Radio-frequency sensor for non-destructive evaluation of composite materials

Abstract

The growing use of advanced composites in critical applications demands reliable, non-destructive testing. Traditional techniques often detect only large defects, missing subtle local property changes that can precede major failures. Microwave spectroscopy offers a promising alternative, probing both structural and dielectric properties with high sensitivity. This paper presents a novel dual-probe radio-frequency sensor (RFlect) using open-ended coaxial probes at 2.4 GHz. Compared to our earlier single-probe design, the dual configuration reduces scanning time and enhances detection of surface and subsurface defects, including minor cavities and inhomogeneities. It employs a cost-effective system comprising a signal generator, directional coupler, and RF power measurement setup to improve noise filtering without relying on expensive vector network analyzers. Integrated with a fused deposition modeling 3D printer’s extruder, the sensor scans samples at millimeter resolution. Experiments confirm its ability to detect both visible and hidden subsurface defects by identifying changes in conductivity and dielectric constant. This work underscores the potential of dual-probe microwave spectroscopy as a cost-effective, high-performance non-destructive evaluation technique for various applications. Future efforts will focus on characterizing high-loss materials, such as biological tissues, thereby expanding the sensor’s use in the biomedical field.

Introduction

The rapid development of modern technologies has resulted in a growing demand for advanced materials that are able to meet the increasingly sophisticated requirements in terms of strength, lightness, and durability^1,2. Composite materials, offering exceptional mechanical properties, proved to be the key solution to these challenges. Composites are widely used in extreme applications such as aerospace, wind turbines³, transportation and biomedical fields⁴, where performance and reliability are crucial.

However, the inherent complexity of composite materials due to the combination of two or more constituent phases (fibers and matrix) in layered architectures, resulting from their anisotropy and heterogeneity⁵, makes them susceptible to a variety of defects and damage^6,7, such as voids and porosity fiber waviness^8,9, impact-induced delaminations¹⁰, fatigue-driven matrix cracking¹¹ and hygro-thermal degradation¹², both during production processes and throughout the operational cycle¹³. To avoid the problems associated with this, it is necessary to introduce a control system at an early stage of the production of composite elements, as well as during their use, to ensure safety^{14, 15–16}. For this purpose, the use of appropriate non-destructive techniques (NDT) is crucial because they allow for a rapid, non-invasive assessment of structural integrity¹⁴.

While traditional NDT solutions are widely employed^{17, 18–19}, they are often limited to visualizing physical structural changes. In composites, the damage process includes many microscopic defects that slowly develop simultaneously throughout the entire volume of the material¹⁷, leading not only to mechanical degradation, but also to changes in properties such as electrical conductivity and thermal stability^14,20,21. These types of damage are difficult to detect using optical methods and techniques that visualize the physical structure of the material²⁰.

A promising approach in this field is microwave spectroscopy. Unlike traditional methods, microwave spectroscopy enables the detection of both structural changes and variations in local properties, such as dielectric constant^21,22. This makes it a highly sensitive and precise tool for evaluating composite materials, even at very early stages of degradation. In our previous work²³, we introduced a simple, cost-effective radio-frequency sensor capable of detecting damage with millimeter accuracy in 3D printed components. This system, based on an open-ended coaxial probe technique, has the ability to detect a wide range of problems - from mechanical defects caused by extrusion errors, through the detection of subsurface defects caused by changes in infill, to inhomogeneities in the conductivity of samples printed with doped filaments. The reflection sensing mode makes it possible to integrate the radio-frequency sensor with the 3D printer as an extruder attachment. It was designed to evaluate flat samples and adapted to scan subsequent layers of 3D prints, enabling the examination of more complex geometries. The system used a single coaxial probe, which made the scanning time largely dependent on the size and complexity of the examined area.

Fig. 1 [Images not available. See PDF.]

Measurement capabilities of the RFlect system. (a) Undersurface defects detection - RFlect image with a photo of a sample containing prefabricated surface and subsurface defects. The inset highlights an internal defect that is invisible on the surface. (b) Submillimeter lateral resolution - high-resolution photo of a sample designed in Autodesk Fusion to contain parallel walls ranging from 1.0 mm to 0.4 mm thick along with the image obtained with RFlect system. The insets show that the probe distinguishes walls 0.4 mm thick. The visible blurring of the wall images is related to the scanning step, which is 1 mm and is determined by the parameters of the 3D printer to which the RFlect system is mounted. The selection of a commonly available printer allowed us to demonstrate that defect detection by the RFlect system is effective even with non-specialist equipment.

