A tactile texture recognition method, system and terminal based on biological interpretability

Abstract

The present invention discloses a tactile texture recognition method, system and terminal based on biological interpretability, the method comprising: constructing a tactile sensor according to the distribution information of a biosensor and collecting tactile data; preprocessing the tactile data to obtain an expected sparse pulse signal and a target sparse pulse signal, and inputting the spurious neural network model for iterative reverse training to optimize the spurious neural network model; converting the task tactile data into a task sparse pulse signal and then inputting the spurious neural network model for forward propagation to output a recognition result. The present invention extracts features from multi-channel tactile data through an adaptive graph topology convolution kernel to obtain pulse signals with different labels, and trains a spurious neural network model according to the difference between the two and preset supervised learning rules, iteratively converges the weight parameters of the spurious neural network model, and finally obtains a network model for tactile texture recognition tasks in a convenient and efficient manner.

Description

Technical Field

[0001] The present invention relates to the technical field of tactile texture recognition, and in particular to a tactile texture recognition method, system, terminal and computer-readable storage medium based on biological interpretability. Background Art

[0002] At present, robotics technology plays an indispensable role in improving social productivity, and the bionic tactile sensing system is an important component of the robot's perception of the external physical environment and feedback regulation. It includes the key technologies of the system's ability to imitate the biological perception of the external environment. In order to imitate the operation of the biological peripheral and central neural systems with artificial neural networks, bionic classifiers represented by artificial neural networks (ANN) are widely used in the field of pattern recognition.

[0003] However, traditional ANNs cannot imitate the nonlinearity and spatiotemporal characteristics of the biological brain very well, and are biologically inaccurate. Their signal patterns are continuous analog signals rather than nerve impulse signals. Therefore, contrary to the low power consumption of the human tactile perception process, the use of traditional ANNs and training algorithms represented by Error Back Propagation (BP) will generate large computational overhead and energy consumption.

[0004] In addition, unlike ANN, the spiking neurons in SNN (Spiking Neural Network) are only activated when sending or receiving pulses, which greatly reduces the computational energy consumption of SNN in inference and prediction, and its energy consumption can be dynamically adjusted by adjusting the membrane potential threshold. However, due to the long length of the data set converted into a pulse signal, especially for high-density signals such as event signals, the computational cost of training an SNN is much higher than that of ANN.

[0005] Therefore, the prior art still needs to be improved and developed. Summary of the invention

[0006] The main purpose of the present invention is to provide a tactile texture recognition method, system, terminal and computer-readable storage medium based on biological interpretability, aiming to solve the problems of low tactile texture recognition accuracy and low computational efficiency caused by preprocessing of tactile data obtained by irregularly connected receptors in the prior art, as well as excessively high training costs caused by training a SNN.

[0007] To achieve the above object, the present invention provides a tactile texture recognition method based on biological interpretability, and the tactile texture recognition method based on biological interpretability comprises the following steps:

[0008] Acquire distribution information of a target biosensor, connect a plurality of constructed tactile sensors according to a specific topological structure according to the distribution information, and receive a plurality of tactile data collected by the plurality of tactile sensors;

[0009] Constructing a simple neuron model, inputting a plurality of the tactile data into the simple neuron model, and outputting a plurality of sparse pulse signals;

[0010] Constructing an adaptive convolutional network, inputting the plurality of sparse pulse signals into the adaptive convolutional network for feature extraction, and outputting a plurality of corresponding desired sparse pulse signals with desired labels;

[0011] Constructing a spiking neural network model, inputting a plurality of the desired sparse pulse signals into the spiking neural network model, and outputting a plurality of corresponding target sparse pulse signals with actual labels;

[0012] Inputting a plurality of the desired sparse pulse signals and a plurality of the target sparse pulse signals into the pulse neural network model for iterative reverse training until a weight parameter of the pulse neural network model is within a preset threshold range, thereby obtaining an optimized target pulse neural network model;

[0013] The task tactile data in the recognition task is acquired, the task tactile data is converted into a task sparse pulse signal, the task sparse pulse signal is input into the target pulse neural network model, and the recognition result is output.

[0014] Optionally, in the tactile texture recognition method based on biological interpretability, the distribution information includes: a plurality of node information and a plurality of node connection information;

[0015] The step of acquiring distribution information of a target biosensor, connecting a plurality of constructed tactile sensors according to a specific topological structure according to the distribution information, and receiving a plurality of tactile data collected by the plurality of tactile sensors specifically includes:

[0016] Acquire the plurality of node information and the plurality of node connection information of the target biosensor, and determine the specific topological structure according to the plurality of node information and the plurality of node connection information;

[0017] The plurality of constructed tactile sensors are connected according to the specific topological structure, and a plurality of tactile data of a plurality of channels collected by the plurality of tactile sensors are received.

