Classification of Embroidered Conductive Stitches Using a Structural Neural Network

Abstract

This study presents a machine learning-based framework for classifying five embroidered stitch patterns—straight, zigzag, joining, satin, and wave—under 10% tensile strain, aiming to enhance their utility in smart textile circuits. Electrical conductivity was derived from resistance data and standardized using Z-score normalization. Conductivity sequences were first analyzed with PCA and Random Forest classifiers, then classified using a structural artificial neural network model. The model employed a structurally informed filter design, reflecting stitch-wise signal periodicity to capture time-varying electrical patterns under cyclic strain. It achieved a test accuracy of 97.33%, with F1-scores above 0.83 for all classes and perfect scores in three. Partial confusion between wave and zigzag patterns was observed due to their similar curved geometry and signal profiles. These results validate the discriminative power of conductivity-based features and demonstrate the potential of structure-aware neural networks for identifying dynamic stitched circuits in smart textiles.

1. Introduction

In recent years, the field of smart wearables has grown rapidly, with smart textiles emerging as a particularly attractive branch due to their comfort, flexibility, and seamless integration into everyday clothing [1,2,3,4]. Textiles provide a unique platform for developing devices that demand deformability and user compatibility, making them suitable for both laboratory research and consumer applications [5,6]. Progress in this area has led to numerous innovations such as fiber-based sensing elements, textile transistors, and actuating fibers [7,8,9]. Depending on the target function, multiple strategies can be employed to embed electronics within fabrics, enabling cross-disciplinary collaboration between design, textile engineering, and electronics [10].

Conductive yarns, in particular, are valuable for creating electrical pathways within textile systems because of their compatibility with diverse fabric structures [4]. They have been successfully utilized in applications ranging from embroidered antennas [11,12] and wireless energy transfer modules [13] to RFID-based solutions [14]. Among the different textile processing techniques, embroidery is especially attractive since it permits pattern customization and reliable attachment of conductive paths. Through embroidery, wires, sensors, and other circuit elements can be integrated directly into the fabric, offering both electrical connectivity and design freedom [15,16]. Compared to weaving or printing, embroidery requires less complex processing and allows greater flexibility in circuit layout, which makes it well-suited for wearable electronics.

Despite this potential, earlier embroidery-related research has mainly concentrated on decorative, structural, or reinforcement purposes [17,18,19,20]. Systematic analysis of how stitch geometries behave electrically under deformation is still rare, and most previous studies emphasized mechanical properties such as tensile reinforcement or impact resistance rather than conductivity [21,22,23]. Understanding how stitches influence electrical behavior is critical for the design of reliable textile circuits, yet machine learning-based approaches for classifying such behaviors remain largely underexplored. For example, embroidered smart garments with alarm systems have been reported, but these works focused on demonstrating functionality rather than providing structured electrical evaluation.

In particular, the use of conductive threads has proven highly effective for establishing electrical interconnections in textile-based systems. These threads are compatible with various fabric structures, enhancing their versatility and applicability. Applications of conductive threads include embroidered antennas, wireless power transfer systems, and RFID technologies. Embroidery is one of the most common and practical methods for incorporating conductive threads into fabrics, as it allows for customizable patterns and reliable integration. Embroidered stitches enable straightforward circuit construction by embedding electronic components—such as wires, switches, sensors, or other devices—into textile substrates while imparting electrical conductivity using conductive threads Compared to weaving, knitting, or printing techniques, embroidery offers greater flexibility in circuit design and ease of fabrication, making it a promising approach for smart textile circuits [24]. Other studies analyzed resistive behavior of embroidered elements [25] or introduced multi-layer stitching for antenna efficiency [26], but data-driven classification of stitch-level conductivity has not been fully addressed.

Deep learning provides an opportunity to capture nonlinear patterns in time-varying signals that traditional methods cannot easily detect. Convolutional neural networks (CNNs) have already shown success in various textile and sensor-related tasks, including recognition of fabric deformation, bio-signal analysis, and flexible electronic signal processing [27,28,29,30,31]. However, their application to the classification of embroidered stitch conductivity remains limited.

