TinyML - EMBEDDED AI

Our smartphones have an inertial measurement unit (IMU) which groups several sensors: the accelerometer (linear accelerations), the gyroscope (rotation speed) and the magnetometer (magnetic field). In this project, the main sensor is the accelerometer, which measures acceleration on 3 axes (X, Y, Z) in m/s² or in g.

Since we are going to solve a classification problem associated with smartphone movements here, we create two distinct classes: Shake vertically and Shake horizontally.

To obtain a performant model, acquiring a sufficient volume of data (the dataset) is crucial. We perform 4 acquisitions of 30 seconds for each class.


Signal acquisition starts automatically on the smartphone. At the end, we obtain curves showing the time-evolution of measured accelerations along the three directions.

After acquiring all data, we divide them into three sets: the Training Set (75%) to train the AI, the Validation Set to fine-tune the estimator parameters, and the Testing Set (25%) to evaluate final predictions on unseen data.


In order to transform the continuous signal into usable data, we use the sliding window method. This principle consists of cutting the signal into fixed-size segments. If the shift step is smaller than the window size, the overlap allows generating more samples and avoids missing a gesture between two cuts.

Before training, we extract spectral features (FFT). Analyzing the spectrum helps identify dominant frequencies. For example:
- Rotation: Significant energy between 2-4 Hz with a single peak.
- Walking: Energy at 1-2 Hz, regular rhythmic peak.
- Fall: Broadband signal (Energy across all frequencies).
- Rest: Near-zero energy.


These extracted features are then visualized as clusters of different colors in the Feature Explorer. If the points form well-separated groups, it means our features allow distinguishing movements effectively.

We configure training with 30 epochs (the model will see all data 30 times) and a learning rate of 0.0005. At each cycle, it adjusts its weights to minimize classification error.

The neural network architecture includes an input layer, 2 hidden layers of 20 neurons and 10 neurons respectively to extract non-linear patterns, and an output layer handling final classification.

It is at this stage that Edge Impulse uses, among others, a gradient descent algorithm to adjust weights and biases associated with each artificial neuron.

The model here has a precision (accuracy) of 100%. This means that 100% of samples from training data had their output class correctly predicted.

To go more in detail, we can observe the confusion matrix:
- In columns, the classes predicted by the algorithm.
- In rows, the classes known because they are labeled in the training data.

The Data explorer allows us to find samples for which the prediction was wrong. Here, we can see that no sample had its class mispredicted.

We obtain a performance summary similar to the previous one. Logically, performance is slightly less satisfactory than on the training dataset (recall that the algorithm had "never seen" the testing data).

In this example, the accuracy obtained, above 97.60%, is satisfactory.

Here the Feature explorer allows us to find samples for which the prediction was wrong. We can see that 6 samples had their class mispredicted.

We now enter the use phase (inference) of the AI algorithm which was previously trained and tested.

First option with Edge Impulse: "Live Classification" – suitable for quick model testing. Here I performed a vertical movement during inference and we can see that the model predicted more vertical movement than horizontal movement.

At the end of the acquisition, the measurement appears on the computer screen. We find the split into multiple samples with the sliding window method, which brings multiple samples for a single measurement. For each sample, the algorithm performs a prediction of membership to one of the classes. This prediction is characterized by a probability (between 0 and 1) of belonging to each class.

Deploy my created AI model to my phone and use it as a WebApp to detect movements.