A Fully Polarimetric Radar System for Non-Destructive Testing of Fiber Glass Layers

During the production of fiber composites, individual layers can be displaced, which can drastically influence the load-bearing capacity of the object to be manufactured. Existing systems in the field of non-destructive testing of fiber composites are either impractical to use in everyday production or do not offer the possibility of examining the structure of the individual fiber layers. For this reason, we present a novel sensor concept that can detect typical defects caused during manufacturing. The concept is based on a millimeter wave SAR approach using a fully polarimetric FMCW radar sensor working at a center frequency of 60.5 GHz, with 17 GHz bandwidth. The sensor offers a phase noise of −77 dBc/Hz at 1 MHz offset from the carrier, with chirp times of 1 ms over the whole bandwidth, while maintaining a RMS frequency deviation of 1.24 ppm. Using a polarimetric decomposition scheme, we can detect the displacement of individual fiber layers, although the thickness of the layers is below the theoretically achievable range resolution of the radar.

Abstract-During the production of fiber composites, individual layers can be displaced, which can drastically influence the load-bearing capacity of the object to be manufactured. Existing systems in the field of non-destructive testing of fiber composites are either impractical to use in everyday production or do not offer the possibility of examining the structure of the individual fiber layers. For this reason, we present a novel sensor concept that can detect typical defects caused during manufacturing. The concept is based on a millimeter wave SAR approach using a fully polarimetric FMCW radar sensor working at a center frequency of 60.5 GHz, with 17 GHz bandwidth. The sensor offers a phase noise of -77 dBc/Hz at 1 MHz offset from the carrier, with chirp times of 1 ms over the whole bandwidth, while maintaining a RMS frequency deviation of 1.24 ppm. Using a polarimetric decomposition scheme, we can detect the displacement of individual fiber layers, although the thickness of the layers is below the theoretically achievable range resolution of the radar.

I. INTRODUCTION
I N RECENT years, the development of lightweight and strong composite materials has enabled considerable progress [1] in constructing efficient plants in the aerospace sector [2] and renewable energies, such as wind turbines [3]. As one of the essential components of composite materials, fiberglass is currently preferred due to its cost efficiency and ease of handling. Various manufacturing options exist [4]. They all share that different layers of knitted and woven glass fibers are superimposed in a defined arrangement. A matrix element, for example, polypropylene or epoxy resin, is added by vacuum injection or impregnation followed by curing. During this process, various problems can massively impair the final strength of the manufactured object [5]. These include, Manuscript received 22 March 2023; revised 2 June 2023; accepted 10 July 2023. Date of publication 17 July 2023; date of current version 3 August 2023. This work was supported in part by the European Union and the Federal State North Rhine-Westphalia in the frame of the European Regional Development Fund (ERDF) for the project "FiberRadar" under Grant EFRE-0801473; and in part by the frame of the "terahertz.NRW". The project "terahertz.NRW" is receiving funding from the programme "Netzwerke 2021", an initiative of the Ministry of Culture for example, fiber waviness [6] or the displacement of the glass fiber layers among each other [7]. For this reason, a practical non-destructive inspection for such defects is essential, which ideally should be carried out before hardening to avoid costintensive reworking. Although many inspection systems for wind turbine blades are currently available [8], [9], they are not suitable for monitoring during the manufacturing process. The most common method in the industrial sector for testing wind turbine blades is the ultrasonic C-scan due to its good representation for the geometric analysis of the defects [10]. However, the sensor head requires direct contact with the measured object since air is a very poor sound transmitter at high frequencies. An additional transmission medium is usually required representing a significant drawback for its use during manufacturing. Other alternatives are thermography and X-ray-based computed tomography (CT) [11]. Whereas CT scans provide fine resolution, they have the major disadvantage of ionizing radiation. In particular, this technology is unsuitable for use in an industrial environment with large components, as a separate measuring chamber is required. Moreover, restrictions for the use of thermography represent the heating process of the material and the presence of dirt on the surface [12].
Due to the continuous development of radar measurement systems, they offer an exciting alternative to the methods above. They provide a fine resolution, are operational even in harsh environmental conditions, and are capable of realtime processing. Thus, radar represents the standard in various fields, such as automotive sensor technology [13], [14], safety sensor technology [15], [16], and material characterization [17]. In the field of non-destructive testing, their influence is rapidly growing [18], [19], and recent applications also include fiber material scanning [20], [21], [22]. All approaches have in common that they can only detect interference points in the order of the magnitude of the range resolution. Individual layers of the fiberglass are therefore not detectable, so that, for example, a displacement of the layers relative to each other would not be recognizable. Moreover, all systems use only a single polarization, but additional information can be obtained from the fiberglass under investigation by recording and evaluating both polarizations. These are well-known techniques, e.g., in airborne synthetic aperture radar [23], [24] or to detect hidden security scanning threats [16], but they were never used in the field of non-destructive testing.
