Tactile perception analysis method and system based on tensor decomposition and temporal convolution network

Abstract

The application relates to a kind of haptic perception analysis method and system based on tensor decomposition and time convolution network;Method includes: using the flexible tactile sensor covering whole hand and corresponding multi-channel pressure detection circuit to collect multidimensional dynamic hand pressure tactile data;Tensor decomposition is carried out on multidimensional dynamic hand pressure tactile data, and low-order tensor data is obtained;The tensor data is used to represent the haptic interaction process when hand and object interact;The tensor data is input into time convolution network to identify the haptic interaction process and output hand pressure distribution chart.The application realizes the accuracy identification of hand dynamic haptic interaction task.

Description

Technical Field

[0001] This invention belongs to the field of tactile perception technology, specifically relating to a tactile perception analysis method and system based on tensor decomposition and temporal convolutional networks. Background Technology

[0002] Touch is one of the five basic human senses, an important aspect of the somatosensory system, and a crucial channel for humans to perceive the external environment. Our hands provide us with various information about our surroundings. Through tactile feedback combined with proprioception, we are able to perform many daily tasks using our hands. Research shows that when the sense of touch is lost, human grasping movements become inaccurate and unstable.

[0003] Due to the complex curved shape of the human hand, limited installation space, and diverse objects to be manipulated, sensors made of traditional rigid materials are ill-suited for tactile data acquisition tasks. Compared to vision and hearing, the application of emerging machine learning in the field of tactile sensing is relatively underdeveloped. Understanding the high-dimensional tactile information during human interaction with external objects, extracting features and relationships from it, and then recognizing tactile interaction tasks requires high spatial and temporal resolution and large-scale tactile datasets as a foundation; however, tactile acquisition, analysis, and recognition systems that meet these requirements are still rare. Summary of the Invention

[0004] In view of the above analysis, the present invention aims to disclose a tactile perception analysis method and system based on tensor decomposition and temporal convolutional networks, which is used to solve the problems of pressure tactile acquisition, processing and display.

[0005] This invention discloses a tactile perception analysis method based on tensor decomposition and temporal convolutional networks, comprising:

[0006] A flexible tactile sensor covering the entire hand and a corresponding multi-channel pressure detection circuit are used to collect multi-dimensional dynamic hand pressure tactile data;

[0007] Tensor decomposition is performed on multidimensional dynamic hand pressure tactile data to obtain low-order tensor data; the tensor data is used to characterize the tactile interaction process when the hand interacts with an object.

[0008] The tensor data is input into a temporal convolutional network to identify the tactile interaction process and output a hand pressure distribution map.

[0009] Furthermore, methods for tensor decomposition of multidimensional dynamic hand pressure tactile data include:

[0010] 1) Organize multidimensional dynamic hand pressure tactile data into a three-dimensional data array;

[0011] 2) Data modeling of the three-dimensional data array is performed using a tensor decomposition model;

[0012] 3) After fitting and estimating the modeling data, third-order tensor data representing the tactile interaction process when the hand interacts with the object are obtained.

[0013] Furthermore, the three axes of the three-dimensional data array correspond to a single flexible tactile sensor, the test time, and the number of tests, respectively.

[0014] Furthermore, the tensor decomposition model decomposes the three-dimensional data array into a set of three-factor matrices;

[0015] The three-factor matrix includes a sensor factor matrix W, a time factor matrix B, and an experiment factor matrix A; wherein the dimension of the sensor factor matrix W is N×R; and the sensor factors in matrix W are represented as follows:

[0016] The dimension of the time factor matrix B is T×R; the time factors in matrix B are represented as follows:

[0017] The dimension of the experimental factor matrix A is K×R; the experimental factors in matrix A are represented as follows:

[0018] n = 1, ..., N; t = 1, ..., T; k = 1, ..., K; r = 1, ..., R; N is the total number of flexible tactile sensors; T is the total duration of the experiment; K is the total number of experiments; R is the rank of the third-order tensor decomposition.

