A method and device for tactile object recognition based on a novel tactile map

Abstract

The present application provides a tactile object recognition method and device based on a new tactile map, which relates to the field of tactile object recognition of intelligent robots. The method includes: acquiring tactile data, and dividing the tactile data into a training data set and a test data set; based on a network tree algorithm, constructing the training data set into a network tree tactile map; constructing an SNN tactile object recognition model, and optimizing the object recognition model using an error function and a Dropout layer to obtain an optimized SNN tactile object recognition model; training the optimized SNN tactile object recognition model according to the network tree tactile map to obtain a trained optimized SNN tactile object recognition model; inputting the test data set into the trained optimized SNN tactile object recognition model to obtain a predicted object category. In this way, the processing efficiency and recognition accuracy of the model for tactile data can be improved.

Description

Technical Field

[0001] The present application relates to the field of tactile object recognition of intelligent robots, and in particular to a tactile object recognition method and device based on a novel tactile map. Background Art

[0002] As a key way for robots to interact with the physical environment, tactile perception is crucial for robots to achieve operational tasks in complex environments. How to improve the accuracy of robot tactile recognition and rapid response has become a major problem to be solved. Most existing tactile recognition algorithms are based on convolutional neural networks and residual network models, which require the fusion of other information such as vision and hearing for recognition. They are slow when processing large amounts of data and cannot meet real-time requirements. At the same time, tactile data contains high-dimensional and nonlinear information. Existing methods cannot efficiently extract tactile feature information in a single tactile recognition mode, and their computational efficiency and recognition accuracy still need to be improved. In solving the core problems of "low accuracy" and "slow response", SNN can process information in the form of pulses, which has advantages in processing complex dynamic and real-time data of tactile perception.

[0003] However, how to effectively use SNN for complex tactile object recognition, especially in large-scale tactile data processing and real-time applications, remains a challenge. On the one hand, current tactile graph construction methods are often too simple to fully capture the subtle details and texture changes on the surface of objects, limiting the performance of tactile recognition systems. On the other hand, the choice of impulse approximation function in SNN has a direct impact on the computational efficiency and recognition accuracy of the network, but existing methods lack effective solutions in balancing these two key indicators. Summary of the invention

[0004] The present application provides a tactile object recognition method and device based on a novel tactile map, aiming to improve the processing efficiency and recognition accuracy of the model for tactile data.

[0005] According to a first aspect of the present application, a tactile object recognition method based on a novel tactile map is provided, the method comprising:

[0006] Acquire tactile data, and divide the tactile data into a training data set and a test data set;

[0007] Based on a network tree algorithm, constructing the training data set into a network tree tactile graph;

[0008] Constructing an SNN tactile object recognition model, and optimizing the object recognition model using an error function and a Dropout layer to obtain an optimized SNN tactile object recognition model; wherein the optimized SNN tactile object recognition model includes a topological graph convolution layer, a LIF neuron excitation layer, a fully connected layer, a Dropout layer, and a voting layer;

[0009] Training the optimized SNN tactile object recognition model according to the network tree tactile graph to obtain a trained optimized SNN tactile object recognition model;

[0010] The test data set is input into the trained optimized SNN tactile object recognition model to obtain the predicted object category.

[0011] According to the above aspects and any possible implementation, an implementation is further provided, wherein the training data set is constructed into a network tree tactile graph based on a network tree algorithm, comprising:

[0012] Preprocessing the tactile data in the training data set, and defining n touch point coordinates of the network tree according to the preprocessed tactile data;

[0013] Constructing an adjacency matrix according to the coordinates of the n touch points of the network tree;

[0014] Initializing an undirected graph, updating the undirected graph according to the adjacency matrix, and constructing a network tree tactile graph according to the updated undirected graph.

[0015] According to the above aspects and any possible implementation manner, an implementation manner is further provided, wherein the step of constructing an adjacency matrix according to the coordinates of the n touch points of the network tree comprises:

[0016] Calculating the Euclidean distance between each pair of touch points according to the coordinates of the n touch points of the network tree;

[0017] Initialize the all-zero adjacency matrix, where the size of the all-zero adjacency matrix is ​​n n;

[0018] It is determined whether the distance between each pair of touch points is less than a preset distance threshold. If so, the corresponding position of the all-zero adjacency matrix is ​​set to 1 to obtain an updated adjacency matrix.

