LNG loading valve identification method and system oriented to complex conditions

Abstract

The invention discloses an LNG loading valve identification method and system oriented to complex conditions. The invention provides a light weight multi-scale detection network LitePAN (Personal Area Network)-YOLO (Young Line Occurrence). The network is constructed based on a YOLOv8 framework, an improved Vanilla module is introduced to enhance the nonlinear expression ability of features, a C2f-ParNet module is designed to enhance cross-layer gradient flow and channel interaction, efficient multi-scale feature fusion is achieved, and a GSConv-based LitePAN structure is adopted in a neck structure to reduce the calculated amount and maintain the feature aggregation efficiency. Meanwhile, the detection head adopts an anchor-free decoupling design, and a WIoU v3 loss function is combined to improve bounding box regression and high IoU positioning precision. Compared with YOLOv8-n, the LitePAN-YOLO has the advantages that the parameter quantity of the LitePAN-YOLO is reduced by 7.7%, the calculated quantity is reduced by 5.9%, and mAP50-95 and mAP95 are respectively improved by 0.59% and 1.30%. The method shows higher robustness and detection precision under the scenes of dense small targets, large attitude change and complex illumination, and provides reliable visual perception support for loading and unloading operation of the LNG collaborative robot.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of LNG loading operation target detection, and particularly relates to a LNG loading valve recognition method and system under complex conditions. BACKGROUND

[0002] With the in-depth development of the concept of intelligent manufacturing, the LNG loading pry system is gradually realizing automation and intelligentization. At present, a variety of automatic loading products represented by pneumatic auxiliary arms, hydraulic semi-automatic arms and electrically-driven full-automatic arms have been formed in the market, which are used to realize automatic docking of the loading pry and the tank truck interface. However, in the loading process, the valve operation at the end of the LNG tank truck still needs manual intervention, which limits the full automation level of the system. As a key control element in the process of LNG storage, transportation and loading, the recognition and positioning accuracy of the valve directly affects the safety and operation efficiency of the system.

[0003] Due to the large number of tank truck valves, complex structure and significant shape difference, the traditional recognition method based on rules or geometric features performs poorly in the face of complex scenes such as size variation, view transformation and light interference. In addition, the traditional feature extraction method relies on a pre-set trajectory to perform a fixed action, which is difficult to meet the demand for accurate recognition and positioning of multiple types of valves in the filling operation. In recent years, the development of deep learning technology has provided a new solution for industrial visual detection. Convolutional neural network (CNN) can automatically extract multi-layer semantic features, and has stronger generalization ability and robustness than traditional methods.

[0004] Among the many deep learning-based target detection algorithms, the YOLO series model is widely used in industrial visual detection tasks due to its end-to-end structure, high detection speed and excellent detection accuracy. Versions such as YOLOv5, YOLOv7 and YOLOv8 have effectively improved the target detection performance by introducing residual structures, attention mechanisms and feature fusion strategies. However, there are still challenges in directly applying the general YOLO model to the valve recognition scene: high similarity between valve categories, variable poses, complex lighting, large proportion of small targets, and strong background interference, which result in insufficient precision and robustness of the model in feature extraction and multi-scale fusion. SUMMARY

[0005] The purpose of the present application is to solve the problems existing in the prior art, and a lightweight, multi-scale feature fusion valve detection network LitePAN-YOLO is proposed. The model is based on YOLOv8, and improved Vanilla modules are introduced into the backbone network to enhance the expression ability of feature extraction. In the detection head, an efficient feature fusion structure is used to realize the full interaction of multi-scale information.

[0006] The technical solution for achieving the object of the application is: on the one hand, a LNG loading valve identification method under complex conditions is provided, and the method comprises the following steps:

[0007] Step 1, constructing a valve data set for model training;

[0008] Step 2, constructing a valve detection network LitePAN-YOLO with light weight and multi-scale feature fusion;

[0009] Step 3, training the valve detection network LitePAN-YOLO using the valve data set;

[0010] Step 4, identifying the LNG loading valve under complex conditions based on the trained valve detection network LitePAN-YOLO model.

