Wireless and Flexible Tactile Sensing Array Based on an Adjustable Resonator with Machine‐Learning Perception

Intelligent soft robotics and wearable electronics require flexible, wireless radio frequency (RF) pressure sensors for human‐like tactile perception of their moving parts. Existing devices face two challenges for array extension: the construction of sensitive units over a limited area and the handling of resonant peaks overlapping within the channel width. Herein, a simply adjustable RF‐resonator‐based tactile array (RFTA) is reported, in which the initial frequency of each resonator unit is regulated by doping polydimethylsiloxane (PDMS) dielectric layers with various concentrations of multiwalled carbon nanotubes (MWCNTs). An array is constructed using four sensor units with a frequency interval of 15 MHz and a multi‐layer micropyramid structure is employed to obtain a low detection limit (<1 Pa) and high sensitivity (17.49 MHz kPa‐1 in the low‐pressure range). A machine‐learning‐based strategy identifies tactile positions precisely via a one‐time S11 reading, achieving 98.5% accuracy with six stimulation modes. Furthermore, the RFTA distinguishes six objects during the grasping process when installed on a soft manipulator. The device shows considerable potential to be extended for flexible moving scenarios and high‐integrated tactile sensing systems for soft robotics.

Intelligent soft robotics and wearable electronics require flexible, wireless radio frequency (RF) pressure sensors for human-like tactile perception of their moving parts. Existing devices face two challenges for array extension: the construction of sensitive units over a limited area and the handling of resonant peaks overlapping within the channel width. Herein, a simply adjustable RF-resonator-based tactile array (RFTA) is reported, in which the initial frequency of each resonator unit is regulated by doping polydimethylsiloxane (PDMS) dielectric layers with various concentrations of multiwalled carbon nanotubes (MWCNTs). An array is constructed using four sensor units with a frequency interval of 15 MHz and a multi-layer micropyramid structure is employed to obtain a low detection limit (<1 Pa) and high sensitivity (17.49 MHz kPa -1 in the low-pressure range). A machine-learningbased strategy identifies tactile positions precisely via a one-time S11 reading, achieving 98.5% accuracy with six stimulation modes. Furthermore, the RFTA distinguishes six objects during the grasping process when installed on a soft manipulator. The device shows considerable potential to be extended for flexible moving scenarios and high-integrated tactile sensing systems for soft robotics.

Introduction
Advances in intelligent soft robotics and wearable electronics have resulted in a growing interest in the development of electronic skin (e-skin) for tactile monitoring platforms, which requires large-scale flexible sensor arrays and suitable data postprocessing to build accurate and versatile interfaces www.advelectronicmat.de unit, which results in a large contradiction between the frequency shift range and the bandwidth of the communication channel. [33] For large-scale arrays, it is challenging to provide a sufficiently wide operating band. It is difficult to obtain devices with a large span of the working band through the same fabrication process, and this span results in higher requirements for the reading equipment.
In this paper, we report a wireless and flexible RFresonator-based tactile array (RFTA) with adjustable resonator units and machine-learning perception. A sample strategic routing was proposed to adjust the initial resonant frequency of each unit by doping polydimethylsiloxane (PDMS) dielectric layers with various concentrations of multiwalled carbon nanotubes (MWCNTs). The RFTA consisted of four sensor units with a frequency interval of 15 MHz between each device. A multilayer micropyramid structure was employed to obtain a low detection limit (≈1 Pa) and high sensitivity (17.48 MHz kPa −1 in the low-pressure range). Furthermore, machine-learning methods such as decision tree, support vector machine (SVM), artificial neural network (ANN), k-nearest neighbor (kNN), and naive Bayes, were introduced to identify the tactile position precisely from the return loss (S11) spectrum. The proposed array was demonstrated through installation on a soft manipulator, and more than six object shapes could be distinguished during the grasping process. The device shows considerable potential for applications in wireless and battery-free tactile sensing systems for soft robotics and machinery.

