Groundwater monitoring method and device based on quadruped robot and electronic equipment

Abstract

The application provides a groundwater monitoring method and device based on a quadruped robot and electronic equipment, and relates to the technical field of environmental monitoring. The quadruped robot is provided with an RGB camera and a thermal infrared camera. The groundwater monitoring method comprises the following steps: based on a K-means algorithm, a wellhead is clustered and divided into multiple operation subareas through wellhead longitude and latitude coordinates; a path planning is performed on each operation subarea by minimizing a path cost function, and an operation subarea target path is obtained; the quadruped robot is controlled to move according to each operation subarea target path respectively, and RGB images and thermal infrared images are obtained; the RGB images and the thermal infrared images are input into a multi-modal data fusion extraction model, and wellhead information is obtained; and the quadruped robot is controlled to sample through wellhead types, wellhead position segmentation maps and wellhead state evaluation scores. The application realizes improved groundwater monitoring efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of environmental monitoring, in particular to a groundwater monitoring method and device based on a quadruped robot and an electronic device. BACKGROUND

[0002] Groundwater monitoring refers to data collection of monitoring wells, and further analysis of dynamic changes of groundwater to master the status of groundwater resources. At present, groundwater monitoring has long relied on manual sampling, which requires personnel to collect water samples on site, or to perform simple measurements, or to take them back to the laboratory for analysis. Due to the wide distribution of monitoring wells, especially the wide distribution and remote location (such as forest areas, mining areas, deserts, and protected areas) of monitoring wells, the traffic is inconvenient and the environment is harsh, resulting in low efficiency of manual sampling, and further resulting in low efficiency of groundwater monitoring. SUMMARY

[0003] The problem solved by the present application is the low efficiency of manual sampling caused by the wide distribution of monitoring wells, and the low efficiency of groundwater monitoring.

[0004] To solve the above problems, the present application provides a groundwater monitoring method, device and electronic equipment based on a quadruped robot.

[0005] In a first aspect, the present application provides a groundwater monitoring method based on a quadruped robot, wherein the quadruped robot is provided with an RGB camera and a thermal infrared camera; the groundwater monitoring method comprises:

[0006] Based on the K-means algorithm, the wellheads are clustered and divided into multiple operation subareas through the latitude and longitude coordinates of the wellheads;

[0007] Each operation subarea is path planned by minimizing the path cost function to obtain the corresponding operation subarea target path;

[0008] The quadruped robot is controlled to move according to each operation subarea target path respectively, and the RGB camera and the thermal infrared camera of the quadruped robot are controlled to perform image acquisition to obtain RGB images and thermal infrared images;

[0009] The RGB images and the thermal infrared images are input into a multi-modal data fusion extraction model to obtain groundwater wellhead information, wherein the groundwater wellhead information includes wellhead type, wellhead position segmentation map and wellhead state evaluation score;

[0010] The quadruped robot is controlled to sample through the wellhead type, the wellhead position segmentation map and the wellhead state evaluation score.

[0011] Optionally, the RGB image and the thermal infrared image are input into a multi-modal data fusion extraction model to obtain underground water wellhead information, comprising:

[0012] The RGB image and the thermal infrared image are subjected to feature complementary fusion processing to obtain corresponding RGB image fusion features and thermal infrared image fusion features;

[0013] An infrared image average pixel value is obtained from the thermal infrared image;

[0014] The infrared image average pixel value, a weather label, and an illumination intensity value are input into an adaptive weight generation self-network to obtain a fusion weight coefficient;

[0015] The fusion weight coefficient comprises:

[0016] ,

[0017] wherein, is an RGB fusion weight, is a thermal infrared fusion weight, AWS is the adaptive weight generation self-network, and M is a vector composed of the infrared image average pixel value, the weather label, and the illumination intensity value;

[0018] The RGB image fusion features and the thermal infrared image fusion features are subjected to weighted fusion using the fusion weight coefficient to obtain underground water wellhead fusion features;

[0019] The underground water wellhead fusion features comprise:

[0020] ,

[0021] wherein, is the underground water wellhead fusion features, is the RGB image fusion features, is the thermal infrared image fusion features;

[0022] The wellhead type, the wellhead position segmentation map, and the wellhead state evaluation score are obtained from the underground water wellhead fusion features.

[0023] Optionally, the RGB image and the thermal infrared image are subjected to feature complementary fusion processing to obtain corresponding RGB image fusion features and thermal infrared image fusion features, comprising:

[0024] The RGB image is input into an RGB feature extraction branch to obtain RGB image high-level features;

[0025] The thermal infrared image is input into a thermal infrared feature extraction branch to obtain thermal infrared image high-level features;

[0026] obtaining the RGB image fusion feature and the thermal infrared image fusion feature according to the RGB image high-level feature and the thermal infrared image high-level feature based on the cross-modal feature mutual enhancement module;

[0027] The RGB image fusion feature includes:

[0028]

[0029] wherein, is the RGB image fusion feature on the cth channel, is the RGB image high-level feature on the cth channel, is an RGB image fusion feature weight;

[0030] The thermal infrared image fusion feature includes:

[0031]

[0032] wherein, is the thermal infrared image fusion feature on the cth channel, is the thermal infrared image high-level feature on the cth channel, is a thermal infrared image fusion feature weight.

[0033] Optionally, after the wellhead type, the wellhead position segmentation map and the wellhead state evaluation score are obtained through the underground water wellhead fusion feature, the method further includes:

[0034] When the wellhead state evaluation score is excellent, the fusion weight coefficient is updated through a preset thermal infrared fusion weight to obtain an updated fusion weight coefficient;

[0035] When the wellhead state evaluation score is poor, the fusion weight coefficient is updated through a preset RGB modal fusion weight to obtain an updated fusion weight coefficient;

[0036] The RGB image fusion feature and the thermal infrared image fusion feature are weighted and fused through the updated fusion weight coefficient to obtain a new underground water wellhead fusion feature.

[0037] Optionally, the multiple operation sub-zones are obtained by clustering and dividing the wellheads through the wellhead latitude and longitude coordinates, including:

[0038] Features of each wellhead are constructed through the wellhead latitude and longitude coordinates to obtain a corresponding multi-dimensional feature vector;

[0039] The multi-dimensional feature vector includes:​​

[0040] ,

[0041] wherein, is the multi-dimensional feature vector, and is the wellhead longitude and latitude coordinate, is the task priority, is the environmental accessibility coefficient, and is the feature vector weight factor;

[0042] by minimizing the intra-cluster squared sum iteration solution to all the multi-dimensional feature vectors, a plurality of the job partitions are obtained, wherein each of the job partitions comprises at least one partition wellhead.

