A robot autonomous inspection and abnormal disposal method for high-risk environment

Abstract

This invention discloses a method for autonomous robot inspection and anomaly handling in high-risk environments, belonging to the field of autonomous robot inspection technology. This method constructs a multimodal data acquisition system to obtain laser point clouds, multi-view images, and operating environment parameters. It utilizes a multi-scale voxel pyramid to extract geometric features and fast point feature histogram features, combined with a two-stage fusion recognition model to achieve coarse identification and fine verification of abnormal areas, effectively reducing the misjudgment rate of inherent equipment. Furthermore, it constructs a graph model, using random forest posterior probability as the fidelity term and histogram intersection kernel similarity as the smoothing term, and optimizes the classification results through multi-label graph cut. Finally, it combines a multi-factor energy consumption prediction model with Bayesian linear regression to dynamically plan inspection paths and execution intervals, ensuring that the robot safely, accurately, and autonomously completes inspection and anomaly handling tasks in high-risk environments.

Description

Technical Field

[0001] This invention relates to the field of autonomous robot inspection technology, and in particular to a method for autonomous robot inspection and anomaly handling in high-risk environments. Background Technology

[0002] With the improvement of industrial automation and intelligence, inspection robots are gradually replacing manual labor in high-risk scenarios such as power, chemical, and underground utility tunnels, undertaking tasks such as equipment status monitoring and environmental anomaly identification. However, in practical applications, robots still face several technical bottlenecks during autonomous inspections, especially in anomaly identification and task scheduling.

[0003] Existing technologies, such as patent CN113902990B, propose an anomaly warning method and system for inspection robots. This method constructs a two-level convolutional neural network model to first identify "straight line segments / non-straight line segments," and then distinguish between "foreign objects" and "reference objects." This method reduces the probability of misclassifying inherent equipment such as vibration dampers and insulators as foreign objects to some extent. However, this solution still has limitations in practical applications: when new reference objects or temporary equipment not labeled in the training set appear in the inspection environment, the second identification model may still misclassify them as foreign objects, leading to frequent false stops or false alarms by the robot. Furthermore, model updates rely on cloud bandwidth; in high-risk environments with limited communication, it is impossible to optimize the identification capabilities in a timely manner, resulting in repeated misjudgments.

[0004] To address the novel foreign object recognition problem, existing patent CN115100497B employs an unsupervised contrastive learning method, judging the presence of abnormal objects based on differences in image features, thus avoiding reliance on fixed categories. However, this technical solution has extremely high requirements for environmental consistency. Image alignment deviations caused by navigation errors during multiple inspections by the robot can easily lead to distorted feature extraction, resulting in misjudging normal lighting changes or angular shifts as anomalies. Especially in high-risk scenarios with sparse textures, the failure rate of feature point matching increases, and the robustness of anomaly detection decreases significantly.

[0005] Furthermore, regarding the scheduling of verification tasks after the discovery of abnormal equipment, existing patent CN106597136B proposes an intelligent inspection method for abnormal equipment based on a substation inspection robot. This method dynamically calculates the inspection interval based on historical mileage, power consumption, and the number of abnormal equipment. However, its power estimation formula uses historical averages and does not consider dynamic factors such as real-time terrain slope, wind speed, and robot load, leading to significant deviations in power prediction. In high-risk environments, if the robot stops midway in a dangerous area due to insufficient power, it may cause safety hazards.

[0006] In summary, existing technologies for robotic inspection in high-risk environments still suffer from problems such as high misjudgment rates in foreign object identification, poor adaptability to novel foreign objects, and inaccurate power management. Therefore, there is an urgent need for a method for autonomous robotic inspection and anomaly handling in high-risk environments to address the shortcomings of existing technologies. Summary of the Invention

[0007] The purpose of this invention is to provide a method for autonomous inspection and anomaly handling of robots in high-risk environments, which solves the problems of high misjudgment rate of foreign object identification, poor adaptability to new foreign objects, and inaccurate power scheduling in existing technologies.

