Interdigitated capacitive strain sensor enables precise yoga-inspired motion tracking

Abstract

We report a fully textile-based interdigitated capacitive (IDC) strain sensor for wearable posture recognition. The sensor is lightweight, comfortable, and washable, and maintained stable electromechanical performance during 6,500 strain cycles, with absolute value of the gauge factor ranging from 0.67 to 0.81 across the tested strain conditions. Integrated into clothing, the sensor captured multidimensional electrical responses, including capacitance-, resistance-, and phase-related signals, during four biomechanically distinct yoga postures: High Lunge, Lunge Forward, Squat, and Tree. Using a supervised machine-learning pipeline combining handcrafted sensor-channel features and MiniROCKET time-series features, record-level classification achieved an accuracy of 94.4% and a macro-averaged F1 score of 94.2%. Record-level five-fold cross-validation further showed stable performance across folds, supporting the robustness of the classification framework within the present controlled cohort. One-vs-rest receiver operating characteristic analysis yielded area under the curve values of 90%, 94%, 99%, and 99% for High Lunge, Lunge Forward, Squat, and Tree, respectively. Channel ablation analysis identified capacitance-derived features as the most informative predictors. Statistical and temporal analyses further showed posture-dependent differences in signal magnitude, normalized response, and waveform dynamics, while supervised low-dimensional projection revealed distinct class structure. These results demonstrate the feasibility of textile-based IDC sensing for wearable movement monitoring applications such as such digital fitness.

Introduction

Flexible and stretchable wearable strain sensors that convert mechanical deformations to electrical signals have attracted considerable attention due to their flexibility and mechanical sensitivity^1–4. Unlike rigid sensors, which can withstand less than 5% strain⁵, soft wearable strain sensors could be used to perform motion detection on curved surfaces^6,7. Sensitive and stretchable wearable strain sensors capable of enduring continuous and repetitive mechanical deformations without performance loss are highly valuable for practical wearable technology applications^8,9.

Strain sensors, such as resistive strain gauges^10,11, capacitive strain sensors^12–15, piezoresistive sensors^16–18, and optical strain sensors^19–21, are designed to measure mechanical deformation by applying strain. Resistive and piezoelectric sensors have been extensively studied and shown to provide accuracy and reliability in strain-sensing applications^22–25. Among them, capacitive strain sensors are particularly promising, as they have good stretchability and sensitivity and modest hysteresis²⁶. The typical parallel plate-type capacitive strain sensor usually facilitates a linear relationship between sensitivity and strain with a theoretically maximum gauge factor GF value of 1 and minimal hysteresis behavior²⁷. A promising structure for capacitive strain sensing is the interdigitated capacitive (IDC) sensor, which responds to mechanical stress by changing their geometric characteristics^28,29. They are typically designed to detect changes in the capacitance of an object’s proximity^30,31 due to factors such as variations in the dielectric properties of the material surrounding the sensor³². Supporting their potential for IDC sensors that can be used in wearable strain sensing applications, as they can conform to curved surfaces with minimal crosstalk with pressure^33–35. Integrating them into a flexible or deformable material can increase their sensitivity and achieve accurate measurements close to the body in contrast to typical manufacturing methods on limited inflexible substrates affixed to flat surfaces, which often leads to limited stretchability^31,35–38. This makes them suitable for real-time measurement of physical activity³⁹. As the popularity of wearable electronics grows, there is a need for sensors that can stretch and bend without being damaged and can be used on different surfaces with various shapes^40–45. The adaptability of stretchable sensors to intricate forms and distortions makes their direct integration into garments possible, providing design flexibility for a wide range of wearable electronic applications⁴⁶. Strain sensors have been integrated into textiles produced using various methods, such as weaving^47,48, embroidery^49–52, and knitting^34,53,54. Due to the omnipresence of clothing in daily life, textile-based systems are particularly attractive for various activities and applications, such as sports and healthcare, as they provide comfort and can mimic natural movements⁵⁵.

Wearable e-textiles can be divided into four generations of development. The first generation is based on a challenging design, with parts such as antennas and electronic components stitched or attached to the garment’s surface using conductive yarn^56,57. This results in bulky and rigid products, making them impractical for everyday use. In second-generation wearables, conductive yarns are knitted or woven into the textile to make it electronically functional. Industrial-scale knitting machines are used to intertwine conductive yarns into various stitch patterns, creating fully functional and comfortable e-textiles^54,58. Third-generation e-textiles have embedded sensors and other electronic components⁵⁹. They are more durable, comfortable, and reliable than first- and second-generation e-textiles and provide better functionality. Fourth-generation e-textiles should be soft and even more comfortable, flexible, and durable⁶⁰. Developing this next generation of e-textiles requires significant innovation but also presents substantial growth opportunities. In the coming decade, wearable textile technology will increasingly interact with the Internet of Things, artificial intelligence, human-machine interfaces, and cloud technology⁶¹. However, despite global efforts to produce wearable sensors, several challenges remain in areas such as stretchability, ease and efficiency of production, breathability, unobtrusive design, ease of use, integration into various clothing items, and, in addition, endurance, environmental bio-compatibility, longevity after environmental exposure, and integration into textile substrates.

As e-textiles evolve from laboratory-scale prototypes toward real-world wearable applications, recent work has further highlighted the role of wearable sensing platforms in enabling continuous physiological monitoring and digital health analytics, emphasizing the need for robust and comfortable sensing technologies integrated directly into garments⁶².

In this study, we develop a wearable IDC sensor that adapts to body movements and the natural flexibility of textiles and can thus be used seamlessly in everyday activities. We precisely fabricated this sensor using laser cutting and an easy and scalable technique. The interdigitated design enables strain sensing and demonstrates stable performance during 6500 strain cycles. The strain-sensitive sensor was also tested under harsh environmental conditions, including multiple washing cycles, high humidity, and varying temperatures.