[0019] In the third-order tensor, represents the data collected by sensor n at time t in experiment k.

[0020] Furthermore, based on the three-factor matrix, the tensor decomposition model models the data from the k-th trial as follows:

[0021]

[0022] Among them, Diag(a k )Will Embedded as diagonal terms of an R×R matrix, It is an N×T data matrix.

[0023] Furthermore, the data from each trial is decomposed into a third-order tensor to obtain the data acquisition matrix X for the k-th trial. k ;X k It is an N×T matrix;

[0024] For data matrix X k and Alternating least squares fitting is used to minimize the square of the reconstruction error as the optimization objective. The three-factor matrix included will be fitted to the desired result. Third-order tensor data to characterize the tactile interaction process when the hand interacts with an object.

[0025] Furthermore, the process of determining the rank in the third-order tensor decomposition includes:

[0026] 1) Using multiple sets of original multidimensional dynamic hand pressure tactile data, tensor decomposition models of different ranks R from 1 to high were established to obtain multiple sets of third-order tensor data under different ranks R.

[0027] 2) For the third-order tensor data under each rank, data fitting is performed to obtain the reconstructed multidimensional dynamic hand pressure tactile data;

[0028] 3) Statistical analysis was performed on the errors between the fitted multidimensional dynamic hand pressure tactile data at different ranks and the original multidimensional dynamic hand pressure tactile data;

[0029] 4) The smallest rank after the error stabilizes is taken as the tensor decomposition rank.

[0030] Furthermore, the temporal convolutional network is a one-dimensional convolutional block temporal convolutional network (TCN).

[0031] The main structure of the Temporal Convolutional Network (TCN) includes causal convolution, dilated convolution, and residual blocks; the tensor data is used as input, and the output is a hand pressure distribution map.

[0032] The present invention also discloses a tactile perception analysis system based on tensor decomposition and temporal convolutional network, comprising: a host computer, a slave computer, a flexible tactile sensor, and a multi-channel pressure detection circuit;

[0033] Flexible tactile sensors are used to sense tactile signals from the hand.

[0034] The multi-channel pressure detection circuit is connected to the flexible tactile sensor to detect the sensed hand tactile signals and obtain multi-dimensional hand pressure tactile signals, which are then transmitted to the lower-level computer.

[0035] The lower-level computer is used to receive multi-dimensional hand pressure tactile signals, select, amplify, filter, and store them to obtain multi-dimensional dynamic hand pressure tactile data; and wait for the multi-dimensional dynamic hand pressure tactile data of the whole hand to be collected before transmitting it to the upper-level computer.

[0036] The host computer is used to set parameters to control the detection of multiple pressure detection circuits; to transmit information with the slave computer and receive multi-dimensional dynamic hand pressure tactile data; and to use the tactile perception analysis method based on tensor decomposition and temporal convolutional network as described in any one of claims 1-8 to perform tensor decomposition and tactile interaction process recognition on the multi-dimensional dynamic hand pressure tactile data to obtain a hand pressure distribution map, which is displayed in real time on the human-computer interaction interface.

[0037] Furthermore, the flexible tactile sensor is an array-distributed flexible thin-film pressure sensor attached to a wearable human device; it includes three parts, namely the index finger-middle finger-ring finger-little finger, thumb, and palm; each part is electrically connected to a multi-channel pressure detection circuit.

[0038] The present invention can achieve the following beneficial effects:

[0039] This invention discloses a tactile perception analysis method based on tensor decomposition and temporal convolutional networks. It uses tensor decomposition for dimensionality reduction and temporal convolutional networks to identify tactile interaction processes from large-scale, high-resolution dynamic hand pressure tactile data. This reveals tactile features representing human grasping strategies, such as spatial correlations between finger regions and circuit dynamics; and achieves accurate identification of dynamic tactile interaction tasks. Attached Figure Description

[0040] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0041] Figure 1 This is a flowchart of the tactile perception analysis method based on tensor decomposition and temporal convolutional networks in an embodiment of the present invention.

