Post-Processing of High Formwork Monitoring Data Based on the Back Propagation Neural Networks Model and the Autoregressive - Moving-Average Model

Yang Yang 0055, Lin Yang, Gang YaoSymmetry 2021 · DOI

TL;DR

This paper introduces a novel approach for analyzing high formwork monitoring data by combining Back Propagation Neural Networks (BPNN) with the Autoregressive-Moving-Average (ARMA) model. The BPNN is used to intelligently determine the optimal order of the ARMA model, leading to more accurate stress trend predictions and improved efficiency compared to traditional methods like AIC and BIC.

What Problem Does This Paper Address?

High formwork systems are critical in construction, and their safety is paramount. Health monitoring systems generate vast amounts of data, but effectively processing this data to predict potential hazards and understand structural behavior remains a challenge. Traditional methods for modeling time-series data, such as ARMA, rely heavily on accurate order estimation. Incorrect order selection can lead to suboptimal models and inaccurate predictions. This paper addresses the need for a more robust and efficient method for post-processing this monitoring data, specifically focusing on identifying the correct ARMA model order to improve stress trend prediction.

Methods

The core of this paper lies in a hybrid approach that leverages the strengths of both BPNN and ARMA models.

  • ARMA Model: The Autoregressive-Moving-Average (ARMA) model is a standard statistical tool for time-series analysis. It models a time series as a linear combination of past values (autoregressive part) and past forecast errors (moving-average part). The key challenge with ARMA is determining the optimal orders (p for AR, q for MA) for the model.
  • Back Propagation Neural Networks (BPNN): BPNNs are a type of feedforward neural network that learns by adjusting its internal weights and biases based on the error between its predictions and the actual values. In this study, BPNNs are employed in a novel way: to identify the optimal order of the ARMA model. The input layer of the BPNN is designed to correspond to potential ARMA model orders, and the network's Mean Squared Error (MSE) is used as the criterion to determine the best fit.
  • Hybrid Approach: The paper proposes using BPNN to simulate the ARMA process and estimate its order. By training BPNNs with different input layer configurations (representing different ARMA orders), the network's MSE on validation data helps pinpoint the most suitable ARMA model order. This approach is designed to handle both linear and nonlinear sequences, offering an advantage over traditional methods that primarily focus on linear relationships.

Datasets

The paper utilizes simulated datasets generated through Monte Carlo simulations to evaluate the proposed method. These simulations involve creating time series based on ARMA models with varying coefficients, sequence lengths, and model orders. The authors also apply their method to actual measured stress data from a high formwork system.

Key Insights and Ideas

  • BPNN as an Order Selector: The most significant insight is the successful application of BPNN not just for modeling but specifically for determining the optimal order of an ARMA model. This moves beyond traditional statistical criteria like AIC and BIC, which can be less robust.
  • Improved Accuracy: The BPNN-based order selection consistently achieves an accuracy rate of over 90%, significantly outperforming AIC and BIC, especially for higher-order models.
  • Robustness to Model Complexity: The BPNN approach demonstrates resilience to variations in ARMA model coefficients and sequence lengths, maintaining high accuracy across different scenarios.
  • Efficiency Gains: By accurately identifying the ARMA model order, the proposed method can increase processing efficiency by approximately 50% compared to traditional methods.
  • Handling Non-Causality: The study shows that when time series data doesn't meet causality requirements (e.g., during the loading phase of formwork), the BPNN approach, after appropriate data transformation (differencing), still provides superior results.

Key Results

  • Order Estimation Accuracy: BPNN consistently achieved over 90% accuracy in estimating ARMA model orders across various simulations, while AIC and BIC performed significantly worse, often below 10% for higher-order models.
  • Performance Comparison: In simulations, BPNN demonstrated superior performance in reducing the envelope width of stress trends compared to AIC and BIC, indicating a more compact and accurate representation of the data. For example, in one case, BPNN reduced the envelope width by 83.08% compared to AIC.
  • Efficiency: The proposed method increased processing time series efficiency by about 50% when determining the model order.
  • Real-World Application: When applied to actual high formwork stress data, the BPNN method provided a more compact representation of the stress trend compared to AIC and BIC.

Limitations and Future Work

  • Computational Resources: The authors acknowledge that calculating models for all possible orders, coefficients, and sequence lengths was limited by computational resources.
  • Accidental Factors: The current model does not account for accidental factors that might influence structural behavior.
  • Future Work: The authors suggest future research should focus on incorporating the impact of accidental factors into model building and exploring methods for predicting periodic data. They also plan to further investigate the effects of model order on different judgment criteria.

Summary updated

BibTeX
@article{yang2021post,
  author = {Yang Yang and Lin Yang and Gang Yao},
  title = {{Post-Processing of High Formwork Monitoring Data Based on the Back Propagation Neural Networks Model and the Autoregressive - Moving-Average Model}},
  journal = {Symmetry},
  year = {2021},
  doi = {10.3390/sym13081543},
}
Figure