[0043] Optionally, the path planning for each of the job partitions by minimizing the path cost function obtains the corresponding job partition target path, comprising:

[0044] the path planning for each of the job partitions obtains a plurality of corresponding job partition candidate paths;

[0045] by performing cost calculation on each of the job partition candidate paths, a corresponding path cost is obtained;

[0046] wherein, the path cost comprises:

[0047] ,

[0048] wherein, is the adjacent inter-well distance between the tth wellhead and the t+1th wellhead, is the priority of the tth wellhead, is the access order of the partition wellhead i in the path, and M is the number of partition wellheads, is the path cost;

[0049] all the job partition candidate paths are screened using the path cost to obtain the job partition target path.

[0050] Optionally, the control of the quadruped robot to sample by the wellhead type, the wellhead location segmentation map and the wellhead state evaluation score comprises:

[0051] the wellhead area contour point set is extracted by the wellhead location segmentation map to obtain the wellhead 2D coordinate;

[0052] based on the pre-calibrated camera intrinsic parameter matrix, the wellhead 2D coordinate is converted into a normalized three-dimensional coordinate;

[0053] convert the normalized three-dimensional coordinates into quadruped robot coordinates based on a hand-eye calibration matrix;

[0054] control the quadruped robot to sample by the wellhead type, the quadruped robot coordinates and the wellhead state evaluation score.

[0055] In a second aspect, the present application provides a quadruped robot-based underground water monitoring device, the quadruped robot being provided with an RGB camera and a thermal infrared camera; the underground water monitoring device comprising:

[0056] a clustering and partitioning module configured to cluster and partition wellheads based on K-means algorithm to obtain a plurality of operation partitions;

[0057] a path planning module configured to plan a path for each operation partition by minimizing a path cost function to obtain a corresponding operation partition target path;

[0058] an image acquisition module configured to control the quadruped robot to move according to each operation partition target path respectively, and control the RGB camera and the thermal infrared camera of the quadruped robot to acquire images to obtain RGB images and thermal infrared images;

[0059] a feature fusion and extraction module configured to input the RGB images and the thermal infrared images into a multi-modal data fusion and extraction model to obtain underground water wellhead information, wherein the underground water wellhead information comprises a wellhead type, a wellhead position segmentation map and a wellhead state evaluation score;

[0060] a quadruped robot control module configured to control the quadruped robot to sample by the wellhead type, the wellhead position segmentation map and the wellhead state evaluation score.

[0061] In a third aspect, the present application provides an electronic device comprising a memory and a processor;

[0062] the memory is configured to store a computer program;

[0063] the processor is configured to, when executing the computer program, implement the quadruped robot-based underground water monitoring method of the first aspect.

[0064] In a fourth aspect, the present application provides a computer readable storage medium, the storage medium storing a computer program, and when the computer program is executed by a processor, the quadruped robot-based underground water monitoring method of the first aspect is implemented.

[0065] The groundwater monitoring method based on the quadruped robot, the device and the electronic equipment have the beneficial effects that: the K-means algorithm is used to cluster and divide the spatially dispersed wellheads through the latitude and longitude coordinates of the wellheads, so that the widely distributed monitoring points are automatically divided into multiple logical clear operation subareas, and the subjectivity and inefficiency of manual grouping are avoided. The path planning is independently carried out for each subarea by minimizing the path cost function, the operation subarea target path is generated, the empty distance of the quadruped robot and the task completion time are reduced. The quadruped robot is controlled according to the operation subarea target path, the quadruped robot has stronger terrain adaptability, can approach the wellhead positions that are difficult to reach by traditional vehicles or artificial access, and expands the monitoring coverage. The RGB camera and the thermal infrared camera are used to collect images, the multi-modal data fusion extraction model is input, the groundwater wellhead information is obtained, the complementary information of the two modes is comprehensively utilized, and the image accuracy is improved. The quadruped robot is controlled through the wellhead type, the wellhead position segmentation graph and the wellhead state evaluation score to carry out sampling, the groundwater sampling efficiency is improved, and then the groundwater monitoring efficiency is improved. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 A flowchart of a groundwater monitoring method based on a quadruped robot according to an embodiment of the present application;

[0067] Figure 2 A schematic diagram of a multi-modal data fusion extraction model according to an embodiment of the present application;

[0068] Figure 3 A structural schematic diagram of a groundwater monitoring device based on a quadruped robot according to an embodiment of the present application;

[0069] Figure 4 A structural schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0070] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings. Although some embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms, and should not be interpreted as being limited to the embodiments described herein, on the contrary, these embodiments are provided to make the present application more thorough and complete. It should be understood that the drawings and embodiments of the present application are only for illustrative purposes, and are not intended to limit the scope of protection of the present application.

[0071] It should be understood that each step described in the method embodiments of the present application can be executed in different order and / or in parallel. In addition, the method embodiments can include additional steps and / or omit the execution of the steps shown. The scope of the present application is not limited in this respect.

[0072] The term "include" and variations thereof as used herein mean "to include, without limitation"; the term "based on" means "based at least in part on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; the term "optional" means "optional in at least one embodiment". Related terms such as "one(s) of the aforementioned" and "one(s) of the aforementioned items" are to be understood to mean "one or more of the aforementioned", "one or more of the aforementioned items", etc. Other definitions will be given in the description that follows. It needs to be noted that the terms "first", "second", etc. mentioned in the present application are only used to distinguish different devices, modules or units, and are not intended to limit the order or interdependence of the functions performed by these devices, modules or units.

[0073] It needs to be noted that the modification of "one" or "multiple" mentioned in the present application is illustrative rather than restrictive, and those skilled in the art should understand that "one or more" should be understood unless otherwise explicitly indicated in the context.

[0074] The names of the messages or information exchanged between the devices in the embodiments of the present application are only for illustrative purposes, and are not intended to limit the scope of the messages or information.

[0075] As shown in Figure 1 The underground water monitoring method based on the quadruped robot provided by the embodiments of the present application is shown in the figure, and the quadruped robot is provided with an RGB camera and a thermal infrared camera; the underground water monitoring method comprises the following steps.

[0076] In step 110, based on the K-means algorithm, the wellheads are clustered and divided by the latitude and longitude coordinates of the wellheads to obtain a plurality of operation sub-zones.

[0077] Specifically, a method combining large-area clustering and small-area optimization is adopted, and first, the K-means algorithm is used to cluster and divide the spatially dispersed wellheads, which are reasonably divided into a plurality of operation sub-zones. K-means is a classic unsupervised clustering algorithm, which is used to divide a data set into K clusters, so that each data point belongs to the cluster where the nearest cluster center (centroid) is located.

[0078] In step 120, the path planning of each operation sub-zone is performed by minimizing the path cost function, and the corresponding operation sub-zone target path is obtained.

[0079] Specifically, one mobile quadruped robot is put into each operation sub-zone for operation, and in each operation sub-zone, a mobile quadruped robot path planning method based on priority and cross-entropy theory is designed and adopted. Combining the priority information of the wellheads, a mobile quadruped robot path planning model based on cross-entropy theory is constructed, and the path planning is performed to minimize the operation path cost.