To address this gap, the present study investigates the electrical responses of five embroidered stitch geometries—straight, zigzag, joining, satin, and wave—under tensile deformation and proposes a structural convolutional neural network (S-CNN) to classify their signal characteristics. The proposed framework combines conventional methods such as PCA and Random Forest with a structural neural architecture designed to reflect the periodicity and geometry of stitched signals. Figure 1 outlines the overall research workflow.

By establishing a systematic, data-driven approach to stitch classification, this work contributes to the design of next-generation textile electronics. The findings not only highlight the feasibility of tailoring stitch types to specific sensor or interconnect functions but also provide a basis for integrating machine learning with embroidery techniques to realize advanced wearable systems.

2. Materials and Methods

2.1. Sensor Fabrication

For sensor fabrication, a stretchable fabric named SP20D (POWERFIT, Seoul, Republic of Korea), composed of 92% polyester and 8% spandex, was used as the base textile. It is composed of 92% polyester and 8% spandex. The polyester is a blend of PET75/36 DTY (ASKIN, HYOSUNG Corporation, Seoul, Republic of Korea) and PET80/48 ITY at a 1F:1F ratio. This fabric has a basis weight of 153 g/m². Unlike plain-woven fabrics, this textile is knit-based with inherent elasticity provided by spandex, making it particularly suitable for wearable applications. In this context, knit structural descriptors such as courses per inch (CPI), wales per inch (WPI), or loop density would be more appropriate than the conventional cover factor (K) used for woven fabrics. All fabric samples were cut along the wale direction to ensure consistency of specimen preparation. Consequently, when tensile deformation was applied perpendicularly to the wale direction, the stitches experienced strain predominantly along the course direction. Although anisotropy between wale and course directions in knit structures could potentially influence mechanical deformation and electrical response, this study controlled such effects by maintaining a fixed stitch orientation and test direction across all specimens. The woven morphology is presented in the SEM image shown in Figure 2. Each sensor was implemented through a single line of embroidered thread. The study incorporated five distinct stitch patterns: straight, zigzag, joining, satin, and wave. To circuit integration, we employed five stitching patterns: straight stitches as linear types and zigzag, joining, satin and wave stitches as curved types. All stitch patterns in this study exhibit an identical cross-sectional configuration, as they were fabricated using the same sewing machine settings and mechanism. The variation among stitch patterns is limited to the surface-level patterns determined by the machine’s programmed designs, with no impact on the depth or penetration profile of the stitches. Figure 3 presents the cross-sectional view of the stitches.

For stitching, we employed a silver-coated conductive yarn (AMANN Silver-Tech, Houston, TX, USA) comprising 34 filaments. Its core consists of a polyester and polyamide blend, and it has a specified linear resistance of approximately 530 Ω/m and a linear density of 28 tex. Sensor stitching was performed at the center of the textile using a Brother Innov-is 95 sewing machine (Cranleigh, UK), which can execute multiple stitch patterns. Each stitch pattern was applied with a consistent total length of 40 mm, though variations in width and geometry arose depending on the specific stitch pattern. The length of a single stitch was set to 2 mm. For linear stitches, the stitch width was approximately 1 mm, while curved stitches exhibited a width of around 1~5 mm. The fabrication process is illustrated in Figure 2.

2.2. Data Acquisition Protocol

Resistance measurements were performed using a universal testing machine (UTM). Each stitched sample had a total length of 40 mm, and a gauge length of 4 mm was used to achieve 10% tensile strain. The testing was carried out at a speed of 0.67 mm/s, with each specimen subjected to 100 repeated loading cycles. Five stitch patterns were evaluated, and five specimens were prepared for each type. The results from the 100 cycles per specimen were averaged to derive representative values for analysis.

To analyze electrical conductivity (σ), resistivity ($\rho$) was first calculated using the standard formula Equation (1), and conductivity was then derived using Equation (4). In the resistivity equation, R denotes resistance, A is the cross-sectional area, and l is the length. For each stitch pattern, variations in cross-sectional area (A) and length (l) under tensile strain were observed, and continuous resistivity values were computed accordingly. These time-dependent resistivity values were then converted to continuous conductivity (σ) using Equation (4) and applied to deep learning-based pattern classification. The deformation of each stitched sensor during tensile testing is shown in Figure 4.

$$\rho ={\displaystyle \frac{RA}{l}}$$

(1)

$$∆A=A_0-A_f$$

(2)

$$∆l=l_f-l_0$$

(3)

$$\sigma ={\displaystyle \frac{1}{\rho }}$$

(4)