[0008] To achieve the above objectives, this invention provides a method for autonomous robot inspection and anomaly handling in high-risk environments, comprising the following steps: Step S1: Construct a multimodal data acquisition system to collect three types of core data: raw laser point cloud data, multi-view image data, and robot operation and environmental parameter data, forming a standardized dataset; Step S2: Construct a multi-scale voxel pyramid neighborhood structure based on the original laser point cloud data. Extract features based on eigenvalues, neighborhood geometric features, projection features, and fast point feature histogram features from each layer of the neighborhood structure, and aggregate the features from each layer to form a multi-scale feature vector. Step S3: Input the multi-scale feature vector and multi-view image data into the two-stage fusion recognition model. Detect potential abnormal regions through the coarse recognition stage, and then calculate the feature difference degree for regions with confidence levels below the threshold through the fine verification stage to obtain the initial classification result. Step S4: Construct a graph model based on the initial classification results. Use the posterior probability soft label of the random forest classifier as the fidelity term and the fast point feature histogram feature similarity calculated by the histogram intersection kernel as the smoothing term. Solve the energy function minimization through the multi-label graph cut method to obtain the optimized classification results. Step S5: Identify abnormal equipment based on the optimized classification results, construct a multi-factor energy consumption prediction model by combining robot operation and environmental parameter data, optimize the model weight coefficients through Bayesian linear regression, dynamically plan the inspection path and execution interval, and complete the abnormal handling. In step S2, the fast point feature histogram features are used to construct the multi-scale feature vector in step S3 and the smoothing term calculation in step S4. The feature difference calculation in the fine verification stage in step S3 takes the multi-scale feature vector generated in step S2 as input. The fidelity term of the graph model in step S4 takes the initial classification result generated in step S3 as input. The basis for abnormal device identification in step S5 is the optimized classification result generated in step S4.

[0009] Preferably, the three types of core data collected in step S1 specifically include: Raw laser point cloud data: includes , , 3D coordinates, intensity information Number of echoes and category tags Acquired through lidar scanning, it is used to construct the neighborhood structure of a multi-scale voxel pyramid; Multi-view image data: including RGB images Depth images Infrared thermal imaging Images captured from multiple angles are obtained through a binocular depth camera and an infrared camera, and are used for coarse recognition and fine verification of the two-stage fusion recognition model. Robot operation and environmental parameter data: including real-time battery level. mileage ,speed Topographic slope Wind speed ,temperature and GPS coordinates Data is acquired through the robot's own sensors and external environmental sensors, and used for dynamic path planning and energy consumption prediction.

[0010] Preferably, the construction of the multi-scale voxel pyramid neighborhood structure in step S2 specifically includes: S2.1 Construction A multi-scale voxel pyramid with layers, where the radius of the voxel at layer 0 is... , No. The radius of the layer voxel is ,in Each layer contains the same number voxels; S2.2 will process each point in the original point cloud. Establishing space in voxel clouds of different scales Neighbors form a multi-scale neighborhood structure. ,in Point In the In the layered voxel pyramid A set of neighboring points.

[0011] Preferably, for nodes In the Neighborhood of the layer Calculate the local covariance matrix : ; in, The first in the neighborhood One point, For the neighborhood centroid; for the covariance matrix Perform singular value decomposition to obtain eigenvalues. Then, the following geometric features are calculated: ; ; ; ; ; ; in, These are the eigenvalues ​​of the covariance matrix. Represents the largest eigenvalue. Indicates intermediate feature values. Represents the smallest eigenvalue. Represents linear characteristics. Indicating planarity, Indicates scattering characteristics, Indicates anisotropy. Represents feature entropy, It represents the change in curvature.

[0012] Preferably, step S2, which extracts fast point feature histogram features, specifically includes: For query point In the Neighborhood of the layer Calculate its simplified point feature histogram ; Fast point feature histogram features are obtained through weighted accumulation. : ; in, For the current query point, For the neighborhood The first in Neighboring points, For query point To nearby points Euclidean distance, Neighboring points The simplified point feature histogram.