[0011] Further, in step 1, the valve data set for model training is constructed, specifically including:

[0012] In a complex environment, a camera is used to capture and collect common valves to form an initial valve data set;

[0013] The images in the initial data set are subjected to enhancement processing to form an enhanced valve data set; the enhancement processing includes flipping, mirroring and scaling operations;

[0014] The images in the enhanced valve data set are preprocessed, screened and labeled to obtain a final valve data set;

[0015] The final valve data set is divided into a training set, a validation set and a test set.

[0016] Further, in step 2, the valve detection network LitePAN-YOLO with light weight and multi-scale feature fusion is constructed, specifically including:

[0017] (1) The backbone part in the overall architecture is constructed based on the YOLOv8 framework, and an improved Vanilla module is introduced for fine feature extraction of images;

[0018] (2) Efficient multi-scale feature fusion is achieved by designing a C2f-ParNet module to construct a light weight feature fusion network structure, including: the LitePAN structure is used to improve the traditional PANet, the high-level features of the image are spliced with the adjacent low-level features through up-sampling operation, and the GSConv module is used to integrate the fused features in the down-sampling stage;

[0019] (3) In the Head part, the anchor-free detection mechanism of YOLOv8 and the decoupling head are used in combination with the WIoU v3 loss function to output three prediction feature maps at different scales, so as to separate the classification and regression tasks.

[0020] Furthermore, the improved Vanilla module specifically includes:

[0021] Based on the original Vanilla module, the original 4×4 convolution is replaced with two cascaded 3×3 convolutional layers;

[0022] A non-linear activation function is introduced between two consecutive 3×3 convolutional layers;

[0023] The parameters of a convolutional layer are expressed as follows:

[0024]

[0025] In the formula, This represents the total number of parameters in the convolutional layer; This is the kernel size, i.e., the side length of the kernel. and These represent the number of channels in the input and output feature maps, respectively.

[0026] Furthermore, in step 2, the C2f-ParNet module adds parallel bottleneck and channel recalibration strategies to the C2f branch.

[0027] Furthermore, the C2f-ParNet module specifically comprises:

[0028] The input is first passed through a 1×1 standard convolution to adjust the number of channels and achieve initial feature compression;

[0029] Subsequently, the features are divided into two parts along the channel dimension through a Split operation. One part of the features is directly retained as the main branch, while the other part is input to the ParNet_Bottleneck module for feature extraction. The ParNet_Bottleneck module includes two layers of 3×3 convolutions and ParNetAttention, which are used to enhance local representations and adaptively model channel dependencies while maintaining spatial resolution. ParNetAttention adopts a three-branch parallel fusion strategy: a 1×1 convolution branch for channel consistency modeling; a 3×3 convolution branch for local detail extraction; and an SSE branch that generates channel weights through global average pooling and gating mechanisms and multiplies them element-wise with the input. The sum of the three branches is then activated by SiLU to obtain attention-enhanced features.

[0030] After the bottleneck outputs and the retained backbone branches at each stage are concatenated along the channel dimension, feature integration is completed through 1×1 convolution to obtain the network output;

[0031] Based on the trade-off between LitePAN fusion path computation and channel interaction, the two 3×3 standard convolutions at the Neck position are replaced with GSConv modules.

[0032] Furthermore, the GSConv module comprises two parallel branches: standard convolution (SC) focuses on modeling cross-channel correlations, while depthwise separable convolution (DSC) captures spatial context with lower computational cost; each branch generates... After each channel is concatenated, the channels are stitched together, and the channel shuffle module promotes the uniform diffusion of channel information from the SC in the overall features.

[0033] Furthermore, based on the task and network depth, the number of bottlenecks and whether to enable identity residual connections can be set as needed to achieve a balance between computational overhead and discriminability.