[0019] According to the above aspects and any possible implementation manner, an implementation manner is further provided, in which the undirected graph is updated according to the adjacency matrix, including:

[0020] Create a new undirected graph and add a node to the undirected graph;

[0021] Determine the value of each element in the adjacency matrix;

[0022] If the element value is 1, an edge connecting the corresponding node is added to the undirected graph to obtain the updated undirected graph.

[0023] According to the above aspects and any possible implementation, an implementation is further provided, which constructs an SNN tactile object recognition model, and optimizes the object recognition model using an error function and a Dropout layer to obtain an optimized SNN tactile object recognition model, including:

[0024] Initializing the SNN tactile object recognition model and using the network tree tactile graph as the model input layer;

[0025] Establishing a topological convolutional layer under the network tree tactile graph module;

[0026] The ERF error function is used to approximate the pulse data and obtain the membrane potential update dynamics of the spiking neuron;

[0027] Taking the membrane potential update dynamics of the spiking neuron as input, adding a fully connected layer, reducing the neurons of the neural network, and obtaining tactile information features;

[0028] Add a dropout layer to regularize the SNN tactile object recognition model;

[0029] Add voting layer to decode network output.

[0030] According to the above aspects and any possible implementation, an implementation is further provided, wherein the pulse data is approximately expressed using the ERF error function as follows:

[0031] ;

[0032] in, Indicates pulse data, Indicates that the pulse data is approximated using the ERF error function. Represents the parameter that controls the smoothness of the gradient during back propagation, represents the error function.

[0033] According to the above aspects and any possible implementation, a further implementation is provided, wherein the ERF error function It is expressed as:

[0034] ;

[0035] Here, t represents time.

[0036] According to the above aspects and any possible implementation, an implementation is further provided, wherein the optimized SNN tactile object recognition model is trained according to the network tree tactile graph, including:

[0037] Inputting the network tree tactile graph into the topological graph convolution layer of the optimized SNN tactile object recognition model to extract the spatial features of the tactile data;

[0038] generating a pulse sequence in the LIF neuron excitation layer through a LIF model according to the spatial characteristics of the tactile data;

[0039] Inputting the pulse sequence into a fully connected layer to obtain the pulse count of the LIF neuron;

[0040] In the voting layer, the spike counts of all LIF neurons are aggregated and the final object category is determined by a majority voting mechanism.

[0041] According to the above aspects and any possible implementation manner, an implementation manner is further provided, wherein the method further includes:

[0042] Compare the predicted object category with the real object category and calculate the object recognition consistency probability;

[0043] If the consistency probability does not reach a preset threshold, the SNN tactile object recognition model is optimized.

[0044] According to a second aspect of the present application, a tactile object recognition device based on a novel tactile map is provided, the device comprising:

[0045] A data set division module, used for acquiring tactile data and dividing the tactile data into a training data set and a test data set;

[0046] A tactile map construction module, used to construct the training data set into a network tree tactile map based on a network tree algorithm;

[0047] An object recognition model construction module is used to construct an SNN tactile object recognition model, and optimize the object recognition model using an error function and a Dropout layer to obtain an optimized SNN tactile object recognition model; wherein the SNN tactile object recognition model includes a topological graph convolution layer, a LIF neuron excitation layer, a fully connected layer, a Dropout layer, and a voting layer;

[0048] An object recognition model training module is used to train the optimized SNN tactile object recognition model according to the network tree tactile graph to obtain a trained optimized SNN tactile object recognition model;

[0049] The object category prediction module is used to input the test data set into the trained optimized SNN tactile object recognition model to obtain the predicted object category.

[0050] Compared with the prior art, this application achieves the following beneficial effects:

[0051] (1) By converting tactile signals into a graph structure, the SNN model can learn the correlation and local features between tactile signals. By constructing a tactile graph through a network tree method, it can efficiently represent the spatial relationship of tactile signals, which helps the model better understand the structure and pattern of tactile data, thereby improving the prediction performance of the model. At the same time, the network tree tactile graph construction module can flexibly adapt to tactile datasets of different shapes and sizes. Even if the tactile dataset changes, the graph structure can be quickly adjusted to adapt to the new data.

