Analysis method of humidity distribution in humidity sensing plane and humidity-sensitive electronic skin

Abstract

The present invention relates to the field of electrical impedance imaging, and is an analysis method for humidity distribution in a humidity sensing plane and a humidity-sensitive electronic skin. The method comprises: obtaining a data set and performing preprocessing; constructing a humidity distribution imaging model, comprising an input layer, a pre-reconstruction module, an encoder module, a decoder module, a contour constraint branch, and an output layer; using the data set to train the humidity distribution imaging model to form a humidity distribution reconstruction model having two outputs, a humidity reconstruction image and a humidity reconstruction contour image; the input layer inputs boundary voltage data of the humidity sensing plane; the pre-reconstruction module generates an initial humidity reconstruction result, and the encoder module extracts humidity-related multi-level multi-scale features based on the pre-reconstruction module; the decoder module performs decoding according to the multi-level multi-scale features extracted by the encoder module; and the contour constraint branch is used to obtain a humidity reconstruction image with a clear outline. The present invention can focus on the reconstruction in the target area, thereby improving the effect of humidity reconstruction.

Description

Technical Field

[0001] The invention relates to the field of electrical impedance imaging, and in particular to an analysis method for humidity distribution in a humidity sensing plane and a humidity-sensitive electronic skin. Background Art

[0002] Electronic skin is a new type of electronic device that simulates the human skin's ability to sense external stimuli (pressure, temperature, humidity) through the integration and feedback of electrical signals. In the past few decades, electronic skin has attracted global attention due to its broad application prospects in intelligent robots, health monitoring, wearable devices, and human-computer interaction. Among them, humidity-sensitive electronic skin can obtain humidity information in the environment and can be used for respiratory monitoring, wound healing assessment, and the production of non-contact multi-level switches and flexible non-contact control screens, etc., which has important research significance and application value.

[0003] Humidity-sensitive electronic skin generally consists of three parts: a flexible humidity sensor array, a data acquisition unit, and an information processing unit. The flexible humidity sensor array can be divided into two working mechanisms: resistive and capacitive, which include m rows and n columns of flexible humidity sensor units. However, the layout of m rows and n columns of flexible humidity sensor units means that there must be criss-crossing electrical connections inside the humidity-sensitive electronic skin, which increases the difficulty of preparing humidity-sensitive electronic skin and limits the detection range of electronic skin. Secondly, the dense internal connections may cause crosstalk, which has a negative impact on the humidity sensing function. In addition, this also limits the sensing range of the humidity-sensitive electronic skin to m×n positions, that is, the spatial resolution of humidity perception is low. In summary, crosstalk and low spatial resolution are major challenges faced in the development of humidity-sensitive electronic skin.

[0004] Electrical impedance imaging is a technology that uses the electrical impedance characteristics of materials to reproduce the conductivity distribution inside an object in the form of an image. It has the advantages of simple operation, non-destructive testing, and low instrument cost. Therefore, electrical impedance technology has been widely used in medical functional imaging. However, due to the sensitive initial value of traditional electrical impedance imaging algorithms, directly using it on electronic skin will lead to problems such as poor reconstruction accuracy, poor repeatability, and unclear boundaries, which greatly affects the imaging effect. Summary of the invention

[0005] In view of this, the first object of the present invention is to provide an analysis method for humidity distribution in a humidity sensing plane, which reduces the adverse effects of crosstalk on humidity sensing accuracy by introducing contour constraints into the electrical impedance imaging neural network and using different loss functions, improves the spatial resolution of imaging, and expands the humidity sensing space from the existing m×n point type to the entire two-dimensional plane. And achieve a good mapping of the boundary voltage to the humidity distribution, and better reproduce the humidity distribution in the humidity sensing plane, including position, range, contour, etc.

[0006] Based on the same inventive concept, a second object of the present invention is to provide a moisture-sensitive electronic skin.

[0007] The first object of the present invention can be achieved by adopting the following technical solutions:

[0008] A method for analyzing humidity distribution in a humidity sensing plane, characterized in that it comprises the following steps:

[0009] Get the data set and preprocess it;

[0010] Constructing a humidity distribution imaging model, wherein the humidity distribution imaging model includes an input layer, a pre-reconstruction module, an encoder module, a decoder module, a contour constraint branch, and an output layer;

[0011] The humidity distribution imaging model is trained using the data set to form a humidity distribution reconstruction model with two outputs: humidity reconstruction image and humidity reconstruction contour image;

[0012] The input layer is used to input the boundary voltage data of the humidity sensing plane, and the output layer outputs the humidity reconstructed image and the corresponding contour image; the pre-reconstruction module is used to generate the initial humidity reconstruction result, and the encoder module is used to further extract humidity-related multi-level and multi-scale features based on the pre-reconstruction module; the decoder module decodes according to the multi-level and multi-scale features extracted by the encoder module; the contour constraint branch is used to obtain a humidity reconstructed image with clear contours.

