An intelligent tactile perception system and method based on electrical impedance imaging and deep learning

Abstract

The application discloses an intelligent tactile perception system and method based on electrical impedance tomography and deep learning, and the system comprises a signal excitation and data acquisition module, a hydrogel sensing medium, a CLU-Net deep learning reconstruction module and a tactile information recognition module, and the method comprises the following steps: a signal excitation circuit excites the hydrogel sensing medium with constant alternating current; multi-channel signal excitation and multi-point voltage RMS value measurement are performed; an initial conductivity distribution is reconstructed through an EIT reconstruction algorithm GREIT; the CLU-Net deep learning reconstruction module is trained; the trained module is used to optimize the initial conductivity distribution; and a tactile pressure field image is reconstructed through conductivity distribution analysis. The application effectively solves the application problem of traditional EIT technology in flexible materials, improves the imaging accuracy and robustness, and is suitable for the fields of flexible electronic skin, intelligent artificial limb, robot skin and the like.

Description

Technical Field

[0001] This invention relates to an intelligent tactile sensing system and method based on electrical impedance imaging and deep learning, belonging to the field of flexible electronic skin technology. The system can be applied to the reconstruction and recognition of high-precision tactile pressure fields, and is particularly suitable for flexible media with complex deformation characteristics. Background Technology

[0002] Electrical impedance tomography (EIT) is a non-invasive imaging technique that uses an electrode array to excite and detect signals at multiple points in a conductive medium, thereby reconstructing the internal conductivity distribution of the medium. This technology has been widely used in medical diagnosis, industrial inspection, and intelligent sensing, primarily to monitor changes in the internal conductivity of a target object, thus reflecting its physical state and structural information. However, traditional EIT techniques are often limited by the limited number of electrodes and the relatively complex reconstruction model, and are susceptible to interference from environmental noise and signal distortion, making it difficult to further improve imaging accuracy and spatial resolution.

[0003] In recent years, with the rapid development of new intelligent sensing technologies such as electronic skin and flexible sensors, the market demand for large-area, highly flexible tactile sensing systems has continued to grow. Traditional EIT methods have revealed many application challenges when adapting to these new flexible devices, especially in the problem of reconstructing the conductivity of flexible and easily deformable materials, where the shortcomings of traditional algorithms are particularly prominent.

[0004] Currently, traditional EIT reconstruction mostly employs linear iterative algorithms or least squares methods. These algorithms have limited effectiveness when dealing with complex conductivity distributions, nonlinear medium characteristics, and dynamically changing scenarios. In flexible electronics and electronic skin applications, the material conductivity exhibits a nonlinear response to external deformation. Traditional EIT reconstruction is prone to significant reconstruction errors under complex deformations, and resolution is difficult to guarantee. Taking flexible and sensitive materials such as hydrogels as an example, their conductivity changes are not only related to pressure but also affected by the coupling effects of various mechanical forces such as tension and bending. Traditional algorithms struggle to accurately characterize these complex electromechanical coupling relationships.

[0005] In addition, the computational efficiency and real-time performance of existing EIT reconstruction algorithms remain key bottlenecks restricting their application. Although some studies have introduced physical constraint-based improvement strategies, such as Tikhonov regularization and total variational regularization, it is still difficult to completely suppress the effects of noise and reconstruction bias caused by complex media, resulting in significant shortcomings in real-time and high-resolution imaging requirements. Overall, traditional EIT technology still has considerable room for improvement in its overall performance for imaging tasks involving flexible and dynamically deformable materials.

[0006] A search revealed Chinese invention patent application CN117494523A, which discloses a multimodal flexible tactile sensor and tactile sensing method based on electrical impedance imaging. The sensor has several identical electrodes uniformly arranged around the inner wall of an Ecoflex rectangular groove. The Ecoflex rectangular groove is filled with a flexible conductive material, and its upper surface is covered with a rectangular insulating cover to form the multimodal flexible tactile sensor. The tactile sensing method utilizes a multimodal deep learning imaging network to train a conductivity image P containing touch shape and intensity information, along with its boundary measurement voltage data, to obtain a multimodal deep learning imaging model. This model is used to predict the conductivity reconstruction image and touch shape and intensity information for arbitrary boundary measurement voltages. This invention enables multifunctional detection of touch position, force intensity, and shape, thereby achieving simultaneous perception of tactile shape and intensity, enhancing the practicality of tactile interaction systems.