Due to this, we want to present a new sensor concept for monitoring defects in the fiberglass composite materials manufacturing process to exploit these possibilities. It comprises This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ a fully polarimetric frequency-modulated continuous wave (FMCW) radar sensor and a corresponding image reconstruction scheme. The radar sensor operates at a center frequency of 60.5 GHz and an adequate bandwidth of 17 GHz. It includes a baseband chirp generator, which ensures short measurement times of 1 ms per polarization with high linearity of 1.24 ppm and low phase noise of −77 dBc at the transceivers' output 1 MHz offset from the carrier.
The algorithmic part is based on a polarimetric decomposition scheme. To this end, we use the four channels of the radar and apply an eigenvalue decomposition approach. Here we use a modified H/α/SPAN decomposition [25], which is mainly based on the Cloude-Pouttier decomposition (also known as H/α/A decomposition) [23]. Further approaches are based on the 4-component scattering power decomposition with rotation of the coherency matrix [16], [24]. Although all methods have in common that they use a decomposition scheme of the coherency matrix to map the image data in the color space. However, we will observe that H/α/SPAN gives excellent results for defect detection and that the entropy (the parameter H in H/α/SPAN) contains additional information worth evaluating.
The structure of the article is as follows. In the following section, we want to take a closer look at the intended application by introducing the defect types which may occur during manufacturing. Furthermore, we motivate the choice of the system parameters. We describe the radar sensor in Chapter II. In Chapter III, we introduce the imaging algorithm. Finally, we verify the concept and the decomposition scheme by measuring probes of different defect types and evaluating the results.

II. APPLICATION AND SYSTEM DISCUSSION
Typical defects that arise during the manufacturing process of fiber composite materials are, e.g., fiber misalignment or voids, cuts or ondulations, cf. Fig. 1. These defects may result in fracture or delamination in the presence of loads. Due to the superimposition of multiple glass fiber layers, detecting defects below the surface by sight is impossible before introducing the corresponding resin. We refer to Section V-B for images of the defects. At first, we need a sufficiently fine resolution to detect a layer with a misalignment or an undulation. Hence, the cross-range resolution should be less than the distance between two fibers. A rough estimation of this distance (unidirectional layer, warp direction) is approximately 2.2 mm.
The resolution in cross-range for the radar image is given by Here λ denotes the wavelength and φ the 3 dB antenna beam width. At 60 GHz we obtain a maximal cross-range resolution of δ x y,max = 1.25 mm < 2.2 mm, which is sufficient. It follows that a minimal opening angle of φ min = 70 • is needed. Concerning the range resolution, we note that, ideally, we want to resolve each fiber layer. The thickness is 0.5 mm to 0.7 mm. However, the range resolution is given by Here c = c 0 / √ ε r is the electromagnetic wave's phase velocity in the fiber material, c 0 denotes the speed of light in vacuum, and ε r the electric permittivity of the fiber material. Note the influence of the permittivity in (2). Assuming ε r = 6 as in [26] would imply bandwidths larger than B = 87 GHz, which are only possible for THz or Sub-THz systems. Unfortunately, the systems available on the market do not have sufficient output power or amplification of the reflected signal, so the high attenuation of the fiberglass in the terahertz range can be compensated only with difficulty to assure a sufficiently high penetration depth [27]. For this reason, the system must operate with a lower bandwidth than would be necessary to determine each layer uniquely. Due to its use in the industrial sector, the desired sensor should be in a frequency range that does not require approval.
For this reason, the frequency band for short-range devices 57 GHz to 64 GHz was chosen. This frequency range ensures an acceptable trade-off between resolution and penetration capabilities. Moreover, it is well-known that the amount of scanning positions increases quadratically with the frequency (which is due to a spatial sampling condition). This is often a limiting factor in view of the system's real-time capabilities. Even if we may not resolve each layer, we still may hope to detect possible defects as the final image consists of the superposition of the layers in one resolution cell.
We use a new full polarimetric approach to increase the dynamic range and, thus, the system's detection capabilities. To this end, we note that layers consisting of homogeneous materials will almost exclusively preserve the polarization of the electromagnetic wave, where changes in shape and volume and multiple reflections will also give cross-polarimetric terms. Thus, misalignment defects will likely be visible in the copolarimetric images, whereas undulation and short layers will likely be in the cross-polarimetric images.

III. RADAR SENSOR A. General Concept
As a basis for our development, we choose the system concept of an FMCW radar. FMCW radars offer a precise multi-target resolution due to their high possible bandwidth and constant output power simultaneously [28]. Due to its principle, the classic FMCW radar is only able to send and receive one polarization direction at the same time. The overall system must be extended by a second channel consisting of a different TX and RX path to transmit and receive a second polarization direction. Figure 2 shows the overall concept of our fully polarimetric FMCW radar.
