinations help most. Clear calibration, honest error analysis, and a real embedded deployment matter more than flashy code. If you can show the system still works after sensor noise or drift, your project starts to look like real engineering.

Project Variations

• Test the same sensor fusion idea on irrigation water, not drinking water, and predict salinity-related quality categories.
• Swap the neural network for a simpler model, such as random forest or linear regression, and compare accuracy against the embedded approach.
• Add a temperature sensor and study whether compensation improves predictions across cold, room-temperature, and warm samples.

Learn More

• NIH PubMed: Search for review articles on water quality sensors, sensor fusion, and machine learning in environmental monitoring.
• USGS Water Science School: Read plain-language explanations of water chemistry, turbidity, and conductivity measurements.
• NOAA National Water Prediction Service: Look for background on water monitoring, environmental data, and measurement context.
• NASA Earthdata: Search for remote sensing and water-quality related data resources that help frame monitoring problems.
• MIT OpenCourseWare: Find free classes on machine learning, embedded systems, and signal processing.
• US EPA Drinking Water page: Review public guidance on drinking-water indicators and common contaminants.