Bragg grating-based dexterous finger tactile sensor and preparation and application thereof

Abstract

The application discloses a kind of based on Bragg grating nimble finger tactile sensor and its preparation and application, wherein, tactile sensor includes elastic matrix, first optical fiber and second optical fiber, first optical fiber is linearly horizontally through elastic matrix and is arranged, second optical fiber is horizontally through elastic matrix and is arranged in U type, and first optical fiber and second optical fiber are horizontally perpendicular intersection, first optical fiber is provided with a plurality of first Bragg grating, and second optical fiber is provided with a plurality of second Bragg grating.Wherein, the application of tactile sensor is specifically FMEM demodulation model based on deep learning neural network, the spectrum information of different positions, different directions, different sizes under the force exerted on the nimble finger tactile sensor based on Bragg grating is demodulated, to realize the demodulation of multidimensional force information.The advantages of the application are: small, simple to manufacture, high resolution, can realize accurate demodulation from spectral signal to multidimensional force signal, meet the requirements of nimble hand to tactile perception.

Description

Technical Field

[0001] This invention relates to the field of tactile sensor technology, specifically to a dexterous finger tactile sensor based on a Bragg grating, its fabrication, and its application. Background Technology

[0002] In the development of intelligent robot technology, tactile perception is crucial for the robot's fine manipulation and adaptability to complex environments. High-performance tactile sensors help robots perceive their external environment more accurately, enabling stable grasping, object manipulation, and safe human-robot interaction. Therefore, tactile sensing technology has received widespread attention in fields such as medical rehabilitation, industrial automation, and service robots. However, existing tactile sensing technologies still face many challenges in terms of measurement accuracy, environmental adaptability, and multi-dimensional information perception capabilities, hindering their development in complex application scenarios.

[0003] Currently, tactile sensors mainly include piezoresistive, capacitive, triboelectric, and piezoelectric types. Piezoresistive sensors are low-cost and simple in structure, but are susceptible to environmental noise and have relatively low measurement accuracy. Capacitive sensors offer high resolution but are prone to electromagnetic interference, resulting in poor stability. Triboelectric sensors are self-powered and suitable for dynamic measurements, but their signals are easily affected by environmental factors, leading to unstable output. Piezoelectric sensors have fast response times and are suitable for high-frequency dynamic measurements, but their performance is limited in static pressure detection. Furthermore, while vision- and optical-based tactile sensing technologies can provide rich information, they are susceptible to occlusion and changes in lighting conditions, and their data processing complexity is high.

[0004] In recent years, fiber Bragg grating (FBG) sensors have become a hot topic in tactile sensing research due to their advantages such as resistance to electromagnetic interference, small size, high sensitivity, and corrosion resistance. FBG sensors sense external forces by detecting changes in spectral signals and can achieve multi-point distributed measurements on a single fiber, making them suitable for robotic tactile sensing.

[0005] However, despite the numerous advantages of fiber Bragg grating (FBG) technology, it still faces challenges in application. First, traditional peak detection methods are susceptible to noise interference, leading to measurement errors, especially exhibiting poor stability in complex environments. Second, most FBG sensors struggle to simultaneously and accurately sense multidimensional signals such as the magnitude, direction, and position of force, while multidimensional force information is crucial for the precise manipulation of robots. Furthermore, the structural design of FBG sensors is relatively fixed; improving flexibility and integration to adapt them to diverse application scenarios is also a key focus of current research.

[0006] It is evident that existing tactile sensors still suffer from the following shortcomings: limited resolution, making it difficult to achieve multi-dimensional force detection. Summary of the Invention

[0007] To address the shortcomings in the prior art, the present invention aims to propose a Bragg grating-based dexterous finger tactile sensor capable of multidimensional force detection, and to propose an FMEM demodulation model based on a deep learning neural network for demodulating the spectral information of applied forces on the Bragg grating-based dexterous finger tactile sensor under different positions, directions, and magnitudes, so as to achieve demodulation of multidimensional force information.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] In a first aspect, a dexterous finger tactile sensor based on a Bragg grating is provided, comprising an elastic substrate, a first optical fiber, and a second optical fiber. The first optical fiber is arranged in a straight line and horizontally penetrates the elastic substrate, and the second optical fiber is arranged in a U-shape and horizontally penetrates the elastic substrate. The first optical fiber and the second optical fiber intersect vertically. A plurality of first Bragg gratings are disposed on the first optical fiber, and a plurality of second Bragg gratings are disposed on the second optical fiber.

[0010] Furthermore, the elastic matrix is ​​made of PDMS material.