[0018] Optionally, the tactile texture recognition method based on biological interpretability, wherein the step of constructing a simple neuron model, inputting a plurality of tactile data into the simple neuron model, and outputting a plurality of sparse pulse signals, specifically comprises:

[0019] Construct a simple neuron model, input all the tactile data into the simple neuron model, and output the membrane potential corresponding to each tactile data:

[0020] ;

[0021] ;

[0022] like ;

[0023] in, Indicates at time The membrane potential, Indicates at the current moment The membrane potential, Indicates at time The recovery variable, Indicates the membrane potential at the current moment The recovery variable, represents the membrane potential recovery speed, Indicates at time The input current, Indicates at the current moment The input current, The unit is millivolt;

[0024] If there are multiple membrane potentials that are not lower than the threshold, multiple pulses are released to obtain multiple sparse pulse signals corresponding to the multiple membrane potentials.

[0025] Optionally, the tactile texture recognition method based on biological interpretability, wherein the step of constructing an adaptive convolutional network, inputting a plurality of the sparse pulse signals into the adaptive convolutional network for feature extraction, and outputting a plurality of corresponding desired sparse pulse signals with desired labels, specifically comprises:

[0026] Constructing an adaptive convolutional network, and inputting the plurality of sparse pulse signals and the information of the specific topological structure into the adaptive convolutional network;

[0027] According to the specific topological structure, feature extraction is performed on all the sparse pulse signals in the adaptive convolutional network to obtain feature extraction results:

[0028] ;

[0029] ;

[0030] in, Indicates Output feature map, Indicates The number of features input to each vertex of the layer, Indicates The operational definition of an adaptive convolution kernel is: Indicates The first node of all nodes in the layer input features, represents the learnable bias vector, represents the degree of polynomial approximation adopted, Indicates Adaptive convolution kernels The coefficients of the polynomial filter, represents the highest power of the polynomial, represents the normalized adjacency matrix, express of Power;

[0031] According to the feature extraction result, the plurality of sparse pulse signals are labeled in the adaptive convolutional network, and a plurality of corresponding expected sparse pulse signals with expected classification labels are output.

[0032] Optionally, the method for tactile texture recognition based on biological interpretability, wherein the step of constructing a spiking neural network model, inputting a plurality of the desired sparse pulse signals into the spiking neural network model, and outputting a plurality of corresponding target sparse pulse signals with actual labels, specifically comprises:

[0033] Constructing a spiking neural network model, wherein the spiking neural network model includes a fully connected layer network with a preset number of layers, and creating multiple learning windows on all the fully connected layer networks;

[0034] Inputting a plurality of the desired sparse pulse signals into the spiking neural network model, and outputting a plurality of corresponding target sparse pulse signals with actual labels;

[0035] Wherein, each of the learning windows in the pulse neural network model processes one expected sparse pulse signal at a time.

[0036] Optionally, the tactile texture recognition method based on biological interpretability, wherein the multiple expected sparse pulse signals and the multiple target sparse pulse signals are input into the spiking neural network model for iterative reverse training until the weight parameters of the spiking neural network model are within a preset threshold range, thereby obtaining an optimized target spiking neural network model, specifically includes:

[0037] Determining difference information between a plurality of the desired sparse pulse signals and a plurality of corresponding target sparse pulse signals;

[0038] The difference information is input into the spiking neural network model for multiple times for back propagation, so as to iteratively train the spiking neural network model, and the weight parameters of the spiking neural network model are output multiple times until the weight parameters are within a preset threshold range, thereby obtaining weight change parameters:

[0039] ;

[0040] ;

[0041] ;

[0042] in, Indicates Layer Network Layer The neurons in the previous network layer The weight change parameter between neurons, Indicates learning efficiency, Indicates the deviation signal, Indicates The number of neurons in a layer, Indicates The previous network layer of the layer The actual pulse arrival time of each neuron is Indicates The previous network layer of the layer The expected pulse arrival time of each neuron is Indicated in The values ​​in different situations, Indicates The average time of pulse emission in the previous network layer of the layer, represents the number of neurons in the output layer, represents the interval that distinguishes target and non-target neurons, Represents the neuron that outputs the actual label;

[0043] ;

[0044] ;

[0045] ;

[0046] in, represents the gradient of the output layer error back propagation in the fully connected layer, represents the relationship between the postsynaptic spike firing time and weight, and They represent the proportional factors of control enhancement and suppression respectively, and represent the presynaptic and postsynaptic spike firing times, respectively. represents the first parameter controlling the weight update, represents the time interval coefficient for weight update, and Respectively represent the maximum weight and current weight of the weight constraint, represents the second parameter controlling the weight update;

[0047] ;

[0048] ;

[0049] in, Represents the recursive formula for the back propagation gradient of the hidden layer error in the fully connected layer, Indicates The previous network layer of the layer The actual pulse arrival time of each neuron is Indicates The previous network layer of the layer The actual pulse arrival time of each neuron;

[0050] The pulse neural network model is updated in real time according to the weight change parameters to obtain an optimized target pulse neural network model.

