Enterprise AI Analysis

Comprehensive Data Collection & Refinement

The dataset for this study was meticulously gathered from experiments conducted on polymethyl methacrylate (PMMA) observation window samples within a high-pressure test chamber, simulating deep-sea environments up to 80 MPa. Six strategically placed strain gauges (both external and internal) monitored strain distribution, while high-precision pressure sensors recorded real-time load changes.

A total of 13,149 sets of time-dependent strain and load data were collected across pressurization, holding, and depressurization phases, repeated twice to capture comprehensive behavior under varying loads.

Crucially, raw strain data underwent a nonlinear correction, and the maximum principal stress was then accurately calculated using material-specific elastic modulus and Poisson's ratio, providing high-quality, reliable input for subsequent machine learning models.

Hybrid AI for Superior Stress Prediction

This research evaluates three machine learning algorithms for maximum principal stress prediction:

• Gaussian Process Regression (GP): A non-parametric model utilizing a Radial Basis Function (RBF) kernel. While effective for simpler prediction issues, it exhibits clear limitations in modeling high-dimensional and complex nonlinear tasks, particularly in capturing critical features with significant stress variations.
• CNN-LSTM: A hybrid neural network combining Convolutional Neural Networks (CNNs) for extracting local features and short-term dependencies from input data, and Long Short-Term Memory (LSTM) networks for capturing long-term temporal dependencies in sequential data.
• Transformer-CNN-BiLSTM: This advanced hybrid architecture first processes raw pressure and strain data via a CNN to extract local patterns. These features are then fed into a BiLSTM layer to capture dynamic relationships and global temporal features. A Transformer encoder, with its Query-Key-Value (QKV) attention mechanism and multi-head attention, models global interactions among features and incorporates positional encoding for sequential awareness, making it highly adept at learning complex temporal patterns and non-linear relationships.

Rigorous Performance Evaluation & Results

Model performance was quantified using Mean Squared Error (MSE), Mean Absolute Error (MAE), and Root Squared Residual (RSR). The evaluation used two distinct datasets:

• Dataset 1 (Rich Features): Included external load time-steps, strain data from two external gauges, and maximum principal stress values from internal gauges.
• Dataset 2 (Simplified Features): Contained only external load time-steps and associated characteristics.

The Transformer-CNN-BiLSTM model consistently outperformed both CNN-LSTM and GP models across all metrics and both datasets. On Dataset 1, it achieved an R² value of 0.9995, demonstrating exceptional accuracy. It showed reductions of 69.03% in MSE, 24.11% in MAE, and 44.3% in RSR compared to the CNN-LSTM model, highlighting its superior capability in capturing complex nonlinear relationships and global dependencies. Even with simplified features (Dataset 2), the hybrid Transformer model maintained robust performance, accurately predicting peaks and troughs where other models struggled significantly.