Hydrogen refueling port position recognition system and automatic hydrogen refueling system based on cryogenic refueling

Abstract

This invention pertains to cryogenic refueling technology, specifically a hydrogen refueling port pose recognition system and an automated hydrogen refueling system based on cryogenic refueling. The hydrogen refueling port pose recognition system includes an image acquisition device for acquiring images of the hydrogen refueling port; a coarse localization unit for extracting features from the hydrogen refueling port image, fusing local and global features to obtain features with CNN-style and Transformer-style characteristics, and classifying these features to obtain two-dimensional position information of the hydrogen refueling port; and a precise localization unit for acquiring three-dimensional point cloud information of the hydrogen refueling port based on the two-dimensional position information, combining the three-dimensional point cloud information and the three-dimensional model information of the hydrogen refueling port, using `create_surface_model` to create a template, locating the template in the 3D scene, and calculating the pose of the hydrogen refueling port. This invention enables automated cryogenic hydrogen refueling, effectively reducing labor costs at hydrogen refueling stations and improving the safety level of hydrogen refueling at these stations.

Description

Technical Field

[0001] This invention belongs to the field of cryogenic refueling technology, specifically to cryogenic hydrogen refueling technology, and more specifically, to a hydrogen refueling port position recognition system and an automatic hydrogen refueling system based on cryogenic refueling. Background Technology

[0002] Refueling stations are refueling facilities for new energy vehicles, providing fuel such as liquefied natural gas (LNG) and hydrogen. Currently, most refueling stations use manual methods for cryogenic hydrogen and liquid refueling. However, this method is labor-intensive, operates in harsh environments, and is prone to errors during prolonged outdoor operations in low temperatures, posing significant safety risks associated with hydrogen-related operations. Summary of the Invention

[0003] To address the aforementioned problems in the prior art, this invention provides a hydrogen refueling port position recognition system and an automatic hydrogen refueling system based on cryogenic refueling, which can realize automatic hydrogen refueling under low temperatures, effectively reduce the labor costs of hydrogen refueling stations, and improve the safety level of hydrogen refueling at hydrogen refueling stations.

[0004] To achieve the above objectives, the present invention provides a hydrogen refueling port pose recognition system, comprising:

[0005] Image acquisition device, used to acquire images of the hydrogen filling port;

[0006] The hydrogen refueling port coarse localization unit extracts features from the hydrogen refueling port image, fuses local and global features to obtain features with CNN style and Transformer style, and classifies the features to obtain the two-dimensional location information of the hydrogen refueling port.

[0007] The hydrogen refueling port precise positioning unit obtains the three-dimensional point cloud information of the hydrogen refueling port based on the two-dimensional position information of the hydrogen refueling port. Combining the three-dimensional point cloud information and the three-dimensional model information of the hydrogen refueling port, it uses create_surface_model to create a template, finds the template in the 3D scene, and calculates the pose of the hydrogen refueling port.

[0008] Preferably, the coarse positioning unit for the hydrogenation port includes:

[0009] The stem module is used to extract initial local features from the hydrogen refueling port image;

[0010] The two branches are connected to the stem module. The two branches include a CNN branch for extracting local features from the initial local features and a Transformer branch for extracting all features from the initial local features.

[0011] FCU, the bridged dual-branch module, is used to interactively fuse local features extracted by the CNN branch with global features extracted by the Transformer branch.

[0012] The first classifier is used to receive all features from the CNN branch, classify them, and output the results.

[0013] The second classifier is used to receive all features from the Transformer branch, classify them, and output the results.

[0014] The output module is used to add the output of the first classifier to the output of the second classifier to obtain the coarse localization result.

[0015] Preferably, the stem module consists of a 7×7 convolutional layer with a stride of 2 and a 3×3 max pooling layer with a stride of 2.

[0016] Preferably, the CNN branch adopts a feature pyramid structure, which is divided into four stages. Each stage consists of multiple convolutional blocks, and each convolutional block contains n bottlenecks. Following the definition in ResNet, each bottleneck contains a 1×1 convolutional layer, a 3×3 spatial convolutional layer, a 1×1 convolutional layer, and residual connections between the input and output. The Transformer branch contains N repeating transformer blocks. Each transformer block consists of a multi-head self-attention layer and an MLP layer. Layer norm is applied before each layer, and there are residual connections in both the self-attention layer and the MLP layer.

