A method of correcting fabric pressure sensor array test results

Abstract

This invention relates to a method for correcting the detection results of a fabric pressure sensor array. An improved U-Net convolutional neural network model is established to correct the detection results, effectively eliminating the problem of inaccurate pressure distribution cloud map display results caused by crosstalk. This invention allows the use of the simplest voltage divider circuit as the basic measurement circuit for a piezoresistive fabric pressure sensor array, significantly reducing the complexity of the measurement circuit. The method is simple and easy to operate, and the output pressure distribution cloud map results are accurate, effectively improving the impact of crosstalk on the pressure distribution cloud map display results obtained by the piezoresistive fabric pressure sensor array.

Description

TECHNICAL FIELD

[0001] The present application relates to a flexible sensor and artificial intelligence technology, in particular to a method for correcting detection results of a fabric pressure sensing array. BACKGROUND

[0002] The pressure distribution measurement of the piezoresistive flexible fabric pressure sensing array requires to quickly measure the resistance value of each sensing unit node in a large-scale resistance network. Since each sensing unit actually forms a parallel loop with other sensing units, directly measuring the resistance value of the sensing unit through the electrodes connected to both ends of the sensing unit actually obtains the total resistance value of the sensing unit and its surrounding sensing units in parallel, resulting in inaccurate detection results. In addition to the crosstalk problem caused by the above mechanism, due to the flexibility of the fabric, when a certain sensing unit is subjected to pressure, due to the continuous deformation characteristics of the fabric, the sensing unit adjacent to the sensing unit actually not subjected to pressure will also be deformed to change the resistance value, resulting in distorted pressure detection results. In addition, due to the limitations of the fabric pressure sensing array manufacturing process, abnormal sensing units appear in the sensing array, resulting in noise points on the generated pressure detection cloud map, which interferes with data analysis. The method commonly used to solve the crosstalk problem of the fabric pressure sensing array is to improve the design in the hardware aspect, such as including the sensor structure, the measurement circuit, etc., but these methods will either reduce the sensitivity of the sensing unit, or the effect of solving the crosstalk problem is not good, or a very complex hardware circuit needs to be designed. SUMMARY

[0003] In view of the problems existing in the pressure distribution measurement of the piezoresistive flexible fabric pressure sensing array, a method for correcting the detection results of the fabric pressure sensing array is proposed to effectively eliminate the problem of inaccurate display of the pressure distribution cloud map output by the fabric pressure sensing array caused by the crosstalk phenomenon.

[0004] The technical scheme of the present application is: a method for correcting the detection results of a fabric pressure sensing array, specifically comprising the following steps:

[0005] 1) constructing a U-Net convolutional neural network model, and modifying the U-Net convolutional neural network model into an image-to-image regression prediction model to obtain an improved U-Net convolutional neural network model;

[0006] 2) placing the fabric pressure sensing array directly above a pressure sensing array with accurate detection results, and simultaneously collecting multiple pressure distribution cloud maps under the same pressure;

[0007] 3) using a median filtering algorithm to process outliers of each pressure cloud map collected by the fabric pressure sensing array;

[0008] 4) using the pressure distribution cloud diagram obtained after processing in step 3) as the input image of the improved U-Net convolutional neural network model established in step 1), using the pressure distribution cloud diagram under the same pressure collected by the pressure sensing array with accurate detection results obtained in step 2) as the output image of the improved U-Net convolutional neural network model, using the data augmentation function to expand the input image and the output image, and constructing a training data set for training the improved U-Net convolutional neural network model;

[0009] 5) training the improved U-Net convolutional neural network model using the training data set until convergence is reached;

[0010] 6) using the fabric pressure sensing array to measure the pressure distribution of the fabric, and inputting the collected pressure distribution cloud diagram into the improved U-Net convolutional neural network model trained in step 5) to output a corrected pressure distribution cloud diagram.