[0011] Furthermore, the length, width, and height of the elastic matrix are 30 mm, 20 mm, and 2 mm, respectively; the diameter of the first optical fiber and the second optical fiber is 125 μm; and the length of the first Bragg grating and the second Bragg grating is 4 mm.

[0012] Secondly, a method for fabricating a dexterous finger tactile sensor based on a Bragg grating is provided, comprising the following steps:

[0013] S1. Embed the first and second optical fibers into the 3D printing mold according to the designed length;

[0014] S2. Add A and B to a stirring device in a ratio of 13:1 and mix evenly to obtain a PDMS mixture; wherein, A is a polydimethylsiloxane prepolymer and B is a hydrogen-containing polysiloxane crosslinking agent.

[0015] S3. Place the PDMS mixture obtained in step S2 into a vacuum chamber to degas and remove bubbles, and obtain a uniform PDMS mixture without bubbles.

[0016] S4. Add the uniform, bubble-free PDMS mixture obtained in step S3 into the 3D printing mold in step S1 in which the first and second optical fibers are embedded, and cure it at room temperature (26°C) for 24 hours. After demolding, the dexterous finger tactile sensor based on the Bragg grating is obtained.

[0017] Thirdly, a system for demodulating spectral signals from a Bragg grating-based dexterous finger tactile sensor is provided, comprising an FBG demodulator, wherein an FMEM demodulation model based on a deep learning neural network is embedded within the FBG demodulator.

[0018] Furthermore, the FMEM demodulation model based on deep learning neural networks includes an input layer, a patch embedding layer, a position encoding layer, a Transformer encoder layer, a flattening layer, a linear layer, and a multilayer perceptron layer.

[0019] The input layer, serving as the data input entry point for the deep learning neural network, is used to perform preliminary data standardization processing on the raw spectral data collected from the Bragg grating-based dexterous finger tactile sensor, converting it into a data format suitable for neural network processing, and providing a unified data foundation for subsequent processing.

[0020] The Patch embedding layer is used to divide the data output from the input layer into multiple small regions (patches).

[0021] The location encoding layer is used to map each small region patch to a high-dimensional feature space and add location information to each small region patch, so that the neural network can distinguish patches at different locations and preserve the sequence order relationship of the spectral data, ensuring that the neural network can identify location-related features in the spectral data.

[0022] The Transformer encoder layer is used to capture long-distance dependencies between different parts of the spectral data through a multi-head attention mechanism and to extract key feature patterns and correlations in the spectral data through multi-head attention and feedforward network concatenation processing.

[0023] The flattening layer is used to flatten the multidimensional feature representation output by the Transformer encoder layer into a one-dimensional vector.

[0024] The linear layer is used to receive the one-dimensional vector output by the flattening layer, and to map the high-level features extracted by the Transformer to the feature space related to the force signal, so as to provide a linear transformation representation of appropriate dimension for subsequent processing.

[0025] The multilayer perceptron layer is used to receive the feature representations output by the linear layer, further process these features through nonlinear transformation, perform regression and classification tasks, and convert the processed features into the required multidimensional force information, realizing a complete demodulation mapping from spectral features to force information.

[0026] Furthermore, the Transformer encoder layer consists of a multi-head attention sublayer, a regularization sublayer, and a feedforward network sublayer; the multilayer perceptron layer consists of multiple fully connected layers and nonlinear activation functions.

[0027] Fourthly, a method for demodulating the spectral signal of a dexterous finger tactile sensor based on a Bragg grating is provided, the method being implemented using the aforementioned FMEM demodulation model based on a deep learning neural network.

[0028] Furthermore, the method includes the following steps:

[0029] T1. Acquire spectral signals under different positions, directions, and magnitudes of applied force on a Bragg grating-based dexterous finger tactile sensor using an FBG demodulator;

[0030] T2. Input the spectral signal acquired in step T1 into the FMEM demodulation model based on deep learning neural network. Demodulate the input spectral signal through the FMEM demodulation model based on deep learning neural network to obtain the demodulation result of multidimensional force information applied to the dexterous finger tactile sensor based on Bragg grating.

[0031] Furthermore, in step T2, the input spectral signal is demodulated using an FMEM demodulation model based on a deep learning neural network to obtain the demodulated result of multidimensional force information applied to the dexterous finger tactile sensor based on a Bragg grating. This specifically includes the following steps:

[0032] T201. The spectral signal acquired in step T1 is received through the input layer, i.e., the raw spectral data. The raw spectral data is then subjected to preliminary data standardization processing in the input layer to convert it into a data format suitable for neural network processing.

[0033] T202. In the Patch embedding layer, the spectral data obtained in step T201, which is suitable for neural network processing, is divided into multiple small regions (patches).