[0013] Furthermore, in the humidity distribution imaging model:

[0014] The pre-reconstruction module includes a fully connected layer and a simple autoencoder. The input boundary voltage data is resized in the fully connected layer and then input into the simple autoencoder for processing.

[0015] The encoder module includes a convolutional layer, a batch normalization layer, 13 reverse residual blocks, and a pyramid pooling block; the 1st, 3rd, 6th, and 13th reverse residual blocks output 4 multi-level features backward respectively; the pyramid pooling block contains five branches, which are used to extract local features by dilated convolution and global features by global average pooling, and then multi-scale feature fusion is performed through the convolution layer to output global features;

[0016] A decoder module is used to decode the multi-level features and global features obtained by the encoder module to obtain a humidity reconstructed image with clear contours;

[0017] The contour constraint branch is used to fuse the shallow features and the deep features obtained by interpolating the global features to obtain the humidity reconstructed contour image.

[0018] Furthermore, the last convolutional layer of the decoder branch and the contour constraint branch uses the sigmoid function as the activation function, and the other convolutional layers use ReLU as the activation function.

[0019] Furthermore, the data set is used to train the humidity distribution imaging model, and the reconstruction error is used as the loss function for the humidity reconstructed image output, and the focal loss is used as the loss function for the humidity reconstructed contour image.

[0020] Furthermore, the reconstruction error is calculated as follows:

[0021] Different weights are assigned to the target and background of the real humidity distribution image in the data set and the humidity reconstructed image output by the humidity distribution reconstruction model, so that the target center area obtains the highest weight, the transition area between the target and the background is second, and the background area obtains the lowest weight;

[0022] The mean square error between the real humidity distribution image and the humidity reconstructed image after weight assignment is calculated as the reconstruction error.

[0023] Furthermore, the reconstruction error is calculated as follows:

[0024]

[0025] The filter is a matrix of all 1s with a size of (3, 3, 1, 1), y_true is the true humidity distribution image, y_pred is the humidity reconstructed image, and N is the number of batch training;

[0026] Furthermore, the focal loss is calculated as follows:

[0027] Loss2=-αy(1-p) γ log(p)-(1-α)(1-y)p γ log(1-p)

[0028] Where y is a binary label, y=0 means it does not belong to the contour area, y=1 means it belongs to the contour area, p is the model prediction value, α and γ are constant parameters, specifically, α is 0.2, and γ is 0.4.

[0029] Furthermore, the total loss is calculated as follows:

[0030] Loss = Loss1 + β * Loss2

[0031] Where β is a constant parameter. Specifically, the value of β is 0.36.

[0032] Furthermore, the obtaining of the data set and preprocessing thereof include:

[0033] Establish a finite element model, set the conductivity value of each grid of the finite element model according to the moisture resistance effect curve, simulate the experimental conditions to generate a large-scale voltage data sample and the corresponding humidity distribution map, and obtain a simulation data set;

[0034] The boundary voltage of the flexible humidity sensing unit in the real environment is collected, and the corresponding humidity distribution map is produced to obtain the experimental data set.

[0035] Furthermore, the use of the data set to train the humidity distribution imaging model includes:

[0036] Use simulated data sets to train the humidity distribution imaging model. First, use large-scale simulated data samples to train the algorithm model, so that the algorithm model can learn the general mapping relationship from boundary voltage to internal conductivity, and obtain the initial weights of the pre-reconstruction module, encoder module, decoder module, and contour constraint branch.

[0037] The weights of the pre-reconstruction module and the encoder module are frozen, and the special mapping relationship that is helpful for the main task of humidity reconstruction is learned through small-scale experimental data. The decoder module and the contour constraint branch are further trained using the experimental dataset, and the weights of the decoder module and the contour constraint branch are updated to further improve the humidity reconstruction accuracy.

