TL;DR
This paper presents a novel temperature compensation method for six-axis force/torque (F/T) sensors, crucial for maintaining accuracy in extreme environments like space. By combining an optimized Least-Square Support Vector Machine (LSSVM) with an improved ensemble technique called trimmed bagging, the researchers significantly reduced temperature-induced measurement errors, making F/T sensors more reliable for critical applications.
What Problem Does This Paper Address?
Force/torque sensors are essential for robotic manipulators, providing feedback for precise control. However, these sensors are highly susceptible to temperature variations, leading to significant "characteristic drift." This drift causes a discrepancy between the measured and actual force/torque values, severely degrading the sensor's accuracy and reliability, especially in environments with wide temperature fluctuations, such as outer space or deep-sea operations. While hardware-based compensation methods exist, they are often costly and inflexible. Software-based compensation, which models the relationship between temperature and sensor output, offers a more adaptable solution. Existing software methods, however, face challenges with complex nonlinear drifts and can be computationally intensive or prone to local optima during parameter tuning. This paper aims to address these limitations by developing a robust and accurate software-based temperature compensation method.
Methods
The core of the proposed method is an ensemble learning framework built upon Least-Square Support Vector Machine (LSSVM) as the base learner. LSSVM is chosen for its ability to handle nonlinear regression problems effectively. To optimize LSSVM's performance, the researchers employ a hybrid Whale Optimization Algorithm (hWOA). This hybrid approach integrates the Simulated Annealing (SA) algorithm into the standard WOA to enhance its ability to escape local optima and improve global search capabilities.
The ensemble framework utilizes an improved trimmed bagging technique. Unlike standard bagging, which averages the outputs of all base learners, trimmed bagging selectively averages the outputs of the "best" performing base learners. The "improvement" comes from an adaptive trimming strategy that dynamically determines the optimal trim portion by minimizing the variance of the quoted errors of the base learners, rather than using a fixed trim portion. This adaptive approach aims to improve the robustness of the ensemble model.
Furthermore, the paper uses cross-validation and inverse quoted error (invQE) to more accurately evaluate the fitness of the base learners and the overall model.
Datasets
The study utilized data collected from a high-low temperature experiment specifically designed for this research. The input to the model consisted of ambient temperature (T), and the output was the voltage signal (V) from the six-axis F/T sensor. The dataset was divided into 80% for training and 20% for testing. The experiment involved varying the temperature from -30°C to 70°C across thirteen sampling points.
Key Insights and Ideas
- Hybrid WOA for Robust Optimization: Integrating Simulated Annealing into the Whale Optimization Algorithm (hWOA) is crucial for overcoming the tendency of WOA to get trapped in local optima, leading to better parameter tuning for LSSVM.
- Adaptive Trimmed Bagging for Stability: The adaptive trimming strategy in the ensemble method is key. By dynamically selecting the best base learners based on their performance (quoted error), the method achieves better generalization and robustness compared to fixed trimming or standard averaging.
- Ensemble Learning for Improved Accuracy: Combining multiple LSSVM models through trimmed bagging significantly enhances predictive accuracy, especially on unseen data (testing set), compared to individual LSSVM models.
- Minimizing Variance for Optimal Ensemble: The adaptive trimming strategy's focus on minimizing the variance of base learner errors is an elegant way to ensure that the ensemble model is not overly influenced by outliers or poorly performing learners.
Key Results
The proposed ensemble hWOA-LSSVM based on improved trimmed bagging achieved a dramatic reduction in measurement errors caused by temperature variations.
- The maximum absolute measurement error for the Fx dimension was reduced from 3.34% of full scale (FS) to 3.9 × 10⁻³%FS.
- After compensation, the measurement errors of all dimensions showed minimal changes with temperature variation, indicating a highly stable and accurate sensor output.
- The Fz and My dimensions exhibited the best compensation performance, with maximum errors less than 2.2 × 10⁻⁵%FS and 1.7 × 10⁻⁴%FS, respectively.
- The ensemble model demonstrated superior performance on the testing set compared to single learner models like WOA-LSSVM and WOA-RBFNN, indicating better generalization.
Limitations and Future Work
- Complexity: The authors acknowledge that the proposed ensemble method, while effective, is complex in its structure.
- Future Work: The researchers plan to focus on reducing the model's complexity and integrating the presented compensation method into more compact six-axis F/T sensors.
Summary updated