[0007] A search revealed Chinese invention patent application CN118443191A, which discloses a tactile sensor and tactile sensing method that enhances the sensitivity of the internal region. This primarily addresses the problem of reduced sensitivity in the internal region of existing impedance imaging tactile sensors, particularly in areas far from the electrodes. The sensor structure of this invention features a flat cylindrical flexible insulating container with radius R. An electrode numbered 0 and M identical electrodes arranged in a ring around the 0 electrode are positioned at the center of the cylinder's bottom surface. N identical electrodes, numbered 1 to N, are then uniformly arranged circumferentially on the cylindrical surface. The container is filled with a flexible conductive material, and its upper surface is covered with a thin film for encapsulation. The tactile sensing method involves alternating excitation modes for electrodes 1 to N, with the potential of electrode 0 set to zero. The voltages of the remaining N-2 circumferential electrodes and M bottom electrodes are collected to obtain a boundary voltage matrix. An impedance imaging algorithm is used to detect the touch force and touch position on the sensor surface. Summary of the Invention

[0008] This invention aims to address the shortcomings of traditional EIT technology in the application of flexible and deformable materials. It proposes an intelligent tactile sensing system and method based on electrical impedance imaging and deep learning. By incorporating deep learning algorithms, this invention can achieve high-resolution, low-noise, and highly robust imaging effects, further optimize the accuracy of pressure field reconstruction, and thus provide more accurate and real-time tactile information reconstruction services for flexible electronic skin and various intelligent sensing devices.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] An intelligent tactile sensing system based on electrical impedance imaging and deep learning includes a signal excitation and data acquisition module, a hydrogel sensing medium, a CLU-Net deep learning reconstruction module, and a tactile information recognition module. The hydrogel sensing medium is connected to the signal excitation and data acquisition module. The CLU-Net deep learning reconstruction module and the tactile information recognition module are integrated into a computer terminal. The signal excitation and data acquisition module is connected to the computer terminal.

[0011] The signal excitation and data acquisition module includes a multi-channel signal excitation circuit and a data acquisition system. The multi-channel signal excitation circuit is connected to the hydrogel sensing medium and is used to apply alternating current to the hydrogel sensing medium. The data acquisition system is connected to the hydrogel sensing medium and a computer terminal respectively and is used to collect the root mean square (RMS) voltage between each electrode on the surface of the hydrogel sensing medium in real time and transmit the collected data to the computer terminal synchronously.

[0012] The hydrogel sensing medium has piezoresistive effect and electrical conductivity. When mechanical stress is applied to it from the outside, the hydrogel sensing medium will change its conductivity accordingly.

[0013] The CLU-Net deep learning reconstruction module relies on convolutional neural networks (CNN), U-Net networks, and long short-term memory networks (LSTM) to perform refined processing on the collected EIT data.

[0014] The tactile information recognition module relies on the reconstructed conductivity distribution map to provide tactile feedback for different application scenarios.

[0015] A smart tactile sensing method based on electrical impedance imaging and deep learning includes the following steps:

[0016] S1. A multi-channel signal excitation circuit provides constant alternating current excitation to the hydrogel sensing medium.

[0017] S2. The data acquisition system collects multi-point measured voltage RMS values;

[0018] S3. Initial conductivity distribution reconstruction is performed using the GREIT reconstruction algorithm (EIT).

[0019] Training of S4 and CLU-Net deep learning reconstruction modules;

[0020] S5. Optimize the initial conductivity distribution using the trained modules;

[0021] S6. Tactile pressure field image reconstructed through conductivity distribution analysis.

[0022] As a preferred technical solution of the present invention, step S1 is specifically as follows:

[0023] The electrode array of the multi-channel signal excitation circuit is uniformly distributed on the surface of the hydrogel to achieve multi-point current excitation. The frequency and amplitude of the excitation signal are specifically optimized in conjunction with the conductivity characteristics of the hydrogel sensing medium. The low-frequency signal of the multi-channel signal excitation circuit interacts with the hydrogel through the electrode array. When the hydrogel sensing medium is subjected to mechanical stress, its conductivity will change accordingly, thereby causing the voltage signal between the electrodes to change synchronously. The correlation between the two is as follows:

[0024] (1);

[0025] in: It's voltage. It is electric current. It is the length of the current path. It is the cross-sectional area of ​​the current. It is electrical conductivity.

[0026] As a preferred technical solution of the present invention, step S2 is specifically as follows:

[0027] The RMS values ​​of the voltage at multiple points on the surface of the hydrogel sensing medium are acquired using a data acquisition system. For AC signals, the RMS value provides an equivalent electrical value to that of a DC signal, and its calculation formula is as follows:

[0028] (2);

[0029] in: It is the root mean square value of the voltage signal. It is in time Voltage signal at the location, It is the period of the signal;

[0030] For a periodic AC voltage signal, the RMS value simplifies to:

[0031] (3);

[0032] in, It is the peak voltage of the signal.

[0033] As a preferred embodiment of the present invention: the data acquisition system includes a sampling circuit, and the sampling signal accuracy of the sampling circuit can be characterized by the following formula:

[0034] (4);

[0035] in, It's the signal-to-noise ratio. It is signal power. It is noise power.

