An optimization method for osteoarticular rehabilitation training

Abstract

The present invention belongs to the field of rehabilitation training, and specifically refers to an optimization method for bone and joint rehabilitation training. The method includes information extraction, model generation, and memory replacement. This solution uses a bone model generation method, which uses multiple depth cameras to obtain patient limb data from different angles, generates and continuously updates a 3D human bone model, realizes continuous tracking of human postures, and improves the stability and flexibility of the 3D human bone model; uses a memory replacement method to enable the long short-term memory network to continuously learn video data when the patient is undergoing bone and joint rehabilitation training. On the premise of not forgetting past knowledge, dynamically adjust the long short-term memory network structure, construct new node paths and remove unimportant nodes to learn new joint movement knowledge and improve the continuous learning ability of the long short-term memory network.

Description

Technical Field

[0001] The present invention relates to the field of rehabilitation training, and specifically refers to an optimization method for osteoarticular rehabilitation training. Background Art

[0002] The optimization method for osteoarticular rehabilitation training refers to a method of optimizing the rehabilitation training of patients through a 3D human bone model.

[0003] Existing patient pose capture technologies have problems that the 3D joint position data generated by depth cameras all have certain errors and occlusions; in addition, traditional long short-term memory networks have the fatal drawback of catastrophic forgetting, and there are technical problems that they are not suitable for analyzing and learning rehabilitation training videos with large amounts of data, high real-time requirements, and fast update speeds. Summary of the Invention

[0004] In view of the above situation, to overcome the defects of the prior art, the present invention provides an optimization method for osteoarticular rehabilitation training. Aiming at the problems that existing patient pose capture technologies have certain errors and occlusions in the 3D joint position data generated by depth cameras, this solution adopts a bone model generation method, uses multiple depth cameras to obtain patient limb data from different angles, generates and continuously updates a 3D human bone model, realizes continuous tracking of human postures, and improves the stability and flexibility of the 3D human bone model; aiming at the disadvantage of catastrophic forgetting in traditional long short-term memory networks and the technical problem that they are not suitable for analyzing and learning rehabilitation training videos with large amounts of data, high real-time requirements, and fast update speeds, this solution adopts a memory replacement method to enable the long short-term memory network to continuously learn video data during the patient's osteoarticular rehabilitation training, dynamically adjust the long short-term memory network structure without forgetting past knowledge, construct new node paths and remove unimportant nodes to learn new joint movement knowledge, and improve the continuous learning ability of the long short-term memory network.

[0005] The technical solution adopted by the present invention is as follows: An optimization method for osteoarticular rehabilitation training provided by the present invention includes the following steps:

[0006] Step S1: Information extraction, using M depth cameras to obtain the depth information of the patient's limbs from different angles;

[0007] Step S2: Model generation, using a bone model generation method to generate and update a 3D human bone model;

[0008] Step S3: Memory replacement, optimizing the long short-term memory network using a memory replacement method, and using the long short-term memory network to infer the similarity between the activities of body parts and the standard human postures of medical rehabilitation training to detect the patient's execution of joint rehabilitation training instructions.

[0009] Further, in step S2, the skeletal model generation method specifically includes the following steps:

[0010] Step S21: Process the depth information using a pre-trained CNN network to generate 2D joint positions, and convert the 2D joint position data into 3D coordinates according to the camera parameters to obtain a 3D human skeletal model;

[0011] Step S22: Use all 3D human skeletal models to calculate joint positions, obtain joint position vectors, and classify the joints in the 3D human skeletal model according to anatomical structure functions to form joint groups;

[0012] Step S23: Continuously track the human body posture.

[0013] Further, in step S23, the continuous tracking of the human body posture specifically includes the following steps:

[0014] Step S231: Calculate the projection angle of the joint, and the formula used is as follows:

[0015] ;

[0016] In the formula, represents the projection angle of the joint, represents the joint position vector, represents the joint position vector of the th depth camera, represents the unit direction vector from the joint to the depth camera, represents the direction vector from the joint to the depth camera;

[0017] Step S232: Calculate the occlusion degree of the joint. Specifically, a reference threshold is preset. If the projection angle is less than the reference threshold , it is considered that the joint is not occluded, and the occlusion degree is set to 1, denoted as . If the projection angle is greater than or equal to the reference threshold , it is considered that the joint is occluded, and the occlusion degree calculation formula is as follows:

