1. INTRODUCTION
With the rapid development of intelligent robots, an increasing number of these machines are employed to replace humans in various engineering tasks. For instance, tasks such as accurately grasping objects, detecting product status, and transporting materials in hazardous environments significantly enhance production safety and product quality. Flexible tactile sensors, which are essential components of intelligent robots, endow these machines with human-like tactile perception capabilities [1,2]. These sensors simulate the human skin’s perception process by converting external stimuli into bioelectrical signals that are then transmitted to the brain [3]. Capable of detecting a range of surface stimuli, these skin-like sensors can assess both intensity and pattern.
To achieve tactile perception, flexible sensors that integrate various sensing principles have been successfully developed. Electrical-based sensors, which commonly utilize piezoresistive [4,5], piezoelectric [6,7], capacitive [8,9], and electromagnetic [10,11] structures, are employed to convert tactile information into electrical signals. Among these, resistance-based sensors are typically fabricated from elastic and conductive materials. However, due to material constraints, these sensors face challenges in achieving high-sensitivity detection across a wide range [4,5]. In contrast, piezoelectric sensors are primarily suited for the dynamic measurement of external forces rather than static sensing applications [6,7]. Conversely, capacitance-based sensors have demonstrated skin-like sensing capabilities in tactile detection. These sensors are characterized by low power consumption, minimal temperature sensitivity, high sensitivity, and reduced detection limits, thereby providing significant advantages in detecting weak forces [8,9]. Nevertheless, edge capacitance can considerably impact their operation, leading to issues such as crosstalk and increased hysteresis. Furthermore, electromagnetic-based sensors, developed utilizing the Hall effect, offer a wide dynamic response range; they typically necessitate a large physical volume and are susceptible to interference from external stray field noise [10,11]. While a highly integrated sensor array based on the aforementioned electrical sensors has been successfully developed, achieving high-density, skin-like detection of tactile force and position, several challenges persist. These include the complexity of array circuits, the high density of wiring, and electromagnetic interference.
With the rapid advancement of optical fiber sensing technology, attributable to its advantages such as compact size, resistance to electromagnetic interference, high flexibility, and ease of networking, significant achievements have been realized in domains including flexible wearable devices, intelligent robotics, and biomedicine [12–14]. Fiber optic sensing structures embedded in flexible materials present a novel methodology for the development of distributed tactile sensing. Due to its flexibility and robust networking capabilities, the fiber Bragg grating (FBG) is frequently employed as a sensing element in tactile sensors. For example, a soft tactile sensor utilizing FBG arrays embedded in polymers has been proposed [15–17]. These sensors can effectively detect forces and spatial positions on the skin when combined with machine learning algorithms. Experimental research has also demonstrated their successful application in various human-computer interaction interfaces and robotic arms for identifying keys and specific positions. In comparison to high-density electronic skin, the FBG-based array touch sensor can achieve tactile detection with a small amount of arrangement of FBGs. Additionally, due to its wavelength division multiplexing capabilities, it effectively addresses complex wiring challenges. Nevertheless, the reliance on a single wavelength demodulation scheme, which is solely dependent on changes in axial strain, limits the FBG array sensor’s force sensitivity and spatial resolution. Despite the low-density configuration of multiple FBGs, the associated costs remain substantial. The micro/nanofiber (MNF) exhibits a significant evanescent field and typically exhibits high sensitivity, making it a viable option for tactile sensor detection methods. A tactile sensor featuring 2×2 MNF arrays embedded in thin layers of polydimethylsiloxane (PDMS) has been proposed [18]. It estimates force by analyzing the output loss from the MNF. However, it is important to highlight that despite the simplicity of the detection method, the extremely small diameter of the fiber significantly compromises the mechanical strength and spectral stability of the sensor, which poses challenges for long-term monitoring. Additionally, a fiber optic soft sensor based on speckle pattern analysis is employed for the detection of distributed tactile forces [19,20]. This kind of sensor typically employs a polymeric material as an elastic pad, incorporating multimode fibers to create a tactile sensing system, and analyzes the resultant speckle field using an image processing algorithm. While some researchers have suggested utilizing a single fiber to achieve distributed tactile detection with a simplified sensing structure, the speckle pattern necessitates a complex spatial light path for acquisition, which restricts its applicability in various scenarios.