[0034] Furthermore, the WIoU v3 loss function is expressed as:

[0035]

[0036] in,

[0037]

[0038]

[0039]

[0040]

[0041]

[0042] In the formula, These represent the WIoU v3 loss function and the WIoU v1 loss function, respectively. Indicates the non-monotonic focusing coefficient. Represents the outlier parameter. and All of these represent adjustable hyperparameters, whose values ​​are optimized and adapted according to the characteristics of different models. Represents the center coordinates of the true bounding box. and These are the width and height of the minimum bounding box, respectively. Indicates the center coordinates of the prediction box, symbol express and It has been separated from the computation graph; Indicates the IoU loss value. This represents the monotonic focusing coefficient.

[0043] On the other hand, a valve identification system for LNG loading under complex conditions is provided, the system comprising:

[0044] The first module constructs a valve dataset for model training;

[0045] The second module constructs a lightweight, multi-scale feature fusion valve detection network, LitePAN-YOLO.

[0046] The third module is used to train the valve detection network LitePAN-YOLO using the valve dataset;

[0047] The fourth module is used to identify LNG loading valves under complex conditions based on the trained valve detection network LitePAN-YOLO model.

[0048] Compared with the prior art, the significant advantages of this invention are:

[0049] (1) To meet the fully automated requirements of liquefied natural gas (LNG) loading skid systems, this invention addresses the problem of insufficient identification and positioning accuracy of tank truck valves under conditions of high category similarity, large attitude changes, and complex lighting. A lightweight multi-scale detection network, LitePAN-YOLO, is proposed. The proposed LitePAN-YOLO model for valve identification introduces an improved Vanilla module in the backbone network to enhance the nonlinear representation of features. In the feature fusion network, a lightweight LitePAN structure is constructed by combining GSConv and ParNet-C2f to achieve bidirectional, efficient multi-scale feature aggregation and strengthen cross-layer gradient flow. The detection head maintains an anchor-free decoupling design to obtain more stable regression performance.

[0050] (2) Compared with the baseline model YOLOv8, the present invention improves mAP50-95 by 0.59%, mAP95 by 1.3%, and reduces the number of parameters by 7.7%. The method of the present invention exhibits higher robustness and stability when dealing with densely distributed small targets, large attitude changes, and complex lighting conditions. It can achieve accurate online identification and positioning of valves in LNG loading scenarios, providing reliable visual perception support for collaborative robot operations. It has strong engineering application value and practical deployment potential.

[0051] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0052] Figure 1This is a schematic diagram of the overall architecture and feature flow of a LitePAN-YOLO network in one embodiment.

[0053] Figure 2 This is a diagram of an improved Vanilla module structure in one embodiment.

[0054] Figure 3 This is a structural diagram of an improved C2f-ParNet module in one embodiment.

[0055] Figure 4 This is a diagram of the ParNet_Bottlenck structure in one embodiment.

[0056] Figure 5 This is a diagram of the ParNetAttention structure in one embodiment.

[0057] Figure 6 This is a structural diagram of GSConv in one embodiment.

[0058] Figure 7 This is a schematic diagram of image samples in a dataset constructed in one embodiment, wherein... Figure 7 In the diagram, (a) to (d) represent a check valve, a butterfly valve, a gate valve, and an expansion joint, respectively.

[0059] Figure 8 This is a schematic diagram of the ablation experiment results in one embodiment. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0061] It should be noted that if the embodiments of the present invention involve descriptions such as "first" and "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0062] In one embodiment, a method for identifying LNG loading valves under complex conditions is provided, the method comprising the following steps:

[0063] Step 1: Construct a valve dataset for model training;

[0064] Step 2: Construct a lightweight, multi-scale feature fusion valve detection network, LitePAN-YOLO;

[0065] Step 3: Train the valve detection network LitePAN-YOLO using the valve dataset;

[0066] Step 4: Based on the trained valve detection network LitePAN-YOLO model, identify LNG loading valves under complex conditions.

[0067] Furthermore, in one embodiment, step 1, constructing a valve dataset for model training, specifically includes:

[0068] In complex environments (not limited to changes in light and stacked parts), a camera is used to photograph and collect data on common valves, forming an initial valve dataset;

[0069] The images in the initial dataset are enhanced to form an enhanced valve dataset; the enhancement process includes operations such as flipping, mirroring, and scaling.