Generally, two consecutive chirps are needed to measure the polarization change induced by the device-under-test (DUT). The first generated chirp is forwarded to the TX H antenna over the switch, and the TX V antenna is disabled. The RX H and the RX V antenna receive the reflected signal simultaneously. After, the signals are mixed down using two mixers to the intermediate frequency (IF). The IF signal is forwarded to two analog-to-digital (AD) converters. Subsequently, the first chirp is finished, the switch is set to the TX V antenna, and the TX H antenna is disabled. The second chirp is generated, reflected by the DUT, and received by the RX (H and V) paths.
This basic concept faces several challenges and requirements in imaging the fiber mesh. As mentioned in section I, the attenuation of the fiber mesh is pretty high, which corresponds to the low energy of the reflected signal, especially for the cross-polarization (HV/VH) signal. The low energy can be compensated with a higher output power of the TX paths, but direct crosstalk between the TX and RX antennas can lead to clipping in the mixers or the AD converters. The best way to compensate for this is to physically spread all TX and RX paths, corresponding to a distributed Multiple Input Multiple Output (MIMO) system. However, this leads to new challenges for the chirp generator and the RF path. Due to the high needed bandwidth and output frequency, the chirp can not be directly generated in the desired output RF range at the antenna. This high frequency would lead to high losses in the connections between the generator, power splitters, and antennas. Instead, the chirp has to be generated at a much lower frequency. After, it is multiplied close to the antennas to the desired output frequency. A direct modulation at the antenna ports, e.g., using phase-locked loops (PLLs) [29], would be possible but results in a bad coherence. Using only one chirp generator solves this problem.
Another critical factor is the length of the chirp. For large scanning surfaces, the measurement time should be as fast as possible. Fast frequency chirps correspond to a high IF frequency, corresponding to higher sampling rates in the AD conversion. We used chirp times of 1 ms as a tradeoff between the amount of data and the measurement time. Generating the frequency chirp for the FMCW system is the most complex part. Ideally, the chirp has a constant d f dt . Unfortunately, every real system has nonlinearities [30]. These causes range detection errors due to sidelobes and spurs in the IF frequency. So, these nonlinearities should be kept as low as possible. The last requirement on the transmitted chirp is the phase noise. Phase noise represents random fluctuations in the output signal's phase, which causes errors in the range profile [31]. So, the phase noise should also be kept as low as possible.
One last consideration is that the overall system architecture should be as flexible as possible. The system's flexibility is tied to the chirp generator and TX/RX modules. The output chirp should be controllable regarding the frequency range, the chirp length, and the frequency pattern. The TX and RX paths for both polarization planes should be physically separated from each other to adjust the distance between each antenna. Moreover, the antenna has to be replaceable to change the gain. Figure 3 shows the complete schematic design of the radar sensor with physically separated TX and RX channels, each for horizontal and vertical polarization. To minimize crosstalk, each channel is housed in a separate package. In addition, the viewing range can be adjusted by changing the distance between the modules.

B. Building Blocks
An external script is first used to generate the required chirp data by division factors. These are transmitted via the Ethernet interface to a field programmable gate array (FPGA), which temporarily stores the complete chirp. A separate channel in the FPGA triggers the chirp generation externally to start a new measurement. The FPGA now reads the generated division factors and modulates the chirp generator via a parallel 8-bit interface. Due to a common 100 MHz reference, the FPGA and the chirp generator always work synchronously. The generated chirp at the output of the chirp generator is not at the desired center frequency of 60 GHz but only at 937.5 MHz, which corresponds to a division factor of 64. This significantly lower frequency makes it possible to route the chirp via conventional SMA cables without producing any significant losses, making the physically separated structure possible in the first place. The distribution to the respective TX and RX modules is done via a standard Wilkinson divider, which is not shown in the figure.
In the TX and RX modules, multiplication to the actual center frequency of 60 GHz must now be performed. The multiplication is done using one PLL, each with a multiplication factor of 64, which is almost fully integrated on a monolithic microwave integrated circuit (MMIC) [32]. Custom MMICs had to be used to utilize the maximum bandwidth available in this frequency range fully. A further reduction of the bandwidth would lead to even worse image results and thus make detecting the fiber layers even more difficult. An overview of the developed MMICs is shown in Tab. I.
These TX-and RX-MMICs enable a maximum bandwidth of 17 GHz at 60.5 GHz center frequency. Only the loop filter and the voltage regulators were placed on a separate board within the module. In a TX module, the high output frequency is transferred and radiated directly to the antennas with low loss using a developed waveguide coupling [33]. The use of standard WR-15 waveguides makes it easy to change antennas. The TX polarization direction is switched via additional channels on the FPGA so that only one direction is emitted at a time, but both directions are received.
There are now two RX modules in the receiver path for the H and V polarizations. In these modules, the generated chirp is multiplied by a factor of 64 with the aid of a PLL again. In addition, there are mixers in the MMICs to mix the reflected signal received via the antennas down to the IF signal using the multiplied chirp. In addition, amplification is performed here with low noise amplifiers (LNA).