In this work, we present a distributed sensing solution that utilizes a single optical fiber, integrating an annular single mode-thin core mode-single mode (STS) optical fiber structure embedded in a PDMS flexible material for tactile sensing. This innovative design effectively converts stimuli applied to any area of the surface through the flexible pad into microbending or stretching of the fiber structure, leading to minor variations in the wavelength and intensity of the interference spectra. A random forest (RF) algorithm is employed to train a mapping model between mechanical stimulation and the interference spectrum, enabling the simultaneous identification of applied force and contact position across multiple areas. Furthermore, due to the annular sensing structure, the sensor can accurately detect tactile stimulation within a millinewton force and a square millimeter area, with the potential for expansion to larger skin areas. To validate the sensor’s characteristics, experimental studies are conducted to assess its temperature anti-interference capability, response time, and reliability performance. The results demonstrate that the sensor exhibits rapid response, strong reliability, and effective temperature interference resistance within a specified temperature range. Additionally, the sensor is adaptable for multifunctional interaction, as evidenced by the development of an intelligent braille button interface designed to assist visually impaired individuals in achieving fast and efficient character input through tactile button touch.
2. DESIGN AND FABRICATION OF THE SOFT SENSOR
A. Design of Soft Sensor Integrating STS Structure
External mechanical stimulation is detected by receptors in the human body in the skin, such as slow-adapting receptors (SA receptors), which then convert this stimulation into potential information [21]. SA receptors are situated close to the skin surface and exhibit a high sensitivity to skin indentation. They are densely distributed in sensitive regions of the skin, allowing for the precise detection of force information. This ability aids in the discrimination of the position and magnitude of force. The receptors encode sensory information and transmit it to the brain through nerve fibers [22]. The brain processes this information by analyzing and interpreting action potentials to make decisions and generate responses, as shown in Fig. 1(a).
Fig. 1. (a) Design inspiration of soft sensor. (b) The schematic diagram of the STS structure. (c) Artificial tactile cognitive systems that mimic biological system.
Inspired by human skin receptors, an artificial intelligence flexible sensor utilizing optical fiber sensing technology is proposed. The sensor is created by integrating a fiber optic sensing structure into a soft material, producing a soft and flexible pad that can be easily wrapped around or attached to robot surfaces. The sensor exhibits a multi-dimensional sensing function akin to human skin, allowing it to simulate the sensory process of human skin. When a mechanical stimulus impacts the soft sensor’s surface, it is translated into an optical signal. This optical signal is processed and encoded before being transmitted to a pre-set algorithm model in spectral data form. The algorithm model conducts logical analysis on the data, similar to the human brain, and provides analysis results to the human-computer interaction interface for decision-making and subsequent actions, as illustrated in Fig. 1(c).
An annular STS optical fiber structure fabricated by splicing a 60 mm long thin core fiber (TCF) with a 5/125 µm core/cladding diameter ratio between two standard single mode fibers (SMFs) with an 8.2/125 µm core/cladding diameter ratio is adopted as the sensing element, as shown in Fig. 1(b). To enhance position differentiation, the sensor’s center does not align with the center of the fiber ring structure. Compared to the traditional linear structure, the annular optical fiber sensing structure can significantly enhance the contact area between the optical fiber structure and the soft material, while also eliminating the size constraints imposed by the packaging material due to the sensing structure. As a result, the sensor can be easily adjusted in size to meet specific engineering requirements. Due to the mismatch in fiber core diameters, incident light is divided into two parts at the splice point. One part travels along the core of the fiber, while the other part enters the cladding of the thin-core fiber, exciting some cladding modes. At the second splicing point, some of the cladding modes couple into the core and interfere with the fundamental mode in the output SMF, forming a Mach–Zehnder interferometer (MZI). Its output intensity can be expressed as [23]
where ??.? and ??:? indicate the light intensities of the core and ? th-order mode in the core of the output SMF, respectively. ?=? represents the effective refractive index (ERI) difference between the fundamental mode and ? th-order mode. ? represents the wavelength of the propagating light in the vacuum, and ? represents the length of the TCF.
Given the flexibility and softness of the sensor, PDMS, a polymer material with a low Young’s modulus and high Poisson’s ratio, was chosen as the wrapping material for the sensing element due to its good resilience. PDMS serves as a medium to transmit external forces to the optical fiber structure when external stimulation is applied to the sensor surface. Variations in the force and position applied to the sensor result in alterations to the fiber structure’s length, coupling ratio, and refractive index. Axial strain in the fiber structure leads to modifications in the sensing fiber’s length, while bending of the optical fiber structure induces changes in the coupling ratio at the fusion splice point. Meanwhile, when the optical fiber structure deforms, the refractive index of silica, the material of the optical fiber, changes due to the photoelastic effect. Taking into account the impact of the force on the sensor, the output light intensity can be represented as [24]
where ??.? is the intensity of incident light. ?? represents axial stress, ?? represents radial stress, and σ represents total stress. ?11 and ?12 refer to the coupling coefficients of light coupling from the core of input SMF into the core and cladding of the TCF at the first splicing point, respectively. ?21 and ?22 represent the coupling coefficients of light in the core and cladding of the TCF recoupling back to the core of the output SMF at the second splicing point, respectively.