[0051] Optionally, the tactile texture recognition method based on biological interpretability, wherein the step of acquiring task tactile data in a recognition task, converting the task tactile data into a task sparse pulse signal, inputting the task sparse pulse signal into the target pulse neural network model, and outputting a recognition result, specifically includes:

[0052] Receiving a plurality of task tactile data collected by the tactile sensor, inputting the plurality of task tactile data into the simple neuron model and the adaptive convolutional network in sequence, and outputting a plurality of task sparse pulse signals with desired labels;

[0053] All the task sparse pulse signals with expected labels are input into the target pulse neural network model for forward propagation, and the recognition result of the recognition task is output.

[0054] In addition, to achieve the above-mentioned purpose, the present invention also provides a tactile texture recognition system based on biological interpretability, wherein the tactile texture recognition system based on biological interpretability includes:

[0055] A tactile data acquisition module, used to obtain distribution information of a target biosensor, connect a plurality of constructed tactile sensors according to a specific topological structure according to the distribution information, and receive a plurality of tactile data acquired by the plurality of tactile sensors;

[0056] A first pulse signal generating module, used for constructing a simple neuron model, inputting a plurality of tactile data into the simple neuron model, and outputting a plurality of sparse pulse signals;

[0057] A second pulse signal generating module is used to construct an adaptive convolutional network, input a plurality of the sparse pulse signals into the adaptive convolutional network for feature extraction, and output a plurality of corresponding desired sparse pulse signals with desired labels;

[0058] A third pulse signal generating module is used to construct a pulse neural network model, input a plurality of the desired sparse pulse signals into the pulse neural network model, and output a plurality of corresponding target sparse pulse signals with actual labels;

[0059] A model optimization module, used for inputting a plurality of the expected sparse pulse signals and a plurality of the target sparse pulse signals into the pulse neural network model for iterative reverse training until a weight parameter of the pulse neural network model is within a preset threshold range, thereby obtaining an optimized target pulse neural network model;

[0060] The tactile texture recognition module is used to obtain task tactile data in the recognition task, convert the task tactile data into a task sparse pulse signal, input the task sparse pulse signal into the target pulse neural network model, and output the recognition result.

[0061] In addition, to achieve the above-mentioned purpose, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a tactile texture recognition program based on bio-interpretability stored in the memory and executable on the processor, wherein the tactile texture recognition program based on bio-interpretability, when executed by the processor, implements the steps of the tactile texture recognition method based on bio-interpretability as described above.

[0062] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a tactile texture recognition program based on bio-interpretability, and when the tactile texture recognition program based on bio-interpretability is executed by a processor, the steps of the tactile texture recognition method based on bio-interpretability as described above are implemented.

[0063] In the present invention, the distribution information of the target biosensor is obtained, and according to the distribution information, the multiple constructed tactile sensors are connected according to a specific topological structure, and the multiple tactile data collected by the multiple tactile sensors are received; a simple neuron model is constructed, the multiple tactile data are input into the simple neuron model, and multiple sparse pulse signals are output; an adaptive convolutional network is constructed, the multiple sparse pulse signals are input into the adaptive convolutional network for feature extraction, and multiple corresponding expected sparse pulse signals with expected labels are output; a pulse neural network model is constructed, the multiple expected sparse pulse signals are input into the pulse neural network model, and multiple corresponding target sparse pulse signals with actual labels are output; the multiple expected sparse pulse signals and the multiple target sparse pulse signals are input into the pulse neural network model for iterative reverse training until the weight parameters of the pulse neural network model are within the preset threshold range, thereby obtaining an optimized target pulse neural network model; the task tactile data in the recognition task is obtained, the task tactile data is converted into a task sparse pulse signal, the task sparse pulse signal is input into the target pulse neural network model, and the recognition result is output. The present invention performs feature extraction on multi-channel tactile data through an adaptive graph topology convolution kernel to obtain pulse signals with different labels, and trains a pulse neural network model based on the difference between the two and preset supervised learning rules, and iteratively converges the weight parameters of the pulse neural network model, finally obtaining a network model for tactile texture recognition tasks in a convenient and efficient manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a flow chart of a preferred embodiment of the tactile texture recognition method based on biological interpretability of the present invention;

[0065] Figure 2 It is a network schematic diagram of a preferred embodiment of the tactile texture recognition method based on biological interpretability of the present invention;

[0066] Figure 3 It is a schematic diagram of feature extraction of a preferred embodiment of the tactile texture recognition method based on biological interpretability of the present invention;

[0067] Figure 4 It is a schematic diagram of the training process of a preferred embodiment of the tactile texture recognition method based on biological interpretability of the present invention;

[0068] Figure 5 It is a schematic diagram of the principle of a preferred embodiment of the tactile texture recognition system based on biological interpretability of the present invention;

[0069] Figure 6 Schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solution and advantages of the present invention clearer and more specific, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0071] The tactile texture recognition method based on biological interpretability described in the preferred embodiment of the present invention is as follows: Figure 1 As shown, the tactile texture recognition method based on biological interpretability includes the following steps:

[0072] Step S10: acquiring distribution information of the target biosensor, connecting the plurality of constructed tactile sensors according to a specific topological structure according to the distribution information, and receiving a plurality of tactile data collected by the plurality of tactile sensors.

