Neural network path planning method and system for nasal digestive tract catheterization robot

Abstract

The invention relates to the technical field of medical robot control, in particular to a neural network path planning method and system for a nasal digestive tract catheterization robot. The method comprises the following steps: establishing a digestive tract three-dimensional point cloud model; determining the pose of a path planning starting point and an end point, and generating a path point sequence; constructing an encoder network for extracting point cloud features of the digestive tract, and obtaining a trained encoder network; training the planning network by adopting the path point sequence to obtain a trained planning network; an encoder network and a planning network are combined with an alimentary canal point cloud model obtained in real time, and a path point sequence is obtained; calling a path planning network to update a global path according to environment information sensed by a sensor; the system comprises an encoder network module, a planning network module, a neural network path planning module and a continuous learning framework module. Through the mode, the efficiency of the robot in a complex, narrow and dynamic change environment is improved, and the real-time performance in an operation is met.

Description

Technical Field

[0001] The present invention relates to the field of medical robot control technology, and in particular to a neural network path planning method and system for a transnasal digestive tract catheterization robot. Background Art

[0002] Nasal intubation is one of the core technologies in clinical nursing and is widely used in enteral nutrition support, gastrointestinal decompression, drug delivery and other treatments. However, the current intubation method mainly relies on blind insertion technology, that is, the nurse measures and estimates the insertion depth of the catheter and inserts the catheter into the patient's digestive tract without direct visual guidance. This method has significant limitations, such as insufficient positioning accuracy, high risk of tissue damage, misplacement of the catheter tip and intubation failure.

[0003] Existing technologies have attempted to improve the success rate of catheterization through auxiliary technologies, such as gastroscopy, X-ray, ultrasound, and magnetic field tracking. Although these technologies have improved the accuracy of catheterization, they also have problems such as high cost, complex operation, poor patient comfort, and long-term radiation risks. In addition, most existing catheter devices are based on rigid designs and lack the flexibility to adapt to the complex digestive tract environment, making it difficult to accurately control the catheter tip in dynamically changing anatomical structures. To address these problems, rigid-flexible coupled large aspect ratio soft robots provide a new solution.

[0004] However, in the transnasal digestive tract catheterization robot system, the unstructured environment, dynamic motion characteristics and narrow and curved paths of the digestive tract make the robot inefficient in complex, narrow and dynamically changing environments, and cannot meet the real-time requirements of the operation. Summary of the invention

[0005] The purpose of the present invention is to provide a neural network path planning method and system for a transnasal digestive tract catheterization robot, aiming to solve the technical problems that robots in the prior art are inefficient in complex, narrow and dynamically changing environments and cannot meet the real-time requirements during surgery.

[0006] To achieve the above object, the present invention adopts a neural network path planning method for a nasal digestive tract catheterization robot, comprising the following steps:

[0007] Establish a 3D point cloud model of the digestive tract;

[0008] Determine the starting and ending positions of the path planning and generate a sequence of path points;

[0009] Construct an encoder network for extracting digestive tract point cloud features, and use the established digestive tract model to train the encoder network to obtain a trained encoder network;

[0010] The planning network is trained using a path point sequence to obtain a trained planning network;

[0011] The encoder network and planning network are combined with the digestive tract point cloud model acquired in real time to obtain the path point sequence;

[0012] Based on the environmental information perceived by the sensor, the path planning network is called to update the global path.

[0013] Among them, in the step of establishing the digestive tract three-dimensional point cloud model:

[0014] Construct a three-dimensional model of the digestive tract. After the point cloud is produced, the format of the three-dimensional module is converted into a point cloud format to obtain a three-dimensional point cloud model of the digestive tract. The data construction method includes using medical imaging data and using simulation tools.

[0015] Among them, in the step of determining the starting and ending positions of the path planning and generating a sequence of path points:

[0016] Predefine a variety of starting and target points to generate a high-quality path from the starting point to the target point; the features in the path point sequence include: actual path length, path smoothness, and minimum distance to obstacles.

[0017] Among them, in the step of using the path point sequence to train the planning network and obtain the trained planning network:

[0018] The potential representation of the environment, the starting point and the target point of the path provided by the encoder network are obtained and defined as input. The input is concatenated into a vector. During the training process, the vector is forward propagated through the multi-layer network to generate a path point, and it is repeated as the new current point. The iteration ends with one of the following situations: generating a path or reaching the vicinity of the target point.

