Abstract
HP austenitic stainless steel undergoes microstructural aging due to prolonged exposure to oxidizing and corrosive atmospheres in steam reforming furnaces. The derived aging states are classified by its service temperature and microstructural markers and monitoring it is important to residual life assessment. In this regard it was used a portable Eddy Current inspection system with the aid of machine learning classification tools, characterizing aging states in HP steel in real-time. The classification profile of a 12-meter tube was acquired, validated through Field Metallurgical Replication. The developed Eddy Current inspection system successfully differentiates three regions, revealing a progression of aging states.
Introduction
Steam reforming furnaces play a vital role in the large-scale synthesis of commercial bulk hydrogen and are composed of multiple vertically oriented tubes cast of HP austenitic stainless steel. In its as-cast condition, the HP steel microstructure consists of an austenitic matrix with a network of interdentritic primary carbides promoting suitable mechanical properties for high temperature applications [1].
Reformer tubes endure prolonged exposure to oxidizing and corrosive atmospheres at 600–1000 °C, impacting microstructure and causing in-service aging. Temperature variations along the tube’s height create a gradient of aging stages (I to VI) based on wall-temperature and specific microstructural markers, including primary carbide coarsening, precipitation of secondary carbides, and etc [1–3].
Assessing the HP steel tubes’ residual life must take this microstructure variations in consideration due to the correlation of creep damage, a primary failure mechanism, with most pronounced states V and VI, resulting from exposure to the highest temperatures. Determining the tube’s aging state is crucial for accurately assessing its remaining service life [4].
Non-Destructive Testing (NDT) is extensively employed in evaluating structural integrity of industrial components. To address the present challenge, it was used a portable Eddy Current inspection system that acquires data and classifies results in real-time employing the Support Vector Machine (SVM) algorithm. The inspection system was assembled into a framework then attached to an inspection vehicle capable of scanning tubes longitudinally, which meets the requirements for on-site inspections.
The objective is the characterization of a full-scale HP steel tube aged in service, extracted from a steam reforming furnace, using electromagnetic non-destructive testing, predicting its aging state profile. To achieve this, a Support Vector Machine (SVM) classification model was trained with calibration samples then employed in real-time within a portable eddy current inspection system. The findings obtained were validated through Field Metallurgical Replication.
Methodology
The methodology employed to determine the aging states of HP steel tubes follows a two-parts inspection process. First part, the calibration samples with known aging states are inspected to create a dataset, which undergoes pre-processing and feature selection. Next, a classification model was trained using the machine learning tool Support Vector Machine (SVM). In the second part, this trained model is employed to determine the aging state when inspecting a tube with unknown aging state. Then a Field Metallurgical Replication (replica) was taken to confirm the class predicted in a single site of the tube.
The calibration samples consist of three tube sections, each measuring 200 mm in length, initially part of the same reformer tube, positioned at different heights, resulting in aging at varying temperatures for 135,000h of service. Table 1 presents the correlation between service temperature, aging state, and the height of the tube column from which the calibration samples were extracted.
Information about the calibration samples used to form the dataset
Information about the calibration samples used to form the dataset
Another entire tube was characterized by its aging state using the trained SVM model (2nd step of inspection), it underwent 160,000 h of service and measures 12 meters in length. All samples described above (tube sections and entire tube) are made of Nb-modified HP austenitic stainless steel which corresponds to 34.01% of Ni, 26.03% of Cr, 1.89 of Si, 1.26 of Mn, 1.24% of Nb and 0.48% of C with Fe in balance.
The subsequent sections will approach the eddy current inspection system utilized for data collection and the experimental setup used to inspect the full-size tube, followed by the discussion of the results pertaining to the previously mentioned methodology.
The portable eddy current inspection system consists in a pair of hybrid sensors and an electronic circuit board, that connects directly in a 110 V power outlet, requiring only a notebook with Matlab software for operation, rendering it portable for in situ applications and enhancing signal quality. The system amplifies and digitizes the signal, obtaining parameters through Fast Fourier Transform (FFT) — amplitude, phase, and offset from both sensors. The sketch of the electronic circuit system is represented in Fig. 1.
The hybrid sensor features an excitation coil and a pickup Hall sensor with external magnetic saturation due to the ferromagnetic oxidized external surface [5]. This sensor type demonstrated its capability for inspecting HP austenitic steels, as evidenced by the results presented by Arenas et al. [5]. In this paper, are used two hybrid sensors employing an external magnetic saturation in only one of them (using permanent magnets of ±300 mT), having both responses from the external oxide layer and the inner microstructure. Nonetheless, the permanent magnets don’t induce saturation in the Hall sensor.The use of a sensor without saturating the external ferromagnetic surface is interesting, on this case, because exposure to higher temperatures presents a thicker external surface and the ferromagnetism increases with its thickness, resulting in a stronger magnetic response for tube positions exposed to higher temperatures [6,7], therefore this magnetic response can positively increment the aging states dataset.