[0013] Preferably, step S3 specifically includes: S3.1 Recognition Stage: A lightweight convolutional neural network is used to process RGB images in multi-view image data. Perform target detection and output the set of abnormal regions. Each region contains category confidence scores. With bounding box information; S3.2 Detailed verification stage: Confidence level of the identification results Below the threshold For the region, a fine-grained verification model based on contrastive learning is invoked to generate the multi-scale feature vector corresponding to that region. Feature vectors of the corresponding regions in the historical standard library Compare the features and calculate the feature difference: ; in, This represents the multi-scale feature vector of the current region. For the feature vector dimension, This represents the multi-scale feature vector of the corresponding region in the historical standard library. express Norm, This represents the th feature vector of the current feature vector. dimension, Represents the first eigenvector of the reference eigenvector dimension; S3.2 Obtain the initial classification result: If If it is a real foreign object, it is determined to be a real foreign object; otherwise, it is classified as an inherent equipment or environmental change. This is a dynamic threshold with an initial value of 0.35.

[0014] Preferably, step S4, which involves constructing the graphical model and solving for the minimization of the energy function, specifically includes: S4.1 Constructing a Graph Model ,in The set of vertices corresponds to each point in the point cloud dataset. ; For undirected edges between adjacent vertices, constructed based on spatial neighborhood relationships, if two points... and If the Euclidean distance in space is less than a preset threshold, an edge connection is established; The weight of the edge connecting adjacent points; S4.2 Constructing a multi-label energy function : ; in, Assign a function to the label. The weighting coefficients for balancing the fidelity and smoothing terms have a range of values. ; S4.3 Fidelity Item Posterior probability soft labels using a random forest classifier: ; in, For point Multi-scale feature vectors Points output by the random forest classifier Category tags The posterior probability, Point The assigned category label; S4.4 Smoothing Item The similarity of fast point feature histogram features calculated using the histogram intersection kernel is used as the edge weight: ; in, For indicator functions, when The time value is Otherwise ; For the edge The weight is calculated using the following formula: ; in, For point Fast feature histogram feature vectors, For point Fast feature histogram feature vectors, The dimension of the fast feature histogram is usually taken as... , For point No. histogram values ​​of dimension, For point No. histogram values ​​of dimension, Represents the kernel function for histogram intersection; S4.5 passed -expansion Algorithm for solving multi-label energy function Minimize the value to obtain the optimized classification result.

[0015] Preferably, the multi-factor energy consumption prediction model in step S5 is: ; in: The total energy consumption required for this predicted abnormal inspection is expressed in milliampere-hours (mAh). The total planned path length, in meters, is calculated from GPS coordinates in robot operation and environmental parameter data. The number of abnormal devices is calculated from the classification results optimized in step S4; For the first The terrain slope of each inspection point, in degrees, is obtained from robot operation and environmental parameter data; For the first The load weight of each inspection point is set according to the power consumption of the detection equipment. This is an environmental compensation factor, measured in milliampere-hours, based on wind speed. and temperature calculate: ; in, Based on the compensation value, and This is the environmental compensation coefficient. Reference temperature; These are the weighting coefficients to be optimized.

[0016] Preferred weighting coefficient Obtained through Bayesian linear regression optimization, specifically including: Constructing the feature matrix and historical energy consumption vector : ; in, This refers to the number of historical inspections. For the first The length of the route for each inspection. For the first The number of abnormal devices during the next inspection. For the first The actual energy consumption of each inspection; Assume the weight coefficients follow a normal prior distribution. ,in Let be the prior mean vector. The prior covariance matrix is ​​given; the observation noise follows a normal distribution. ,in To observe the noise variance; posterior distribution The mean and covariance are: ; ; in, The posterior mean vector represents the optimized weight coefficients. ; Let be the posterior covariance matrix, representing the uncertainty of the weight coefficients; The value is 0.01.

[0017] Preferably, the dynamic planning of the inspection path and execution interval in step S5 specifically includes: If the current remaining battery power Dynamic path replanning is triggered when the following conditions are met: ; in, The current remaining battery power is obtained from robot operation and environmental parameter data; Set as the safety threshold. 15%; Dynamic path replanning adopts A* Algorithm, energy cost function for: ; in, For a certain edge in the path, This is the planar projection distance of the side, in meters; This represents the elevation difference between the two endpoints, in meters. and These are cost coefficients, with values ​​of 1.0 and 0.5 respectively. The formula for dynamically adjusting the inspection interval is: ; This is the execution interval for this abnormal inspection, in minutes; The baseline interval is 30 minutes. This is the adjustment coefficient, with a value of 0.8. The charging rate, measured in milliampere-hours per minute, is obtained from robot operation and environmental parameter data. The function ensures that the increment of the interval is non-negative.