Results

Figure 1 presents the full architecture of our fully-textile interdigitated capacitive strain sensor and the custom electromechanical testing setup. Figure 1a shows the design of the fully textile interdigitated capacitive strain sensor together with a zoomed-in SEM image of the conductive textile fibers. Figure 1b illustrates the in-plane strain-induced deformation of the sensor structure accompanied by changes in the electric field distribution. We begin by quantifying the sensor’s baseline capacitance and gauge factor under controlled uniaxial strain (0–30%). Next, we evaluate dynamic response, cycle-to-cycle stability over 6500 loading-unloading cycles, and environmental robustness (humidity and temperature). Together, these measurements demonstrate the sensor’s high sensitivity, repeatability, and suitability for wearable motion tracking and movement recognition.

Fig. 1. Conceptual design of the fully-textile interdigitated capacitive (IDC) strain sensor.

Both the electrodes and the substrate are fabricated entirely from Lycra fibers, and the electric field distribution above the surface, showing the fringing fields between electrode fingers. Adapted and modified with permission from ACS Applied Materials & Interfaces, 2017³⁵.

In an IDC sensor, two sets of electrodes are interleaved, and the resulting capacitance depends on the number of electrodes, their width, spacing, and the material properties (simplified analytical expression)³⁶.

$$C_0={\epsilon }_0{\epsilon }_r\frac{\left(2l_0-W_0+2w_0\right)}{d_0}{t}_0\left(2n-1\right)$$ 1

where W₀ is the total width of the IDC sensor, w₀ is the width of the electrodes, l₀ is the length of the electrodes, t₀ is the electrode thickness, d₀ is the spacing between adjacent electrodes, and n is the number of electrodes.

When the strain gauge is subjected to a strain ε in the x-direction, perpendicular to the electrodes, the geometrical term (2l₀ − W₀ + 2w₀) is assumed to decrease by a factor of (1 − νε). Similarly, the electrode thickness t₀ is expected to decrease by a factor of (1 − νε), where ν is the Poisson ratio. In contrast, the electrode spacing d₀ increases by a factor of (1 + ε) relative to its initial value.

$$C_{{\mathbf{IDC}}_x}={\epsilon }_0{\epsilon }_r\frac{\left(1-\nu \epsilon \right)\left(2l_0-W_0+2w_0\right)\left(1-\nu \epsilon \right)}{d_0\left(1+\epsilon \right)}{t}_0\left(2n-1\right)$$ 2

$$G{F}_{{\mathbf{IDC}}_x}=\frac{\mathrm{\Delta }C/C_0}{\epsilon }=\frac{1}{\epsilon }\left[\frac{\left(1-\nu \epsilon \right)\left(1-\nu \epsilon \right)}{\left(1+\epsilon \right)}-1\right]$$ 3

When a IDC strain gauge is subjected to strain ε in the y-direction along the electrodes, the term 2l₀ − W₀ + 2w₀ which is the charging between two opposite electrodes increased by factor (1 + νε) and t₀ decrease by a factor (1 − νε), while, d₀ is expressed in decreasing by a factor of (1 − ε) from the initial values.

$$C_{{\mathbf{IDC}}_y}={\epsilon }_0{\epsilon }_r\frac{\left(1-\nu \epsilon \right)\left(2l_0-W_0+2w_0\right)\left(1+\nu \epsilon \right)}{d_0\left(1-\epsilon \right)}{t}_0\left(2n-1\right)$$ 4

$$G{F}_{{\mathbf{IDC}}_y}=\frac{\mathrm{\Delta }C/C_0}{\epsilon }=\frac{1}{\epsilon }\left[\frac{\left(1-\nu \epsilon \right)\left(1+\nu \epsilon \right)}{\left(1-\epsilon \right)}-1\right]$$ 5

Based on the above analysis when the IDC strain sensor is stretched in plane, x- or y-directions, the GF is independent of the intermediate transducer dimensions (e.g., spacing, length, and thickness of the electrodes) and depends only on the strain (ϵ) and the elastic properties, or the Poisson ratio, of the materials used. Stretching the capacitor in x-direction (along the electrode fingers) increases the gap between adjacent fingers while reducing their thickness and length. Combined, the changes produced by applying strain decrease the capacitance of the IDC sensor. In contrast, stretching the capacitor in y-direction (across the electrode fingers), the distance between any two electrodes decreases, and their thickness also decreases while their length increases. Combined, these changes increase the capacitance of the IDC sensor. This analysis shows that the sensitivity of the IDC sensor varies depending on the direction of strain. As a result, the sensor exhibits anisotropic mechanical behavior, resulting in different stretching behaviors along the x- and y-direction⁶³. The directional difference is illustrated in Supplementary Fig. S1. Upon application of a potential difference, a uniform electric field is generated between the positive and negative electrodes. COMSOL simulations confirm the presence of the electric field across the electrode surfaces (Supplementary Fig. S2). The remainder of this study focuses on the IDC sensor’s response under tensile strain applied along the direction exhibiting the higher gauge factor. A schematic illustration of the sensor under relaxed and stretched conditions, as well as a summary of the sensor dimensions, are provided in Supplementary Fig. S3 and Supplementary Table S1, respectively.

Experimental characterization of the IDC strain sensor

Multifunctional wearable electronics are attractive because they can be lightweight, soft, flexible, stretchable, air-permeable, low-cost, and compatible with scalable fabrication and garment integration^64,65. The performance of capacitive sensors depends strongly on the constituent materials, and the choice of textile substrate should therefore reflect the requirements of the intended wearable application². On this basis, we selected a low-cost textile substrate based on Spandex fibers together with conductive silver fabric as the electrode material. The textile substrate is inherently breathable and lightweight (200 g/m²) and exhibits high extensibility, with reported stretchability of ~80% in the transverse direction and 70% in the longitudinal direction, which is advantageous for wearable motion sensing compared with less extensible textile fibers such as cotton, silk, or wool^66,67. In addition, its textile nature enables conformal integration onto curved and dynamic body surfaces using simple joining and garment-integration methods. These properties make the proposed IDC architecture well suited for wearable electronics and soft robotic applications in which mechanical adaptability is essential⁶⁸. In sportswear applications, Lycra fabrics containing elastane fibers can conform closely to body contours and recover their original shape after stretching, making them attractive for future health-monitoring and high-performance textile systems. The dielectric textile used here also showed stable behavior across a broad frequency range, as shown in Supplementary Fig. S4.