[0080] Step 130, control the quadruped robot to move according to each target path of the work partition respectively, and control the RGB camera and the thermal infrared camera of the quadruped robot to collect images to obtain RGB images and thermal infrared images.

[0081] Specifically, the mobile quadruped robot moves to the specified position according to the target path of the work partition, and the quadruped robot collects images after reaching the rough target area planned by satellite navigation. Among them, the mobile quadruped robot carries two modal sensors of RGB camera and thermal infrared camera, controls the RGB camera and thermal infrared camera to collect images, and obtains RGB images and thermal infrared images.

[0082] Step 140, input the RGB images and the thermal infrared images into a multi-modal data fusion extraction model to obtain underground water wellhead information, wherein the underground water wellhead information includes wellhead type, wellhead position segmentation map and wellhead state evaluation score.

[0083] Specifically, the underground water wellhead fusion features are sent to the decoder for upsampling and fine prediction, and finally a pixel-level semantic segmentation map is output. The semantic segmentation map is used to accurately identify the position and contour of the wellhead, and provides high-precision spatial guidance for subsequent robot arm operation. In addition, the type label of the well cover and the wellhead state evaluation label are also output. The wellhead type includes circular rotary opening type and square hook lifting type, corresponding to labels 0 and 1 respectively, which are output by the wellhead type branch of the model. The wellhead state evaluation score includes "good" and "bad", corresponding to labels 0 and 1 respectively, which are output by the wellhead state branch of the model. "Good" corresponds to clean wellhead, which does not need human intervention; "bad" corresponds to serious corrosion or covering by vegetation, soil and other objects, which needs human intervention for cleaning.

[0084] Step 150, control the quadruped robot to sample through the wellhead type, the wellhead position segmentation map and the wellhead state evaluation score.

[0085] In some more specific embodiments, a multi-modal data fusion multi-task network is embedded in a mobile quadruped robot to identify and locate the "H" logo pasted on the wellhead, while outputting the wellhead type, wellhead position segmentation map and wellhead state evaluation score. The quadruped robot is also provided with a light-sensitive environment sensor, a mechanical arm, a multi-parameter water quality probe, a sampling pump and a sampling tube, and a sample box. The wellhead position segmentation map is used to guide the subsequent mechanical arm operation, the wellhead type is used to guide the subsequent mechanical arm selection, and the wellhead state evaluation score is used to indicate the wellhead state and whether manual maintenance is needed. The quadruped robot first controls the mechanical arm to open the well cover through the wellhead type, wellhead position segmentation map and wellhead state evaluation score. After the well cover is opened, the mechanical arm can perform two operations: control the multi-parameter water quality probe to measure in-situ at a specified depth underground, and transmit the measurement values back through satellite communication; or, control the sampling pump and sampling tube to extract water samples at a specified depth underground and store them in the internal sample box for further testing in the laboratory. Among them, if the in-situ measurement data is abnormal, the monitoring priority of the corresponding wellhead can be raised to achieve rapid response to pollution events. After the task is completed, the sampling device is recovered, the well cover is re-covered, and the image is collected again to confirm that the well cover has returned to its original state.

[0086] In this embodiment, based on the K-means algorithm, the spatially dispersed wellheads are clustered and divided by latitude and longitude coordinates, automatically dividing the widely distributed monitoring points into multiple logically clear operation partitions, avoiding the subjectivity and inefficiency of manual grouping. For each partition, the path planning is independently performed by minimizing the path cost function to generate the operation partition target path, reducing the empty distance of the quadruped robot and the task completion time. According to the operation partition target path of each operation partition, the quadruped robot is controlled to move, and the quadruped robot has stronger terrain adaptability, can approach the wellhead positions that are difficult for traditional vehicles or manual to reach, and expands the monitoring coverage. The RGB camera and thermal infrared camera are used for image acquisition, and the multi-modal data fusion extraction model is input to obtain groundwater wellhead information, and the complementary information of the two modalities is comprehensively utilized to improve the image accuracy. The quadruped robot is controlled for sampling by the wellhead type, wellhead position segmentation map and wellhead state evaluation score, improving the efficiency of groundwater sampling and thus improving the efficiency of groundwater monitoring.

[0087] Optionally, as shown in Figure 2 the RGB image and the thermal infrared image are input into a multi-modal data fusion extraction model to obtain groundwater wellhead information, including:

[0088] complementary feature fusion processing is performed on the RGB image and the thermal infrared image to obtain corresponding RGB image fusion features and thermal infrared image fusion features;

[0089] an infrared image average pixel value is obtained from the thermal infrared image;

[0090] inputting the infrared image average pixel value, the weather label and the illumination intensity value into an adaptive weighting sub-network to obtain a fusion weight coefficient;

[0091] The fusion weight coefficient comprises:

[0092] ,

[0093] The fusion weight coefficient comprises: is an RGB fusion weight, is a thermal infrared fusion weight, AWS is the adaptive weighting sub-network, and M is a vector composed of the infrared image average pixel value, the weather label and the illumination intensity value;

[0094] The RGB image fusion feature and the thermal infrared image fusion feature are weighted and fused by using the fusion weight coefficient to obtain an underground water wellhead fusion feature.

[0095] The underground water wellhead fusion feature comprises:

[0096] ,

[0097] The underground water wellhead fusion feature comprises: is the underground water wellhead fusion feature, is the RGB image fusion feature, is the thermal infrared image fusion feature.

[0098] The wellhead type, the wellhead position segmentation map and the wellhead state evaluation score are obtained through the underground water wellhead fusion feature.

[0099] Specifically, an infrared image average pixel value, a weather label and an illumination intensity value are input into an adaptive weighting sub-network (AWS) to obtain a fusion weight coefficient. A multi-layer perception is used for the adaptive weighting sub-network, the input of which is an environment metadata vector M, and the output is two fusion weight coefficients, wherein the fusion weight coefficient comprises:

[0100] ,

[0101] The fusion weight coefficient comprises: is an RGB fusion weight, is a thermal infrared fusion weight, AWS is the adaptive weighting sub-network, and M is an environment metadata vector composed of the infrared image average pixel value, the weather label and the illumination intensity value, wherein:

[0102] ,

[0103] wherein, .

[0104] The RGB image fusion features and the thermal infrared image fusion features are weighted and fused by using a fusion weight coefficient to obtain underground water wellhead fusion features.

[0105] In some more specific embodiments, under sufficient light conditions, the network will assign a higher weight to the RGB modality (i.e. tending to 1); while in night, bad weather or scenes with occlusion, the weight of the thermal infrared modality will be significantly increased (i.e. tending to 1).