2.3. Structural Convolutional Neural Network

The structural convolutional neural network (S-CNN) is a multi-layered model that can learn discriminative features directly from sequential input signals [32]. Rather than relying on manual feature design, it automatically identifies informative temporal characteristics. This type of architecture has been applied in diverse fields, including vision-based recognition, biomedical signal analysis, and motion detection, due to its effectiveness in classification tasks with moderate complexity. Because the network is relatively shallow, training and deployment are less demanding compared with deeper models, which reduces the risk of overfitting when dealing with consistent and well-controlled data such as that collected from tensile testing [33]. In addition, its limited computational requirements and simple implementation make it attractive for a wide range of practical scenarios, especially where efficient recognition of key patterns is needed [34,35,36]. For these reasons, the S-CNN framework was selected in the present study, and its structure is depicted in Figure 5.

To classify five different stitch patterns based on their electrical conductivity behavior, continuous data was collected from 100 repeated tensile tests. To minimize errors during training, the conductivity data were standardized using Z-score normalization Equation (5), where $\mu$ and $\sigma$ denote the mean and standard deviation of each feature, respectively.

Importantly, the filter size of the model was empirically set to 4, reflecting the average signal periodicity observed in the stitch structures, especially in patterns with shorter loop intervals or denser stitch distributions. This selection was inspired by stitch features such as stitch count per inch, peak spacing in the conductivity curve, and waveform symmetry. This model was designed to facilitate more effective feature extraction from the dataset and is referred to as the model.

The S-CNN was designed with seven convolutional blocks, each including convolution, normalization, ReLU, and dropout layers. After these, a global average pooling layer reduces dimensionality, followed by a fully connected layer and a classification output using SoftMax activation [37]. The SoftMax function normalizes the outputs of K logits into a probability distribution where all values are between 0 and 1 and the total equals 1, as expressed in Equation (6) [38].

In addition to deep learning, classification performance was assessed using principal component analysis (PCA), and confusion matrices derived from a random forest classifier. S-CNN’s classification accuracy, loss values, and confusion matrix were also evaluated. All computations were carried out in MATLAB 2023a (MathWorks Inc., Natick, MA, USA). For each stitch pattern, 5800 data samples were prepared, and the dataset was divided into training, validation, and test sets at a ratio of 7:1.5:1.5. The learning rate was set to 0.001, the mini-batch size to 4, and the number of epochs to 100.

$$Z={\displaystyle \frac{x-\mu }{\sigma }}$$

(5)

$$\sigma {\left(\mathit{z}\right)}_i={\displaystyle \frac{e^{z_i}}{{\sum }_{j=1}^K{e}^{z_j}}} for i=1, \dots , K and \mathit{z}=\left(z_1, z_2, \dots , z_K\right)ϵR^K$$

(6)

3. Results and Discussion

3.1. Characterization

Five different stitch patterns were designed, and five samples were produced for each type [39]. All stitched sensors were produced at identical lengths. To ensure precise measurement of length, conductive tape (Gyeonggi-do, Republic of Korea) was attached to both ends of each stitch, enabling more accurate length control for all sensors. The tape was cut into strips with dimensions of 3 mm in width and 10 mm in length. The stitch patterns, arranged from top to bottom, include straight, zigzag, joining, satin, and wave patterns. Figure 6 presents both the schematic illustration and photographs of the stitched sensors.

To further examine the anisotropic characteristics of the knit substrate, additional tensile tests were conducted in both the wale and course directions. As shown in Figure 7, the wale direction exhibited slightly higher load and stress responses compared to the course direction. At 50% strain, the maximum stress reached approximately 0.015 kPa in the wale direction, while the course direction showed about 0.013 kPa. Similarly, the load–displacement curve indicated that the wale direction sustained a higher load (≈0.065 kgf) than the course direction (≈0.050 kgf) at a displacement of 90 mm. These results confirm that the knit fabric possesses inherent anisotropy depending on orientation. However, the difference between the two directions was relatively small, suggesting that stitch performance is not critically affected by the choice of whale or course alignment. In this study, sensors were designed such that tensile deformation was applied along the course direction. Given the slightly greater extensibility and higher sensitivity observed in the course orientation, this configuration is considered more suitable for embroidered strain sensors, as it enables measurable electrical responses under practical loading conditions without requiring excessive mechanical force.