[0036] As a preferred technical solution of the present invention, step S3 is as follows:

[0037] The GREIT algorithm, based on the EIT model, solves the relationship between boundary voltage and internal conductivity using a mathematical model, generating a preliminary conductivity distribution map. By minimizing the following objective function, a more realistic conductivity distribution is calculated:

[0038] (5);

[0039] in, It is the spatial distribution vector of conductivity to be solved. express The mathematical mapping relationship between the boundary voltage and the boundary voltage. These are the measured voltage data. It is a regularization parameter used to find the most suitable conductivity distribution. This ensures that the calculated voltage is as close as possible to the actual measured voltage, and that the conductivity distribution satisfies physical constraints.

[0040] As a preferred technical solution of the present invention, step S4 is specifically as follows:

[0041] S41. Construct the training sample set as follows:

[0042] The preliminary conductivity distribution map and the corresponding target conductivity distribution map or tactile pressure field label map are used to form training sample pairs;

[0043] S42. Preprocess the training samples as follows:

[0044] The preliminary conductivity distribution map is organized into network input data and matched with the corresponding label map; in the dynamic tactile perception scenario, the preliminary conductivity distribution map at continuous time moments is composed into an image sequence input according to time; the input samples are normalized; at the same time, the output label map is constrained to make it consistent with the network output value range, thereby obtaining standardized training samples;

[0045] S43. Establish the CLU-Net deep learning reconstruction model, as detailed below:

[0046] A deep learning reconstruction model CLU-Net, consisting of a convolutional neural network (CNN), a long short-term memory network (LSTM), and a U-Net network, is constructed. The CNN is used to extract spatial features, the LSTM is used to model temporal features, and the U-Net network is used to fuse multi-scale features and output an optimized conductivity distribution map.

[0047] S44. Model training and parameter optimization, as detailed below:

[0048] The training samples are input into the CLU-Net deep learning reconstruction model, and the target image is used as the supervision signal. The composite loss function is constructed by the mean squared error loss, structural similarity loss and total variational loss to calculate the error. The model parameters are then iteratively updated using the backpropagation algorithm.

[0049] S45. Model Validation and Selection, as detailed below:

[0050] The training samples are divided into training set, validation set and test set according to a preset ratio; during the training process, the performance of the CLU-Net deep learning reconstruction model is evaluated using the validation set, and the parameters of the CLU-Net deep learning reconstruction model are tuned according to the validation results; when the validation error reaches the optimal value, the corresponding network parameters are saved as the final model.

[0051] As a preferred technical solution of the present invention, step S5 is as follows:

[0052] S51. Spatial feature extraction is performed on the initially reconstructed image using a convolutional neural network (CNN), as follows:

[0053] Convolutional Neural Networks (CNNs) receive initial reconstructions of conductivity maps through input layers, and then extract spatial features from shallow to deep layers through multiple convolutional layers, with the number of convolutional kernels increasing layer by layer.

[0054] Pooling layers are used after convolutional layers to reduce computation, shrink feature map size, and retain core information.

[0055] By stacking multiple sets of convolutional and pooling layers, local and global features of the image are extracted step by step, and finally a high-order feature map is obtained.

[0056] S52. The high-order feature map will be used as the input to the Long Short-Term Memory (LSTM) network, as follows:

[0057] The input to a Long Short-Term Memory (LSTM) network is time-series data. Therefore, the output of a Convolutional Neural Network (CNN) needs to be transformed into a form suitable for LSTM processing. Specifically, this involves converting the spatial feature map output by the CNN into a format suitable for LSTM processing. Flattened into a two-dimensional time series data matrix The conversion is as follows:

[0058] (6);

[0059] It is the height of the image. It is the width of the image. It is the number of channels. L The number of spatial regions is represented by D, which is equivalent to treating each pixel as a time step. D represents the number of features per time step, which is equivalent to the number of channels.

[0060] The S53 and U-Net networks integrate the spatial feature maps extracted by the convolutional neural network (CNN) with the time series data matrix extracted by the LSTM to complete image restoration and generate the final conductivity distribution map.

[0061] As a preferred technical solution of the present invention, step S6 is specifically as follows:

[0062] The tactile information recognition module identifies and analyzes the reconstructed tactile pressure field image, accurately interprets the variation law of the conductivity of the hydrogel sensing medium, and feeds back the analyzed tactile information to the terminal application system.

[0063] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0064] 1. High-precision tactile information reconstruction: This system deeply integrates deep learning algorithms with EIT technology, effectively breaking through the limitations of traditional EIT methods in flexible media applications, and achieving high-resolution, refined tactile pressure field reconstruction. Relying on the CLU-Net deep learning module, and optimizing the entire EIT data processing workflow with CNN, U-Net, and LSTM networks, it significantly improves imaging resolution and reconstruction accuracy, ensuring stable adaptation even to the tactile sensing needs of complex flexible materials under dynamic loading conditions.