[0106] In this optional embodiment, unlike traditional multi-modal fusion methods, the fusion weight of the network is not fixed or only learned from image features, but is dynamically generated by a lightweight adaptive weight generation network from a set of real-time collected environmental metadata. The metadata includes the illumination intensity value, the weather label, and the average pixel value of the infrared image, which together constitute a quantitative description of the current perceived environment, so that the network can adaptively adjust the weight according to the change of the environment.

[0107] Optionally, as Figure 2 shown, the feature complementary fusion processing on the RGB image and the thermal infrared image to obtain corresponding RGB image fusion features and thermal infrared image fusion features comprises:

[0108] inputting the RGB image into an RGB feature extraction branch to obtain RGB image high-level features;

[0109] inputting the thermal infrared image into a thermal infrared feature extraction branch to obtain thermal infrared image high-level features;

[0110] obtaining the RGB image fusion features and the thermal infrared image fusion features based on a cross-modal feature mutual enhancement module according to the RGB image high-level features and the thermal infrared image high-level features;

[0111] wherein, the RGB image fusion features comprise:

[0112] ,

[0113] wherein, is the RGB image fusion feature on the cth channel, is the RGB image high-level feature on the cth channel, is an RGB image fusion feature weight;

[0114] wherein, the thermal infrared image fusion features comprise:

[0115] ,

[0116] wherein, is the thermal infrared image fusion feature on the cth channel, is the thermal infrared image high-level feature on the cth channel, is a thermal infrared image fusion feature weight.

[0117] Specifically, the RGB image and the thermal infrared image are respectively processed by two independent backbone feature extraction branches, the RGB image and the thermal infrared image are respectively input into the RGB feature extraction branch and the thermal infrared feature extraction branch, and the RGB image high-level feature and the thermal infrared image high-level feature are obtained, and the feature extraction branch is used to represent a plurality of convolution blocks. Based on a cross-modal feature mutual enhancement module (Mutual Enhancement Module, MEM), the and are complementarily fused, so that the features of one modality guide and enhance the feature learning of the other modality, mine complementary information, and obtain the RGB image fusion feature and the thermal infrared image fusion feature .

[0118] wherein, the RGB image fusion feature comprises:

[0119] ,

[0120] wherein, is the RGB image fusion feature on the cth channel, is the RGB image high-level feature on the cth channel, is an RGB image fusion feature weight, c=1, 2, 3.

[0121] ,

[0122] ,

[0123] ,

[0124] ,

[0125] wherein, is the RGB image fusion feature, is the RGB image high-level feature, is an RGB image fusion feature weight, and are learnable parameters, is the compressed intermediate variable, is an activation function, and Bn() is a normalization operation, is a linear transformation weight matrix, is the feature value on each channel after is the feature value after the average operation, is the feature value after the average operation on the cth channel, is the feature map on the cth channel, H and W are the height and width of the feature map, respectively, is and is the feature map obtained by element-wise addition.

[0126] The thermal infrared image fusion features include:

[0127] ,

[0128] ,

[0129] wherein, is the thermal infrared image fusion feature on the cth channel, is the thermal infrared image high-level feature on the cth channel, is a thermal infrared image fusion feature weight.

[0130] In this optional embodiment, the cross-modal feature mutual enhancement module is trained using a large amount of data, so that it can adaptively adjust the information confidence of the two modal data corresponding branches. In a working scene with good lighting conditions and clear background, the information confidence of the RGB image corresponding branch is high, and the RGB image recognition result is mainly used; in a complex night scene, it completely relies on the thermal infrared data; in a scene with poor lighting conditions, the feature information of the two branches is fused and processed, and finally an optimal result is output.

[0131] Optionally, as shown in Figure 2 after the wellhead type, the wellhead position segmentation map and the wellhead state evaluation score are obtained through the underground water wellhead fusion feature, the method further includes:

[0132] When the wellhead state evaluation score is excellent, the fusion weight coefficient is updated by a preset thermal infrared fusion weight, to obtain an updated fusion weight coefficient;

[0133] When the wellhead state evaluation score is poor, the fusion weight coefficient is updated by a preset RGB modal fusion weight, to obtain an updated fusion weight coefficient;

[0134] The RGB image fusion features and the thermal infrared image fusion features are weighted and fused by the updated fusion weight coefficients, and new underground water wellhead fusion features are obtained.

[0135] In some more specific embodiments, for the case of wellhead state evaluation as "poor", the network will adopt a special segmentation enhancement strategy to maintain segmentation accuracy. This strategy is based on the fact that when the wellhead is partially covered by mud or corroded, the visual texture information provided by the RGB modality has been severely degraded or even misleading. The heat capacity and thermal conductivity of the wellhead are usually significantly different from the surface coverings (such as mud, grass, etc.). This difference will cause the outline of the wellhead to be highlighted in the thermal infrared image, forming a wellhead outline enhancement effect. Therefore, when the wellhead state is evaluated as "poor", the adaptive weight generation self-network AWS will make a decision adjustment, which will trigger the start of a weight reconstruction subroutine, output an extreme weight configuration, and significantly increase the fusion weight of the thermal infrared modality , set to 0.7 or higher, to suppress the weight coefficient , so that the segmentation decoding process mainly relies on the thermal structure outline of the thermal infrared image that is not affected by the wellhead surface covering, to maximize the segmentation accuracy. When the wellhead state is evaluated as "excellent", the adaptive weight generation subnetwork AWS will maintain or further increase the fusion weight of the RGB modality , set to 0.5 or higher. In this way, the segmentation decoding process will mainly rely on the rich visual texture information provided by the RGB image, so as to obtain more accurate and accurate segmentation results in good environments. The total loss function of the network is composed of the following two parts, including:

[0136] ,

[0137] wherein, is the total loss, is the difference between the predicted segmentation map and the true label, is the wellhead type classification loss, which adopts a multi-class cross-entropy loss function, and is used to identify the wellhead type (such as circular rotary opening type, square hook lifting type, etc.), is the wellhead state evaluation classification loss, which adopts a binary classification cross-entropy loss function, and is used to judge whether the wellhead state is "excellent" or "poor", is the weight parameter, is a regularization loss used to prevent the weight generation network from being too extreme, to prevent the AWS network from always outputting extreme weights.

[0138] In this optional embodiment, the adaptability, robustness, and accuracy of groundwater wellhead information perception are significantly improved by dynamically adjusting the multimodal fusion weights based on the wellhead state assessment score. The cross-modal feature enhancement module is trained using a large amount of data, enabling it to adaptively adjust the information confidence levels of the corresponding branches of the two modalities. In well-lit and clear background scenarios, the information confidence level of the RGB image's corresponding branch is high, and the RGB image recognition results are the primary focus. In complex nighttime scenarios, however, the system relies entirely on thermal infrared data.