[0034] T203. In the position encoding layer, each small region patch obtained in step T202 is mapped to a high-dimensional feature space, and position information is added to each small region patch to preserve the sequence order relationship of the spectral data.

[0035] T204. The spectral data output in step T203 is processed through the Transformer encoder layer. First, the long-distance dependencies between different parts of the spectral data are captured using a multi-head attention mechanism. Then, the key feature patterns and correlations in the spectral data are extracted through the concatenation of multi-head attention and feedforward networks.

[0036] T205. The multidimensional feature representation output by the Transformer encoder layer is flattened into a one-dimensional vector through a flattening layer.

[0037] T206: Receives a one-dimensional vector output from the flattening layer through a linear layer, maps the high-level features extracted by the Transformer to the feature space related to the force signal, and outputs a linear transformation representation of appropriate dimension.

[0038] T207 receives the feature representations output by the linear layer through the multilayer perceptron layer, and further processes these features through nonlinear transformation to perform regression and classification tasks. Then, the processed features are converted into the required multidimensional force information, thus realizing the complete demodulation mapping from spectral features to force information.

[0039] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0040] (1) A dexterous finger tactile sensor based on Bragg grating is proposed. It is small in size, simple to manufacture, and has high resolution. It can realize high-resolution force sensing and multi-dimensional force detection and is easy to integrate into the fingertip of a robot.

[0041] (2) A deep learning neural network-based FMEM demodulation model is proposed, which combines a patch strategy to treat spectral data as time-series data and use the correlation between data to derive multidimensional force information from spectral data with limited variation. It can simultaneously achieve multidimensional accurate prediction of the magnitude, position and direction of force.

[0042] (3) A FMEM demodulation model based on deep learning neural network is proposed. The Transformer encoder layer adopts a multi-head attention mechanism to process the sequence in parallel, which reduces information redundancy and can capture long-distance dependence and location-specific features well. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in this embodiment, the accompanying drawings used in the description of the embodiment will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram of the structure of an embodiment of the dexterous finger tactile sensor based on a Bragg grating provided by the present invention;

[0045] Figure 2 This is a schematic diagram showing the distribution of the first and second optical fibers within the elastic matrix.

[0046] Figure 3This is a flowchart illustrating the fabrication process of an embodiment of the dexterous finger tactile sensor based on a Bragg grating provided by the present invention.

[0047] Figure 4 This is a sensitivity test schematic diagram of an embodiment of the dexterous finger tactile sensor based on a Bragg grating provided by the present invention;

[0048] Figure 5 This is a sensitivity test result diagram of an embodiment of the dexterous finger tactile sensor based on a Bragg grating provided by the present invention;

[0049] Figure 6 This is a schematic diagram of the structure of the FMEM demodulation model based on deep learning neural network provided by the present invention;

[0050] Figure 7 This is an experimental setup designed to evaluate the performance of the Bragg grating-based dexterous finger tactile sensor provided in this invention;

[0051] Figure 8 yes Figure 1 A schematic diagram of a tactile sensor embodiment where the surface is divided into 24 position units;

[0052] Explanation of reference numerals in the attached drawings: 100, elastic matrix; 200, first optical fiber; 300, second optical fiber; 400, first Bragg grating; 500, second Bragg grating; 600, 3D printing mold; 700, stirring device; 800, vacuum chamber. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0054] Example 1: See Figure 1 and Figure 2 This invention proposes an embodiment of a dexterous finger tactile sensor based on Bragg gratings, including an elastic substrate 100, a first optical fiber 200, and a second optical fiber 300. The first optical fiber 200 is arranged horizontally through the elastic substrate 100 in a straight line, and the second optical fiber 300 is arranged horizontally through the elastic substrate 100 in a U-shape. The first optical fiber 200 and the second optical fiber 300 intersect vertically. Three first Bragg gratings 400 are arranged on the first optical fiber 200, and four second Bragg gratings 500 are arranged on the second optical fiber 300.

[0055] Specifically, in Example 1, the elastic matrix 100 is made of PDMS material. The Bragg grating-based dexterous finger tactile sensor provided by this invention selects PDMS, which exhibits excellent bending elasticity and structural stability, as the matrix material, which can effectively ensure reliable force transmission and consistent sensor behavior under various load conditions.