We fabricated seven nominally identical IDC strain sensors, each used for a defined set of test conditions to assess performance across different operating scenarios. A detailed description of the fabrication process is provided in Supplementary Fig. S5. Under ambient conditions, the mean initial capacitance was 4.00 ± 1.3 pF. Because interdigitated capacitive sensors are sensitive to geometric variation, even small differences in electrode placement or spacing can influence the baseline capacitance. We therefore assessed reproducibility under identical conditions by collecting parallel datasets from three independently fabricated sensors and comparing their electromechanical responses statistically. The LCR meter was operated at an excitation frequency of 1 kHz, while the LabVIEW-based data acquisition system sampled the electrical response at 100 Hz. This sampling rate provided sufficient temporal resolution to capture strain-induced changes while minimizing noise and parasitic effects associated with the slow deformations investigated in this study. To measure sensitivity, the samples were clamped in a linear testing system, as illustrated in Supplementary Fig. S6 and schematically in Fig. 2a, and strain levels of 10%, 20%, and 30% were applied along the x-direction. This orientation was maintained throughout all tests to ensure reproducibility. Across three repeated measurements (Supplementary Fig. S7), consistent sensor performance was observed. The mean initial capacitance varied from 3.75 to 3.76 pF, corresponding to relative differences below 0.2%, with standard deviations of ~2.4 × 10⁻¹³ F. Skewness and kurtosis were also similar across measurements, further supporting the precision and repeatability of the sensor response.

Fig. 2. Characterization of the mechanical behavior of the interdigitated capacitive sensor under different strain conditions.

a Schematic illustration of the mechanical characterization setup. The IDC sensor was fabricated via a facile and efficient laser cutting method and characterized under controlled environmental conditions using tensile testing and an LCR meter. The purpose of this experimental setup was to synchronously obtain the sensor’s mechanical data and electrical signal output. b Hysteresis curves of the IDC sensor under 50 loading/unloading cycles of 10%, 20%, and 30% strain in the x-direction. Inset: Enlarged view of the change in the sensor’s capacitance after linear fitting. c Output characteristics of the sensor when subjected to 6000 30% strain cycles in the x-direction. Relative change in capacitance under cyclic 30% sinusoidal strain at 1 Hz. Insets: Enlarged views of the first 2000–2020 and 5050–5070 cycles. d Cyclic tensile testing under 10%, 20%, and 30% strain in the x-direction repeated 50 times.

IDC sensors are inherently sensitive to in-plane strain because mechanical deformation alters the spacing and alignment of adjacent electrodes. When the substrate is stretched, these geometric changes produce measurable variations in capacitance, making IDC architectures well suited for strain sensing. With appropriate geometric design, the capacitive response can be tuned to the desired strain range⁶³. The interdigitated geometry, which consists of alternating electrode fingers patterned on a substrate, increases both the effective sensing area and the fringing electric-field interaction, thereby enhancing strain sensitivity. A gauge factor GF of −0.81 was obtained at 10% strain, demonstrating performance consistent with established textile-based sensing technologies, indicating their suitability for practical applications. In absolute terms, the gauge factor was highest at 10% strain and gradually decreased as strain increased from 20% to 30% (Fig. 2b, d). This suggests high sensitivity to low strains, enabling the detection of different movement amplitudes. Overall, the IDC sensor had GF of −0.67 under 30% strain, −0.71 under 20% strain, and −0.81 under 10% strain. Sensitivity GF and stretchability alone offer limited insight into the overall performance of a sensor, highlighting the need to consider other parameters, such as hysteresis, linearity, fatigue, and dynamic/static state. The equation governing sensitivity, GF = (ΔC/C₀)/ε, is valid for linear elastic dielectric materials between electrodes and does not capture nonlinear variations in compliance or other factors that influence the sensor’s performance. Herein, in most real-world scenarios, the sensitivity would exhibit non-linear behavior, with nonlinear material or geometric effects⁶⁹.

To evaluate cyclic hysteresis, we applied 50 loading–unloading cycles at strain levels of 10%, 20%, and 30%. The sensor exhibited measurable hysteresis, as shown by the enclosed area between the stretching and release curves in Figure 2b. In each cycle, the absolute value of ΔC/C₀ decreased during stretching and increased during release. As expected, hysteresis increased with increasing strain. A detailed description of the hysteresis-area calculation under the different strain conditions is provided in Supplementary Fig. S8 and Supplementary Table S2. The hystereis-related calculations were performed based on Supplementary Eq. (1) and Supplementary Eq. (2). These hysteresis effects can be attributed to strain-induced changes in the dielectric properties of the material⁷⁰. Hysteresis is often associated with internal friction, arising from microscopic energy dissipation caused by sliding and interactions among particles within the material⁷¹. Another possible contributor is electrostriction, which results from ion polarization within a material⁷² and may account for part of the hysteretic response. Although the Lycra dielectric used in this study is not expected to exhibit a strong electrostrictive effect, such behavior may emerge under repeated mechanical and electrical stimulation. Encapsulation has been suggested as a practical strategy to reduce hysteresis and improve cyclic reliability^71,73. As shown in the inset of Figure 2b, the IDC strain sensor also exhibited an approximately linear response in relative capacitance, ΔC/C₀, over the tested range. We further evaluated mechanical durability by subjecting the sensor to more than 6000 stretch–release cycles using a mechanical tester. The sensor demonstrated excellent repeatability and stability over 6000 cycles at 30% strain (Fig. 2c), while maintaining consistent gauge factors across 10%, 20%, and 30% strain conditions (Fig. 2d). However, translating these controlled laboratory results to large-scale manufacturing remains challenging, because the present characterization was performed under tightly regulated conditions and may not fully capture variability introduced during mass production. In practical settings, additional factors may affect sensor performance, including fabrication variability, garment integration, and environmental exposure. These considerations motivate the subsequent evaluation of environmental robustness, long-term stability, and integration into wearable textile systems.

