Ground surface liquid accumulation identification method, device and system based on combination of multi-expert mechanism and positive sample generation

Abstract

The invention discloses a ground area liquid accumulation identification method, device and system based on combination of a multi-expert mechanism and positive sample generation. The method comprises the following steps: acquiring original pictures of all inspection point locations without ground area liquid accumulation; obtaining a to-be-detected image of any inspection point position shot in the inspection process; inputting the to-be-detected image into a pre-constructed and trained target detection model based on a multi-expert mechanism to obtain a candidate hydrops target frame; carrying out edge expansion processing on the candidate hydrops target frame; intercepting a candidate original target frame corresponding to the position on the original picture; and matching the candidate effusion target frame with the candidate original target frame, and judging whether effusion is detected in the to-be-detected image according to a matching result. According to the method, different effusion target types or environmental conditions are processed through a plurality of expert models, the accuracy and robustness of detection are improved, and false positive targets are effectively eliminated, false detection is reduced and the false alarm rate is reduced by comparing the candidate targets with the corresponding areas in the original picture.

Description

Technical Field

[0001] The present invention belongs to the technical field of robot inspection, and in particular relates to a method, device and system for identifying liquid accumulation on the ground based on a combination of a multi-expert mechanism and positive sample generation. Background Art

[0002] Identification of liquid accumulation on the ground in industrial chemical scenes is a vital task. The accumulated liquid may contain harmful chemicals. If it is not detected and treated in time, it may cause serious safety accidents such as leakage, explosion or fire. It is not only related to production safety and environmental protection, but also directly affects the economic benefits and social responsibility of enterprises. Through effective identification and management of liquid accumulation on the ground, accidents can be prevented, pollution can be reduced, costs can be reduced, corporate image can be enhanced, and sustainable development can be promoted. Therefore, the development of advanced liquid accumulation recognition technology has important practical significance and long-term value.

[0003] Early industrial liquid accumulation detection mainly relied on manual inspections. This method is time-consuming and labor-intensive, and is easily affected by human factors, and the detection results are not accurate enough. With the development of technology, fixed water level sensors have begun to be used for liquid accumulation monitoring. Although this method can provide continuous data, its coverage is limited and cannot fully reflect the situation of the entire industrial area. In addition, it may be difficult to install fixed sensors in complex terrain or special environments. Sensors may be affected by environmental factors such as dust and humidity, resulting in inaccurate data. Sensors are exposed to harsh environments for a long time, are prone to failure, and require regular maintenance and replacement.

[0004] In recent years, deep learning technology has been widely used in the field of image recognition, including effusion detection. For example, the Yolo network model can be used to analyze images taken by the camera to effectively identify the effusion area. However, the detection accuracy and recall rate of a single Yolo network model will decrease in complex environments, especially in industrial environments, where factors such as lighting and occlusion will affect the detection effect. Moreover, in complex backgrounds, the model may have false detections (misjudging non-effusion areas as effusions) and missed detections (failing to detect actual effusions), both of which will affect the reliability of detection. Summary of the invention

[0005] Technical purpose: In response to the above technical problems, the present invention proposes a method, device and system for identifying ground fluid accumulation based on a combination of a multi-expert mechanism and positive sample generation. It uses multiple expert models to process different types of fluid accumulation targets or environmental conditions, thereby improving the accuracy and robustness of detection. By comparing candidate targets with corresponding areas in the original image, it effectively eliminates false positive targets, reduces false detections, and reduces false alarm rates.

[0006] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution:

[0007] A method for identifying liquid accumulation on the ground based on a combination of a multi-expert mechanism and positive sample generation comprises the following steps:

[0008] Get the original image without ground liquid at each inspection point n ;

[0009] Obtain the image to be inspected of any inspection point taken during the inspection process;

[0010] The image to be detected is input into a pre-built and trained target detection model based on a multi-expert mechanism to obtain a candidate effusion target frame T, wherein the training data set of the target detection model based on the multi-expert mechanism includes the original image O n Positive sample images with effusion generated by combining large image model simulation;

[0011] Expand the candidate effusion target frame T;

[0012] Retrieve the original image of the inspection point corresponding to the image to be detected, and intercept the candidate original target frame T' corresponding to the position of the candidate effusion target frame T on the retrieved original image;

[0013] Use feature matching method to match the candidate effusion target frame T and the candidate original target frame T', and output the matching result;

[0014] It is determined whether fluid accumulation is detected in the image to be detected based on the matching results. Specifically, it is determined whether the number of feature matching similarity points is less than a preset matching point threshold N. If so, it is determined that fluid accumulation is detected. If not, it is determined that fluid accumulation is not detected.