[0038] The second object of the present invention can be achieved by adopting the following technical solutions:

[0039] A moisture-sensitive electronic skin, comprising:

[0040] A flexible humidity sensing unit, comprising a humidity sensing plane made of a whole piece of flexible humidity-sensitive material, and a peripheral measurement circuit connected to a data acquisition unit;

[0041] The data acquisition unit is connected to the peripheral measurement circuit, and sets the corresponding driving and acquisition modes according to the principle of electrical impedance imaging technology to measure the boundary voltage data, and transmits it to the humidity reconstruction unit for further processing and analysis;

[0042] The humidity reconstruction unit is used to input the boundary voltage data into the humidity distribution reconstruction model established according to the above analysis method to obtain the humidity distribution position, range, contour and other information of the humidity sensing plane.

[0043] The present invention has the following beneficial effects compared with the prior art:

[0044] 1. The present invention innovatively combines electrical impedance imaging technology with flexible humidity sensors to provide an executable technical solution for the preparation of humidity-sensitive electronic skin.

[0045] 2. The present invention replaces the existing m-row and n-column flexible humidity sensing array with a whole flexible humidity sensing plane, which reduces the electrical connections inside the humidity sensing unit, simplifies the manufacturing process of the flexible humidity sensor, increases the detection range of the electronic skin, and avoids the adverse effects of crosstalk on humidity perception accuracy.

[0046] 3. The present invention proposes a novel and lightweight humidity reconstruction method model. The memory and total parameter volume of the model are 18.6% of the currently published advanced reconstruction algorithms FC-UNet and 90% of EIM-CNN. In addition, the model maintains excellent reconstruction performance while meeting the lightweight requirements. Experimental results show that using the image correlation coefficient (ICC) as the evaluation index, the ICC obtained by the model provided by the present invention in the self-built simulation data set is 0.905, which is slightly better than FC_UNet (0.890) and significantly better than EIM-CNN (0.863).

[0047] 4. The present invention proposes an improved version of the reconstruction error, which uses convolution operation to expand the receptive field by replacing the pixel-level error with the local error, which can further stimulate the learning potential of the model and make the loss value curve smoothly decrease. In addition, by setting different weights for the target area, transition area, and background area, the model can pay more attention to the reconstruction in the target area during training, thereby improving the effect of humidity reconstruction.

[0048] 5. The electronic skin prepared by the present invention based on electrical impedance imaging technology, deep learning and other technologies can improve the spatial resolution of existing moisture-sensitive electronic skin, because it can obtain the resistance (conductivity) distribution of the two-dimensional plane of the sensing area, which can be further corresponded to the humidity distribution in the two-dimensional plane within the sensing area, that is, the humidity sensing space is expanded from the existing m×n point type to the entire two-dimensional plane, and can further realize the function of superhuman perception, that is, the humidity distribution in the two-dimensional plane can be specific to the humidity corresponding position, range, and contour information that is currently difficult for humans to perceive. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a framework diagram of the Mobile-EIT network model involved in the solution of Example 1 of the present invention.

[0050] Figure 2 It is a diagram of various convolution structures involved in the Mobile-EIT network model in the solution of Example 1 of the present invention.

[0051] Figure 3 It is a schematic diagram of the improved loss function weight setting method involved in the embodiment 1 of the present invention.

[0052] Figure 4This is a schematic diagram of the structure of the humidity-sensitive electronic skin involved in Example 2 of the present invention.

[0053] Figure 5 This is a system block diagram of a data acquisition unit according to Embodiment 2 of the present invention.

[0054] Figure 6 It is a schematic diagram of the collection method of relative excitation and adjacent measurement involved in Example 2 of the present invention.

[0055] Figure 7 It is a schematic diagram of a humidity distribution image involved in Example 2 of the present invention. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0057] Embodiment 1:

[0058] This embodiment provides a method for analyzing humidity distribution in a humidity sensing plane, comprising the following steps:

[0059] S10, obtain the data set and perform preprocessing, including:

[0060] S11. Obtain a simulation data set. A finite element model is established based on MATLAB and EIDORS software. The conductivity value of each grid of the finite element model is set according to the moisture resistance effect curve of the flexible moisture-sensitive material. The experimental conditions are simulated to generate a large-scale voltage data sample and a corresponding label map (actual humidity distribution map). When setting the conductivity value, the range, position, and contour of the distribution of high conductivity values ​​(corresponding to high humidity) and low conductivity values ​​(corresponding to low humidity) are random to meet the needs of sample diversity, and in the process of generating voltage data samples, a Gaussian white noise with a signal-to-noise ratio of 30 to 50 dB is randomly added to improve the noise resistance of the algorithm model.