[0056] Specifically, in Embodiment 1, the length, width, and height of the elastic substrate 100 are 30 mm, 20 mm, and 2 mm, respectively. To balance the robustness and sensitivity of the fiber Bragg grating (FBG), this embodiment selects an elastic substrate 100 with a flexible encapsulation thickness of 2 mm to encapsulate the first Bragg grating 400 and the second Bragg grating 500 on the first fiber 200 and the second fiber 300. Furthermore, the surface of the elastic substrate 100 is systematically divided into 4×6=24 position units in a 4x6 grid, as shown below. Figure 8 As shown in the figure (AX represents the 24 position units), and the size of each position unit is 5×5 mm².

[0057] Specifically, in this embodiment 1, the diameters of the first optical fiber 200 and the second optical fiber 300 are both 125 μm, and the lengths of the first Bragg grating 400 and the second Bragg grating 500 are both 4 mm.

[0058] More specifically, in this embodiment 1, the center wavelengths of the three first Bragg gratings 400 are 1543.77 nm, 1545.31 nm and 1546.34 nm, respectively; and the center wavelengths of the four second Bragg gratings 500 are 1542.73 nm, 1545.21 nm, 1546.82 nm and 1548.68 nm, respectively.

[0059] More specifically, in this embodiment 1, two adjacent first Bragg gratings 400 are spaced 1 mm apart; the four second Bragg gratings 500 are spaced 6 mm, 70 mm, and 6 mm apart respectively.

[0060] Example 2: See Figure 3 This invention provides a flowchart of the fabrication process for a dexterous finger tactile sensor based on a Bragg grating, which specifically includes the following steps:

[0061] S1. Embed the first optical fiber 200 and the second optical fiber 300 into the 3D printing mold 600 according to the designed length;

[0062] S2. Add A and B adhesives in a ratio of 13:1 to the stirring device 700 and mix them evenly to obtain a PDMS mixture; wherein, A adhesive is a polydimethylsiloxane prepolymer and B adhesive is a hydrogen-containing polysiloxane crosslinking agent.

[0063] S3. Place the PDMS mixture obtained in step S2 into a vacuum chamber at 800°C to degas and remove bubbles (i.e., any air bubbles) to obtain a uniform PDMS mixture without bubbles.

[0064] S4. The uniform, bubble-free PDMS mixture obtained in step S3 is added to the 3D printing mold 600 in step S1 in which the first optical fiber 2 and the second optical fiber 3 are embedded, and cured at room temperature of 26°C for 24 hours. After demolding, a dexterous finger touch sensor based on Bragg grating (referred to as Fibertouch sensor) is obtained.

[0065] Specifically, the A and B adhesives are mixed in a 13:1 ratio to ensure the grating's sensitive response to strain while maintaining the sensor's flexibility.

[0066] The present invention provides a dexterous finger tactile sensor based on fiber Bragg gratings, which achieves tactile sensing by embedding multiple fiber Bragg gratings (FBGs) into an elastic material (i.e., an elastic matrix 100), where pressure can be converted into tensile strain on the optical fibers. The sensitivity characterization of these embedded fiber Bragg gratings (FBGs) is crucial for evaluating their force measurement capabilities, as it directly affects the sensor's ability to convert mechanical deformation into measurable optical signals.

[0067] To verify the sensitivity of the dexterous finger tactile sensor based on a Bragg grating provided in this invention, the applicant conducted corresponding tests using the aforementioned embodiment of the dexterous finger tactile sensor based on a Bragg grating as the test object. Test method: as follows... Figure 4 As shown, by applying force at five different locations (1, 2, 3, 4, and 5) on the surface of the aforementioned dexterous finger tactile sensor embodiment based on a Bragg grating, and through a system calibration process, the fiber Bragg grating (FBG) response was measured within a controlled force application range (0 to 15 N). Linear regression analysis was performed on wavelength shift versus applied force at the five measurement points (1, 2, 3, 4, and 5). Sensitivity test results are shown below. Figure 5 As shown.

[0068] pass Figure 5 It can be seen that the slopes of the linear fits at points 1 and 4 are very similar, reflecting their positional symmetry relative to the fiber Bragg grating (FBG). Points 2 and 3, being farther from the FBG, show slightly different sensitivities; point 5, located directly above the FBG, has a slope of 107.75 pm / N. These sensitivity test results demonstrate that the FBG responds even without applying force directly above it, proving that FBGs can be used as sensing materials in the design of tactile sensors.

[0069] Example 3: This embodiment of the invention proposes a system for demodulating the spectral signal of a dexterous finger tactile sensor based on a Bragg grating, comprising an FBG demodulator, which internally embeds an FMEM demodulation model based on a deep learning neural network; wherein, FMEM is an abbreviation for force multi-information estimation model.

[0070] Specifically, in this embodiment 3, see... Figure 6 The FMEM demodulation model based on deep learning neural networks includes an input layer, a patch embedding layer, a position encoding layer, a Transformer encoder layer, a flattening layer, a linear layer, and a multilayer perceptron layer.