The two IF signals are fed back via SMA cable near the FPGA, where they can be filtered and digitized. The digitized IF signals are forwarded via Ethernet for further image processing. For maximum noise immunity, the entire system is fully differential. This applies to the chirp generator's output and the IF signals.
The following subsections describe and explain all individual blocks in detail.
1) Chirp Generator: High demands are placed on the chirp generator regarding its linearity and phase noise since these are directly reflected in the output of all TX-antennas. To minimize the linearity error in the TX/RX modules, the effective frequency deviation of the modulated reference signal should be as slight as possible. By generating a high reference frequency using a PLL and then modulating it by a frequency divider in the feed-forward path, a minor frequency step is more possible than possible with a single modulated PLL. Generating a high reference frequency in the GHz range while maintaining sufficient phase noise requires unique components. Moreover, the frequency divider in the feed-forward path has to handle the high reference frequency while supporting high-speed modulation of the division factors. Figure 4 shows the block diagram of the chirp generator. An oven-controlled-crystal oscillator (OCXO) of type Abracon AOCJY2 with an output frequency of 100 MHz is fed via an active clock buffer of type Analog Devices LTC6957 to the FPGA and another PLL located in the chirp generator. The PLL is based on the TX-MMIC, which is a special MMIC implemented in Infineon's SiGe:C BiCMOS B11HFC consisting of a phase-frequency-detector (PFD) [34], a static divider with a division factor of 512, and a voltage-controlledoscillator (VCO) [32]. An external loop filter now controls the VCO with the help of a tuning voltage U Tune so that a static output frequency of f PLL = 51.2 GHz with deficient phase noise is generated.
The generated signal is then fed to a divide-by-N frequency divider based on the dual-modulus concept [35]. This divider's unique feature is its high possible input frequency of up to 92 GHz and its control interface. An adapted interface provides 8-parallel differential CML lines to dynamically select divider values from 12 to 259 and thus generate output frequencies from 198 MHz to 4.2 GHz. After the completion of each divider cycle, the new divider value is read in via the N + 1 data lines. This allows dynamic modulation of the output frequency. The N + 1 data lines are directly connected to the transceiver outputs of the FPGA, which can provide very high byte rates of up to 18 GBytes/s.
The required division factors are precalculated to generate a nearly ideal frequency chirp. Due to the complexity of the calculation, this can not be done in the FPGAs logic [36]. The precalculated division factors are converted into a byte stream, taking setup and hold times into account and written to a DDR4 memory via Ethernet using a network protocol. If a new chirp is started, the div-by-N divider is first transferred to a defined state via its reset pin, and then the byte stream is read from the DDR4 memory and forwarded to the transceiver outputs. Due to the common clock reference of the OCXO with the div-by-N divider and the FPGA, the correct division value is provided at the right time via clever and intelligent logic. The final distribution of the chirp is done using two external Wilkinson dividers combined with two amplifiers.
2) TX Module: Special requirements are also imposed on the TX modules concerning their possible bandwidth, linearity, and phase noise. In this measurement scenario, fast frequency chirps of around 1 ms must be driven over the entire bandwidth while maintaining sufficient output power. An additional unique feature is the possibility of deactivating one module at a time. Figure 5 shows the block diagram of a TX module. The basic structure also consists of a PLL, the basis of which is the TX-MMIC developed by previous works [32]. Due to the high possible bandwidth of the VCO, a particular compensation network was used, which adjusts the gain of the PFD in proportion to the tuning voltage so that the product of the gain of the PFD and the VCO remains constant [37]. This consists of two resistors and a diode. Only this compensation network enables the PLL to achieve the fast chirp times of 1 ms and high linearity of 1.24 ppm. The entire PLL was simulated at its center frequency (Fig. 6), and the loop filter was designed for a bandwidth of 5 MHz and a sufficiently high phase reserve of 70 • . The function of the phase noise shown here represents the additive phase noise neglecting the chirp generator. The jitter in the integrated range from 100 Hz to 100 MHz is 43 fs.
A unique challenge is the deactivation of a module. In principle, it would be possible to disconnect the VCO from its power supply, but the duty cycle would be too high to achieve a fast switching of both TX modules. For this reason, the VCO is not switched off electrically but kept in the lower, unused frequency range by injecting an additional voltage U Lock . This way, there is no frequency overlap of the transmitted chirp of the other module. When switching it on again, the entire system has to wait for only approximately 500 ns, corresponding to the settling time of the PLL. The toggling, controlled via the FPGA, is done with the help of an analog switch of the type Analog Devices ADG1419.