Furthermore, when force is applied to the sensor surface, the length of the optical fiber structure and surface stress are altered, impacting the ERI of the mode and causing wavelength drift. While the refractive index (RI) of the optical fiber changes with stress, the contribution of waveguide dispersion is considered insignificant compared to the photoelastic effect and length influence. Thus, by considering the principle of the MZI, the wavelength shift induced by force can be represented as
By analyzing Eqs. (2) and (3), it becomes evident that varying forces of different magnitudes and locations applied on the sensor surface result in different stress distributions across the optical fiber structure via the flexible material. Consequently, the variability in the output affects both the wavelength and intensity of the spectrum, enabling a single sensor to possess distributed detection of force and location.
To determine the contact force and positioning of the soft sensor surface, a sensor model along with the corresponding applied external force was developed using Comsol Multiphysics. The sensor surface is uniformly divided into nine regions, where force is applied to the sensor’s surface, which is depicted in Fig. 2. This system of soft sensors effectively localizes the contact source by leveraging the spectral output from an annular-shaped fiber optic structure. In Fig. 2(a), a top view illustrates various example positions on the sensor surface under a force of 0.1 N, with the force application location highlighted in yellow. Position 1 signifies that the force is applied to the center of the sensor rather than directly to the fiber structure. Positions 2 and 3 denote the force exerted directly on the fiber structure and the splicing points, respectively. Position 4 illustrates the application of force to two consecutive edge positions on the sensor. The theoretical simulation of stress distribution and deformation at positions 1–4 on the sensor surface is conducted by applying force. The ?-axis is defined as perpendicular to the sensor surface, while the ?- and ?-axes are parallel to it. To simplify the calculation, the optical fiber structure is approximated by the ring structure in the simulation. A perfect bonding between the optical fiber and the encapsulation polymer is assumed in the simulation. The second line in Fig. 2 depicts the stress distribution diagram of the fiber structure in the XY plane at different stress positions, while the third line illustrates the stress distribution and deformation in the YZ or XZ plane at the stress concentration position. The red dot indicates the location of the splicing points. The results indicate that stress in the fiber structure increases when the location is applied near the fiber (positions 2, 3, 4). The stress distribution on the ring varies due to the different force positions. Specifically, at position 1, stress is not evenly distributed due to the ring’s center deviating from the sensor’s center, leading to stress concentration in the deviated direction. The stress at the stress concentration position will be slightly higher for the two-point force at position 4 compared to a single-region force. Furthermore, the fiber undergoes radial bending as a result of stress. And the application of force above the fiber causes the bending to occur along the ?-axis. The results indicate that the bending position varies based on the location of the force application. Additionally, the bending degree is slightly greater for two-region forces compared to single-region forces. Variations in bending degrees can alter the RI of the fiber and the coupling ratio at the splicing points of the fiber structure. Consequently, the stress distribution and bending of the annular structure, induced by forces acting at various positions on the sensor surface, are reflected in the output spectrum. Theoretical analysis indicates that the proposed sensor possesses the capability for multi-region detection, positioning it as a promising candidate for applications in distributed detection.
Fig. 2. Diagram of the simulation results for four example positions where the forces (0.1 N) are applied. The graphs in the first line show the locations of the four applied forces. The graphs in the second line show the stress distribution of the annular fiber structure in the XY plane. The graphs in the third line show the stress distribution and deformation in the YZ or XZ plane at the stress concentration position.
B. Fabrication of the Soft Sensor
To ensure the sensor’s flexibility, the decision is made to utilize PDMS material as the sensor’s packaging material, with the detailed preparation process outlined in Fig. 3(a). The PDMS solution is formulated by the main agent and curing agent (SYLGARDTM 184 Silicone Elastomer) in a ratio of 10:1, subsequently undergoing centrifugation to eliminate air bubbles before being poured into a ?+??m?m square titanium mold. The solution is subjected to curing at 100°C for 20 min to create a soft sensor base measuring 3.2 mm in thickness. Following this, the annular STS structure is placed on the soft base and fixed using two PDMS blocks. To avoid the formation of two symmetrical positions leading to identical optical fiber surface stress, intentional asymmetry is introduced by positioning the center of the annular structure off-center from the sensor center. To ensure complete encapsulation of the fiber structure by PDMS, the PDMS solution is dripped again until the thickness is up to 5.4 mm. Employing the same curing procedure, the solution is subjected to a temperature of 100°C for 20 min, followed by removal from the mold. Ultimately, a soft sensor characterized by high flexibility and durability is fabricated, as shown in Fig. 3(b).
