Pipelines are important components of nuclear industry equipment. Unexpected pipeline failures can cause serious economic losses and safety hazards. Pipelines are inevitably affected by various factors such as fatigue, corrosion, and impact during operation, which causes pipeline damage (Wang et al., 2023; Eastvedt et al., 2022; Spandonidis et al. , 2022). Identifying pipeline damage in real time is critical to ensuring safety, maintaining optimal performance, and reducing maintenance costs. Non-destructive damage identification technology based on ultrasound guided waves has been widely used for online health monitoring and identification of pipeline structural damage (Zhang et al., 2018; Datta and Sarkar, 2016). This popularity is mainly due to the remarkable properties of ultrasound guided waves, such as long transmission distance, low energy consumption, and high sensitivity to small defects ( Yang et al., 2023 ). Nevertheless, guided waves are highly sensitive to changes in environmental and operating conditions (EOC), such as variations in temperature, humidity, and flow rates. These changes can have a significant impact on guided wave propagation, resulting in the addition of wave modes and changes in wave speed and attenuation rate. As a result, such sensitivity can mask structural anomalies and cause false alarms or false negatives in the monitoring system, severely compromising its stability. Pipelines in the nuclear industry generally operate under complex environmental and operating conditions with constantly changing temperatures and pressure loads. This sensitivity to EOC limits the application of guided wave structural health monitoring in real nuclear industrial pipelines. Over the past two decades, many researchers have utilized modern data analysis techniques to overcome these challenges and enhance the potential of guided waves in damage detection (e.g., Zang et al., 2022; Zhang et al., 2022).
In this regard, two main methods are available to detect the health status of structures through data processing: model-based methods and data-driven methods (Chen et al., 2023; Humer et al. , 2022). Model-based methods require comprehensive mathematical and applied mechanics knowledge to create structural models. Furthermore, simulating models of structures operating under realistically varying EOCs is often too complex and impractical (Gorgin et al., 2020; Eybpoosh et al., 2017). Conversely, data-driven methods establish a baseline structural condition model based on past measurement data. The baseline model and data from online sensors are then used to determine the current health state of the structure. Additionally, the baseline model can be updated in real time as structural conditions change. Therefore, this paper applies a data-driven framework to study it. Traditional data-driven damage identification methods rely heavily on feature engineering, which requires manual extraction and selection of useful features from raw data as input for classifying and determining structural health status. Yes (Gawde et al., 2023). Accurate detection of structural damage highly depends on the extracted/selected features. The extraction/selection of damage-sensitive features mainly depends on the structure properties, geometry, boundary conditions, and damage type, and these features can change over time (Liu et al. ., 2023).
To address the aforementioned issues, researchers have attempted to leverage the powerful representation capabilities of deep learning models to facilitate automatic feature learning (Zhang et al., 2023). The widely adopted process of feature learning can be divided into two consecutive steps: (1) signal decomposition, and (2) the use of neural networks for feature recognition and extraction. This represents a modular learning paradigm (feature engineering and fault classification are performed separately). From the perspective of the entire model, these methods cannot guarantee the optimization of model parameters. Based on this, end-to-end learning was introduced in the field of damage identification. Build integration pipelines that directly map inputs to outputs (Huang et al., 2023a; Ajagekar and You, 2021; Wang et al., 2021). The integrated pipeline optimizes all model weights simultaneously, improving diagnostic accuracy.
Considering the aforementioned considerations, this paper proposes an integrated pipeline damage identification method based on particle swarm optimization (PSO), bidirectional gated iterative unit (BiGRU), and attention mechanism. This method utilizes a bidirectional gate iterative unit and an attention mechanism to extract features from the time-domain data of the original pipeline guided waves to classify the pipeline state. Additionally, the PSO algorithm is utilized to obtain the optimal relevant hyperparameters for the model. This study focuses on the real operating conditions of pipeline structures in the nuclear industry, conducting pipeline guided wave monitoring and damage identification experiments under time-varying environmental conditions. The effects of temperature, load, and temperature-load coupling on damage identification are analyzed using ultrasonic guided wave signals collected under comparable pipeline conditions. The proposed method is applied to online monitoring and damage identification of pipeline structures in the nuclear industry.
The innovations and contributions of this research are listed below.
- (1)
The proposed method improves the efficiency of identifying pipeline damage under various temperature and pressure loading conditions. This feature allows the method to reduce the influence of changes in temperature and pressure loading conditions on the damage identification results, ensuring high accuracy and reliability even under varying temperature and pressure loading conditions. The robustness of the proposed method to guided wave signals improves its reliability in environments characterized by fluctuations in temperature and pressure loads and represents an advance in the field of pipeline damage identification in the nuclear industry.
- (2)
The proposed method operates independently of the need for prior knowledge about structural properties or specific environmental and operating conditions. This property highlights its good adaptability to a wide range of structures, requiring only recalibration of the model by training signals from both damaged and undamaged structures. This inherent flexibility allows for seamless application across different structural contexts, making this method adaptable to diverse scenarios without requiring specific details regarding temperature loading parameters in nuclear industry pipeline environments. Confirms robustness and versatility.
- (3)
A damage detection method for nuclear industrial pipelines using PSO-based BiGRU-Attention model is proposed. This model directly uses the original ultrasound guided wave signal as input and allows BiGRU to automatically extract damage features. The integration of attention layers allocates more attention to critical damage features and emphasizes their influence on outcomes. Additionally, PSO is employed to optimize the structure of his BiGRU-Attend model. This methodology minimizes the influence of manual intervention on damage detection and experimentally validates its superior performance compared to previous studies.
The following sections are organized as follows: Section 2 presents a comprehensive review of previous research and discusses the theoretical basis of this paper. Section 3 details the proposed approach and the architectural design of the model. Section 4 introduces the nuclear industry pipeline testbed, including details of data acquisition and preparation methods. In Section 5, we validated the model using datasets collected under different working conditions. Finally, Section 6 concludes the study.