[0139] Optionally, the step of clustering the wellhead using its latitude and longitude coordinates to obtain multiple operational zones includes:

[0140] By constructing features for each wellhead using its latitude and longitude coordinates, a corresponding multidimensional feature vector is obtained.

[0141] The multidimensional feature vector includes:

[0142] ,

[0143] in, For the multidimensional feature vector, and The latitude and longitude coordinates of the wellhead are... As a task priority, Environmental accessibility coefficient, and The eigenvector weighting factor;

[0144] By iteratively solving for minimizing the sum of squares within the cluster on all the multidimensional feature vectors, multiple job partitions are obtained, wherein each job partition includes at least one partition wellhead.

[0145] Specifically, a multidimensional feature vector is constructed for each wellhead. Characterization is performed for cluster analysis, where the multidimensional feature vectors include:

[0146] ,

[0147] in, For the multidimensional feature vector, and The coordinates of the wellhead are latitude and longitude. Task priority is determined by hydrogeological experts based on information such as the degree of groundwater pollution, hydrogeologically sensitive areas, and areas with historical data anomalies, and assigned a value (e.g., an integer from 1 to 10). The higher the priority, the greater the weight that wellhead will be considered during clustering. is the environmental accessibility coefficient, which is used to reflect the difficulty of reaching the wellhead (for example, 1 for flat areas, 1.2 for light vegetation coverage, and 1.5 for heavy vegetation or complex terrain). and is the feature vector weight factor, which is used to flexibly adjust the importance of different factors.

[0148] Then the number of partitions is determined. According to the total number of wellheads (N) of this task and the reasonable working load (Q) of a single robot, the number of partitions K is automatically calculated, and the calculation formula is: K = round up (N / Q). The feature vector set constructed above is taken as input, and K clusters and corresponding centroids are obtained by iterative solution by minimizing the within-cluster sum of squares. The similarity between clusters is measured by Euclidean distance, and the centroid is updated according to the mean value of the samples in the cluster. In addition, during the clustering process, the algorithm constraint is used to ensure that the number of wellheads contained in each partition generated finally is roughly equivalent, avoiding the generation of work partitions with a large difference in task amount, so as to realize the load balancing of the multi-robot system.

[0149] Optionally, the path planning for each work partition by minimizing the path cost function to obtain the corresponding work partition target path comprises:

[0150] The path planning for each work partition obtains a plurality of work partition candidate paths;

[0151] The path cost of each work partition candidate path is calculated to obtain the corresponding path cost;

[0152] The path cost comprises:

[0153] ,

[0154] wherein, is the adjacent well distance between the tth wellhead and the (t+1)th wellhead, is the priority of the tth wellhead, is the access order of the partition wellhead i in the path, and M is the number of partition wellheads, is the path cost;

[0155] The path cost is used to screen all the work partition candidate paths to obtain the work partition target path.

[0156] Specifically, one mobile quadruped robot is put into each work partition for work, and there are M wellheads in each work partition, and the path is represented as a permutation The path cost comprises:

[0157] ,​

[0158] wherein, is the distance between adjacent wells, is the priority of the zoned wellhead i, is the visiting order of the zoned wellhead i in the path, and M is the number of zoned wellheads, is the path cost. The optimization objective is to minimize the path cost function. The priority represents the importance of each wellhead, which can be specified by hydrogeology experts. The more important the wellhead, , the larger the priority. The priority is specified by hydrogeology experts according to information such as the degree of groundwater pollution, hydrogeological sensitive areas, and historical data anomaly areas. For example, wellhead i needs to be monitored before wellhead j, then .

[0159] In some more specific embodiments, the zoned wellhead coordinates and priorities are first output, and the path probability distribution is initialized. An iterative update step is performed, first sampling, generating N candidate paths from the current distribution . Second, evaluation is performed, calculating the cost of each path . High-quality paths are screened, and the top P percent of candidate paths are selected as the elite sample set , i.e. candidate paths with lower cost. Parameter update, minimizing cross-entropy, equivalent to maximizing the log-likelihood of the elite sample, including:

[0160] ,

[0161] wherein, is the model parameter at time t+1 . Iteration until the distribution converges, obtaining the optimal path considering the priority, which is used as the operation zoned target path.

[0162] In this optional embodiment, wellhead priority and visiting order are introduced into the path cost function, and the algorithm will automatically tend to place high-priority wellheads at the front of the path to reduce the overall cost, achieving an intelligent scheduling strategy and improving emergency response capabilities. At the same time, the path cost also includes the distance between adjacent wells, ensuring that the total path length is reasonable. Based on the multi-candidate path screening mechanism, intelligent path planning is achieved, significantly improving the operation efficiency of the quadruped robot in wide-area groundwater monitoring.

[0163] Optionally, the control of the quadruped robot to sample through the wellhead type, the wellhead location segmentation map, and the wellhead state evaluation score includes:

[0164] extracting wellhead region contour point sets through the wellhead location segmentation map to obtain wellhead 2D coordinates;

[0165] Converting the wellhead 2D coordinates into normalized three-dimensional coordinates based on the pre-calibrated camera intrinsic matrix;

[0166] Converting the normalized three-dimensional coordinates into quadruped robot coordinates based on the hand-eye calibration matrix;

[0167] Controlling the quadruped robot to sample through the wellhead type, the quadruped robot coordinates, and the wellhead state evaluation score.

[0168] In some more specific embodiments, first, the wellhead contour and center point coordinates are extracted, and after the multi-task network outputs the pixel-level wellhead segmentation map, the contour point set of the wellhead region is extracted from the segmentation map, and the 2D coordinates (u, v) of the wellhead center in the image are calculated through the contour point set:

[0169] ,

[0170] wherein u is the horizontal direction coordinate, v is the vertical direction coordinate, is the horizontal direction coordinate of the i-th contour point, where z is the vertical direction coordinate of the i-th point, and N is the number of contour points. Meanwhile, the network outputs the wellhead type label, which is used for subsequent selection of robotic arm tools and cover opening strategies. Meanwhile, the network outputs the wellhead type label, which is used for subsequent selection of robotic arm tools and cover opening strategies. Using the pre-calibrated camera intrinsic matrix K, the 2D coordinates (u, v) of the image center are converted to normalized 3D coordinates (x_c, y_c, 1) in the camera coordinate system, and the spatial ray direction from the camera optical center to the wellhead center is determined. Through the hand-eye calibration matrix T_cam_base (this matrix describes the fixed transformation relationship between the camera coordinate system and the robot base coordinate system) obtained by pre-calibration, the 3D ray obtained in the previous step is converted from the camera coordinate system to the robot base coordinate system. Assuming that the wellhead is located on the ground plane, in the robot base coordinate system, this ground plane can be approximately represented by the equation Z = 0. The spatial ray converted to the base coordinate system is intersected with this ground plane equation, and the 3D coordinates (X, Y, Z) of the wellhead center in the robot base coordinate system are calculated. In order to guide the robotic arm to approach the manhole cover in the correct posture, the normal vector (Nx, Ny, Nz) of the manhole cover needs to be estimated. By default, it is assumed that the manhole cover is installed horizontally, so its normal vector is (0, 0, 1) by default. Finally, the system sends the calculated 3D coordinates (X, Y, Z) and normal vector (Nx, Ny, Nz) to the robotic arm. The wellhead state includes "good" and "bad", corresponding to labels 0 and 1, respectively, which are output by the wellhead state branch of the model. "Good" corresponds to a clean wellhead that does not require human intervention; "bad" corresponds to a wellhead that is severely corroded or covered with vegetation, soil, etc., and requires human intervention for cleaning. At the same time, the system automatically selects a matching cover opening tool from the tool library and calls the corresponding cover opening action from the action library according to the recognized wellhead type. For circular rotating opening manhole covers, a high-torque three-finger gripper is used to perform a rotating opening action. For square hook lifting manhole covers, a hook-shaped tool with force feedback is used to perform a vertical lifting action.