[0042] Figure 2 This is a schematic diagram of the tensor decomposition process in an embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of the temporal convolutional network structure in an embodiment of the present invention;

[0044] Figure 4 This is a block diagram of a tactile perception analysis system based on tensor decomposition and temporal convolutional networks in an embodiment of the present invention.

[0045] Figure 5 This is a schematic block diagram showing the components and connections of the lower-level machine in an embodiment of the present invention. Detailed Implementation

[0046] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and, together with the embodiments of the present invention, serve to illustrate the principles of the present invention.

[0047] Example 1

[0048] One embodiment of the present invention discloses a tactile perception analysis method based on tensor decomposition and temporal convolutional networks, such as... Figure 1 As shown, it includes:

[0049] Step S1: Use a flexible tactile sensor covering the entire hand and a corresponding multi-channel pressure detection circuit to collect multi-dimensional dynamic hand pressure tactile data;

[0050] Step S2: Perform tensor decomposition on the multidimensional dynamic hand pressure tactile data to obtain low-order tensor data; the tensor data is used to characterize the tactile interaction process when the hand interacts with the object.

[0051] Step S3: Input the tensor data into a temporal convolutional network to identify the tactile interaction process and output a hand pressure distribution map.

[0052] Specifically, in step S1, the flexible tactile sensor is an array-distributed flexible thin-film pressure sensor attached to a wearable device for the human body; it includes three parts, namely the index finger, middle finger, ring finger, little finger, thumb, and palm.

[0053] The flexible tactile sensor is electrically connected to a multi-channel pressure detection circuit. The multi-channel pressure detection circuit is used to detect the hand tactile signal from the flexible tactile sensor in real time to obtain multi-dimensional hand pressure tactile signals, and then perform signal conditioning to obtain multi-dimensional hand pressure tactile data.

[0054] By using a flexible tactile sensor covering the entire hand and a corresponding multi-channel pressure detection circuit, the bending motion of the fingers is detected. This enables both joint bending activities and the control of high-precision, fine movements, as well as the acquisition of high-resolution, large-scale tactile data.

[0055] Specifically, such as Figure 2 As shown, in step S2, the method for tensor decomposition of multidimensional dynamic hand pressure tactile data includes:

[0056] 1) Organize multidimensional dynamic hand pressure tactile data into a three-dimensional data array;

[0057] The three axes of the three-dimensional data array correspond to a single flexible tactile sensor, the test time, and the number of tests, respectively.

[0058] 2) Data modeling of the three-dimensional data array is performed using a tensor decomposition model;

[0059] The tensor decomposition model decomposes a three-dimensional data array into a set of three-factor matrices;

[0060] The three-factor matrix includes the sensor factor matrix W; the time factor matrix B; and the test factor matrix A.

[0061] The sensor factor matrix W has a dimension of N×R; the sensor factors in matrix W are represented as follows:

[0062] The dimension of the time factor matrix B is T×R; the time factors in matrix B are represented as follows:

[0063] The dimension of the experimental factor matrix A is K×R; the experimental factors in matrix A are represented as follows:

[0064] n = 1, ..., N; t = 1, ..., T; k = 1, ..., K; r = 1, ..., R; N is the total number of flexible tactile sensors; T is the total duration of the experiment; K is the total number of experiments; R is the rank of the third-order tensor;

[0065] In the third-order tensor, represents the data collected by sensor n at time t in experiment k.

[0066] Based on the three-factor matrix, the tensor decomposition model models the data for each trial as follows:

[0067]

[0068] Among them, Diag(a k ) will a k Embedded as diagonal terms of an R×R matrix; It is an N×T data matrix.

[0069] Decompose the third-order tensor using the data from each trial to obtain X. k An N×T matrix containing tensor data from trial k;

[0070] 3) Fit and estimate the modeling data to obtain third-order tensor data that characterizes the tactile interaction process when the hand interacts with the object.