Mechanism and Analysis
As shown in the schematic in Figure 1a, each sensor unit of the RFTA is constructed from a copper spiral inductor and parallel capacitor, forming an inductor-capacitor resonator. The multilayer micropyramid MWCNT/PDMS structure was introduced as the dielectric of the capacitor, which resulted in a high sensitivity, low detectable threshold, and wide working range. More importantly, with the highly mobile electrons and long conductive paths, doping with MWCNTs can achieve better electronics polarization, thereby increasing the permittivity. Consequently, it become possible to adjust the initial resonant frequency of the sensor without modifying the inductor geometry by quantitatively altering the sensing capacitance's initial value through doping and changing the dielectric constant. Thus, the mutual inductance between adjacent units is the same, which effectively reduces the mutual interference and enhances the availability of signals. As shown in the reading effect diagram in Figure S1 (Supporting Information), in a large working range, there is no visible clutter, so the influence of mutual interference is regarded as eliminated. Finally, wireless communication was realized by a readout antenna that magnetically couples with the sensor, and the resonant frequency of the sensor was obtained from the return loss (S11) spectrum. As shown in Figure 1c,d, machine-learning methods were applied to the perceptual processes to solve the problem of peaks overlap. Subsequently, the RFTA was adopted in prospective applications, such as e-skin, soft robot hands, motion detection, and tire pressure monitoring.
In this design, the relationship between the resonant frequency and the permittivity of the dielectric can be described as follows. According to the equivalent circuit as shown in Figure 1a, the resonant frequency ( f ) and the quality factor (Q) of the RF resonant sensor is expressed as [31] 1 2 where L and C stand for the inductance of the coil and the total capacitance of the sensor, respectively. In our design, L is determined by the coil geometry and is regarded as a constant. The equation can be written as To elucidate the relationship between permittivity and capacitance, the total capacitance (C) was divided into several capacitances of the individual pyramid units (C i ). The pyramid of the layers was considered to be parallel to each other, and the layers were considered to be series. The relationship can be expressed as where n is the pyramid unit number of each layer, and m is the number of layers. C i can be calculated using the simplified electric circuit model and was roughly considered as the parallel of three parts, the composites contribution part (C e1 ), combined action part (C mix ), and air contribution part (C air2 ), as depicted in Figure 1b. The relationship can be expressed as [11,34] i e air m ix where C e , C air , and C mix can be calculated using the parallelplate capacitor formula C s d ε = . Here, s represents the capacitance area and d represents the dielectric thickness, which are determined by the geometric form. C mix was calculated by the continuous integration of the composite to air proportion, and the terms can be presented respectively as  where ε air and ε e represent the permittivities of the air and MWCNTs/PDMS composite; l 1 , l 2 , and l 3 are the lengths of the divided capacitors C e1 , C e2 , and C air2 , respectively; d e and d air are the dielectric thicknesses of the composite and air, respectively; k, k 1 , k 2 , k 3 , and k 4 are constant terms. At the initial state, l 1 , l 2 , l 3 , and h are the given parameters, and hence, Equation As evident from the equation, the modulation of the initial resonant frequency ( f initial ) can be realized by the quantitative adjustment of the composite permittivity (ε e ). At different pressures, the pyramids were compressed into different configurations, leading to the change in l 1 and h, and thus the frequency of the device.  Figure 2a). The MWCNTs/PDMS-doped compound was then spin-coated on a template with a sacrificial polyvinyl alcohol (PVA) layer to construct a pyramid dielectric layer ( Figure S2, Supporting Information). The first step involved cleaning the pyramid template and then spin coating PVA (1:10) on the template surface as the sacrificial layer. After vacuum treatment and curing for 30 min, the PDMS/MWCNT compound was spin coated and cured in 55 °C for 8 h. The PVA sacrificial layer was dissolved to release the pyramid structure layer from the template, which was then peeled off. Finally, the layers were stacked, and polyurethane (PU) tape was used to fix the multilayers.