Fig. 3. (a) Schematic diagram of the preparation process for soft sensor. (b) Photographs of the sensor exhibiting flexible and deformable characteristics. (c) Touch process of the indenter. (d) The RF model architecture for contact force and position recognition. (e) Experimental spectra for the four representative examples.
3. METHOD
A. Experimental Setup and Protocol
The experiment utilizes an electrical pressure tester (ZQ-990) with a resolution of 0.01 N to exert force on the sensor surface. A square indenter measuring ?+??m?m is used to replicate the minimum characteristic size of objects manipulated in precise grasping tasks (10 mm) [25]. The sensor is affixed to a two-dimensional displacement platform (Physik Instrumente, M-413) with a resolution of 0.2 µm to precisely manipulate the position of the indenter relative to the sensor surface. One end of the sensor is linked to a broadband light source (BBS) covering wavelengths ranging from 1525 to 1565 nm, while the other end is connected to an optical spectral analyzer (OSA, YOKOGAWA, AQ6370 with 0.02 nm resolution) for real-time spectral data acquisition.
In the experiments, the sensor surface is equally divided into nine blocks in the XY plane with a size of ?+??m?m for every block. The indentation trajectories for both a single block and two blocks are illustrated in Fig. 3(c). In this figure, the red line denotes the path of the indenter’s movement. A force of 0.1 N is sequentially loaded/unloaded at various blocks along the S-shaped path. The procedure involves automatic force-controlled indentation, starting from block 1 with a step size of 10 mm, covering a total of nine blocks. Subsequently, forces on two different regions are tested. A force of 0.1 N is to be maintained on block 1, with additional 0.1 N forces applied at various locations along the S-shaped path from block 2 until all remaining eight blocks have been traversed. Subsequently, the constant force on block 1 is removed, and a constant force of 0.1 N is applied to block 2. Additionally, a 0.1 N force is applied along the S-shaped path to each block, except for block 2. Following a consistent methodology, all possible combinations of two regions on the sensor surface are tested with simultaneous force application, resulting in a total of 36 scenarios. Applying consistent experimental methodology, forces are exerted on three distinct blocks of the sensor surface, repeated on four different blocks, and subsequently extended to encompass nine areas. A total of 511 configurations are generated to represent various locations. The applied force ranges from 0.1 to 0.2 N with a step size of 0.1 N (two forces per position). To assess the sensor’s reliability, each loaded/unloaded force cycle is repeated four times, yielding a dataset of 4088 groups of data (511 location schemes and two forces, repeated four times).
B. Machine-Learning-Based Perception Model
Figure 3(e) presents the experimental spectra for the four representative examples from the previously described experiments. The diagram includes the spectrum of the sensor in air, as well as the spectra obtained when 0.1 N force is applied at positions 1, 2, and 3 as shown in Fig. 2. The experimental results demonstrate that variations in force and position lead to slight changes in the output spectra. These findings are consistent with the theoretical analysis. To obtain comprehensive data on position and force, an identification tool utilizing a machine-learning-based random forest model is employed, as illustrated in Fig. 3(d). The spectral data is incorporated as a dataset in the algorithm, with the light intensity at each sampling wavelength serving as a defining characteristic of the data matrix. RF randomly and replacefully selects N samples to create N training subsets, with each dataset being utilized to construct a decision tree. This results in each decision tree containing overlapping yet distinct information, thereby enhancing the diversity among decision trees. During the construction of decision trees, ? features are randomly chosen from a pool of ? features within the sample, with the optimal features being utilized to recursively partition nodes until further division is no longer feasible. By employing recursive feature selection and node splitting, multiple decision trees are generated in parallel and independently to construct a random forest, thereby enhancing the model’s stochasticity. The collective output of these decision trees is subsequently evaluated through a voting mechanism, culminating in an integrated classification result. The differences in regional and force characteristics result in minor alterations in output responses, ultimately producing a high degree of similarity in output spectra. This similarity can lead to an overreliance on the training data within the model, impeding its capacity to generalize effectively to novel data and resulting in overfitting. The utilization of random sampling in constructing decision trees within RF significantly mitigates the risk of overfitting and enhances the model’s robustness and generalization capabilities. Additionally, the collaborative efforts of multiple decision trees in classifying outcomes contribute to improved accuracy and stability of the model.
