High Three-Dimensional Detection Accuracy in Piezoelectric-Based Touch Panel in Interactive Displays by Optimized Artificial Neural Networks

Shuo Gao, Yanning Dai, Vasileios Kitsos, Bo Wan, Xiaolei Qu2019 · DOI

TL;DR

This paper presents an optimized Artificial Neural Network (ANN) architecture for piezoelectric-based touch panels, achieving over 97% accuracy in both touch location and force detection while significantly reducing computational cost. This advancement addresses the limitations of existing force-sensing technologies, paving the way for more responsive and efficient interactive displays.

What Problem Does This Paper Address?

Traditional 2D touch sensing (capacitive and resistive) is being challenged by the growing demand for 3D force touch capabilities in modern devices. While piezoelectric-based touch panels offer a promising, component-light solution for 3D sensing, their practical application has been hindered by inconsistent force-voltage responsivity across the panel. This inconsistency leads to reduced detection accuracy. Previous ANN-based approaches have shown acceptable accuracy but suffer from high computational costs, making them unsuitable for resource-constrained devices like smartphones. This paper aims to bridge this gap by developing an optimized ANN that delivers high accuracy with low computational overhead for piezoelectric touch panels.

Methods

The core of this work lies in optimizing an Artificial Neural Network (ANN) for piezoelectric touch panel data. The authors employed a fully connected ANN architecture, consisting of an input layer, multiple hidden layers, and an output layer.

Key methodological aspects include:

  • Data Pre-processing: A four-step pipeline was developed to extract meaningful features from the raw piezoelectric signals:
  • DC Offset Removal: Subtracting the mean voltage to center the signal.
  • Envelope Detection: Using the Hilbert Transform to capture the positive and negative voltage variations.
  • Multi-channel Noise Suppression: Averaging signals across all nine channels to reduce noise.
  • Peak Detection: Identifying local maximum voltage values to represent touch events.
  • ANN Optimization: The authors systematically investigated the impact of ANN hyperparameters on detection accuracy and computational cost. This included:
  • Cost Function Selection: Comparing Mean Squared Error, Categorical Cross-Entropy, and Binary Cross-Entropy loss functions.
  • Hyperparameter Tuning: Experimentally determining the optimal number of hidden layers and nodes for both force and location classification.
  • Computational Cost Analysis: The study analyzed both training time cost (measured experimentally) and classification time cost (analyzed theoretically based on the number of parameters).

Datasets

The study utilized data collected from four volunteers performing touch events on a custom-built piezoelectric touch panel prototype. Each volunteer performed 1350 taps at nine specified locations with three different force levels. A "touch set" comprised 50 taps per location per force level. The dataset was then split into training (80%), validation (10%), and testing (10%) sets. No new benchmark datasets were introduced; rather, data was generated specifically for this research.

Key Insights and Ideas

  • Binary Cross-Entropy is Key: For both force and location classification, the binary cross-entropy cost function emerged as superior, demonstrating better accuracy and less susceptibility to overfitting compared to mean squared error and categorical cross-entropy.
  • Structure-Accuracy Trade-off: The study revealed a clear relationship between ANN complexity (number of layers and nodes) and accuracy. While increasing complexity generally improves accuracy, there's a point of diminishing returns. For force classification, 3 hidden layers and 32 nodes were found to be optimal, while for location classification, 3 hidden layers and 64 nodes provided the best balance.
  • Computational Cost Scales with Complexity: As expected, increasing the number of layers and nodes significantly increases both training time and classification time. The paper quantifies this relationship, showing exponential growth in parameters and classification time with network size.
  • Achieving High Accuracy with Low Cost: The core breakthrough is demonstrating that an optimized ANN can achieve >97% accuracy for both force and location detection with a classification time of under 1 ms, making it practical for real-time interactive applications.

Key Results

  • Detection Accuracy: The optimized ANN achieved a test set accuracy of 97.7% for force classification and 97.0% for location classification.
  • Computational Cost: The classification time for one touch event was measured at 0.98 ms for an ANN with 5 hidden layers and 512 nodes per layer, which is significantly lower than previous ANN-based approaches.
  • Optimal ANN Structure:
    • For force classification: 3 hidden layers, 32 nodes per layer.
  • For location classification: 3 hidden layers, 64 nodes per layer.
  • Cost Function Performance: Binary cross-entropy yielded the highest accuracy (98.5% for force level estimation).

Limitations and Future Work

  • Volunteer Variability: The study acknowledges that the strength of the three force levels are highly dependent on individual volunteers' physical body conditions and touch habits. While data from four volunteers was used, a larger and more diverse dataset could further enhance generalizability.
  • Generalizability to Different Hardware: The current results are based on a specific prototype. Future work could explore the ANN's performance on different piezoelectric touch panel hardware configurations and materials.
  • Real-world Application Testing: While the paper demonstrates high accuracy and low computational cost, further validation in actual interactive display systems and under various environmental conditions would be beneficial.
  • Exploration of Other ANN Architectures: Although a fully connected ANN was used, exploring other architectures like Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs) for feature extraction or end-to-end learning could potentially yield further improvements.

Summary updated

BibTeX
@misc{gao2019high,
  author = {Shuo Gao and Yanning Dai and Vasileios Kitsos and Bo Wan and Xiaolei Qu},
  title = {{High Three-Dimensional Detection Accuracy in Piezoelectric-Based Touch Panel in Interactive Displays by Optimized Artificial Neural Networks}},
  year = {2019},
  doi = {10.3390/s19040753},
}
Figure