[0019] Among them, in the step of obtaining a path point sequence by using the encoder network and the planning network in combination with the digestive tract point cloud model acquired in real time:

[0020] The point cloud data of the digestive tract is acquired through real-time sensors, a three-dimensional point cloud representation is generated, and normalization processing is performed; wherein the normalization processing includes mean normalization and norm normalization;

[0021] Input the point cloud data into the trained encoder network to extract the low-dimensional latent representation;

[0022] The generated potential representation, current path point and path end point are concatenated into an input vector and passed to the planning network. The next path point is generated through forward propagation. This path point is used as a new starting point and repeatedly input into the network to gradually generate a complete path point sequence until it reaches the vicinity of the target point.

[0023] Among them, in the step of obtaining a path point sequence by using the encoder network and the planning network in combination with the digestive tract point cloud model acquired in real time:

[0024] During the path generation process, the feasibility of the path is dynamically checked, and if an infeasible segment is detected in the path, the path is adjusted.

[0025] The present invention also provides a neural network path planning system for a nasal digestive tract catheterization robot, comprising an encoder network module, a planning network module, a neural network path planning module and a continuous learning framework module; wherein:

[0026] The encoder network module is used to encode environmental information into a low-dimensional potential representation;

[0027] The planning network module is used to generate an incremental path point sequence based on the potential representation of the environment, the starting point and the target point;

[0028] The neural network path planning module is used to generate an initial path by combining the starting point, the end point and the potential representation of the environment;

[0029] The continuous learning framework module is used to train the encoder network module and the planning network module in an end-to-end manner to adapt to the continuous influx of new data.

[0030] The present invention discloses a neural network path planning method and system for a transnasal digestive tract catheterization robot, the method comprising: establishing a three-dimensional point cloud model of the digestive tract; determining the posture of the starting point and the end point of the path planning, and generating a path point sequence; constructing an encoder network for extracting digestive tract point cloud features, and training the encoder network with the established digestive tract model to obtain a trained encoder network; using the path point sequence to train the planning network to obtain a trained planning network; using the encoder network and the planning network in combination with the digestive tract point cloud model acquired in real time to obtain a path point sequence; calling the path planning network to update the global path according to the environmental information sensed by the sensor; the system adopts: the encoder network module is used to encode the environmental information into a low-dimensional potential representation; the planning network module is used to generate an incremental path point sequence based on the environmental potential representation, the starting point and the target point; the neural network path planning module is used to generate an initial path in combination with the starting point, the end point and the environmental potential representation; the continuous learning framework module is used to train the encoder network module and the planning network module in an end-to-end manner to adapt to the continuous inflow of new data; through the above-mentioned method, the efficiency of the robot in a complex, narrow and dynamically changing environment is improved to meet the real-time performance in surgery. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0032] Figure 1 It is a flowchart of the steps of the neural network path planning method for the nasal digestive tract catheterization robot of the present invention.

[0033] Figure 2 It is a schematic diagram of the working mode of the planning network and the encoder network of the present invention.

[0034] Figure 3 It is a schematic diagram of the working mode of the neural network of the present invention.

[0035] Figure 4 It is a structural principle diagram of the electronic equipment of the present invention. DETAILED DESCRIPTION

[0036] Here, exemplary embodiments are described in detail, and examples thereof are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application.

[0037] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The singular forms of "a", "said" and "the" used in this application and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0038] It should be understood that although the terms first, second, third, etc. may be used in the present application to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0039] See also Figure 1 to Figure 3 ,in Figure 1 is a flowchart of the steps of the neural network path planning method for the nasogastric catheterization robot. Figure 2This is a schematic diagram of how the planning network and encoder network work. Figure 3 This is a diagram of how a neural network works.

[0040] The present invention provides a neural network path planning method for a nasal digestive tract catheterization robot, comprising the following steps:

[0041] S100: Establish a three-dimensional point cloud model of the digestive tract.

[0042] In this embodiment, a three-dimensional model of the digestive tract is constructed as the environmental basis for the robot catheterization path planning. After the point cloud is produced, the format of the three-dimensional module is converted into a point cloud format to obtain a three-dimensional point cloud model of the digestive tract; wherein the data construction method includes the following two methods: one is to use the patient's medical imaging data to generate a high-precision point cloud model of the digestive tract through three-dimensional reconstruction technology; the other is to manually construct an anatomical structure model through a simulation tool to flexibly adjust and expand the scene. After the point cloud is generated, it needs to be formatted into a point cloud format that can be processed by the path planning algorithm.