[0017] Preferably, when training the CNN branch and the Transformer branch, two cross-entropy functions with the same importance weights in the loss function are used to supervise the first classifier and the second classifier respectively, so as to obtain features that have both CNN style and Transformer style.

[0018] Preferably, the specific steps for creating the template are as follows: Multiple cameras are evenly distributed in a 3D scene to simulate shooting the 3D model of the hydrogen refueling port from different angles, obtaining different angle information of the 3D model of the hydrogen refueling port, and storing the views from all perspectives in a 3D model; denoising and filtering are performed on the 3D point cloud information of the hydrogen refueling port obtained based on the 2D position information of the hydrogen refueling port to obtain point cloud information containing only the hydrogen refueling port; the point cloud information containing only the hydrogen refueling port is combined with the 3D model information obtained from the 3D model of the hydrogen refueling port to obtain the final model; the second-order mean or the mean of the central moments of the final model is calculated using create_surface_model, and an affine transformation is performed to generate the template coordinate system to obtain the template.

[0019] Preferably, the specific steps for finding the template and calculating the hydrogen refueling port pose in the 3D scene are as follows:

[0020] Uniform sampling is performed on template instances in the 3D scene, with 50-150 scene points sampled for each template instance. A set of key points is selected from the sampled scene points, and the number of key points selected is controlled by the parameter KeyPoint.

[0021] For each selected key point, under the assumption that the key point is located on the surface of the object, the optimal pose of the template is calculated. The pose with the highest score among all key points is selected as the approximate pose, which is the hydrogen filling port pose.

[0022] To achieve the above objectives, an automated hydrogen refueling system based on cryogenic refueling includes a cryogenic explosion-proof hydrogen refueling robot and the aforementioned hydrogen refueling port pose recognition system. The image acquisition device of the hydrogen refueling port pose recognition system is located outside the hydrogen refueling robot, while the coarse positioning unit and precise positioning unit of the hydrogen refueling port pose recognition system are located inside the hydrogen refueling robot. The image acquisition device employs a cryogenic explosion-proof industrial 3D camera. The precise positioning unit of the hydrogen refueling port is connected to a control module located inside the hydrogen refueling robot. The control module controls the movement of the robotic arm of the hydrogen refueling robot based on the hydrogen refueling port pose calculated by the precise positioning unit. The robotic arm drives the hydrogen refueling nozzle mounted on the robotic arm to automatically dock with the hydrogen refueling port position for automated hydrogen refueling.

[0023] Furthermore, it also includes a pose correction module located inside the hydrogen refueling robot. When the pose of the hydrogen refueling port calculated by the precise positioning unit is not within the flexible pose space of the robotic arm, it considers the isotropic shape of the hydrogen refueling port, generates a hydrogen refueling port pose matrix based on the hydrogen refueling port pose, and generates multiple new poses rotating around the z-axis while ensuring that the z-axis orientation remains unchanged. It then performs inverse kinematics calculation on the new poses to calculate the cost function of each set of joint solutions and determine the optimal pose. Here, the z-axis is the position and orientation of the end of the sixth axis of the robotic arm in the base coordinate system of the robotic arm.

[0024] Furthermore, it also includes a motion correction module located inside the hydrogen refueling robot. When the error between the calculated hydrogen refueling port pose and the actual hydrogen refueling port pose affects coaxial alignment, the module uses the contact force feedback between the end of the robotic arm and the hydrogen refueling port to compensate for the hole-axis alignment movement when coaxial alignment fails, so as to ensure the completion of the hydrogen refueling movement.

[0025] Preferably, the hydrogen refueling port precise positioning unit sets a minimum score threshold. If the score exceeds the minimum threshold, the calculated hydrogen refueling port pose is transmitted to the control module, and the control module controls the robotic arm movement based on the hydrogen refueling port pose.

[0026] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0027] (1) The hydrogen refueling port pose recognition system of the present invention adopts a combination of 2D deep learning and 3D template matching. It uses 2D deep learning for coarse positioning and 3D point cloud matching for precise positioning, thereby improving the positioning accuracy of the hydrogen refueling port.