For the best possible interference suppression of the individual modules within the MMIC and the loop filter among each other, all voltage supplies were built separately and decoupled with low-noise linear regulators of the type Linear LT3045. The TX module is implemented using two mechanically separate circuit boards combined in an aluminum housing for the best possible shielding (Fig. 7). The differential loop filter, the compensation network, and the power supplies are on a daughterboard. Due to the low frequencies, this is a standard FR-4 board connected to the RF board via connectors. The RF board, made of gold-plated Rogers RO-3003, houses the MMIC and the waveguide transition. The RF board is double-layered and has a cutout for the MMIC, which is mounted directly to the package for the best possible cooling. The signal lanes of the MMIC are connected to the RF board via bonding wires to minimize additional losses.
To achieve a high output power and, at the same time, high flexibility regarding the used antenna, a novel coupling from the MMIC to a WR-15 waveguide has been realized. Since the output stage of the MMIC operates differentially, a single-ended signal is first generated via a rat-race coupler. This signal is transferred via a taper to a substrate-integrated waveguide (SIW) and via a stepped-impedance converter to the actual WR-15 waveguide, enabling simple antenna replacement. The stepped impedance converter is milled directly into the aluminum housing, enabling low-loss coupling [33].
3) RX Module and Receive Path: The RX module is used to receive the reflected signal, mix it down and then amplify it, which must be done with particularly low noise. The RX module is realized similarly to a TX module, whereby only Fig. 7. Picture of the manufactured TX module. A module contains an aluminum housing, which also provides the waveguide transition, an RF board to accommodate the MMIC, and a daughter board, which mainly provides the loop filter and the power supplies. the developed MMIC was changed. The output frequency of the VCO is not forwarded to the antenna. Instead, it is guided to the LO port of a developed mixer. On the other hand, the RF input of the mixer is connected to the antenna via the familiar waveguide transition. This produces the desired LO signal, amplified with low noise via an LNA.
The final IF signal is differentially fed from the RX module via SMA cable. Deactivating the module is unnecessary because both polarization directions are received simultaneously. Due to multiple reflections and additional mixing products, the IF signal must be sufficiently filtered so that mirror frequencies cannot occur during analog-to-digital conversion. A separate five-stage anti-aliasing filter in multiple-feedback architecture takes care of it. The differential filter has a cut-off frequency of 1 MHz with an additional gain of 20 dB in the passband. The ADA4945 operational amplifier from Analog Devices made the entire filter as low-noise as possible. The AD7626 from Analog Devices was used as the analog-todigital converter. Due to its high sampling rate of 10 MS/s at 16 bit, it offers a sufficiently high dynamic range to sample the IF signal with sufficient accuracy.
The FPGA reads the ADCs out completely synchronously with the chirp generator. The two ADCs of the RX modules are driven in phase by additional clock buffers. The digitized IF signals are then buffered, fragmented within the FPGA and prepared for further transport via Ethernet. Due to the high amount of data, a unique network protocol based on UDP was developed to transfer the data in real time over the Gigabit Ethernet interface. This protocol was implemented exclusively via the FPGA logic without using additional microcontrollers or system-on-chips.
Due to the large amount of data, the data must be transmitted in fragments. For this purpose, 300 AD samples of both RX paths are combined. Each datagram first consists of a 16-byte header. Then the current polarization direction (1 Byte), the fragmentation (2 Bytes), and the position index (4 Bytes) of the scanner are transmitted. This is followed by the 600 Samples (4 Bytes each). The second TX polarization is transmitted after successfully sending the first TX polarization.

C. Characterization
In what follows, the most critical parameters of the radar sensor were characterized, which were essential for further imaging. First, a check of the output frequency over the desired chirp time of 1 ms was performed, which can be seen in Fig. 8. For this purpose, the output signal of a TX module was divided into the baseband using the integrated frequency divider in the MMIC. Subsequently, the divided signal was measured in transient mode using a Rohde&Schwarz FSWP phase noise analyzer and compensated again concerning the divider factor. As can be seen, the radar sensor meets the requirement of 1 ms chirp times while maintaining a the bandwidth of 17 GHz. In addition, the deviation from an ideal frequency chirp is shown in Fig. 8. This satisfies an RMS frequency error of 21 kHz with an RMS nonlinearity of 1.24 ppm and thus exhibits sufficiently high linearity for imaging. Further analysis of the frequency deviation showed that the measuring principle mainly limits it. This is due to the frequency divider, which produces additional harmonics. However, no suitable measurement method is currently available since we consider high bandwidths with simultaneously short sweep times. It is assumed that the overall system will achieve even better performance.
The phase noise of the chirp generator and the TX/RX modules were analyzed at a fixed output frequency at the chirp generator. Figure 9 shows the phase noise at an output frequency of 60.68 GHz at the TX module and an output frequency of 948 MHz at the output of the chirp generator. For this purpose, the chirp generator was not modulated but set to a fixed division factor of 54. This factor corresponds approximately to the center frequency of the overall system. The phase noise was again measured using the Rohde&Schwarz FSWP. Harmonic mixers were used at the TX module waveguide port to characterize the output signal. The chirp generator's output was measured directly at the FSWP. As shown in Fig. 9, the phase noise of the overall system is mainly determined by the chirp generator, whereas the PLL in the TX module has little noticeable influence. An additional simulation of the  phase noise at the TX module, taking the chirp generator as a reference, shows good agreement. The measured jitter at the TX module corresponds to 953 fs over an integrated range of 100 Hz -100 MHz. The OCXO mainly determines the phase noise in the chirp generator.