[0065] 2. Rapid Dynamic Response and High Robustness: This invention combines the physical constraints of EIT technology with the data analysis advantages of deep learning, enabling the system to quickly respond to various complex pressure changes and dynamic operating conditions. Deep learning algorithms can extract nonlinear mapping relationships from boundary voltage data, and, combined with time-series modeling, accurately capture the dynamic changes in tactile signals, effectively improving the robustness of the system in practical applications. Simultaneously, the system can flexibly adapt to changes in the conductivity of materials under different stress states, ensuring stable progress and rapid response in the imaging process.

[0066] 3. Excellent Adaptability and Stability: Thanks to the optimized capabilities of the deep learning module, the system can accurately reconstruct the conductivity distribution under different tactile perception scenarios, demonstrating excellent scene adaptability and operational stability. Deep learning methods can efficiently handle the inherent nonlinear and time-varying characteristics of hydrogel materials, ensuring the accuracy and consistency of imaging results under various loading conditions. Furthermore, the inclusion of physical consistency constraints further enhances system stability, effectively avoiding imaging errors caused by noise interference and missing data.

[0067] 4. Improved Tactile Recognition Accuracy: The optimization of the EIT data reconstruction process by deep learning networks not only improves image resolution but also enables the rapid generation of high-quality tactile images in practical applications. By integrating multiple neural network architectures such as CNN, LSTM, and U-Net, the system can efficiently extract both spatial and temporal features of the tactile field, helping devices such as flexible electronic skin, intelligent prostheses, and robotic skin achieve accurate tactile perception.

[0068] 5. Reduce material waste and improve efficiency: By leveraging deep learning to optimize EIT data processing, the system can achieve accurate conductivity reconstruction under complex deformation conditions, avoiding reconstruction errors that traditional algorithms are prone to in complex scenarios. Simultaneously, it reduces resource waste caused by low resolution and noise interference in data acquisition and image reconstruction, effectively improving the overall efficiency of tactile information reconstruction.

[0069] 6. Wide Applicability: This system has broad application prospects in fields such as electronic skin, intelligent prostheses, and robotic skin. Through the cross-disciplinary integration of EIT and deep learning, it can provide precise tactile feedback for flexible sensors, wearable devices, and other applications, adapting to diverse application scenarios such as intelligent sensing, medical monitoring, virtual reality, and augmented reality.

[0070] 7. High degree of automation and scalability: This system achieves a high degree of automation in tactile information reconstruction through the collaborative operation of the deep learning module and the EIT algorithm. It can autonomously process multi-channel EIT data and complete rapid reconstruction based on the optimized deep learning algorithm, significantly reducing manual intervention. In the event of data anomalies or system failures, the system will automatically trigger alarms and shut down, ensuring comprehensive operational safety and reliability. Attached Figure Description

[0071] Figure 1 This is a schematic diagram of the overall hardware of an intelligent tactile sensing system based on electrical impedance imaging and deep learning.

[0072] Figure 2 This is a circuit diagram of the multi-channel signal excitation circuit in this invention;

[0073] Figure 3 This is a structural diagram of the CLU-Net deep learning reconstruction module in this invention;

[0074] Figure 4 This is a flowchart of an intelligent tactile perception method based on electrical impedance imaging and deep learning.

[0075] List of reference numerals in the attached diagram:

[0076] 1. Signal excitation and data acquisition module; 2. Hydrogel sensing medium; 3. CLU-Net deep learning reconstruction module; 4. Tactile information recognition module. Detailed Implementation

[0077] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0078] like Figure 1 As shown, this invention proposes an intelligent tactile sensing system based on electrical impedance imaging and deep learning, comprising a signal excitation and data acquisition module 1, a hydrogel sensing medium 2, a CLU-Net deep learning reconstruction module 3, and a tactile information recognition module 4. The hydrogel sensing medium 2 is connected to the signal excitation and data acquisition module 1. The CLU-Net deep learning reconstruction module 3 and the tactile information recognition module 4 are integrated into a computer terminal. The signal excitation and data acquisition module 1 is connected to the computer terminal.

[0079] The signal excitation and data acquisition module 1 includes a multi-channel signal excitation circuit and a data acquisition system, wherein:

[0080] The multi-channel signal excitation circuit is connected to the hydrogel sensing medium 2 and is used to apply alternating current to the hydrogel sensing medium 2. The frequency and amplitude of the excitation signal used by the multi-channel signal excitation circuit are optimized in combination with the conductivity characteristics of the hydrogel material itself, so as to reduce the data distortion problem caused by excessively high signal frequency.

[0081] The data acquisition system is connected to the hydrogel sensing medium 2 and the computer terminal respectively, and is used to collect the root mean square (RMS) voltage between each electrode on the surface of the hydrogel sensing medium 2 in real time, and transmit the collected data to the computer terminal synchronously.