[0018] Therefore, the present invention employs the above-mentioned method for autonomous robot inspection and anomaly handling in high-risk environments, which has the following beneficial effects: (1) This invention effectively solves the problem that traditional methods easily misjudge inherent equipment such as vibration dampers and insulators as foreign objects by constructing a two-stage fusion recognition model of "coarse identification + fine verification". The coarse identification stage quickly filters suspicious areas, and the fine verification stage uses contrastive learning to compare the features of suspicious areas with historical standard libraries, and judges whether they are real foreign objects by the feature difference. This two-level verification mechanism enables the robot to accurately distinguish between inherent equipment and foreign objects, greatly reducing the frequency of false alarms and invalid shutdowns, ensuring the continuous execution of inspection tasks, and avoiding frequent human intervention and inspection interruptions caused by misjudgment.

[0019] (2) This invention does not rely on a pre-defined fixed foreign object category, but instead uses geometric features extracted from multi-scale voxel pyramids and fast point feature histogram features, combined with contrastive learning, to determine feature differences. When the robot encounters a novel foreign object it has never seen before, the system can automatically identify its abnormal attributes based on feature differences without waiting for manual updates to the category library.

[0020] (3) This invention constructs a multi-factor energy consumption prediction model, comprehensively considering environmental factors such as path length, number of abnormal devices, terrain slope, load weight, and wind speed and temperature, to accurately predict the power required for abnormal verification tasks. When the predicted power is insufficient to complete the task, the system automatically triggers dynamic path replanning, prioritizing the inspection of high-risk devices or nearby devices, and can intelligently adjust the execution interval to ensure that the robot safely returns to recharge before running out of power. This mechanism effectively avoids the safety hazards caused by the robot stopping midway in high-risk areas due to insufficient power, and improves the overall reliability and safety of inspection tasks.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is an overall flowchart of a robot autonomous inspection and anomaly handling method for high-risk environments according to the present invention. Figure 2 This is a schematic diagram of the neighborhood structure of the multi-scale voxel pyramid according to an embodiment of the present invention; Figure 3 The following are comparison charts of the classification results before and after optimization in the embodiments of the present invention, wherein (a) is a comparison chart of classification accuracy and (b) is a comparison chart of foreign object false alarm rate. Detailed Implementation

[0023] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] Example 1 This embodiment was verified in a high-risk area of ​​a chemical industrial park. The area includes typical high-risk scenarios such as storage tank area, pipeline area, and power distribution room. The terrain includes flat roads (slope 0°~5°), gentle slopes (5°~15°), and steep slopes (>15°). The area contains existing equipment such as vibration dampers, insulators, and valves, as well as foreign objects such as potentially fallen tool bags and temporary warning signs.

[0025] The configuration of the inspection robot platform is shown in Table 1: Table 1: Inspection Robot Platform Configuration Table

[0026] The robot's maximum speed is 1.2 m / s. Before the experiment, 2000 frames of point cloud data and corresponding multi-view images were collected to construct a training dataset, labeled with three categories: "foreign object," "inherent equipment," and "background." An additional 500 frames were collected as a test set.

[0027] Please see Figure 1 The specific execution process is as follows: Step S1, Multimodal Data Acquisition: The robot travels along the preset inspection route, simultaneously collecting three types of core data: 1. Raw laser point cloud data: Each frame contains approximately 100,000 points. Each point records the X, Y, and Z three-dimensional coordinates (unit: meters), intensity information, echo count, and offline labeled classification tags, which are used for subsequent multi-scale voxel pyramid neighborhood structure construction.

[0028] 2. Multi-view image data: The binocular depth camera acquires RGB images and depth images, and the infrared camera acquires infrared thermal images. At the same time, the camera is controlled by a gimbal to shoot from multiple angles, which is used for coarse recognition and fine verification of the two-stage fusion recognition model.

[0029] 3. Robot operation and environmental parameter data: Real-time recording of current battery level, mileage, speed, terrain slope, wind speed, temperature and GPS coordinates, used for subsequent dynamic path planning and energy consumption prediction.

[0030] This embodiment completed 12 full inspection tasks, with each inspection collecting approximately 1500 frames of multimodal data.