In addition, the phase noise was also measured at a distance of 1 MHz from the carrier over the entire output frequency range of the TX module. For this purpose, fixed division factors from 49 to 64 were set on the chirp generator. As shown in Fig. 10, this corresponds approximately to the ideal logarithmic curve. This indicates that the compensation network used within the TX and RX PLLs significantly improves the phase noise at the edge regions of the output frequency. A further approximation to the maximum output frequency was impossible due to the chirp generator's integer divider values.
A TX module was connected to a Keysight U8489A power meter via a waveguide adapter to determine the output power. An external signal generator of type Rohde&Schwarz SMB40 covered the entire frequency range. The output power was compensated concerning the waveguide adapter and additional cables. As seen in Fig. 10, a sufficiently high output power is achieved at the waveguide of the TX module and assumes an average output power of 3.8 dBm. It is also clear that the maximum output frequency is 69.27 GHz since the PLL no longer operates stably above this frequency. The slight beat of the output power is mainly due to the transition between MMIC, the bond wires, and the PCB. However, this is corrected by further calibration.
After calibration, the range resolution of the system was measured. For this purpose, a metal plate was positioned as a reflector at 50 cm from the system. The calibrated raw data was windowed using a Tukey window function (α = 0.25), and the 6 dB width of the resulting peak was measured. This results in a range resolution of 11.755 mm. The achievable lobe width adding the correction factors for the window function was 11.747 mm. The remaining error was 8 µm, showing the high linearity of the overall system.

IV. IMAGING AND POLARIMETRIC DECOMPOSITION
We want to recall the basic ideas of radar imaging briefly. The radar image arises as a fusion of multiple measurements achieved by a so-called aperture. To create an aperture, one may move a single radar sensor (SAR = synthetic aperture radar) along a specific trajectory or use a complex antenna structure (MIMO). A combination of these two methods is also possible. The image reconstruction is achieved with a digital focus. We use a synthetic aperture approach with the above full-polarimetric radar in the present case. Hence, we obtain four different data sets, one for each channel. We recall that each channel corresponds to a different combination of transmitter and receiver, which differ in their polarization. The subsequent signal processing steps are depicted in Fig. 11.
In the first step, we reconstruct the four images corresponding to each channel using a standard digital focusing algorithm. Thus, we obtain four images corresponding to each data set. We denote by S H H and S V V the co-polarization, whereas S H V and S V H correspond to the cross-polarization. In matrix form, we have In the next step, we introduce surface orthogonal coordinates to extract the corresponding fiber layers individually. As the surface's orientation depends on the experimental setup, we need to use the image data to calculate the change of coordinates. In the present case, we apply the ideas in [38] to the co-polarimetric image S V V (x, y, z). Hence, the surface z = z(x, y) is given by the voxel of maximal amplitude in z-direction, i.e., we have In a next step we may approximate the surface z(x, y) by a plane and introduce new coordinates (x ′ , y ′ , z ′ ) such that the object's surface corresponds to z ′ = 0. We refer to [38] for the details.
In what follows, we want to fuse the four polarimetric channels to obtain an output that enables the end-user to detect the various defects. To this end, we require that the orientation of the fiber must be visible to detect a possible misalignment. Moreover, the color should represent the involved scattering processes to detect possible undulations. We follow the ideas in [25] and modify the corresponding H/α/SPAN decomposition in [25]. To this end we introduce the coherency matrix which will be the main subject of interest in what follows. Then we have with Here t p is the Pauli-vector, which describes the weighting of the different scattering processes in image one cell. Roughly speaking, simple reflection corresponds to the first entry, double reflection to the second, and diffuse scattering with a polarization change to the third entry.
In the next step, the scattering matrix is weighted over adjacent image cells to determine the three main scattering processes in each part of the image. Then each of them will be a combination of the three basic (simple and double reflection, diffuse scattering) ones. An eigenvalue decomposition of the coherency matrix achieves this. Let λ 1 > λ 2 > λ 3 (13) be eigenvalues corresponding to the eigenvectors of the weighted coherency matrixT . The magnitude of the eigenvalue determines the strength of the scattering process. The eigenvector is this kind of process. In what follows, we denote the total backscattered power and the entropy. Then H measures the number of involved scattering processes. The original Cloude Pottier decomposition also made use of the anisotropy, which corresponds to the parameter A in the original H/α/A decomposition. However, for small values of H (which holds true in our case), the anisotropy contains no meaningful information [25]. Next, we define the parameter α as Here v (3) i is the third component of the i-th eigenvector. Note that this definition differs from the usual one in [23] and [25], where the authors use the first component of the eigenvector instead of the third. This is because these quantities define the image's color; thus, they should depend on diffuse scattering in our case.