[0070] The images in the enhanced valve dataset are preprocessed (to handle complex conditions and adapt to the input requirements of subsequent models), filtered, and labeled to obtain the final valve dataset.

[0071] The final valve dataset is divided into training, validation, and test sets.

[0072] Here, the preprocessing includes, but is not limited to: removing environmental noise during the acquisition process by median filtering or Gaussian filtering, adjusting image brightness using histogram equalization algorithm to overcome uneven illumination, and normalizing image size to uniformly adjust it to the fixed pixel size required by the model.

[0073] Furthermore, in one embodiment, step 2, constructing a lightweight, multi-scale feature fusion valve detection network LitePAN-YOLO, specifically includes:

[0074] (1) Based on the YOLOv8 framework, the backbone part of the overall architecture is constructed, and the improved Vanilla module is introduced to extract fine features of the image;

[0075] (2) By designing the C2f-ParNet module, efficient multi-scale feature fusion is achieved, and a lightweight feature fusion network structure is constructed, including: using the LitePAN structure to lighten and improve the traditional PANet, the high-level features of the image are spliced ​​with the adjacent low-level features through upsampling operation, and the GSConv module is used to integrate the fused features in the downsampling stage.

[0076] (3) In the Head part, the anchor-free detection mechanism of YOLOv8 and the decoupling head are used in combination with the WIoU v3 loss function to output three prediction feature maps at different scales, so as to separate the classification and regression tasks.

[0077] In summary, this network is built on the YOLOv8 framework, introducing an improved Vanilla module to enhance the nonlinear representation of features, designing a C2f-ParNet module to strengthen cross-layer gradient flow and channel interaction, achieving efficient multi-scale feature fusion, and employing a GSConv-based LitePAN structure in the neck structure to reduce computational cost while maintaining feature aggregation efficiency. Simultaneously, the detection head adopts an anchor-free decoupled design and incorporates the WIoU v3 loss function to improve bounding box regression and high IoU localization accuracy.

[0078] Preferably, in some embodiments, combined with Figure 2 The improved Vanilla module is specifically as follows:

[0079] (1) Based on the original Vanilla module, the original 4×4 convolution is replaced with two cascaded 3×3 convolutional layers;

[0080] Here, the original 4×4 convolution is replaced with two cascaded 3×3 convolutional layers to improve the network's non-linear expressive power while maintaining an equivalent receptive field. Specifically:

[0081] According to the formula for calculating the receptive field:

[0082]

[0083] in The effective receptive field is the spatial range that the network can perceive on the input feature map. and These are the receptive field sizes for the first and second convolutional layers, respectively.

[0084] For two 3×3 convolutional layers, we can obtain:

[0085]

[0086] It can be seen that two 3×3 convolutions significantly enhance the ability to capture feature details while maintaining the equivalent of a single 5×5 receptive field.

[0087] (2) Introduce a non-linear activation function between two consecutive 3×3 convolutional layers to enable the network to have stronger feature representation capabilities under the same computational conditions;

[0088] (3) The parameters of the convolutional layer are expressed as follows:

[0089]

[0090] In the formula, This represents the total number of parameters in the convolutional layer. This refers to the kernel size, which is the side length of the convolution kernel. and These represent the number of channels in the input and output feature maps, respectively.

[0091] Here, after replacing the 4×4 convolution with two 3×3 convolutions, the number of parameters changes as follows:

[0092]

[0093] The improved Vanilla module achieves higher feature density and information extraction capability while maintaining similar computational complexity. Although the number of parameters in individual modules increases slightly, the overall number of parameters in the LitePAN-YOLO network is reduced by approximately 7.7% compared to the baseline model due to adjustments and fusion of channel numbers in subsequent structures, resulting in significant optimization of computational efficiency.

