Deep Monte Carlo Quantile Regression for Quantifying Aleatoric Uncertainty in Physics-informed Temperature Field Reconstruction

Xiaohu Zheng, Wen Yao, Zhiqiang Gong, Yunyang Zhang, Xiaoyu Zhao, Tingsong Jiang

TL;DR

This paper introduces a novel "Deep Monte Carlo Quantile Regression" (Deep MC-QR) method to reconstruct temperature fields and quantify data noise uncertainty. It leverages physics-informed training to reduce reliance on labeled data and employs quantile regression to estimate uncertainty, offering a more robust solution for real-world engineering applications where perfect data is scarce.

What Problem Does This Paper Address?

Accurately understanding temperature distribution is crucial for the health and safety of complex engineering systems like aircraft and satellites. Traditional methods rely on limited sensor data, which can't provide a complete picture. While Convolutional Neural Networks (CNNs) are good at reconstructing these temperature fields (Temperature Field Reconstruction - TFR), they typically require vast amounts of labeled training data. Furthermore, real-world sensor data is often noisy, leading to "aleatoric uncertainty" (uncertainty inherent in the data itself). Standard CNNs struggle to quantify this type of uncertainty, and obtaining noiseless, labeled data for training is a significant challenge in practical engineering. This paper aims to bridge this gap by developing a method that can reconstruct temperature fields accurately with less labeled data and, crucially, quantify the aleatoric uncertainty introduced by data noise.

Methods

The core of the proposed solution is the "Deep MC-QR" method, which combines several key ideas:

  • Physics-Informed CNN Training: The method incorporates physical knowledge, specifically the Laplace equation governing steady-state heat conduction, into the CNN's training process. This "physics-informed" approach allows the model to learn an accurate temperature field reconstruction surrogate without needing any labeled temperature field data.
  • Quantile Regression for Uncertainty: Instead of just predicting a single temperature value, the Deep MC-QR method uses quantile regression. This technique estimates the conditional distribution of the output, allowing it to predict values at different quantiles.
  • Monte Carlo Sampling of Quantile Levels: During training, the method generates "quantile level images" by randomly sampling quantile levels from a uniform distribution. These, along with the input temperature monitoring point images, form the training data.
  • Aleatoric Uncertainty Quantification: In the prediction phase, the trained model uses Monte Carlo sampling of these quantile levels. By generating multiple predictions for the same input, the method can estimate the aleatoric uncertainty by calculating the variance (standard deviation) of these predictions. This effectively captures the uncertainty arising from noisy input data.
  • U-Net Architecture: The paper utilizes a U-Net architecture, known for its effectiveness in image-to-image tasks, for both feature extraction and temperature field prediction.

Datasets

The paper describes generating a dataset for training and evaluation.

  • Synthetic Dataset: A synthetic dataset was generated using a "recon-data-generator." This dataset consists of 15,000 MP (Monitoring Point) temperature images.
  • Noise Injection: To simulate real-world conditions, five different types of noise were added to the temperature values of the monitoring points:
  • Gaussian noise with varying standard deviations (0.3² and 0.5²).
  • Uniform noise within different ranges (-0.3 to 0.3, and -1 to 1).
  • Data Split: The generated dataset was split into training (10,800 samples), validation (1,200 samples), and testing (3,000 samples) sets. The ground truth temperature fields were also generated for the testing data.

Key Insights and Ideas

  • Physics as a Data Augmenter: Integrating physical laws (like the Laplace equation) into the training process is a powerful way to reduce the need for extensive labeled data, a common bottleneck in many scientific machine learning applications.
  • Quantile Regression as an Uncertainty Gauge: Shifting from point predictions to quantile predictions is an effective strategy for quantifying aleatoric uncertainty. It provides a richer understanding of the prediction's reliability.
  • Monte Carlo for Practical Uncertainty Estimation: Using Monte Carlo sampling with randomly generated quantile levels provides a practical and effective way to estimate aleatoric uncertainty in a trained neural network.
  • Noise Propagation is Complex: The study highlights that data noise doesn't just affect the immediate vicinity of the noisy data points; it can propagate and accumulate, leading to larger uncertainties in unexpected areas.

Key Results

  • Accurate Reconstruction: The proposed Deep MC-QR method achieved highly accurate temperature field reconstructions, as evidenced by very small RMSE, MAE, and MRE values, and R² values greater than or equal to 0.99995 across different noise conditions.
  • Effective Uncertainty Quantification: The method successfully quantified aleatoric uncertainty, with the uncertainty maps showing larger values around the noisy monitoring points. The quantified uncertainties were also shown to be approximately equal to the standard deviations of the injected noises.
  • Reduced Data Dependency: The physics-informed approach enabled accurate reconstruction without requiring labeled temperature fields, a significant advantage over traditional supervised learning methods.
  • Demonstrated Noise Effects: The experiments clearly illustrated how data noise can lead to increased aleatoric uncertainty in both monitored and unmonitored areas, and how this uncertainty can propagate.

Limitations and Future Work

  • Computational Cost: While not explicitly detailed as a limitation, Monte Carlo sampling for uncertainty quantification inherently increases computational cost during the prediction phase.
  • Generalizability to Other PDEs: The current work focuses on the Laplace equation. Future work could explore extending the physics-informed approach to other partial differential equations (PDEs) relevant to different physical phenomena.
  • Epistemic Uncertainty: The paper primarily focuses on aleatoric uncertainty. Future research could investigate incorporating methods to quantify epistemic uncertainty (uncertainty due to model limitations or lack of data in certain regions) alongside aleatoric uncertainty.
  • Real-World Data Validation: While synthetic data with injected noise was used, validating the method on actual noisy sensor data from real-world systems would further strengthen its practical applicability.

Summary updated

Abstract

For the temperature field reconstruction (TFR), a complex image-to-image regression problem, the convolutional neural network (CNN) is a powerful surrogate model due to the convolutional layer's good image feature extraction ability. However, a lot of labeled data is needed to train CNN, and the common CNN can not quantify the aleatoric uncertainty caused by data noise. In actual engineering, the noiseless and labeled training data is hardly obtained for the TFR. To solve these two problems, this paper proposes a deep Monte Carlo quantile regression (Deep MC-QR) method for reconstructing the temperature field and quantifying aleatoric uncertainty caused by data noise. On the one hand, the Deep MC-QR method uses physical knowledge to guide the training of CNN. Thereby, the Deep MC-QR method can reconstruct an accurate TFR surrogate model without any labeled training data. On the other hand, the Deep MC-QR method constructs a quantile level image for each input in each training epoch. Then, the trained CNN model can quantify aleatoric uncertainty by quantile level image sampling during the prediction stage. Finally, the effectiveness of the proposed Deep MC-QR method is validated by many experiments, and the influence of data noise on TFR is analyzed.
BibTeX
@inproceedings{zheng2022deep,
  author = {Xiaohu Zheng and Wen Yao and Zhiqiang Gong and Yunyang Zhang and Xiaoyu Zhao and Tingsong Jiang},
  title = {{Deep Monte Carlo Quantile Regression for Quantifying Aleatoric Uncertainty in Physics-informed Temperature Field Reconstruction}},
  booktitle = {IJCNN},
  year = {2022},
  url = {https://openreview.net/forum?id=JCu_E8UJ676},
}
Figure