Finally, we can define each pixel and its coordinates in the color space. To this end, we use the HUE color model, which has as parameter the color C HUE , the saturation S HUE and the illumination L HUE . We set (after normalization of P and H ): L HUE = 10 · log 10 P.
Here frac(·) denotes the fractional part and ν 1 , ν 2 are parameters, which we choose both around 0.8. Note that the color is defined in a way such that the absence of diffuse scattering corresponds to blue color. This convention is based on airborne SAR, where simple reflections mainly correspond to sea clutter.

A. Measurement Setup
For the verification, we want to use a SAR approach, i.e., the sensor is mounted to the scanner platform and moved along a predefined trajectory. The scanner platform is depicted in Figure 12. It consists of two linear units, one for horizontal and the other for vertical movement.
The linear units can be moved along a total size of 80 cm × 80 cm in xand y-direction. For the measurement, we moved the radar along a zig-zag trajectory with a continuous movement in the vertical direction. A spatial sampling rate of 2 mm was used to avoid spatial aliasing. In the case of the maximal velocity (0.4 m s −1 ) this gives a total measurement duration of approximately 22 min. We stress that the platform is mainly intended for validation purposes. Hence lower velocities for the prototype were used.
The TX and RX modules were mounted on separate brackets on the front side of the scanner (Fig. 13) and the other peripherals on the backside. The arrangement of the TX/RX modules was chosen so the RX modules always have the same distance to the TX module. We used standard waveguide antennas with a 3 dB beam angle of 25 • for our test setup.
Once the radar is integrated, we perform a calibration to compensate for the frequency response of the radar. The latter may be degraded e.g. due to non-linearities of the frequency ramp or dispersive effects. For a monostatic radar, the calibration is well-known: a measurement is performed using a predefined object (corner, metal plate, dihedral reflector) and the radar's frequency response is compared to the expected one. Then a calibration factor may be calculated as in [39]. We note that this approach may easily be applied to calibrate the co-polarimetric channels of the full-polarimetric radar. Then, the calibration of the cross-polarimetric channels may be achieved using a rotatable dihedral reflector [40]. Correcting the radar's frequency response is only one step. Due to the distributed antenna concept, we have to compensate for the spatial offset between the four channels. In this case, we use a predefined object (e.g. Siemens star) again and perform a Radar prototype (with gainhorn antennas) integrated into the measurement setup. SAR measurement of the object. Then the spatial offset of the four images may be calculated using a cross-correlation approach.

B. Imaging Results
To validate the concept, we consider three samples, each corresponding to a different type of defect, cf. Fig. 14 (a)-(c). Each sample consists of 10 layers, 8 of them having the same direction 0 • and 2 with defects. Sample A has a misalignment in the fiber orientation in Layers 5 and 6, whereas Sample B has an undulation in Layers 5 and 6. Sample C has incomplete layers, i.e., these fiberglass layers only cover half of the surface. The layers of each sample are given in Tab. III. Each layer has a size of approximately 30 cm × 30 cm.
As multiple layers are placed on the defective layer, the defects are not visible from the front surface. Finally, for the measurement, we fix the samples to a wooden plate and covered them by vacuum foil. A vacuum pump was used to extract the air to achieve conditions as realistic as possible. In Fig. 14 (d)-(i), the radar images for two different channels are depicted. Figures (d)-(f) correspond to the maximum projection of a co-polarized channel, whereas Figures (g)-(i) correspond to a cross-polarized channel. We note that these two channel are sufficient for a qualitative evaluation of the results, for the decomposition scheme all four channels are used. We note that the co-polarized images look similar since the main part of the energy is reflected back from the surface of the object. Thus, the misalignment in Figure ( III MEASUREMENT SAMPLES AND LAYER CONFIGURATION incomplete layer seems visible but might not be significant for possible detection. As expected, the dynamical range of the cross-polarized images are significantly lower than of the co-polarized. However, the cross-polarized images can reveal the subsurface structures since the object's surface becomes almost invisible. Hence, the undulation in Figure (e) becomes clearly detectable. Also, the fiber misalignment in Figure (d) becomes visible. However, the incomplete layer is not detectable in Figure (f). In Fig. 15, the results of the decomposition scheme for the layers are depicted. We note that we use here surface orthogonal coordinates in contrast to global coordinates in Fig. 14. We consider two layers of different depths below the surface. We assume a permittivity of ε r = 6 as in [26] to determine the depth of the layer. An incorrect estimate of the refraction effect results in a wrong estimate of the object's range information, cf. also [41]. Figures (a)-(c) are located 1.53 mm below the surface and correspond to the 2nd or 3rd fiber layer. Figures (d)-(f) are located 3.26 mm below the surface and correspond to the 6th layer. Here we applied zero-padding to evaluate layers below the resolution limit. In such a way we may observe the change of the different layers. One observes that the overall fiber structure may be clearly detected in all images. For Sample