[0015] Preferably, in the training data set of the target detection model based on the multi-expert mechanism, the positive sample images with effusion are generated by the following steps:

[0016] Pre-collecting images containing effusion at each inspection point, and marking each image containing effusion, wherein the marked information includes the location, shape and size of the effusion, so as to form an image data set containing effusion;

[0017] Pre-train a large image generation model and fine-tune it according to formula (1):

[0018]

[0019] Among them, Z0 is the original clean sample encoding information, ∈ is the added Gaussian noise, and Z t is the sample encoding information after t iterations, and α is a hyperparameter;

[0020] The image dataset containing effusion is input into the image generation model to generate positive samples, specifically:

[0021] O final = G finetuned (β EncodeText(P)+(1-β) EncodeImage(O)) (2)

[0022] Among them, P is text information, which describes the shape, size, and location of the effusion; O is the original image, EncodeText is the text encoding module, EncodeImage is the image encoding module, and β is a hyperparameter.

[0023] Preferably, it is characterized in that the steps of constructing and training the target detection model based on the multi-expert mechanism are as follows:

[0024] Using Yolov11 as the basic network, a target detection model including a backbone network and a detection head is constructed. The MOE layer is inserted under the EPSA layer of the backbone network. The MOE layer includes multiple expert models, each of which focuses on processing a predetermined type of effusion target or an effusion target under preset environmental conditions. A routing mechanism network module is designed between the EPSA layer and the MOE layer to assign weights and feature maps learned by the network to each expert model of the MOE layer. The routing mechanism network module is defined as follows:

[0025] e i =R(f(F i )) (3)

[0026] Among them, R is the routing module, e i is the expert model index assigned to, F i is the feature map, and f is the function of transformation processing;

[0027] Use the gradient descent method to optimize the CIoU loss function, update the model parameters, and obtain the target detection model based on the multi-expert mechanism. The CIoU loss function is defined as follows:

[0028]

[0029] Where N represents the number of expert models, y′ i is the prediction result of the i-th expert modeler, y is the true label, and w i is the weight assigned to the i-th expert model by the routing mechanism network module, and CIou is the standard loss function in Yolo.

[0030] Preferably, it is characterized in that the steps of obtaining the candidate effusion target T are as follows:

[0031] Input the image to be detected into the fully trained target detection model to obtain the feature map F i, the routing mechanism network module is assigned to the corresponding expert model for reasoning, and the expert model obtains F after reasoning i ', defined as follows;

[0032]

[0033] Among them, E is the expert model, e i is the expert model index assigned to, w i Automatically assign network modules to expert models for routing mechanisms The weight of F i ′ is the expert model Processed feature map;

[0034] The feature map F processed by the expert model i 'Send it to the detection head, continue reasoning, and finally obtain the candidate effusion target box T.

[0035] Preferably, the candidate effusion target T′ and the original candidate target T are feature matched according to equations (6) and (7), and the number of matching feature points is calculated:

[0036]

[0037] Where i and j are the indices of T' and T respectively, x is the eigenvector, is the i-th eigenvector of T', is the jth eigenvector of T, dot is the vector click operation, Linear is the linear transformation operation, S ij Match the similarity score matrix between the feature points of the current T' and T;

[0038]

[0039] Among them, τ represents the similarity threshold, I represents the indicator function, when S ij <τ, I is 1, otherwise I is 0, and M is the number of points in the match.

[0040] A device for identifying liquid accumulation on the ground based on a combination of a multi-expert mechanism and positive sample generation, comprising:

[0041] The original image acquisition module is used to obtain the original image without ground liquid at each inspection point. n ,

[0042] The image acquisition module to be detected is used to obtain the image to be detected of any inspection point taken during the inspection process;

[0043] The target detection model based on the multi-expert mechanism is used to input the image to be detected and obtain the candidate effusion target box T. Its training data set includes the original image On Positive sample images with effusion generated by combining large image model simulation;

[0044] An edge expansion processing module is used to expand the candidate effusion target frame T;

[0045] An image capture module is used to retrieve the original image of the inspection point corresponding to the image to be detected, and capture the candidate original target frame T' corresponding to the position corresponding to the candidate effusion target frame T on the retrieved original image;

[0046] A feature matching module uses a feature matching method to match the candidate effusion target frame T and the candidate original target frame T' and outputs the matching result;

[0047] The output module is used to determine whether fluid accumulation is detected in the image to be detected based on the matching results. Specifically, it is used to determine whether the number of feature matching similarity points is less than a preset matching point threshold N. If so, it is determined that fluid accumulation is detected; if not, it is determined that fluid accumulation is not detected.

