In the world of AI and machine learning, cameras play a big role in data collection and inference. However, a significant section of industrial automation still relies on non-image sensors such as accelerometers, IMUs, and pressure sensors. These sensors typically generate time series data, which can be used to detect patterns, predict failures, and optimize performance.
At Embedded World 2025, I showcased a demo that highlights how Edge Impulse can work with time series data to train efficient models that run on low-power MCUs. This demo features the Nordic Thingy:53, which detects vibration patterns and anomalies in a DC motor setup. Let's dive into how it works!
Hardware Setup
The demo consists of the following components:
- Edge Impulse — used to collect sensor data, train a machine learning model, and deploy it to the device
- Nordic Thingy:53 — a powerful embedded device with built-in sensors, including an accelerometer
- Arduino Board with a Motor Driver — generates PWM signals to control a 12V DC motor. Romeo v2.2 R3
- DC motor with the gearbox removed
- A trolley wheel
- A prototyping plate square
The Thingy:53 is not physically connected to the motor circuit. Instead, it passively detects vibrations from the motor using its built-in accelerometer, allowing for an independent and non-intrusive monitoring system.
How It Works
The goal of this demo is to classify different motor vibration patterns and detect anomalies when something is off.
1. Data Collection & Training
I first collected time series accelerometer data from the Thingy:53 while the motor ran at different patterns. Using Edge Impulse, I:
- Labeled the data according to different motor speed patterns
- Trained a machine learning model to classify the patterns
- Deployed the trained model back to the Thingy:53 for real-time inference
2. Pattern Recognition with LED Indicators
Once the model was deployed, the Thingy:53 could classify different motor patterns based on vibration data. To visualize the classification:
- Pattern 1 → Blue LED
- Pattern 2 → Amber LED
- Pattern 3 → Green LED
As the motor stabilizes in a pattern, the corresponding LED illuminates, confirming the correct classification.
3. Anomaly Detection
Anomalies occur when the motor exhibits vibrations that do not match any trained patterns. To demonstrate this:
- I attached a small piece of blue tack to the motor to simulate an unbalanced load, mimicking real-world industrial failures (e.g., worn-out belts or misaligned pulleys)
- With the anomaly present, the LED turned red, and the anomaly score increased, indicating an irregular vibration pattern
- Removing the blue tack restored normal operation, with LEDs switching back to their expected colors
Industrial Use Cases
This demo represents a low-power, real-time solution for industrial condition monitoring. Possible applications include:
- Predictive Maintenance — Detecting early signs of mechanical wear before failures occur
- Quality Control — Ensuring machinery operates within expected vibration parameters
- Safety Monitoring — Identifying anomalies in factory equipment to prevent hazards
Edge Impulse enables engineers to train and deploy tinyML models on low-power devices, bringing AI-driven insights to industrial automation.
This demo illustrates how machine learning on non-image sensors can provide valuable insights for industrial applications. Using Edge Impulse and Nordic Thingy:53, we successfully classified motor vibration patterns and detected anomalies, all running on a low-power embedded device.
If you're interested in building similar solutions, check out Edge Impulse and start deploying machine learning models on embedded hardware today!
🚀 Let’s bring intelligence to the edge! Sign up for your free Developer Plan account today.