www.advelectronicmat.de
Second, the copper foil was embossed into a pattern (Figure 2b), and the patterned coil was conformally transferred to the glass sheet with PVA and spin-coated Ecoflex mixture to build stable cross-linking between copper foil and Ecoflex. The patterned copper foil was submerged in the Ecoflex and fully protected by flexible materials so that the entire device was flexible and could be stretched and twisted without damage. Figure 2c depicts the assembly process in which the micropyramid dielectric layer was placed upon the bottom electrode plate before overturning the upper plate and then stuck with glue to form a stable sandwiched architecture. Finally, four sensors with different doping concentrations of the dielectric were self-encapsulated by the stickiness between their Ecoflex substrates to construct a square 2 × 2 pixelated array.
The optical images of the formed micropyramid structure layer, the fabricated sensor, and as-formed RFTA are shown in Figure 2. A pressure testing machine was used to exert pressure on the sensor surface, an LCR meter was used to measure the capacitance to calculate the dielectric permittivity, and a vector network analyzer (VNA) was used to read the return loss (S11) curve of the devices. A field-emission scanning electron microscope (SEM) was used to characterize the morphology of the multilayer micropyramid MWCNTs/PDMS structure.

Effect of Doping MWCNTs Doping on the Resonant Frequency
As the doping process used was fully mechanical, the compound elasticity became significantly worse beyond 5 wt%, and posed a challenge for uniform doping, therefore, a doping threshold of 0 to 5% doping was chosen. Within this threshold, the initial frequency of ≈45 MHz varies according to the selected multilayer pyramid dielectric structure; consequently, the interval of 15 MHz obtained the best effect for quaternary arrays. Moreover, according to this doping method, the dielectric constant of the dielectric layer is continuously tunable; therefore, the required step size can be flexibly designed www.advelectronicmat.de as needed. Before fabricating the device, we calculated the dielectric constant of the compound with different doping concentrations, and the relationship is shown in Figure S3 (Supporting Information). The doping concentration and dielectric constant exhibited a linear relationship with a slope of 1.66. Figure 3a shows the SEM cross-section morphology of the MWCNT/PDMS dielectric layer, where the MWCTNs were well embedded into the PDMS with uniform distribution. In addition, the doped MWCNTs remained in a good elongated shape with cross-links (inset of Figure 3a-iii), which formed complex and long conductive paths to enhance the electron polarization. Figure 3b shows the relative permittivity of dielectric with four different MWCNT concentrations (0, 1.5, 2.9, and 4.8 wt%). As the ratio increased, the permittivity increased and steadily, thus, four devices with an initial frequency interval of 15 MHz could be fabricated (Figure 3c). Figure 3d shows the response curves of the above devices, which show a similar response trend, indicating that appropriate doping hardly disrupts the elastic characteristics of the sensing dielectric. The sensitivity of all the devices can be regarded in three stages, corresponding to the low pressure (0-4 kPa, 15.75-18.19 MHz kPa −1 ), medium pressure (4-20 kPa, 1.75-2.34 MHz kPa −1 ), and high pressure (20-60 kPa, 0.45-0.5 MHz kPa −1 ) ranges. Therefore, the method provides a prospective option for array scaling without the need for complex micromachining, only requiring the replacement of the filler. In future studies, we plan to focus on this topic, exploring the impact of doping processes and parameters, as well as more sophisticated arrays.