Based on the original concept²³, this paper presents a novel, dual-probe radio frequency (RF) sensor (RFlect) for inspecting surface and subsurface layers and detecting subtle changes in material properties at significantly faster scanning speeds. The newly developed dual-probe RFlect sensor offers several significant improvements: (i) two open-ended coaxial probes simultaneously record two different points on the sample, which allows for approximately a two-fold reduction in total scanning time while maintaining lateral and spatial resolution; (ii) the use of 2.4 GHz frequency, combined with optimized probe geometry, increases the effective inspection depth to ~ 4 mm in carbon fiber laminates while maintaining lateral accuracy at the submillimeter level; (iii) dedicated electronics, including a low-cost continuous wave signal generator, directional coupler, and dual logarithmic power detectors, eliminate the need for a vector network analyzer, reducing hardware costs by more than 90%. All the above-mentioned advantages of the RFlect sensor translate into faster, more sensitive and economical microwave inspection of the surface and subsurface areas of composite structures (Fig. 1), which predisposes this technique for routine use in industry. Especially since the detection of subsurface defects complements the results of the commonly known ultrasonic imaging, covering the dead zone characteristic of this technique²⁴.

The paper is organized as follows. In Sect. 2, we provide an overview of the microwave spectroscopy method, detailing the construction and design of experimental setups. This section also includes a description of the prepared samples and outlines the measurement procedures employed to ensure accurate and reliable results. Section 3 presents the experimental results, showcasing the sensors’ performance across different scanning configurations. Section 4 concludes the work with a summary of the most important findings and presents perspectives for future work.

Methodology

Microwave spectroscopy is a powerful technique for characterizing the electromagnetic properties of materials by analyzing their response to microwave signals. This technique is particularly useful for analyzing dielectric and conductive materials, as it provides insight into their permittivity, conductivity, and loss characteristics. Depending on the measurement configuration, microwave spectroscopy can be performed in transmission or reflection mode. In the present work, the open-ended coaxial probe method in reflection mode is used. This technique is employed in near-field microwave spectroscopy for the characterization of local dielectric properties, enabling high resolution, non-destructive analysis based on the reflection of an incident signal. A detailed description of the NDT microwave methodology and its implementation in the near-field microwave imaging system was provided in our previous work²³. The present study utilizes three experimental setups for microwave reflection measurements, as depicted in Fig. 2. In each configuration, the RF signal is generated at 2.4 GHz and transmitted to the probe. Based on previous research, the systems were extended to incorporate dual-probe configuration, enabling sequentially alternating measurements. The first configuration employs a vector network analyzer (VNA) with two ports. This setup is referred to as “VNA 2-port” throughout the text. The second setup integrates RF switch controlled by microcontroller, enabling measurements from two probes using a single VNA port. This configuration is referred to as “VNA 1-port”. The third one, referred to as the “RFlect system”, replaces the VNA with an RF single-frequency continuous-wave (CW) signal generator and a directional coupler, while the reflected signal is recorded using an RF power measurement module. The collected data is digitized using an analog-to-digital converter and processed by a microcontroller, providing a cost-effective alternative to the previous systems using a VNA. VNA based configurations are expensive even in the case of a budget (around $1000) VNA and the proposed change reduces the cost of components by 90%, which is not insignificant for the universality of the developed solution. The next goal of testing different systems was to obtain good image quality without the need for tedious data processing. Subsequent configurations were created to achieve the assumed goal. In Sect. 3, we show and discuss the differences in the quality of the obtained images. RFlect gives results with a significantly lower signal-to-noise ratio. In addition, it also improves portability and the possibility of integration with various scanning devices. The VNA analyzer requires a rather permanent setup due to the possibility of mechanical damage to the ports when bending the connecting cables. The RFlect system, which is currently a prototype, will eventually constitute an integral whole for direct mounting on the scanning element.

Fig. 2 [Images not available. See PDF.]

Schematic representation of three experimental setups: a) Vector Network Analyzer (VNA) with two ports for simultaneous two-probes measurements, b) single-port VNA with RF Switch controlled by microcontroller, and c) RFlect: a low-cost system using RF signal generator (Analog Devices, ADF4351), directional coupler (self-made), RF power measurement module (Mini-circuits ZX47-60-S+) and microcontroller (Arduino Uno) for data processing.

Figure 3 shows the probes made of a flanged N-type socket and used in all the discussed measurement systems. The simulated electric field distribution in a 5 mm thick sample generated by the probes for the frequency f = 2.4 GHz is shown in Fig. 3a, while Fig. 3b shows these probes placed in a holder enabling mounting to the 3D printer’s extruder. The simulation results show that the probes will be able to effectively detect structures in the subsurface region to a depth of at least 4 mm.