[0071] The input layer, serving as the data input entry point for the deep learning neural network, is used to perform preliminary data standardization processing on the raw spectral data collected from the Bragg grating-based dexterous finger tactile sensor, converting it into a data format suitable for neural network processing, thus providing a unified data foundation for subsequent processing.

[0072] The Patch embedding layer is used to divide the data output from the input layer into multiple small regions (patches).

[0073] The location encoding layer is used to map each small region patch to a high-dimensional feature space and add location information to each small region patch, enabling the neural network to distinguish patches at different locations and preserve the sequential order of spectral data, ensuring that the neural network can identify location-related features in spectral data;

[0074] The Transformer encoder layer is used to capture long-range dependencies between different parts of the spectral data through a multi-head attention mechanism and to extract key feature patterns and correlations in the spectral data through multi-head attention and feedforward network concatenation.

[0075] The flattening layer is used to flatten the multidimensional feature representation output by the Transformer encoder layer into a one-dimensional vector.

[0076] The linear layer receives the one-dimensional vector output from the flattening layer and maps the high-level features extracted by the Transformer to the feature space related to the force signal, providing a linear transformation representation of appropriate dimension for subsequent processing.

[0077] The multilayer perceptron layer receives the feature representations output by the linear layer, further processes these features through nonlinear transformations, performs regression and classification tasks, and converts the processed features into the required multidimensional force information, realizing a complete demodulation mapping from spectral features to force information.

[0078] Specifically, the Transformer encoder layer consists of a multi-head attention sublayer, a regularization sublayer, and a feedforward network sublayer. The multi-head attention sublayer, by computing multiple sets of different attention weights in parallel, can simultaneously focus on different positions and feature representations of the input sequence, thereby capturing richer contextual information. The regularization sublayer, through techniques such as dropout, can reduce the risk of model overfitting and enhance the model's generalization ability. The feedforward network sublayer, composed of two linear transformations and a nonlinear activation function, is used to nonlinearly transform the features extracted by the multi-head attention sublayer, further improving the model's expressive power. In this invention, the multi-head attention sublayer can effectively capture important spectral features and peak shift patterns in long-term optical signals because it uses multiple attention heads working in parallel; that is, each attention head learns various aspects of the wavelength-intensity relationship.

[0079] Specifically, a multilayer perceptron layer consists of multiple fully connected layers and nonlinear activation functions.

[0080] Example 4: A method for demodulating the spectral signal of a dexterous finger tactile sensor based on a Bragg grating, specifically implemented through an FMEM demodulation model based on a deep learning neural network.

[0081] Specifically, in this embodiment 4, the method for demodulating the spectral signal of a Bragg grating-based dexterous finger tactile sensor includes the following steps:

[0082] T1. Acquire spectral signals under different positions, directions, and magnitudes of applied force on a Bragg grating-based dexterous finger tactile sensor using an FBG demodulator;

[0083] T2. Input the spectral signal acquired in step T1 into the FMEM demodulation model based on deep learning neural network. Demodulate the input spectral signal through the FMEM demodulation model based on deep learning neural network to obtain the demodulation result of multidimensional force information applied to the dexterous finger tactile sensor based on Bragg grating.

[0084] Specifically, in step T2 of embodiment 4, the input spectral signal is demodulated using an FMEM demodulation model based on a deep learning neural network to obtain the demodulated result of multidimensional force information applied to the dexterous finger tactile sensor based on a Bragg grating. This includes the following steps:

[0085] T201. The input layer receives the spectral signal acquired in step T1, i.e. the raw spectral data, and performs preliminary data standardization processing on the raw spectral data in the input layer to convert it into a data format suitable for neural network processing.

[0086] T202. In the Patch embedding layer, the spectral data obtained in step T201, which is suitable for neural network processing, is divided into multiple small regions (patches).

[0087] T203. In the position encoding layer, each small region patch obtained in step T202 is mapped to a high-dimensional feature space, and position information is added to each small region patch to preserve the sequence order relationship of the spectral data.

[0088] T204. The data output from the position encoding layer is processed through the Transformer encoder layer. First, the long-distance dependencies between different parts of the spectral data are captured by the multi-head attention mechanism. Then, the key feature patterns and correlations in the spectral data are extracted by the cascaded processing of multi-head self-attention and feedforward networks.

[0089] T205. The multidimensional feature representation output by the Transformer encoder layer is flattened into a one-dimensional vector through a flattening layer.

[0090] T206: Receives a one-dimensional vector output from the flattening layer through a linear layer, maps the high-level features extracted by the Transformer to the feature space related to the force signal, and outputs a linear transformation representation of appropriate dimension.