In addition to sensitivity, hysteresis, linearity, and resistance to mechanical durability, integration into textiles requires additional properties, such as washability. In many application scenarios, electronic textiles will need regular cleaning procedures, as they are required to be worn on clothes. To evaluate the durability after multiple washing cycles, we assess the washability of our fabricated sensor by measuring relative changes in its capacitance, applying the conditions outlined in the IEC 63203-204-1:2023 standard. The procedure involved subjecting the samples to several wash cycles under identical conditions. Specifically, we washed the samples 25 times with a 2 kg load using an industrial detergent for 30 min at 30 °C and a spinning speed of 400 rpm and then air dried them. We measured the capacitance of the samples before and after each wash cycle. The recorded data are at intervals after washing (6, 12, 18, 24 cycles) are shown in Supplementary Fig. S9. There are also SEM images at 1.00 kV with a scale bar of 100 μm. In that figure, we provide the capacitance values at 30% strain, including the pre-washed sample (0 washing-cycles). The capacitance results exhibit an unusual trend, showing a higher capacitance after 24 washing cycles. This trend is crucial to understanding the internal mechanism of electro-fibers that led to this observation. The performance of the capacitive sensors largely depends on the fibers made out of nylon, and any changes in the fiber structure or any chemical modifications will lead to changes in the performance. The chemical structure of Nylon is polyamide with repeating amide groups, which can form hydrogen bonds with water, making Nylon able to absorb moisture⁷⁴. When nylon absorbs moisture, it undergoes a physical change, e.g. swelling and increasing the space between electric fibers. However, conductivity decreased after the 24th washing cycle. That could potentially arise because the conductive pathways on the surface are damaged/oxidized, which could come from environmental exposure to the tough environment with detergent and mechanical stresses of other clothing. Especially when the conductive fibers on the fabric surface are damaged after surface exposure. A more detailed description of the capacitance mechanism and water regain is provided in Supplementary Eqs. (3) and (4) for completeness, although no quantitative calculations were performed. Generally, in the literature, various procedures for observing endurance after multiple washing cycles (washability) are described, however, standards, different testing methods are compared and evaluated, and there is an urgent need for standardized testing, which provides e-textile to fairly lead to conclusions and broader observations^75,76.

Given the broad range of wearable sensing applications for long-term and everyday monitoring, stability under environmental stressors such as temperature, humidity, and sweat is essential. The sensor showed a stable response over the tested temperature range of 20–45 ^∘C under controlled chamber conditions. To assess the effect of moisture, distilled water was sprayed twice onto the sensor surface, which increased the capacitance from 5.37 pF to 4.5 nF. To evaluate the influence of sweat, synthetic sweat (DIN 53160-2:2010-10) was applied to the participant’s forearm, and the sensor response was recorded under resting conditions. After sweat application, the baseline capacitance increased from 9.64 pF to 32.54 pF at 1 kHz and room temperature 25⁰C (higher frequencies was used to reduce electrode polarization and improve measurement stability.)   (Supplementary Fig. S10). These increases are consistent with strong moisture uptake by the textile substrate. Absorbed liquid replaces air within the porous textile matrix, thereby increasing the effective permittivity of the sensing layer because distilled water and artificial sweat have much higher dielectric constants than air. Further details are provided in the Supplementary Information under “Electric Field of the IDC Sensor”. In addition, the ionic content of artificial sweat may further influence the dielectric behavior of the sensing layer. In both conditions, the capacitance returned approximately to baseline after 2 h. A comparison table of textile-based sensors with washing durability and/or environmental testing is provided in the Supplementary Table S3.

Wearable posture recognition using IDC sensors

The supervised classification model used a combined feature representation consisting of handcrafted electrical sensing features from seven sensor channels (Co, CP, PHASE, RS, DC, DC/Co, and DC/Co (%)) together with MiniROCKET-derived time-series features extracted from the capacitance signal. These complementary features capture both static electrical properties and temporal dynamics of the textile IDC sensor response.

The confusion matrix shown in Fig. 3a demonstrates strong classification performance across all four posture classes. At the record level, only two misclassifications were observed, resulting in an overall accuracy of 94.4% and a macro-averaged F1 score of 94.2%. All Squat and Tree recordings were correctly identified, while one Lunge Forward instance was misclassified as High Lunge and another as Squat. The results indicate reliable discrimination between the different movement patterns captured by the wearable IDC sensors.

Fig. 3. Record-level classification performance for yoga posture recognition using textile IDC sensors.

a Record-level confusion matrix illustrating classification results across the four movement classes (High Lunge, Lunge Forward, Squat, and Tree). b Channel ablation study comparing the contribution of individual electrical sensing channels to classification accuracy. Capacitance-derived signals (Co, DC/Co, and DC/Co (%)) provide strong predictive capability, while models relying solely on phase or resistance signals perform substantially worse. c Record-level five-fold cross-validation accuracy presented as a boxplot, demonstrating consistent model performance across folds. d One-vs-rest ROC curves for each posture class, showing strong discriminative capability of the trained classifier.

To further investigate the contribution of individual electrical sensing channels, a channel ablation analysis was conducted (Fig. 3b). Models were trained using individual signals and selected combinations of channels to evaluate their relative importance for posture recognition. The full multi-channel configuration achieved the highest performance with a record-level accuracy of 91.7%. Among the individual sensing modalities, the CP signal yielded the strongest predictive capability (88.9%), followed by capacitance-derived parameters including Co and normalized capacitance ratios (DC/Co and DC/Co (%)), each achieving approximately 86.1% accuracy. In contrast, models relying solely on resistance (RS) or phase (PHASE) signals showed substantially lower accuracies of 77.8% and 66.7%, respectively. These findings confirm that capacitance-based measurements provide the dominant sensing mechanism for posture recognition using textile IDC electrodes, while impedance phase and resistance signals provide complementary but weaker discriminatory information.

The robustness of the classification framework was further evaluated using record-level five-fold cross-validation (Fig. 3c). The cross-validation results demonstrate stable performance across folds, with accuracies consistently above 90% and low variance, indicating reliable generalization of the model across different subsets of the dataset.

Receiver operating characteristic (ROC) analysis (Fig. 3d) further illustrates the discriminative capability of the trained model. The one-vs-rest ROC curves show strong separation between posture classes, with area under the curve (AUC) values of 90% for High Lunge, 94% for Lunge Forward, 99% for Squat, and 99% for Tree. These results demonstrate that the wearable IDC sensing system captures distinct electrical signatures associated with different biomechanical postures, enabling accurate posture classification using the proposed machine learning pipeline.