[0043] S200: Determine the starting point and end point positions of the path planning, and generate a path point sequence.

[0044] In this embodiment, a variety of starting points and target points are predefined to generate a high-quality path from the starting point to the target point; wherein the features in the path point sequence include: actual path length, path smoothness, and minimum distance to obstacles.

[0045] In the above process, in order to generate matching training samples for the path planning task, a comprehensive data generation process needs to be designed for a variety of actual scenarios and possible path conditions. Specifically, a set of diverse starting and target points are first predefined in the simulation environment. These points should cover key locations in the digestive tract, such as different target areas from the nasal entrance to the stomach, from the stomach to the duodenum, etc. This diverse configuration ensures that the model can adapt to the planning needs of various catheterization tasks. In the data generation process, an expert planner is used to generate a high-quality path from the starting point to the target point. The expert planner can generate a set of continuous state points based on the geometric characteristics of the environment, dynamic constraints, and obstacle distribution to form a path point sequence. These path point sequences not only record the motion trajectory of the robot from the starting point to the target point, but also contain additional information such as motion direction, acceleration, and path curvature, providing detailed reference for the subsequent path generation model. In order to enhance the practicality and diversity of training data, it is also necessary to simulate a variety of possible digestive tract anatomical structures and dynamic change scenarios in the simulation environment. For example, the geometry of the simulation environment is adjusted for different patient groups (such as children, adults, and patients with pathological variations), including digestive tract diameter, curvature, and branching structure. In addition, dynamic scenes such as breathing and peristalsis can be simulated to make the path planning task closer to the actual operation environment. In dynamic scenes, obstacles may move over time and the path may need to be adjusted in real time. By introducing these changes, the generated data can better train the model to cope with complex situations in real operations. In order to further improve the quality of path planning and the adaptability of the model, it is particularly important to accurately annotate the generated path data. Each path point sequence needs to record the following features: actual path length (used to evaluate the optimality of the path), path smoothness (measured by analyzing the angle changes of adjacent path points), and minimum distance to obstacles (as a safety indicator). This annotation information is not only used for training the path generation model, but also serves as an important reference for the path optimization module. Finally, in order to avoid overfitting and improve the generalization ability of the model, data augmentation technology is used to expand the generated data. Diverse data samples can be generated by applying random perturbations to path points, changing the point cloud density, and increasing environmental noise. In addition, some complex environments with random dynamic obstacles can be designed to simulate challenging scenarios that may be encountered in actual catheterization tasks. After this series of data generation and enhancement operations, a set of high-quality training data with wide coverage, rich annotation information, and both static and dynamic characteristics will be constructed, providing solid data support for path planning tasks.

[0046] S300: constructing an encoder network for extracting digestive tract point cloud features, and training the encoder network with the established digestive tract model to obtain a trained encoder network; wherein the encoder network is used to encode the cloud data into a low-dimensional potential representation.

[0047] In this embodiment, the encoder network is responsible for encoding the input complex point cloud data into a low-dimensional latent representation for path generation for subsequent path planning. The input of the encoder network is 5351×3-dimensional point cloud data, each point contains x, y, z coordinates, and the output is a 60-dimensional latent representation. First, the point cloud data is normalized to reduce the impact of data bias and imbalance on model performance. The normalization step includes calculating the mean, standard deviation, and norm of the point cloud data, and then standardizing the data with these values, and finally flattening the three-dimensional point cloud into a one-dimensional vector of 1×5351×3. The encoder network adopts the structure of a feedforward neural network, which contains multiple layers of fully connected layers to gradually compress the feature dimensions. The design of the hidden layer consists of multiple layers of neurons, 4096, 1024, and 256, etc., which gradually extract and condense features, and finally output a 60-dimensional latent representation through a fully connected layer. The training goal of the network is to minimize the reconstruction loss Ensure that the potential representation Z can effectively preserve the geometric characteristics of the point cloud environment. The Adam optimizer is used for training, with an initial learning rate of 0.001, and the network parameters are optimized by batch gradient descent. The training termination condition is that the reconstruction error on the validation set is lower than the preset threshold of 10 -3 , or the loss function converges after multiple iterations. In the verification phase, different dynamic scenes and complex environments are used to test the compression ability and generalization of the potential representation, and the model performance is evaluated by comparing the similarity between the original point cloud and the reconstructed point cloud. The trained and verified Enet can accurately extract the point cloud environment features and provide an efficient and reliable low-dimensional environment representation for the path planning module.