A, the misalignment of the fiber structures of 10 • becomes visible for deeper structures, cf. Figure (d). Moreover, we observe a slight change of color due to the change in the scattering process with higher depth. In Figures (b) and (e), we may clearly detect the undulation of Sample B below the surface. However, due to the (limited) bandwidth of the radar, the undulation is already visible in Figure (b) at a depth of 1.53 mm. Its influence becomes stronger with increasing depth, until its final location is reached. Finally, the incomplete layer in Sample C may hardly be detected in Figures (c) and (f). To detect the defect in Sample C, we want to investigate the influence of the values of the entropy H . We do not consider P and α defined in (15) resp. (17) as they indicate the color of the image cell or its luminosity. Thus, their influence should be already clearly visible in Fig. 15. In Fig. 16, the results of the entropy (normalized to its maximum in the layer) for the deeper layer at 3.26 mm are depicted. One observes that the misalignment of the fiber layers has an impact on the entropy, but it is not significant enough to deduce any conclusions. However, the undulation and the incomplete fiber layer are clearly visible in the entropy images. The absolute values for the entropy give further insight. For Sample A (Figure (a)), the maximum of the entropy is 0.263, whereas for Sample B (Figure (b)), we have 0.358 (+36%), and for Sample C (Figure (c)) even 0.635 (+141%). Hence, the incomplete fiber layers may also be detected considering the absolute value of the entropy. In particular, the entropy seems to represent a very good indicator of possible defects in materials. This would be worth investigating in forthcoming research. For comparison, we briefly want to investigate an additional sample. Here we also consider an incomplete fiber layer as in Sample C. However, the incomplete layers should also contain a change in the fiber orientation. More precisely, Layers 1 to 4 and 7-10 contain fiberglass material with 0 • fiber orientation, whereas Layers 5-6 contain an incomplete layer with 45 • fiber orientation. The results for the cross-polarization and the polarimetric decomposition in a depth of 3.26 mm are depicted in Fig. 17. In this case, the incomplete layer is clearly visible and may be detected.

VI. COMPARISON WITH STATE-OF-THE-ART SYSTEMS
In Tab. IV, current radar systems developed for detecting flaws in fiberglass were compared with our sensor system. In [20], a 140 GHz FMCW system with 60 GHz bandwidth was tested in transmission as a quasi-optical system and in reflection with SAR approach at flaws more significant than 6 mm. It was found that these imperfections were detected in the quasi-optical system but not in the SAR approach. It should be noted that the transmission approach is unsuitable   for industrial use since the final shape, for example, a wind turbine blade, would be too misshapen. Detection of fiber layers to each other was not performed. In [21], the detachment and ondulation of fiber layers could be detected. However, this approach does not provide an image but only a damage indicator. This approach is unsuitable for control during the  fabrication of the fiberglass because this approach can only detect a change in the structure under load. In [22] and [42], network analyzers were used in the near-field region. In principle, detection of the interference points was possible, but only in the order of the magnitude of the range resolution. Thus, these approaches are unsuitable for detecting the fiber layers to each other. Furthermore, using a network analyzer in an industrial test environment is challenging. Our approach, consisting of the full-polarimetric radar and the polarimetricdecomposition scheme, demonstrated for the first time the ability to detect single missing or misaligned fiber layers, even though they were below the range resolution. The sensor concept thus has a unique point compared to existing systems. Table IV compares our radar system with other polarimetric radar systems. It can be seen that our system offers the highest bandwidth with the lowest phase noise. Due to the high bandwidth and high linearity, it also offers the best range resolution of 11.75 mm, making it ideal for close-range SAR applications.

VII. CONCLUSION
We have introduced a sensor solution consisting of a fully-polarimetric radar and a corresponding decomposition scheme to monitor fiber composite materials during manufacturing. The radar sensor combines unique features such as high bandwidth of 17 GHz with fast chirp times of 1 ms while maintaining low phase noise of −77 dBc/Hz at 1 MHz offset from the carrier and high linearity of 1.24 ppm. The imaging algorithm could extend previously known decomposition schemes. Hence, through their combination, we could first detect typical defects that arise during the manufacturing process, which are below the range resolution of the radar. We are convinced that the techniques may trigger several research points based on these investigations. This includes e.g. the hardware development of fully-polarimetric massive MIMO systems to reduce the overall scanning time or the use of polarimetric techniques for non-destructive testing. Apart from the overall decomposition scheme, the entropy seems to be a very promising method for a defect investigation. Moreover, the development of an efficient scanning device or the use of AI methods for the detection of defects seems to be the next step. To conclude, using radar as an alternative or additional sensor for non-destructive testing has proven to be very efficient. We are convinced these techniques could considerably impact current research in this area.