Fig. 4. (a) Verification results and (b) test results of contact position and applied force for different algorithms.
4. RESULTS AND DISCUSSION
A. Perception Model Training Results
Models I, II, and III are representative of training models utilizing varying quantities of data, with a training data to test data ratio of 8:2. Model I exclusively utilizes data from Cycle 1 for training and validation, Model II integrates data from Cycles 1 and 2 for training and validation, and Model III incorporates data from Cycles 1, 2, and 3 for training and validation. In each set of models, the data from Cycle 4 is designated as the test data. The test accuracy of the contact position and applied force for each model is presented in Table 1. The findings indicate that models trained with a larger dataset exhibit improved accuracy.
To assess the recognition performance of different algorithms, a comparison was conducted between the support vector machine (SVM), K-nearest neighbors (KNN), back propagation neural network (BP), linear discriminant analysis (LDA), multilayer perceptron (MLP), and RF. The recognition accuracy of each algorithm by using Model III is presented in Fig. 4. Figure 4(a) displays the verification results, while Fig. 4(b) illustrates the test results. The findings indicate that the RF algorithm effectively addresses sensor force and position recognition challenges, achieving a recognition accuracy exceeding 96%. This phenomenon can be primarily ascribed to the subtle alterations in spectra caused by the force and position response, leading to a significant approximation in the output spectrum that is susceptible to overfitting. RF incorporates randomness and ensemble learning techniques, effectively mitigating issues related to local optimization stemming from high data similarity. Consequently, it is better suited for addressing intricate recognition problems in high-dimensional and complex datasets.
B. Force Response
The result presented in Table 1 demonstrates that the sensor achieves 100% force detection accuracy when the contact force interval is 0.1 N. To further assess the sensor’s capacity to detect contact forces, experiments are conducted to evaluate force and contact position in nine blocks ranging from 0.1 to 0.2 N with a step of 0.01 N. Data is collected at contact force intervals of 0.02 and 0.01 N, and the sensor recognition accuracy in three distinct training models is presented in Table 2. Among these, Model IV utilizes the data from Cycle 1 for both training and verification, while Model V employs the data from Cycles 1 and 2 for the same purposes. In contrast, Model VI incorporates the data from Cycles 1, 2, and 3 for training and verification. Notably, the test data for all three models is sourced from the data of Cycle 4. It is observed that even with a reduction in contact force resolution to 0.02 N, the sensor exceeds 99% accuracy in force recognition. Additionally, while an increase in training data improved identification accuracy, accuracy fading is observed with further reductions in contact force resolution. Thus, the proposed sensor exhibits great recognition capabilities at a force resolution of 0.02 N.
Table 2. Sensing Accuracy of Contact Force for Different Training Models
To comprehensively understand the force response characteristics of the sensor, the recognition capabilities of the sensor across a broader range of force are evaluated experimentally. The experiments assess force and contact position across nine blocks, ranging from 0.06 to 0.1 N and from 0.2 to 0.26 N, with a step of 0.02 N. Models VII and VIII are developed to differentiate the data corresponding to the two force ranges of 0.06–0.1 N and 0.2–0.26 N. Model IX encompasses all data with a step size of 0.02 N within the range of 0.06–0.26 N. The data from Cycles 1, 2, and 3 is utilized for training and verification, while the data from Cycle 4 is employed for testing. The identification results are presented in Table 3, which indicates that although the force detection range of the sensors is expanded, their force recognition accuracy consistently exceeds 98%. This implies that enhancing the sensor’s force capabilities does not significantly affect its recognition accuracy. The range of force detection can be tailored to meet specific requirements, thereby offering significant potential for a broad range of force detection applications.