[0052] (2) The pulse excitation module uses the ERF error function to approximate the derivative of the LIF neuron activation function. The ERF error function provides a continuous derivative. Compared with the traditional step function, it can smoothly activate the pulse feature, which helps the gradient to flow steadily during the back propagation process and avoids the problem of gradient vanishing or exploding. In particular, it can achieve effective weight updates, especially when processing complex tactile signals, and can converge to the optimal solution faster. At the same time, the use of the ERF function imitates the behavior of biological neurons to a certain extent, especially in the generation and propagation of pulse signals, which enables the model to simulate the working principle of the nervous system more realistically.

[0053] (3) The dropout layer in the regularization module further enhances the generalization ability of the SNN model, effectively avoids overfitting, and ensures that the model can maintain a high level of recognition accuracy when faced with complex and diverse tactile data.

[0054] It should be understood that the contents described in the Summary of the Invention are not intended to limit the key or important features of the embodiments of the present application, nor are they intended to limit the scope of the present application. Other features of the present application will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The above and other features, advantages and aspects of the embodiments of the present application will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present application. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which:

[0056] Figure 1 A flow chart of a tactile object recognition method based on a novel tactile map according to an embodiment of the present application is shown;

[0057] Figure 2 A structural block diagram of the SNN tactile object recognition model according to an embodiment of the present application is shown;

[0058] Figure 3 shows a network tree tactile diagram constructed according to an embodiment of the present application;

[0059] Figure 4shows the comparison between the training loss and the test loss of the tactile object recognition model according to an embodiment of the present application, and the final tactile object prediction accuracy (%);

[0060] Figure 5 shows a training loss graph of an SNN tactile object recognition model according to an embodiment of the present application;

[0061] Figure 6 A test loss diagram of an SNN tactile object recognition model according to an embodiment of the present application is shown;

[0062] Figure 7 shows an optimal tactile object recognition accuracy diagram of the SNN tactile object recognition model according to an embodiment of the present application;

[0063] Figure 8 The figure shows the best recognition accuracy and the average calculation time per iteration according to an embodiment of the present application.

[0064] Fig. 9 A block diagram of a tactile object recognition device based on a novel tactile map according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0066] In addition, the term "and / or" in this application is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0067] This application proposes a tactile object recognition method based on a novel tactile map, the method comprising:

[0068] Acquire tactile data and divide the tactile data into a training data set and a test data set; construct the training data set into a network tree tactile graph based on a network tree algorithm; construct an SNN tactile object recognition model, and optimize the object recognition model using an error function and a Dropout layer to obtain an optimized SNN tactile object recognition model; train the optimized SNN tactile object recognition model according to the network tree tactile graph to obtain a trained optimized SNN tactile object recognition model; input the test data set into the trained optimized SNN tactile object recognition model to obtain a predicted object category. In this way, the processing efficiency and recognition accuracy of the model for tactile data can be improved.

[0069] A specific embodiment of a tactile object recognition method based on a novel tactile map provided by the present application is described below in conjunction with the accompanying drawings.

[0070] Embodiment 1

[0071] Figure 1 FIG. 4 is a flow chart of a tactile object recognition method based on a novel tactile map according to an embodiment of the present application. Figure 1 As shown, the method includes:

[0072] S100: Acquire tactile data, and divide the tactile data into a training data set and a test data set.

[0073] In some embodiments, tactile data can be acquired through a tactile sensor, and the acquired tactile data can be divided into a training data set and a test data set. Specifically, it can be divided into 80% of the training data set and 20% of the test data set, and it is ensured that the proportion of tactile data of each category in the training data set and the test data set is roughly the same. In this way, it helps to avoid the tactile object recognition model being biased towards a certain category during the training process, thereby ensuring the generalization ability of the tactile object recognition model on different categories.

[0074] S120: Based on a network tree algorithm, construct the training data set into a network tree tactile graph.