[0071] The data from each trial is decomposed into a third-order tensor to obtain the data matrix X of the k-th trial. k ;X k It is an N×T matrix;

[0072] For data matrix X k and Alternating least squares fitting is used to minimize the square of the reconstruction error as the optimization objective. The three-factor matrix included will be fitted to the desired result. Third-order tensor data to characterize the tactile interaction process when the hand interacts with an object.

[0073] The goal is to minimize the square of the reconstruction error:

[0074]

[0075] The estimated value of the factor matrix is ​​obtained by alternating least squares (ALS). The motivation of ALS is to fix two factor matrices and optimize them as a least squares subproblem, which is convex and has a closed-form solution.

[0076] Preferably, by determining the size of the rank R, all tactile interaction processes can be described by a linear combination of an appropriate number of low-order components.

[0077] Specifically, the rank determination process of the third-order tensor decomposition includes:

[0078] 1) Using multiple sets of original multidimensional dynamic hand pressure tactile data, tensor decomposition models of different ranks R from 1 to high were established to obtain multiple sets of third-order tensor data under different ranks R.

[0079] 2) For the third-order tensor data under each rank, data fitting is performed to obtain the reconstructed multidimensional dynamic hand pressure tactile data;

[0080] The data fitting method used for reconstruction can be any existing fitting method, which is not limited here and does not affect the scope of protection of this invention. For example, using... Perform fitting.

[0081] 3) Statistical analysis was performed on the errors between the fitted multidimensional dynamic hand pressure tactile data at different ranks and the original multidimensional dynamic hand pressure tactile data;

[0082] The statistical error is an error fraction. χ represents the original multidimensional dynamic hand pressure tactile data; Reconstructed multidimensional dynamic hand pressure tactile data.

[0083] 4) The smallest rank after the error stabilizes is taken as the tensor decomposition rank;

[0084] The error stabilizing means that the fluctuation of the error is less than the set threshold. Through theoretical analysis and specific tests, it can be seen that as the rank R increases, the error counted in step 3) gradually decreases and then stabilizes, fluctuating within a certain range.

[0085] The rank determined through the above process satisfies both the requirement of minimizing the reconstruction error of the third-order tensor data and ensuring a minimum data size, thus reducing the data burden on subsequent processing.

[0086] In step S3, for tactile interaction tasks with high spatial resolution and long historical information, a Temporal Convolutional Network (TCN) with one-dimensional convolutional blocks is selected as the neural network. Using a TCN to recognize tactile interaction tasks offers high accuracy and strong generalization performance. Compared to traditional methods that rely on manual feature extraction and machine learning classifiers, TCN offers higher accuracy and is easier to use.

[0087] Specifically, the Temporal Convolutional Network (TCN) takes the tensor data obtained in step S2 as input and outputs a hand pressure distribution map.

[0088] The main structure of a Temporal Convolutional Network (TCN) includes causal convolution, dilated convolution, and residual blocks.

[0089] Temporal convolutions are built on two principles: the network produces an output that is the same length as the input, and information cannot be passed from the future to the past. To achieve the first point, TCN uses a one-dimensional convolutional network architecture.

[0090] In this architecture, the length of each hidden layer is equal to the length of the input layer, and the length of subsequent layers is the length of the input layer after adding zero padding to the previous layers. To achieve the second point, TCN uses causal convolution. The output at time t is the convolution of the elements of the previous layer up to time t. Unlike traditional convolutional neural networks, causal convolution cannot see future data. This is called causal convolution because it is a strictly time-constrained model. That is, TCN = 1D FCN + causal convolution.

[0091] The causal convolution processing is achieved by shifting the output of a regular convolution by several time steps.

[0092] The simple stacking of causal convolutions results in a linear relationship between the length of the captured sequence and the network depth, which is not conducive to applying this structure to sequence tasks. To address this limitation, TCN also introduces a dilated convolution structure to provide the receptive field of the causal convolution processing with several orders of magnitude greater receptive field. For a one-dimensional input sequence... And a filter f: The dilated convolution operation F on sequence elements s is defined as:

[0093]

[0094] Where d is the dilation factor, k is the filter size, and sd·i indicates the past direction. Therefore, dilation is equivalent to introducing a fixed stride between every two adjacent filters. If d = 1, the dilated convolution becomes a normal convolution. A larger dilation factor allows the upper layer output to represent a wider range of inputs, effectively increasing the network's receptive field. This gives two ways to increase the receptive field of a TCN: choosing a larger filter size k and increasing the dilation factor d.