[0091] T207 receives the feature representations output by the linear layer through the multilayer perceptron layer, and further processes these features through nonlinear transformation to perform regression and classification tasks. Then, the processed features are converted into the required multidimensional force information (such as the magnitude, position, and direction of the force), thus realizing the complete demodulation mapping from spectral features to force information.

[0092] More specifically, in step T201, the original spectral data undergoes preliminary data standardization in the input layer to convert it into a data format suitable for neural network processing. This includes the following steps:

[0093] The first step is to perform data augmentation on the raw input spectral data in the input layer, including random jittering and scaling, to enhance the robustness of the model and prevent overfitting.

[0094] The second step involves preprocessing the enhanced data and treating it as a multivariate time series. It's important to note that the data from each Bragg grating (FBG) is propagated through a separate channel based on its wavelength reflectance measurement. Since the Bragg grating-based dexterous finger haptic sensor contains multiple FBGs, it generates multiple data channels. To meet the data format requirements suitable for neural network processing, each data channel can be considered a univariate time series. Thus, the raw spectral data input to the input layer from the Bragg grating-based dexterous finger haptic sensor (Fibertouch sensor) can be treated as a multivariate time series. Furthermore, while the multiple data channels of the Fibertouch sensor share the same Transformer architecture in the neural network, their forward propagation processes remain independent of each other.

[0095] More specifically, in step T202, the spectral data obtained in step T201, which is suitable for neural network processing, is divided into multiple small regions (patches) in the Patch embedding layer. The specific method is as follows:

[0096] For each data channel, the univariate time series is first instantiated to standardize the data distribution; then, the normalized series is divided into blocks with appropriate stride settings. Specifically, each input univariate spectral signal... First, the data is divided into blocks, which may overlap or not; then these blocks are joined together to generate a block sequence. ,in It is the number of blocks. , Indicates the length of the block. This represents stride, which is the non-overlapping area between two consecutive blocks.

[0097] More specifically, in step T203, each small region patch from step T202 is mapped to a high-dimensional feature space in the position encoding layer, and position information is added to each small region patch to preserve the sequential order of the spectral data; the specific method is as follows:

[0098] In the positional encoding layer, each block sequence output by the Patch embedding layer is first processed. Mapped to dimension The latent space is then processed through a trainable linear projection. And apply a learnable additive positional encoding. To preserve the time order of spectral blocks .

[0099] More specifically, in step T204, the data output from the position encoding layer is processed by the Transformer encoder layer, and the specific method is as follows:

[0100] First, each small region patch is fed into the Transformer encoder layer. Then, the multi-head attention sublayer of the Transformer encoder layer computes multiple sets of different attention weights in parallel, while simultaneously focusing on different positions and feature representations of the input sequence to capture richer contextual information. Next, the regularization sublayer of the Transformer encoder layer reduces the risk of model overfitting and enhances the model's generalization ability through techniques such as dropout. Finally, the feedforward network sublayer of the Transformer encoder layer performs nonlinear transformation on the features extracted by the multi-head attention sublayer to further improve the model's expressive power.

[0101] In this invention, the multi-head attention sublayer can effectively capture important spectral features and peak shift patterns in long-term optical signals because multiple attention heads work in parallel, i.e., each attention head learns various aspects of the wavelength-intensity relationship.

[0102] In this invention, in order to calculate the attention score, that is, to obtain the attention output. In multi-head attention, for each attention head, the input information is first linearly transformed to obtain the query matrix. Key matrix Sum matrix Then, by querying the matrix Key matrix Sum matrix The attention output of each attention head is calculated.

[0103] Among them, the query matrix Key matrix Value matrix The calculation formula is as follows:

[0104]

[0105] in, , , ;

[0106] The attention output of each attention head can be obtained using the following formula:

[0107] .

[0108] More specifically, in step T206, the linear layer receives the feature vector output by the flattening layer and maps it to the feature space related to the force signal. The specific method is to perform a linear transformation through a trainable weight matrix and bias vector to compress the high-dimensional feature space into a lower-dimensional force signal feature representation space.

[0109] More specifically, in step T207, the multilayer perceptron layer receives the feature representation output by the linear layer, realizing the final demodulation mapping from spectral features to force information. The specific method is as follows:

[0110] Three parallel branch network structures are designed, responsible for the force signal regression task (force magnitude) and two classification tasks (force location and direction), respectively. For the regression task, a three-layer fully connected network is used, with ReLU non-linear activation functions between each layer, and the last layer outputting the predicted force magnitude. For the classification task, a two-layer fully connected network is used, with the first layer using the ReLU activation function and the second layer using the Softmax function to predict the probability distribution of the force direction and location. Furthermore, during the training phase, a multi-task loss function is used to simultaneously optimize the regression loss (mean squared error) and the classification loss (cross-entropy), with weights of 1.0 and 2.0, respectively.