[0018] ;

[0019] In the formula, represents the occlusion degree of the th joint in the th joint group in the th 3D human skeletal model, represents the joint position vector, represents the joint position vector of the th depth camera, represents the projected angle of the current joint, represents the reference threshold;

[0020] Step S233: Evaluate the reliability of the joint group. Specifically, traverse and calculate the confidence of the joints in all joint groups, and calculate the confidence of the joint group according to the confidence of the joints and the occlusion degree. The formula used is as follows:

[0021] ;

[0022] In the formula, represents at time step , the confidence of the th joint group in the th 3D human skeleton model, represents the index of the 3D human skeleton model, represents the index of the joint group, represents the th joint group in the th 3D human skeleton model, represents the index of the joint, represents the th joint group in the th 3D human skeleton model, represents the confidence of the th joint in the th 3D human skeleton model, th joint group, represents the occlusion degree of the

[0023] Step S234: Calculate the comprehensive confidence. It is used to calculate the comprehensive confidence of all 3D human skeleton models according to the confidence of the joint groups. Specifically, first calculate the ratio of the confidence of each joint group in each 3D human skeleton model to the cumulative sum of the confidence of all joint groups to obtain the average confidence, and then calculate the cumulative sum of the average confidence of all 3D human skeleton models to obtain the comprehensive confidence;

[0024] Step S235: Update the joint position vector at the current time step. Specifically, add the joint position vector at the previous time step and the comprehensive confidence at the current time step to obtain the joint position vector at the current time step;

[0025] Step S236: Update the comprehensive confidence at the current time step. Specifically, repeat steps S231 to S234 to obtain the comprehensive confidence, and integrate the 3D human skeleton models generated by the depth camera according to their respective comprehensive confidences to obtain the final 3D human skeleton model.

[0026] Further, in step S3, the memory replacement method specifically includes the following steps:

[0027] Step S31: Extract activity features. Specifically, use a 3DCNN network as a feature extractor, and use the feature extractor to extract new long-term features and short-term features related to human activities from the 3D human bone model in real time, and record the new long-term features and short-term features as input data;

[0028] Step S32: Network replacement.

[0029] Further, in step S32, the network replacement specifically includes the following steps:

[0030] Step S321: Update the long-term memory weight vector. The formula used is as follows:

[0031] ;

[0032] ;

[0033] In the formula, represents the maximum value of the long-term memory weight vector, represents the updated value of, represents the minimum value of the long-term memory weight vector, represents the updated value of, represents the learning rate, represents the input data, represents the current time step at the value of the and respectively represent and at the dimension of, represents the global maximum value, and outputs a and scalar value that is the largest among all dimensions, represents the element-wise maximum value, and outputs a vector with the same dimension as where each dimension represents the larger value of and in that dimension;

[0034] The long-term memory weight vector is used to store the boundary values of long-term features during the learning iteration process of the long short-term memory network;

[0035] Step S322: Calculate the The matching error between a node and the current input data is calculated using the following formula:

[0036] ;

[0037] In the formula, represents the matching error of the -th node in the long short-term memory network for the input data at the current time step t, represents the square of the Euclidean distance, represents the -th node's weight vector;

[0038] Step S323: Matching nodes. Specifically, select the best-matching node corresponding to the current input data. The best-matching node is the node that is most similar to the input data, i.e., the node with the smallest matching error. The formula used is as follows:

[0039] ;

[0040] In the formula, represents the best-matching node of the current input data, represents the index of the node that makes the smallest;

[0041] Step S324: Calculate the activation value of the best-matching node. The activation value is used to measure the response intensity of the node to the input data. The formula used is as follows:

[0042] ;

[0043] In the formula, represents the activation value of the best-matching node, represents the matching error of the best-matching node to the input data;

[0044] Step S325: Preset the activation threshold. Specifically, if the activation value is less than the preset threshold, it means that the current best-matching node is not sufficient to represent the input data well, and a new node needs to be generated. Execute step S326. If the activation value is greater than or equal to the preset threshold, it means that the current best-matching node can represent the input data, and there is no need to generate a new node;