[0073] By obtaining the distribution information of the target biosensors, such as the distribution and connection of finger sensors on human skin, the tactile sensors are linked according to the corresponding topological structure to collect multi-channel tactile data, so as to imitate the operation of biological nerve centers in the tactile perception process and improve the recognition accuracy of the perception system.

[0074] Specifically, multiple node information and multiple node connection information of the target biosensor are obtained, and the specific topological structure is determined according to the multiple node information and multiple node connection information; multiple constructed tactile sensors are connected according to the specific topological structure, and multiple tactile data of multiple channels collected by the multiple tactile sensors are received.

[0075] Among them, the distribution information includes: multiple node information and multiple node connection information. It aims to solve the problem of preprocessing tactile data obtained by irregularly connected receptors (linking the tactile sensors according to the corresponding topological structure based on the distribution information of the target biosensors). Since the general data processing methods are ineffective for processing tactile signals, multiple tactile sensors are linked according to the corresponding topological structure through a topological adaptive graph convolutional network. On the basis of high performance, it can also significantly reduce energy consumption and improve the recognition accuracy of tactile perception, making the entire method biologically interpretable.

[0076] Step S20, constructing a simple neuron model, inputting a plurality of the tactile data into the simple neuron model, and outputting a plurality of sparse pulse signals.

[0077] The collected tactile data is converted into an encodable pulse signal through a simple neuron model for training the pulse neural network model.

[0078] Specifically, a simple neuron model is constructed, all the tactile data are input into the simple neuron model, and the membrane potential corresponding to each tactile data is output:

[0079] ;

[0080] ;

[0081] like ;

[0082] in, Indicates at time The membrane potential, Indicates at the current moment The membrane potential, Indicates at time The recovery variable, Indicates the membrane potential at the current moment The recovery variable, represents the membrane potential recovery speed, Indicates at time The input current, Indicates at the current moment The input current, It represents the unit of millivolt; if there are multiple membrane potentials that are not lower than the threshold, multiple pulses are released to obtain multiple sparse pulse signals corresponding to the multiple membrane potentials.

[0083] Among them, Figure 2 As shown, the tactile data is converted into a pulse sequence through a simple neuron model. In this embodiment, the Izhikevich neuron model (a commonly used neuron model) is preferably used, which can improve biological interpretability and reduce the computational cost as much as possible. After obtaining the membrane potential, it can be converted into a pulse sequence by means of a threshold. For example, the threshold voltage is set to 35 If the membrane potential at a certain moment reaches or exceeds this threshold, the neuron immediately sends a pulse and resets its membrane potential, so that a pulse signal corresponding to the membrane potential signal (i.e., a sparse pulse signal) can be obtained.

[0084] Step S30, construct an adaptive convolutional network, input the multiple sparse pulse signals into the adaptive convolutional network for feature extraction, and output multiple corresponding expected sparse pulse signals with expected labels.

[0085] After encoding the multi-channel original tactile signals into pulse signals, an adaptive topological graph convolutional network is used to extract features from the sparse pulse signals corresponding to the tactile signals collected by all sensors and label the sparse pulse signals.

[0086] Specifically, an adaptive convolutional network is constructed, and information of a plurality of the sparse pulse signals and the specific topological structure is input into the adaptive convolutional network; according to the specific topological structure, feature extraction is performed on all the sparse pulse signals in the adaptive convolutional network to obtain a feature extraction result:

[0087] ;

[0088] ;

[0089] in, Indicates Output feature map, Indicates The number of features input to each vertex of the layer, Indicates The operational definition of an adaptive convolution kernel is: Indicates The first node of all nodes in the layer input features, represents the learnable bias vector, represents the degree of polynomial approximation adopted, Indicates Adaptive convolution kernels The coefficients of the polynomial filter, represents the highest power of the polynomial, represents the normalized adjacency matrix, express of power; according to the feature extraction result, labeling the multiple sparse pulse signals in the adaptive convolutional network, and outputting multiple corresponding expected sparse pulse signals with expected classification labels.

[0090] Among them, Figure 3 As shown, according to a specific topological structure, feature extraction is performed on all the sparse pulse signals, and they are labeled as expected sparse pulse signals with expected classification labels based on the feature extraction results. Based on this, the pulse neural network model can be trained, and sparse pulse signals with actual classification labels can be obtained in the voting layer. Based on the sparse pulse signals with different labels, the network can be trained to obtain a tactile texture recognition system with biological interpretability.

[0091] Step S40, constructing a pulse neural network model, inputting a plurality of the desired sparse pulse signals into the pulse neural network model, and outputting a plurality of corresponding target sparse pulse signals with actual labels.

[0092] After feature extraction, each sparse pulse signal will be trained through a three-layer fully connected layer network FC (Fully Connected Layer). Its neuron model is a special IF model (Integrate and Fire) and each neuron sends only one pulse within a learning window, further reducing computational overhead.