[0061] S12, obtaining an experimental data set. A humidity sensing plane is made, and metal electrodes are arranged around the humidity sensing plane. An excitation current is injected into the humidity sensing plane through the metal electrodes, and the voltage between any two electrodes on the edge of the humidity sensing plane is measured, thereby collecting the boundary voltage of the flexible humidity sensing plane in the real environment and obtaining an experimental data set.

[0062] S20, construct a humidity distribution imaging model, such as Figure 1As shown, the humidity distribution imaging model is based on the Mobile-EIT algorithm, which adopts a data-to-image reconstruction strategy, that is, it does not need to use a traditional reconstruction algorithm (such as Tikhonov) to pre-reconstruct the boundary voltage, but directly uses a deep learning algorithm for end-to-end training. The humidity distribution imaging model includes:

[0063] Input layer, input boundary voltage data with size (1,192);

[0064] The pre-reconstruction module is mainly used to generate an initial humidity reconstruction result, including a fully connected layer (FC Layer) with 4096 nodes and a simple 7-layer autoencoder. The significance of this setting is: while ensuring the initial reconstruction effect, avoid using multiple fully connected layers to reduce the number of parameters; the fully connected layer output is resized to (64,64,1) and then output to the simple autoencoder for processing; the 7-layer structure of the autoencoder is a convolutional layer (Conv) with a convolution kernel size of 16, a batch normalization layer, a maximum pooling layer (MaxPooling) with a step size of 2, a convolutional layer with a convolution kernel size of 16, an upsampling layer (UpSampling) with a size of 2, and a convolutional layer with a convolution kernel size of 16; all convolutional layers and fully connected layers use the Relu function as the activation function;

[0065] The encoder module further extracts humidity-related multi-level and multi-scale features based on the pre-reconstruction module, including a convolution layer with a convolution kernel size of 3×3 and a number of 32, a batch normalization layer, 13 reverse residual blocks, and a pyramid pooling block; the setting of the reverse residual block refers to the first 13 reverse residual blocks of MobileNet2; the 1st, 3rd, 6th, and 13th reverse residual blocks output 4 multi-level features skip1(64,64,16), skip2(32,32,24), skip3(16,16,32), sk respectively. ip4(8,8,96); The pyramid pooling block contains five branches, which respectively extract local features with dilation rates of 1, 2, 3, 4, convolution kernel size of 3×3, and 96 dilated convolution layers, and extract global features with global average pooling layers; then multi-scale feature fusion is performed with convolution kernel size of 1×1 and 128 convolution layers, and global features are output skip5(8,8,128); The convolution operation in the reverse residual block and the dilated convolution in the pyramid pooling block both use Relu as the activation function, and are implemented using depthwise separable convolution to reduce the amount of calculation;

[0066] The decoder module is used to decode the multi-level and multi-scale features obtained by the encoder module to obtain a more refined humidity reconstruction result. It first performs two convolution operations on skip5(8,8,128) with a convolution kernel size of 64 and a size of 2 upsampling to obtain the local humidity reconstruction result up1(16,16,64). It concatenates skip3(16,16,32) and up1(16,16,64), and then performs two convolution operations and upsampling with a size of 2 to obtain up2(32,32,48). Similarly, it continues to convolve s kip2(32,32,48) is concatenated with up2(32,32,24) to obtain up3(64,64,32); then skip1(64,64,16) is concatenated with up3(64,64,32), and two convolution operations are performed using a convolution kernel size of 24 convolution layers. Finally, the final humidity reconstructed image is obtained by a convolution kernel size of 1 and a convolution layer with a sigmoid function as the activation function. Except for the last convolution layer, the remaining convolution layers use Relu as the activation function and are implemented using depthwise separable convolution to reduce the amount of calculation.

[0067] The contour constraint branch shares the structure and weights of the pre-reconstruction module and the encoder module with the decoder. The contour constraint branch is mainly used to alleviate the non-sharp boundaries and poor target contour reconstruction caused by the defects of the convolutional neural network itself, so as to ensure good edge preservation of the target area. The contour constraint branch selects the shallow feature skip1 (64, 64, 16) and the deep feature skip6 (64, 64, 128) for fusion, and then performs two convolution operations with a convolution kernel size of 3×3 and 128 convolution layers. Finally, the contour of the humidity reconstructed image is obtained through a convolution layer with a convolution kernel size of 1×1 and 1 number and an activation function of sigmoid function. Among them, skip6 is obtained by interpolation of skip5 (8, 8, 128). Except for the last convolution layer, the remaining convolution layers all use Relu as the activation function, and are all implemented using depthwise separable convolution to reduce the amount of calculation.