[0045] Step S326: New node generation. Specifically, construct a new node. The initial weight of the generated new node is the mean of the sum of the input data and the weight vector of the best-matching node;

[0046] Step S327: Update the regularization degree. The formula used is as follows:

[0047] ;

[0048] In the formula, Indicates the regularization degree of the node, Indicates the regularization degree of the node after update, Indicates the node 's time decay factor, Indicates the regularization control parameter;

[0049] Step S328: Update the weight vector of the node, and the formula used is as follows:

[0050] ;

[0051] In the formula, Indicates the node 's weight vector, Indicates the weight vector of the node after update ; Indicates the node 's learning rate;

[0052] Step S329: Node removal. Specifically, calculate the removal threshold. If the regularization degree of any node is greater than or equal to the removal threshold, keep the node. If the regularization degree of any node is less than the removal threshold, consider the node unimportant and remove the node. The formula for calculating the removal threshold is as follows:

[0053] ;

[0054] In the formula, Indicates the removal threshold, Indicates the average value of the regularization degrees of all nodes, Indicates the standard deviation of the regularity distribution.

[0055] The beneficial effects achieved by the present invention using the above solution are as follows:

[0056] (1) Aiming at the problem that the 3D joint position data generated by the depth camera in the existing patient pose capture technology all have certain errors and occlusions, this solution adopts a bone model generation method, uses multiple depth cameras to obtain patient limb data from different angles, generates and continuously updates the 3D human bone model, realizes continuous tracking of the human pose, and improves the stability and flexibility of the 3D human bone model;

[0057] (2) Aiming at the technical problem that the traditional long short-term memory network has the drawback of catastrophic forgetting and is not suitable for analyzing and learning rehabilitation training videos with large amounts of data, high real-time requirements, and fast update speeds, this solution adopts a memory replacement method to enable the long short-term memory network to continuously learn video data during the patient's osteoarticular rehabilitation training. Without forgetting past knowledge, the structure of the long short-term memory network is dynamically adjusted, new node paths are constructed, and unimportant nodes are removed to learn new joint movement knowledge and enhance the continuous learning ability of the long short-term memory network. Description of the Drawings

[0058] Figure 1 It is a schematic flowchart of an optimized method for osteoarticular rehabilitation training provided by the present invention;

[0059] Figure 2 It is a schematic flowchart of step S2;

[0060] Figure 3 It is a schematic flowchart of step S23;

[0061] Figure 4 It is a schematic flowchart of step S3;

[0062] Figure 5 It is a schematic flowchart of step S32;

[0063] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. Detailed Embodiments

[0064] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments; based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0065] Embodiment 1, refer to Figure 1 , an optimized method for osteoarticular rehabilitation training provided by the present invention, the method includes the following steps:

[0066] Step S1: Information extraction, using M depth cameras to obtain the depth information of the patient's limb from different angles;

[0067] Step S2: Model generation, using a bone model generation method to generate and update a 3D human bone model;

[0068] Step S3: Memory replacement. Optimize the long short-term memory network using the memory replacement method, and use the long short-term memory network to infer the similarity between the activities of body parts and the standard human postures in medical rehabilitation training, and detect the implementation of the joint rehabilitation training doctor's advice by the patient.

[0069] Embodiment 2. Refer to Figures 1 to 2 , this embodiment is based on the above embodiment. In step S2, the method for generating the bone model specifically includes the following steps:

[0070] Step S21: Process the depth information using a pre-trained CNN network to generate 2D joint positions, and convert the 2D joint position data into 3D coordinates according to the camera parameters to obtain a 3D human bone model;

[0071] Step S22: Use all 3D human bone models to calculate joint positions, obtain joint position vectors, and classify the joints in the 3D human bone model according to the anatomical structure and function to form joint groups;

[0072] Step S23: Continuous tracking of human postures.

