Fault Restoration of Six-Axis Force/Torque Sensor Based on Optimized Back Propagation Networks

Xuhao Li, Lifu Gao, Xiaohui Li, Huibin Cao, Yuxiang Sun2022 · DOI

TL;DR

This paper proposes a novel method for restoring damaged six-axis force/torque (F/T) sensors by leveraging the inherent coupling effects between sensor dimensions. Using a Particle Swarm Optimization (PSO)-tuned Back Propagation Neural Network (BPNN), the system can accurately predict the output of broken sensor channels based on the readings from the remaining functional ones, enhancing sensor robustness and reducing replacement costs.

What Problem Does This Paper Address?

Six-axis F/T sensors are crucial components in robotics and automation, providing essential feedback for tasks like precise manipulation in challenging environments (e.g., deep sea, space). However, these sensors are susceptible to damage from adverse conditions, leading to the failure of one or more measurement dimensions. Replacing an entire F/T sensor due to a single or dual-channel failure is costly and impractical. While decoupling methods exist to mitigate the negative effects of coupling on measurement precision, this paper proposes to exploit these coupling effects. The core problem addressed is how to restore the functionality of a damaged F/T sensor by accurately predicting the output of the broken channels using the correlated outputs of the intact channels.

Methods

The paper introduces a fault restoration method based on an optimized Back Propagation Neural Network (BPNN). Here's a breakdown of the approach:

  • Back Propagation Neural Networks (BPNNs): BPNNs are used as the core model to learn the complex, non-linear relationships between the functional and damaged dimensions of the F/T sensor. The network takes the voltage outputs of the intact dimensions as input and predicts the voltage output of the damaged dimension(s).
  • Particle Swarm Optimization (PSO): A critical aspect of this work is the optimization of BPNN hyperparameters. PSO is employed to tune parameters such as the number of hidden layers, the number of neurons in each hidden layer, and the maximum training iterations. This optimization is crucial because manual tuning of BPNN hyperparameters is time-consuming and often suboptimal. PSO's meta-heuristic nature helps find optimal configurations that balance model complexity and performance, preventing under-fitting or over-fitting.
  • Two Restoration Schemes: The method is designed to handle two scenarios:
  • One-Dimensional Restoration: Restoring a single damaged dimension using the outputs of the other five.
  • Two-Dimensional Restoration: Restoring two damaged dimensions using the outputs of the remaining four.
  • Dataset Preparation: Data is collected from coupling experiments where loads are applied to individual dimensions, and the resulting voltage outputs of all six dimensions are recorded. This data captures the inherent coupling effects. Approximately 70% of this data is used for training the PSO-BPNN models, and the remaining 30% is reserved for testing.

Datasets

The paper does not introduce a new benchmark dataset. Instead, it utilizes data collected from a novel six-axis F/T sensor prototype developed by the Institute of Intelligent Machines (IIM), Chinese Academy of Sciences (CAS). This data was generated through controlled coupling experiments, where specific loads were applied to the sensor to capture the inter-dimensional correlations.

Key Insights and Ideas

  • Leveraging Coupling for Restoration: The most significant insight is the shift from mitigating coupling effects to actively using them for fault restoration. The paper demonstrates that the correlated outputs of intact sensor channels can be effectively used to infer the behavior of damaged channels.
  • PSO for BPNN Hyperparameter Optimization: The integration of PSO for optimizing BPNN hyperparameters is a key methodological contribution. This approach automates and improves the selection of network architecture and training parameters, leading to better restoration performance compared to standard BPNNs.
  • Feasibility of Two-Dimensional Restoration: The success in restoring two damaged dimensions, in addition to one, highlights the robustness of the proposed method and its potential for handling more complex failure scenarios.
  • Restoration Accuracy Approaching Original Precision: The experimental results indicate that the restored sensor outputs can achieve a level of precision comparable to the original, undamaged sensor, validating the practical utility of the method.

Key Results

The paper presents quantitative results in terms of Mean Squared Error (MSE) and Quote Errors (QE) for both one- and two-dimensional restoration schemes, comparing the proposed PSO-BPNN method against a standard BPNN.

  • Testing Errors (MSE):
    • For one-dimensional restoration, the average testing MSE was approximately $1.1693 \times 10^{-5}$. The maximum MSE was $5.553 \times 10^{-5}$.
  • For two-dimensional restoration, the average testing MSE was approximately $3.4205 \times 10^{-5}$. The maximum MSE was $1.218 \times 10^{-4}$.
  • Quote Errors (QE):
    • For one-dimensional restoration, the average testing QE was $8.800 \times 10^{-3}$. The maximum QE was $3.600 \times 10^{-2}$.
  • For two-dimensional restoration, the average testing QE was $8.200 \times 10^{-3}$. The maximum QE was $2.42 \times 10^{-2}$.
  • Comparison: In most cases, the PSO-BPNN method achieved significantly lower testing errors (both MSE and QE) compared to the standard BPNN, demonstrating its superior performance in fault restoration.

Limitations and Future Work

  • Indirect Coupling Effects: The authors acknowledge that the current model primarily focuses on direct coupling effects. They suggest that future research should explore the "inner mechanism of coupling effects" and how indirect coupling might influence restoration accuracy.
  • Damage Detection: The current work assumes that damage is known and focuses on restoration. Future work could involve developing methods to detect damage in the F/T sensors.
  • Generalizability: While the method is presented for a six-axis F/T sensor, further investigation into its applicability to other types of multi-axis F/T sensors with minor modifications is suggested.

Summary updated

BibTeX
@misc{li2022fault,
  author = {Xuhao Li and Lifu Gao and Xiaohui Li and Huibin Cao and Yuxiang Sun},
  title = {{Fault Restoration of Six-Axis Force/Torque Sensor Based on Optimized Back Propagation Networks}},
  year = {2022},
  doi = {10.3390/s22176691},
}
Figure