[0075] In some embodiments, when the network tree tactile graph is constructed for the divided training data set in step S2, the following steps are included:

[0076] Preprocessing the tactile data in the training data set, and defining n touch point coordinates of the network tree according to the preprocessed tactile data;

[0077] Constructing an adjacency matrix according to the coordinates of the n touch points of the network tree;

[0078] Initializing an undirected graph, updating the undirected graph according to the adjacency matrix, and constructing a network tree tactile graph according to the updated undirected graph.

[0079] In some embodiments, in the above step S22, constructing an adjacency matrix according to the coordinates of the n touch points of the network tree includes the following steps:

[0080] Calculating the Euclidean distance between each pair of touch points according to the coordinates of the n touch points of the network tree;

[0081] Initialize the all-zero adjacency matrix, where the size of the all-zero adjacency matrix is ​​n n;

[0082] It is determined whether the distance between each pair of touch points is less than a preset distance threshold. If so, the corresponding position of the all-zero adjacency matrix is ​​set to 1 to obtain an updated adjacency matrix.

[0083] In some embodiments, in the above steps, updating the undirected graph according to the adjacency matrix includes the following steps:

[0084] Create a new undirected graph and add a node to the undirected graph;

[0085] Determine the value of each element in the adjacency matrix;

[0086] If the element value is 1, an edge connecting the corresponding node is added to the undirected graph to obtain the updated undirected graph.

[0087] S130: constructing an SNN tactile object recognition model, and optimizing the object recognition model using an error function and a Dropout layer to obtain an optimized SNN tactile object recognition model; wherein the optimized SNN tactile object recognition model includes a topological graph convolution layer, a LIF neuron excitation layer, a fully connected layer, a Dropout layer and a voting layer.

[0088] Among them, an SNN tactile object recognition model is constructed, which includes a topological graph convolution layer, two fully connected layers, three LIF neuron excitation layers and a voting layer. The model architecture is: input tactile pulse data obtained by the tactile sensor → use the network tree method to construct a tactile graph containing tactile pulse information → topological graph convolution layer (3×3×64) → LIF neuron excitation layer → fully connected layer (128) → LIF neuron excitation layer → fully connected layer (256) → LIF neuron excitation layer → voting layer → output recognition label, and use the error function and Dropout layer to optimize the object recognition model.

[0089] Optimization principle: The error function can smoothly activate the pulse signal features in the tactile data, avoid the gradient vanishing or gradient exploding problems of the SNN model during the back propagation process, and enable the SNN model to converge to the optimal solution faster; the dropout layer in the regularization module can enhance the generalization ability of the SNN model, effectively avoid overfitting, and ensure that the SNN model can still maintain a high degree of recognition accuracy when facing complex and diverse tactile data.

[0090] In some embodiments, constructing an SNN tactile object recognition model, and optimizing the object recognition model using an error function and a Dropout layer to obtain an optimized SNN tactile object recognition model includes the following steps:

[0091] Initializing the SNN tactile object recognition model and using the network tree tactile graph as the model input layer;

[0092] Establishing a topological convolutional layer under the network tree tactile graph module;

[0093] The ERF error function is used to approximate the pulse data and obtain the membrane potential update dynamics of the spiking neuron;

[0094] Taking the membrane potential update dynamics of the spiking neuron as input, adding a fully connected layer, reducing the neurons of the neural network, and obtaining tactile information features;

[0095] Add a dropout layer to regularize the SNN tactile object recognition model;

[0096] Add voting layer to decode network output.

[0097] So far, the construction and optimization of the SNN tactile object recognition model has been completed.

[0098] In some embodiments, in the above step S33, the pulse data may be approximated using the ERF error function as follows:

[0099] ;

[0100] in, Indicates pulse data, Indicates that the pulse data is approximated using the ERF error function. Represents the parameter that controls the smoothness of the gradient during back propagation, represents the error function.

[0101] In some embodiments, the ERF error function It is expressed as:

[0102] ;

[0103] Here, t represents time.

[0104] S140: Training the optimized SNN tactile object recognition model according to the network tree tactile graph to obtain a trained optimized SNN tactile object recognition model.

