This study presents a novel and non-destructive technique for the detection of weak welds between the pole piece and the cell terminal in an electric vehicle battery module. The proposed method uses a system comprising non-contact vibration excitation and a laser displacement sensor for weak-weld detection. A pulsed gas jet is reported for the first time to generate transient vibrations at the weld region. The resulting micro-scale vibrations are measured with a high-precision laser displacement sensor, and the peak deflection is correlated with the weld strength. A microcontroller-based system is used to control the gas flow to create a pulsed jet, and the resulting transient vibrations are analysed using the NI LabVIEW data acquisition system to determine the maximum deflection. Destructive testing is subsequently conducted on each joint using a universal testing machine to determine the weld strength. A total of 150 joints were inspected using the proposed non-destructive and conventional destructive methods. A strong correlation was observed between the maximum peak deflection and the weld strength, and receiver operating characteristic analysis was performed to evaluate classification performance. The selected peak-deflection threshold of 0.065 mm correctly identified all weak joints, as determined by the universal testing machine, in the present dataset. The analysis achieved 100% sensitivity and 97.33% classification accuracy, while a limited number of acceptable-strength joints were conservatively flagged for further inspection. Accordingly, the proposed method is used as a conservative production-screening tool in which elevated deflection indicates a potential weak weld or mechanically abnormal condition requiring further evaluation.
Magnetic sensors are essential components for electromagnetic nondestructive testing (NDT). In practical applications, the diverse operating principles and performance characteristics make magnetic sensor selection challenging. This review analyzes mainstream magnetic sensors used in NDT, including induction coils, Hall effect sensors, magnetoresistive sensors, giant magneto-impedance sensors, magnetoelectric composite sensors, fluxgate magnetometers, superconducting quantum interference devices, optically pumped magnetometers, and nitrogen-vacancy center sensors. Their operating principles, applications in major NDT techniques and performance characteristics are summarized, with magnetic noise discussed in relation to measurement conditions and practical defect detectability. In addition, comparisons are presented in terms of representative magnetic noise, output sensitivity, advantages, and limitations. Based on the comparisons, key considerations for sensor selection and future research directions are also discussed.
Raw-signal health indicators are validated by a pooled cross-validated coefficient against a reference measurement, often on few physical units. We measure how such a coefficient depends on its units by deleting one unit at a time on four public run-to-failure archives: Li-ion batteries, bearings, milling tools and fatigued lap joints. One deletion moves the pooled coefficient of determination by 0.20 to 3.14 and changes its sign in three domains, while the sign of the mean per-unit correlation survives every deletion in 30 of 40 distinct tests. The unit count also sets a floor, since a lone unanimous cell needs ten units to reach q<0.05 after false-discovery correction over 40 tests. Four confounders are then isolated under an unchanged grouped protocol. Record duration alone reaches R2=+0.478±0.146 on battery capacity; a bearing reference from a channel the descriptors never see lowers that domain’s error by 44%; a per-unit reference frame multiplies the milling coefficient by 2.9; one corrupt record in 137 carries 55% of the uncorrected milling result. Two mechanisms reproduce when induced where they do not occur. The corrections also add evidence: a battery model that names discharge duration reaches R2=+0.754±0.115 and keeps its sign under every deletion. Battery subsets show that unit count alone can leave the sign unresolved, and in simulation the units needed vary with between-unit heterogeneity. On these archives a coefficient belongs in print beside its deletion range, read as a sensitivity, not an uncertainty.