[0171] Then, the robot arm introduces a force-position hybrid mechanism to achieve robust operation when performing the cover opening action. Real-time monitoring of joint torque force sensor signals. When excessive resistance is detected (exceeding a set threshold), the robot arm will automatically pause and perform a micro-motion strategy, with small amplitude reciprocating motion, trying to "loosen" the manhole cover. If multiple attempts still fail to open, it is determined to be an abnormal manhole cover, and the position and image are recorded and reported to the background through the satellite to request expert manual intervention. After the manhole cover is opened, the robot arm can perform two operations: control the multi-parameter water quality probe to a specified depth underground for in-situ measurement, and transmit the measurement data back through satellite communication; or, control the sampling pump and sampling tube to extract water samples at a specified depth underground and store them in the internal sample box for further testing in the laboratory. Among them, if the in-situ measurement data is abnormal, the monitoring priority of the corresponding wellhead can be increased to achieve rapid response to pollution events. After the task is completed, the sampling device is recovered, and the manhole cover is resealed. The visual positioning subsystem confirms again that the manhole cover has returned to its original state.

[0172] In some more specific embodiments, the groundwater monitoring method based on the quadruped robot further comprises: online difficult case mining, labeling and collection, after each identification task is completed, the mobile quadruped robot transmits the image frames with identification confidence lower than the threshold (i.e. "difficult case" images) and the corresponding environmental data (time, location, light intensity value, weather label, etc.) to the cloud server through satellite communication. Cloud model optimization and incremental update, an online model training platform is deployed in the cloud, which continuously uses the above-mentioned "difficult case" images collected in real environment for incremental learning and continuous optimization of the model. The optimized model weight is transmitted to the mobile quadruped robot through satellite communication, and the iterative update of the model is completed.

[0173] In this optional embodiment, the wellhead position segmentation map is used for coordinate conversion, which significantly improves the accuracy of subsequent quadruped robot control. The operation strategy of the quadruped robot is flexibly adjusted according to the specific requirements of different types of wellheads, increasing the adaptability of the system. From the identification of wellhead position, coordinate conversion to the final sampling operation, all are completed automatically by the system, reducing labor costs. Not only enhances the flexibility and automation level of the system, but also guarantees the quality and reliability of the collected data. In addition, through online difficult case mining, labeling and collection, the entire robot system has the ability of continuous optimization. Using cloud model optimization and incremental update, the experience of the quadruped robot going to different regions and facing different styles of groundwater monitoring wells can be shared through the cloud, so as to continuously adapt to more extensive and complex working scenarios, greatly improving the scalability of the system.

[0174] As Figure 3As shown, the underground water monitoring device based on the quadruped robot provided by the embodiment of the application is provided with an RGB camera and a thermal infrared camera; the underground water monitoring device comprises:

[0175] The clustering and partitioning module 10 is configured to cluster and partition wellheads based on the K-means algorithm to obtain a plurality of operation partitions through the latitude and longitude coordinates of the wellheads;

[0176] The path planning module 20 is configured to plan a path for each operation partition by minimizing a path cost function to obtain a corresponding operation partition target path;

[0177] The image acquisition module 30 is configured to control the quadruped robot to move according to each operation partition target path respectively, and control the RGB camera and the thermal infrared camera of the quadruped robot to acquire images to obtain an RGB image and a thermal infrared image;

[0178] The feature fusion and extraction module 40 is configured to input the RGB image and the thermal infrared image into a multi-modal data fusion and extraction model to obtain wellhead information of underground water, wherein the wellhead information of underground water comprises a wellhead type, a wellhead position segmentation map and a wellhead state evaluation score;

[0179] The quadruped robot control module 50 is configured to control the quadruped robot to perform sampling through the wellhead type, the wellhead position segmentation map and the wellhead state evaluation score.

[0180] The underground water monitoring device based on the quadruped robot of the embodiment is used to implement the underground water monitoring method based on the quadruped robot as described above, and has the same advantages as the underground water monitoring method based on the quadruped robot compared with the prior art, which will not be described here again.

[0181] Optionally, the feature fusion and extraction module 40 is specifically configured to perform feature complementary fusion processing on the RGB image and the thermal infrared image to obtain corresponding RGB image fusion features and thermal infrared image fusion features;

[0182] An infrared image average pixel value is obtained through the thermal infrared image;

[0183] The infrared image average pixel value, a weather label and an illumination intensity value are input into a self-adaptive weight generation self-network to obtain a fusion weight coefficient;

[0184] The fusion weight coefficient comprises:

[0185] ,

[0186] wherein, is an RGB fusion weight, For thermal infrared fusion weights, AWS generates the adaptive weights from the network, and M is a vector composed of the average pixel value of the infrared image, the weather tag, and the light intensity value.

[0187] The fusion weight coefficients are used to weight and fuse the RGB image fusion features and the thermal infrared image fusion features to obtain the groundwater wellhead fusion features.

[0188] The groundwater wellhead fusion features include:

[0189] ,

[0190] in, The groundwater wellhead fusion feature, The RGB image fusion features, The thermal infrared image fusion features;

[0191] The wellhead type, wellhead location segmentation map, and wellhead status evaluation score are obtained through the groundwater wellhead fusion features.

[0192] Optionally, the feature fusion extraction module 40 is specifically used to: input the RGB image into the RGB feature extraction branch to obtain high-level features of the RGB image;

[0193] The thermal infrared image is input into the thermal infrared feature extraction branch to obtain high-level features of the thermal infrared image;

[0194] Based on the cross-modal feature mutual enhancement module, the RGB image fusion feature and the thermal infrared image fusion feature are obtained according to the high-level features of the RGB image and the high-level features of the thermal infrared image;

[0195] The RGB image fusion features include:

[0196] ,

[0197] in, The RGB image fusion feature on the c-th channel, The high-level features of the RGB image on the c-th channel, Weights for RGB image fusion features;

[0198] The thermal infrared image fusion features include:

[0199] ,

[0200] in, The thermal infrared image fusion feature on the c-th channel, a high-level feature of the thermal infrared image on a cth channel, a thermal infrared image fusion feature weight.