[0048] Preferably, the device further comprises: an image generating large model based on the original image O n , simulate and generate positive sample images with effusion.

[0049] Preferably, the target detection model based on the multi-expert mechanism includes a backbone network and detection, the backbone network uses the Yolo model as the basic model, and the MOE layer is inserted under the EPSA layer of the backbone network. The MOE layer includes multiple expert models, each of which focuses on processing a predetermined type of effusion target or an effusion target under preset environmental conditions; a routing mechanism network module is provided between the EPSA layer and the MOE layer, which is used to assign weights and feature maps learned by the network to each expert model of the MOE layer, and the routing mechanism network module is defined as follows:

[0050] e i =R(f(F i )) (3)

[0051] Among them, R is the routing module, e i is the expert model index assigned to, F i is the feature graph.

[0052] A patrol robot system is provided with the ground liquid recognition device based on the combination of multi-expert mechanism and positive sample generation.

[0053] Beneficial effects: Due to the adoption of the above technical solution, the present invention has the following beneficial effects:

[0054] (1) The present invention uses multiple expert models to handle different effusion target types or environmental conditions. Each expert model focuses on a specific task, which improves the accuracy and robustness of detection. It is more flexible and accurate than a single model. At the same time, the large image model is used to generate positive samples, which can effectively increase the diversity of training data, enable the model to better generalize to unseen data, and improve the adaptability of the model.

[0055] (2) The present invention utilizes the inspection robot to construct fixed inspection point information and introduces feature matching as a post-processing verification step. By comparing the candidate targets with the corresponding areas in the original image, false positive targets can be effectively excluded, false detections can be reduced, and the false alarm rate can be lowered.

[0056] (3) The present invention is highly automated from the construction of inspection points, generation of positive samples, training of target detection models to the final identification of liquid on the ground. Each module can be optimized and upgraded independently, which is convenient for the long-term maintenance and development of the system, reduces the need for manual intervention, and reduces operating costs. In addition, the present invention can also be expanded to detection scenarios of other industrial-grade complex environments such as flames and smoke. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flow chart of the method for identifying liquid on the ground based on the combination of multi-expert mechanism and positive sample generation of the present invention;

[0058] Figure 2 This is a positive sample image example 1 generated by the image generation large model of the present invention;

[0059] Figure 3 This is a positive sample image example 2 generated by the image generation large model of the present invention;

[0060] Figure 4 Schematic diagram of the MOE-EPSA structural block based on the multi-expert mechanism of the present invention;

[0061] Figure 5 This is a network structure diagram of the multi-expert mechanism module of the present invention. DETAILED DESCRIPTION

[0062] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0063] Embodiment 1

[0064] like Figures 1 to 5 As shown, this embodiment proposes a method for identifying liquid on the ground based on a combination of a multi-expert mechanism and positive sample generation, comprising the following steps:

[0065] S1. Inspection robot system builds multiple inspection points G n, take the original picture of all points without effusion O n ;

[0066] S2, using the original image O n Combined with the large image model simulation, positive sample images with effusion are generated for subsequent target detection model training and testing;

[0067] S3, constructing a target detection model based on a multiple expert (MOE) mechanism, and training it with the positive sample images of effusion simulated and generated in step S2;

[0068] S4, the inspection robot performs inspection and arrives at a certain inspection point G i After that, the image P to be detected at the point is captured and inferred through the trained target detection model to obtain the candidate effusion target T, where the candidate effusion target frame T mainly refers to the area frame on the image that is judged to be effusion, and is used to mark the possible location of effusion;

[0069] S5: After expanding the position of the candidate effusion target T by a certain multiple, i The corresponding position of the image is intercepted to obtain T', and T and T' are matched by feature matching. If the number of matching points is higher than a certain threshold N, false detection of effusion can be eliminated.