[0031] Step S2: Multi-scale voxel pyramid feature extraction: like Figure 2 A three-layer multi-scale voxel pyramid was constructed. The voxel radius of the 0th layer was 0.1m, the 1st layer was 0.2m, and the 2nd layer was 0.4m. Each voxel layer contained the same number of 32 neighboring points.

[0032] For each query point in the original point cloud, spatial K-nearest neighbors are established in voxel clouds at different scales to form a multi-scale neighborhood structure. Within each neighborhood, the local covariance matrix is ​​calculated, and eigenvalues ​​are obtained through singular value decomposition. The following geometric features are then extracted: linear features, planarity features, scattering features, anisotropic features, feature entropy, and curvature variation. Linear features: Linear features reflect the degree to which a point cloud exhibits a one-dimensional linear distribution locally. When Much larger and hour, A value close to 1 indicates that points within the neighborhood are distributed along a straight line, commonly found on the edges of pipes and wires, or on the surface of rod-shaped equipment. In this embodiment, the pipe support area... The average value is 0.82, which is significantly higher than that of the flat area (0.15), and can be used to distinguish between rod-shaped structures and planar structures.

[0033] Planarity: Planarity reflects the degree to which a point cloud exhibits a two-dimensional planar distribution locally. When and Much larger hour, A higher value indicates that the point distribution is approximately planar, commonly found on walls and equipment casings. In this embodiment, the surface area of ​​the storage tank... The average value is 0.67, while at pipe connections the value drops to 0.23, which can be used to identify the main surface of the equipment.

[0034] Scattering characteristics: Scattering characteristics reflect the degree to which a point cloud exhibits a three-dimensional discrete distribution locally. When the three characteristic values ​​are similar in magnitude, it indicates a uniform point distribution, commonly found in vegetation, cluttered areas, or complex structural regions. In this embodiment, scattered debris within the inspection area... The average value is 0.31, while the flat ground... With a value of only 0.08, it can be used to identify unstructured foreign object regions.

[0035] Anisotropy characteristics: Anisotropy characteristics comprehensively reflect the degree of directional differences in point cloud distribution. The larger the value, the stronger the extension of the point cloud in a certain direction. In this embodiment, the vibration damper area... The average value is 0.71, exhibiting a clear directional characteristic; while the spherical equipment area... The average value is only 0.35, indicating strong isotropy.

[0036] Feature entropy: Feature entropy measures the complexity of the local structure of a point cloud. When the three feature values ​​are uniformly distributed, A higher value indicates a complex point cloud structure in the neighborhood; when a certain feature value dominates, A lower value indicates a simple and regular structure. In this embodiment, the insulator string region... The average value is 0.42, and the structure is regular; however, the point cloud near the pipeline leak point is scattered due to liquid splashing. With an average value of 0.68, the increased entropy value can be used as an auxiliary indicator for anomaly detection.

[0037] Curvature variation: Curvature variation reflects the degree of curvature of a local surface. The larger the value, the greater the curvature of the region where that point is located, and the more curved the surface. In this embodiment, at the pipe bend... The average value is 0.12, which is much higher than the 0.03 for straight pipe sections, and can be used to identify the geometric feature points of the equipment.

[0038] Multi-scale feature fusion effect: Taking a vibration damper as an example, the linear features extracted at layer 0 (radius 0.1m) This reflects its slender structure; planarity features extracted in the second layer (radius 0.4m). This reflects its relative relationship with the surrounding plane. By aggregating the features at the three scales, a complete geometric description of the vibration damper is formed, making it distinguishable from similarly shaped foreign objects (such as a dropped slender tool) in the feature space.

[0039] Simultaneously, fast point feature histogram features are extracted, with a dimension of 33. The geometric features extracted from each layer are aggregated with the FPFH features to form a 512-dimensional multi-scale feature vector, which serves as the input for subsequent recognition and optimization.

[0040] Step S3: Two-stage fusion recognition: (1) Coarse recognition stage: The lightweight convolutional neural network MobileNetV3 is used to detect objects in RGB images and output a set of potential abnormal regions. Each region contains category confidence and bounding box information.