Optimization of the Sensitive Pyramid Layer
The multilayer micropyramid structure was simulated with a series of geometric parameters using COMSOL, a finite element analysis (FEA) software, to optimize the device sensitivity. Figure 4a-c shows the relationship between the exerted pressure and the resonant frequency of the device with a single pyramid layer. The pressure sensitivity was significantly improved with an increase in the pyramid separation density, whereas it was barely affected by its height. Furthermore, the pressure sensitivities of different pyramid layers were simulated and compared. The stress distribution and deformation diagrams for the single, double, and triple pyramid layers, are shown in Figure 4d-f, respectively. In our device, the subordinate pyramid supports the upper membrane. Thus, the membrane bending co-acts with pyramid tip compression, which considerably amplifies the deformation. As shown in Figure 4g-i, the changes in the dielectric thickness (Δd) at a stress of 1 kPa are 0.07, 0.2, and 0.35 mm for the single-layer, double-layer, and triple-layer structures, respectively, suggesting amplified capacitance changes, which lead to a higher sensitivity. The pressure sensitivity (S) is defined as the frequency shift (Δf s ) per unit pressure (ΔP), S = Δf s /ΔP. From the simulated S11 curves under different pressures (Figure 4j-l), the pressure sensitivities were calculated as 3.2, 4, and 8 MHz kPa −1 , respectively. The material parameters of PDMS and Ecoflex are presented in Tables S1 and S2 (Supporting Information), and the detailed simulation results of the structural parameters, initial  Figure 5a,b shows the response curve of the resonant frequency changes for the devices with different separation densities and the numbers of layers. All the curves show three sensitivity ranges, and the sensitivities gradually decrease with increasing pressure. As shown in Figure 5a, the augmented separation of each pyramid significantly improves the pressure sensitivity of the devices, which is consistent with the FEA results. Figure 5b shows the sensitivity curves of the devices with different layers based on the optimized height (1:1) and separation density (1:2.5). The sensitivity of the sensor with triple pyramid layers reaches 17.48 MHz kPa −1 in the low-pressure range, which is five times that of the sensor with a single layer (3.66 MHz kPa −1 ). The sensitivity is 1.75 MHz kPa −1 in the mid-pressure range and 0.5 MHz kPa −1 in the high-pressure range. The experimental results of the sensitivity (S), initial frequency ( f 0 ), and shift value under 60 kPa (Δf ) are summarized in Tables S5 and S6 (Supporting Information).

Sensitivity Test of the Devices
This device exhibited a very low detection threshold. The response curve to a tiny applied force from 0.4 to 40 Pa is shown in Figure 5c, where the response to a pressure as low as 1 Pa (20 mg) can be clearly captured, which is approximately the weight of an ant. In addition, the pressure response range is an important factor in determining the tactile perception performance. The sensor exhibited a recoverable and Figure 4. Optimization of the sensitive pyramid layer by using COMSOL. a) Simulated response results of the single-layer sensors with various separation densities (three separation densities corresponding to 4 × 4, 5 × 5, and 6 × 6 pyramids arranged on the area). b) Sensitivity comparison diagram of the simulation results of single-layer sensors with various separation densities and height ratios. c) Simulated pressure response of the single-layer sensors with various height ratios. d-f) Simulated stress distribution of the single-layer, double-layer, and triple-layer devices when a pressure of 1000 Pa was applied on the sensor up-surface. g-i) Deformation distribution and thickness change of the dielectric layer (Δd) under 1000 Pa. j-k) S11 curves and the calculated sensitivity of the device response to applied pressures in the range from 0 to 1000 Pa.

www.advelectronicmat.de
reliable response under extremely high pressure. As shown in Figure 5d, during trampling by the tester (corresponding to a loading pressure of ≈2400 kPa), the S11 curves and frequency shift were visible and consistent when the tester picked up and put down a notebook. Figure 5e shows that the resonant frequency of the sensor remains 283 MHz during tensile from 0% to 20%, which reflects of the frequency stability under stretching conditions. We conducted 500 cycles and extracted the peak value of each S11 spectrum to draw the cycle fluctuation curves. The results show that when a force of 0.2 N was applied or withdrawn, the fluctuation error of the sensor was less than 0.5% and most of the data points were within the region of two standard deviations above and below around the mean value (Mean ± 2SD), and the response showed no fatigue decay ( Figures S4 and S5, Supporting Information). The mean, standard deviations, maximum and minimum value were summarized in Table S7 (Supporting Information). The sensitivity, detection limit, and pressure range of our sensor are compared with several recent works [18,23,26,28,[35][36][37][38][39][40] as listed in Figure 5f and Table S8 (Supporting Information), showing the advantage of the multilayer pyramid design.