[0070] Further, in step S2, using the original image O n Combining the large image model to generate positive samples includes the following steps:

[0071] S21: Collect in advance a dataset of images of real industrial scenes containing effusions, and annotate them. By annotating the location of the effusion (for example, adding a frame), the effusion area can be accurately located to help the model learn the specific location of the effusion. The annotated text description can provide more detailed information about the effusion, such as by describing the shape, size, color, etc. of the effusion, which helps the model better understand and learn the various characteristics of the effusion. Compared with unannotated images, the advantage of adding annotated images to generate positive samples is that the annotation can help the model understand the existence and characteristics of the effusion more accurately, thereby improving the accuracy of the generated image; through detailed annotation, more diverse effusion scenes can be generated, covering different locations, shapes, sizes, etc., making the generated data richer and more diverse. For the problem of lack of positive samples in complex industrial environments, detailed annotation can help the generation model create effusion samples that meet the actual situation, thereby improving the recall rate of the algorithm for detecting effusions. In the present invention, the pre-trained image generation large model Stable Diffusion is used for fine-tuning. The generation large model does not need to be fine-tuned and trained for each inspection. The fine-tuned generation large model is defined as follows,

[0072]

[0073] Among them, Z0 is the original clean sample encoding information, ∈ is the added Gaussian noise, and Z t is the sample encoding information after t iterations, t∈[1,N], N represents the number of iterations, and α is a hyperparameter.

[0074] S22. In the inference stage, by adjusting the background area, combining text prompts and normal images, a large model is generated to generate effusions of any shape. The industrial scene has a complex environment and lacks positive samples. Generating rich and diverse positive sample data through simulation helps improve the recall rate of the algorithm.

[0075] Examples of effusion formation include Figure 2 ,The process of generating positive samples through the image large model is as follows:

[0076] O final = G finetuned (β EncodeText(P)+(1-β) EncodeImage(O)) (2)

[0077] Among them, P is text information, which may include the shape, size, and location of the effusion, O is the original image, EncodeText and EncodeImage are encoding modules, and β is a hyperparameter.

[0078] Furthermore, in step S3, constructing a target detection model based on a multi-expert mechanism includes the following steps:

[0079] S31. During the training phase, the data set for training the target detection model includes the simulated effusion data generated in step S2, the real effusion data set of industrial scenes accumulated in advance, and the background data of some industrial scenes. The pre-accumulated industrial scene effusion data set is the real effusion data accumulated in the related work before the project, and is the positive sample data set together with the simulated effusion data. All effusion targets in the above data sets need to be labeled, and the labeling information includes the location and category of the effusion target.

[0080] S32. Use Yolov11 as the basic model because Yolov11 has significantly reduced the number of parameters while maintaining excellent performance compared to the previous Yolo series, and the detection effect is even better. This breakthrough indicates that Yolov11 will run more efficiently and quickly on edge devices, and perform well in terms of real-time and accuracy.

[0081] S33. In order to further improve the accuracy and robustness of detection, a target detection model based on a multi-expert mechanism is constructed. The MOE layer is added to the EPSANet structure block of the Yolov11 backbone network to obtain the MOE-EPSA structure block. The structure diagram is shown in the figure Figure 3 As shown, each expert model E i Focus on processing specific types of effusion targets or specific environmental conditions. For example, expert E1 focuses on effusion of small targets, expert E2 focuses on effusion of large targets, and expert E3 prefers occluded effusion.

[0082] S34. Design a routing mechanism network module to automatically learn the weights assigned to each expert through overall model training. The feature graph learned by the network will be controlled by the routing mechanism network module to which expert models it is assigned. Multiple experts can be selected for comprehensive testing, which helps the network to perform adaptive learning adjustments and automatically assign expert models. The routing mechanism network module contains multiple levels of decision logic, automatically adjusts the allocation strategy through learning, and dynamically adjusts the weights assigned to each expert based on the content of the input feature graph, annotation information, and historical learning results.

[0083] MOE and routing network structure Figure 4 As shown, the definition is as follows:

[0084] e i =R(f(F i )) (3)

[0085] Where R is the routing module, e i is the expert model index assigned to, F i is the feature map, f is used to map the feature map F i Function that performs transformation processing.

[0086] S35. Due to the existence of the MoE layer, the feature map will be assigned to multiple experts for processing. Each expert may produce different prediction results, so these results need to be integrated. Therefore, the CIoU function definition needs to be modified as follows:

[0087]

[0088] Where N is the number of experts, y′ i is the prediction result of the i-th expert model, y is the true label, and w i is the weight assigned by the routing module to the i-th expert model, and CIou is the standard loss function in Yolo.