Fig. 3 [Images not available. See PDF.]

Probes made of a flanged N-type socket. a) Simulation of the electric field generated by two coaxial probes interacting with a sample, conducted using CST Studio Suite, and b) photography of the probes mounted in a 3D printed holder.

In Table 1, we show a comparison of the actual reflection parameters (difference between the recorded signal intensity in the presence of a sample and its intensity without a sample) of the N-type probe used here and the SMA-type probe used in our previous work²³. The N-type probe successfully records the signal reflection from the sample structures for probe-to-sample distances from 0.1 mm to 4 mm, while the SMA-type probe works effectively only in the range from 0.1 mm to 0.5 mm. Furthermore, the intensity measurements for the SMA-type probe are burdened with relatively large errors and for probe-to-sample distances above 0.5 mm their values are within the measurement uncertainty. This clearly demonstrates that the larger aperture and field range generated by N-type probes make them much more effective at inspecting subsurface objects even when operated at increased distances from the sample.

Table 1. Reflection parameters as a function of the probe-to-sample distance for N-type and SMA-type probes.

Probe – sample separation distance [mm]	Reflection parameters [mV]
	N-type probe	SMA-type probe
0.1	(3.12 ± 0.09)	(1.66 ± 0.16)
0.5	(0.95 ± 0.09)	(1.03 ± 0.11)
1.0	(0.71 ± 0.03)	(0.12 ± 0.16)
1.5	(0.50 ± 0.09)	(0.31 ± 0.2)
2.0	(0.36 ± 0.07)	(0.33 ± 0.17)
2.5	(0.33 ± 0.09)	(0.31 ± 0.15)
3.0	(0.17 ± 0.06)	(0.20 ± 0.16)
4.0	(0.19 ± 0.04)	(0.30 ± 0.17)

Experimental measurements were performed on specially designed samples with prefabricated defects in the form of cavities. Each sample was printed using a ferromagnetic composite material containing dispersed iron particles (Iron Composite PLA filament). The use of such a material ensures a sufficiently strong interaction with the probe’s electric field, facilitating the unambiguous interpretation of the results obtained in the tested measurement systems.

All samples were designed in Fusion 360, exported as an STL file, processed in FlashForge slicing software to generate layer-by-layer G-code, and printed on a FlashForge 5 M printer equipped with a 0.4 mm brass nozzle. The slicer settings included a 0.2 mm layer height with a 0.3 mm initial layer height, a 0.4 mm line width, and walls set to 0.8 mm thickness with two lines and zero horizontal expansion. Four bottom and four top layers were used with ironing disabled, while the infill pattern followed a grid configuration with a 15% overlap (Fig. 4). A Dimafix was used to ensure the adhesion of the build plate. Initially, the fan speed was 0%, reaching 100% from the second layer. The nozzle temperature was set to 215 °C, the platform temperature was maintained at 55 °C, and printing was performed at a speed of 60 mm/s.

Fig. 4 [Images not available. See PDF.]

Sample design visualization (FlashForge slicing software) based on STL file: a) surface, b) inside; c) photography of a sample made on the basis of the presented design using Iron Composite PLA filament.

In all three setups, samples were scanned under identical conditions to ensure comparability between datasets. Since maintaining a constant distance between the probe and the measured object is crucial for obtaining stable and reliable measurement data, all measurements were performed at a height of 0.1 mm above the samples. The measuring distance between the probe and the sample was determined based on previous studies²³, during which simulations were performed to analyze the distribution of the probe’s electric field in the presence of the sample and to determine the most optimal probe-sample position. The alignment of the probes was carried out using the built-in motion control system of a 3D printer. The probes were securely attached to the printer’s extruder, enabling precise height adjustment through the printer’s stepper motor system. During the scanning process, the samples remain fixed on the printer bed, ensuring the stability and evenness of the tested surface.

Before starting the actual measurements, a brief calibration routine is performed before each scan. Reference measurements are taken at several points on the sample to establish a common baseline. The resulting offsets are averaged to calculate correction factors that must be applied to the data during measurement to ensure that both probes maintain a common baseline.

Data is collected using custom Python software with time, serial, os, numpy, and pandas modules. The software controls the positioning of the 3D printer, manages the RF switch sequence, and triggers ADC data acquisition on the microcontroller. The microcontroller is programmed to send a single request and take one measurement at a time. Once the stepper motor has positioned the probe, the script instructs the RF switch to receive data from RF port 1 connected to the first probe. It then sends a start signal to the Arduino for the first ADC acquisition, switches RF to port 2 connected to the second probe, and triggers the second acquisition. After each acquisition, the software puts the data in the appropriate place in the array, and then it repositions the probes to the next measurement points.