[0028] (2) The hydrogen refueling port pose recognition system of the present invention uses a combination of CNN branch and Transformer branch in deep learning. Based on multi-level Transformer, traditional convolution is inserted between the layers of the network. Local and global features of the image are extracted hierarchically by traditional convolution + global attention. At the same time, a feature coupling unit (i.e. FCU) is designed to fuse local and global features, which enhances the visual representation capability in an interactive way and improves the coarse positioning accuracy of the hydrogen refueling port.

[0029] (3) The hydrogen refueling port pose recognition system of the present invention calculates the hydrogen refueling port pose by using a 3D template matching algorithm based on create_surface_model to obtain the six-dimensional pose information of the hydrogen refueling port, thereby achieving accurate positioning of the hydrogen refueling port.

[0030] (4) The automatic hydrogen refueling system of the present invention includes a hydrogen refueling robot and a hydrogen refueling port posture recognition system. The hydrogen refueling robot is a low-temperature explosion-proof hydrogen refueling robot made of low-temperature explosion-proof parts, and the image acquisition device is a low-temperature explosion-proof industrial 3D camera, which can realize automatic recognition, compliant docking and explosion-proof and cold-proof of hydrogen refueling port under low temperature conditions.

[0031] (5) The automatic hydrogen refueling system of the present invention is equipped with a pose correction module. When the calculated pose of the hydrogen refueling port is not within the flexible pose space of the robotic arm, the isotropic shape of the hydrogen refueling port is considered, and the pose is corrected while ensuring that the z-axis orientation remains unchanged, so that the robotic arm can complete the hydrogen refueling task more flexibly.

[0032] (6) The automatic hydrogen refueling system of the present invention is further provided with a motion correction module. When the error between the calculated hydrogen refueling port pose and the actual hydrogen refueling port pose affects the coaxial alignment, the contact force feedback between the end of the robotic arm and the hydrogen refueling port is used to compensate for the hole-axis alignment movement when the coaxial alignment fails, so as to ensure the completion of the hydrogen refueling movement. Attached Figure Description

[0033] Figure 1 This is a structural block diagram of the hydrogen refueling port pose recognition system described in Embodiment 1 of the present invention;

[0034] Figures 2a-2c This is a structural diagram of the coarse positioning unit for the hydrogenation port described in Embodiment 1 of the present invention;

[0035] Figure 3 This is a flowchart illustrating the template creation process described in Embodiment 1 of the present invention;

[0036] Figures 4a-4c This is a schematic diagram illustrating the process of creating a template in Embodiment 1 of the present invention;

[0037] Figure 5 This is a flowchart illustrating the pose correction module's pose correction process as described in Embodiment 2 of the present invention.

[0038] Figure 6 This is a schematic diagram of the position and orientation of the end of the sixth axis of the robotic arm in the base coordinate system of the robotic arm according to Embodiment 2 of the present invention;

[0039] Figure 7 This is a flowchart illustrating the alignment motion correction module described in Embodiment 2 of the present invention.

[0040] Figures 8a-8d This is a schematic diagram of the alignment correction process of the motion correction module described in Embodiment 2 of the present invention;

[0041] Figure 9 This is a schematic diagram of the trajectory of the alignment motion corrected by the motion correction module described in Embodiment 2 of the present invention.

[0042] In the figure, 1 is the image acquisition device, 2 is the coarse positioning unit for the hydrogen refueling port, 21 is the STEM module, 22 is the CNN branch, 23 is the Transformer branch, 24 is the FCU, 25 is the first classifier, 26 is the second classifier, 27 is the output module, 3 is the precise positioning unit for the hydrogen refueling port, 31 is the template generation module, and 32 is the pose calculation module. Detailed Implementation

[0043] The present invention will now be described in detail through exemplary embodiments. However, it should be understood that, without further description, elements, structures, and features in one embodiment may be advantageously incorporated into other embodiments.

[0044] In the description of this invention, it should be noted that the terms "upper," "lower," "left," "right," "inner," and "outer," etc., indicate the orientation or positional relationship based on the positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0045] Example 1: See Figure 1 This embodiment provides a hydrogen refueling port pose recognition system, including:

[0046] Image acquisition device 1 is used to acquire images of the hydrogen filling port;

[0047] The hydrogen refueling port coarse localization unit 2 extracts features from the hydrogen refueling port image, fuses local and global features to obtain features with CNN style and Transformer style, and classifies the features to obtain the two-dimensional location information of the hydrogen refueling port.