Figure 4 presents a multi-level characterization of posture-specific responses derived from textile-based interdigitated capacitive sensors. Four yoga postures–High Lunge, Lunge Forward, Squat, and Tree–were compared across distributional features, normalized temporal trajectories, and supervised low-dimensional projections to assess whether the sensor outputs preserve posture-dependent structure in both amplitude and feature space.

Fig. 4. Feature-level characterization of yoga postures using capacitive sensing.

a Distribution of ${\log }_{1p}\left(\mid \mathrm{\Delta }C\mid \right)$ across movements. b Distribution of normalized capacitance changes, ${\log }_{1p}\left(\mid \mathrm{\Delta }C/C_0\mid \right)$, across movements. c Normalized temporal traces of ΔC within a 10 s window starting at the 8th second of the recording. d Supervised linear discriminant analysis (LDA) projection of participant-level feature vectors, color-coded by posture, illustrating visually separated class structure in the reduced feature space.

Figure 4a shows the distribution of ${\log }_{1p}\left(\mid \mathrm{\Delta }C\mid \right)$, representing the magnitude of posture-related capacitance changes across the four movements. Tree displayed the largest upper-range values because of a prominent high-response outlier, whereas the remaining postures were concentrated near lower values with partially overlapping interquartile ranges. Despite this compression in scale, the overall Kruskal–Wallis test indicated a significant between-posture effect (p = 0.0283), supporting the presence of posture-dependent differences in the raw capacitance magnitude.

To reduce sensitivity to baseline capacitance differences across recordings, Fig. 4b presents the normalized feature ${\log }_{1p}\left(\mid \mathrm{\Delta }C/C_0\mid \right)$. After normalization, the distributions remained posture-dependent, with Tree and Lunge Forward generally showing higher values than High Lunge and Squat. The global Kruskal–Wallis test again remained significant (p = 0.0231), indicating that normalization preserved discriminative information while placing the responses on a more comparable relative scale.

Temporal structure is illustrated in Fig. 4c, which shows normalized ΔC trajectories within a fixed 10 s interval beginning at the 8th second of each recording. The posture classes exhibited distinct temporal patterns despite partial overlap. High Lunge showed a marked late negative deflection, Lunge Forward exhibited repeated positive excursions, Squat displayed broader mid-window fluctuations, and Tree showed comparatively smoother oscillatory behavior. These differences suggest that posture information is encoded not only in response magnitude but also in the temporal evolution of the capacitive signal.

Figure 4d visualizes participant-level feature vectors after supervised linear discriminant analysis. In contrast to an unsupervised manifold view, this projection emphasizes posture-specific separation by using class information during dimensionality reduction. The four movements formed visually distinct groups, with High Lunge and Squat occupying the upper half of the display and Lunge Forward and Tree shifted downward, while each class retained a compact local spread around its centroid. This representation indicates that the extracted participant-level features contain sufficient structure to support posture discrimination in a reduced feature space.

Taken together, Fig. 4a–d shows that textile-based capacitive sensing captures complementary posture-related information across magnitude, normalization, temporal dynamics, and multivariate feature structure. These results support the use of interdigitated textile capacitive sensors for wearable posture monitoring and provide an interpretable basis for downstream classification and movement analysis.

Discussion

Recent literature demonstrates the use of interdigitated capacitive designs implemented through various fabrication approaches. Fujimoto et al. fabricated IDCs using aerosol jet printing (AJP), which precisely deposits silver nanoparticle ink on a flexible polymer substrate³⁶. A different approach enables highly miniaturized sensors using lithographically defined microchannels that are subsequently filled with liquid metal (LM). However, liquid metal systems may raise practical concerns, including potential leakage and delamination due to insufficient adhesion between layers and achieves durability at 30% strain with stretchability up to 100%⁷⁷. Other studies have employed polydimethylsiloxane (PDMS) substrates patterned with silver nanowires (AgNWs) using capillary force lithography (CFL), demonstrating gauge factors of approximately GF = −1.57 at 30% strain and durability up to 1000 cycles³⁵. For example, Kim et al. patterned AgNWs using CFL to produce stretchable thin films capable of sensing strains up to 50%³⁵. In a separate study, conductive fabrics were embedded within a PDMS substrate using a layer-by-layer method; however, detailed characterization of strain sensitivity was not reported⁵⁹. Other work has explored conductive elastomers based on barium titanate-Ecoflex composites, achieving sensitivities approaching 100% strain²⁸. In these approaches, textile structures were embedded within elastomer matrices to improve reproducibility, achieving gauge factors of ~0.8 at 50% strain³³.

Table 1 summarizes these representative studies, including their materials, strain sensitivity, and mechanical durability. Despite promising performance, a key limitation of many previously reported systems is the limited attention given to breathability, biocompatibility, and robustness under real garment conditions. Materials such as Ag nanowires³⁵, and carbon black composites²⁸ are often embedded within polymer matrices (PDMS⁷⁷, silicone^28,33). In contrast, the present work employs conductive textile electrodes integrated directly within a fabric architecture, enabling improved flexibility, comfort, and garment compatibility.

Table 1.

Comparison of interdigitated capacitive flexible strain sensors reported in the literature for motion monitoring

Reference	Substrate	Electrode material	Absolute GF	Strain (%)	Strain Cycles
77	PDMS	Liquid metal	0.30–0.69	30–100	7000
35	PDMS	AgNWs	2.0	30	1000
59	PDMS	Conductive woven fabric	0.87	NR	NR
28	BT/Ecoflex	CB/Ecoflex	0.5	100	1000
33	Ecoflex	Silver-coated conductive textile	0.8	50	500
This work	Lycra fabric	Conductive silver fabric	0.67–0.81	10–30	6500

PDMS polydimethylsiloxane, AgNWs silver nanowires, BT barium titanate, CB carbon black.

Previous studies have also suggested that the elastic operating range of textile capacitive sensors can be extended by incorporating dielectric elastomer layers^33,71. For example, combining textiles with silicone dielectric layers can improve mechanical resilience while maintaining capacitive sensing capability. Such hybrid approaches can support extreme deformation ranges (50–100% strain), which may be advantageous for dynamic body movements. Nevertheless, even within the moderate strain range explored in this work (10–30%), the proposed IDC textile sensor demonstrates reliable electrical responses and strong mechanical durability, supporting its suitability for wearable motion monitoring applications.

