TL;DR
This paper introduces ASTRAS360, a novel miniaturized absolute rotary encoder that uses machine learning to achieve high accuracy by analyzing captured images. By leveraging ML for both classifying image sectors and regressing precise angles, the system demonstrates robustness against mechanical imperfections and eccentricity, paving the way for more accurate and compact angular sensing in robotics and other applications.
What Problem Does This Paper Address?
Traditional high-accuracy rotary encoders (REs) are often bulky, limiting their use in miniaturized systems like robotic surgical instruments. While existing absolute REs exist, their accuracy typically decreases with size. The authors aim to develop a miniaturized, high-accuracy, single-turn absolute rotary encoder that overcomes these limitations. The core challenge lies in accurately determining the rotation angle from captured images, which are subject to optical artifacts and mechanical imperfections. Furthermore, achieving real-time processing on embedded systems is crucial for many applications.
Methods
The paper presents a machine learning-based approach for linearization and error compensation of their novel ASTRAS360 encoder. The measurement principle relies on capturing an image that uniquely identifies the rotation angle. This involves a two-step process:
- Sector Classification: The captured image is first classified into one of 16 sectors based on its color. Several ML algorithms were evaluated for this, including Convolutional Neural Networks (CNNs), k-Nearest Neighbors (kNN), decision trees, and Support Vector Machines (SVMs). CNNs, in particular, showed excellent performance by directly processing image data.
- Angle Regression: Once the sector is identified, the precise rotation angle is regressed. The authors explored polynomial regression, Feed-Forward Neural Networks (FFNNs), and Gaussian processes. A key insight is the use of "feature engineering," where the raw image data is transformed into meaningful inputs for the ML models. This includes color intensity vectors, weighted hue histograms, and crucially, the shift of the captured shadow image relative to a reference image.
A significant aspect is the use of extra inputs, such as the direction of rotation, to compensate for mechanical imperfections like the play in the rotary mount. They also investigated the encoder's tolerance to eccentric mounting, a common issue in real-world applications.
Datasets
The authors developed a custom calibration setup to automatically generate labeled training data. This setup involved:
- The ASTRAS360 prototype.
- A high-accuracy commercial optical encoder (Renishaw) as a calibration standard.
- A stepper motor for controlled rotation.
This process yielded a dataset of 16,786 samples. For training and testing, this dataset was divided into:
- Classification: 75% for training (12,590 samples) and 25% for testing (4,196 samples).
- Regression: Similar splits were used, with a distinction made between "clockwise" and "counter-clockwise" datasets to account for potential directional biases.
- Eccentricity Testing: A subset of the "Ecc0" (zero eccentricity) dataset was used for training, while other datasets (Ecc1, Ecc2, Ecc4) with varying degrees of eccentricity were used for testing.
Key Insights and Ideas
- ML for Image-Based Sensing: Machine learning is not just an add-on but a core component for interpreting the complex image data from the ASTRAS360 encoder, enabling both classification and precise regression.
- Feature Engineering is Crucial: The performance of the ML models, especially for regression, is highly dependent on the chosen input features. The "shift" of the shadow image proved to be a particularly effective feature.
- Leveraging Extra Inputs for Robustness: Incorporating additional information, like the direction of rotation, allows ML algorithms to actively compensate for mechanical imperfections, significantly improving accuracy.
- Eccentricity Tolerance: The ASTRAS360 system demonstrates a notable tolerance to eccentric mounting, a critical factor for practical deployment in real-world scenarios.
- Miniaturization Potential: The successful application of ML to this image-based principle opens doors for developing highly accurate, yet compact, absolute rotary encoders.
Key Results
- Classification Accuracy: Achieved 100% accuracy for sector classification using various ML algorithms, including CNNs, kNN, decision trees, and SVMs.
- Regression Accuracy:
- The model function regression (Equation 1) yielded a sigma of 187 arcsec and a peak-to-peak error of 2700 arcsec.
- Polynomial regression achieved a sigma of 11 (157) arcsec and a peak-to-peak error of 136 (200) arcsec (values in parentheses for counter-clockwise data).
- FFNN and Gaussian processes achieved even better results, with FFNN showing a sigma of 8.6 arcsec and a peak-to-peak error of 60 arcsec.
- Eccentricity Tolerance: The system maintained high accuracy even with eccentricities up to 4mm. For instance, with 1mm eccentricity (Ecc1), the accuracy was 99.8%, and the systematic error sigma was 104 arcsec.
- Speed: The ML algorithms were designed to be fast enough for real-time applications, with some classification algorithms achieving prediction times as low as 2.5 ms (decision tree, excluding histogram computation).
Limitations and Future Work
- Mechanical Imperfections: While ML can compensate for some mechanical imperfections (like rotation direction), the paper acknowledges that the play in the rotary mount can still introduce systematic errors, especially when the direction of rotation changes.
- Feature Engineering Dependency: The success of the regression models was heavily reliant on effective feature engineering, particularly the "shift" feature.
- Future Work:
- Development of a miniaturized version of ASTRAS360, potentially using a miniaturized image sensor and a similar measurement principle but with a two-degree-of-freedom pivot.
- Integration of additional sensors (e.g., thermometer, pressure sensor) to compensate for a wider range of environmental variables.
- Further exploration of ML algorithms and feature sets for even greater accuracy and robustness.
- Implementation on embedded platforms like FPGAs for real-time processing.
Summary updated