[0094] Furthermore, in one embodiment, in step 2, the C2f-ParNet module incorporates a parallel bottleneck and channel recalibration strategy into the C2f branch. This combination significantly improves the model's detection performance and robustness against valve targets of different sizes while achieving efficient feature flow.

[0095] Preferably, in some embodiments, the C2f-ParNet module has the following structure: Figure 3 As shown, this addresses the issue of insufficient feature extraction capabilities of the C2f module in complex detection scenarios. Specifically:

[0096] The input is first passed through a 1×1 standard convolution (Conv1×1) to adjust the number of channels and achieve initial feature compression;

[0097] Subsequently, the features are divided into two parts along the channel dimension using a Split operation. One part of the features is directly retained as the main branch, while the other part is input into the ParNet_Bottleneck module for feature extraction. The ParNet_Bottleneck module consists of two 3×3 convolutional layers and ParNetAttention, as shown below. Figure 4 As shown, ParNetAttention is used to enhance local representations and adaptively model channel dependencies while maintaining spatial resolution; among them, ParNetAttention is as follows: Figure 5 As shown, a three-branch parallel fusion strategy is adopted: a 1×1 convolutional branch for channel consistency modeling; a 3×3 convolutional branch for local detail extraction; and an SSE branch that generates channel weights through global average pooling and gating mechanism and multiplies them element-wise with the input. The three branches are added together and then activated by SiLU to obtain attention-enhanced features.

[0098] After the bottleneck outputs and the retained backbone branches at each stage are concatenated along the channel dimension, feature integration is completed through 1×1 convolution to obtain the network output;

[0099] Based on a trade-off between LitePAN fusion path computation and channel interaction, the two 3×3 standard convolutions at the Neck position are replaced with GSConv modules, such as... Figure 6 As shown, the GSConv module contains two parallel branches: standard convolution (SC) focuses on modeling cross-channel correlations, and depthwise separable convolution (DSC) captures spatial context with lower computational cost; each branch generates After each channel is concatenated, the channels are stitched together, and the channel shuffle module promotes the uniform diffusion of channel information from the SC in the overall features.

[0100] Here, the GSConv module enables efficient interaction and fusion between feature channels, and LitePAN can complete bidirectional feature transfer from top to bottom and from bottom to top with lower computational cost.

[0101] Here, compared to using DSC alone or in large quantities Compared to pointwise convolution, this design better preserves cross-channel interactions while maintaining lower FLOPs, thus achieving a better trade-off between accuracy and efficiency.

[0102] Preferably, in some embodiments, the number of bottlenecks and whether to enable identity residual connections are set as needed based on the task and network depth to achieve a balance between computational overhead and discriminability.

[0103] Preferably, in some embodiments, the WIoU v3 loss function is expressed as:

[0104]

[0105] in,

[0106]

[0107]

[0108]

[0109]

[0110]

[0111] In the formula, These represent the WIoU v3 loss function and the WIoU v1 loss function, respectively. Indicates the non-monotonic focusing coefficient. Represents the outlier parameter. and All of these represent adjustable hyperparameters, whose values ​​are optimized and adapted according to the characteristics of different models. Represents the center coordinates of the true bounding box. and These are the width and height of the minimum bounding box, respectively. Indicates the center coordinates of the prediction box, symbol express and It has been separated from the computation graph; Indicates the IoU loss value. This represents the monotonic focusing coefficient.

[0112] Here, an outlier parameter is introduced to measure the quality of the anchor frame. and based on A non-monotonic focusing coefficient was constructed. Smaller This indicates a high-quality anchor frame, in which case a smaller allocation is used. This allows Bounding Box Regression (BBR) to focus more on medium-quality anchor boxes; while larger ones... This indicates a low-quality anchor frame, in which case a smaller allocation is required. This is to prevent low-quality samples from generating excessively large and harmful gradients. This allocation mechanism allows the model to focus more on anchor boxes of average quality, thereby improving the model's localization ability.

[0113] In one embodiment, a valve identification system for LNG loading under complex conditions is provided, the system comprising:

[0114] The first module constructs a valve dataset for model training;

[0115] The second module constructs a lightweight, multi-scale feature fusion valve detection network, LitePAN-YOLO.