Table 3. Sensing Accuracy of Contact Force for Different Training Models
C. Spatial Response
The spatial response capability of the sensor is assessed through a series of tests. The sensor surface is divided into blocks of varying areas in the XY plane. To ensure uniform force distribution across each block during the experiment, a slide matching the size of each block is placed accordingly. The contact area of a square indenter, measuring ?+??m?m, is selected to ensure it does not exceed the dimensions of the smallest subdivided block. The sensor surface is uniformly divided into 16 blocks, each with an equal area of ?+??m?m. A 0.1 N contact force is successively applied to various blocks of the sensor along the S-shaped path until all 16 blocks have been traversed. Each loaded/unloaded force cycle is repeated four times. Subsequently, the sensor surface is divided into 25, 36, 49, 64, and 81 equal-area regions, and the aforementioned experiment is replicated for each region. The first three cycles of experimental data are utilized for training and verification, while the fourth cycle serves as the test dataset, and the result is shown in Fig. 5(a). The results indicate that when the region area is reduced to ?+??m?m, the detection accuracy remains above 95%. However, further reduction in the block area leads to a significant decline in detection accuracy. These findings suggest that the proposed sensor is capable of achieving spatial resolution detection at ?+??m?m.
Fig. 5. (a) The impact of varying block quantities on recognition accuracy across different models. (b) Schematic diagram of scanning range and spatial resolution of XY plane. (c) The influence of temperature variations on recognition accuracy across different models.
Furthermore, the proposed sensor exhibits the capability to adjust the interference length according to specific engineering requirements, thereby expanding the detection area. Modifying the length of the sensing fiber TCF is beneficial for increasing the detection area. However, altering the sensing length leads to a variation in the free spectral range (FSR) of the output spectrum, a relationship that can be mathematically expressed as [26]
As demonstrated by Eq. (4), an extension in sensing length results in a heightened spectral density. Within the context of conventional spectral detection methods, an increase in the FSR can lead to the superposition of interference peaks, which are induced by spectral drift during the detection process, thereby making it difficult for observational efforts. However, the application of algorithms can effectively mitigate this issue. The increase in the FSR implies that a greater amount of interference information is embedded within the spectrum. Consequently, the RF algorithm can extract more feature information from this enriched dataset for training and recognition, thereby enhancing the accuracy of recognition.
The experiments assessed the sensor’s potential capability to detect extensive areas. Soft sensors of varying dimensions are fabricated by using TCFs of different lengths. An annular structure is created by splicing SMF on both sides of a 90 mm long TCF, serving as the sensing structure. This sensing structure is embedded within PDMS measuring ?+??m?m, designated as Sensor 2. Another sensor, featuring an annular structure formed by a 110 mm long TCF, is embedded in PDMS with dimensions of ?+??m?m, designated as Sensor 3. The distance between the annular structure and the upper surface of the sensor is 3.2 mm. The soft sensor described above has dimensions of ?+??m?m and is made of a 60 mm long TCF, designated as Sensor 1. The surfaces of the three sensors are uniformly divided into nine blocks in the XY plane. The dimensions of each block for the three sensors are ?+??m?m, ?+??m?m, and ?+??m?m, respectively. The scanning range and spatial resolution of the XY plane are shown in Fig. 5(b). To ensure uniform force distribution, a slide matching the size of each area is placed on the surface. The indenter traverses each block sequentially along an S-shaped path at a force of 0.1 N, cycling four times for each load and unload cycle. The first three cycles of experimental data are utilized for training and verification, while the fourth cycle serves as the test dataset, achieving a recognition accuracy above 95% for Sensor 2 and Sensor 3. This indicates that the sensor can adapt its detection area based on varying requirements, demonstrating potential for large-area detection. However, as the area of the partition block is further reduced and the detection area is expanded, the sensor’s detection accuracy may encounter challenges.
The above experiments are conducted in a controlled environment at a constant temperature of 24°C. To assess the impact of temperature on sensor detection accuracy, multiple sets of position recognition experiments are performed at varying temperature ranges. These experiments followed the same methodology as the force and position identification experiments, just under different temperature ranges. The initial experiment is conducted at 24°C, while four verification experiments are carried out within specific temperature ranges: 24.4–25.3°C (T1), 22.9–23.8°C (T2), 21–22.5°C (T3), and 19.7–20.4°C (T4). For Model X, 80% of the initial experimental data was randomly selected as the training set, with 20% of the data from each of the four temperature ranges designated as the test set. Concurrently, for Model XI, 80% of the initial experimental data and 80% of the data from each temperature range were randomly selected as the training set, while the remaining 20% of the data in each of the four temperature ranges served as the test set. The recognition accuracies of the two training models are illustrated in Fig. 5(c). The results indicate that as the temperature difference from the initial experiment increases, the recognition accuracy declines significantly. This suggests that variations in temperature have a substantial effect on recognition accuracy. Incorporating experimental data that includes temperature information can greatly enhance the sensor’s recognition accuracy. Notably, the recognition accuracy of Model XI exceeds 95%, demonstrating that the sensor can maintain high-precision recognition of multi-region locations within a narrow temperature range.