[0011] Further, the step 1) is specifically implemented as follows:

[0012] The constructed U-Net convolutional neural network model performs 4 times of down-sampling and 4 times of up-sampling on the input pressure distribution cloud diagram, and the network has a total of 4 layers. First, the input image is subjected to image feature extraction through 64 convolution kernels and a ReLU activation function, then down-sampling is performed, and then convolution and activation are performed through 128 convolution kernels, and so on to complete 4 times of down-sampling to obtain 1024 feature layers with a size of 8x8. From bottom to top, the obtained 1024 feature layers with a size of 8x8 are subjected to deconvolution through 512 convolution kernels to constantly expand the picture, and the corresponding up-sampled feature layers are increased by means of the picture during down-sampling, and then convolution kernel convolution operation is performed, and so on. Through 4 times of up-sampling, the number of feature image pixels is increased, and finally the picture size and accuracy are restored. On the basis of the U-Net convolutional neural network model, the Softmax normalization layer and the Segmentation classification layer at the end of the U-Net convolutional neural network model are deleted, then a Regression regression layer is added and connected with the last convolution layer, the number of filters of the last convolution layer is modified to 3, and the total number of layers is 33, to obtain the improved U-Net convolutional neural network model. The input image size is 128x128x3, and after multiple convolution and pooling, the output image size is still 128x128x3.

[0013] Further, after the median filtering algorithm is used to process the abnormal values of each pressure cloud diagram collected by the fabric pressure sensing array in step 3), the peak signal-to-noise ratio (PSNR) is used to evaluate the abnormal value removal effect of the pressure cloud diagram before and after processing by the median filtering algorithm.

[0014] Furthermore, in step 5), during the training of the improved U-Net convolutional neural network model using the training dataset, the change in pixel size M in the pressure cloud map under each convolution operation is calculated. (h或w) :

[0015] In the formula, h is the number of rows, w is the number of columns, n is the number of pixels, p is the number of pixel rows and columns for edge padding, f is the pixel size of the convolution kernel, and s is the stride size;

[0016] The cross-entropy loss function E is used to calculate the difference in performance between good and bad models during training:

[0017]

[0018] In the formula, X is the input image to be processed, and p l(x) (X) represents the actual output probability, Ω represents the image dataset, l represents the true label of each pixel when Ω approaches {1,...,K}, K is a natural number, and w(X) is the introduced weight map used to compensate for different frequencies of image pixels during training.

[0019]

[0020] In the formula, w c To balance the probability of different images appearing, d1 is the distance between the feature to be segmented in the image and the nearest boundary, d2 is the distance between the feature to be segmented in the image and the second nearest boundary, w0 is the exponential term weight update coefficient, and σ is the standard deviation.

[0021] Furthermore, in the process of training the improved U-Net convolutional neural network model using the training dataset, the training set is divided into mini-batches for training the model. The mini-batch stochastic gradient descent method is used to obtain the minimum value of the loss function, ensuring that as many samples as possible are input into the model in each round of iterative training.

[0022] Furthermore, the fabric pressure sensing array fabrication method in step 2) is as follows:

[0023] The first silver-plated conductive yarn and the first non-conductive polyester yarn are symmetrically sewn in an array along the warp direction of the fabric substrate on the front and back sides of the fabric substrate, respectively. The array spacing of the first silver-plated conductive yarn and the first non-conductive polyester yarn is 10mm. The first silver-plated conductive yarn and the first non-conductive polyester yarn have a first locking buckle for interlocking connection at the intersection of the non-warp and weft arrays.

[0024] Subsequently, a composite solution consisting of polyurethane / N,N-dimethylformamide solution and conductive carbon black was printed onto the first silver-plated conductive yarn and the nearby fabric using screen printing and then dried to form a pressure sensing unit array.

[0025] Finally, using the lockstitch method, the second silver-plated conductive yarn and the second non-conductive polyester yarn are symmetrically sewn in an array along the weft direction of the fabric substrate on the front and back sides of the fabric substrate, respectively. The array spacing of the second silver-plated conductive yarn and the second non-conductive polyester yarn is 10mm. The second silver-plated conductive yarn and the second non-conductive polyester yarn have a second locking buckle for interlocking connection at the intersection of the non-warp and weft arrays. The pressure sensing unit is located at the intersection of the first silver-plated conductive yarn and the second silver-plated conductive yarn on the front side of the fabric substrate, so that the conductive yarn and the sensing unit are connected by circuit, thus obtaining a fabric pressure sensing array with 32×32 sensing units.