[0116] The third module is used to train the valve detection network LitePAN-YOLO using the valve dataset;

[0117] The fourth module is used to identify LNG loading valves under complex conditions based on the trained valve detection network LitePAN-YOLO model.

[0118] Specific limitations regarding the LNG loading valve identification system under complex conditions can be found in the limitations of the LNG loading valve identification method under complex conditions described above, and will not be repeated here. Each module in the aforementioned LNG loading valve identification system under complex conditions can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0119] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements:

[0120] Step 1: Construct a valve dataset for model training;

[0121] Step 2: Construct a lightweight, multi-scale feature fusion valve detection network, LitePAN-YOLO;

[0122] Step 3: Train the valve detection network LitePAN-YOLO using the valve dataset;

[0123] Step 4: Based on the trained valve detection network LitePAN-YOLO model, identify LNG loading valves under complex conditions.

[0124] For specific limitations on each step, please refer to the limitations on the LNG loading valve identification method under complex conditions mentioned above, which will not be repeated here.

[0125] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being implemented when executed by a processor:

[0126] Step 1: Construct a valve dataset for model training;

[0127] Step 2: Construct a lightweight, multi-scale feature fusion valve detection network, LitePAN-YOLO;

[0128] Step 3: Train the valve detection network LitePAN-YOLO using the valve dataset;

[0129] Step 4: Based on the trained valve detection network LitePAN-YOLO model, identify LNG loading valves under complex conditions.

[0130] For specific limitations on each step, please refer to the limitations on the LNG loading valve identification method under complex conditions mentioned above, which will not be repeated here.

[0131] As a specific example, the invention will be further verified and illustrated in one embodiment.

[0132] The four most common types of valves were selected as the target for identification: check valve, butterfly valve, gate valve, and expansion joint.

[0133] The simulation of complex warehouse environments also includes image samples of varying lighting and stacked parts.

[0134] Use a USB camera to photograph and capture images of selected marine valve parts. Figure 7 This shows some image samples from the dataset.

[0135] During data processing, the acquired images underwent various enhancement processes, including flipping, mirroring, and scaling.

[0136] Then, image processing, filtering, and labeling steps are performed to obtain the valve dataset used for model training.

[0137] The collection contains 3,600 images, including 650 images each of individual check valves, butterfly valves, gate valves, and expansion joints, and 250 images each of the four types of valves arranged normally, in sunlight, in darkness, and stacked together.

[0138] Next, all the collected marine valve images were divided into a training set (2520 images), a validation set (720 images), and a test set (360 images) in a ratio of 7:2:1.

[0139] Based on a self-made dataset, ablation experiments were conducted on the LitePAN-YOLO model to evaluate its effectiveness in valve target detection tasks.

[0140] Comparative experiments were conducted with other similar algorithms to ensure the repeatability of the experiments and the accuracy of the results.

[0141] The version and configuration information of the environment required for the experiment, as well as the relevant parameter settings used throughout the training phase, are given in Tables 1 and 2.

[0142] Table 1. Environmental configuration required for the experiment

[0143]

[0144] Table 2 Training Parameter Settings

[0145]

[0146] Based on the implementation results, see Table 3 below and Figure 8 As shown, the proposed LitePAN-YOLO model demonstrates continuous performance improvement across all ablation stages of the self-built LNG valve dataset, with YOLOv8-n as the baseline model. Replacing the original single-layer 4×4 convolution with two concatenated 3×3 convolutions (Vanilla) reduces the number of parameters and computational cost by 17.23% and 16.18% respectively (from 2.685M / 6.8G to 2.222M / 5.7G), while simultaneously increasing mAP95 from 47.11 to 47.99 (+0.88).