[0089] S36. Optimize the CIoU loss function through the gradient descent method, thereby updating the model parameters and obtaining a target detection model based on a multi-expert mechanism.

[0090] Further, in step S4, the method for obtaining the candidate effusion target T includes the following steps:

[0091] S41. In the inference stage, the image to be detected I is input into the well-trained detection model to obtain the feature map Fi , and the routing module determines which experts will perform reasoning to obtain F i ', sent to the subsequent detection head, the network continues to reason to obtain the final candidate target box T, which is defined as follows;

[0092]

[0093] Among them, w i is the weight automatically assigned to an expert model by the routing module, E is the expert model, e i is the expert model index assigned to, F i ′ is the feature map processed by the expert model, F i 'After subsequent processing by the detection head network, a more accurate bounding box and a higher confidence level will be obtained.

[0094] Further, in step S5, the candidate frame ground liquid recognition method includes the following steps:

[0095] S51, expand the position of the candidate target T of the inspection point by a certain multiple and then add it to the original image O i The corresponding position of is intercepted to obtain T', and the default setting in the present invention is 1.35 times. The reason for intercepting the original image after expanding the edge is that the shooting position of the robot has a certain error. Expanding the edge and intercepting can prevent the candidate target from being lost in the original image;

[0096] S52, feature matching is performed on the candidate effusion target T' and the original candidate target T by an efficient and accurate feature matching method, and the number of matching feature points is calculated. The formula is defined as follows (6):

[0097]

[0098] Where i and j are the indices of T' and T respectively, x is the eigenvector, is the i-th eigenvector of T', is the jth eigenvector of T, dot is the vector click operation, Linear is the linear transformation operation, S ij The similarity score matrix is ​​matched for the feature points of the current T' and T. In the matching process, in order to accurately match the features, it is necessary to rely on the image pixel information inside T and T'.

[0099]

[0100] Among them, τ is a threshold, I(S ij <τ) is the indicator function, when S ij <τ, it is 1, otherwise it is 0, and M is the number of points in the match.

[0101] S53, if the matched feature points exceed a certain threshold value N, it is considered that T is the original target existing in T', and the candidate target is a false detection item. If the matched feature points do not exceed the threshold value N, it is considered that T is not the original target existing in T', and the candidate target is a correctly detected effusion, which is expressed as the following formula. Through the post-processing method of feature matching, the occurrence of false detection in complex environments is reduced, and the target is identified more accurately.

[0102]

[0103] Among them, Flag is whether effusion is detected, M is the number of similar points of feature matching, and N is the constant threshold.

[0104] The present invention adopts a method of combining a multi-expert mechanism with positive sample generation, which can overcome the influence of factors such as illumination and occlusion and improve the detection effect. The multi-expert mechanism scheme uses multiple expert models to handle different effusion target types or environmental conditions, so that even in the case of illumination changes or partial occlusion, it can accurately identify ground effusion. By using a large image model to generate additional positive samples, the diversity and quantity of training data are increased, which not only helps the model learn more about the appearance of effusion, but also enhances the model's ability to understand illumination changes and occlusion, thereby improving its generalization ability and accuracy.

[0105] Embodiment 2

[0106] This embodiment proposes a device for identifying liquid on the ground based on a combination of a multi-expert mechanism and positive sample generation, and executes the identification method in the first embodiment. The device includes:

[0107] Inspection point construction module, used to construct multiple inspection points in the inspection area n ;

[0108] The original image acquisition module is used to obtain the original image without ground surface liquid at each inspection point. n ,

[0109] The image acquisition module to be detected is used to obtain the image to be detected of any inspection point taken during the inspection process;

[0110] The target detection model based on the multi-expert mechanism is used to input the image to be detected and obtain the candidate effusion target box T. Its training data set includes the original image O n Positive sample images with effusion generated by combining large image model simulation;

[0111] An edge expansion processing module is used to expand the candidate effusion target frame T;

[0112] An image capture module is used to retrieve the original image of the inspection point corresponding to the image to be detected, and capture the candidate original target frame T' corresponding to the position corresponding to the candidate effusion target frame T on the retrieved original image;

[0113] A feature matching module uses a feature matching method to match the candidate effusion target frame T and the candidate original target frame T' and outputs the matching result;

[0114] The output module is used to determine whether effusion is detected in the image to be detected according to the matching result, specifically, to determine whether the number of feature matching similarity points is less than a preset matching point threshold N. If so, it is determined that effusion is detected; if not, it is determined that effusion is not detected;

[0115] Image generation model, based on the original image O n , simulate and generate positive sample images with effusion.