[0073] Further, in step S23, the continuous tracking of human postures specifically includes the following steps:

[0074] Step S231: Calculate the projection angle of the joint. The formula used is as follows:

[0075] ;

[0076] In the formula, represents the projection angle of the joint, represents the joint position vector, represents the th joint position vector of the depth camera, represents the unit direction vector from the joint to the depth camera, represents the direction vector from the joint to the depth camera;

[0077] Step S232: Calculate the occlusion degree of the joint. Specifically, a reference threshold is preset. If the projection angle is less than the reference threshold , it is considered that the joint is not occluded, and the occlusion degree is set to 1, denoted as . If the projection angle is greater than or equal to the reference threshold , it is considered that the joint is occluded, and the occlusion degree calculation formula is as follows:

[0078] ;

[0079] In the formula, represents the The occlusion degree of the joint in the th joint group of a 3D human bone model, represents the joint position vector, represents the joint position vector of the th depth camera, represents the projection angle of the current joint, represents the reference threshold;

[0080] Step S233: Evaluate the reliability of the joint group. Specifically, traverse and calculate the confidence of the joints in all joint groups, and calculate the confidence of the joint group according to the confidence of the joints and the occlusion degree. The formula used is as follows:

[0081] ;

[0082] In the formula, represents the confidence of the th joint group of the st 3D human bone model at time step , represents the index of the 3D human bone model, represents the index of the joint group, represents the st number of joints detected in the th joint group of the st 3D human bone model, represents the index of the joint, represents the confidence of the th joint in the th st th joint group of the st 3D human bone model,

[0083] Step S234: Calculate the comprehensive confidence, which is used to calculate the comprehensive confidence of all 3D human bone models according to the confidence of the joint groups. Specifically, first calculate the ratio of the confidence of each joint group in each 3D human bone model to the cumulative sum of the confidence of all joint groups to obtain the average confidence, and then calculate the cumulative sum of the average confidence of all 3D human bone models to obtain the comprehensive confidence;

[0084] Step S235: Update the joint position vector at the current time step. Specifically, add the joint position vector at the previous time step and the comprehensive confidence at the current time step to obtain the joint position vector at the current time step;

[0085] Step S236: Update the comprehensive confidence of the current time step. Specifically, repeat steps S231 to S234 to obtain the comprehensive confidence, and integrate the 3D human bone models generated by the depth camera according to their respective comprehensive confidences to obtain the final 3D human bone model.

[0086] Example 4, refer to Figures 1 to 4 , based on the above example, in step S3, the memory replacement method specifically includes the following steps:

[0087] Step S31: Extract activity features. Specifically, use a 3DCNN network as a feature extractor, and use the feature extractor to extract new long-term features and short-term features related to human activities from the 3D human bone model in real time, and record the new long-term features and short-term features as input data;

[0088] Step S32: Network replacement.

[0089] Example 5, refer to Figures 1 to 5 , based on the above example, in step S32, the network replacement specifically includes the following steps:

[0090] Step S321: Update the long-term memory weight vector, and the formula used is as follows:

[0091] ;

[0092] ;

[0093] In the formula, represents the maximum value of the long-term memory weight vector, represents the updated value of represents the minimum value of the long-term memory weight vector, represents the updated value of represents the learning rate, represents the input data, represents the current time step at the value of the and respectively represent and the dimension of represents the global maximum value, and outputs a and scalar value that is the largest among all dimensions of represents the element-wise maximum value, and outputs a vector that is the same as Vectors with the same dimension, where each dimension represents the larger value under that dimension and ;

[0094] The long-term memory weight vector is used to store the boundary values of long-term features during the learning iteration of the long short-term memory network;

[0095] Step S322: Calculate the matching error between the th node and the current input data, and the formula used is as follows:

[0096] ;

[0097] In the formula, represents the matching error of the th node in the long short-term memory network to the input data at the current time step t, represents the square of the Euclidean distance, represents the weight vector of the th node;

[0098] Step S323: Match nodes. Specifically, select the best matching node corresponding to the current input data. The best matching node is the node most similar to the input data, that is, the node with the smallest matching error. The formula used is as follows:

[0099] ;

[0100] In the formula, represents the best matching node of the current input data, represents the index of the node that makes the smallest;

[0101] Step S324: Calculate the activation value of the best matching node. The activation value is used to measure the response intensity of the node to the input data. The formula used is as follows:

[0102] ;

[0103] In the formula, represents the activation value of the best matching node, represents the matching error of the best matching node to the input data;

[0104] Step S325: Preset an activation threshold. Specifically, if the activation value is less than the preset threshold, it means that the current best matching node is not sufficient to represent the input data well, and a new node needs to be generated. Execute step S326. If the activation value is greater than or equal to the preset threshold, it means that the current best matching node can represent the input data, and there is no need to generate a new node;

[0105] Step S326: Generation of a new node. Specifically, construct a new node, and the initial weight of the generated new node is the mean of the sum of the input data and the weight vector of the best-matching node.