[0026] Furthermore, in step 2), the pressure sensor array with accurate detection results adopts the RX-M3232L type piezoresistive flexible thin film pressure sensor pad, and the number of its sensing units and the center distance between adjacent sensing units are the same as those of the fabric pressure sensor array.

[0027] The beneficial effects of this invention are as follows: This invention provides a method for correcting the detection results of a fabric pressure sensor array. It utilizes machine learning to solve the crosstalk phenomenon of the piezoresistive fabric pressure sensor array. The method allows the use of the simplest voltage divider circuit as the basic measurement circuit for the piezoresistive fabric pressure sensor array, significantly reducing the complexity of the measurement circuit. The method is simple and easy to operate, and the output pressure distribution cloud map results are accurate. It can effectively improve the impact of crosstalk on the display results of the pressure distribution cloud map obtained by the piezoresistive fabric pressure sensor array. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the surface structure of the fabric pressure sensing array on one side of the present invention, which contains the sensing unit.

[0029] Figure 2 This is a schematic diagram of the surface structure on the opposite side of the surface where the sensing unit of the fabric pressure sensing array of Embodiment 1 of the present invention is located.

[0030] Figure 3 This is a pressure distribution cloud map obtained when no pressure is applied to the fabric pressure sensing array in Embodiment 1 of the present invention;

[0031] Figure 4 This is a pressure distribution cloud map of a cylindrical object placed on a fabric pressure sensing array according to Embodiment 1 of the present invention.

[0032] Figure 5 This is a structural diagram of the U-Net convolutional neural network model in Embodiment 1 of the present invention;

[0033] Figure 6This is the cloud map after processing the pressure distribution cloud map collected by the fabric pressure sensing array without any object placed, using the median filtering algorithm in Embodiment 1 of the present invention;

[0034] Figure 7 This is a cloud map obtained by processing the pressure distribution cloud map of a cylindrical object placed by the fabric pressure sensing array using the median filtering algorithm in Embodiment 1 of the present invention.

[0035] Figure 8 This is an error curve diagram of the training of the improved U-Net convolutional neural network model of Embodiment 1 of the present invention;

[0036] Figure 9 Pressure distribution cloud maps under the same pressure conditions were collected for comparison in this invention. Detailed Implementation

[0037] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0038] Figure 1 and Figure 2 A schematic diagram of the structure of the piezoresistive fabric pressure sensing array 1 of this embodiment is shown. The fabrication process and structure of the piezoresistive fabric pressure sensing array 1 are as follows: A first silver-plated conductive yarn 2 and a first non-conductive polyester yarn 7 are symmetrically sewn in an array along the warp direction of the fabric substrate 5 on the front and back sides of the fabric substrate 5 using a lockstitch method. The array spacing of the first silver-plated conductive yarn 2 and the first non-conductive polyester yarn 7 is 10mm. At the intersection of the non-warp and weft arrays of the first silver-plated conductive yarn 2 and the first non-conductive polyester yarn 7, there are first locking buckles 6 for interlocking connection. Subsequently, a composite solution composed of polyurethane / N,N-dimethylformamide solution and conductive carbon black is printed onto the first silver-plated conductive yarn 2 and the nearby fabric using a screen printing method, and then dried to form a pressure sensing array. The pressure sensing unit array is then constructed. Finally, using a lockstitch method, the second silver-plated conductive yarn 4 and the second non-conductive polyester yarn 8 are symmetrically sewn in an array along the weft direction of the fabric substrate 5 on the front and back sides of the substrate. The array spacing of the second silver-plated conductive yarn 4 and the second non-conductive polyester yarn 8 is 10mm. At the intersection of the non-warp and weft arrays of the second silver-plated conductive yarn 4 and the second non-conductive polyester yarn 8, there is a second locking buckle 9 for interlocking connection. The pressure sensing unit 3 is located on the front side of the fabric substrate 5 at the intersection of the first silver-plated conductive yarn 2 and the second silver-plated conductive yarn 4, enabling the conductive yarn and the sensing unit to achieve circuit connection, thus producing a fabric pressure sensing array 1 with 32×32 sensing units 3. The fabric pressure sensing array 1 is then connected to the first silver-plated conductive yarn 2 and the second silver-plated conductive yarn 8...Figure 1 , 2 The data acquisition and wireless transmission module (not shown) is connected to construct a fabric pressure sensor array system, which can transmit the acquired data to... Figure 1 , 2 A pressure distribution cloud map can be generated on a computer or mobile device (not shown) for viewing and analysis.