[0116] Among them, the target detection model based on the multi-expert mechanism includes a backbone network and detection. The backbone network uses the Yolov11 model as the basic model. The MOE layer is inserted under the EPSA layer of the backbone network. The MOE layer includes multiple expert models, each of which focuses on processing a predetermined type of effusion target or an effusion target under preset environmental conditions. A routing mechanism network module is provided between the EPSA layer and the MOE layer to assign weights and feature maps learned by the network to each expert model of the MOE layer. The routing mechanism network module is defined as follows:

[0117] e i =R(f(F i )) (3)

[0118] R is the routing module, e i is the expert model index assigned to, F i is the feature map, and f is the function of transformation processing.

[0119] In another embodiment of the present invention, a patrol robot system is provided, which is provided with a ground liquid accumulation recognition device based on a combination of a multi-expert mechanism and positive sample generation as in the second embodiment.

[0120] The present invention improves the accuracy and robustness of detection by constructing a target detection model based on a multi-expert mechanism. At the same time, the generalization ability and adaptability of the model are improved through the method of inspection point construction and positive sample generation. In addition, the present invention reduces the occurrence of false detection in complex environments through the post-processing method of feature matching, identifies targets more accurately, and solves the problem that some previous methods use sensors to sense effusion, but fixed sensors can only monitor the situation at specific locations and cannot fully cover the entire industrial area. It makes up for the deficiency that some other existing methods, even if they combine the robot inspection method, do not use the information of the original collected pictures.

[0121] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any form, and any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. A method for identifying liquid on the ground based on a combination of a multi-expert mechanism and positive sample generation, characterized in that: The following steps are involved: Construct multiple inspection points in the inspection area; Get the original image without ground liquid at each inspection point n ; Obtain the image to be inspected of any inspection point taken during the inspection process; The image to be detected is input into a pre-built and trained target detection model based on a multi-expert mechanism to obtain a candidate effusion target frame T, wherein the training data set of the target detection model based on the multi-expert mechanism includes the original image O n Positive sample images with effusion generated by combining large image model simulation; Expand the candidate effusion target frame T; Retrieve the original image of the inspection point corresponding to the image to be detected, and intercept the candidate original target frame T' corresponding to the position of the candidate effusion target frame T on the retrieved original image; Use feature matching method to match the candidate effusion target frame T and the candidate original target frame T', and output the matching result; It is determined whether fluid accumulation is detected in the image to be detected based on the matching results. Specifically, it is determined whether the number of feature matching similarity points is less than a preset matching point threshold N. If so, it is determined that fluid accumulation is detected. If not, it is determined that fluid accumulation is not detected.

2. According to claim 1, a method for identifying liquid accumulation on the ground based on a combination of a multi-expert mechanism and positive sample generation, characterized in that: In the training data set of the target detection model based on the multi-expert mechanism, positive sample images with effusion are generated by the following steps: Pre-collecting images containing effusion at each inspection point, and marking each image containing effusion, wherein the marked information includes the location, shape and size of the effusion, so as to form an image data set containing effusion; Pre-train a large image generation model and fine-tune it according to formula (1): Among them, Z0 is the original clean sample encoding information, ∈ is the added Gaussian noise, and Z t is the sample encoding information after t iterations, and α is a hyperparameter; The image dataset containing effusion is input into the image generation model to generate positive samples, specifically: O final =G finetuned (βEncodeText(P)+(1-β)EncodeImage(O))(2) Among them, P is text information, which describes the shape, size, and location of the effusion; O is the original image, EncodeText is the text encoding module, EncodeImage is the image encoding module, and β is a hyperparameter.