[0082] The hydrogel sensing medium 2 exhibits both piezoresistive effect and electrical conductivity. When external mechanical stress is applied to it, the hydrogel sensing medium 2 undergoes a corresponding change in conductivity. The hydrogel material used in the hydrogel sensing medium 2 combines piezoresistive effect and excellent electrical conductivity, enabling it to accurately respond to various mechanical stresses such as external pressure, tension, and deformation, thereby causing corresponding changes in conductivity. Furthermore, the material properties of this hydrogel material are highly compatible with human skin and flexible electronic devices, possessing excellent biocompatibility and flexibility, making it suitable for various complex deformation application scenarios.

[0083] The CLU-Net deep learning reconstruction module 3 relies on convolutional neural networks (CNN), U-Net networks, and long short-term memory networks (LSTM) to perform refined processing on the acquired EIT data. The CLU-Net deep learning reconstruction module 3 uses deep learning algorithms to optimize the conductivity distribution reconstruction process, solving the limitations of traditional EIT methods in flexible media, improving imaging accuracy and resolution. The CLU-Net deep learning reconstruction module 3 has strong robustness and fast dynamic response capabilities, and can adapt to the conductivity changes of materials under different stress states, ensuring the stability of the imaging process.

[0084] The tactile information recognition module 4 relies on the reconstructed conductivity distribution map to complete the recognition and application of tactile information. The tactile data obtained can be applied to scenarios such as intelligent prostheses, robotic skin, and electronic skin, providing high-precision tactile feedback support for these devices.

[0085] like Figure 4 As shown, to realize the intelligent electrical impedance imaging system of the present invention, the present invention also proposes an intelligent tactile perception method based on electrical impedance imaging and deep learning, including the following steps:

[0086] S1. A multi-channel signal excitation circuit provides constant alternating current excitation to the hydrogel sensing medium 2.

[0087] S2. The data acquisition system collects multi-point measured voltage RMS values;

[0088] S3. Initial conductivity distribution reconstruction is performed using the GREIT reconstruction algorithm (EIT).

[0089] S4, CLU-Net deep learning reconstruction module 3 training;

[0090] S5. Optimize the initial conductivity distribution using the trained modules;

[0091] S6. Tactile pressure field image reconstructed through conductivity distribution analysis.

[0092] The above method is used to implement the system as a whole, proceeding sequentially from signal excitation, data acquisition, deep learning reconstruction to tactile information recognition.

[0093] In the initial stage of system implementation, a constant alternating current is applied to the hydrogel sensing medium 2 using a multi-channel signal excitation circuit. This multi-channel signal excitation circuit consists of a single-ended to differential circuit, a voltage follower, and a Howland circuit, as shown in the specific structure below. Figure 2As shown, the electrode array of the multi-channel signal excitation circuit is uniformly distributed on the surface of the hydrogel sensing medium 2 to achieve multi-point current excitation and voltage signal acquisition. The frequency and amplitude of the excitation signal are specifically optimized in conjunction with the conductivity characteristics of the hydrogel material to ensure stable and distortion-free signal transmission. This type of low-frequency signal interacts with the hydrogel through the electrode array. When the hydrogel is subjected to mechanical stresses such as external pressure, tension, and deformation, its conductivity will change accordingly, thereby causing the voltage signal between the electrodes to change synchronously. The relationship between the two is shown below:

[0094] (1);

[0095] in: It's voltage. It is electric current. It is the length of the current path. It is the cross-sectional area of ​​the current. It is electrical conductivity. The value remains unchanged, therefore The changes mainly come from Electrical conductivity.

[0096] The system employs a multi-channel signal excitation mode. Through the data acquisition system, it can acquire RMS values ​​of the voltage at multiple points on the hydrogel surface. For AC signals, the RMS value provides an equivalent electrical value to the DC signal, facilitating the acquisition of more comprehensive voltage data. The calculation formula is as follows:

[0097] (2);

[0098] in: It is the root mean square value of the voltage signal. It is in time Voltage signal at the location, This is the period time of the signal. For a periodic AC voltage signal such as a sine wave, the RMS value can be simplified to:

[0099] (3);

[0100] in, It is the peak voltage of the signal.

[0101] After signal excitation, the acquired voltage RMS data is obtained in real time through the data acquisition system and then transmitted to the system's deep learning unit via the ADC module. The accuracy and stability of the data acquisition system directly determine the overall imaging effect. Therefore, a dedicated sampling circuit is designed to ensure accurate voltage signal measurement. Simultaneously, a high signal-to-noise ratio design effectively resists external environmental noise interference. The sampling signal accuracy can be characterized by the following formula:

[0102] (4);

[0103] in, It's the signal-to-noise ratio. It is signal power. It is noise power.