[0039] Figure 3 The diagram shows the pressure distribution contours obtained when no pressure is applied to the fabric pressure sensing array 1. Figure 3 As can be seen, although no pressure is applied to the sensor array 1, some sensor units 3 still display abnormal pressure values. These abnormal results are often related to factors such as short circuits caused by the first silver-plated conductive yarn 2 and the second silver-plated conductive yarn 4 being connected at the sensor unit 3 during the fabric pressure sensor array 1 manufacturing process, and poor printing quality of individual sensor units causing their resistance values ​​to differ significantly from other sensor units.

[0040] Figure 4 The diagram shows the pressure distribution cloud map when a cylindrical object is placed on the fabric pressure sensor array 1. As can be seen from the diagram, apart from abnormal pressure detection values ​​caused by malfunctioning sensing units on the fabric pressure sensor array 1, the pressure distribution cloud map... Figure 4 The pressure distribution cloud map was also distorted due to crosstalk issues.

[0041] The detection results of the fabric pressure sensing array 1 are corrected using an improved U-Net convolutional neural network model, including the following steps:

[0042] Step 1: Construct a U-Net convolutional neural network model and modify it into an image-to-image regression prediction model to obtain an improved U-Net convolutional neural network model. Specifically, the constructed U-Net convolutional neural network model performs 4 downsampling and 4 upsampling operations on the input pressure distribution cloud map, with a total of 4 network layers. Figure 5 The structure diagram of the U-Net convolutional neural network model is shown. Figure 5 The left half of the diagram represents the process of first extracting image features from the input image using 64 convolutional kernels and the ReLU activation function, then performing downsampling, followed by convolution and activation using 128 convolutional kernels, and so on, to complete four downsampling operations, resulting in 1024 8×8 feature layers. Figure 5The right side of the image is processed from bottom to top, with 1024 8×8 feature layers deconvolved using 512 convolutional kernels to continuously enlarge the image. To prevent information loss, the downsampled image from the left side is used to add to the upsampled feature layers on the right, which are then convolved again, and so on. Through four upsampling operations, the number of pixels in the feature image is increased, ultimately restoring the image size and accuracy. Based on the existing U-Net convolutional neural network model, the final Softmax normalization layer and Segmentation classification layer are removed, and a Regression layer is added connected to the last convolutional layer. The number of filters in the final convolutional layer is modified to 3, resulting in a total of 33 layers, yielding an improved U-Net convolutional neural network model. The input image size is 128×128×3, and after multiple convolutions and pooling, the output image size remains 128×128×3.

[0043] Step 2: Place the fabric pressure sensor array 1 directly above the pressure sensor array with accurate detection results, and simultaneously collect pressure distribution cloud maps under the same pressure conditions multiple times. Specifically, in this embodiment, the pressure sensor array with accurate detection results uses an RX-M3232L type piezoresistive flexible film pressure sensor pad. The RX-M3232L type piezoresistive flexible film pressure sensor pad is made by combining organic polymer materials with a flexible film through precision printing and encapsulation. It has 32×32 sensing units, with a center-to-center distance of 10mm between adjacent sensing units. Each sensing unit has infinite resistance when not under pressure, and the sensing units adopt a distributed independent design. When a sensing unit is under pressure, the adjacent sensing units that are not under pressure will not deform, and therefore will not change their resistance, thus eliminating crosstalk effects. When collecting input and output datasets, the fabric pressure sensor array 1 is placed directly above the RX-M3232L type piezoresistive flexible film pressure sensor pad to simultaneously collect pressure cloud maps under the same pressure conditions. In this embodiment, 2160 pressure distribution cloud maps were collected using the fabric pressure sensor array 1 and the RX-M3232L type piezoresistive flexible thin film pressure sensor pad, respectively.

