State monitoring technology is essential for fault prognostics and health management (PHM) of power electronic devices, especially wire-bonded IGBT modules that are prone to bond wire lift-off and solder aging. In recent years, electromagnetic voiceprint (EMVP) signals generated during switching transients have emerged as a promising, inherently non-intrusive indicator of device health. However, their use in PHM has been largely constrained by two gaps: the lack of a clear physical mechanism linking internal degradation to EMVP features, and the reliance on manual, experience-driven interpretation of EMVP patterns, which limits accuracy and generalization.

This work builds a complete EMVP-based state monitoring framework that bridges physical mechanisms, diagnostic indicators, and intelligent data-driven evaluation. First, we investigate how IGBT health status influences EMVP signals by developing mechanistic models that explicitly describe the impact of bond wire lift-off and solder aging on current distribution, Lorentz forces, and the resulting EMVP response. Simulation models are constructed to analyze EMVP variations under different degradation scenarios and chip locations in multi-chip modules, and typical EMVP features—such as peak-to-peak amplitude and peak value—are proposed as sensitive diagnostic indicators of bond wire degradation and localization. These models and indicators are validated experimentally using a double-pulse test platform and a dedicated EMVP acquisition system, demonstrating that EMVP features increase monotonically with fault severity and exhibit chip-dependent patterns. Importantly, the EMVP sensors can be placed outside the module package, enabling a truly non-destructive and non-intrusive monitoring solution.

On this physical and experimental foundation, we further address the challenge of robust, automated state identification. A spatiotemporal feature fusion cross-attention neural network is proposed to recognize IGBT aging states from EMVP signals under diverse operating conditions, achieving over 95% accuracy in experimental studies. To enhance applicability in realistic scenarios where labeled data are limited, a transfer learning strategy is introduced, significantly improving model effectiveness and generalization on small-sample datasets.

Overall, this integrated research transforms EMVP from a promising but poorly understood signal into a theoretically grounded, experimentally validated, and algorithmically empowered tool for non-destructive state monitoring. It provides a new pathway for high-sensitivity, non-intrusive reliability evaluation of power electronic devices, and represents a meaningful breakthrough for the fields of non-destructive testing and PHM-oriented state monitoring.

Additive manufacturing (AM) is becoming increasingly adapted across a wide range of industries, specifically aerospace and defense, due to its ability to produce complex geometries at a reduced cost and reduced fabrication time. Due to manufacturing variability inherent to the additive process, there is a need for automated inspection that can be integrated during the production process. This work presents a feature detection methodology for additively manufactured components with a large number of internal features. Results will focus on two common AM approaches, metal powder-bed laser sintering and fused filament fabrication with a polymer filament carrier of a metal powder. While metal powder-bed AM is higher cost, it is more accepted in industry. The fused filament approach comes at a reduced cost but is not well understood due to the manufacturing process requiring post sintering along with the thermal or chemical removal of the polymer carrier. Regardless, both approaches have found hesitation in acceptance due to manufacturing uncertainty and historical challenges from failed components. Inspection is performed utilizing both a conventional ultrasonic spherical transducer and from an ultrasonic phased array probe, and results are compared against those from X-ray computed tomography. The uniqueness of this work is the automated feature extraction from the captured ultrasonic waveforms. Adding to the challenge for feature extraction is that a large number of similar and/or repeating features are fabricated within the structures. The specific application of interest is the use of multiple highly complex internal surfaces. The standard inspection approach is to find a defect, such as a void or thin profile within a larger region. In the present study we seek to quantify each of the individual features that are separated by a length scale that is similar to or smaller than the feature of interest itself, without any user interaction. The analysis approach utilizes a hybrid of waveform homogenization and smoothing coupled with automated feature extraction through a sequence of steps that include multidimensional Fourier transforms, edge detection, localized homogenization, and time-shifting. The results presented include flat parts and parts with a radius of curvature nominally 37.5 mm from both powder bed and fused filament fabrication. Each component has between 50 and 200 individual features, and the results presented show that every feature can be detected. More importantly, the geometric profile of each feature is quantified, and typical errors are found to be less than 1 mm across a dominant geometric parameter (i.e., diameter, rectangle side, hexagon edge-to-edge distance) for components that range in thickness from 3 mm to 10 mm. As the components inspected are additively manufactured, there is a significant difficulty caused by the internal scatter. This challenge will be highlighted by providing an example of a high precision CNC micro machined flat component fabricated from billet stock.

In this study, a large number of test specimens (10cm x 10cm x 40cm) used in the concrete freeze-thaw test method (JIS A 1148) are used to verify whether thermographic index based on non-contact thermal behavior using a forced heat source can contribute to the quality evaluation of concrete surfaces.
Using a forced heat source (heater), superheated steam is generated from a wet state on the partial surface of a concrete rectangular column test specimen, and the magnitude of the temperature rise was related to the surface roughness, density, and ultrasonic velocity.
The results show an experiment in which changes in temperature rise on the concrete surface are related to its quality.
A hierarchical structure of a machine learning evaluation scheme is also constructed.
That includes the degree of Irregular appearance, the presence or absence of internal reinforcing bars, the presence or absence of simulated defects (foam beads) and surface cavities, the ultrasonic velocity, moisture content, and density of the test specimen as learning elements, and the inference applicability of thermographic index based on machine learning is examined.
Based on the results of inference calculations, an evaluation of thermographic index is proposed for determining whether repair inspection should be performed using surface impregnation materials.