[0093] Specifically, a pulse neural network model is constructed, wherein the pulse neural network model includes a fully connected layer network with a preset number of layers, and multiple learning windows are created on all the fully connected layer networks; multiple expected sparse pulse signals are input into the pulse neural network model, and multiple corresponding target sparse pulse signals with actual labels are output; wherein each of the learning windows in the pulse neural network model processes one expected sparse pulse signal at a time.

[0094] Among them, the use of pulse neural networks can effectively reduce computing costs, and its biological accuracy is significantly higher than that of ANN (Artificial Neural Network). The pulse neural network is only activated when sending or receiving pulses, which greatly reduces its computing energy consumption in reasoning and prediction. Its energy consumption can also be dynamically adjusted by adjusting the membrane potential threshold, so it has a lower reasoning computing cost.

[0095] Step S50, inputting the multiple expected sparse pulse signals and the multiple target sparse pulse signals into the pulse neural network model for iterative reverse training until the weight parameters of the pulse neural network model are within a preset threshold range, thereby obtaining an optimized target pulse neural network model.

[0096] Among them, the update rule of the fully connected neural network weights in the pulse neural network model can adopt the more biologically interpretable STDP rule (Spike-Timing-Dependent Plasticity) to better simulate the memory and cognitive process of the human brain when receiving external stimuli (for example, a finger touching the surface of an object).

[0097] Specifically, the difference information between the multiple expected sparse pulse signals and the multiple corresponding target sparse pulse signals is determined; the difference information is input into the pulse neural network model for back propagation multiple times to iteratively train the pulse neural network model, and the weight parameters of the pulse neural network model are output multiple times until the weight parameters are within a preset threshold range, thereby obtaining the weight change parameters:

[0098] ;

[0099] ;

[0100] ;

[0101] in, Indicates Layer Network Layer The neurons in the previous network layer The weight change parameter between neurons, Indicates learning efficiency, Indicates the deviation signal, Indicates The number of neurons in a layer, Indicates The previous network layer of the layer The actual pulse arrival time of each neuron is Indicates The previous network layer of the layer The expected pulse arrival time of each neuron is Indicated in The values ​​in different situations, Indicates The average time of pulse emission in the previous network layer of the layer, represents the number of neurons in the output layer, represents the interval that distinguishes target and non-target neurons, Represents the neuron that outputs the actual label;

[0102] ;

[0103] ;

[0104] ;

[0105] in, represents the gradient of the output layer error back propagation in the fully connected layer, represents the relationship between the postsynaptic spike firing time and weight, and They represent the proportional factors of control enhancement and suppression respectively, and represent the presynaptic and postsynaptic spike firing times, respectively. represents the first parameter controlling the weight update, represents the time interval coefficient for weight update, and Respectively represent the maximum weight and current weight of the weight constraint, represents the second parameter controlling the weight update;

[0106] ;

[0107] ;

[0108] in, Represents the recursive formula for the back propagation gradient of the hidden layer error in the fully connected layer, Indicates The previous network layer of the layer The actual pulse arrival time of each neuron is Indicates The previous network layer of the layer The actual pulse arrival time of each neuron; the pulse neural network model is updated in real time according to the weight change parameter to obtain an optimized target pulse neural network model.

[0109] Among them, Figure 4 As shown in the figure, after obtaining the sparse pulse signals with the expected label and the actual label, according to the difference between the sparse pulse signals of the two labels, the pulse neural network is back-propagated multiple times to realize the iterative training of the pulse neural network model, until the weight parameter of the pulse neural network model is within the preset threshold, and it is defined as the weight change parameter, and then the pulse neural network model is updated in real time according to the weight change parameter; Among them, the learning rule based on spike time plasticity (STDP) describes the relationship between the synaptic connection strength and the time difference between the pre- and post-protrusion pulses, which is more biologically interpretable, can better describe the process of the human brain to establish memory and thinking, and reproduce the response pattern of people to certain external stimuli, and become a better solution for dealing with the strong interference and high-density data sets of the project; and through this learning rule, it is no longer necessary to update the weights between all synapses at each moment, but only update the weights of the synapses that are excited before and after in a specific time (learning window), which effectively solves the problem of high training cost of training a SNN, and on this basis, it can also effectively deal with high-density and high-noise data sets, with the advantages of stability and efficiency.

[0110] Step S60, acquiring task tactile data in the recognition task, converting the task tactile data into a task sparse pulse signal, inputting the task sparse pulse signal into the target pulse neural network model, and outputting the recognition result.

[0111] Specifically, multiple task tactile data collected by the tactile sensor are received, and the multiple task tactile data are input into the simple neuron model and the adaptive convolutional network in sequence, and multiple task sparse pulse signals with desired labels are output; all the task sparse pulse signals with desired labels are input into the target pulse neural network model for forward propagation, and the recognition result of the recognition task is output.