[0044] Step 3: Use a median filtering algorithm to process outliers in each pressure contour image acquired by the fabric pressure sensor array 1. In this embodiment, the pixel template size used in the median filtering algorithm is 128×128. Figure 6 This shows the effect of processing the pressure distribution cloud map acquired by the fabric pressure sensor array 1 when no object is placed on it using the median filtering algorithm. Figure 7 This shows the effect of processing the pressure distribution cloud map of a cylindrical object placed on it, acquired by the fabric pressure sensor array 1, using a median filtering algorithm.Figure 3 and Figure 4 The comparison clearly shows that the median filtering algorithm has a very significant effect on removing outliers in the pressure distribution contour map. The peak signal-to-noise ratio (PSNR) is used to evaluate the processing effect of the median filtering algorithm. Specifically, the mean square error (MSE) of the two pressure contour maps before and after processing is calculated first.

[0045]

[0046] In the formula, X is the image before processing, Y is the image after processing, H is the number of pixels in the height direction of the image, and W is the number of pixels in the width direction of the image.

[0047] Next, calculate the peak signal-to-noise ratio (PSNR) value:

[0048]

[0049] In the formula, n is the number of bits in the image pixels.

[0050] Table 1 lists the criteria for evaluating image quality based on Peak Signal-to-Noise Ratio (PSNR). (The table then lists the criteria for each criterion.) Figure 3 and Figure 6 , Figure 4 and Figure 7 The peak signal-to-noise ratio (PSNR) of the pressure cloud map was calculated, and the results are shown in Table 2. Table 2 shows that after processing with the median filtering algorithm, the image quality of the pressure cloud map reached the "good" level shown in Table 1. However, due to crosstalk, the distortion in the pressure cloud map was not eliminated by the median filtering algorithm's handling of outliers.

[0051] Table 1

[0052]

[0053] Table 2

[0054]

[0055] Step 4: Use the pressure distribution cloud map obtained in Step 3 as the input image for the improved U-Net convolutional neural network model, and use the pressure distribution cloud map under the same pressure, simultaneously acquired by a pressure sensing array with accurate detection results, as the output image for the improved U-Net convolutional neural network model. Use data augmentation functions to expand both the input and output images, constructing a training dataset for training the improved U-Net convolutional neural network model. Specifically, the data augmentation function used is the `imadjust` function, expanding the input and output images to 10,800 images each, for a total of 21,600 images. Of these, 15,120 images are used as the training dataset for training the improved U-Net convolutional neural network model.

[0056] Step 5: Train the improved U-Net convolutional neural network model using the training dataset until convergence is achieved. Specifically, during training, it is necessary to calculate the change M of pixel size in the pressure cloud map under each convolution operation. (h或w) :

[0057]

[0058] In the formula, h represents the number of rows, w represents the number of columns, n represents the number of pixels, p represents the number of pixel rows and columns for edge padding, f represents the pixel size of the convolution kernel, and s represents the stride size.

[0059] The cross-entropy loss function E is used to calculate the difference in performance between good and bad models during training:

[0060]

[0061] In the formula, X is the input image to be processed, and p l(x) (X) represents the actual output probability, Ω represents the image dataset, l represents the true label of each pixel when Ω approaches {1,...,K}, K is a natural number, and w(X) is the introduced weight map used to compensate for different frequencies of image pixels during training.

[0062]

[0063] In the formula, w c To balance the probability of different images appearing, d1 is the distance between the feature to be segmented in the image and the nearest boundary, d2 is the distance between the feature to be segmented in the image and the second nearest boundary, w0 is the exponential term weight update coefficient, and σ is the standard deviation.

[0064] To improve training speed, mini-batch stochastic gradient descent is used to minimize the loss function, ensuring that as many samples as possible are input into the model in each iteration. In this embodiment, the mini-batch size is set to 128, meaning the 7560 samples in the training set are divided into 59 mini-batches of 128 samples each for model training. The solver used for training is Adam, the learning rate is set to 0.001, the validation frequency is 50 times per epoch, the number of training epochs is 50, and the runtime environment is a single-core GPU.