[0048] The hydrogen refueling port precise positioning unit 3 obtains the three-dimensional point cloud information of the hydrogen refueling port based on the two-dimensional position information of the hydrogen refueling port. It combines the three-dimensional point cloud information and the three-dimensional model information of the hydrogen refueling port with the create_surface_model to create a template, finds the template in the 3D scene, and calculates the pose of the hydrogen refueling port.

[0049] Specifically, see [link to relevant documentation] Figure 1 , and see Figures 2a-2c The coarse positioning unit 2 for the hydrogenation port includes:

[0050] stem module 21 is used to extract initial local features from the hydrogen filling port image;

[0051] The two branches are connected to the stem module. The two branches include a CNN branch 22 for extracting local features from the initial local features and a Transformer branch 23 for extracting all features from the initial local features.

[0052] FCU 24, bridging the two branches, serves as a bridging module to interactively fuse the local features extracted by CNN branch 22 with the global features extracted by Transformer branch 23.

[0053] The first classifier 25 is used to receive all features from the CNN branch, classify them, and output the results.

[0054] The second classifier 26 is used to receive all features from the Transformer branch, classify them, and output the results.

[0055] Output module 27 is used to add the output of the first classifier and the output of the second classifier to obtain the coarse localization result.

[0056] Specifically, in this embodiment, the stem module consists of a 7×7 convolutional layer with a stride of 2 and a 3×3 max pooling layer with a stride of 2.

[0057] Specifically, in this embodiment, the CNN branch adopts a feature pyramid structure, divided into four stages. Each stage consists of multiple convolutional blocks, and each convolutional block contains n bottlenecks. Following the definition in ResNet, each bottleneck contains a 1×1 convolutional layer, a 3×3 spatial convolutional layer, a 1×1 convolutional layer, and a residual connection between the input and output. In this embodiment, n is set to 1 in the first convolutional block and satisfies ≥2 in the subsequent N-1 convolutional blocks.

[0058] Transformers project image patches into vectors in a single step, resulting in the loss of local details. In contrast, CNNs use overlapping feature maps through which convolutional kernels slide, offering the possibility of preserving fine local features. Therefore, CNN branches can continuously provide local feature details to transformer branches.

[0059] Specifically, in this embodiment, the Transformer branch contains N repeating Transformer blocks, each Transformer block consisting of a multi-head self-attention layer and an MLP layer. Layernorm is applied before each layer, and there are residual connections in both the self-attention layer and the MLP layer.

[0060] It's worth noting that the feature maps generated by the stem module are compressed into 14×14 patch embeddings without overlap, through a linear projection layer, which is a 4×4 convolution with a stride of 4. Considering that the CNN branches (3×3 convolutions) encode local features and spatial location information, positional embedding is no longer needed. This helps improve image resolution for downstream vision tasks.

[0061] Specifically, when training the CNN and Transformer branches, two cross-entropy functions with the same importance weights in the loss function are used to supervise the first and second classifiers respectively, in order to obtain features that have both CNN and Transformer styles.

[0062] It's important to note that, given the feature map in the CNN branch and the patch embedding in the Transformer branch, eliminating the misalignment between them is a crucial issue. To address this, this embodiment uses the FCU to interactively couple local features with the global representation. On one hand, we must recognize that the feature dimensions of the CNN and Transformer are inconsistent. The CNN feature map has dimensions C×H×W (C, H, and W are channels, height, and width), while the patch embeddings have a shape of (K+1)×E, where K, 1, and E represent the number of image patches, class tokens, and embedding dimensions, respectively. When fed into the Transformer branch, the feature map first needs to be aligned with the number of channels in the patch embedding using a 1×1 convolution. Then, a downsampling module (see...) is used... Figure 2a Complete spatial dimension alignment. Finally, see Figure 2bThe feature map is then fitted with patch embeddings. When feeding back from the Transformer branch to the CNN branch, the patch embeddings need to be upsampled to align the spatial scale. Then, the channel dimensions are aligned with the dimensions of the CNN feature map using 1×1 convolutions and added to the feature map. Simultaneously, LayerNorm and BatchNorm modules are used to regularize the features. On the other hand, there is a significant semantic gap between the feature map and the patch embedding: the feature map is collected from local convolutional operators, while the patch embedding is aggregated globally. Therefore, FCU is applied in each block (except the first) to progressively fill in the semantic gaps.