[0112] Among them, after completing the training of the spiking neural network model, it can be used for the recognition task of tactile texture recognition. After encoding and converting the tactile data obtained in the recognition task through the IF neuron, forward propagation is performed in the spiking neural network model. When the membrane potential reaches the specified threshold, the neuron generates a pulse, and the membrane potential returns to the resting potential. In this time-based encoding, the higher the synaptic weight of the connection and the more active the neuron, the earlier the pulse is sent in each time window, which corresponds to the excitatory neuron with a higher encoding frequency in the speed-based encoding:

[0113] ;

[0114] in, Indicates that the current neuron is at time The membrane potential at Indicates the current neuron at the current moment The membrane potential at represents the weight between the current neuron and all neurons in the previous layer, represents transpose, Indicates that the current neuron is All presynaptic spikes at that moment.

[0115] The present invention performs feature extraction on multi-channel tactile data through an adaptive graph topology convolution kernel to obtain pulse signals with different labels, and trains a pulse neural network model based on the difference between the two and preset supervised learning rules, and iteratively converges the weight parameters of the pulse neural network model, finally obtaining a network model for tactile texture recognition tasks in a convenient and efficient manner.

[0116] Furthermore, if Figure 5 As shown, based on the above-mentioned tactile texture recognition method based on biological interpretability, the present invention also provides a tactile texture recognition system based on biological interpretability, wherein the tactile texture recognition system based on biological interpretability includes:

[0117] A tactile data acquisition module 51 is used to obtain distribution information of a target biosensor, connect a plurality of constructed tactile sensors according to a specific topological structure according to the distribution information, and receive a plurality of tactile data collected by the plurality of tactile sensors;

[0118] A first pulse signal generating module 52, used for constructing a simple neuron model, inputting a plurality of tactile data into the simple neuron model, and outputting a plurality of sparse pulse signals;

[0119] A second pulse signal generating module 53 is used to construct an adaptive convolutional network, input the plurality of sparse pulse signals into the adaptive convolutional network for feature extraction, and output a plurality of corresponding desired sparse pulse signals with desired labels;

[0120] A third pulse signal generating module 54 is used to construct a pulse neural network model, input a plurality of the desired sparse pulse signals into the pulse neural network model, and output a plurality of corresponding target sparse pulse signals with actual labels;

[0121] A model optimization module 55 is used to input the plurality of expected sparse pulse signals and the plurality of target sparse pulse signals into the pulse neural network model for iterative reverse training until the weight parameter of the pulse neural network model is within a preset threshold range, thereby obtaining an optimized target pulse neural network model;

[0122] The tactile texture recognition module 56 is used to obtain task tactile data in the recognition task, convert the task tactile data into a task sparse pulse signal, input the task sparse pulse signal into the target pulse neural network model, and output the recognition result.

[0123] Furthermore, if Figure 6 As shown, based on the above-mentioned tactile texture recognition method and system based on biological interpretability, the present invention also provides a terminal accordingly, and the terminal includes a processor 10, a memory 20 and a display 30. Figure 6 Only some components of the terminal are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0124] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal. In other embodiments, the memory 20 may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the terminal. Further, the memory 20 may also include both an internal storage unit of the terminal and an external storage device. The memory 20 is used to store application software and various types of data installed in the terminal, such as the program code of the installation terminal. The memory 20 may also be used to temporarily store data that has been output or is to be output. In one embodiment, a tactile texture recognition program 40 based on biological interpretability is stored on the memory 20, and the tactile texture recognition program 40 based on biological interpretability can be executed by the processor 10, thereby realizing the tactile texture recognition method based on biological interpretability in the present application.

[0125] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor or other data processing chip, used to run the program code or process data stored in the memory 20, such as executing the tactile texture recognition method based on bio-interpretability.

[0126] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, etc. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.

[0127] In one embodiment, when the processor 10 executes the tactile texture recognition program 40 based on biological interpretability in the memory 20, the steps of the tactile texture recognition method based on biological interpretability as described above are implemented.

[0128] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a tactile texture recognition program based on biological interpretability, and when the tactile texture recognition program based on biological interpretability is executed by a processor, the steps of the tactile texture recognition method based on biological interpretability as described above are implemented.

[0129] In summary, the present invention provides a tactile texture recognition method based on biological interpretability and related equipment, the method comprising: obtaining distribution information of a target biosensor, connecting a plurality of constructed tactile sensors according to a specific topological structure according to the distribution information, and receiving a plurality of tactile data collected by the plurality of tactile sensors; constructing a simple neuron model, inputting a plurality of the tactile data into the simple neuron model, and outputting a plurality of sparse pulse signals; constructing an adaptive convolutional network, inputting a plurality of the sparse pulse signals into the adaptive convolutional network for feature extraction, and outputting a plurality of corresponding expected sparse pulse signals with expected labels; constructing a pulse neural network model, inputting a plurality of the expected sparse pulse signals into the pulse neural network model, and outputting a plurality of corresponding target sparse pulse signals with actual labels; inputting a plurality of the expected sparse pulse signals and a plurality of the target sparse pulse signals into the pulse neural network model for iterative reverse training until the weight parameters of the pulse neural network model are within a preset threshold range, thereby obtaining an optimized target pulse neural network model; obtaining task tactile data in a recognition task, converting the task tactile data into a task sparse pulse signal, inputting the task sparse pulse signal into the target pulse neural network model, and outputting a recognition result. The present invention performs feature extraction on multi-channel tactile data through an adaptive graph topology convolution kernel to obtain pulse signals with different labels, and trains a pulse neural network model based on the difference between the two and preset supervised learning rules, and iteratively converges the weight parameters of the pulse neural network model, finally obtaining a network model for tactile texture recognition tasks in a convenient and efficient manner.