[0105] In some embodiments, the training of the optimized SNN tactile object recognition model according to the network tree tactile graph comprises the following steps:

[0106] Inputting the network tree tactile graph into the topological graph convolution layer of the optimized SNN tactile object recognition model to extract the spatial features of the tactile data;

[0107] generating a pulse sequence in the LIF neuron excitation layer through a LIF model according to the spatial characteristics of the tactile data;

[0108] Inputting the pulse sequence into a fully connected layer to obtain the pulse count of the LIF neuron;

[0109] In the voting layer, the spike counts of all LIF neurons are aggregated and the final object category is determined by a majority voting mechanism.

[0110] S150: Inputting the test data set into the trained optimized SNN tactile object recognition model to obtain a predicted object category.

[0111] In some embodiments, the method further comprises:

[0112] Compare the predicted object category with the real object category and calculate the object recognition consistency probability;

[0113] If the consistency probability does not reach a preset threshold, the SNN tactile object recognition model is optimized.

[0114] Embodiment 2

[0115] Combine the following Figure 2-Figure 8 , a specific embodiment of the tactile object recognition method based on the novel tactile map provided in the present application is described in detail.

[0116] The tactile object recognition method based on the novel tactile map provided in the present application may include the following steps:

[0117] Step s1: Divide the dataset into training dataset and test dataset.

[0118] Step s2: constructing a tactile graph, and constructing a network tree tactile graph based on the network tree;

[0119] Step s3: Initialize the SNN tactile object recognition model and use the network tree tactile graph module constructed in step s2 as the model input layer;

[0120] Step s4: establishing a topological convolutional layer under the network tree tactile layer to adapt to the topological structure of the network tree tactile graph so as to more effectively propagate and aggregate tactile information in the network;

[0121] Step s5: Use the error function (ERF) to approximate the LIF neuron to realize the tactile information pulse excitation and obtain the membrane potential update dynamics of the pulse neuron;

[0122] Step s6: taking the result of step 5 as input, adding a fully connected layer, reducing the neurons of the neural network to obtain tactile information features;

[0123] Step s7: Regularize the SNN tactile object recognition model through the dropout layer to enhance the generalization ability of the model and prevent overfitting;

[0124] Step s8: Decode the network output through the voting layer. Each output label is first associated with a neuron in the voting layer.

[0125] Step s9: Output the predicted class, which is the class associated with the neuron with the largest average number of votes in the time window.

[0126] Specifically, in step s1, there are 36 test categories in total, 20 samples in each object category, and 720 samples in total. The 720 samples can be divided into 80% training data set and 20% test data set, and ensure that the proportion of tactile data of each category in the training data set and the test data set is roughly the same. In this way, it helps to avoid the tactile object recognition model from being biased towards a certain category during the training process, thereby ensuring the generalization ability of the tactile object recognition model on different categories.

[0127] Specifically, when the network tree tactile graph is constructed for the divided tactile training data in step s2, the following steps are included:

[0128] Step s21: Import numpy, torch_geometric, pytorch, networkx and sklearn libraries to prepare for building the network tree tactile graph;

[0129] Step s22: Define the coordinates of 39 touch points on the tactile sensor as points, points = np.array([[-6, 0], [-5.3, -3], [-5.3, 3], [-4.6, -7.8], [-4.6, 7.8], [-3.5, 0], [-3.05, -5.2], [-3.05, 5.2], [-3.1, -1.75], [-3.1, 1.75], [-1.6, -8.9], [-1.75, -3], [-1.6, 8.9], [-1.75, 3], [-1.5, 0], [-0.7, -1.3], [-0.7, 1.3], [0, -6],[0, 6],[0, , 3], [6, 0]]);

[0130] Step s23: Calculate the number of points and store it in the variable n_points;

[0131] Step s24: Initialize an all-zero adjacency matrix adj_matrix, whose size is the same as the number of points, to record whether there is an edge between two points in the graph;

[0132] Step s25: Set the threshold value to 3.5 to determine the distance between two points, thereby adding an edge to the graph;

[0133] Step s26: Start a double loop to traverse all points and calculate the Euclidean distance dist between each pair of points;

[0134] Step s27: if the distance between two points is less than the set threshold, then the value of the corresponding position in the adjacency matrix is ​​set to 1, indicating that there is an edge between the two points. At the same time, since the graph is undirected, values ​​need to be set in both directions of the matrix.