[0111] Specifically, in this invention, a patching method is chosen to learn local spectral features and reduce computational complexity in order to uncover the correlations between spectral signals. The input spectral signal is divided into multiple non-overlapping blocks, each containing frequency information within a certain range. This block-based processing allows the model to focus on capturing patterns and variations within specific frequency bands. By learning these local features, the model can identify the correlation structures and dependencies between different frequency bands, thus more accurately representing the complete spectral characteristics. Furthermore, compared to directly processing the entire high-dimensional spectral data, the patching method significantly reduces the number of parameters and computational load, improving training and inference efficiency and enabling the model to run efficiently even in resource-constrained environments.

[0112] In addition, to evaluate the performance of the Bragg grating-based dexterous finger touch sensor (Fibertouch sensor for short) designed in this invention, the applicant also conducted a series of force application experiments and designed a special experimental setup for this purpose, see [link to relevant documentation]. Figure 7As shown, this experimental setup consists of an FBG demodulator 001, a push-pull force gauge 002, a tension / compression testing platform 003, an XY-axis displacement platform 004, an inclination measuring device 005, and a large-angle tilting platform 006. The FBG demodulator 001 is located beside the tension / compression testing platform 003 and is used to generate laser excitation and acquire the reflection spectrum from the Bragg grating (FBG) in the Fibertouch sensor 008. The push-pull force gauge 002 is located on the upper part of the tension / compression testing platform 003 and is used to transmit the signal to the Fibertouch sensor. The sensor 007 applies force; the XY-axis displacement platform 004 is located at the lower part of the tensile / compression testing machine 003 and directly below the push-pull force gauge 002, and is used to enable the Fibertouch sensor 007 to move along the X and Y axes; the large-angle tilt platform 006 is located at the upper part of the XY-axis displacement platform 004 and directly below the push-pull force gauge 002, and is used to adjust the angle of the Fibertouch sensor 007; the tilt measuring device 005 is located at one end of the large-angle tilt platform 006, and is used to display the tilt angle of the large-angle tilt platform 006.

[0113] During testing, the Fibertouch sensor 007 was first mounted on the large-angle tilt platform 006 and connected to the FBG demodulator 001. Then, according to actual needs, the position of the Fibertouch sensor 007 was adjusted sequentially via the XY-axis displacement platform 004 and the angle of the Fibertouch sensor 007 was adjusted via the large-angle tilt platform 006. Subsequently, a force was applied to each position unit of the Fibertouch sensor 007 sequentially via the push-pull force gauge 002. During this process, the FBG demodulator 001 recorded in real time the reflection spectrum, Bragg wavelength shift, and corresponding force and position information generated by each position unit.