[0130] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or terminal including the element.

[0131] Of course, those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware (such as a processor, a controller, etc.) through a computer program, and the program can be stored in a computer-readable storage medium that can be read by a computer, and the program can include the processes of the above-mentioned method embodiments when executed. The computer-readable storage medium can be a memory, a disk, an optical disk, etc.

[0132] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A tactile texture recognition method based on biological interpretability, characterized in that: The tactile texture recognition method based on biological interpretability includes: Acquire distribution information of a target biosensor, connect a plurality of constructed tactile sensors according to a specific topological structure according to the distribution information, and receive a plurality of tactile data collected by the plurality of tactile sensors; Constructing a simple neuron model, inputting a plurality of the tactile data into the simple neuron model, and outputting a plurality of sparse pulse signals; Constructing an adaptive convolutional network, inputting the plurality of sparse pulse signals into the adaptive convolutional network for feature extraction, and outputting a plurality of corresponding desired sparse pulse signals with desired labels; Constructing a spiking neural network model, inputting a plurality of the desired sparse pulse signals into the spiking neural network model, and outputting a plurality of corresponding target sparse pulse signals with actual labels; Inputting the plurality of expected sparse pulse signals and the plurality of target sparse pulse signals into the pulse neural network model for iterative reverse training until the weight parameter of the pulse neural network model is within a preset threshold range, thereby obtaining an optimized target pulse neural network model; The task tactile data in the recognition task is acquired, the task tactile data is converted into a task sparse pulse signal, the task sparse pulse signal is input into the target pulse neural network model, and the recognition result is output.

2. The tactile texture recognition method based on biological interpretability according to claim 1 is characterized in that: The distribution information includes: a plurality of node information and a plurality of node connection information; The step of acquiring distribution information of a target biosensor, connecting a plurality of constructed tactile sensors according to a specific topological structure according to the distribution information, and receiving a plurality of tactile data collected by the plurality of tactile sensors specifically includes: Acquire the plurality of node information and the plurality of node connection information of the target biosensor, and determine the specific topological structure according to the plurality of node information and the plurality of node connection information; The plurality of constructed tactile sensors are connected according to the specific topological structure, and a plurality of tactile data of a plurality of channels collected by the plurality of tactile sensors are received.

3. The tactile texture recognition method based on biological interpretability according to claim 1, characterized in that: The constructing of the simple neuron model, inputting the plurality of tactile data into the simple neuron model, and outputting a plurality of sparse pulse signals specifically includes: Construct a simple neuron model, input all the tactile data into the simple neuron model, and output the membrane potential corresponding to each tactile data: ； ； like ; in, Indicates at time The membrane potential, Indicates at the current moment The membrane potential, Indicates at time The recovery variable, Indicates the membrane potential at the current moment The recovery variable, represents the membrane potential recovery speed, Indicates at time The input current, Indicates at the current moment The input current, The unit is millivolt; If there are multiple membrane potentials that are not lower than the threshold, multiple pulses are released to obtain multiple sparse pulse signals corresponding to the multiple membrane potentials.

4. The tactile texture recognition method based on biological interpretability according to claim 2 is characterized in that: The step of constructing an adaptive convolutional network, inputting the plurality of sparse pulse signals into the adaptive convolutional network for feature extraction, and outputting a plurality of corresponding desired sparse pulse signals with desired labels specifically includes: Constructing an adaptive convolutional network, and inputting the plurality of sparse pulse signals and the information of the specific topological structure into the adaptive convolutional network; According to the specific topological structure, feature extraction is performed on all the sparse pulse signals in the adaptive convolutional network to obtain feature extraction results: ； ； in, Indicates Output feature maps, Indicates The number of features input to each vertex of the layer, Indicates The operational definition of an adaptive convolution kernel is: Indicates The first node of all nodes in the layer Input features, represents the learnable bias vector, represents the degree of polynomial approximation adopted, Indicates Adaptive convolution kernels The coefficients of the polynomial filter, represents the highest power of the polynomial, represents the normalized adjacency matrix, express of Power; According to the feature extraction result, the plurality of sparse pulse signals are labeled in the adaptive convolutional network, and a plurality of corresponding expected sparse pulse signals with expected classification labels are output.