[0104] After data acquisition, the traditional EIT reconstruction algorithm GREIT is first used to perform a preliminary reconstruction of the conductivity distribution of the voltage data. This step relies on a mathematical model to solve the correlation between boundary voltage and internal conductivity, generating a rough conductivity distribution map. By minimizing the following objective function, a more realistic conductivity distribution is calculated:

[0105] (5);

[0106] in, It is the spatial distribution vector of conductivity to be solved. express The mathematical mapping relationship between the boundary voltage and the boundary voltage. These are the measured voltage data. It is a regularization parameter used to find the most suitable conductivity distribution. This approach aims to make the calculated voltage as close as possible to the actual measured voltage, and to ensure that the conductivity distribution meets physical constraints. However, due to factors such as the nonlinear characteristics of hydrogels and external noise, the initially reconstructed conductivity images often suffer from blurring and errors. Traditional algorithms struggle to achieve accurate restoration. Therefore, subsequent deep learning optimization steps are needed to refine the initial results, compensate for the shortcomings of traditional EIT methods, and further improve reconstruction accuracy and image quality.

[0107] like Figure 3 As shown, CLU-Net deep learning reconstruction module 3 plays a core optimization role. The training of CLU-Net deep learning reconstruction module 3 is as follows:

[0108] The training sample set is constructed as follows:

[0109] The preliminary conductivity distribution map and the corresponding target conductivity distribution map or tactile pressure field label map are used to form training sample pairs;

[0110] The training samples are preprocessed as follows:

[0111] The preliminary conductivity distribution map is organized into network input data and matched with the corresponding label map; in the dynamic tactile perception scenario, the preliminary conductivity distribution map at continuous time moments is composed into an image sequence input according to time; the input samples are normalized; at the same time, the output label map is constrained to make it consistent with the network output value range, thereby obtaining standardized training samples;

[0112] The CLU-Net deep learning reconstruction model is established as follows:

[0113] A deep learning reconstruction model CLU-Net, consisting of a convolutional neural network (CNN), a long short-term memory network (LSTM), and a U-Net network, is constructed. The CNN is used to extract spatial features, the LSTM is used to model temporal features, and the U-Net network is used to fuse multi-scale features and output an optimized conductivity distribution map.

[0114] Model training and parameter optimization are detailed below:

[0115] The training samples are input into the CLU-Net deep learning reconstruction model, and the target image is used as the supervision signal. The composite loss function is constructed by the mean squared error loss, structural similarity loss and total variational loss to calculate the error. The model parameters are then iteratively updated using the backpropagation algorithm.

[0116] Model validation and selection are detailed below:

[0117] The training samples are divided into training set, validation set and test set according to a preset ratio; during the training process, the performance of the CLU-Net deep learning reconstruction model is evaluated using the validation set, and the parameters of the CLU-Net deep learning reconstruction model are tuned according to the validation results; when the validation error reaches the optimal value, the corresponding network parameters are saved as the final model.

[0118] The initial conductivity distribution is optimized using the trained CLU-Net deep learning reconstruction module 3, as follows:

[0119] First, a convolutional neural network (CNN) is used to extract spatial features from the initially reconstructed image. CNNs can autonomously learn and capture local features of conductivity changes, and their effectiveness is particularly pronounced in capturing conductivity distribution changes in complex deformation regions. The CNN receives the initially reconstructed conductivity map through an input layer, and then extracts spatial features from shallow to deep layers through multiple convolutional layers. The number of convolutional kernels increases layer by layer, from an initial 32 to 64, 128, and 256. Pooling layers are then used after the convolutional layers to reduce computational cost, shrink the feature map size, and retain core, effective information. Through the stacking of multiple convolutional and pooling layers, local and global features of the image are gradually extracted, ultimately yielding a high-order feature map.

[0120] Next, the high-order feature maps extracted from the convolutional neural network (CNN) layers will be used as input to the long short-term memory (LSTM) network. Since the input to the LSTM network is time-series data, the output of the CNN needs to be transformed into a form suitable for LSTM processing. Specifically, this involves converting the spatial feature maps output by the CNN into a format suitable for LSTM processing. Flattened into a two-dimensional time series data matrix The conversion is as follows:

[0121] (6); L The number of spatial regions

[0122] It is the height of the image. It is the width of the image. It is the number of channels. L The number of spatial regions is represented by D, which is equivalent to treating each pixel as a time step. D represents the number of features at each time step, which is equivalent to the number of channels.

[0123] Long Short-Term Memory (LSTM) networks can model the changes in conductivity during the dynamic deformation of hydrogels, accurately capture the temporal characteristics of tactile signals, especially the temporal patterns of pressure and deformation propagation, and finally output feature representations containing temporal information, providing data support for subsequent image optimization.