D. Sensor Performance
To evaluate the sensor’s rapid response and recovery capabilities in reaction to sudden mechanical stimulation, we conducted experimental assessments of both the response time and recovery time of the sensor. When a contact force of 0.1 N is applied to the center of the sensor (position 1 in Fig. 2) for a duration of 2 s, a significant change in the output spectrum is observed. The specific dip is continuously monitored at a sampling interval of 0.2 s, as illustrated in Fig. 6(a). The results reveal that the response time from loading to stability is 0.4 s, while the recovery time from unloading to stability is approximately 0.5 s. These findings further illustrate that the proposed sensor exhibits rapid response and recovery capabilities, making it suitable for real-time monitoring of tactile signals. Meanwhile, mechanical stability is also a critical parameter for sensors. To assess the reliability of the sensor, we conducted 100 load/unload cycles under consistent conditions, maintaining a holding time of 2 s for each cycle, as illustrated in Fig. 6(b). The results indicate that the sensor maintains stable signal output even after undergoing numerous cycles. Following the completion of 100 cycles, the sensor is retested for force and position identification, achieving an identification accuracy exceeding 95%. This demonstrates that the proposed sensor exhibits exceptional durability and reliability. Furthermore, the repeatability of the sensor was experimentally assessed by applying contact forces of varying magnitudes at the center of the sensor (position 1 in Fig. 2). The contact force is increased from 0.1 to 0.2 N and then decreased back to 0.1 N in a single cycle, with increments of 0.02 N. This loading and unloading cycle is performed three times. The specific dip in the output spectrum is monitored, and linear fittings for the three cycles are generated, as illustrated in Fig. 6(c). In this context, “PA” denotes the process of force increase, while “RR” signifies the process of force decrease. The results demonstrate that the sensor exhibits excellent repeatability, with a repeatability error of less than 2% across the three cycles.
Fig. 6. (a) Dynamic response characteristics of tactile force for the sensor. (b) Force reliability test of the sensor. (c) Repeatability test of sensor.
Fig. 7. (a) Representation of basic Braille. (b) The FR model architecture for intelligent Braille touch interaction. (c) Schematic diagram of the intelligent Braille keyboard. (d) Confusion matrix of number recognition. (e) Intelligent Braille touch interface based on the soft sensor. The examples of the successful input of Braille “F” and “:”.
E. Intelligent Braille Key Interaction
Based on the force positioning function of the soft sensor, we have developed an interface for intelligent Braille key recognition. To enhance accessibility for visually impaired individuals, the standard Braille intelligent keyboard employs six raised keys to represent frequently used letters, numbers, and special characters. When multiple keys are pressed concurrently, they correspond to specific letters and symbols, as illustrated in Fig. 7(a) [27]. Additionally, when the “number” key is activated, numerical input is generated. The bottom line in Fig. 7(a) illustrates the key combinations associated with various numbers. In this study, we expand the Braille intelligent keyboard to include nine raised keys, incorporating three additional keys designated for “Delete”, “Blank space”, and “Enter”. The arrangement of these keys is shown in Fig. 7(b). To facilitate tactile interaction for visually impaired users, a circular protrusion with a radius of 5 mm is fabricated from PDMS and affixed to nine blocks on the surface of the sensor. This results in the development of a multifunctional intelligent touch key designed specifically for the blind, utilizing a soft sensor. A multifunctional touch interface is implemented within the LabVIEW-MATLAB environment. When a finger touches the button, the soft sensor rapidly generates and maintains a signal, which is then transmitted as a response signal to a personal computer. The RF model predicts effective button interactions by analyzing real-time data. It is noticed that signal fluctuations caused by false touches or material creep can lead to misjudgments in practical applications. To mitigate this issue, time and force thresholds have been established to determine the user’s subjective operation. Specifically, an action is considered effective and subsequently input into the recognition system only when the action duration exceeds 0.4 s and the applied force surpasses 0.1 N. Thus, the stability of the system can be enhanced, and the erroneous identification resulting from inadvertent user contact can be effectively prevented. The judgment flow chart is illustrated in Fig. 7(c). The recognition accuracy of all key combinations presented in Fig. 7(a) is evaluated. Due to the extensive volume of letter and character data, the confusion matrix contains a substantial amount of information, which may hinder observation. Here, only the confusion matrix for digital recognition is presented, as shown in Fig. 7(d). The matrix indicates that the recognition rate for numeric keys is 96%. An analysis of the confusion matrix reveals that misjudgments primarily arise from similar key combinations. Figure 7(e) presents the photographs illustrating the testing of an intelligent Braille recognition system with the letter “F” and the character “:”. The highlighted section within the figure demonstrates the real-time Braille touch interface. The upper section of the interface displays the spectrum generated by the buttons in real-time, while the lower section presents the recognized keys by the algorithm alongside the generated characters. The experimental results validate the sensor’s effectiveness for Braille input. For visually impaired users, the proposed intelligent Braille key interaction system is capable of real-time detection of the user’s touch on the keyboard, thereby enabling rapid and efficient character input.