[0063] Adding a Transformer structure to a traditional CNN can combine local features based on CNNs with global features based on Transformers to enhance representation learning and improve detection and localization results.

[0064] See also Figure 1 The hydrogen refueling port precise positioning unit 3 includes:

[0065] Template generation module 31 is used to obtain the three-dimensional point cloud information of the hydrogen filling port based on the two-dimensional location information of the hydrogen filling port, and to create a template by combining the three-dimensional point cloud information of the hydrogen filling port and the three-dimensional model information of the hydrogen filling port using create_surface_model;

[0066] The pose calculation module 32 is used to find the template in the 3D scene and calculate the pose of the hydrogen refueling port.

[0067] See Figure 3 and Figures 4a-4c The specific steps for creating the template are as follows: Multiple cameras are evenly distributed in a 3D scene to simulate and capture images of the 3D model of the hydrogen refueling port from different angles, obtaining different angle information of the 3D model of the hydrogen refueling port. Views from all perspectives are stored in a single 3D model. The 3D point cloud information of the hydrogen refueling port, obtained from the 2D position information of the hydrogen refueling port, is denoised and filtered to obtain point cloud information containing only the hydrogen refueling port. This point cloud information is combined with the 3D model information obtained from the 3D model of the hydrogen refueling port to obtain the final model. The second-order mean or mean central moment of the final model is calculated using the create_surface_model algorithm, and an affine transformation is performed to generate the template coordinate system to obtain the template. Specifically, in this embodiment, 12 cameras are evenly distributed in the 3D scene. It should be noted that a coordinate system also needs to be set along the insertion direction of the hydrogen refueling gun to facilitate the hydrogen refueling insertion operation.

[0068] Specifically, the steps for finding the template in the 3D scene and calculating the pose of the hydrogen refueling port are as follows:

[0069] Uniform sampling is performed on template instances in the 3D scene, with 50-150 scene points sampled for each template instance. A set of keypoints is then selected from the sampled scene points, and the number of keypoints selected is controlled by the parameter KeyPoint. For example, setting the parameter value to 0.5 will use 50% of the sampled scene points as keypoints. More keypoints will improve the accuracy of the matching results, but will reduce the matching speed. Generally, a value below 0.5 is selected.

[0070] For each selected key point, under the assumption that the key point is located on the surface of the object, the optimal pose of the template is calculated. The pose with the highest score among all key points is selected as the approximate pose, which is the hydrogen filling port pose.

[0071] It should be noted that before performing approximate pose matching, the point cloud information is first denoised to remove useless information from the point cloud information.

[0072] The hydrogen refueling port pose recognition system described in this embodiment uses a combination of 2D deep learning and 3D template matching. It uses 2D deep learning for coarse positioning and 3D point cloud matching for precise positioning, thereby improving the positioning accuracy of the hydrogen refueling port.

[0073] Example 2: This example provides an automated hydrogen refueling system based on cryogenic refueling, including a cryogenic explosion-proof hydrogen refueling robot and the hydrogen refueling port pose recognition system described in Example 1. The image acquisition device of the hydrogen refueling port pose recognition system is located outside the hydrogen refueling robot, while the coarse positioning unit and the precise positioning unit of the hydrogen refueling port pose recognition system are located inside the hydrogen refueling robot. The image acquisition device uses a cryogenic explosion-proof industrial 3D camera. The precise positioning unit of the hydrogen refueling port is connected to a control module located inside the hydrogen refueling robot. The control module controls the movement of the robotic arm of the hydrogen refueling robot based on the hydrogen refueling port pose calculated by the precise positioning unit. The robotic arm drives the hydrogen refueling gun located on the robotic arm to automatically dock with the hydrogen refueling port position for automated hydrogen refueling.

[0074] Specifically, the hydrogen refueling robot is constructed from low-temperature resistant and explosion-proof components, and is a six-axis robot, meaning it has six degrees of freedom. In this embodiment, the robot's robotic arm has a working radius of ≥1500mm, a load capacity of ≥15kg, a safety factor of ≥1.5, and the electrical control cabinet with the control module must comply with EN ISO13849 PLd safety standards. The robot itself must pass both Chinese and EU ATEX explosion-proof certifications and can be used in Class 1 hazardous gas environments.