[0201] Optionally, the feature fusion extraction module 40 is specifically configured to: when the wellhead state evaluation score is excellent, update the fusion weight coefficient by a preset thermal infrared fusion weight to obtain an updated fusion weight coefficient;

[0202] when the wellhead state evaluation score is poor, then update the fusion weight coefficient by a preset RGB modal fusion weight to obtain an updated fusion weight coefficient;

[0203] weighting fusion of the RGB image fusion feature and the thermal infrared image fusion feature by the updated fusion weight coefficient, and updating to obtain new underground water wellhead fusion features.

[0204] Optionally, the clustering partition module 10 is specifically configured to: perform feature construction on each wellhead by the wellhead longitude and latitude coordinates to obtain a corresponding multi-dimensional feature vector;

[0205] wherein the multi-dimensional feature vector comprises:

[0206] ,

[0207] wherein, the multi-dimensional feature vector, and the wellhead longitude and latitude coordinates, a task priority, an environmental accessibility coefficient, and a feature vector weight factor;

[0208] by minimizing the intra-cluster sum of squares and iteratively solving all the multi-dimensional feature vectors, a plurality of the work partitions are obtained, wherein each work partition comprises at least one partition wellhead.

[0209] Optionally, the path planning module 20 is specifically configured to: perform path planning on each work partition to obtain a corresponding plurality of work partition candidate paths;

[0210] by performing cost calculation on each work partition candidate path, a corresponding path cost is obtained;

[0211] wherein the path cost comprises:

[0212] ,

[0213] wherein, is the adjacent well distance between the tth wellhead and the (t+1)th wellhead, is the priority of the tth wellhead, is the access order of the partition wellhead i in the path, and M is the number of partition wellheads, is the path cost;

[0214] All the operation partition candidate paths are screened by using the path cost, to obtain the operation partition target path.

[0215] Optionally, the quadruped robot control module 50 is specifically configured to: extract wellhead region contour point sets from the wellhead position segmentation map to obtain wellhead 2D coordinates;

[0216] Based on the pre-calibrated camera intrinsic parameter matrix, the wellhead 2D coordinates are converted into normalized three-dimensional coordinates;

[0217] Based on the hand-eye calibration matrix, the normalized three-dimensional coordinates are converted into quadruped robot coordinates;

[0218] The quadruped robot is controlled to sample by the wellhead type, the quadruped robot coordinates and the wellhead state evaluation score.

[0219] As shown in Figure 4 The embodiment of the present application provides an electronic device 400, which comprises a memory 410 and a processor 420; the memory 410 is used for storing a computer program; the processor 420 is used for realizing the underground water monitoring method based on the quadruped robot when the computer program is executed.

[0220] Alternatively, an electronic device 400 comprises a memory 410 and a processor 420 coupled to the memory 410; the memory 410 is configured to store a computer program; the processor 420 is configured to execute the following operations when the computer program is executed:

[0221] Based on the K-means algorithm, the wellheads are clustered and divided by wellhead longitude and latitude coordinates to obtain a plurality of operation partitions;

[0222] The path of each operation partition is planned by minimizing the path cost function, to obtain the corresponding operation partition target path;

[0223] The quadruped robot is controlled to move according to each operation partition target path respectively, and the RGB camera and the thermal infrared camera of the quadruped robot are controlled to perform image acquisition, to obtain RGB images and thermal infrared images;

[0224] input the RGB image and the thermal infrared image into a multi-modal data fusion extraction model to obtain groundwater wellhead information, wherein the groundwater wellhead information comprises wellhead type, wellhead position segmentation map and wellhead state evaluation score;

[0225] The four-legged robot is controlled through the wellhead type, the wellhead position segmentation map and the wellhead state evaluation score to perform sampling.

[0226] The embodiment of the present application provides a computer readable storage medium, and the storage medium stores a computer program.

[0227] Alternatively, a non-volatile computer readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor performs the following operations:

[0228] Based on a K-means algorithm, a plurality of operation sub-zones are obtained by clustering and dividing wellheads through wellhead longitude and latitude coordinates;

[0229] A path planning is performed on each operation sub-zone by minimizing a path cost function to obtain a corresponding operation sub-zone target path;

[0230] The four-legged robot is controlled to move according to each operation sub-zone target path respectively, and the RGB camera and the thermal infrared camera of the four-legged robot are controlled to perform image acquisition to obtain an RGB image and a thermal infrared image;

[0231] The RGB image and the thermal infrared image are input into a multi-modal data fusion extraction model to obtain groundwater wellhead information, wherein the groundwater wellhead information comprises wellhead type, wellhead position segmentation map and wellhead state evaluation score;

[0232] The four-legged robot is controlled through the wellhead type, the wellhead position segmentation map and the wellhead state evaluation score to perform sampling.

[0233] An electronic device 400, which can be a server or a client of the present application, will now be described, which is an example of a hardware device that can be applied to aspects of the present application. The electronic device 400 is intended to represent various forms of digital electronic computer devices such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device 400 can also represent various forms of mobile devices such as personal digital assistants, cellular telephones, smart phones, wearable devices, and other similar computing devices. The components shown here, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the present application described and / or claimed in this document.

[0234] The electronic device 400 includes a computing unit that can perform various appropriate actions and processes in accordance with a computer program stored in a read-only memory (ROM) or a computer program loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The computing unit, the ROM, and the RAM are connected to each other through a bus. An input / output (I / O) interface is also connected to the bus.

[0235] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware by a computer program, and the program can be stored in a computer readable storage medium. When the program is executed, the processes of the above-mentioned embodiments can be included. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM), or a random access memory (RAM), etc. In this application, the units described as separate components can be or can not be physically separated, and the components shown as units can be or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments of the present application. In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0236] Although the present application is disclosed as above, the protection scope of the present application is not limited to this. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present application, and these changes and modifications will fall within the protection scope of the present application.

Claims

1. A groundwater monitoring method based on a quadruped robot, characterized by, The four-legged robot is provided with an RGB camera and a thermal infrared camera; the underground water monitoring method comprises: Based on the K-means algorithm, the wellheads are clustered and divided into multiple operation subareas through the latitude and longitude coordinates of the wellheads; The path planning is performed on each operation subarea by minimizing the path cost function, and the corresponding operation subarea target path is obtained; The four-legged robot moves according to each operation subarea target path respectively, and the RGB camera and the thermal infrared camera of the four-legged robot are controlled to perform image acquisition, and the RGB image and the thermal infrared image are obtained; The RGB image and the thermal infrared image are input into a multi-modal data fusion extraction model to obtain underground water wellhead information, wherein the underground water wellhead information comprises wellhead type, wellhead position segmentation map and wellhead state evaluation score; The four-legged robot is controlled through the wellhead type, the wellhead position segmentation map and the wellhead state evaluation score to perform sampling.

