Skatelligence

The mechanical design of Skatelligence V2 centers around a central module housed in a custom 3D-printed enclosure. This enclosure contains the core components of the system, including the custom PCB, battery/BMS, and one IMU. The housing is mounted to the skater's lower back using velcro straps, and is designed shell is designed for durability and wearability, with accessible ports for charging the internal battery and reflashing the microcontroller when needed. Four JST connectors are contained in the central housing, providing wired connections to the external sensor modules—two mounted on the skater's wrists and two on the skates.

The electrical system of Skatelligence V2 is centered around a custom-designed PCB featuring an STM32 microcontroller. This microcontroller communicates with five MPU6050 IMUs using I2C, routed through a TCA9548A I2C multiplexer due to the limited I2C channels on the microcontroller. Power is supplied by a 1S lithium-ion battery, housed within the central module. To meet the voltage requirements of the system, the battery output is boosted to 5V before being delivered to the microcontroller, which steps it back down to 3.3V to power the other components. Charging is handled by an onboard TP4056 module, which provides safe and efficient battery management. This configuration enables a compact yet robust electrical setup, supporting reliable operation throughout extended skating sessions.

The embedded system on Skatelligence V2 is built around an STM32 microcontroller and programmed using FreeRTOS to manage concurrent tasks. The firmware is responsible for initializing communication with the IMUs, establishing BLE or communication, and handling real-time sensor data collection, storage, and transmission. During operation, the system polls five MPU6050 IMUs. Sensor data is collected at a frequency of 100 Hz, with each reading capturing three axes of linear acceleration and three axes of angular velocity per sensor. These readings are stored as 16-bit integers and written to internal card in 6KB binary files, capturing 1 second of recording. FreeRTOS allows for multithreading, enabling the system to record and transmit data simultaneously. One core is dedicated to collecting and storing IMU readings in real-time, while a second core runs handles the asynchronous data transmission over BLE to a local server. If no connection is available, files continue to be stored locally and are uploaded once connectivity is restored.

The data processing pipeline in Skatelligence V2 transforms raw IMU readings into clean, interpretable data that can be used for analysis and jump classification. After the embedded system logs IMU data as binary files, these files are passed through a multi-step processing routine. The first step involves reading and scaling the raw binary data to physical units—acceleration in Gs and angular velocity in degrees per second. Once scaled, a sixth-order Butterworth low-pass filter is applied to each data stream to remove high-frequency noise and produce smoother signals for analysis. This filtering ensures that only meaningful motion patterns are retained, which is critical for reliable jump detection. The filtered data is then visualized using a custom-built PyQt-based GUI. The interface supports both raw and processed data views, displaying all six measurement axes (3 acceleration + 3 gyroscopic) for each of the five sensors. Users can scroll through data files, track the latest recordings in real time, and select and examine detected jumps. Jump detection is performed using a state-machine approach on the torso's vertical acceleration signal. The system identifies a jump when it detects a sequence consisting of a takeoff phase (acceleration spike), an air phase (acceleration drop), and a landing phase (second spike). Additional criteria, such as a minimum airtime, sufficient rotational motion, and a vertical force peak from one of the skate-mounted IMUs, are used to validate the jump. Once identified, a jump is extracted as a fixed-length sequence of readings and saved as an individual binary file for further analysis or classification.

The AI component of Skatelligence focuses on classifying figure skating jumps using the processed data from the previous step. Due to the logistical limitations of collecting extensive, high-quality data, we first augmented our dataset through synthetic data generation. Initial augmentation involved shifting and adding noise to existing samples. We further expanded the dataset using a Generative Adversarial Network (GAN), trained to generate realistic IMU sequences for six jump types—single Axel, Flip, Loop, Lutz, Salchow, and Toe Loop—based on manually labeled examples. The classification model itself is a multi-layer Recurrent Neural Network (RNN) using Long Short-Term Memory (LSTM) units. The network consists of four stacked LSTM layers followed by a fully connected layer that outputs a jump classification across six categories. Each input to the model is a 150x30 matrix, representing 150 time steps of data from all six axes of five sensors. After training on both real and synthetic data, the model achieved an accuracy of approximately 80% on a held-out validation set. As the dataset continues to grow—particularly with more examples from different skaters and additional jump types—we expect to see continued improvements in model performance and generalizability.