[0048] S400: Using the path point sequence to train the planning network to obtain a trained planning network.

[0049] In this embodiment, the potential representation of the environment, the starting point and the target point of the path provided by the encoder network are obtained and defined as input, and the input is concatenated into a vector. During the training process, the vector is forward propagated through a multi-layer network to generate a path point, and it is repeatedly iterated as the new current point. The iteration ends with one of the following situations: generating a path or reaching the vicinity of the target point.

[0050] In the above process: the planning network is trained using the path point sequence obtained by traditional methods to obtain a trained planning network. The planning network adopts a multi-layer perceptron architecture, which aims to init and target point c goalGenerate a sequence of waypoints. The network input consists of three parts: the potential representation Z of the environment provided by the encoder network (dimension is 1×60), the starting point and the target point of the path (both are 1×3 three-dimensional coordinates), which are concatenated into a vector of size 1×66. The core of the planning network is a deep fully connected network (FCN), which contains multiple layers of linear transformations, each layer is followed by a PReLU activation function and Dropout regularization to enhance the nonlinear expression ability of the network and prevent overfitting. The specific structure includes 8 hidden layers connected after the input layer, and the number of neurons decreases layer by layer: 1280, 896, 512, 384, 256, 128, 64 and 32. Finally, a linear layer is used to map the hidden representation to the output layer to generate a single waypoint (dimension is 1×3). The network is built using PyTorch's nn.Sequential to ensure clear hierarchy and easy scalability. During the training process, the hyperparameters of the planning network are set to batch size (batchsize) 100, learning rate (learning rate) 0.001, and training rounds (numepochs) 200. The waypoint sequence generated by Oracle Planner is used as the supervision target, and the waypoint prediction error is used as the loss function to calculate the Euclidean distance between the predicted waypoint and the target waypoint:

[0051]

[0052] in and c i are the prediction point and the target point respectively, and N is the number of path points. During the training process, the concatenated input vector is forward propagated through a multi-layer network to generate path points, and it is repeated as the new current point until the path is completely generated or reaches the vicinity of the target point. The Adam optimizer is used in training, combined with Dropout and PReLU to improve the generalization performance, and the stability and convergence of the network are ensured through 200 training rounds. In the verification stage, the path generation error, smoothness and arrival success rate of the planning network are evaluated through path prediction tests in a variety of scenarios to ensure its application reliability and adaptability in dynamic and complex environments. The deep structure and reasonable training configuration of the planning network enable it to efficiently generate high-precision path point sequences, providing key support for path planning.

[0053] S500: Utilize the encoder network and planning network combined with the digestive tract point cloud model acquired in real time to obtain a path point sequence.

[0054] S600: Calling a path planning network to update a global path according to environmental information sensed by the sensor.

[0055] In this embodiment, point cloud data of the digestive tract is acquired by real-time sensors, a three-dimensional point cloud representation is generated, and normalization processing is performed; wherein the normalization processing includes mean normalization and norm normalization; the point cloud data is input into a trained encoder network to extract a low-dimensional potential representation; the generated potential representation, the current path point, and the path end point are concatenated as an input vector, which is passed to the planning network, and the next path point is generated through forward propagation. This path point is used as a new starting point and repeatedly input into the network to gradually generate a complete path point sequence until it reaches the vicinity of the target point; during the path generation process, the feasibility of the path is dynamically checked, and if an infeasible segment in the path is detected, the path is adjusted.