Reading Range Extension
A copper ring is commonly used to read resonant characteristics. However, its readout area and quality are hampered by the poor coupling between the ring and the spiral inductance of the sensor. In this study, we designed a planar monopole antenna to extend the reading area and quality. The readout

www.advelectronicmat.de
performance was tested at different vertical heights and horizontal offsets, as shown in Figure 5g. A magnitude of −2 dB at a horizontal offset of 3 cm was obtained by the monopole antenna, whereas the signal received by the ring antenna was hardly read at an offset greater than 1.5 cm. Furthermore, the monopole antenna received a magnitude greater than −4 dB at heights of 2-5 cm heights, but the magnitude of the ring antenna decreased sharply from 2.5 cm, and reduce to less than −1 dB at 4 cm (Figure 5h).

Tactile Recognition Based on Machine Learning
If the working band of a sensor unit covers others, the resonant peaks will overlap. When the array resonance peaks overlap, it is difficult to determine each peak value, but machine-learning methods can help solve this problem. Machine-learning method is used to capture more comprehensive features of the spectrum, such as the waveform changes. This ability has been proven by significant achievements in the field of image recognition. Thus, even if the resonance peaks overlap, the identified missions can be quickly and effectively achieved by the algorithm. Consequently, this benefit can also help improve the utilization of finite frequency bands, that is, to increase the scale in a certain frequency band. This can also be combined with the physical basis of expanding the frequency band to increase the scale of units further. Figure 6 shows the 2 × 2 RFTA and the recognition results for the tactile position. When a metal weight of 50 g (≈2 kPa stimulation) was used to simulate the tactile action on one of the units in the array, the corresponding resonant peak shifted significantly, whereas the peaks of other units remained around their initial positions (Figure 6b). The magnitude reduction accompanied by a frequency downshift was a general pattern in the experiments. First, this is due to the introduction of an additional parasitic load and the reduction of the resonant quality factor. But this phenomenon is not always negative and can lead to more features that enhance machine-learning cognition. Second, as the position of the antenna is not completely fixed, its spatial position within the four array units affects its The S11 curves show each frequency initial peak by shadow and its offset positions by arrows. The resonant frequency peaks change with the response of a particular sensor unit to the external force loading. f 0 represents the initial resonant frequency of the sensor, and f represents the frequency of the sensor during the pressing test. c) Process of machine learning, including data standardizing for input, training and cross-validation, and predicting for output. d) Accuracy comparison of several machine-learning methods for pressure position recognition. www.advelectronicmat.de relative magnitude scale. In the elongated pictures, the weaker peak appears less apparent, but it still has a magnitude larger than −1 dB, which is sufficient for determining the resonant peak. Therefore, the reduction in the resonance peak magnitude did not cause a problem for our sensor array.
Five typical machine-learning strategies were adopted, including decision tree, support vector machine (SVM), artificial neural network (ANN), k-nearest neighbor (kNN), and naive Bayes, to recognize the tactile position rapidly and accurately. As shown in Figure 6c, datasets captured from nonstimulated, single-stimulated, and bi-stimulated states were input into the programs for training, validation, and testing. In the experiment, 180 sets of data were used as the training sets and fivefold crossvalidation was performed to verify the feasibility of each algorithm. A larger test set of data from 600 independent experiments was then used to verify and compare the generalizations of the learning methods. Figure 6d summarizes the recognition accuracy of the 12 typical classifiers by cross-validation and set-aside tests. The cross-validation results show that the experimental data of the training set were collected under limited experimental conditions (such as reading distance and environmental noise), therefore, the accuracy was extremely high (100%). In terms of device level, the array and recognition methods are easily competent, and benefit from the high stability of the sensor array under experimental conditions. Subsequently, the data of the test set were collected without limiting the reading distance and were input into the previously trained model for testing. The two-color comparison in Figure 6d shows that KNN and ANN methods have good generalization and can still achieve accuracies of 98% and 98.5% for tactile position recognition in completely unfamiliar data and test environments. However, some algorithms have poor generalization after training, such as NBC (100% to 69.9%) and Gau-SVM (only 60.9%). As the global feature of the S11 curve is handed by the machine-learning method, the recognition process can be completed by reading the spectrum at one time, which is more effective and accurate than the ordinary method that only considers the frequency parameters, particularly when the resonant peaks overlap.