Raw data, captured as dimensionless numbers by digitizing millivolt signals with an 18-bit ADC, are first converted to 2D pixel arrays and then processed in Python using NumPy, SciPy, and Matplotlib. For all acquired images, the sample region is identified using an intensity gradient algorithm, and then a 2D fast Fourier transform is applied to the surrounding background. The low-frequency components responsible for probe motion are then removed and an inverse Fourier transform (IFT) is applied to obtain a flattened baseline. Subtracting this baseline value from the original background provides an estimate of the noise level, which is then removed from the raw data. It is worth emphasizing that the measurement results using RFlect cover a signal amplitude range of about 1200 units, which is four times greater than the 300 units recorded in VNA-based systems, so artifacts resulting from cable movement in the RFlect system are negligible. The final step is to rescale the data, for example from probe 2, to match the amplitude range of probe 1, which will ensure consistency of contrast and sensitivity across the data set.

The images of the test sample obtained in this way can now be subjected to numerical processing to smooth them and facilitate the identification of various features. In the present case, we used intensity thresholding to darken the brightest points by 2–5% of the value. This step helped increase contrast around deeper subsurface defects without creating artifacts. The brightest pixels correspond to locations where the probe-to-sample distance was shortest, such as over a small surface bump. Cutting off the high end of the intensity scale helps to reveal subsurface defects. Measurements performed on a sample with prefabricated air-filled holes of 4 mm diameter placed at different depths below the surface starting from 1 mm showed that the signal attenuation varies from 45% of the reference value (no defect) for a defect position at a depth of 1 mm to almost 97% for a defect located at a depth of 3.5 mm. By lowering the maximum signal value, reflections from shallow structures can be removed and only deeper defects can be shown. Such a low-pass amplitude gate operates solely on signal strength and does not require time analysis. Thanks to this, in contrast to ultrasonic flaw detectors, the presented method does not have a dead zone near the surface (Table 2). On the other hand, increasing the lower intensity threshold suppresses background noise and improves the visibility of surface roughness of different origins. This means that skillful numerical processing allows for the analysis of the image of the tested sample in the range from the surface to the penetration depth of the waves generated by the probes with millimeter accuracy.

Results and discussion

This section compares the results for three measurement systems (Fig. 2) used to detect details in the test sample. All results are compared according to the power of the reflected signal, because the third (Fig. 2c) tested system allows only such measurements.

In the first measurement system we used a two-port VNA, in such a way that each of the two ports was connected to a separate sensor (Fig. 2a). The obtained results show the inconsistency of the recorded signals (Fig. 5a). For the first probe, the signal reflected from the sample is greater than the background, while the second probe shows the opposite relationship. The resulting images are asymmetric with inconsistent contrast levels. Moreover, the low-budget VNA analyzer used in this system has a limited signal amplitude span of recorded values, which not only reduces contrast, but also increases the impact of thermal noise and mechanical disturbances, e.g., caused by cable movement. To remedy the problems revealed in the dual-port measurement system, the VNA used there was replaced with a single-port analyzer with an RF switch controlled by a microcontroller (Fig. 2b). The image of the test sample obtained in this system (Fig. 5b) was still not satisfactory, although the basic inconsistencies observed earlier had been eliminated. However, one can still notice the limited precision of the recorded signal visible in the image in the form of noise.

Drawing on these experiences, we decided that the situation could be improved by using a more sophisticated and expensive VNA analyzer or developing our own measurement configuration without such an analyzer. The configuration of the new RFlect setup, shown in Fig. 2c, includes an RF signal generator, a designed directional coupler, an RF power measurement system, and an RF switch from the previously discussed system. Measurements in the new system were fully controlled by a microcontroller, which significantly improved imaging quality (Fig. 5c). Moreover, the use of an 18-bit converter ensured an appropriate range of recorded values, which resulted in greater contrast and more accurate reproduction of the details of the tested sample (Fig. 5d). Overall, there was a significant improvement in the quality of the image compared to those obtained in two previous attempts for VNA-based systems. However, it is still possible to notice a difference in the imaging by each probe and therefore it was necessary to investigate which element of the hardware has a real impact on the measurement result. To achieve this, a number of tests were carried out, changing the signal propagation channels within the system built with a switch (port 1, 2) - cables (1, 2) - probes (1, 2).