[0124] Subsequently, the U-Net network integrates the spatial feature map extracted by the convolutional neural network (CNN) with the time series data matrix extracted by the long short-term memory network (LSTM) to complete image restoration and generate a high-resolution conductivity distribution map.

[0125] The U-Net network mainly consists of three parts: an encoder, a decoder, and a bottleneck layer. The encoder, relying on multiple convolutional and pooling layers, extracts multi-level features from the image, gradually reducing the image size and increasing the depth of the feature map. The decoder, using convolutional and unpooling layers, gradually restores the image resolution, and with skip connections, retains image edge information, achieving excellent detail restoration in areas of abrupt changes in conductivity, ensuring the clarity and accuracy of the final conductivity distribution map. The bottleneck layer between the encoder and decoder compresses and refines core features, retaining key image information. Through this deep learning optimization process, the system effectively overcomes the limitations of traditional EIT methods in reconstruction accuracy and resolution, outputting more accurate tactile pressure field imaging results.

[0126] After the conductivity distribution is reconstructed, the tactile information recognition module 4 identifies and analyzes the reconstructed tactile pressure field image, accurately interprets the variation pattern of hydrogel conductivity, and feeds back the analyzed tactile information to terminal application systems such as intelligent prostheses, robotic skin, and electronic skin. In practical applications, the system can control the operation of robots and prostheses based on the tactile pressure field distribution data, enabling them to accurately perceive and adapt to various external tactile stimuli, achieving more refined and intelligent tactile feedback.

[0127] Through the entire implementation process, this invention not only effectively solves the pain points of traditional EIT technology in the field of flexible materials, but also optimizes the conductivity distribution reconstruction process by relying on deep learning methods, creating a high-resolution and highly robust tactile information reconstruction system with broad application potential in various scenarios.

[0128] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A smart tactile sensing method based on electrical impedance imaging and deep learning, characterized in that, Includes the following steps: S1. The multi-channel signal excitation circuit excites the hydrogel sensing medium (2) with a constant alternating current. S2. The data acquisition system collects multi-point measured voltage RMS values; S3. Initial conductivity distribution reconstruction is performed using the GREIT reconstruction algorithm (EIT). Step S3 is as follows: The GREIT algorithm, based on the EIT model, solves the relationship between boundary voltage and internal conductivity using a mathematical model, generating a preliminary conductivity distribution map. By minimizing the following objective function, a more realistic conductivity distribution is calculated: （5）； in, It is the spatial distribution vector of conductivity to be solved. express The mathematical mapping relationship between the boundary voltage and the boundary voltage. These are the measured voltage data. It is a regularization parameter used to find the most suitable conductivity distribution. Furthermore, the conductivity distribution satisfies physical constraints; S4, CLU-Net deep learning reconstruction module (3) training; S5. Optimize the initial conductivity distribution using the trained modules; S6. Tactile pressure field image reconstructed through conductivity distribution analysis.

2. The intelligent tactile sensing method based on electrical impedance imaging and deep learning according to claim 1, characterized in that, Step S1 is as follows: The electrode array of the multi-channel signal excitation circuit is uniformly arranged on the surface of the hydrogel to achieve multi-point current excitation. The frequency and amplitude of the excitation signal are specifically optimized in combination with the conductivity characteristics of the hydrogel sensing medium (2). The low-frequency signal of the multi-channel signal excitation circuit interacts with the hydrogel through the electrode array. When the hydrogel sensing medium (2) is subjected to mechanical stress, its conductivity will change accordingly, thereby driving the voltage signal between the electrodes to change synchronously. The relationship between the two is as follows: （1）； in: It's voltage. It is electric current. It is the length of the current path. It is the cross-sectional area of ​​the current. It is electrical conductivity.

3. The intelligent tactile sensing method based on electrical impedance imaging and deep learning according to claim 1, characterized in that, Step S2 is as follows: The RMS value of the multi-point voltage on the surface of the hydrogel sensing medium (2) is acquired through a data acquisition system. For AC signals, the RMS value can provide an equivalent electrical value to the DC signal. The calculation formula is as follows: （2）； in: It is the root mean square value of the voltage signal. It is in time Voltage signal at the location, It is the period of the signal; For a periodic AC voltage signal, the RMS value can be simplified as: （3）； in, It is the peak voltage of the signal.

4. The intelligent tactile sensing method based on electrical impedance imaging and deep learning according to claim 3, characterized in that, The data acquisition system includes a sampling circuit, and the sampling signal accuracy of the sampling circuit can be characterized by the following formula: （4） in, It's the signal-to-noise ratio. It is signal power. It is noise power.