[0065] The root mean square error (RMSE) is used as the evaluation metric for training effectiveness.

[0066]

[0067] In the formula, x i f(x) is the input value. i ) represents the model's predicted value, y iis the true value, and m is the number of samples.

[0068] Figure 8 The error curve of the model training is shown. In the first 100 iterations, the RMSE value rapidly decreased from around 105 to around 20, indicating a significant reduction in training error. With continued training, the error gradually decreased from around 20 to a final value of 7.108. During this process, the model's training accuracy was continuously refined, with weight parameters adjusted in each iteration until convergence was achieved.

[0069] Step Six: Collect pressure distribution cloud maps using the fabric pressure sensor array 1. Input the collected pressure distribution cloud maps into the improved U-Net convolutional neural network model trained in Step Five, and output the corrected pressure distribution cloud map. In this embodiment, the fabric pressure sensor array 1 collected pressure data for four types of objects: palm, foot, cylindrical object, and rectangular wooden stick, and obtained pressure distribution cloud maps for each. These were then input into the improved U-Net convolutional neural network model trained in Step Five to obtain the corrected pressure distribution cloud map output. To compare and evaluate the correction results, pressure distribution cloud maps under the same pressure conditions were collected, and the results are as follows. Figure 9 As shown. Figure 9 (a) shows the pressure distribution cloud map of the palm. Figure 9 (b) shows the foot pressure distribution cloud map. Figure 9 Image (c) shows the pressure distribution contour map of the cylindrical object. Figure 9 Image (d) shows the pressure distribution cloud map of the rectangular wooden stick. Figure 9 The first column shows the pressure distribution cloud map directly acquired by the fabric pressure sensor array 1, including crosstalk effects. The second column shows the pressure distribution cloud map output by the trained improved U-Net model. The third column shows the pressure distribution cloud map acquired by the RX-M3232L piezoresistive flexible thin-film pressure sensor pad without crosstalk effects. Figure 9 As can be seen, the pressure distribution cloud map directly acquired by the fabric pressure sensor array 1 has a poor display effect, especially under the pressure of irregularly shaped palms and soles, where the pressure distribution display results are significantly affected by crosstalk. However, after correction by the improved U-Net convolutional neural network model, a pressure distribution cloud map display effect similar to that of the crosstalk-free RX-M3232L type piezoresistive flexible film pressure sensor pad can be obtained. This indicates that the established improved U-Net convolutional neural network model has achieved good training results and effectively eliminates the influence of crosstalk in the fabric pressure sensor array 1 on the pressure cloud map display results.