[0147] Table 3 Ablation Experiment Results

[0148]

[0149] As shown in Table 4, the proposed LitePAN-YOLO model outperforms the baseline model YOLOv8-n in overall performance. While maintaining a lightweight structure, the number of parameters in LitePAN-YOLO decreased from 2.685M to 2.478M (a reduction of approximately 7.7%), and the computational cost decreased from 6.8 GFLOPs to 6.4 GFLOPs (a reduction of approximately 5.9%), while mAP50-95 increased from 86.78% to 87.37%, and mAP95 increased from 47.11% to 48.41%. This indicates that the model effectively improves detection accuracy and localization capability while significantly reducing its size.

[0150] Table 4. Performance comparison results of the LitePAN-YOLO algorithm proposed in this invention with other YOLO series algorithms.

[0151]

[0152] The results demonstrate that enhancing the nonlinear expressive power of features through lightweight convolution stacking can significantly improve feature representation performance without increasing computational burden. Introducing the ParNet-C2f module improves mAP50–95 to 87.31 (+0.53) and mAP95 to 48.21, validating the effectiveness of parallel multi-scale feature fusion in strengthening cross-layer gradient flow. Introducing GSConv into the neck structure reduces the number of parameters and computational cost by 7.70% and 5.88% respectively compared to the baseline model, while maintaining mAP50 at 99.49, achieving mAP50–95 of 87.01 and mAP95 of 48.12, indicating a better trade-off between accuracy and efficiency. Finally, after incorporating the WIoU v3 loss function, mAP95 further improved to 48.41 (+1.30 from the baseline), and mAP50–95 increased to 87.37, indicating that this loss function can effectively focus on medium-quality samples, thereby improving localization accuracy under high IoU thresholds. Overall, the Vanilla module delivers significant feature representation gains while maintaining lightweight design, ParNet-C2f enhances feature interaction capabilities, GSConv achieves efficient lightweight feature aggregation, and WIoU v3 improves the robustness of bounding box regression. The synergistic effect of these modules achieves an optimal balance between detection accuracy and computational efficiency. The LitePAN-YOLO algorithm proposed in this invention achieves an excellent balance between accuracy, efficiency, and model size, validating the effectiveness of lightweight multi-scale detection networks in complex industrial scenarios.

[0153] In summary, the model proposed in this invention improves the mAP50–95 and mAP95 of the valve dataset by 0.59% and 1.3%, respectively, and reduces the number of parameters by about 7.7%. While ensuring real-time performance, it significantly improves the accuracy and stability of valve identification, providing reliable technical support for the intelligent and safe operation of LNG automated loading systems.

[0154] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention without departing from its spirit and scope should be included within the protection scope of the present invention.

Claims

1. A method for identifying LNG loading valves under complex conditions, characterized in that, The method includes the following steps: Step 1: Construct a valve dataset for model training; Step 2: Construct a lightweight valve detection network, LitePAN-YOLO, that integrates multi-scale features; Step 3: Train the valve detection network LitePAN-YOLO using the valve dataset; Step 4: Based on the trained valve detection network LitePAN-YOLO model, identify LNG loading valves under complex conditions.

2. The LNG loading valve identification method under complex conditions according to claim 1, characterized in that, Step 1 involves constructing a valve dataset for model training, specifically including: In complex environments, cameras are used to capture images of common valves, forming an initial valve dataset. The images in the initial dataset are enhanced to form an enhanced valve dataset; the enhancement process includes flipping, mirroring, and scaling operations. The images in the enhanced valve dataset are preprocessed, filtered, and labeled to obtain the final valve dataset. The final valve dataset is divided into training, validation, and test sets.

3. The LNG loading valve identification method under complex conditions according to claim 1, characterized in that, Step 2 involves constructing a lightweight, multi-scale feature fusion valve detection network, LitePAN-YOLO, which specifically includes: (1) Based on the YOLOv8 framework, the backbone part of the overall architecture is constructed, and the improved Vanilla module is introduced to extract fine features of the image; (2) Efficient multi-scale feature fusion is achieved by designing the C2f-ParNet module to construct a lightweight feature fusion network structure, including: using the LitePAN structure to lighten and improve the traditional PANet, the high-level features of the image are spliced ​​with the adjacent low-level features through upsampling operation, and the GSConv module is used to integrate the fused features in the downsampling stage. (3) In the Head part, the anchor-free detection mechanism of YOLOv8 and the decoupling head are used in combination with the WIoU v3 loss function to output three prediction feature maps at different scales, so as to separate the classification and regression tasks.