[0056] In the above steps, the trained encoder network and planning network are used in combination with the real-time digestive tract point cloud model to generate a path point sequence. First, the point cloud data of the digestive tract is acquired through real-time sensors to generate a 5351×3 three-dimensional point cloud representation, and normalization is performed, including mean normalization and norm normalization, to ensure the consistency of data distribution. Then, the point cloud data is input into the trained encoder network to extract the low-dimensional latent representation Z to efficiently describe the complex digestive tract geometric environment. Based on the generated latent representation Z, the current path point and the end point of the path, the three are concatenated into an input vector and passed to the planning network to generate the next path point through forward propagation. This path point is used as a new starting point and repeatedly input into the network to gradually generate a complete path point sequence until it reaches the vicinity of the target point. During the path generation process, the feasibility of the path is dynamically checked. If an infeasible segment in the path is detected (such as crossing an obstacle or not meeting the planned area), the path is adjusted in combination with the real-time point cloud and OraclePlanner. OraclePlanner can generate a global optimized path based on the point cloud environment and the target point, ensure the safety and feasibility of the path, and integrate the optimized path into the complete path sequence. At the same time, the high-quality path generated by OraclePlanner is used as the input of the continuous learning module to dynamically update the model parameters of the planning network. Through the gradient memory playback technology, the continuous learning module maintains the memory library of old task samples to prevent catastrophic forgetting caused by new data learning. Combining real-time path data with historical samples in the memory library, the planning network is optimized using a joint loss function to achieve a balance between new task learning and old task performance. Finally, the complete path point sequence is output as a set of continuous state points, describing the motion trajectory from the starting point to the target point, and verifying its smoothness, feasibility and dynamic obstacle avoidance capabilities. By introducing OraclePlanner and the continuous learning mechanism, the present invention continuously improves the adaptability and reliability of the path planning system to ensure its efficient performance in complex dynamic environments.

[0057] The present invention provides a neural network path planning system for a nasal digestive tract catheterization robot, comprising an encoder network module, a planning network module, a neural network path planning module and a continuous learning framework module; wherein:

[0058] The encoder network module is used to encode environmental information into a low-dimensional potential representation;

[0059] The planning network module is used to generate an incremental path point sequence based on the potential representation of the environment, the starting point and the target point;

[0060] The neural network path planning module is used to generate an initial path by combining the starting point, the end point and the potential representation of the environment;

[0061] The continuous learning framework module is used to train the encoder network module and the planning network module in an end-to-end manner to adapt to the continuous influx of new data.

[0062] In this embodiment, the encoder network module encodes the environment information into a low-dimensional potential representation:

[0063] The input of the encoder network module is the digestive tract point cloud model, the network architecture is a fully connected neural network, and the output is a low-dimensional embedding in the generated latent space, which is used as a representation of environmental information for the planning network. The training method is divided into independent training and end-to-end training, where independent training uses an autoencoder architecture for unsupervised learning to learn the latent space representation from point cloud data. The loss function includes reconstruction error and parameter regularization terms, and its expression is:

[0064]

[0065] End-to-end training uses the objective loss function of the path planning task for gradient backpropagation.

[0066] The planning network module generates an incremental sequence of path points based on the potential representation of the environment, the starting point and the target point:

[0067] The rules of path planning are learned through neural networks, and feasible paths from the starting point to the target point are gradually predicted in a low-dimensional latent space. The input of the planning network module includes the potential representation of the environment generated by the encoder network module, as well as the coordinate information of the starting point and the target point, and the output is a sequence of path points. The internal structure of the planning network module is usually composed of a multi-layer fully connected network or a recursive neural network. Through these network levels, the planning network module establishes a mapping relationship between input features and path points. During the working process, the planning network module adopts an incremental planning method, predicting the first path point from the starting point, and then using the predicted path point as part of the next input, iteratively generating a complete path until the target point is reached. The loss function expression of the planning network module is:

[0068]

[0069] The neural network path planning module combines the starting point, the end point and the potential representation of the environment to generate an initial path:

[0070] First, the encoder network module is used to encode the point cloud input of the digestive tract to generate a potential representation of the environment. A bidirectional neural planner is used to combine the starting point, the end point and the potential representation of the environment to generate an initial path.

[0071] The path is optimized by the lazy state compression algorithm. Check whether the optimized path is feasible. If it is feasible, the path is returned directly. If the path is not feasible, the path replanning phase is entered. Replanning is divided into two phases. The first phase is replanning based on the neural network. In the loop of replanning iteration Nr times, the replanning function is used to regenerate the path by combining the existing path and the potential representation of the environment. The generated path is optimized again by lazy state compression. Check the feasibility of the path. If the generated path is feasible, the path is returned. The second phase is hybrid replanning. If no feasible path is found in the first phase, hybrid replanning based on the expert planner is enabled. The replanning function is called again to regenerate the path by combining the path, the potential representation of the environment and the expert planner. The regenerated path is optimized again by lazy state compression. Check the feasibility of the path. If the path is feasible, the generated path is returned.