[0068] The output layer is used to output the humidity reconstruction image (i.e., Output A) and the humidity reconstruction contour image (i.e., Output B) both of size (64, 64, 1).

[0069] In this embodiment, the contour constraint branch is introduced into the model to solve the problem that the target in the reconstruction result looks blocky due to the spatial invariance of the convolutional neural network, resulting in the target's contour not being clear enough. The experimental results show that the addition of the contour branch can effectively improve the humidity reconstruction effect. The image correlation coefficient between the humidity reconstructed image output by the model and the real humidity reconstructed image on the self-built data set is increased from 0.894 to 0.905.

[0070] In this embodiment, the humidity distribution imaging model introduces the Relu nonlinear factor to avoid the disappearance of the gradient and the problem of overfitting, and batch normalization alleviates the internal covariate displacement by adding a normalization step and a scale and displacement step before each layer of nonlinearity, which can speed up the training speed and improve the reconstruction performance; and the humidity reconstructed image output by the output layer has the same size as the humidity reconstructed contour image.

[0071] In this embodiment, Figure 1 The various convolutions of the humidity distribution imaging model shown in the figure have corresponding structures as follows Figure 2 As shown, the essence is depth-wise separable convolution, which can reduce the computational complexity of the algorithm model. Figure 2 Parameter a refers to the number of convolution kernels, and parameter b refers to the size of the convolution kernel. For Inverted-Block and SepConv_BN, the size of the convolution kernel is fixed, and the number of convolution kernels can be adjusted.

[0072] S30, using the data set to train a humidity distribution imaging model to form a humidity distribution reconstruction model having two outputs: a humidity reconstruction image and a humidity reconstruction contour image, specifically including:

[0073] S31. Divide the simulated data set into a training set, a validation set, and a test set in a ratio of 7:2:1, input the training set and the validation set into the humidity distribution imaging model for training and learning, and use the test set to evaluate the performance of the algorithm model obtained by training.

[0074] S32, continuously adjust the learning rate, training epoch, learning strategy and loss function parameters in the training phase, and select and save the residual neural network algorithm model that has completed training and obtained the best performance based on the test set. In this embodiment, the Adam optimization algorithm is selected, and the learning rate is set to 10 -4 , epoch is 120, batch_size is 8, improved reconstruction error and focal loss are selected as loss functions, and the learning rate adjustment rule is set as follows: if the total loss of the validation set does not decrease within 5 epochs, the learning rate is reduced by 0.1, and the Early Stopping rule is set as follows: if the total loss of the validation set does not decrease within 10 epochs, stop training.

[0075] Specifically, the loss function is calculated as follows:

[0076] The humidity reconstructed image uses the improved reconstruction error as the loss function, and the output humidity reconstructed contour image uses the focal loss as the loss function;

[0077] For the improved reconstruction error, different weights are set for the background and the target, and then the reconstruction error is calculated; Figure 3 As shown in the figure, the weight setting method is to convolve an all-one matrix (also called an all-one template) of size (3, 3, 1, 1) with the real humidity distribution image and the model reconstructed humidity distribution image, so that the target center area obtains the highest weight, the transition area between the target and the background obtains the second highest weight, and the background area obtains the lowest weight; the significance of the improvement is: first, through the 3×3 convolution operation, the local error is used instead of the pixel-level error to expand the receptive field; second, by setting different weights for the target area, transition area, and background area, the model can be prompted to pay more attention to the reconstruction in the target area during training, thereby improving the effect of humidity reconstruction.

[0078] The calculation process of the improved reconstruction error is shown in the following formula (1), where filter is the all-one matrix mentioned above, y_true is the true humidity distribution image, y_pred is the humidity distribution image reconstructed by the model, and N is the number of batch training;

[0079]

[0080] For focal loss, the calculation process is shown in the following formula (2), where y is a binary label 0 or 1 (0 means it does not belong to the contour area, 1 means it belongs to the contour area), p is the model prediction value, in this embodiment, the parameter α is 0.2, and γ is 0.4;

[0081] Loss2=-αy(1-p) γ log(p)-(1-α)(1-y)p γ log(1-p) (2)

[0082] For the total loss, the calculation process is shown in the following formula (3). In this embodiment, the parameter β is set to 0.36;

[0083] Loss=Loss1+β*Loss2 (3)

[0084] S33. Divide the experimental data set into a training set and a test set in a ratio of 8:2, perform secondary training on the algorithm model based on the idea of ​​transfer learning, use the weights of each parameter of the algorithm model saved in step S32 as the initial weights, and input the training set samples of the experimental data set into the algorithm model constructed in step S20 again for training. During training, only the weights of the decoder and contour constraint branches are updated.