4. The LNG loading valve identification method under complex conditions according to claim 3, characterized in that, The improved Vanilla module is specifically as follows: Based on the original Vanilla module, the original 4×4 convolution is replaced with two cascaded 3×3 convolutional layers; A non-linear activation function is introduced between two consecutive 3×3 convolutional layers; The parameters of a convolutional layer are expressed as follows: ； In the formula, This represents the total number of parameters in the convolutional layer; This is the kernel size, i.e., the side length of the kernel. and These represent the number of channels in the input and output feature maps, respectively.

5. The LNG loading valve identification method under complex conditions according to claim 3, characterized in that, In step 2, the C2f-ParNet module adds parallel bottleneck and channel recalibration strategies to the C2f branch.

6. The LNG loading valve identification method under complex conditions according to claim 5, characterized in that, The C2f-ParNet module is specifically as follows: The input is first passed through a 1×1 standard convolution to adjust the number of channels and achieve initial feature compression; Subsequently, the features are divided into two parts along the channel dimension through a Split operation. One part of the features is directly retained as the main branch, while the other part is input to the ParNet_Bottleneck module for feature extraction. The ParNet_Bottleneck module includes two layers of 3×3 convolutions and ParNetAttention, which are used to enhance local representations and adaptively model channel dependencies while maintaining spatial resolution. ParNetAttention adopts a three-branch parallel fusion strategy: a 1×1 convolution branch for channel consistency modeling; a 3×3 convolution branch for local detail extraction; and an SSE branch that generates channel weights through global average pooling and gating mechanisms and multiplies them element-wise with the input. The sum of the three branches is then activated by SiLU to obtain attention-enhanced features. After the bottleneck outputs and the retained backbone branches at each stage are concatenated along the channel dimension, feature integration is completed through 1×1 convolution to obtain the network output; Based on the trade-off between LitePAN fusion path computation and channel interaction, the two 3×3 standard convolutions at the Neck position are replaced with GSConv modules.

7. The LNG loading valve identification method under complex conditions according to claim 6, characterized in that, The GSConv module comprises two parallel branches: Standard Convolution (SC) focuses on modeling cross-channel correlations, while Depthwise Separable Convolution (DSC) captures spatial context with lower computational cost; each branch generates... After each channel is concatenated, the channels are stitched together, and the channel shuffle module promotes the uniform diffusion of channel information from the SC in the overall features.

8. The LNG loading valve identification method under complex conditions according to claim 6, characterized in that, Based on the task and network depth, set the number of bottlenecks and whether to enable identity residual connections as needed to achieve a balance between computational overhead and discriminability.

9. The LNG loading valve identification method under complex conditions according to claim 3, characterized in that, The WIoU v3 loss function is expressed as follows: ； in, ； ； ； ； ； In the formula, These represent the WIoU v3 loss function and the WIoU v1 loss function, respectively. Indicates the non-monotonic focusing coefficient. Represents the outlier parameter. and All of these represent adjustable hyperparameters, whose values ​​are optimized and adapted according to the characteristics of different models. Represents the center coordinates of the true bounding box. and These are the width and height of the minimum bounding box, respectively. Indicates the center coordinates of the prediction box, symbol express and It has been separated from the computation graph; Indicates the IoU loss value. This represents the monotonic focusing coefficient.

10. An LNG loading valve identification system for complex conditions based on the method of any one of claims 1 to 9, characterized in that, The system includes: The first module constructs a valve dataset for model training; The second module constructs a lightweight, multi-scale feature fusion valve detection network, LitePAN-YOLO. The third module is used to train the valve detection network LitePAN-YOLO using the valve dataset; The fourth module is used to identify LNG loading valves under complex conditions based on the trained valve detection network LitePAN-YOLO model.

Images