[0072] Bidirectional neural planner and replanning are important modules for achieving path generation and optimization, working together at different stages to ensure the feasibility and optimization effect of the final path. First, the bidirectional neural planner is responsible for generating the initial path through bidirectional neural planning. Specifically, two paths are initialized from the starting point and the end point respectively, and then through loop iteration, the planning network module is used to predict the next point of the path and add it to the current path. At the same time, the steerTo function is called to try to connect the end of the current path to the end of the target path. If the connection is successful, the two paths are spliced ​​into a complete path and returned. Otherwise, the starting point and the end point of the path are exchanged and the iteration is continued. If no feasible path is found after all iterations are completed, an empty path is returned. After the path generation is completed, if the path is still not feasible, the replanning module is entered, each segment of the path is checked point by point, and the current point is connected to the next point through the steerTo function. If the connection fails, the current two points are set as beacon nodes, and the beacon nodes are used as similar initial and end positions for bidirectional neural planning. If the number of cycles is less than Nr, the bidirectional neural planning module is used for local replanning. If the number of cycles is greater than Nr, the expert planner is selected for replanning. If the replanning is successful, the regenerated subpath will be added to the new path; if the replanning fails, an empty path is returned to indicate planning failure. Finally, the replanning module returns the updated path as the final optimization result.

[0073] The continuous learning framework module trains the encoder network module and the planning network module in an end-to-end manner to adapt to the continuous inflow of new data:

[0074] In the continuous learning framework module, the encoder network module and the planning network module are trained in an end-to-end manner to adapt to the continuous inflow of new data. The input data consists of state and target value pairs, expressed as:

[0075] (s1,y1,…,s i ,y i ,…,s N ,y N );

[0076] Where s=(c t ,c r ,x obs ) is the input of the neural network path planning module, including the current state of the robot c t 、Target state c r , and obstacle information x obs The goal y is the next state c provided by the expert planner. t+1The neural network path planning module in continuous learning will face the problem of catastrophic forgetting. When the gradient is updated for new data, the previously learned knowledge may be erased. In order to describe the problem of catastrophic forgetting, the concept of Backward Transfer is introduced. Suppose that on a dataset M at time t-1 and t, the average success rate based on expert examples is A M,t-1 and A M,t . The reverse migration B is defined as:

[0077] B=A M,t -A M,t-1 ;

[0078] If B ≥ 0, it means that after learning the new task, the performance of the old task is improved (forward reverse migration). If B < 0, it means that after learning the new task, the performance of the old task is reduced (catastrophic forgetting). Use the gradient memory replay method to deal with the problem of catastrophic forgetting. Gradient memory replay uses a memory bank M containing finite continuous data to ensure that the performance of the old task will not be reduced when the model is updated, while allowing forward reverse migration. For the neural network path planning module, the gradient memory replay is adapted to the regression problem, and its optimization goal is:

[0079]

[0080] in, is the squared error loss function, f θ Represents the neural network path planning module at time t. To ensure the constraints, gradient memory replay projects the updated gradient to the nearest gradient that meets the constraints through the following gradient projection formula:

[0081]

[0082] Among them, g M is the gradient on the memory data, representing the knowledge of the old task. The continuous learning framework adopts a global distribution matching strategy to fill the memory through random sampling techniques to capture the global distribution of the data stream as closely as possible. When a new sample is received, the probability is used:

[0083]

[0084] The old samples in the memory bank are replaced with probabilities, thereby dynamically updating the memory bank. In addition, the neural network path planning module also maintains a replay buffer B * , which is used to further alleviate catastrophic forgetting by learning from old samples. This replay mechanism can significantly improve model performance and jointly train the data of old and new tasks in each replay cycle.

[0085] Regarding the system in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0086] For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present application. A person of ordinary skill in the art can understand and implement it without creative work.

[0087] Accordingly, the present application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned high-reliability learning resource security sharing method based on big data. Figure 4 As shown, a hardware structure diagram of a high-reliability learning resource security sharing system based on big data provided by an embodiment of the present invention is provided for any device with data processing capabilities, except Figure 4 In addition to the processor, memory and network interface shown, any device with data processing capability in which the apparatus in the embodiment is located may also include other hardware, generally based on the actual functions of the device with data processing capability, which will not be described in detail.