The classification results demonstrate that a machine learning model trained on textile-derived electrical sensing features can reliably distinguish between biomechanically distinct yoga postures. The record-level confusion matrix (Fig. 3a) shows strong classification performance across the four movements, with an overall accuracy of 94.4% and a macro-averaged F1 score of 94.2%. Most postures were classified correctly, with only minor confusion observed between High Lunge and Lunge Forward, which share similar lower-body configurations.

Receiver operating characteristic (ROC) analysis further confirms the discriminative capability of the trained classifier (Fig. 3d). The area under the curve (AUC) values reached 90% for High Lunge, 94% for Lunge Forward, and 99% for both Squat and Tree, indicating strong separability of the learned feature representations across posture classes.

The channel ablation analysis (Fig. 3b) provides insight into the relative contribution of different sensing channels. Models using the complete set of sensor signals achieved the highest performance (91.7% record-level accuracy). However, capacitance-derived channels alone–particularly Co, CP, and the normalized capacitance change (DC/Co)–retained comparable predictive capability. In contrast, models relying solely on resistance (RS) or phase (PHASE) signals showed lower classification accuracy. These results suggest that capacitance-based measurements capture the dominant biomechanical deformation patterns associated with posture changes in textile-integrated sensors.

These findings should nevertheless be interpreted within the scope of a controlled proof-of-concept study. The cohort comprised 9 participants after one withdrawal, which is sufficient to demonstrate feasibility, signal discriminability, and repeatable sensing behavior, but not to establish broad population-level generalizability. Because the present evaluation was conducted at the record level, future work should extend validation to larger and more diverse cohorts using subject-wise evaluation protocols to determine how well the proposed sensing and classification framework generalizes across unseen individuals.

These classification outcomes complement the signal-level analyses presented in Fig. 4, which reveal the underlying distributions and temporal dynamics responsible for posture separability. The distributions of log-transformed capacitance changes, ${\log }_{1p}\left(\mid \mathrm{\Delta }C\mid \right)$ (Fig. 4a), demonstrate that posture-induced deformation of the textile sensor produces distinct electrical responses across movements. Even after logarithmic compression–which reduces dynamic range and mitigates the influence of extreme values–posture-specific differences remain visible. This confirms that the capacitive signal reliably captures biomechanical strain patterns generated during different yoga movements.

Normalization of the signal relative to baseline capacitance, expressed as ${\log }_{1p}\left(\mid \mathrm{\Delta }C/C_0\mid \right)$ (Fig. 4b), further highlights the robustness of the sensing modality. The preservation of distinct posture-dependent distributions after normalization indicates that the extracted features are not dominated by baseline sensor offsets or subject-specific capacitance levels. This robustness is particularly important for wearable sensing systems, where variations in garment placement, fabric tension, and individual physiology can introduce substantial baseline variability.

Temporal analysis of normalized capacitance signals within a fixed 10 s window (Fig. 4c) reveals additional posture-dependent dynamics. While all movements generate measurable signal fluctuations, their temporal patterns differ in both amplitude modulation and waveform smoothness. For example, Squat movements tend to produce smoother signal evolution, whereas Lunge Forward exhibits sharper transient variations. These posture-specific temporal signatures suggest that incorporating dynamic waveform features could further enhance classification performance, particularly during movement transitions.

The supervised LDA projection (Fig. 4d) provides a geometric interpretation of the extracted feature space. The clear separation of posture clusters demonstrates that the engineered features capture discriminative structure enabling reliable classification. The covariance ellipses illustrate compact within-class distributions and distinct class centroids, indicating strong inter-class separability. This low-dimensional representation also suggests that accurate posture recognition can be achieved with a relatively compact feature set, which is advantageous for real-time wearable sensing applications with limited computational resources.

Despite the strong performance observed in controlled experimental conditions, several challenges remain for real-world deployment. Variability arising from sensor displacement, garment looseness, perspiration, and individual movement styles could influence capacitance readings and potentially reduce classification accuracy. Addressing these challenges may require adaptive calibration strategies, larger multi-subject datasets, or multimodal sensing approaches that combine textile capacitive sensing with inertial or optical measurements.

The ablation results also highlight opportunities for hardware optimization. Because capacitance-related channels provide most of the predictive capability, future wearable systems may reduce sensing complexity and power consumption by focusing on a smaller subset of highly informative capacitance features. Such simplification could enable lightweight, energy-efficient smart garments capable of real-time posture recognition.

Overall, the results establish a clear pathway for textile-based capacitive sensors in movement recognition systems. By integrating signal amplitude characteristics, temporal dynamics, and discriminative feature embeddings, the proposed framework enables reliable posture monitoring using fabric-integrated sensors. Future work should extend this analysis to larger populations, dynamic movement sequences, and unconstrained environments to fully evaluate system generalizability and practical deployment potential.

Future work should evaluate the sensing framework in a larger and more diverse cohort and across a broader range of physical activities. Such validation will be important for assessing generalizability beyond the controlled postures studied here and for extending the approach to multi-joint and dynamic movements in unconstrained environments.

Wireless integration and energy-efficient edge computing are promising directions for enabling real-time, continuous data acquisition in remote health-monitoring scenarios^78–81. Moreover, integrating complementary sensing modalities, such as inertial and optical sensors, may improve robustness and support adaptive sensing in mobile, cluttered, or noisy settings. Collectively, these developments could advance the platform toward practical deployment in personalized healthcare, rehabilitation, and human–machine interaction.

Methods

Computational modeling and simulation

For the analytical study, we developed scripts using MATLAB R2018. The electric field distribution of the sensing modalities was simulated using COMSOL Multiphysics 6.1 with the Electrostatics interface. The simulation model reproduced the physical geometry of the fabricated sensors to ensure realistic boundary conditions and electrode configurations.

Materials

ElectroLycra 234-309A (76% nylon, 24% elastic fiber) was purchased from MindSets (Saffron Walden, UK) and used as the conductive textile material. This fabric is stretchable in both directions and has a surface resistivity of less than 1 Ω/sq when unstretched. Spandex textile with a thickness of 0.50 mm and one-directional stretchability was purchased from Spitzentraum (Schänis, Switzerland) and used as the dielectric layer.