[0095] Residual connections allow the network to pass information across layers. In this embodiment, the Temporal Convolutional Network (TCN) creates a residual block to replace the convolutional layers. A residual block contains two layers of weights that have undergone weight-normalized convolutions and ReLU nonlinear mappings. Furthermore, the TCN adds Dropout after each dilated convolution within the residual block for regularization.

[0096] o = Activation(x + F(x))

[0097] Here, F represents a series of transformations, x is the input vector of the residual block, and o is the output of the residual block. This has been repeatedly proven to be beneficial for very deep networks. In general residual connections, the input is directly added to the output, but in TCN, the input and output can have different widths. To address this difference, TCN uses an additional 1×1 convolution to ensure that element-wise addition ⊕ results in tensors of the same shape.

[0098] More specifically, in the Temporal Convolutional Network (TCN), the size of the receptive field is determined based on the amount of input third-order tensor data, the number of extended convolutional layers, the dilation factor corresponding to each extended convolutional layer, the size of the convolutional kernel, and the number of stacks.

[0099] The specific formula is as follows:

[0100] R field =1+2·(K) size -1)·N stack ·∑ i d i ;

[0101] Among them, R field To determine the size of the sensing field, K size N is the size of the convolution kernel; stack d represents the number of stacks; i is the dilation factor corresponding to the i-th extended convolutional layer.

[0102] When the input third-order tensor data is 1500, the specific structural parameters of the temporal convolutional network TCN in this embodiment are as follows:

[0103]

[0104] The network structure of a temporal convolutional network is as follows: Figure 3 As shown.

[0105] Specifically, the hand pressure distribution map output in step S3 can reflect the tactile interaction process.

[0106] The hand pressure distribution map can guide improvements to the tactile sensing device. The low-dimensional representation of the spatial distribution allows for adjusting the sensor point density at corresponding hand locations, while the temporal dynamics can guide the search for a suitable sampling frequency. This helps determine important parameters such as sensor density, resolution, location, and bandwidth, leading to a more rational sensor layout.

[0107] In summary, the tactile perception analysis method based on tensor decomposition and temporal convolutional networks in this embodiment has the following effects:

[0108] 1. By using a flexible tactile sensor covering the entire hand and a corresponding multi-channel pressure detection circuit, the bending motion of the fingers is detected, which can realize both joint bending activities and high-precision fine motion control, and achieve the acquisition of high-resolution large-scale tactile data.

[0109] 2. Tensor decomposition was used to obtain the spatial distribution, content and manner of external information in hand movements and dynamic patterns during human-object interaction from the complex high-dimensional dynamics of hand pressure tactile data;

[0110] 3. By finding a suitable rank, all tactile interaction processes can be described by a linear combination of an appropriate number of low-order components.

[0111] 4. The results of tensor decomposition can guide improvements to tactile sensing devices. The low-dimensional representation of spatial distribution allows for adjustments to sensor point density at corresponding hand locations, and temporal dynamics can guide the search for suitable sampling frequencies. Such research may help determine important parameters such as sensor density, resolution, location, and bandwidth.

[0112] 5. For tactile interaction tasks with high spatial resolution and long historical information, temporal convolutional networks are used for tactile interaction task recognition. This approach achieves high accuracy while also exhibiting strong generalization performance. Compared to traditional methods that rely on manual feature extraction followed by machine learning classifiers, this method offers higher accuracy and ease of use.

[0113] Example 2

[0114] This embodiment discloses a tactile perception analysis system based on tensor decomposition and temporal convolutional networks, such as... Figure 4 As shown, it includes: a host computer, a slave computer, a flexible tactile sensor, and a multi-channel pressure detection circuit;

[0115] Flexible tactile sensors are used to sense tactile signals from the hand.