Table 4. Comparison Results of Other Braille Sensors
View Table | View all tables in this article
Table 4 presents a comparison of the results between the sensor proposed in this article and representative Braille key sensors reported in the literature. As indicated in the table, various sensing mechanisms can be employed for Braille input recognition. The fabrication of electrical sensors typically necessitates multi-layer composite structures, resulting in a complex preparation process. Furthermore, existing Braille key recognition sensors always require sensor arrays to facilitate multiple key detection [28–30]. Notably, the proposed sensor effectively recognized 26 English letters, 10 numerical digits, and common symbols using a single unit. This approach successfully addresses the complex wiring and crosstalk issues associated with sensor arrays, while also reducing the costs. In contrast, the sensor featuring an annular STS embedded in PDMS offers a simpler structure, facilitates easy fabrication, and achieves high recognition accuracy.
5. CONCLUSION
This paper introduces a novel soft tactile sensor utilizing an optical fiber structure. The sensor employs an optical fiber annular structure, fabricated using STS, which is embedded in PDMS. When the force is applied on the surface of the sensor, the stress is transmitted through the soft material and concentrated on the optical fiber structure. This interaction induces changes in the wavelength and intensity of the output spectrum, thereby enabling the detection of tactile stimuli. To validate this hypothesis, a theoretical model of the sensor was developed. Variations in the force position of the sensor induce distinct strains within the fiber structure, leading to differential output responses. Consequently, the sensor demonstrates the capability to differentiate between various forces and positions.
The sensor surface is partitioned into nine equally sized blocks, and recognition experiments are performed to assess various forces and block positions. The experimental data is analyzed utilizing the RF algorithm, achieving a recognition accuracy exceeding 96% for both force and position. Minor variations in force and position result in subtle changes in the output spectrum, thereby leading to high data similarity. The RF algorithm is particularly well-suited for addressing intricate recognition problems within high-dimensional and complex datasets, as evidenced by its significant advantages over other algorithms in comparative evaluations. Furthermore, in the context of force response recognition, the accuracy of detecting contact forces with the range from 0.1 to 0.2 N reaches 100%, with a force resolution of 0.1 N. To further investigate the sensor’s force recognition capabilities, tests are conducted at a finer resolution of 0.02 N, demonstrating that the recognition accuracy of contact forces remains above 99%. The force detection range is expanded to 0.06–0.26 N, and it is found that the sensor continued to show excellent recognition characteristics. For spatial response recognition, the sensor surface is uniformly segmented into varying numbers of blocks to evaluate the sensor’s performance at different spatial resolutions. When the spatial resolution is decreased to 3.75×3.75?m?m, the sensor sustains an identification accuracy exceeding 95% for contact position. Simultaneously, due to its annular fiber structure, the proposed sensor demonstrates the capability to conform to soft materials and flexibly adjust its size, resulting in enabling large-area skin detection. Experimental results reveal that the position recognition accuracy is still above 95% when the area is expanded to 48×48?m?m. In the traditional method, an increased interference length results in heightened interference, thereby complicating observation. However, this approach can effectively enhance spectral features, making them more conducive to recognition by the RF algorithm. To evaluate the response characteristics of the sensor in real-world conditions, a contact position identification experiment is conducted with a temperature range from 19.7°C to 25.3°C. The results indicate that recognition accuracy diminishes as the temperature difference increases. However, by incorporating data that includes temperature information into the training set of the RF algorithm, the recognition accuracy is effectively improved.
To further investigate the potential application capabilities of soft sensors in enhancing human-computer interaction, this study proposes an intelligent Braille key interactive system. This keyboard employs nine intelligent buttons to facilitate real-time character input and display for visually impaired users. The soft tactile sensor utilizing a single fiber optic structure represents a two-dimensional tactile perception solution. It effectively addresses the challenges related to complex wiring and crosstalk commonly encountered in conventional distributed sensors, thereby holding substantial potential for applications in human-computer interaction and intelligent robotics.