[0041] In this embodiment, the detection results for a certain pipeline area are as follows: confidence level 0.45 for the vibration damper area, confidence level 0.82 for the fallen tool bag area, and confidence level 0.51 for the insulator area.

[0042] (2) Fine-grained verification stage: Set the confidence threshold to 0.6. For regions with a confidence level below 0.6 (vibration dampers and insulators), call the fine-grained verification model based on contrastive learning to compare the multi-scale feature vectors corresponding to the region with the feature vectors of the corresponding regions in the historical standard library and calculate the feature difference. The initial value of the dynamic threshold is 0.35.

[0043] (3) Calculation results: The feature difference degree of the vibration damper area is 0.22, which is lower than the threshold and is judged as inherent equipment; the feature difference degree of the insulator area is 0.18, which is judged as inherent equipment; the confidence degree of the tool kit area is higher than the threshold and passes directly.

[0044] Initial classification results: Based on the combined results of coarse identification and fine verification, a total of 47 suspicious areas were detected in 12 inspections. After fine verification, 32 were classified as inherent equipment and 15 were determined to be real foreign objects.

[0045] Step S4: Graph model optimization: S41 constructs a graph model: the vertex set corresponds to each point in the point cloud dataset, totaling approximately 80,000 vertices. Edges are constructed based on spatial neighborhood relationships: if the spatial Euclidean distance between two points is less than 0.05m, an undirected edge is established to connect them.

[0046] S42 Energy Function Construction: A multi-label energy function is employed to balance the fidelity and smoothing terms. The fidelity term uses posterior probability soft labels from a random forest classifier, consisting of 100 decision trees with a maximum depth of 20. The smoothing term uses the similarity of fast point feature histogram features calculated using the histogram intersection kernel as edge weights.

[0047] S43 Energy Minimization: The energy function is minimized using the α-expansion algorithm to obtain the optimized classification result.

[0048] After optimization, the classification accuracy improved from the initial 87.3% to 94.6%, and the false alarm rate for foreign objects decreased by 62%. The number of false alarms for vibration dampers decreased from 5 to 0, and the number of false alarms for insulators decreased from 4 to 1.

[0049] Step S5: Multi-factor energy consumption prediction and dynamic path planning: S51. Abnormal Equipment Identification: Based on the optimized classification results, count the number of abnormal equipment. In a certain inspection in this embodiment, three abnormal devices were identified: a dropped toolkit, a temporary warning sign, and a pipeline leak.

[0050] S52. Multi-factor energy consumption prediction: Construct an energy consumption prediction model that comprehensively considers path length, number of abnormal devices, terrain slope, load weight, and environmental factors such as wind speed and temperature.

[0051] In this embodiment, the total length of the planned path is... The length is 320 meters, and the number of abnormal devices is [not specified]. There are 3 locations, with varying terrain slopes. The degrees are 2 degrees, 8 degrees, and 15 degrees respectively, with load weights. The power consumption values ​​are set to 1.0, 1.2, and 1.5 respectively. Current ambient wind speed. The speed is 3.2 m / s, and the temperature is... The temperature is 32 degrees Celsius, and the baseline temperature is 20 degrees Celsius. The weighting coefficients optimized by Bayesian linear regression are: path length coefficient. The coefficient for the number of abnormal devices is 2.8. The slope coefficient is 45.0. It is 12.5. Environmental compensation factor. The calculated value is 99.6 mA / h.

[0052] Final predicted total energy consumption It is 1556.85 mAh.

[0053] S53, Dynamic Path Replanning: The current remaining power is 1420 mAh, and the safety threshold is 15% of the predicted energy consumption, which is 233.5 mAh. The sum of the predicted energy consumption and the safety threshold is 1790.35 mAh. The current remaining power is lower than this value, triggering dynamic path replanning.

[0054] S54. Using the A* algorithm, the energy consumption cost function comprehensively considers the planar projection distance and elevation difference. After replanning, the new path length is 345 meters, an increase of 7.8%, but it avoids the 15-degree steep slope section. The total energy consumption is predicted to decrease to 1428 mAh, which meets the remaining power constraint.

[0055] S55, Dynamic Adjustment of Inspection Interval: Base interval 30 minutes, adjustment coefficient 0.8, charging rate 200 mAh / min. Based on the difference between predicted energy consumption and current battery level, the inspection interval is adjusted to 30.55 minutes to ensure the robot has sufficient time to complete charging preparation.