[0116] The multi-channel pressure detection circuit is connected to the flexible tactile sensor to detect the sensed hand tactile signals and obtain multi-dimensional hand pressure tactile signals, which are then transmitted to the lower-level computer.

[0117] The lower-level computer is used to receive multi-dimensional hand pressure tactile signals, select, amplify, filter, and store them to obtain multi-dimensional dynamic hand pressure tactile data; and wait for the multi-dimensional dynamic hand pressure tactile data of the whole hand to be collected before transmitting it to the upper-level computer.

[0118] The host computer is used to manually set parameters to control the acquisition of hand pressure tactile signals: it transmits information with the slave computer and receives multi-dimensional dynamic hand pressure tactile data; it uses the tactile perception analysis method based on tensor decomposition and temporal convolutional network from the previous embodiment to perform tensor decomposition on the multi-dimensional dynamic hand pressure tactile data to obtain low-order tensor data; the tensor data is input into the temporal convolutional network to identify the tactile interaction process and obtain a hand pressure distribution map that is displayed in real time on the human-computer interaction interface.

[0119] Specifically, the flexible tactile sensor is an array-distributed flexible thin-film pressure sensor attached to a wearable human device; it includes three parts, namely the index finger-middle finger-ring finger-little finger, thumb, and palm; each part is electrically connected to a multi-channel pressure detection circuit, which collects tactile signals in real time and transmits the collected data to the lower-level machine in real time.

[0120] like Figure 5 As shown, the lower-level machine includes a digital output module, a storage module, an analog input module, a signal conditioning module, and a USB communication module. The upper-level machine controls the STM32 lower-level machine's data acquisition via C++ programming. The digital output module and analog input module, combined with a multi-channel pressure detection circuit, acquire high-dimensional hand pressure tactile data from the flexible tactile sensor. The signal conditioning module performs selection, amplification, and filtering operations on the acquired tactile data. The storage module transmits the acquired hand pressure tactile data to the upper-level machine via the USB communication module after the entire hand pressure tactile data acquisition is complete.

[0121] The main processor of the lower-level machine uses STMicroelectronics (ST) chips of the STM32F407 series. It uses an array scanning method to collect sensor points in three parts: four fingers, thumb, and palm. It consists of an STM32 embedded microcontroller (including digital output module, analog input module, signal conditioning module, storage module and USB communication module) and a multi-channel pressure detection circuit.

[0122] The host computer is used to manually set parameters to control the acquisition of hand pressure tactile signals: specifically, the setting of parameters such as sampling mode, sampling frequency, sampling gain, and number of sampling points; the host computer transmits information with the slave computer through a USB communication module and displays the results of dimensionality reduction analysis and interactive recognition in real time on the human-computer interaction interface.

[0123] The same technical details and beneficial effects as in the previous embodiment are described in the previous embodiment, and will not be repeated here.

[0124] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A tactile perception analysis method based on tensor decomposition and temporal convolutional networks, characterized in that, include: A flexible tactile sensor covering the entire hand and a corresponding multi-channel pressure detection circuit are used to collect multi-dimensional dynamic hand pressure tactile data; Tensor decomposition is performed on multidimensional dynamic hand pressure tactile data to obtain low-order tensor data; the tensor data is used to characterize the tactile interaction process when the hand interacts with an object. The tensor data is input into a temporal convolutional network to identify the tactile interaction process and output a hand pressure distribution map. Methods for tensor decomposition of multidimensional dynamic hand pressure tactile data include: 1) Organize multidimensional dynamic hand pressure tactile data into a three-dimensional data array; 2) Data modeling of the three-dimensional data array is performed using a tensor decomposition model; 3) After fitting and estimating the modeling data, third-order tensor data representing the tactile interaction process when the hand interacts with the object are obtained; The tensor decomposition model decomposes a three-dimensional data array into a set of three-factor matrices; The three-factor matrix includes the sensor factor matrix. W Time factor matrix B and the experimental factor matrix A; Among them, the sensor factor matrix W The dimension is N × R ;matrix W The sensor factor is expressed as ; Time factor matrix B's The dimension is T×R; matrix B The time factor is expressed as ; The dimension of the experimental factor matrix A is K×R; the matrix A The experimental factor is expressed as ; n =1,…, N ; t =1,…, T ; k =1,…, K ; r= 1,…, R ; N This represents the total number of flexible tactile sensors. T This refers to the total duration of the experiment. K The total number of trials; R Let be the rank of the third-order tensor decomposition; In a third-order tensor, the experiment is represented. k Mid-time t Time sensor n Data collection .