Fig. 5 [Images not available. See PDF.]

Comparison of imaging quality in different measurement setups and photography of the test sample. Results obtained by using a) dual-port VNA setup – block diagram Fig. 2a, (b) single-port VNA with RF switch setup – block diagram Fig. 2b, (c) RFlect - the newly developed setup (unprocessed image with visible measurement artefacts) – block diagram Fig. 2c, b, (d) photography of the sample.

The schematics in Fig. 6 show the tested propagation channels. Thanks to systematic changes in the configuration of the appropriate elements of the signal propagation channels, it was possible to determine how each of them affects the quality of the measured signal.

Fig. 6 [Images not available. See PDF.]

Analysis of the influence of RFlect setup elements (probe, cable, and RF switch port) on measurement quality, aimed to identify the source of measurement artefacts visible in Fig. 5c. The left columns show the measurement configurations, while the right columns present the corresponding images. The greatest impact on the quality of the measured signal is related to insufficient measurement time allocated to a given port of the RF switch.

The images in Fig. 6 present scans of the same reference sample with 3 holes under various measurement configurations. Although the geometry of the sample remains unchanged and defects are clearly visible, the clarity and contrast of the images differ depending on the setup. The characteristic shift in every other row of pixels, also visible in Fig. 5c, can easily be assigned to the first port of the RF switch. This issue arose from insufficient measurement time allocated to that port and increasing the measurement interval successfully resolved the artifact, as seen in Fig. 7c. Moreover, the second coaxial cable tends to exhibit more pronounced noise, indicating a lower dynamic range of its recorded values. This reduced amplitude plausibly explains the higher noise level seen in that channel and may be the result of the non-ideal impedance matching between the cable and the probe, which results in some internal reflections within the setup. In RF/microwave systems, it is well known that precise impedance matching is essential to avoid additional noise and signal degradation, so all parts of the setup must be carefully designed and manufactured to achieve optimal performance.

After addressing the issues related to insufficient measurement time, the system was employed for subsurface defect analysis. The same sample was examined from the opposite site, which featured a plain surface (Fig. 7b). As expected, the sensors detected a subsurface defect located 4 mm beneath the surface. Notably, beyond the intentionally introduced defect, the scans revealed internal infill walls 0.8 mm thick (Fig. 7a). This observation validates the system’s capability not only for detecting subsurface defects but also identifying submillimeter-scale structural irregularities, with performance that is more precise than conventional ultrasonic methods²⁵ and comparable to ultrasonic phased array testing (Table 2). Moreover, by employing coaxial probes that utilize near-field detection, the proposed sensor offers more precise detection of shallow subsurface defects (up to 4 mm) compared to ultrasonic imaging^26,27.

In the field of RF NDT methods, split-ring–resonator (SRR) probes have been proposed both for laboratory mapping of surface dielectric permittivity²¹ and for quality control during additive manufacturing²². In the SRR approach, the resonator must be in physical contact with the element under test. Its sensitivity extends to ≈ 6 mm in the transverse direction from the measuring slit and ≈ 24 mm in depth. In addition, an external vector network analyzer is required²¹. Although the SRR geometry can be miniaturized, in practice the large sampling volume blurs the image of sub-millimeter features and limits defect localization to the immediate surface. RFlect employs two open-ended coaxial probes operating in the near field regime. This non-contact configuration provides ~ 0.5 mm lateral resolution, no surface “dead zone”, and reliable defect detection down to 4 mm under carbon fiber or PETG-ESD laminates. The cost of the RFlect prototype is less than 10% of the cost of a portable ultrasound system and completely eliminates the need for a vector analyzer. Moreover, the dual probe architecture cuts scanning time in half compared to single probe or SRR scanners, enabling rapid inspection of large areas. In summary, RFlect provides more accurate subsurface feature discrimination than contact SRR probes, has higher imaging throughput, and significantly lower design complexity, which predisposes it to be an alternative to current NDT methods.

Table 2. Comparison of the characteristics of three flaw detectors: rflect, conventional ultrasound and ultrasonic phased array.