5. The tactile texture recognition method based on biological interpretability according to claim 1, characterized in that: The constructing of the spiking neural network model, inputting the plurality of the expected sparse pulse signals into the spiking neural network model, and outputting a plurality of corresponding target sparse pulse signals with actual labels specifically includes: Constructing a spiking neural network model, wherein the spiking neural network model includes a fully connected layer network with a preset number of layers, and creating multiple learning windows on all the fully connected layer networks; Inputting a plurality of the desired sparse pulse signals into the spiking neural network model, and outputting a plurality of corresponding target sparse pulse signals with actual labels; Wherein, each of the learning windows in the pulse neural network model processes one expected sparse pulse signal at a time.

6. The tactile texture recognition method based on biological interpretability according to claim 5, characterized in that: The step of inputting the plurality of expected sparse pulse signals and the plurality of target sparse pulse signals into the pulse neural network model for iterative reverse training until the weight parameter of the pulse neural network model is within a preset threshold range to obtain an optimized target pulse neural network model specifically includes: Determining difference information between a plurality of the desired sparse pulse signals and a plurality of corresponding target sparse pulse signals; The difference information is input into the spiking neural network model for multiple times for back propagation, so as to iteratively train the spiking neural network model, and the weight parameters of the spiking neural network model are output multiple times until the weight parameters are within a preset threshold range, thereby obtaining a weight change: ； ； ； in, Indicates Layer Network Layer The neurons in the previous network layer The weight change parameter between neurons, Indicates Layer Network Layer The neurons in the previous network layer The weights between neurons, Indicates learning efficiency, Indicates the deviation signal, Indicates The number of neurons in a layer, Indicates The previous network layer of the layer The actual pulse arrival time of each neuron is Indicates The previous network layer of the layer The expected pulse arrival time of each neuron is Indicated in The values ​​in different situations, Indicates The average time of pulse emission in the previous network layer of the layer, represents the number of neurons in the output layer, represents the interval that distinguishes target and non-target neurons, Represents the neuron that outputs the actual label; ； ； ； in, represents the gradient of the output layer error back propagation in the fully connected layer, represents the relationship between the postsynaptic spike firing time and weight, and They represent the proportional factors of control enhancement and suppression respectively, and represent the presynaptic and postsynaptic spike firing times, respectively. represents the first parameter controlling the weight update, represents the time interval coefficient for weight update, and Respectively represent the maximum weight and current weight of the weight constraint, represents the second parameter controlling the weight update; ； ； in, Represents the recursive formula for the back propagation gradient of the hidden layer error in the fully connected layer, Indicates The previous network layer of the layer The actual pulse arrival time of each neuron is Indicates The previous network layer of the layer The actual pulse arrival time of each neuron; The pulse neural network model is updated in real time according to the weight change parameters to obtain an optimized target pulse neural network model.

7. The tactile texture recognition method based on biological interpretability according to claim 1, characterized in that: The acquiring task tactile data in the recognition task, converting the task tactile data into a task sparse pulse signal, inputting the task sparse pulse signal into the target pulse neural network model, and outputting the recognition result specifically includes: Receiving a plurality of task tactile data collected by the tactile sensor, inputting the plurality of task tactile data into the simple neuron model and the adaptive convolutional network in sequence, and outputting a plurality of task sparse pulse signals with desired labels; All the task sparse pulse signals with expected labels are input into the target pulse neural network model for forward propagation, and the recognition result of the recognition task is output.

8. A tactile texture recognition system based on biological interpretability, characterized in that: The tactile texture recognition system based on biological interpretability includes: A tactile data acquisition module, used to obtain distribution information of a target biosensor, connect a plurality of constructed tactile sensors according to a specific topological structure according to the distribution information, and receive a plurality of tactile data acquired by the plurality of tactile sensors; A first pulse signal generating module, used for constructing a simple neuron model, inputting a plurality of tactile data into the simple neuron model, and outputting a plurality of sparse pulse signals; A second pulse signal generating module is used to construct an adaptive convolutional network, input a plurality of the sparse pulse signals into the adaptive convolutional network for feature extraction, and output a plurality of corresponding desired sparse pulse signals with desired labels; A third pulse signal generating module is used to construct a pulse neural network model, input a plurality of the desired sparse pulse signals into the pulse neural network model, and output a plurality of corresponding target sparse pulse signals with actual labels; A model optimization module, used for inputting a plurality of the expected sparse pulse signals and a plurality of the target sparse pulse signals into the pulse neural network model for iterative reverse training until a weight parameter of the pulse neural network model is within a preset threshold range, thereby obtaining an optimized target pulse neural network model; The tactile texture recognition module is used to obtain task tactile data in the recognition task, convert the task tactile data into a task sparse pulse signal, input the task sparse pulse signal into the target pulse neural network model, and output the recognition result.

9. A terminal, characterized in that: The terminal includes: a memory, a processor, and a tactile texture recognition program based on bio-interpretability stored in the memory and executable on the processor. When the tactile texture recognition program based on bio-interpretability is executed by the processor, the steps of the tactile texture recognition method based on bio-interpretability are implemented as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a tactile texture recognition program based on biological interpretability, and when the tactile texture recognition program based on biological interpretability is executed by a processor, the steps of the tactile texture recognition method based on biological interpretability as described in any one of claims 1 to 7 are implemented.

Images