[0075] The hydrogen refueling port is approximately cylindrical. Since the hydrogen refueling port pose calculated by the precise positioning unit may not be within the flexible pose space of the robotic arm, the automatic hydrogen refueling system described in this embodiment also includes a pose correction module located inside the hydrogen refueling robot. This module, when the hydrogen refueling port pose calculated by the precise positioning unit is not within the flexible pose space of the robotic arm, considers the isotropic shape of the hydrogen refueling port, generates a hydrogen refueling port pose matrix based on the pose, and generates multiple new poses rotating around the z-axis while ensuring the z-axis orientation remains unchanged. Inverse kinematics calculations are then performed on these new poses to calculate the cost function of each set of joint solutions and determine the optimal pose. Here, the z-axis represents the position and pose of the end of the sixth axis of the robotic arm in the robotic arm's base coordinate system. (See [link to documentation] for the position and pose of the end of the sixth axis of the robotic arm in the robotic arm's base coordinate system.) Figure 6 The pose correction module, which adjusts the hydrogen refueling port's position, allows the robotic arm to perform more flexible and complete hydrogen refueling movements. Specifically,

[0076]

[0077] The cost function is expressed as: In the formula, θ i For the joint angle, θ mid The intermediate value for the limit of each joint, θ max The maximum value of the limit for each joint, θ min This is the minimum limit for each joint. It's worth noting that this cost function is for all six axes, calculating the optimal angle value for each axis joint separately. The optimal solution for each joint can be calculated using the loss function.

[0078] The automated hydrogen refueling system described in this embodiment also includes a motion correction module installed inside the hydrogen refueling robot. This module compensates for any misalignment between the calculated and actual hydrogen refueling port positions, using the contact force feedback between the robotic arm's end effector and the hydrogen refueling port to correct the hole-axis alignment during misalignment, thus ensuring successful hydrogen refueling. The method for correcting the hydrogen refueling motion using the motion correction module is described in [link to documentation]. Figure 7 For the correction process, please refer to Figures 8a-8d .

[0079] Specifically, the hydrogen refueling port precise positioning unit sets a minimum score threshold. If the score exceeds this minimum threshold, the calculated hydrogen refueling port pose is transmitted to the control module, which then controls the robotic arm's movement based on the hydrogen refueling port pose. It should be noted that when converting the robotic arm's Cartesian pose to joint angles, a certain joint angle may exceed the limit. The optimal joint solution can be calculated using a loss function, but if the threshold is too high, the calculation will fail. Therefore, by setting a minimum threshold, the joint motion posture of the robotic arm can be obtained while satisfying the joint limits.

[0080] The above embodiments are used to explain the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. An automated hydrogen refueling system based on cryogenic refueling, characterized in that, The system includes a low-temperature explosion-proof hydrogen refueling robot and a hydrogen refueling port pose recognition system. The image acquisition device of the hydrogen refueling port pose recognition system is located outside the robot, while the coarse positioning unit and precise positioning unit of the hydrogen refueling port are located inside the robot. The precise positioning unit is connected to a control module inside the robot. The control module controls the robot's robotic arm movement based on the hydrogen refueling port pose calculated by the precise positioning unit. The robotic arm drives the hydrogen refueling nozzle mounted on it to automatically dock with the hydrogen refueling port for automatic hydrogen refueling. It also includes a pose correction module located inside the hydrogen refueling robot. When the hydrogen refueling port pose calculated by the precise positioning unit is not within the flexible pose space of the robotic arm, it considers the isotropic shape of the hydrogen refueling port, generates a hydrogen refueling port pose matrix based on the port pose, and generates multiple new poses around the z-axis while ensuring the z-axis orientation remains unchanged. Inverse kinematics calculations are performed on these new poses to calculate the cost function of each set of joint solutions and determine the optimal pose. Here, the z-axis represents the position and pose of the end face of the sixth axis of the robotic arm in the base coordinate system. The cost function is expressed as: In the formula, θ i For joint angle, θ mid The intermediate value for the limit of each joint, θ max The maximum limit for each joint. θ min The minimum limit for each joint; it also includes a motion correction module inside the hydrogen refueling robot, which is used to compensate for the hole-axis alignment movement when the error between the calculated hydrogen refueling port pose and the actual hydrogen refueling port pose affects coaxial alignment, by using the contact force feedback between the end of the robotic arm and the hydrogen refueling port, so as to ensure the completion of the hydrogen refueling movement. The hydrogen refueling port pose recognition system includes: An image acquisition device is used to acquire images of the hydrogen refueling port; the image acquisition device employs a low-temperature resistant, explosion-proof industrial 3D camera. The hydrogen refueling port coarse localization unit extracts features from the hydrogen refueling port image, fuses local and global features to obtain features with CNN style and Transformer style, and classifies the features to obtain the two-dimensional location information of the hydrogen refueling port. The hydrogen refueling port precise positioning unit obtains the three-dimensional point cloud information of the hydrogen refueling port based on the two-dimensional position information of the hydrogen refueling port. Combining the three-dimensional point cloud information and the three-dimensional model information of the hydrogen refueling port, it uses create_surface_model to create a template, finds the template in the 3D scene, and calculates the pose of the hydrogen refueling port. The specific steps for creating the template are as follows: Multiple cameras are evenly distributed in a 3D scene to simulate shooting the 3D model of the hydrogen refueling port from different angles, obtaining different angle information of the 3D model of the hydrogen refueling port, and storing the views from all perspectives in a 3D model; denoising and filtering are performed on the 3D point cloud information of the hydrogen refueling port obtained based on the 2D position information of the hydrogen refueling port to obtain point cloud information containing only the hydrogen refueling port; the point cloud information containing only the hydrogen refueling port is combined with the 3D model information obtained from the 3D model of the hydrogen refueling port to obtain the final model; the second-order mean or central moment mean of the final model is calculated using create_surface_model, and an affine transformation is performed to generate the template coordinate system to obtain the template.