3. According to claim 1, a method for identifying liquid on the ground based on a combination of a multi-expert mechanism and positive sample generation is characterized in that: The steps to build and train a multi-expert mechanism-based target detection model are as follows: Using Yolo as the basic network, a target detection model including a backbone network and a detection head is constructed. The MOE layer is inserted under the EPSA layer of the backbone network. The MOE layer includes multiple expert models, each of which focuses on processing a predetermined type of effusion target or an effusion target under preset environmental conditions. A routing mechanism network module is designed between the EPSA layer and the MOE layer to assign weights and feature maps learned by the network to each expert model of the MOE layer. The routing mechanism network module is defined as follows: e i =R(f(F i ))(3) Among them, R is the routing module, e i is the expert model index assigned to, F i is the feature map, f is the function of transformation processing; Use the gradient descent method to optimize the CIoU loss function, update the model parameters, and obtain the target detection model based on the multi-expert mechanism. The CIoU loss function is defined as follows: Where N represents the number of expert models, y ′ i is the prediction result of the i-th expert modeler, y is the true label, and w i is the weight assigned to the i-th expert model by the routing mechanism network module, and CIou is the standard loss function in Yolo.

4. The method for identifying liquid on the ground based on a combination of a multi-expert mechanism and positive sample generation according to claim 1, characterized in that: The steps to obtain the candidate effusion target T are as follows: Input the image to be detected into the fully trained target detection model to obtain the feature map F i , the routing mechanism network module is assigned to the corresponding expert model for reasoning, and the expert model obtains F after reasoning i ', defined as follows; Among them, E is the expert model, e i is the expert model index assigned to, w i Automatically assign network modules to expert models for routing mechanisms The weight of F i ′ is the expert model Processed feature map; The feature map F processed by the expert model i 'Send it to the detection head, continue reasoning, and finally obtain the candidate effusion target box T.

5. The method for identifying liquid on the ground based on a combination of a multi-expert mechanism and positive sample generation according to claim 1, characterized in that: According to equations (6) and (7), feature matching is performed between the candidate effusion target T' and the original candidate target T, and the number of matching feature points is calculated: Where i and j are the indices of T' and T respectively, x is the eigenvector, is the i-th eigenvector of T', is the jth eigenvector of T, dot is the vector click operation, Linear is the linear transformation operation, S ij Match the similarity score matrix between the feature points of the current T' and T; Among them, τ represents the similarity threshold, I represents the indicator function, when S ij <τ, I is 1, otherwise I is 0, and M is the number of points in the match.

6. A device for identifying liquid on the ground based on a combination of a multi-expert mechanism and positive sample generation, characterized in that: include: Inspection point construction module, used to construct multiple inspection points in the inspection area n ; The original image acquisition module is used to obtain the original image without ground surface liquid at each inspection point. n , The image acquisition module to be detected is used to obtain the image to be detected of any inspection point taken during the inspection process; The target detection model based on the multi-expert mechanism is used to input the image to be detected and obtain the candidate effusion target box T. Its training data set includes the original image O n Positive sample images with effusion generated by combining large image model simulation; An edge expansion processing module is used to expand the candidate effusion target frame T; An image capture module is used to retrieve the original image of the inspection point corresponding to the image to be detected, and capture the candidate original target frame T' corresponding to the position corresponding to the candidate effusion target frame T on the retrieved original image; A feature matching module uses a feature matching method to match the candidate effusion target frame T and the candidate original target frame T' and outputs the matching result; The output module is used to determine whether fluid accumulation is detected in the image to be detected based on the matching results. Specifically, it is used to determine whether the number of feature matching similarity points is less than a preset matching point threshold N. If so, it is determined that fluid accumulation is detected; if not, it is determined that fluid accumulation is not detected.

7. The device for identifying liquid on the ground based on a combination of a multi-expert mechanism and positive sample generation according to claim 6, characterized in that: The device also includes: an image generating large model based on the original image. n , simulate and generate positive sample images with effusion.

8. The device for identifying liquid on the ground based on a combination of a multi-expert mechanism and positive sample generation according to claim 6, characterized in that: The target detection model based on the multi-expert mechanism includes a backbone network and detection. The backbone network uses the Yolov11 model as the basic model. The MOE layer is inserted under the EPSA layer of the backbone network. The MOE layer includes multiple expert models, each of which focuses on processing a predetermined type of effusion target or an effusion target under preset environmental conditions. A routing mechanism network module is provided between the EPSA layer and the MOE layer to assign weights and feature maps learned by the network to each expert model of the MOE layer. The routing mechanism network module is defined as follows: e i =R(f(F i ))(3) Among them, R is the routing module, e i is the expert model index assigned to, F i is the feature map, and f is the function of transformation processing.

9. A patrol robot system, characterized in that: A device for identifying liquid on the ground based on a combination of a multi-expert mechanism and positive sample generation as described in any one of claims 6 to 8 is provided.

Images