Pressure Detection for Moving Body and Machine
Based on the flexible and stretchable structure, the proposed sensor closely contacts the skin surface, allowing good responses to both subtle expressions and large-scale joint flexion. As shown in Figure 7a, the sensor was attached to the outside surfaces of the wrist, knee, and elbow to monitor the different bending states. Bending produces an equivalent pressure applied to the sensor and leads to an apparent frequency shift (≈60 MHz) with good repeatability after the body returns to its initial state. To demonstrate the sensing performance for tiny stresses, several application experiments for facial expressions have been conducted, including smiling, frowning, and pronouncing ( Figure 7b). The frequency shifts by 15-25 MHz because of the tiny pressure produced by the expression. These results suggest the application potential of our proposed sensor for prosthetics and humanoid robotics. Figure 7c,d demonstrates real-time tire pressure monitoring using our device. The sensor was implanted between the tread rubber and inner liner of a bicycle tire. The on the right in Figure 7c shows the signal strength from the rotating sensor as it approaches and moves away from the fixed readout antenna. When the sensor is facing the antenna, the peak intensity reaches a maximum value. As the sensor gradually rolls away from the antenna, the signal peak attenuates but the resonant frequency remains the same. As shown in Figure 7d, with the increase in tire pressure from 5 to 40 Psi, the resonant frequency shifts from 191.35 to 180.65 MHz with high stability during the wheel rotation. Thus, tire pressure can be continuously and periodically monitored while riding, which clearly shows the potential for long-term and noninvasive tire pressure monitoring in intelligent driving systems.

Tactile Recognition for Soft Robot Grasping
An RFTA with two tactile pixels was mounted on one finger of a soft robot hand, as shown in Figure 8a, to classify and recognize various geometrical objects. The frequency and waveform of the resonant peaks varied owing to the different force distributions when grasping different objects (Figure 8b). When grasping small objects, the pressure was mainly concentrated on the fingertips (sensor A), as shown in Figure 8b-i. As the volume increased (Figure 8b-i-v), the pressure acting on the lower sensor (sensor B) increased gradually. For an object large enough for intimate contact by the robot hand, the force distribution varied from the outer surface curvature, as shown in Figure 8bv,vi. Thus, we could recognize the objects using the different strength distributions of the sensors.
However, as shown in Figure 8c, it is difficult to extract the resonant peak of an individual unit from the S11 spectrum. The machine-learning methods were utilized to classify six objects based on 350 sets of data for the accurate and efficient recognition of objects. The first 175 sets of data were used as the training set, and the remaining 175 sets of data were used as the test set. As the results presented in Figure 8d, three types of KNNs and a single-layer ANN (sing-ANN) showed the highest accuracy (92.6-94.9%). Figure 8e,f shows the frame schematic and confusion matrix of the single-layer ANN (92.6%) and precision KNN (93.1%). Both methods incorrectly recognized the 4 and 6 cm balls, which may be because of the ambiguous features of the similar-sized balls; however, the 92.6% recognition rate was well within the requirements of tactile sensing. Further improvements can be made in future study based on the co-optimization of the device design and algorithm. Based on the experimental results, applying higher-resolution arrays and more advanced algorithms can help us identify more complex objects.