Complex dielectric permittivity measurements were performed using a Rohde & Schwarz ZXX analyzer equipped with a DAKTL2 3.5 probe (SN 1050). Conductivity measurements were performed using a Fluke 177 digital multimeter (Fluke Corporation, Everett, WA, USA). Laser cutting was carried out using a Rayjet 50 system (Trotec, Marchtrenk, Austria). Capacitance measurements were obtained using a Hioki IM7581-1 LCR analyzer (Hioki E.E. Corporation, Ueda City, Japan). Mechanical tensile testing was performed using an Instron E-Puls E3000 Linear Testing System (Instron, Norwood, MA, USA).

Capacitance measurements were recorded on a PC using a custom LabVIEW acquisition script while maintaining a sensing frequency bandwidth of ~100 Hz. Microscopy images were captured using a Schottky SU5000 field-emission scanning electron microscope (Hitachi High-Tech Corporation, Tokyo, Japan). Temperature testing was performed using a TH-50A environmental chamber (LIB Environmental Simulations Industry, China). Sewing was performed using an Innovis NV-2600 sewing machine (Brother Industries, Nagoya, Japan). Sportswear leggings used for integration of the sensors were purchased from Decathlon (Zurich, Switzerland).

Sensor fabrication

Laser cutting was used to fabricate the textile sensors because it allows accurate and reproducible patterning of the conductive layers. Lycra textile was used as the dielectric material because of its stable dielectric properties across a broad frequency range and its favorable mechanical stretchability. All fabrics were cut parallel to the textile wale direction (Supplementary Fig. S11). The IDC sensor consisted of alternating positive and negative electrodes. To ensure device consistency, the spacing between adjacent electrodes was fixed at 2.50 mm and the electrode width was set to 7.50 mm. This configuration minimized electrode gaps while maximizing the interaction length of the electric field⁶³. The final geometry resulted in an electrode length ~20 times larger than the electrode gap and an electrode width approximately three times the gap.

Mechanical testing

Before each experiment, the linear testing system was calibrated. The maximum applied strain was defined as 30% of the initial sample length. The IDC sensor was cyclically stretched at 10%, 20%, and 30% strain levels while capacitance measurements were recorded at a sampling frequency of 1 kHz using the LCR analyzer.

Custom 3D-printed fixtures were used to secure the textile during mechanical testing and to ensure uniform tension along the stretching direction. The textile samples were inserted into stitched pockets and mounted within the fixtures to maintain consistent alignment. Strain was computed from the raw displacement data using the complete sensor geometry. If w₀ and d₀ represent the electrode width and electrode gap, respectively, the initial sensor length was defined as

$$L_0=6w_0+5d_0.$$ 6

Participants and ethical approval

A total of 10 subjects were recruited for the study. One subject withdrew during the study and was therefore excluded from the final analysis. Consequently, data from 9 participants were included in the reported results.

Each participant performed four yoga movements: High Lunge, Lunge Forward, Squat, and Tree. For each movement, participants completed ~15–20 repetitions while wearing the textile sensor garment.

Ethical approval was obtained from the ETH Zurich Ethics Commission (approval number 25 ETHICS-304). All experimental procedures were conducted in accordance with the relevant guidelines and regulations, and all protocols were approved by the ETH Zurich Ethics Commission. Written informed consent was obtained from all participants. After anonymization, the collected dataset will be made publicly available in an open data repository.

Data acquisition and signal preprocessing

Sensor data were recorded from IDC textile sensors embedded within the garment during each posture. Each participant’s data were stored in a structured recording file containing a separate sheet for each movement.

The primary signal analyzed in this study was the capacitance channel denoted as C(t), corresponding to the measured capacitance signal (Co channel) as a function of time t.

Column names were standardized across recordings to account for variations in naming conventions. Non-numeric entries were converted to missing values and interpolated using linear interpolation. Remaining missing values were filled using forward and backward filling. Recordings containing insufficient valid samples were excluded.

Time vectors were shifted such that each recording started at t = 0. The baseline capacitance was estimated from the first second of each recording using the median value:

$$C_0=\mathrm{median}\left(C\left(t\right):t\le 1\mathrm{s}\right).$$ 7

The baseline-corrected capacitance signal was then defined as

$$\mathrm{\Delta }C\left(t\right)=C\left(t\right)-C_0.$$ 8

To stabilize the dynamic range and emphasize posture-related magnitude variations, logarithmic compression of the absolute signal was applied:

$$z\left(t\right)=\log \left(1+\mid \mathrm{\Delta }C\left(t\right)\mid \right).$$ 9

Record-level feature extraction

Each recording was summarized using descriptive statistics derived from the baseline-corrected signal ΔC(t). These statistics included the mean value, standard deviation, maximum absolute amplitude, 90th percentile of the absolute signal magnitude, and the area under the absolute signal curve.

Dominant peak characteristics were also extracted from the signal, including peak amplitude, peak timing, and rise time relative to the estimated onset of the signal response.

For distributional comparisons, each recording was summarized using the 90th percentile of the log-compressed signal:

$$S=Q_{0.90}\left(z\left(t\right)\right),$$ 10

where Q_0.90( ⋅ ) denotes the 90th percentile operator.

Temporal trace analysis

To analyze posture-dependent temporal patterns, a fixed analysis window of 10 s was extracted from each recording beginning at the eighth second:

$$t\in \left[8,18\right]\mathrm{s}.$$ 11

Signals within this window were resampled onto a uniform grid of 300 time points using linear interpolation. Each trace was then normalized using a robust scaling procedure based on the median and interquartile range of the baseline-corrected signal:‘

$$\tilde{C}\left(t\right)=\frac{\mathrm{\Delta }C\left(t\right)-\mathrm{median}\left(\mathrm{\Delta }C\left(t\right)\right)}{\mathrm{IQR}\left(\mathrm{\Delta }C\left(t\right)\right)}.$$ 12

Mean temporal profiles and the standard error of the mean (SEM) were computed across participants for each posture.