The dataset acquisition was performed through automated scanning using a KUKA robotic arm, resulting in a reduced lift-off distance of 0.4 mm.
Experimental setup
The experimental setup is designed to address the second step of inspection which involves eddy current inspection and real-time classification along the entire length of a 12-meter tube utilizing the pre-trained SVM classification model. To achieve this, the system was securely integrated into a robust framework capable of withstanding mechanical forces. Subsequently, this framework was affixed to an inspection vehicle, enabling longitudinal inspection of the tube as illustrated in Fig. 1.
Scheme representation of the inspection system scanning a tube sample. There are also specified the hybrid sensors, the electronic circuit board and the controlling computer with data acquisition interface.
Classification training
The dataset used to create and train the classification model was formed by the inspection sweep of the whole surface of each calibration sample. It was used the Eddy Current inspection system described in the previous section resulting in a total of six features. In addition, it was collected data with the hybrid probe away from any sample, named lift-off, resulting in a total of four classes: Lift-off, I, III and V. Due to the nature of the materials microstructure the electric and magnetic properties may vary in the same sample [8] resulting in a range of values for each class in each feature.
To select which features were more suitable to differentiate all classes in the classification training, a feature selection process was conducted using the feature filter algorithm RELIEF-F and the two-sample Kolmogorov-Smirnov (KS) test, both implemented in Matlab [9,10]. The feature filter analysis revealed that offset of sensor 2 exhibited the least relevance for the examined dataset, being the only attribute to receive a negative relevance coefficient. The KS test (𝛼 = 0.05) was utilized to assess potential differences in probability distributions among the classes, resulting in the finding that only “offset 2” failed to distinguish all classes, corroborating the first result.
Consequently, five features were chosen to compose the dataset used for training the classification model: amplitude, phase, and offset from sensor 1, as well as amplitude and phase from sensor 2. Each class contains a total of 4500 data points and the graphical representations (Fig. 2) reveal that all four classes can be visually distinguished.
Graphic representation of the dataset by pair of features. ∗Analog-to-Digital Unit.
The selected dataset was utilized to train the classification model for predicting the aging state of new samples. It was used the SVM from the Matlab Classification Learner library, the linear kernel was opted for its fast processing and satisfactory results, where the classification is based on weighted similarity among instances [11]. To assess the classifier’s performance, the k-fold cross-validation technique was applied with k = 10, Table 2 presents the resulting confusion matrix obtained during training.
Confusion matrix derived from the classification model training
The real-time classification of the full-length tube was graphically represented along the sweep direction, corresponding to the height of the tube, as depicted in Fig. 3. The inspection was conducted five times to ensure reproducibility. The reference point for height zero was set at the initial welded joint, situated near the upper portion of the furnace. The graphical representation revealed a progression of aging states, delineated by three distinct regions, with a predominant occurrence of aging state V. In the initial region spanning from 500 to 750 mm, there was a 100% prediction of aging state I, indicating proximity to the as-cast condition of HP austenitic steel, suggesting exposure to temperatures below 600 °C near the furnace ceiling. The subsequent region, from 750 to 1320 mm, indicates a transition with increasing temperatures within the furnace. In this range, approximately 78.85% of predictions indicated aging state III. Finally, from 1320 mm until the conclusion of the inspection, there was a significant presence of aging state V, accounting for approximately 98.32% of the predictions in this area.
To validate the classification outcomes, a replica was executed at the 7,500 mm position, at 2mm-deep, also shown in Fig. 3. Upon analyzing the replica it becomes evident that its microstructure exhibits characteristics indicative of aging state V. The presence of well-coalesced secondary carbides is apparent, although some precipitation is still discernible. Notably, the replica unveils an advanced aging state V microstructure, which is in the process of transitioning towards aging state VI. This progression implies that the exposed temperature reached almost 1000 °C.
On the left: graphic of aging state by tube’s height. On the right: The replica.
The implemented system demonstrated the capability to conduct real-time eddy current inspections and classify distinct aging states in a reformer tube that had been in operation for 160,000 hours. The observed progression of aging states aligns with the expectations documented in the literature, thereby validating the effectiveness of the non-destructive evaluation conducted. Additionally, at a specific position across the tube, its classification was corroborated through a Field Metallurgical Replication, affirming the accuracy of the trained classification model in correctly identifying the aging state.
Footnotes
Acknowledgements
The authors express their gratitude to PETROBRAS and the Brazilian research agencies CAPES and COPPETEC.