[0135] Step s28: Create an empty undirected graph G using the networkx library;

[0136] Step s29: add a node to the graph G, where the node identifier is the sequence number of the point;

[0137] Step s210: traverse the adjacency matrix. If a value in the matrix is ​​1, add an edge connecting the corresponding node in the graph G. This realizes the process of building an adjacency matrix based on a given point set, and then creating and visualizing an undirected graph. The undirected graph is defined as a network tree tactile graph. The final network tree tactile graph is as follows: Figure 3 Specifically, the network tree tactile graph construction code is as follows:

[0138] n_points = points.shape[0]

[0139] adj_matrix = np.zeros((n_points, n_points), dtype=int)

[0140] threshold = 3.5

[0141] # Fill the adjacency matrix

[0142] for i in range(n_points):

[0143] for j in range(i + 1, n_points):

[0144] dist = np.sqrt(np.sum(np.square(points[i] - points[j])))

[0145] if dist <threshold:

[0146] adj_matrix[i][j] = 1

[0147] adj_matrix[j][i] = 1

[0148] # Create a graph

[0149] G = nx.Graph()

[0150] # Add nodes

[0151] G.add_nodes_from(range(n_points))

[0152] # Add edges

[0153] for i in range(n_points):

[0154] for j in range(i + 1, n_points):

[0155] if adj_matrix[i][j]== 1:

[0156] G.add_edge(i, j)

[0157] Specifically, in the implementation of step s3, when the training module loads the network tree tactile graph constructed by the divided tactile training data into the network model for training, the following steps are included:

[0158] Step s31: Import the os, numpy, torch_geometric and pytorch libraries, as well as the network tree tactile graph construction module, and prepare to build the SNN tactile object recognition model;

[0159] Step s32: define the model name as class snn_model(nn.Module);

[0160] Step s33: define the topological graph convolution layer, fully connected layer, voting layer, and the network tree tactile graph module object for generating the graph.

[0161] Step s33: Establish the first topology graph convolution layer, the data input is ((cfg_cnn[0])[0]), and the data output is ((cfg_cnn[0])[1]);

[0162] Step s34: define an approximate activation function activate_faction, which is a PyTorch automatic differentiation function used to simulate the activation behavior of spiking neurons;

[0163] Step s35: After the first layer, a LIF neuron excitation layer is embedded. This layer is similar to the activation function in CNN, which enables the model to realize forward and backward propagation, wherein the forward propagation function is used to generate pulses, and the backward propagation function provides an approximate derivative calculation for use in reverse propagation. Because the tactile signal has nonlinear characteristics, the present invention uses the ERF function to approximate the derivative of the LIF neuron activation function and converts it into linear pulse data, so as to realize the reverse propagation of the pulse data in the network model. The ERF function can be expressed as:

[0164] ;

[0165] in, Indicates pulse data, Indicates that the pulse data is approximated using the approximate ERF function. Represents the parameter that controls the smoothness of the gradient during back propagation, Can be set between 0 and 1 for smoothing. represents the ERF error function, It can be defined as:

[0166] ;

[0167] Wherein, t represents time;

[0168] Step s36: Add a fully connected layer nn.Sequential(nn.Linear(cfg_s[-1] cfg_cnn[-1][1], cfg_fc[1]));

[0169] Step s37: Add the dropout layer nn.Dropout(dropout) and set the dropout parameter to 0.25;

[0170] Step s38: Add an output layer with only one output. The SNN tactile object recognition model is built. The function name is snn_model. The structural diagram of the SNN tactile object recognition model is as follows: Figure 2 shown.

[0171] Step s39: call snn_model model model = snn_model (num_classes, k, device=device);

[0172] Step s310: define the number of categories num_classes as 36, the number of nodes k connected when building the graph as 3, the learning rate learning_rate as 0.001, and the number of training cycles num_epochs as 100;

[0173] Step s311: Import the training set or test set, parse the label from the file name, load the data and convert it into a tensor;

[0174] Step s312: Setting the learning rate decay strategy: When the preset period is reached, the learning rate is multiplied by 0.1 for decay;

[0175] Step s313: Use mean square error loss MSELoss as the loss function and Adam as the optimizer;

[0176] Step s314: Perform training once for each epoch and output the training and testing status of the current epoch.