Statistical analysis

Differences among posture groups were first evaluated using the Kruskal–Wallis test applied to the record-level summary statistic S. Pairwise posture comparisons were then performed using two-sided Mann–Whitney U tests. Holm–Bonferroni correction was applied to control the family-wise error rate for multiple comparisons.

Low-dimensional visualization of posture separability

Participant-level recordings were represented using feature vectors denoted by ${\mathbf{x}}_i\in {\mathbb{R}}^7$, where each vector contains seven sensor-derived features: Co, CP, PHASE, RS, DC, DC/Co, and DC/Co (%). The index i denotes the individual recording.

Feature vectors were standardized to zero mean and unit variance prior to dimensionality reduction. Linear Discriminant Analysis (LDA) was then applied to project the feature vectors into a two-dimensional embedding space that maximizes class separability between the four posture classes:

$${\mathbf{u}}_i=\mathrm{LDA}\left({\mathbf{x}}_i\right),$$ 13

where ${\mathbf{u}}_i\in {\mathbb{R}}^2$ represents the low-dimensional representation of recording i.

The two-dimensional LDA embedding is shown directly without additional transformation. Covariance ellipses and class centroids are displayed to aid visualization of posture separation.

Supervised classification and model evaluation

A supervised machine-learning pipeline was implemented in Python using scikit-learn and sktime to classify four yoga postures: High Lunge, Lunge Forward, Squat, and Tree. The input time-series representation for each sample consisted of 21 channels: seven raw electrical sensing channels (Co, CP, PHASE, RS, DC, DC/Co, and DC/Co (%)), together with their first and second temporal derivatives.

Two complementary feature families were used. The first consisted of handcrafted features extracted from the 21-channel time-series tensor. For each channel, time-domain descriptors, spectral descriptors, and DCT-like coefficients were computed. Specifically, the handcrafted features included mean, standard deviation, minimum, maximum, range, median, interquartile range, root-mean-square value, energy, skewness, kurtosis, peak index, trough index, first-derivative mean, first-derivative standard deviation, and second-derivative energy, together with spectral descriptors (dominant frequency, spectral entropy, spectral centroid, low-, mid-, and high-band power, and first- and second-harmonic power) and 10 DCT-like coefficients. Correlation features among the original seven raw channels and template landmark values sampled from each channel were also included. The second feature family consisted of MiniROCKET-derived multivariate time-series features generated using 10,000 convolutional kernels.

The complete feature representation therefore combined handcrafted features with MiniROCKET-derived time-series features. The classification pipeline consisted of four sequential stages: (i) missing-value imputation using the median, (ii) z-score feature standardization, (iii) feature selection using SelectKBest with mutual-information scoring, and (iv) supervised classification using a hard-voting ensemble classifier.

The final ensemble classifier combined three base learners: logistic regression (max_iter=5000, class_weight=balanced), random forest (n_estimators=500, class_weight=balanced_subsample, random_state=42), and extra trees (n_estimators=700, class_weight=balanced, random_state=42). Voting was performed using hard majority voting across the three classifiers. The number of selected features was set to the top 1400 ranked features, or all features if fewer than 1400 were available.

Model predictions were first obtained at the sample level and then aggregated at the recording level using majority voting:

$${\hat{y}}_r=\mathrm{mode}\{{\hat{y}}_t:t\in r\},$$ 14

where ${\hat{y}}_t$ denotes the predicted class label for sample t and ${\hat{y}}_r$ denotes the final predicted class assigned to recording r.

Classification performance was evaluated using a record-level confusion matrix, overall accuracy, and the macro-averaged F1 score. For the random-split evaluation, data were divided using a stratified 70/30 train–test split with random seed 42. To assess robustness, stratified 5-fold cross-validation was additionally performed using StratifiedKFold with shuffling enabled and random seed 42. In each fold, sample-level predictions were aggregated within each recording using majority voting to produce record-level predictions, and record-level accuracy was summarized across folds.

Multi-class receiver operating characteristic (ROC) curves were computed using a one-versus-rest strategy. Class probabilities were obtained from the logistic-regression branch of the fitted ensemble classifier using the predict_proba function. Probabilities were averaged across all samples belonging to the same recording before ROC computation in order to obtain record-level probability estimates. The area under the ROC curve (AUC) was calculated separately for each posture class.

Channel ablation analysis

To investigate the relative contribution of different sensing modalities, channel-ablation experiments were performed using predefined channel groups: Co only, DC/Co only, DC/Co (%) only, capacitance family (Co + DC/Co + DC/Co (%)), PHASE only, RS only, PHASE + RS, CP only, and all channels. For each selected raw channel, the corresponding first-derivative and second-derivative signals were also included, yielding a subset-specific multichannel time-series representation.

For each ablation subset, handcrafted features were re-extracted and the classification pipeline was retrained using a stratified 70/30 train–test split with random seed 42. Performance was evaluated at both the cycle level and the record level using accuracy and macro-averaged F1 score, with record-level predictions obtained by majority voting across samples belonging to the same recording.

Feature importance was additionally assessed using permutation importance applied to the fitted ensemble classifier, with classification accuracy as the scoring metric, 10 permutation repeats, and a fixed random seed of 42. Importance scores were aggregated according to sensor family and normalized by the number of selected handcrafted features belonging to each channel group. MiniROCKET features were excluded from this sensor-family aggregation so that the resulting importance scores reflected the relative contribution of the interpretable sensor-derived channel families. Record-level classification accuracy obtained from the ablation experiments was then used to rank channel combinations and identify the most informative sensing modalities.

Author contributions

M.P. conceived the ideas and designed the experiments under the guidance of M.E. and C.M.M. M.P. performed the experiments and analyzed them. M.E. performed the statistical analysis and developed the machine learning models. M.P. and M.E. wrote the manuscript. All authors discussed the results and commented on the manuscript.

Data availability

All data are available at the following links: https://github.com/Elgendi/Textile_Capacitive_Sensors_Yoga_Movements/blob/main/npjBiosensing_IDC_2026.ipynb and https://figshare.com/s/a5c0ac9e3a6a371100d3.

Competing interests

All other authors declare no competing interests. The author M.E. serves as an associate editor of npj biosensing and had no role in the peer review or decision to publish this manuscript.

Supplementary information

The online version contains supplementary material available at 10.1038/s44328-026-00101-1.