Abstract: An ultrasonic guided-wave testing approach based on machine learning optimization is proposed to accurately predict surface defects in pipelines. To address the significant drawbacks of traditional feature extraction methods in ultrasonic signal processing, such as high subjectivity and inadequate stability, a PLO-RF feature selection model was developed, incorporating Polar Lights Optimization (PLO) and Random Forest (RF) to enhance feature extraction capabilities by dynamically optimizing feature weights. Building on this, a PLO-CNN-BiLSTM prediction model was created. The PLO algorithm is employed to perform adaptive hyperparameter optimization on the CNN-BiLSTM structure. The convolutional neural network captures spatial local features in ultrasonic signals, while the Bidirectional Long Short-Term Memory (BiLSTM) network extracts temporal dynamic features. This combination facilitates a deep fusion of spatial and temporal features. Experimental results indicated that in the prediction task for pinhole defects (varying from 0 to 8 mm in diameter and 10% to 80% wall thickness in depth) in 304 stainless steel pipelines, the proposed model achieved an average coefficient of determination (R²) of 94.85%, representing an 8.63% improvement over the traditional CNN model (average R² = 86.22%). This result demonstrates the efficacy of the proposed hybrid model in quantifying defect sizes and explaining 94.85% of data variations. By overcoming the bottlenecks in traditional nondestructive testing techniques for the quantitative evaluation of micro and small defects, the research findings provide a solution that offers high precision and real-time performance for the operation and maintenance of intelligent pipelines. Furthermore, they deliver important engineering insights for enhancing the safety and reliability of energy transportation infrastructure.