[0085] S34, continuously adjust the learning rate, training epoch, learning strategy, loss function and other parameter settings in the training phase, and select and save the algorithm model with the best performance in step S33 as the humidity distribution reconstruction model for the final practical application.

[0086] In this embodiment, adding a contour constraint branch to the model effectively improves the humidity reconstruction effect. The image correlation coefficient between the humidity reconstructed image output by the model and the real humidity reconstructed image on the self-built data set is improved from 0.894 to 0.905.

[0087] In this embodiment, a lightweight humidity reconstruction model is constructed using depthwise separable convolutions. The model memory is approximately 19.2MB, and the total number of parameters is approximately 1.54 million, which is 18.6% of the currently published advanced reconstruction algorithm of the same type FC-UNet (the FC-UNet model memory is approximately 103MB, and the total number of parameters is approximately 8.5 million), and 90% of EIM-CNN (where the EIM-CNN model memory is approximately 21.3MB, and the total number of parameters is approximately 1.76 million).

[0088] In this embodiment, the algorithm model is trained twice based on the idea of ​​transfer learning. Its significance lies in that the simulated data samples are simple and easy to obtain, but the training samples generated in the simulated environment may not fully take into account all the interference factors in the real environment, and the experimental data samples and the simulated experimental data samples may not meet the conditions of independent and identical distribution. Therefore, the algorithm model trained with simulated data is directly used to reconstruct the humidity of the experimental data, and the results are generally poor. However, due to the high time cost and difficulty in obtaining experimental data samples, the sample capacity is usually small, which cannot meet the sample size requirements of deep learning. Directly using small-scale experimental data samples is prone to under-fitting or over-fitting of the algorithm model. Therefore, in this embodiment, the algorithm model is first trained with large-scale simulated data samples to allow the algorithm model to learn the general mapping relationship from boundary voltage to internal conductivity; on this basis, the special mapping relationship that is helpful for the main task of humidity reconstruction is learned through small-scale experimental data. The two-stage training strategy used in this embodiment can further improve the accuracy of humidity reconstruction.

[0089] Embodiment 2:

[0090] This embodiment provides a moisture-sensitive electronic skin, such as Figure 4 As shown, including:

[0091] A flexible humidity sensing unit, comprising a humidity sensing plane made of a whole piece of flexible humidity-sensitive material, and a peripheral measurement circuit connected to a data acquisition unit;

[0092] The data acquisition unit is connected to the peripheral measurement circuit, and the corresponding driving and acquisition modes are set according to the principle of electrical impedance imaging technology, so as to collect the boundary voltage data of the humidity sensing plane through the peripheral measurement circuit, and transmit it to the humidity reconstruction unit for further processing and analysis;

[0093] The humidity reconstruction unit is used to construct a lightweight humidity distribution reconstruction model, and input the boundary voltage data into the constructed humidity distribution reconstruction model to analyze the humidity distribution of the humidity sensing plane, including the position, range, contour and other information. The constructed humidity distribution reconstruction model is based on a lightweight humidity reconstruction algorithm (Mobile-EIT).

[0094] In this embodiment, the flexible moisture-sensitive material is made of a knitted fabric made of a mixture of conductive yarn and insulating cotton thread. The fabric is treated with hydrophilicity to improve its water absorption capacity and exhibits different electrical properties at different water absorption rates, that is, the more water a certain part of the fabric absorbs, the smaller the resistance value of that part; the less water a certain part of the fabric absorbs, the greater the resistance value of that part. The conductivity of the mixed conductive yarn of conductive yarn and insulating cotton thread needs to be controlled within the range of 1 to 20 kΩ / cm. In specific implementation, the conductivity of the mixed conductive yarn can be adjusted by adjusting the ratio of conductive yarn and insulating cotton thread.

[0095] In this embodiment, the above-mentioned flexible humidity-sensitive material is cut into a block with a size of 5.6cm×5.6cm as a humidity sensing plane, and 16 metal electrodes are evenly arranged around the humidity sensing plane. The metal electrodes are directly connected to the humidity-sensitive material, that is, the excitation current can be injected into the humidity sensing plane or the voltage between any two electrodes on the side of the humidity sensing plane can be measured through the metal electrodes. The 16 metal electrodes are connected to the multi-way switch in the data acquisition unit, that is, the 16 metal electrodes can be controlled by the multi-way switch in the data acquisition unit. In this embodiment, the shape and size of the humidity sensing plane, as well as the shape and number of the metal electrodes can be adjusted according to the application scenario requirements during specific implementation.