[0177] The classification output is returned through the forward function, and the average pulse obtained by dividing the total pulse of the last time step by the time window size is trained once for each epoch, and the training and test status of the current epoch are output.

[0178] Combination Figure 4-Figure 8 , we can see that the prediction results obtained by this application are:

[0179] like Figure 4 As shown, the training loss of the tactile object recognition model of the present application is 1.689965, the test loss is 1.174933, and the final tactile object prediction accuracy is 89.583%. Figure 5 , Figure 6 The training loss graph and the test loss graph of the SNN tactile object recognition model of the present invention are shown respectively.

[0180] like Figure 7-Figure 8 As shown, the best recognition accuracy of the SNN tactile object recognition model of the present invention is 89.583%, and each iteration takes an average of 708 seconds.

[0181] In summary, the tactile object recognition method based on the new tactile map provided in this application realizes the tactile object recognition of the robot, improves the efficiency of tactile object recognition in a single tactile recognition mode, builds a more efficient and accurate tactile signal processing system, and provides strong technical support for application scenarios such as robot tactile perception and tactile object recognition.

[0182] According to the embodiments of the present application, the following technical effects are achieved:

[0183] (1) By converting tactile signals into a graph structure, the SNN model can learn the correlation and local features between tactile signals. By constructing a tactile graph through a network tree method, it can efficiently represent the spatial relationship of tactile signals, which helps the model better understand the structure and pattern of tactile data, thereby improving the prediction performance of the model. At the same time, the network tree tactile graph construction module can flexibly adapt to tactile datasets of different shapes and sizes. Even if the tactile dataset changes, the graph structure can be quickly adjusted to adapt to the new data.

[0184] (2) The pulse excitation module uses the ERF error function to approximate the derivative of the LIF neuron activation function. The ERF error function provides a continuous derivative. Compared with the traditional step function, it can smoothly activate the pulse feature, which helps the gradient to flow steadily during the back propagation process and avoids the problem of gradient vanishing or exploding. In particular, it can achieve effective weight updates, especially when processing complex tactile signals, and can converge to the optimal solution faster. At the same time, the use of the ERF function imitates the behavior of biological neurons to a certain extent, especially in the generation and propagation of pulse signals, which enables the model to simulate the working principle of the nervous system more realistically.

[0185] (3) The dropout layer in the regularization module further enhances the generalization ability of the SNN model, effectively avoids overfitting, and ensures that the model can maintain a high level of recognition accuracy when faced with complex and diverse tactile data.

[0186] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by the present application.

[0187] The above is an introduction to the method embodiment. The following is a further explanation of the scheme described in this application through an apparatus embodiment.

[0188] Fig. 9 FIG. 4 is a block diagram of a tactile object recognition device based on a novel tactile map according to an embodiment of the present application. Fig. 9 As shown, the device comprises:

[0189] A data set division module 101 is used to obtain tactile data and divide the tactile data into a training data set and a test data set;

[0190] A tactile map construction module 102 is used to construct the training data set into a network tree tactile map based on a network tree algorithm;

[0191] The object recognition model construction module 103 is used to construct an SNN tactile object recognition model, and optimize the object recognition model using an error function and a Dropout layer to obtain an optimized SNN tactile object recognition model; wherein the SNN tactile object recognition model includes a topological graph convolution layer, a LIF neuron excitation layer, a fully connected layer, a Dropout layer and a voting layer;

[0192] The object recognition model training module 104 is used to train the optimized SNN tactile object recognition model according to the network tree tactile graph to obtain a trained optimized SNN tactile object recognition model;

[0193] The object category prediction module 105 is used to input the test data set into the trained optimized SNN tactile object recognition model to obtain the predicted object category.

[0194] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0195] In the technical solution of this application, the acquisition, storage and application of user personal information involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0196] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this application can be executed in parallel, sequentially or in different orders, as long as the expected results of the technical solution of this application can be achieved, and this document is not limited here.

[0197] The above specific implementations do not constitute a limitation on the protection scope of this application. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of this application should be included in the protection scope of this application.