3.1.1. Sensitivity

The experiments were conducted at room temperature (20–25 °C), and changes in electrical resistance were measured within a tensile strain range of 10%. The data represents the average resistance change ratio obtained by stretching and releasing five samples for each stitch pattern over five repeated cycles (Figure 8).

The joining stitch exhibited the lowest resistance change ratio, remaining nearly constant around 0.005 throughout the strain range. The zigzag stitch showed a slightly higher but relatively stable response, with values peaking at around 0.013. The straight and satin stitches demonstrated moderate sensitivity, with resistance change ratios reaching approximately 0.02. Among all, the wave stitch showed the highest increase, with $\Delta R/R_0$ values exceeding 0.04 at 10% strain, indicating the greatest responsiveness to deformation.

These results indicate that the order of patterns with the highest resistance change ratio is wave, satin, zigzag, straight, and joining. Among them, all curved stitch patterns except the joining stitch (i.e., wave, satin, and zigzag) showed relatively high resistance changes, suggesting that stitch geometry affects electrical response under strain.

This trend can be explained by the changes in cross-sectional area ($\Delta A$), based on resistance equation 1. Since the tensile displacement ($\Delta l$) was fixed at 4 mm for all patterns, the variations in $\Delta A$ likely contributed to the differences in resistance response. The measured $\Delta A$ values were approximately 1.5 for the wave stitch, 1.0 for the zigzag stitch, and 0.5 for the joining stitch. These values correlate well with the resistance change ratios of those patterns. However, the satin stitch, despite not having a quantified $\Delta A$ in this study, exhibited a higher resistance change than the joining stitch. This suggests that, in addition to $\Delta A$, other factors—such as local deformation, stitch loop density, or mechanical compliance—may also influence resistance behavior. Thus, while $\Delta A$ is a major determinant, the resistance response under tensile strain should be interpreted as a multi-factorial outcome.

Furthermore, this change in resistance arises as the tension in the conductive thread increases, causing the spacing between the threads to narrow [40]. As the angle of the conductive thread increases [41], the resistance also increases, resulting in higher resistance values for curved stitch patterns compared to linear ones. As shown in Figure 9, part (a) illustrates the initial thread width $\left(w_0\right)$, while part (b) shows the reduced thread width $\left(w_f\right)$ under tensile deformation. The difference, $\Delta w$ $(w_0 - w_f)$, becomes larger when the stitch includes a curved geometry. Consequently, greater resistance changes are observed in stitch patterns with larger curvature.

To statistically validate these differences, a one-way ANOVA was performed on the resistance values across stitch types. The results revealed a highly significant effect of stitch type (F(4, 29,313) = 6.14 × 10⁶, p < 0.001), confirming that the resistance variations among stitch geometries are statistically meaningful.

3.1.2. Durability

For the development of stitch sensor circuits, a stability test was conducted under the same tensile deformation conditions following the measurement of resistance change ratios. The tensile test was performed using a universal testing machine (UTM), with the initial sensor length set to 40 mm and the clamping length set to 4 mm. A total of 100 repeated cycles were applied over a duration of 1200 s. The measurement results are presented in Figure 10. The measured resistance ranges for each stitch pattern were as follows: for the straight stitch, the resistance varied from approximately 12.97 Ω to 13.2 Ω; for the zigzag stitch, from 19.47 Ω to 19.75 Ω; for the joining stitch, from 18.4 Ω to 18.55 Ω; for the satin stitch, from 19.72 Ω to 20.2 Ω; and for the wave stitch, from 16.25 Ω to 16.9 Ω.

Overall, all stitched sensors exhibited relatively stable resistance changes throughout the testing period. No significant drift or permanent increase in resistance was observed over 100 cycles, indicating good durability of the stitched sensors under repeated tensile deformation. The small variation range confirms the potential of these sensors for stable operation in practical applications involving dynamic strain.

3.2. Analysis of Measured Data

To evaluate the electrical conductivity of the stitch-based sensor under tensile deformation, the resistivity measured during testing was converted into conductivity using a standard equation. In general, electrical conductivity $\left(\sigma \right)$ is inversely proportional to resistivity, and it also exhibits an inverse relationship with the relative resistance change. The resulting conductivity values were normalized using Z-score normalization to ensure consistent data processing.