[0096] In this embodiment, the data acquisition unit is designed according to the principle of electrical impedance imaging system, and its system block diagram is as follows: Figure 5As shown. The unit includes: a microcontroller, an excitation current source module, a multi-way switch, a signal processing module, and a communication module. The microcontroller first controls the excitation current source module to output a sinusoidal excitation current with a frequency of 1kHz and a peak-to-peak value of 2mA; then the excitation electrode is selected through the multi-way switch to inject the excitation current into the flexible humidity sensing plane; at the same time, the voltage signal between the electrodes is fed back to the signal processing module by selecting the measurement electrode through the multi-way switch; the signal processing module performs noise reduction filtering on the collected voltage signal and extracts the amplitude of the signal, and finally transmits it to the humidity reconstruction unit through the communication module in the form of Bluetooth communication. The multi-way switch is composed of four MAX306 chips, which respectively control the selection of a pair of excitation electrodes and a pair of measurement electrodes.

[0097] In this embodiment, data acquisition is performed using a relative excitation and adjacent measurement driving method. Figure 6 As shown, the metal electrode r1 and the metal electrode r9 are selected as the excitation electrodes through the multi-way switch to inject the excitation current into the flexible humidity sensing plane, and the metal electrode r2 and the metal electrode r3 are selected as the measuring electrodes in a clockwise manner, and then the metal electrode r3 and the metal electrode r4 are selected as the measuring electrodes. By analogy, 12 voltage data can be measured by stimulating one pair of electrodes; similarly, the metal electrode r2 and the metal electrode r10 are continued to be selected as the excitation electrodes in a clockwise direction to measure the voltage data between adjacent electrodes. By analogy, a total of 16 pairs of excitation electrodes can be selected, and 192 voltage data can be obtained in one complete measurement.

[0098] In this embodiment, the humidity distribution reconstruction model constructed by the humidity reconstruction unit is obtained by the analysis method of humidity distribution in the humidity sensing plane of the first embodiment of the present invention. In the humidity reconstruction unit, the above 192 voltage data will be input into the lightweight humidity distribution reconstruction model to output the humidity distribution image of the flexible humidity sensing plane, whose size is (64,64,1). Figure 7 As shown in FIG. 1 , a schematic diagram of a humidity distribution image is shown. The color column in the figure represents the normalized humidity value, that is, 0 represents a water absorption rate of 0%, which is the lowest humidity, and 1 represents a water absorption rate of 100%, which is the highest humidity. Figure 7 The specific distribution range, location, contour and other information of high humidity areas and low humidity areas can be clearly seen, rather than being limited to obtaining humidity information of m×n location points.

[0099] In summary, the present invention proposes a method for analyzing humidity distribution in a humidity sensing plane, which can establish a new and lightweight humidity distribution reconstruction model that meets the lightweight requirements while maintaining excellent reconstruction performance. At the same time, different functions are used for different output results in the process of model establishment, and the algorithm model is trained twice based on the idea of ​​transfer learning, which can prompt the model to pay more attention to the reconstruction in the target area during the training process, thereby improving the effect of humidity reconstruction.

[0100] The humidity distribution reconstruction model established based on the analysis method of humidity distribution in a humidity sensing plane provided by the present invention can be applied to humidity-sensitive electronic skin to analyze the humidity distribution in the humidity sensing plane, thereby expanding the perception range of the humidity-sensitive electronic skin from the existing m×n point type to the entire two-dimensional plane. It can also further realize the function of superhuman perception, that is, the humidity distribution in the two-dimensional plane can be specific to the position and range corresponding to the humidity, as well as contour information that is currently difficult for humans to perceive.

[0101] Obviously, the embodiments described above are only part of the embodiments of the present invention, rather than all of the embodiments. The present invention is not limited to the details of the above embodiments, and any appropriate changes or modifications made by ordinary technicians in the relevant technical field are deemed to be within the patent scope of the present invention.