Technique	Relative hardware cost	Spatial resolution	Surface/subsurface „dead zone” depth	Work regime	System complexity
RFlect (this work)	< 10% of portable UT	~ 0.5 mm	No dead zone	Non-contact	Low (CW source + coupler + detectors)
Conventional Ultrasound25	100% (~ 5 000 $)	0.5–1 mm (at 5–10 MHz)	4–8 mm	Contact + couplant gel	Moderate (pulser/receiver, transducer)
Ultrasonic Phased Array24	200–300%	0.2–0.5 mm (at 5–15 MHz)	1–2 mm	Contact + couplant gel	High (array controller, beam-forming)

In comparison with the previously presented SMA-based probes²³, the N-type flanged probes described in this study exhibit lower resolution, which is why 0.5 mm scratches (visible in Figs. 5d and 7d) were not detected. This detection accuracy can be improved by reducing the measurement step from 1 mm to a fraction of this value. Moreover, the N-type flanged design of the probe allows for measurement of a larger surface area and provides deeper subsurface detection - up to four times deeper than the SMA-based approach. Consequently, while SMA probes are preferable for applications requiring finer spatial resolution, the N-type socket design offers advantages in scenarios where rapid scanning and deeper penetration are crucial.

Fig. 7 [Images not available. See PDF.]

Results of RFlect measurements taken for both sides of the iron PLA sample, with an invisible and visible prefabricated defect. a) Subsurface image of defects together with internal infill walls and b) photography of bottom side of the sample, c) image of surface defects, together with d) photography of surface of the sample. Both raw images demonstrate the potential of using RFlect for quick and easy inspection of the build quality or detection of internal defects occurring during use.

The analysis of measurements taken for a sample made of PETG-ESD material with defects of different shapes placed at different depths is presented in Figs. 8 and 9. Figure 8 shows the results of measurements taken on both sides of the sample, together with their photographs. First of all, it should be noted that all artifacts, placed respectively 0.25, 0.5, 0.75 and 1 mm below the surface, are visible in a single scan. Then, by comparing the signal reflection caused by the defect with the reflection from the surface, we can estimate the depth of the defect. More precisely, if P_surf is the signal power reflected from the surface and P_def is the signal power reflected from the defect, then the ratio P_def/P_surf defines the relative depth of the defect according to the well-known attenuation profile. The P_def/P_surf values are specific to each material and increase linearly with the distance of the defect from the surface. Figure 9 shows a graph of P_def/P_surf as a function of the depth of the defects in the tested sample. As can be seen, the values of P_def/P_surf for individual defects determined on the basis of the measurement data correspond to their actual depths within the limits of measurement errors.

It is also worth mentioning that the proper use of a band-pass filter allows for precise visualization of defects at specific depths without the need to use time-domain reflectometry. This is another advantage of RFlect over classic RF analysis due to the simplicity and speed of obtaining results.

Fig. 8 [Images not available. See PDF.]

Results of RFlect measurements performed for two sides of PETG - ESD sample with invisible and visible prefabricated defects. (a) image of subsurface defects together with (b) photography of bottom side of the sample, (c) image of surface defects, together with (d) photography of surface of the sample. The defects are placed at different depths (upper star: 0.25 mm, lower star: 0.50 mm, circle: 0.75 mm, square 1.00 mm).

Fig. 9 [Images not available. See PDF.]

Linear dependence of P_def/P_surf ratio on subsurface defect depth of 0.25 mm, 0.50 mm, 0.75 mm and 1.00 mm. P_def is the signal intensity recorded in the defect region, while P_surf is the reference signal from the defect-free surface. Higher P_def/P_surf values indicate a weaker contrast between the defect and the background and deeper defect location.

Conclusion

The investigations presented here demonstrate the effectiveness of the proposed RF-based NDT system for detecting both surface and subsurface irregularities in composite elements. Compared to more established non-destructive testing techniques such as ultrasonic inspection, the configuration introduced in this article offers distinct advantages in terms of measurement accuracy and detecting shallow subsurface defects^27,28 Notably, unlike ultrasonic methods, this system requires no couplant gel between the probe and the tested object, greatly simplifying the measurement process for large surfaces. Furthermore, it is more cost-effective than the previously proposed system²³, and its submillimeter accuracy makes it highly appealing for both hobbyists and professional users.

By employing two probes operating in parallel and incorporating an RF switch, the RFlect setup achieves significantly reduced scanning times (up to half of those required by earlier single-probe or single-port configurations) without compromising imaging resolution. By operating the probe just 0.1 mm above the sample, the optimum working distance identified in our simulations, the system obtains its best signal quality: the power‑based signal to noise ratio around 22 dB. Maintaining high-contrast images requires cautious control of signal integrity, amplitude range, and timing parameters. Hardware choices, such as an 18-bit converter, combined with strategic post-processing, enable the detection of submillimeter-sized features like infill walls and subsurface defects.