[0056] Example 2 To verify the technical effectiveness of this invention, three sets of comparative experiments were set up: Control group 1: Only coarse recognition model was used (no fine verification); Control group 2: Two-stage recognition was used, but graph model optimization was not used; Control group 3: Using a fixed-coefficient energy consumption prediction model (without Bayesian optimization, using empirical coefficients). The specific experimental results are shown in Table 2: Table 2: Comparison of Experimental Results

[0057] Table 2 shows that: In control group 1 (coarse identification only), a foreign object false alarm rate of 18.7% was achieved when 500 frames were tested. Among them, 23% of vibration dampers were misidentified as foreign objects, and 15% of insulators were misidentified as foreign objects. The main reason is that the coarse identification model lacks the ability to distinguish between familiar equipment and foreign objects with similar shapes.

[0058] Control Group 2 (no graph model optimization): The classification accuracy was 87.3%, but the consistency of classification in boundary regions was poor, and there was a significant jagged discontinuity in the classification results of adjacent point clouds. The false alarm rate for foreign objects was 12.8%, which was an improvement over Control Group 1 but still relatively high. The situation of Control Group 2 before and after optimization in this application is as follows: Figure 3 As shown (control group in) Figure 3 The previous version is considered as before optimization, while this invention is considered as after optimization.

[0059] Control Group 3 (Fixed-coefficient energy consumption prediction): A fixed-coefficient energy consumption prediction model was used, with weighting coefficients based on empirical values ​​and no Bayesian optimization. The actual energy consumption deviated from the prediction by 23%. During one inspection, insufficient power prediction caused the robot to run out of power and stop midway on a steep slope, requiring manual intervention. The task completion rate was 91.7%.

[0060] The invention reduces the false alarm rate for foreign objects to 3.2%, significantly decreasing the number of misjudgments of inherent equipment such as vibration dampers and insulators; it successfully identifies novel foreign objects (temporary warning signs) not found in the training set, with a feature difference of 0.42, accurately identifying them as foreign objects; the energy consumption prediction deviation is less than 5%, and no mid-journey power shortage or return failure occurred during 12 inspections; dynamic path replanning reduces actual energy consumption by an average of 12.8%, and the inspection task completion rate is 100%.

[0061] Therefore, this invention employs the aforementioned method for autonomous robot inspection and anomaly handling in high-risk environments. First, a multi-scale voxel pyramid structure is constructed, extracting geometric features based on eigenvalues ​​and fast point feature histogram features from the laser point cloud to form a multi-scale feature vector. Second, a two-stage fusion identification model of "coarse recognition + fine verification" is used to verify the feature difference of low-confidence areas based on contrastive learning, generating initial classification results. Then, a graph model is constructed, using the posterior probability output by a random forest classifier as a fidelity term and the FPFH feature similarity calculated by the histogram intersection kernel as a smoothing term, optimizing the classification results through a multi-label graph cut method. Finally, based on the optimized results, abnormal equipment is identified, and an energy consumption prediction model is constructed by combining multiple factors such as path length, number of devices, terrain slope, wind speed, and temperature. The weight coefficients are optimized through Bayesian linear regression, dynamically adjusting the inspection path and execution interval to achieve accurate and safe anomaly handling.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for autonomous robot inspection and anomaly handling in high-risk environments, characterized in that, Includes the following steps: Step S1: Construct a multimodal data acquisition system to collect three types of core data: raw laser point cloud data, multi-view image data, and robot operation and environmental parameter data, forming a standardized dataset; Step S2: Construct a multi-scale voxel pyramid neighborhood structure based on the original laser point cloud data. Extract features based on eigenvalues, neighborhood geometric features, projection features, and fast point feature histogram features from each layer of the neighborhood structure, and aggregate the features from each layer to form a multi-scale feature vector. Step S3: Input the multi-scale feature vector and multi-view image data into the two-stage fusion recognition model. Detect potential abnormal regions through the coarse recognition stage, and then calculate the feature difference degree for regions with confidence levels below the threshold through the fine verification stage to obtain the initial classification result. Step S4: Construct a graph model based on the initial classification results. Use the posterior probability soft label of the random forest classifier as the fidelity term and the fast point feature histogram feature similarity calculated by the histogram intersection kernel as the smoothing term. Solve the energy function minimization through the multi-label graph cut method to obtain the optimized classification results. Step S5: Identify abnormal equipment based on the optimized classification results, construct a multi-factor energy consumption prediction model by combining robot operation and environmental parameter data, optimize the model weight coefficients through Bayesian linear regression, dynamically plan the inspection path and execution interval, and complete the abnormal handling.