[0070] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for correcting the detection results of a fabric pressure sensor array, characterized in that, Specifically, the steps include the following: 1) Construct a U-Net convolutional neural network model, and modify the U-Net convolutional neural network model into an image-to-image regression prediction model to obtain an improved U-Net convolutional neural network model; the modification includes: based on the U-Net convolutional neural network model, improving the final... Softmax Normalized layers and Segmentation Delete the category layer, then add one. Regression The regression layer is connected to the last convolutional layer, and the number of filters in the last convolutional layer is changed to 3; 2) Place the fabric pressure sensor array directly above the pressure sensor array with accurate detection results, and simultaneously collect pressure distribution cloud maps under the same pressure multiple times. 3) Use the median filtering algorithm to process outliers in each pressure cloud image acquired by the fabric pressure sensor array; 4) Use the processed pressure distribution cloud map obtained in step 3) as the input image of the improved U-Net convolutional neural network model established in step 1), and use the pressure distribution cloud map under the same pressure obtained in step 2) collected by the pressure sensing array with accurate detection results as the output image of the improved U-Net convolutional neural network model. Use data augmentation functions to expand the input and output images to construct a training dataset for training the improved U-Net convolutional neural network model. 5) Train the improved U-Net convolutional neural network model using the training dataset until it converges; 6) Measure the pressure distribution of the fabric using a fabric pressure sensor array, and input the collected pressure distribution cloud map into the improved U-Net convolutional neural network model trained in step 5) to output the corrected pressure distribution cloud map; The fabric pressure sensor array fabrication method in step 2): The first silver-plated conductive yarn (2) and the first non-conductive polyester yarn (7) are symmetrically sewn in an array along the warp direction of the fabric base (5) on the front and back sides of the fabric base (5). The array spacing of the first silver-plated conductive yarn (2) and the first non-conductive polyester yarn (7) is 10 mm. The first silver-plated conductive yarn (2) and the first non-conductive polyester yarn (7) have a first locking buckle (6) for interlocking connection at the intersection of the non-warp and weft arrays. Subsequently, a composite solution consisting of polyurethane / N,N-dimethylformamide solution and conductive carbon black was printed onto the first silver-plated conductive yarn (2) and the fabric nearby using screen printing and then dried to form a pressure sensing unit array. Finally, the second silver-plated conductive yarn (4) and the second non-conductive polyester yarn (8) are symmetrically sewn in an array along the weft direction of the fabric substrate (5) on the front and back sides of the fabric substrate. The array spacing of the second silver-plated conductive yarn (4) and the second non-conductive polyester yarn (8) is 10 mm. The second silver-plated conductive yarn (4) and the second non-conductive polyester yarn (8) have a second locking buckle (9) for interlocking connection at the intersection of the non-warp and weft arrays. The pressure sensing unit (3) is located at the intersection of the first silver-plated conductive yarn (2) and the second silver-plated conductive yarn (4) on the front side of the fabric substrate (5), so that the conductive yarn and the sensing unit are connected by circuit, and a fabric pressure sensing array with 32×32 sensing units (3) is obtained.

2. The method for correcting the detection results of the fabric pressure sensor array according to claim 1, characterized in that, The specific implementation method of step 1) is as follows: The constructed U-Net convolutional neural network model performs four downsampling and four upsampling operations on the input pressure distribution cloud map. The network has a total of four layers. First, the input image is processed by 64 convolutional kernels and the ReLU activation function to extract image features. Then, it is downsampled sequentially, followed by convolution and activation by 128 convolutional kernels, and so on to complete four downsampling operations, resulting in 1024 8×8 feature layers. From bottom to top, the 1024 8×8 feature layers are deconvolved by 512 convolutional kernels to continuously enlarge the image. The downsampled image is used to add corresponding upsampled feature layers, and then convolution is performed again, and so on. Through four upsampling operations, the number of pixels in the feature image is increased, and the image size and accuracy are finally restored. The improved U-Net convolutional neural network model has a total of 33 layers. The input image size is 128×128×3. After multiple convolutions and pooling, the output image size is still 128×128×3.

3. The method for correcting the detection results of the fabric pressure sensor array according to claim 1, characterized in that, Step 3) involves processing outliers in each pressure cloud image acquired by the fabric pressure sensor array using a median filtering algorithm, and then applying the peak signal-to-noise ratio... PSNR To evaluate the effect of median filtering algorithm on the removal of outliers in pressure contour maps before and after processing.

4. The method for correcting the detection results of the fabric pressure sensor array according to claim 1, characterized in that, In step 5), during the training of the improved U-Net convolutional neural network model using the training dataset, the cross-entropy loss function is used. E To calculate the difference in performance between good and bad models during training: ， In the formula, X The input image to be processed. Ω represents the actual output probability, and Ω represents the image dataset. l Let Ω tend towards {1, ...,} K The actual label of each pixel at that time. K For natural numbers, w ( X The weight map is introduced and used to compensate for different frequencies of image pixels during training. ， In the formula, w c To balance the probability of different images appearing, d 1 represents the distance between the feature to be segmented in the image and the nearest boundary. d 2 represents the distance between the feature to be segmented in the image and the second nearest boundary. w 0 represents the index term weight update coefficient. σ The standard deviation is denoted as .

5. The method for correcting the detection results of the fabric pressure sensor array according to claim 4, characterized in that, In the process of training the improved U-Net convolutional neural network model using the training dataset, the training set is divided into mini-batches for training the model. The mini-batch stochastic gradient descent method is used to obtain the minimum value of the loss function, ensuring that as many samples as possible are input into the model in each round of iterative training.

Images