In comparison to our earlier research using SMA-based probes, the new open-ended coaxial probe arrangement provides deeper RF penetration, further enhancing defect detection in thicker or higher-loss composites. The demonstrated success of the dual-probe system suggests clear potential for expansion into larger probe arrays (N × N matrix), enabling faster scanning and wider industrial applicability. Future work will focus on further extending penetration depth to enable accurate characterization of high-loss materials, including biological tissues such as muscles and bioprinted materials and tissue-mimicking structures, thus expanding the system’s utility in diverse fields.

Acknowledgements

We extend our sincere thanks to Joanna Ślot for generously providing assistance with the photographs used in this work. Their contribution was invaluable and greatly enhanced the quality and clarity of the visual elements presented in our manuscript.

Author contributions

M.Ś. led the investigation, provided funding, established the methodology, supervised the work, and contributed to the original manuscript. K.S. performed the investigation, developed the software, conducted data analysis, and wrote the original manuscript text. A.B. also contributed to writing the manuscript. M.J. Ś. prepared the figures and took part in the investigation. I.Z. supervised the project and contributed to writing the manuscript. All authors reviewed the manuscript.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

M.Ś., K.S., and I.Z. are inventors and co-applicants on patent application PCT/PL2024/050038, which covers the system described in this manuscript. The application is currently in progress and has undergone an initial state-of-art check. There are no any other competing interests for any author.

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

1. Lin, X et al. Three-dimensional printed metal-nested composite fuel grains with superior mechanical and combustion properties. Virtual Phys. Prototyp.; 2022; 17, pp. 437-450. [DOI: https://dx.doi.org/10.1080/17452759.2022.2035934]

2. Alarifi, I. M. Simulation and mechanical testing of 3D printing Shin guard composite materials. Virtual Phys. Prototyp.19https://doi.org/10.1080/17452759.2024.2411022 (2024).

3. Dimitrova, M et al. A survey on non-destructive smart inspection of wind turbine blades based on industry 4.0 strategy. Appl. Mech.; 2022; 3, pp. 1299-1326. [DOI: https://dx.doi.org/10.3390/applmech3040075]

4. Biswal, T. S., BadJena, K. & Pradhan, D. Synthesis of polymer composite materials and their biomedical applications. Mater Today Proc. 30, 305–315; (2020). https://doi.org/10.1016/j.matpr.2020.01.567

5. Jollivet, T; Peyrac, C; Lefebvre, F. Damage of composite materials. Procedia Eng.; 2013; 66, pp. 746-758. [DOI: https://dx.doi.org/10.1016/j.proeng.2013.12.128]

6. Tretiak, I; Kawashita, LF; Hallett, SR. Manufacturing composite laminates with controlled void content through process control. J. Reinf. Plast. Compos.; 2024; 43, pp. 16-29.1:CAS:528:DC%2BB2cXhsFGqu7c%3D

7. Bayat, S et al. Temporal enhanced ultrasound: A new method for detection of porosity defects in composites. Compos. Part. A: Appl. Sci. Manuf.; 2023; 164, 107259.

8. Alves, MP; Cimini Junior, CA; Ha, SK. Fiber waviness and its effect on the mechanical performance of fiber reinforced polymer composites: an enhanced review. Compos. Part. A: Appl. Sci. Manuf.; 2021; 149, 106526.1:CAS:528:DC%2BB3MXhs1ejtrbO

9. Chai, EYH; Wang, WC; Christian, WJ. R. Experimental study on the effect of waviness defects on composite material impact dynamics. Compos. Part. B: Eng.; 2024; 283, 111638.

10. Tabatabaeian, A; Jerkovic, B; Harrison, P; Marchiori, E; Fotouhi, M. Barely visible impact damage detection in composite structures using deep learning networks with varying complexities. Compos. Part. B: Eng.; 2023; 264, 110907.

11. Alamry, A. Fatigue damage and analysis of laminated composites: A state-of-the-art. Journal of Engineering Research. Preprint at (2024). https://doi.org/10.1016/j.jer.2024.09.006

12. Hussnain, SM; Shah, SZH; Megat-Yusoff, PSM; Hussain, MZ. Degradation and mechanical performance of fibre-reinforced polymer composites under marine environments: A review of recent advancements. Polym. Degrad. Stab.; 2023; 215, 110452.1:CAS:528:DC%2BB3sXhtlCjt7rE

13. Lu, L. et al. Autonomous intelligent additive manufacturing of continuous fiber-reinforced composites: data-enhanced knowledgebase and multi-sensor fusion. Virtual Phys. Prototyp.19https://doi.org/10.1080/17452759.2024.2412192 (2024).

