E-Scooter Bearing Wear Monitor Using Vibration AI

ISEF Category: Engineering Technology: Statics and Dynamics

Subcategory: Ground Vehicle Systems  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A worn bearing can sound fine for a while, then fail fast. That makes early warning a big deal for any small electric vehicle. You can turn tiny vibrations into a health check for an e-scooter hub motor. This project mixes sensors, signal processing, and machine learning in one clean test.

What Is It?

This project tracks the vibration pattern of an e-scooter hub motor and tries to spot bearing wear before a full failure. Think of the bearing like a wheel on a shopping cart. When the wheel gets rough, the cart still rolls, but the motion gets noisy. A healthy bearing makes one vibration pattern. A worn bearing makes another.

An accelerometer is a sensor that measures motion changes. You mount it on the motor housing and record vibration data while the motor runs. Then you turn that raw signal into features, like Mel-frequency cepstral coefficients (MFCCs), which compress the sound or vibration pattern into numbers a model can read. A one-dimensional convolutional neural network, or 1D-CNN, looks for shape patterns in those numbers and learns to classify healthy versus worn conditions.

Why This Is a Good Topic

This is a strong science fair topic because you can test one clear idea, whether vibration data can reveal bearing wear early. You can measure accuracy, compare feature sets, and check if the model still works on real hardware after training on public datasets. The topic connects to electric bikes, scooters, factories, and any machine that needs preventive maintenance. You can learn signal processing, model validation, and sensor design without building a full vehicle from scratch.

Research Questions

• How does bearing surface roughness affect accelerometer-based classification accuracy for an e-scooter hub motor?
• What is the effect of using MFCC features instead of raw vibration features on early wear detection?
• Does a 1D-CNN outperform a simpler classifier, such as logistic regression or random forest, on the same vibration data?
• To what extent does training on public bearing datasets transfer to a real e-scooter motor with new operating conditions?
• Which sensor mounting position gives the most reliable wear signal on the motor housing?
• How does motor speed change the model's false positive rate when the bearing condition stays the same?

Basic Materials

• E-scooter hub motor or similar small brushless motor.
• Accelerometer sensor module with data logging capability.
• Microcontroller board with enough memory for local inference, such as an ESP32 or Arduino-compatible board.
• Stable motor mount or test rig.
• Power supply matched to the motor.
• Smartphone or camera for documenting setup.
• Laptop for data analysis and model training.
• Safety goggles and hand protection.
• Calipers or a feeler gauge for documenting wear changes.
• Reference healthy bearing and a way to compare roughened conditions.

Advanced Materials

• E-scooter hub motor with removable bearing assembly.
• Triaxial accelerometer with higher sampling rate.
• Data acquisition hardware with synchronized timestamping.
• Vibration isolation base and rigid mounting fixtures.
• Tachometer or encoder for speed control.
• Bearing damage characterization tools, such as microscope imaging or surface roughness measurement.
• Access to a machine shop or bearing press tools.
• Embedded microcontroller with on-device inference support.
• Labeling and validation set from real motor trials.
• Reference samples from MAFAULDA and CWRU datasets for transfer learning.

Software & Tools

• Python: Cleans sensor data, extracts features, and trains the classifier.
• Jupyter Notebook: Lets you compare models and plot vibration patterns in one place.
• TensorFlow Lite: Helps you test compact on-device inference on a microcontroller.
• scikit-learn: Builds baseline classifiers and checks whether the neural network adds value.
• ImageJ: Measures bearing surface changes if you document wear with microscope images.

Experiment Steps

1. Define the failure state you want to detect, such as healthy, lightly worn, and clearly worn bearings.
2. Choose one sensor placement and one motor operating condition set so your data stays comparable.
3. Build a labeling plan that combines public bearing datasets with your own motor measurements.
4. Decide whether to compare raw signals, MFCC features, or both before training the model.
5. Set up a validation plan that tests the model on hardware data it never saw during training.
6. Plan an embedded deployment check so you can measure whether the model fits on the microcontroller and runs fast enough.

Common Pitfalls

• Mixing public dataset records with your own motor data without tracking domain shift, which can make the model look better than it really is.
• Letting the accelerometer move between trials, which changes the vibration signal more than the bearing condition does.
• Using only one worn bearing sample, which teaches the model one part instead of a general wear pattern.
• Testing on the same motor speeds used for training, which hides speed-related false alarms.
• Calling any noisy bearing a failure case, which makes the labels too vague for a real classifier.

What Makes This Competitive

A stronger project does more than report a high accuracy score. You would test whether the model still works when the motor speed changes, when the bearing wear is mild, and when the data comes from a different source than the training set. You could compare raw signals against MFCC features, then add a baseline model to prove the 1D-CNN earns its complexity. A polished entry also explains failure modes, not just success cases.

Project Variations

• Use a bicycle hub motor instead of an e-scooter motor to compare how frame mounting changes vibration signatures.
• Skip MFCC features and test whether wavelet features or summary statistics give better early wear detection.
• Add transfer-learning analysis by training on public datasets first, then fine-tuning on your own motor data.

Learn More

• CWRU Bearing Data Center: Search the Case Western Reserve University bearing dataset and read the fault examples and metadata.
• MAFAULDA Dataset: Search the bearing dataset from University of Ottawa for multi-fault vibration records.
• NASA Prognostics Center of Excellence: Find reliability and remaining-useful-life resources for condition monitoring projects.
• NIST Time Series Analysis resources: Search NIST for guidance on signal processing and model evaluation.
• MIT OpenCourseWare: Search for introductory machine learning, signal processing, and control systems materials.