[0114] Finally, it should be noted that the above description is only an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for demodulating the spectral signal of a Bragg grating-based dexterous finger tactile sensor, characterized in that, The method is implemented using an FMEM demodulation model based on a deep learning neural network. This FMEM demodulation model includes an input layer, a patch embedding layer, a position encoding layer, a Transformer encoder layer, a flattening layer, a linear layer, and a multilayer perceptron layer. The input layer, serving as the data input entry point for the deep learning neural network, performs preliminary data standardization on the raw spectral data collected from the dexterous finger tactile sensor based on a Bragg grating, converting it into a data format suitable for neural network processing. The patch embedding layer divides the data output from the input layer into multiple small patches. The location encoding layer is used to map each small region patch to a high-dimensional feature space and add location information to each small region patch, so that the neural network can distinguish patches at different locations and preserve the sequence order relationship of the spectral data, ensuring that the neural network can identify location-related features in the spectral data. The Transformer encoder layer is used to capture long-distance dependencies between different parts of the spectral data through a multi-head attention mechanism and to extract key feature patterns and correlations in the spectral data through multi-head attention and feedforward network concatenation processing. The flattening layer is used to flatten the multidimensional feature representation output by the Transformer encoder layer into a one-dimensional vector; the linear layer is used to receive the one-dimensional vector output by the flattening layer, map the high-level features extracted by the Transformer to the feature space related to the force signal, and provide a linear transformation representation of appropriate dimension for subsequent processing; the multilayer perceptron layer is used to receive the feature representation output by the linear layer, further process these features through nonlinear transformation, perform regression and classification tasks, and convert the processed features into the required multidimensional force information, realizing a complete demodulation mapping from spectral features to force information; The method includes the following steps: T1. Acquire spectral signals under different positions, directions, and magnitudes of applied force on a Bragg grating-based dexterous finger tactile sensor using an FBG demodulator; T2. Input the spectral signal collected in step T1 into the FMEM demodulation model based on deep learning neural network. Demodulate the input spectral signal through the FMEM demodulation model based on deep learning neural network to obtain the demodulation result of multidimensional force information applied to the dexterous finger tactile sensor based on Bragg grating. Step T2 specifically includes the following steps: T201. The spectral signal acquired in step T1 is received through the input layer, i.e., the raw spectral data. The raw spectral data is then subjected to preliminary data standardization processing in the input layer to convert it into a data format suitable for neural network processing. T202. In the Patch embedding layer, the spectral data obtained in step T201, which is suitable for neural network processing, is divided into multiple small regions (patches). T203. In the position encoding layer, each small region patch obtained in step T202 is mapped to a high-dimensional feature space, and position information is added to each small region patch to preserve the sequence order relationship of the spectral data. T204. The spectral data output in step T203 is processed through the Transformer encoder layer. First, the long-distance dependencies between different parts of the spectral data are captured using a multi-head attention mechanism. Then, the key feature patterns and correlations in the spectral data are extracted through the concatenation of multi-head attention and feedforward networks. T205. The multidimensional feature representation output by the Transformer encoder layer is flattened into a one-dimensional vector through a flattening layer. T206: Receives a one-dimensional vector output from the flattening layer through a linear layer, maps the high-level features extracted by the Transformer to the feature space related to the force signal, and outputs a linear transformation representation of appropriate dimension. T207: The feature representations output by the linear layer are received by the multilayer perceptron layer, and these features are further processed by nonlinear transformation to perform regression and classification tasks. Then, the processed features are converted into the required multidimensional force information, thus realizing the complete demodulation mapping from spectral features to force information. The dexterous finger tactile sensor based on Bragg grating includes an elastic substrate (100), a first optical fiber (200), and a second optical fiber (300). The first optical fiber (200) is arranged horizontally through the elastic substrate (100) in a straight line, and the second optical fiber (300) is arranged horizontally through the elastic substrate (100) in a U-shape. The first optical fiber (200) and the second optical fiber (300) intersect vertically. Multiple first Bragg gratings (400) are arranged on the first optical fiber (200), and multiple second Bragg gratings (500) are arranged on the second optical fiber (300).

2. The method for demodulating the spectral signal of a Bragg grating-based dexterous finger tactile sensor according to claim 1, characterized in that, The Transformer encoder layer consists of a multi-head attention sublayer, a regularization sublayer, and a feedforward network sublayer; the multilayer perceptron layer consists of multiple fully connected layers and nonlinear activation functions.

3. The method for demodulating the spectral signal of a Bragg grating-based dexterous finger tactile sensor according to claim 1, characterized in that, The elastic substrate (100) has a length, width, and height of 30 mm, 20 mm, and 2 mm, respectively. The diameters of the first optical fiber (200) and the second optical fiber (300) are both 125 μm. The lengths of the first Bragg grating (400) and the second Bragg grating (500) are both 4 mm. The elastic substrate (100) is made of PDMS material. The surface of the elastic substrate (100) is systematically divided into 4×6=24 position units in a 4x6 grid, and the size of each position unit is 5×5 mm².

4. A method for fabricating the dexterous finger tactile sensor based on a Bragg grating as described in claim 1, characterized in that, Includes the following steps: S1. Embed the first optical fiber (200) and the second optical fiber (300) into the 3D printing mold (600) according to the designed length; S2. Add A and B to the stirring device (700) in a ratio of 13:1 and mix evenly to obtain a PDMS mixture; wherein, A is a polydimethylsiloxane prepolymer and B is a hydrogen-containing polysiloxane crosslinking agent; S3. Place the PDMS mixture obtained in step S2 into a vacuum chamber (800) to degas and remove bubbles, and obtain a uniform PDMS mixture without bubbles; S4. The uniform, bubble-free PDMS mixture obtained in step S3 is added to the 3D printing mold (600) in which the first optical fiber (200) and the second optical fiber (300) are embedded in step S1, and cured at room temperature (26°C) for 24 hours. After demolding, the dexterous finger touch sensor based on the Bragg grating is obtained.

5. A system applying the method for demodulating the spectral signal of a Bragg grating-based dexterous finger tactile sensor as described in claim 1, characterized in that: It includes an FBG demodulator, which internally embeds the FMEM demodulation model based on a deep learning neural network.

Images