[0106] Step S327: Update the regularization degree. The formula used is as follows:

[0107] ;

[0108] In the formula, represents the regularization degree of the node, represents the regularization degree of the node after update, represents the node 's time decay factor, represents the regularization control parameter;

[0109] Step S328: Update the weight vector of the node. The formula used is as follows:

[0110] ;

[0111] In the formula, represents the weight vector of the node , represents the weight vector of the node after update, represents the learning rate of the node ;

[0112] Step S329: Node removal. Specifically, calculate the removal threshold. If the regularization degree of any node is greater than or equal to the removal threshold, keep the node. If the regularization degree of any node is less than the removal threshold, consider the node unimportant and remove it. The formula for calculating the removal threshold is as follows:

[0113] ;

[0114] In the formula, represents the removal threshold, represents the average value of the regularization degrees of all nodes, represents the standard deviation of the regularity distribution.

[0115] Example 6. Refer to Figures 1 to 5 . Based on the above example, in step S22, the joint group specifically includes a torso group, a head group, a left upper limb group, a right upper limb group, a left lower limb group, and a right lower limb group. The torso group includes the joints of the neck, chest, and waist; the head group includes the head joints; the left upper limb group includes the joints of the left upper limb; the right upper limb group includes the joints of the right upper limb; the left lower limb group includes the joints of the left lower limb; the right lower limb group includes the joints of the right lower limb.

[0116] Example 7. Refer toFigures 1 to 5 , this embodiment is based on the above embodiment. In step S233, .

[0117] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device.

[0118] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

[0119] The above describes the present invention and its embodiments. Such a description is not restrictive. What is shown in the drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. Generally speaking, if those of ordinary skill in the art are inspired by it and design similar structural ways and embodiments without creative efforts without departing from the purpose of the present invention's creation, they shall fall within the protection scope of the present invention.