2. The tactile perception analysis method according to claim 1, characterized in that, Based on the three-factor matrix, the tensor decomposition model will... k The data model for this experiment is as follows: ； in, Will Embedded as R × R The diagonal terms of the matrix, for N × T The data matrix.

3. The tactile perception analysis method according to claim 2, characterized in that, The data from each trial is decomposed into a third-order tensor to obtain the... k Data Matrix of this Experiment ; for N × T matrix; For data matrix and Alternating least squares fitting is used to minimize the square of the reconstruction error as the optimization objective. The three-factor matrix included will be fitted to the desired result. Third-order tensor data to characterize the tactile interaction process when the hand interacts with an object.

4. The tactile perception analysis method according to claim 1, characterized in that, The process of determining the rank in the third-order tensor decomposition includes: 1) Multiple sets of raw, multidimensional, dynamic hand pressure tactile data were used, with rank... R Tensor decomposition models of different ranks from 1 to higher are established, and tensor decomposition is performed to obtain tensor models of different ranks. R Multiple sets of third-order tensor data; 2) For the third-order tensor data under each rank, data fitting is performed to obtain the reconstructed multidimensional dynamic hand pressure tactile data; 3) Statistical analysis was performed on the errors between the fitted multidimensional dynamic hand pressure tactile data at different ranks and the original multidimensional dynamic hand pressure tactile data; 4) The smallest rank after the error stabilizes is taken as the tensor decomposition rank.

5. The tactile perception analysis method according to claim 1, characterized in that, The temporal convolutional network is a one-dimensional convolutional block temporal convolutional network (TCN). The main structure of the Temporal Convolutional Network (TCN) includes causal convolution, dilated convolution, and residual blocks; the tensor data is used as input, and the output is a hand pressure distribution map.

6. A tactile perception analysis system based on tensor decomposition and temporal convolutional networks, characterized in that, include: Host computer, slave computer, flexible tactile sensor and multi-channel pressure detection circuit; Flexible tactile sensors are used to sense tactile signals from the hand. The multi-channel pressure detection circuit is connected to the flexible tactile sensor to detect the sensed hand tactile signals and obtain multi-dimensional hand pressure tactile signals, which are then transmitted to the lower-level computer. The lower-level computer is used to receive multi-dimensional hand pressure tactile signals, select, amplify, filter, and store them to obtain multi-dimensional dynamic hand pressure tactile data; and wait for the multi-dimensional dynamic hand pressure tactile data of the whole hand to be collected before transmitting it to the upper-level computer. The host computer is used to set parameters to control the detection of multiple pressure detection circuits; it also transmits information with the slave computer and receives multi-dimensional dynamic hand pressure tactile data. Using the tactile perception analysis method based on tensor decomposition and temporal convolutional networks as described in any one of claims 1-5, tensor decomposition and tactile interaction process identification are performed on multidimensional dynamic hand pressure tactile data to obtain a hand pressure distribution map, which is then displayed in real time on the human-computer interaction interface.

7. The tactile perception analysis system according to claim 6, characterized in that, The flexible tactile sensor is an array-distributed flexible thin-film pressure sensor attached to a wearable device on the human body; it consists of three parts, namely the index finger-middle finger-ring finger-little finger, thumb, and palm; each part is electrically connected to a multi-channel pressure detection circuit.

Images