2. The automated hydrogen refueling system based on cryogenic refueling as described in claim 1, characterized in that, The hydrogen refueling port precise positioning unit sets a minimum score threshold. If the score exceeds the minimum score threshold, the calculated hydrogen refueling port pose is transmitted to the control module, and the control module controls the robotic arm to move according to the hydrogen refueling port pose.

3. The automated hydrogen refueling system based on cryogenic refueling as described in claim 1, characterized in that, The coarse positioning unit for the hydrogenation port includes: The stem module is used to extract initial local features from the hydrogen refueling port image; The two branches are connected to the stem module. The two branches include a CNN branch for extracting local features from the initial local features and a Transformer branch for extracting all features from the initial local features. FCU, the bridged dual-branch module, is used to interactively fuse local features extracted by the CNN branch with global features extracted by the Transformer branch. The first classifier is used to receive all features from the CNN branch, classify them, and output the results. The second classifier is used to receive all features from the Transformer branch, classify them, and output the results. The output module is used to add the output of the first classifier to the output of the second classifier to obtain the coarse localization result.

4. The automated hydrogen refueling system based on cryogenic refueling as described in claim 3, characterized in that, The stem module consists of a 7×7 convolutional layer with a stride of 2 and a 3×3 max pooling layer with a stride of 2. The CNN branch adopts a feature pyramid structure, divided into four stages. Each stage consists of multiple convolutional blocks, and each convolutional block contains n bottlenecks. Following the definition in ResNet, each bottleneck contains a 1×1 convolutional layer, a 3×3 spatial convolutional layer, a 1×1 convolutional layer, and residual connections between the input and output. The Transformer branch contains N repeated transformer blocks. Each transformer block consists of a multi-head self-attention layer and an MLP layer. Layer norm is applied before each layer, and there are residual connections in both the self-attention layer and the MLP layer.

5. The automated hydrogen refueling system based on cryogenic refueling as described in claim 3, characterized in that, When training the CNN and Transformer branches, two cross-entropy functions with the same importance weights in the loss function are used to supervise the first and second classifiers respectively, so as to obtain features that have both CNN and Transformer styles.

6. The automated hydrogen refueling system based on cryogenic refueling as described in claim 1, characterized in that, The specific steps for finding the template and calculating the hydrogen refueling port pose in the 3D scene are as follows: Uniform sampling is performed on template instances in the 3D scene, with 50-150 scene points sampled for each template instance. A set of key points is selected from the sampled scene points, and the number of key points selected is controlled by the parameter KeyPoint. For each selected key point, under the assumption that the key point is located on the surface of the object, the optimal pose of the template is calculated. The pose with the highest score among all key points is selected as the approximate pose, which is the hydrogen filling port pose.

Images