2. The groundwater monitoring method based on a quadruped robot according to claim 1, characterized by, The RGB image and the thermal infrared image are input into a multi-modal data fusion extraction model to obtain underground water wellhead information, wherein the underground water wellhead information comprises wellhead type, wellhead position segmentation map and wellhead state evaluation score; The RGB image and the thermal infrared image are subjected to feature complementary fusion processing to obtain corresponding RGB image fusion features and thermal infrared image fusion features; An infrared image average pixel value is obtained through the thermal infrared image; The infrared image average pixel value, a weather label and an illumination intensity value are input into an adaptive weight generation self-network to obtain a fusion weight coefficient; The fusion weight coefficient comprises: ， wherein, is an RGB fusion weight, is a thermal infrared fusion weight, AWS is an adaptive weight generation network, and M is a vector composed of the average pixel value of the infrared image, the weather label, and the illumination intensity value. The RGB image fusion features and the thermal infrared image fusion features are subjected to weighted fusion by using the fusion weight coefficient to obtain underground water wellhead fusion features; The underground water wellhead fusion features comprise: ， wherein, is the groundwater wellhead fusion feature, is the RGB image fusion feature, is the thermal infrared image fusion feature; The wellhead type, the wellhead position segmentation map and the wellhead state evaluation score are obtained through the underground water wellhead fusion features.

3. The groundwater monitoring method based on a quadruped robot according to claim 2, characterized by, The RGB image and the thermal infrared image are subjected to feature complementary fusion processing to obtain corresponding RGB image fusion features and thermal infrared image fusion features, comprising: The RGB image is input into an RGB feature extraction branch to obtain RGB image high-level features; The thermal infrared image is input into a thermal infrared feature extraction branch to obtain thermal infrared image high-level features; Based on a cross-modal feature mutual enhancement module, the RGB image fusion features and the thermal infrared image fusion features are obtained according to the RGB image high-level features and the thermal infrared image high-level features; The RGB image fusion features comprise: ， wherein, is the RGB image fusion feature on the cth channel, is the RGB image high-level feature on the cth channel, is the RGB image fusion feature weight; The thermal infrared image fusion features comprise: ， wherein, is the thermal infrared image fusion feature on the cth channel, is the thermal infrared image high-level feature on the cth channel, is a thermal infrared image fusion feature weight.

4. The groundwater monitoring method based on a quadruped robot according to claim 2, characterized by, After the wellhead type, the wellhead position segmentation map and the wellhead state evaluation score are obtained through the underground water wellhead fusion features, the method further comprises: When the wellhead state evaluation score is excellent, the fusion weight coefficient is updated through a preset thermal infrared fusion weight to obtain an updated fusion weight coefficient; When the wellhead state evaluation score is poor, the fusion weight coefficient is updated through a preset RGB modal fusion weight to obtain an updated fusion weight coefficient; The RGB image fusion features and the thermal infrared image fusion features are weighted and fused by the updated fusion weight coefficients, and new underground water wellhead fusion features are obtained.

5. The groundwater monitoring method based on a quadruped robot according to claim 1, characterized by, The wellhead is clustered and divided into multiple operation sub-zones based on the wellhead longitude and latitude coordinates, including: Features of each wellhead are constructed based on the wellhead longitude and latitude coordinates, and corresponding multi-dimensional feature vectors are obtained. The multi-dimensional feature vectors include: ， wherein, is the multi-dimensional feature vector, and is the wellhead latitude and longitude coordinates, is the task priority, is the environmental accessibility coefficient, and is the feature vector weight factor; The multiple operation sub-zones are obtained by iteratively solving the minimum intra-cluster squared sum of all multi-dimensional feature vectors, wherein each operation sub-zone includes at least one sub-zone wellhead.

6. The groundwater monitoring method based on a quadruped robot according to claim 5, characterized by, The path planning of each operation sub-zone is performed by minimizing the path cost function, and the corresponding operation sub-zone target path is obtained, including: The path planning of each operation sub-zone obtains a plurality of operation sub-zone candidate paths; The path cost of each operation sub-zone candidate path is calculated, and the corresponding path cost is obtained; The path cost includes: ， wherein, is the adjacent interwell distance between the tth wellhead and the t+1th wellhead, is the priority of the tth wellhead, is the access order of the zoned wellhead i in the path, and M is the number of zoned wellheads, is the path cost; All operation sub-zone candidate paths are screened by using the path cost, and the operation sub-zone target path is obtained.

7. The groundwater monitoring method based on a quadruped robot according to claim 1, characterized by, The four-legged robot is controlled to sample by using the wellhead type, the wellhead position segmentation map and the wellhead state evaluation score, including: The wellhead region contour point set is extracted from the wellhead position segmentation map to obtain wellhead 2D coordinates; The wellhead 2D coordinates are converted into normalized three-dimensional coordinates based on the pre-calibrated camera intrinsic parameter matrix; The normalized three-dimensional coordinates are converted into four-legged robot coordinates based on the hand-eye calibration matrix; The four-legged robot is controlled to sample by using the wellhead type, the four-legged robot coordinates and the wellhead state evaluation score. 8.A groundwater monitoring device based on a quadruped robot, characterized in that, The four-legged robot is provided with an RGB camera and a thermal infrared camera; The underground water monitoring device includes: A clustering and partitioning module is configured to cluster and divide wellheads into multiple operation sub-zones based on a K-means algorithm and wellhead longitude and latitude coordinates. A path planning module is configured to perform path planning of each operation sub-zone by minimizing a path cost function, and obtain a corresponding operation sub-zone target path. An image acquisition module is configured to control the four-legged robot to move according to each operation sub-zone target path, and control the RGB camera and the thermal infrared camera of the four-legged robot to acquire images, and obtain RGB images and thermal infrared images. A feature fusion and extraction module is configured to input the RGB images and the thermal infrared images into a multi-modal data fusion and extraction model, and obtain underground water wellhead information, wherein the underground water wellhead information includes a wellhead type, a wellhead position segmentation map and a wellhead state evaluation score. A four-legged robot control module is configured to control the four-legged robot to sample by using the wellhead type, the wellhead position segmentation map and the wellhead state evaluation score.

9. An electronic device, comprising: It includes a memory and a processor. The memory is configured to store a computer program. The processor is configured to execute the computer program to implement the four-legged robot-based underground water monitoring method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium has stored thereon a computer program which, when executed by a processor, implements the groundwater monitoring method based on the quadruped robot according to any one of claims 1 to 7.

Images