Claims

1. A bone and joint rehabilitation training optimization method, characterized in that: The method comprises the following steps: Step S1: information extraction, using M depth cameras to obtain depth information of the patient's limbs from different angles; Step S2: Model generation, using a skeleton model generation method to generate and update a 3D human skeleton model; Step S3: memory replacement, using a memory replacement method to optimize the long short-term memory network, using the long short-term memory network to infer the similarity between the activities of body parts and the standard human posture of medical rehabilitation training, and detecting the patient's execution of the doctor's instructions for joint rehabilitation training; In step S2, the skeleton model generation method specifically includes the following steps: Step S21: using a pre-trained CNN network to process the depth information, generate 2D joint positions, and convert the 2D joint position data into 3D coordinates according to camera parameters to obtain a 3D human skeleton model; Step S22: using all 3D human skeleton models, calculating joint positions, obtaining joint position vectors, and classifying the joints in the 3D human skeleton models according to anatomical structure functions to form joint groups; Step S23: Continuous tracking of human body posture; In step S23, the human body posture is continuously tracked, specifically comprising the following steps: Step S231: Calculate the projection angle of the joint using the following formula: ； In the formula, represents the projection angle of the joint, represents the joint position vector, Indicates The joint position vectors of the depth cameras, Represents the unit direction vector from the joint to the depth camera, Represents the direction vector from the joint to the depth camera; Step S232: Calculate the occlusion degree of the joint, specifically, pre-set the reference threshold , if the projection angle is less than the reference threshold , then the joint is considered not to be blocked, and the blocking degree is set to 1, recorded as , if the projection angle is greater than or equal to the reference threshold , then it is considered that the joint is blocked, and the occlusion degree calculation formula is as follows: ； In the formula, Indicates 3D human skeleton model The joint group The degree of occlusion of each joint, represents the joint position vector, Indicates The joint position vectors of the depth cameras, represents the projection angle of the current joint, represents the baseline threshold; Step S233: Evaluate the reliability of the joint group, specifically, traverse and calculate the confidence of the joints in all joint groups, and calculate the confidence of the joint group according to the confidence of the joint and the occlusion degree. The formula used is as follows: ； In the formula, Indicates that at time step At that time, 3D human skeleton model The confidence of each joint group, Represents the index of the 3D human skeleton model, represents the index of the joint group, Indicates 3D human skeleton model The number of detected joints in a joint group, represents the index of the joint, Indicates 3D human skeleton model The joint group The confidence of each joint, Indicates 3D human skeleton model The joint group The degree of occlusion of each joint; Step S234: calculating the comprehensive confidence, which is used to calculate the comprehensive confidence of all 3D human skeleton models according to the confidence of the joint group. Specifically, firstly calculating the ratio of the confidence of each joint group in each 3D human skeleton model to the cumulative sum of the confidence of all joint groups to obtain the average confidence, and then calculating the cumulative sum of the average confidence of all 3D human skeleton models to obtain the comprehensive confidence; Step S235: updating the joint position vector of the current time step, specifically, obtaining the joint position vector of the current time step by adding the joint position vector of the previous time step and the comprehensive confidence of the current time step; Step S236: updating the comprehensive confidence of the current time step, specifically, repeating steps S231 to S234 to obtain the comprehensive confidence, integrating the 3D human skeleton models generated by the depth camera according to their respective comprehensive confidences, and obtaining the final 3D human skeleton model; In step S3, the memory replacement method specifically includes the following steps: Step S31: extracting activity features, specifically, using a 3DCNN network as a feature extractor, using the feature extractor to extract new long-term features and short-term features related to human activities from the 3D human skeleton model in real time, and recording the new long-term features and short-term features as input data; Step S32: network replacement; Furthermore, in step S32, the network replacement specifically includes the following steps: Step S321: Update the long-term memory weight vector. The formula used is as follows: ； ； In the formula, represents the maximum value of the long-term memory weight vector, express The updated value of represents the minimum value of the long-term memory weight vector, express The updated value of represents the learning rate, represents the input data, Indicates the current time step hour In the The value of the dimension, and Respectively and No. Dimensions, Represents the global maximum value and outputs a and The largest scalar value across all dimensions of Represents the maximum value of each element, outputting a vectors of the same dimension, where each dimension represents the and The larger value of The long-term memory weight vector is used to store the boundary value of the long-term feature in the long short-term memory network learning iteration process; Step S322: Calculate the The matching error between a node and the current input data is calculated using the following formula: ； In the formula, Indicates that at the current time step t, the first The matching error of the nodes to the input data is represents the square of the Euclidean distance, Indicates The weight vector of each node; Step S323: Matching nodes, specifically, selecting the best matching node corresponding to the current input data. The best matching node is the node that is most similar to the input data, that is, the node with the smallest matching error. The formula used is as follows: ； In the formula, Represents the best matching node for the current input data, Indicates The index of the smallest node; Step S324: Calculate the activation value of the best matching node. The activation value is used to measure the response strength of the node to the input data. The formula used is as follows: ； In the formula, represents the activation value of the best matching node, Represents the matching error of the best matching node to the input data; Step S325: pre-set an activation threshold. Specifically, if the activation value is less than the preset threshold, it means that the current best matching node is not good enough to represent the input data, and a new node needs to be generated. Step S326 is executed. If the activation value is greater than or equal to the preset threshold, it means that the current best matching node can represent the input data, and there is no need to generate a new node. Step S326: generating a new node, specifically, constructing a new node, wherein the initial weight of the generated new node is the mean of the sum of the input data and the best matching node weight vector; Step S327: Update the regularization degree, the formula used is as follows: ； In the formula, represents the regularization degree of the node, represents the degree of regularization after the node is updated, Representation Node The time decay factor, represents the regularization control parameter; Step S328: Update the weight vector of the node. The formula used is as follows: ； In the formula, Representation Node The weight vector of Indicates the updated node The weight vector of Representation Node The learning rate; Step S329: node removal, specifically, calculating the removal threshold. If the regularization degree of any node is greater than or equal to the removal threshold, the node is retained. If the regularization degree of any node is less than the removal threshold, the node is considered unimportant and is removed. The calculation formula for the removal threshold is as follows: ； In the formula, represents the removal threshold, represents the average regularization degree of all nodes, Represents the standard deviation of the normal distribution.

Images