14. Bulgarevich, DS; Shiwa, M; Furuya, T; Tani, M. Gigahertz time-domain spectroscopy and imaging for non-destructive materials research and evaluation. Sci. Rep.; 2016; 6, 27980.2016NatSR..627980B [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27302877][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4908427]

15. Rao, J., Leong Sing, S., Liu, P., Wang, J. & Sohn, H. Non-destructive testing of metal-based additively manufactured parts and processes: a review. Virtual Phys. Prototyp.18https://doi.org/10.1080/17452759.2023.2266658 (2023).

16. Lu, Q. Y. & Wong, C. H. Applications of non-destructive testing techniques for post-process control of additively manufactured parts. Virtual Phys. Prototyp.12https://doi.org/10.1080/17452759.2017.1357319 (2017).

17. Perkowski, Z; Tatara, K. The use of dijkstra’s algorithm in assessing the correctness of imaging brittle damage in concrete beams by means of ultrasonic transmission tomography. Materials; 2020; 13, 551.2020Mate..13.551P1:CAS:528:DC%2BB3cXht1Kms7jO [DOI: https://dx.doi.org/10.3390/ma13030551] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31979333][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7040618]

18. Szymanik, B; Keo, SA; Brachelet, F; Defer, D. Investigation of carbon fiber reinforced polymer concrete reinforcement ageing using microwave infrared thermography method. Appl. Sci.; 2024; 14, 4331.1:CAS:528:DC%2BB2cXhtFGmt7%2FK [DOI: https://dx.doi.org/10.3390/app14104331]

29. Cieślik, M et al. Multiple reprocessing of conductive PLA 3d-printing filament: rheology, morphology, thermal and electrochemical properties assessment. Materials; 2023; 16, 1307.2023Mate..16.1307C1:CAS:528:DC%2BB3sXjtFGmt7s%3D [DOI: https://dx.doi.org/10.3390/ma16031307] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36770313][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9920316]

20. Khatri, B; Rehra, J; Schmeer, S; Breuer, U; Balle, F. Metal/carbon-fiber hybrid composites - damage evolution and monitoring of isothermal fatigue at low and elevated temperatures. J. Compos. Sci.; 2022; 6, 67.1:CAS:528:DC%2BB38XovVCksr8%3D [DOI: https://dx.doi.org/10.3390/jcs6030067]

21. Lekas, S; Drummond, R; Grant, PS; Duncan, SR. Control of additive manufacturing for radio frequency devices with spatially varying dielectric properties. IEEE Trans. Control Syst. Technol.; 2024; 32, pp. 1579-1589.

22. Isakov, D; Stevens, CJ; Castles, F; Grant, P. A split ring resonator dielectric probe for near-field dielectric imaging’. Sci. Rep.; 2017; 7, 2038.2017NatSR..7.2038I [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28515424][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5435736]

23. Ślot, M., Samolej, K., Bartosik, M., Drabik, P. & Zasada, I. Non-contact microwave sensor for FDM 3D printing quality control. Virtual Phys. Prototyp.19https://doi.org/10.1080/17452759.2024.2360167 (2024).

24. Yeom, YT et al. Improved ultrasonic dead zone detectability of work rolls using a convolutional neural network. Appl. Sci.; 2022; 12, 5009.1:CAS:528:DC%2BB38Xhtl2ns7bP [DOI: https://dx.doi.org/10.3390/APP12105009]

25. Ciecieląg, K et al. Non-destructive detection of real defects in polymer composites by ultrasonic testing and recurrence analysis. Materials; 2022; 15, 7335.2022Mate..15.7335C1:CAS:528:DC%2BB38XislCns7nP [DOI: https://dx.doi.org/10.3390/ma15207335] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36295400][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9611944]

26. Ma, M; Wang, Z; Gao, Z; Jiang, M. Ultrasonic phased array testing and identification of multiple-type internal defects in carbon fiber reinforced plastics based on convolutional neural network. Materials; 2025; 18, 318.1:CAS:528:DC%2BB2MXit1Gnsb4%3D [DOI: https://dx.doi.org/10.3390/ma18020318] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/39859789][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11766589]

27. Zhang, H et al. High quality ultrasonic imaging at low detection frequency for defects in Thick composites. J. Nondestruct Eval; 2025; 44, 21.

28. Gupta, M; Khan, MA; Butola, R; Singari, RM. Advances in applications of non-destructive testing (NDT): A review. Adv. Mater. Process. Technol.; 2022; 8, pp. 2286-2307. [DOI: https://dx.doi.org/10.1080/2374068X.2021.1909332]