[0088] Accordingly, the present application also provides a computer-readable storage medium on which computer instructions are stored, and when the instructions are executed by the processor, the above-mentioned high-trust learning resource security sharing method based on big data is implemented. The computer-readable storage medium can be an internal storage unit of any device with data processing capabilities described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), an SD card, a flash card (Flash Card), etc. equipped on the device. Furthermore, the computer-readable storage medium can also include both an internal storage unit of any device with data processing capabilities and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and can also be used to temporarily store data that has been output or is to be output.

[0089] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the contents disclosed herein. The present application is intended to cover any variations, uses or adaptations of the present application, which follow the general principles of the present application and include common knowledge or customary technical means in the art that are not disclosed in the present application.

[0090] It should be understood that the present application is not limited to the exact construction that has been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof.

Claims

1. A neural network path planning method for a nasal digestive tract catheterization robot, characterized in that: The steps include: Establish a 3D point cloud model of the digestive tract; Determine the starting and ending positions of the path planning and generate a sequence of path points; Construct an encoder network for extracting digestive tract point cloud features, and use the established digestive tract model to train the encoder network to obtain a trained encoder network; The planning network is trained using a path point sequence to obtain a trained planning network; The encoder network and planning network are combined with the digestive tract point cloud model acquired in real time to obtain the path point sequence; Based on the environmental information perceived by the sensor, the path planning network is called to update the global path.

2. The neural network path planning method for a nasal digestive tract catheterization robot according to claim 1, characterized in that: In the steps of building the 3D point cloud model of the digestive tract: Construct a three-dimensional model of the digestive tract. After the point cloud is produced, the format of the three-dimensional module is converted into a point cloud format to obtain a three-dimensional point cloud model of the digestive tract. The data construction method includes using medical imaging data and using simulation tools.

3. The neural network path planning method for a nasal digestive tract catheterization robot according to claim 2, characterized in that: In the steps of determining the starting and ending positions of path planning and generating a sequence of path points: Predefine various starting and target points to generate high-quality paths from the starting point to the target point; The features of the path point sequence include: actual path length, path smoothness, and minimum distance to obstacles.

4. The neural network path planning method for a nasal digestive tract catheterization robot according to claim 3, characterized in that: In the step of using the path point sequence to train the planning network and obtain the trained planning network: The potential representation of the environment, the starting point and the target point of the path provided by the encoder network are obtained and defined as input. The input is concatenated into a vector. During the training process, the vector is forward propagated through the multi-layer network to generate a path point, and it is repeated as the new current point. The iteration ends with one of the following situations: generating a path or reaching the vicinity of the target point.

5. The neural network path planning method for a nasal digestive tract catheterization robot according to claim 4, characterized in that: In the step of obtaining a sequence of waypoints by combining the encoder network and the planning network with the digestive tract point cloud model acquired in real time: The point cloud data of the digestive tract is acquired through real-time sensors, a three-dimensional point cloud representation is generated, and normalization processing is performed; wherein the normalization processing includes mean normalization and norm normalization; Input the point cloud data into the trained encoder network to extract the low-dimensional latent representation; The generated potential representation, current path point and path end point are concatenated into an input vector and passed to the planning network. The next path point is generated through forward propagation. This path point is used as a new starting point and repeatedly input into the network to gradually generate a complete path point sequence until it reaches the vicinity of the target point.

6. The neural network path planning method for a nasal digestive tract catheterization robot according to claim 5, characterized in that: In the step of obtaining a sequence of waypoints by combining the encoder network and the planning network with the digestive tract point cloud model acquired in real time: During the path generation process, the feasibility of the path is dynamically checked, and if an infeasible segment is detected in the path, the path is adjusted.

7. A neural network path planning system for a transnasal digestive tract catheterization robot, applied to the neural network path planning method for a transnasal digestive tract catheterization robot as claimed in claim 1, characterized in that: It includes encoder network module, planning network module, neural network path planning module and continuous learning framework module; among which: The encoder network module is used to encode environmental information into a low-dimensional potential representation; The planning network module is used to generate an incremental path point sequence based on the potential representation of the environment, the starting point and the target point; The neural network path planning module is used to generate an initial path by combining the starting point, the end point and the potential representation of the environment; The continuous learning framework module is used to train the encoder network module and the planning network module in an end-to-end manner to adapt to the continuous influx of new data.

Images