Conclusion
We proposed a simple adjustable RFTA that does not require a complex micromachining process by regulating the initial frequency through doping various concentrations of MWCNTs into the PDMS dielectric layer. Four sensor units with frequency intervals of 15 MHz were constructed as an array, and a multilayer micropyramid structure was employed to obtain a low detection limit (<1 Pa) and wide working range (up to ≈2.4 MPa). To overcome the overlap problem, a machinelearning-based strategy was used for tactile position and object recognition through a one-time reading of the return loss (S11) spectrum of the resonator array, with an accuracy of 98.5% to unfamiliar test sets. Given the flexibility of frequency adjustment methods and rapid tactile recognition strategies, the methodology of this device shows considerable potential for applications in highly integrated and large-scale tactile sensing systems for soft robotics. Applications of a single pressure unit. a) S11 curves responses to the different bending states of limbs, including i) wrist, ii) elbow, and iii) knee. b) Resonant frequency curves corresponding to i) smiling, ii) frowning, and iii) pronouncing. c) Schematic of a tire pressure monitoring system and the S11 magnitudes of three rotation states with the sensor facing the antenna and moving away from the antenna. d) Bicycle tire pressure monitoring experiment, with pressure from 5 to 40 Psi (1 Psi ≈ 6.89 kPa). www.advelectronicmat.de

Experimental Section
Fabrication of Dielectric: MWCNTs (XFM06, XFNano Materials Tech Co., Ltd., Nanjing, China, <8 nm diameter, >95% purity) and PDMS (curing agent ratio 1:15, Sylgard 184 Silicone Elastomer Kit, Dow Corning, Shanghai, China) were mixed thoroughly in ratios of 0-4.8 wt% by agitation and sonication (3 h). The PDMS/MWCNT-doped compound was then spin-coated on a template with a PVA sacrificial layer Figure 8. Grasping test with machine-learning recognition. a) Optical photo of the prepared test platform with sensors, reading antenna, and the soft robot hand. b) Distribution of the pressure density of the upper and lower sensors when the object is 2, 3, 4, 5, and 6 cm spheres and a 6 cm cube. c) Typical S11 curves for unloading and grasping various objects. d) Accuracy comparison of several machine-learning methods for object grasping recognition. e) Schematic of the Sing-ANN classifier and its confusion matrix; the training and test sets are independent 175 groups of data, and the accuracy is 92.6%. f) Schematic of the precision kNN (Pre-KNN) classifier and its confusion matrix; the training and test sets are independent 175 groups of data, and the accuracy is 93.1%.

www.advelectronicmat.de
(50 mg mL −1 , Shandong Yousuo Chemical Tech Co., Ltd., Shandong, China) and vacuumed for 30 min. Subsequently, an ultrasonic bath was used to remove PVA, and the micropyramid dielectric layer was peeled off from the template.
Fabrication of the Sensor: The patterned copper foil was tightly combined with an Ecoflex (Smooth-0030, Smooth-On Inc., Macungie, PA, USA, curing in a vacuum chamber at 80 °C for 60 min) substrate into an inalienable component. The patterned coil was conformally transferred to a glass sheet coated with PVA. Subsequently, the vacuum-treated Ecoflex mixture was uniformly spin-coated and cured (in a vacuum chamber at 80 °C for 60 min) to build stable cross-linking between the copper foil and Ecoflex. The micropyramid dielectric layer was placed upon the bottom electrode plate before overturning the upper plate and then stuck with glue to form a stable sandwich architecture.
Data Acquisition: A pressure-sensitivity test system was constructed using a VNA (Keysight Tech. Co. Ltd.), a set of standard weights (from 1 to 50 g), and a pressure-testing machine (ZQ-990A, 0-10 kg, Zhiqu Precision Instruments) (Figure 3g). The sensor was placed horizontally on an acrylic platform with an RF antenna for data acquisition. The readout antenna was connected to the VNA via a coaxial cable. A field-emission SEM was used to characterize the morphology of the structures.

Supporting Information
Supporting Information is available from the Wiley Online Library or from the author.