Figure 11 illustrates the electrical conductivity values measured during a single cycle of the tensile test. The average conductivity, in descending order, was approximately 1.9 for the wave stitch, 1.8 for the zigzag stitch, 1.7 for the joining stitch, 1.6 for the satin stitch, and 1.4 for the straight stitch. This order does not perfectly reflect an inverse proportionality to the relative resistance change. This discrepancy may be attributed to the fact that electrical conductivity is often evaluated under static conditions, particularly in rigid materials with minimal deformation. In contrast, the measurements in this study were obtained under dynamic tensile motion using a flexible substrate.

Moreover, since the normalized electrical conductivity data exhibits highly similar patterns across different samples, it is difficult to analyze or classify their efficiency or distinctive characteristics based on this data alone. Therefore, initial classification was performed using simple machine learning algorithms such as PCA and Random Forest, and a structured artificial neural network was ultimately employed to enable more effective feature extraction.

To further investigate the distinguishability of the five stitch patterns based on their electrical conductivity, principal component analysis (PCA) was performed. As shown in Figure 12, the 2D projection using the first two principal components (PC1 and PC2) reveals a ring-shaped distribution, where each stitch pattern—joining (j), semi-nonelastic (sn), straight (st), wave (w), and zigzag (z)—forms partially clustered groups. The 3D visualization using PC1, PC2, and PC3 further enhances the class separation, although some overlap remains. These results suggest that the stitch patterns possess distinguishable conductive characteristics that can be captured in reduced-dimensional space, supporting their suitability for classification tasks.

To quantitatively validate the class separability, a Random Forest classifier was applied, and the resulting confusion matrix is presented in Figure 13. The model achieved perfect classification accuracy (100%) for the j, st, w, and z classes. For the sn class, two out of thirty samples were misclassified as w, resulting in an accuracy of 93.3%. Overall, the classifier exhibited excellent performance, with minimal misclassification, confirming that the conductivity-based features are sufficiently discriminative for identifying stitch patterns.

Similarly, a one-way ANOVA was conducted on the conductivity values across stitch types. The analysis revealed a highly significant effect of stitch type (F(4, 29,313) = 2.29 × 10⁶, p < 0.001), confirming that the conductivity differences observed in Figure 9 are statistically meaningful.

3.3. S-CNN-Based Classification of Stitch Patterns

The objective of the structural artificial neural network employed in this study was to learn and classify the distinguishing features of five different stitch patterns. The input data used for training consisted of the Z-score normalized electrical conductivity values described in Figure 8. Normalization facilitates data generalization, which can lead to improved learning efficiency, reduced error rates, and increased classification accuracy compared to raw data. In particular, the architecture was designed with a convolutional filter size, which was empirically selected to reflect the average peak spacing and local periodicity observed in the conductivity signals of curved stitch types such as zigzag and wave. This choice allowed the model to effectively capture repeating micro-patterns inherent to specific stitch geometries, thus reinforcing its ability to differentiate structurally similar patterns.

The classification accuracy and loss values obtained through training are presented in Figure 14. Figure 14a shows the training and validation accuracy over epochs, while Figure 14b illustrates the corresponding training and validation loss curves. The final classification results are summarized in the confusion matrix shown in Figure 15. The overall test accuracy achieved by the model was 97.33%. The validation accuracy was 96.10%, while the test dataset achieved an even higher accuracy, demonstrating that the model was not overfitting and exhibited stable generalization performance. Moreover, an error rate below 3% suggests that further improvements may be attainable through data augmentation or deeper network architectures. Nevertheless, an accuracy of approximately 97% already provides strong evidence of the model’s high performance, thereby confirming its reliability.

The class labels, numbered from 1 to 5 in the confusion matrix, correspond to the stitch patterns in the following order: straight (st), zigzag (z), joining (j), satin (sn), and wave (w). Among these, only the wave stitch (label 5) showed a relatively lower classification accuracy of 85%, with misclassifications primarily occurring with the zigzag stitch (label 2). This confusion is likely due to the similar curved structure of both stitch patterns, as well as the closely matching patterns observed in their electrical conductivity profiles. Despite this partial overlap, the high accuracy of 85% for the wave stitch and the overall classification accuracy of 97.33% indicate that the model achieved reliable performance in identifying stitch patterns.