Claims

1. A method for analyzing humidity distribution in a humidity sensing plane, characterized in that: The following steps are involved: Get the data set and preprocess it; Constructing a humidity distribution imaging model, wherein the humidity distribution imaging model includes an input layer, a pre-reconstruction module, an encoder module, a decoder module, a contour constraint branch, and an output layer; The humidity distribution imaging model is trained using the data set to form a humidity distribution reconstruction model with two outputs: humidity reconstruction image and humidity reconstruction contour image; The input layer is used to input the boundary voltage data of the humidity sensing plane, and the output layer outputs the humidity reconstruction image and the corresponding contour image; the pre-reconstruction module is used to generate the initial humidity reconstruction result, and the encoder module is used to further extract humidity-related multi-level and multi-scale features based on the pre-reconstruction module; The decoder module decodes according to the multi-level and multi-scale features extracted by the encoder module; the contour constraint branch is used to obtain a humidity reconstructed image with clear contours; The data set is used to train the humidity distribution imaging model, and the reconstruction error is used as the loss function for the humidity reconstructed image output, and the focal loss is used as the loss function for the humidity reconstructed contour image; The reconstruction error is calculated as follows: Different weights are assigned to the target and background of the real humidity distribution image in the data set and the humidity reconstructed image output by the humidity distribution reconstruction model, so that the target center area obtains the highest weight, the transition area between the target and the background is second, and the background area obtains the lowest weight; Calculate the mean square error between the real humidity distribution image and the humidity reconstruction image after weight allocation as the reconstruction error; The reconstruction error is calculated as follows: The filter is an all-one matrix of size (3,3,1,1), y_true is the true humidity distribution image, y_pred is the humidity reconstructed image, and N is the number of batch training; The focal loss is calculated as follows: Loss2=-αy(1-p) γ log(p)-(1-α)(1-y)p γ log(1-p) Where y is a binary label, y = 0 means it does not belong to the contour area, y = 1 means it belongs to the contour area, p is the model prediction value, and α and γ are constant parameters.

2. The method for analyzing humidity distribution in a humidity sensing plane according to claim 1, characterized in that: In the humidity distribution imaging model: The pre-reconstruction module includes a fully connected layer and a simple autoencoder. The input boundary voltage data is resized in the fully connected layer and then input into the simple autoencoder for processing. The encoder module includes a convolutional layer, a batch normalization layer, 13 reverse residual blocks, and a pyramid pooling block; the 1st, 3rd, 6th, and 13th reverse residual blocks output 4 multi-level features backward respectively; the pyramid pooling block contains five branches, which are used to extract local features by dilated convolution and global features by global average pooling, and then multi-scale feature fusion is performed through the convolution layer to output global features; A decoder module is used to decode the multi-level features and global features obtained by the encoder module to obtain a humidity reconstructed image with clear contours; The contour constraint branch is used to fuse the shallow features and the deep features obtained by interpolating the global features to obtain the humidity reconstructed contour image.

3. The method for analyzing humidity distribution in a humidity sensing plane according to claim 2, characterized in that: The last convolutional layer of the decoder branch and the contour constraint branch uses the sigmoid function as the activation function, and the other convolutional layers use Relu as the activation function.

4. The method for analyzing humidity distribution in a humidity sensing plane according to claim 1, characterized in that: Get the dataset and preprocess it, including: Establish a finite element model, set the conductivity value of each grid of the finite element model according to the moisture resistance effect curve, simulate the experimental conditions to generate a large-scale voltage data sample and the corresponding humidity distribution map, and obtain a simulation data set; The boundary voltage of the flexible humidity sensing unit in the real environment is collected, and the corresponding humidity distribution map is produced to obtain the experimental data set.

5. The method for analyzing humidity distribution in a humidity sensing plane according to claim 4, characterized in that: The method of using the data set to train the humidity distribution imaging model comprises: The humidity distribution imaging model is trained using a simulated data set to obtain the initial weights of the pre-reconstruction module, encoder module, decoder module, and contour constraint branch; Freeze the weights of the pre-reconstruction module and the encoder module, further train the decoder module and the contour constraint branch using the experimental dataset, and update the weights of the decoder module and the contour constraint branch.

6. A moisture-sensitive electronic skin, characterized in that: include: A flexible humidity sensing unit, comprising a humidity sensing plane and a peripheral measurement circuit; A data acquisition unit, connected to the peripheral measurement circuit, for measuring boundary voltage data; The humidity reconstruction unit is used to input the boundary voltage data into the humidity distribution reconstruction model established according to the analysis method of humidity distribution in the humidity sensing plane according to any one of claims 1-5 to obtain the humidity distribution position, range and contour of the humidity sensing plane.

Images