Claims

1. A tactile object recognition method based on a novel tactile map, characterized in that: include: Acquire tactile data, and divide the tactile data into a training data set and a test data set; Based on a network tree algorithm, constructing the training data set into a network tree tactile graph; Wherein, the training data set is constructed into a network tree tactile graph based on a network tree algorithm, including: Preprocessing the tactile data in the training data set, and defining 39 touch point coordinates of the network tree according to the preprocessed tactile data; Constructing an adjacency matrix according to the coordinates of the 39 touch points of the network tree; Traverse the adjacency matrix. If a value in the matrix is ​​1, add an edge connecting the corresponding node in the graph G. This realizes the process of building an adjacency matrix based on a given point set, and then creating and visualizing an undirected graph. The undirected graph is defined as a network tree tactile graph. The final network tree tactile graph is obtained. Constructing an SNN tactile object recognition model, and optimizing the object recognition model using an error function and a Dropout layer to obtain an optimized SNN tactile object recognition model; wherein the optimized SNN tactile object recognition model includes a topological graph convolution layer, a LIF neuron excitation layer, a fully connected layer, a Dropout layer, and a voting layer; The step of constructing an SNN tactile object recognition model and optimizing the object recognition model using an error function and a Dropout layer to obtain an optimized SNN tactile object recognition model comprises: Initializing the SNN tactile object recognition model and using the network tree tactile graph as the model input layer; Building a topological convolutional layer after the network tree tactile graph; The pulse data is approximated using the ERF error function to obtain the membrane potential update dynamics of the spiking neuron; wherein, the pulse data is approximated using the ERF error function as follows: ； in, Indicates pulse data, Indicates that the pulse data is approximated using the ERF error function. Represents the parameter that controls the smoothness of the gradient during back propagation, represents the error function; Taking the membrane potential update dynamics of the spiking neuron as input, adding a fully connected layer, reducing the neurons of the neural network, and obtaining tactile information features; Add a dropout layer to regularize the SNN tactile object recognition model; Add voting layer to decode network output; Training the optimized SNN tactile object recognition model according to the network tree tactile graph to obtain a trained optimized SNN tactile object recognition model; The test data set is input into the trained optimized SNN tactile object recognition model to obtain the predicted object category.

2. The method according to claim 1, characterized in that The ERF error function It is expressed as: ； Here, t represents time.

3. The method according to claim 1, characterized in that The step of training the optimized SNN tactile object recognition model according to the network tree tactile graph comprises: Inputting the network tree tactile graph into the topological graph convolution layer of the optimized SNN tactile object recognition model to extract the spatial features of the tactile data; generating a pulse sequence in the LIF neuron excitation layer through a LIF model according to the spatial characteristics of the tactile data; Inputting the pulse sequence into a fully connected layer to obtain the pulse count of the LIF neuron; In the voting layer, the spike counts of all LIF neurons are aggregated and the final object category is determined by a majority voting mechanism.

4. The method according to any one of claims 1 to 3, characterized in that: The method further comprises: Compare the predicted object category with the real object category and calculate the object recognition consistency probability; If the consistency probability does not reach a preset threshold, the SNN tactile object recognition model is optimized.

5. A tactile object recognition device based on a novel tactile map, used to implement the tactile object recognition method based on a novel tactile map as claimed in any one of claims 1 to 4, characterized in that: include: A data set division module, used for acquiring tactile data and dividing the tactile data into a training data set and a test data set; A tactile map construction module, used to construct the training data set into a network tree tactile map based on a network tree algorithm; An object recognition model construction module is used to construct an SNN tactile object recognition model, and optimize the object recognition model using an error function and a Dropout layer to obtain an optimized SNN tactile object recognition model; wherein the SNN tactile object recognition model includes a topological graph convolution layer, a LIF neuron excitation layer, a fully connected layer, a Dropout layer, and a voting layer; An object recognition model training module is used to train the optimized SNN tactile object recognition model according to the network tree tactile graph to obtain a trained optimized SNN tactile object recognition model; The object category prediction module is used to input the test data set into the trained optimized SNN tactile object recognition model to obtain the predicted object category.

Images