To further assess the classification performance, the F1-score was also calculated and analyzed for each class. These results are summarized in

Table 1. Classification performance metrics of S-CNN.

Metric	Straight	Zigzag	Joining	Satin	Wave	Macro Average
Precision	1.0	0.8	1.0	1.0	0.88	0.94
Recall	1.0	0.87	1.0	1.0	0.82	0.94
F1-score	1.0	0.84	1.0	1.0	0.85	0.94
Accuracy						0.96

. The straight, joining, and satin stitch patterns achieved perfect F1-scores of 1.000, indicating that the model consistently made correct predictions for all instances in these classes and did not mistakenly assign samples from other classes to them. This reflects the fact that the conductivity patterns associated with these stitches are sufficiently distinct and stable under tensile deformation, enabling the S-CNN to learn and classify them with high confidence.

By contrast, the zigzag (F1 = 0.839) and wave (F1 = 0.848) stitch patterns exhibited slightly lower F1-scores, revealing a degree of confusion between these two structurally similar classes. Specifically, the zigzag class had a relatively lower precision of 0.812, suggesting that some of the samples predicted as zigzag originated from other classes, most notably the wave class. On the other hand, the wave class yielded a slightly higher precision of 0.875 but a lower recall of 0.824, indicating that while most predicted wave samples were correct, a non-negligible portion of actual wave samples were misclassified, again primarily as zigzag.

This mutual misclassification highlights a semantic similarity in signal patterns that challenges the discriminative capacity of the S-CNN when features overlap in both time-domain shape and normalized amplitude. Despite this, both F1-scores remain high, and the trade-off between precision and recall is well-balanced, suggesting that the S-CNN still captures essential pattern characteristics. The structural artificial neural network appears to improve model performance by employing filters of sizes comparable to the signal magnitude. These findings imply that while certain curved stitch patterns may exhibit similar conductive behaviors, the model retains a strong capacity to generalize and identify meaningful distinctions within the data. Furthermore, the high F1-scores obtained for each stitch pattern demonstrate the model’s ability to accurately distinguish between their unique characteristics, suggesting the potential for circuit design tailored to the functional roles of each pattern.

4. Conclusions

This study demonstrated the effective classification of five distinct stitch patterns—comprising both curved and straight geometries—under 10% tensile strain by analyzing their resistance change ratios and Z-score normalized electrical conductivity profiles. The S-CNN was ultimately employed as the final classification model, following preliminary analysis using principal component analysis (PCA) and a Random Forest classifier. The model achieved an overall accuracy of 97.33% and F1-scores exceeding 0.83 across all classes, including perfect scores for three stitch patterns. These results validate the discriminative power of conductivity-based features and highlight the reliability of deep learning in distinguishing textile-integrated sensor patterns.

In particular, the temporal signal evolution of each stitch under cyclic tensile loading was effectively captured through Z-score normalized time-series inputs, reflecting the dynamic electrical behavior of embroidered structures. In addition, the S-CNN architecture was intentionally designed as a structural artificial neural network, enabling pattern recognition that reflects both the geometric configuration of the stitches and the shape of the signal. This approach allows for a structural interpretation of the time-varying conductivity profiles.

Based on these findings, each stitch pattern may be tailored to serve specific functional roles in sensor-integrated circuitry, particularly where differentiated sensitivity or electrical response is required. Moreover, this analytical framework could be extended to real-time signal classification, adaptive textile systems, and multimodal sensing platforms. Nevertheless, the present findings were derived from a fixed tensile strain of 10%, which was deliberately selected to ensure consistent comparison across stitch types. Although this level of deformation was sufficient to capture stable and reproducible electrical responses, it may not fully represent the wider range of strain conditions encountered in practical wearable environments. Accordingly, future studies will aim to investigate broader stitch geometries and deformation ranges, assess performance under higher mechanical loading, and explore embedded processing strategies to enable on-device intelligence. In addition, the proposed framework has strong potential for real-world applications in wearable health monitoring, sports performance tracking, and smart garment–integrated circuit design, where reliable and distinguishable textile-based sensors are essential. Specifically, future work will also consider evaluating overlock stitches, which are widely used in knitted sportswear for their durability and elasticity, to further broaden the applicability of the proposed model to performance-oriented textiles.