5. The intelligent tactile sensing method based on electrical impedance imaging and deep learning according to claim 4, characterized in that, Step S4 is as follows: S41. Construct the training sample set as follows: The preliminary conductivity distribution map and the corresponding target conductivity distribution map or tactile pressure field label map are used to form training sample pairs; S42. Preprocess the training samples as follows: The preliminary conductivity distribution map is organized into network input data and matched with the corresponding label map; in the dynamic tactile perception scenario, the preliminary conductivity distribution map at continuous time moments is composed into an image sequence input according to time; the input samples are normalized; at the same time, the output label map is constrained to make it consistent with the network output value range, thereby obtaining standardized training samples; S43. Establish the CLU-Net deep learning reconstruction model, as detailed below: A deep learning reconstruction model CLU-Net, consisting of a convolutional neural network (CNN), a long short-term memory network (LSTM), and a U-Net network, is constructed. The CNN is used to extract spatial features, the LSTM is used to model temporal features, and the U-Net network is used to fuse multi-scale features and output an optimized conductivity distribution map. S44. Model training and parameter optimization, as detailed below: The training samples are input into the CLU-Net deep learning reconstruction model, and the target image is used as the supervision signal. The composite loss function is constructed by the mean squared error loss, structural similarity loss and total variational loss to calculate the error. The model parameters are then iteratively updated using the backpropagation algorithm. S45. Model Validation and Selection, as detailed below: The training samples are divided into training set, validation set and test set according to a preset ratio; during the training process, the performance of the CLU-Net deep learning reconstruction model is evaluated using the validation set, and the parameters of the CLU-Net deep learning reconstruction model are tuned according to the validation results; when the validation error reaches the optimal value, the corresponding network parameters are saved as the final model.

6. The intelligent tactile sensing method based on electrical impedance imaging and deep learning according to claim 1, characterized in that, Step S5 is as follows: S51. Spatial feature extraction is performed on the initially reconstructed image using a convolutional neural network (CNN), as follows: Convolutional Neural Networks (CNNs) receive initial reconstructions of conductivity maps through input layers, and then extract spatial features from shallow to deep layers through multiple convolutional layers, with the number of convolutional kernels increasing layer by layer. Pooling layers are used after convolutional layers to reduce computation, shrink feature map size, and retain core information. By stacking multiple sets of convolutional and pooling layers, local and global features of the image are extracted step by step, and finally a high-order feature map is obtained. S52. The high-order feature map will be used as the input to the Long Short-Term Memory (LSTM) network, as follows: The input to a Long Short-Term Memory (LSTM) network is time-series data. Therefore, the output of a Convolutional Neural Network (CNN) needs to be transformed into a form suitable for LSTM processing. Specifically, this involves converting the spatial feature map output by the CNN into a format suitable for LSTM processing. Flattened into a two-dimensional time series data matrix The conversion is as follows: （6）； It is the height of the image. It is the width of the image. It is the number of channels. L The number of spatial regions is represented by D, which is equivalent to treating each pixel as a time step. D represents the number of features per time step, which is equivalent to the number of channels. The S53 and U-Net networks integrate the spatial feature maps extracted by the convolutional neural network (CNN) with the time series data matrix extracted by the LSTM to complete image restoration and generate the final conductivity distribution map.

7. The intelligent tactile sensing method based on electrical impedance imaging and deep learning according to claim 1, characterized in that, Step S6 is as follows: The tactile information recognition module (4) is used to identify and analyze the reconstructed tactile pressure field image, interpret the change law of the conductivity of the hydrogel sensing medium (2), and feed back the analyzed tactile information to the terminal application system.

8. A system based on the intelligent tactile sensing method of electrical impedance imaging and deep learning according to any one of claims 1-7, characterized in that, The system includes a signal excitation and data acquisition module (1), a hydrogel sensing medium (2), a CLU-Net deep learning reconstruction module (3), and a tactile information recognition module (4). The hydrogel sensing medium (2) is connected to the signal excitation and data acquisition module (1). The CLU-Net deep learning reconstruction module (3) and the tactile information recognition module (4) are integrated into a computer terminal. The signal excitation and data acquisition module (1) is connected to the computer terminal. The signal excitation and data acquisition module (1) includes a multi-channel signal excitation circuit and a data acquisition system. The multi-channel signal excitation circuit is connected to the hydrogel sensing medium (2) and is used to apply alternating current to the hydrogel sensing medium (2). The data acquisition system is connected to the hydrogel sensing medium (2) and a computer terminal respectively and is used to collect the root mean square (RMS) voltage between each electrode on the surface of the hydrogel sensing medium (2) in real time and transmit the collected data to the computer terminal synchronously. The hydrogel sensing medium (2) has piezoresistive effect and conductivity. When mechanical stress is applied to it from the outside, the hydrogel sensing medium (2) will undergo a corresponding change in conductivity. The CLU-Net deep learning reconstruction module (3) relies on the convolutional neural network CNN, U-Net network and long short-term memory network LSTM to carry out refined processing of the collected EIT data; The tactile information recognition module (4) relies on the reconstructed conductivity distribution map to provide tactile feedback for different application scenarios.

Images