WeldSimPatch is a normal-only anomaly-detection method for pipeline weld radiography. It retains PatchCore's memory-bank and nearest-neighbor scoring pipeline and adds a BYOL-style teacher-predictor module to adapt pretrained patch descriptors to weld radiographs. Hybrid Neighborhood-Pair Similarity (HNPS) supplies a continuous affinity that combines pairwise feature proximity with local-neighborhood consistency. This work does not claim a new anomaly-detection paradigm; it studies a task-specific descriptor adaptation within PatchCore. The dataset contains 5,501 normal radiographs (5,001 for training and 500 for testing) and 500 defective test radiographs with pixel-level masks. On the reported internal split, WeldSimPatch obtained Image AUROC 0.9934, Pixel AUROC 0.8774, and AUPRO 0.7912, compared with 0.9884, 0.8265, and 0.7350 for PatchCore. These differences equal 0.50, 5.09, and 5.62 percentage points, respectively. The evidence is limited to a single internally constructed dataset and does not establish operation independent of factory, device, material, pipe geometry, or exposure condition.
Precise quantification of minute defects in pulsed eddy current (PEC) testing is fundamentally limited by the broadening of defect boundary response signals caused by eddy current field diffusion — a key factor that undermines detection accuracy. To improve quantification precision for such small defects, this paper introduces an adaptive power-law sharpening signal processing approach. This method incorporates an algorithm that dynamically tunes the power-law exponent γ according to normalized peak height, thereby compressing the edges of the response envelope and reducing the impact of boundary effects. Subsequently, piecewise cubic Hermite interpolating polynomial (PCHIP) interpolation is applied to generate a high-density data field. This data field is then combined with K-means clustering for automatic segmentation of the defect region, enabling precise extraction of both the area and two-dimensional coordinates. Experiments were conducted using a pulsed eddy current array sensor to detect multi-layer riveted blocks containing corrosion defects. Experimental results verify that the proposed method effectively suppresses the adverse effects of eddy current diffusion on boundary resolution. For circular corrosion defects with diameters from 5 to 12 mm at identical burial depth, the method achieves a relative area error below 16.3
Aiming at the problems of low signal-to-noise ratio (SNR) of laser ultrasonic signals, weak defect peak characteristics and susceptibility to environmental noise interference in the detection of slab bottom voids in asphalt concrete rigid pavement, a recognition method based on the combination of laser ultrasonic peak reconstruction and unsupervised anomaly detection is proposed. Firstly, an unsupervised autoencoder network model is constructed. The noisy signal is compressed into a low-dimensional hidden space through the encoder and reconstructed using the decoder. Meanwhile, a peak-weighted loss function is introduced to enhance the local features near the defect echo, achieving adaptive denoising and peak amplitude fidelity of the signal. Secondly, principal component analysis is performed on the reconstructed multi-dimensional ultrasonic features for dimensionality reduction. The principal components are extracted to eliminate the residual redundant information, and the local outlier factor algorithm is used in the low-dimensional space to calculate the local reachable density and outlier factor value of each sample point to identify the outlier points with significantly deviated density from the neighborhood, completing the identification of the blank area at the bottom of the plate. The results show that the SNR after denoising by the proposed method is 21.38 dB and the peak amplitude retention rate is 94.3
Space debris in the millimetre-to-centimetre range represents a critical threat for spacecraft operating in low Earth orbit, because these particles are generally too small to be individually tracked by ground-based Space Surveillance and Tracking systems, yet can still cause severe local damage upon impact. Existing on-board debris detectors can provide in-situ impact information, but they are typically implemented as dedicated non-structural payloads, increasing mass, complexity, and integration effort. This work presents a proof-of-concept for a multifunctional structural-sensing approach in which the load-bearing spacecraft composite structure itself functions as the detection system. The concept is realised through the integration of copper wire mesh layers within carbon-fibre-reinforced polymer (CFRP) laminates, with the long-term aim of enabling distributed detection and localisation of debris impacts. Numerical finite element simulations and experimental three-point bending tests were performed to support the mechanical design and manufacturing of the proposed multifunctional composite and to characterise the interlaminar and flexural behaviour of CFRP coupons with and without embedded copper mesh. Experimental results demonstrate that the embedded copper wires have no statistically significant effect on the mechanical performance while enabling reliable monitoring of strain-induced resistance variations using a low-cost Arduino-based data acquisition system. These findings provide a coupon-level proof-of-concept validation of the proposed structural-sensing approach under quasi-static loading conditions. The overall aim is to enable real-time on-orbit awareness of small debris (1–10 mm) impact events, supporting mitigation strategies such as spacecraft attitude adjustment or collision-avoidance manoeuvres to reduce exposure to the debris flux.
Periodic inspection of fatigue cracks in welded steel oil storage tanks is critical for structural integrity. Conventional techniques, such as Magnetic Particle Testing, often necessitate costly and time-consuming coating removal, which generates industrial waste and increases downtime. To address these challenges, this study proposes a novel "Virtual ACFM-type Butterfly Plot Analysis" using Uniform Eddy Current (UEC) probes to enable high-accuracy crack evaluation through coatings without complex hardware. The butterfly plot is a well-known technique associated with Alternating Current Field Measurement (ACFM) probes. The proposed method derives a synthesized horizontal magnetic field signal, denoted as V^'_H , by spatially integrating the vertical magnetic field component, denoted as V_V , measured by a single detection coil of the UEC probe. This mathematical transformation facilitates the generation of ACFM-type "butterfly plots" from UEC signals. The effectiveness of this methodology was validated through finite element method simulations and experimental benchmarks on welded steel plates containing fatigue cracks. A significant advantage is that the integration process functions as a digital spatial filter, effectively mitigating high-frequency noise arising from surface roughness and improving the signal-to-noise ratio in the presence of complex weld bead geometries. Consequently, the synthesized V^'_H signal demonstrated a superior signal-to-noise ratio in detecting fatigue cracks in welded joints compared to direct physical V_H measurements obtained from conventional ACFM probes. The proposed method is highly synergistic with automated measurement systems, where the integration of precise spatial data from a scanner maximizes its diagnostic potential. Furthermore, by utilizing a signal feature plane derived solely from the V_V signal, it is possible to determine whether a detected crack requires repair without performing complex calculations. This research demonstrates that sophisticated ACFM-level diagnostics can be achieved using the UEC probe. Ultimately, the combination of the UEC probe and the proposed virtual butterfly plot analysis offers a robust, efficient, and cost-effective solution for the non-destructive evaluation of coated welded structures in industrial environments.