2. The method for autonomous robot inspection and anomaly handling in high-risk environments according to claim 1, characterized in that, The three types of core data collected in step S1 specifically include: Raw laser point cloud data: includes , , 3D coordinates, intensity information Number of echoes and category tags Acquired through lidar scanning, it is used to construct the neighborhood structure of a multi-scale voxel pyramid; Multi-view image data: including RGB images Depth images Infrared thermal imaging Images captured from multiple angles are obtained through a binocular depth camera and an infrared camera, and are used for coarse recognition and fine verification of the two-stage fusion recognition model. Robot operation and environmental parameter data: including real-time battery level. mileage ,speed Topographic slope Wind speed ,temperature and GPS coordinates Data is acquired through the robot's own sensors and external environmental sensors, and used for dynamic path planning and energy consumption prediction.

3. The method for autonomous robot inspection and anomaly handling in high-risk environments according to claim 2, characterized in that, Step S2, which involves constructing the neighborhood structure of the multi-scale voxel pyramid, specifically includes: S2.1 Construction A multi-scale voxel pyramid with layers, where the radius of the voxel at layer 0 is... , No. The radius of the layer voxel is ,in Each layer contains the same number voxels; S2.2 will process each point in the original point cloud. Establishing space in voxel clouds of different scales Neighbors form a multi-scale neighborhood structure. ,in Point In the In the layered voxel pyramid A set of neighboring points.

4. The method for autonomous robot inspection and anomaly handling in high-risk environments according to claim 3, characterized in that: For nodes In the Neighborhood of the layer Calculate the local covariance matrix : ； in, The first in the neighborhood One point, For the neighborhood centroid; for the covariance matrix Perform singular value decomposition to obtain eigenvalues. Then, the following geometric features are calculated: ； ； ； ； ； ； in, These are the eigenvalues ​​of the covariance matrix. Represents the largest eigenvalue. Indicates intermediate feature values. Represents the smallest eigenvalue. Represents linear characteristics. Indicating planarity, Indicates scattering characteristics, Indicates anisotropy. Represents feature entropy, It represents the change in curvature.

5. The method for autonomous robot inspection and anomaly handling in high-risk environments according to claim 4, characterized in that: Step S2, which extracts fast point feature histogram features, specifically includes: For query point In the Neighborhood of the layer Calculate its simplified point feature histogram ; Fast point feature histogram features are obtained through weighted accumulation. : ； in, For the current query point, For the neighborhood The first in Neighboring points, For query point To nearby points Euclidean distance, Neighboring points The simplified point feature histogram.

6. The method for autonomous robot inspection and anomaly handling in high-risk environments according to claim 5, characterized in that: Step S3 specifically includes: S3.1 Recognition Stage: A lightweight convolutional neural network is used to process RGB images in multi-view image data. Perform target detection and output the set of abnormal regions. Each region contains category confidence scores. With bounding box information; S3.2 Detailed verification stage: Confidence level of the identification results Below the threshold For the region, a fine-grained verification model based on contrastive learning is invoked to generate the multi-scale feature vector corresponding to that region. Feature vectors of the corresponding regions in the historical standard library Compare the features and calculate the feature difference: ； in, This represents the multi-scale feature vector of the current region. For the feature vector dimension, This represents the multi-scale feature vector of the corresponding region in the historical standard library. express Norm, Represents the th eigenvector of the current feature vector dimension, Represents the first eigenvector of the reference eigenvector dimension; S3.2 Obtain the initial classification result: If If it is a real foreign object, it is determined to be a real foreign object; otherwise, it is classified as an inherent equipment or environmental change. This is a dynamic threshold with an initial value of 0.

35.

7. The method for autonomous robot inspection and anomaly handling in high-risk environments according to claim 6, characterized in that, Step S4, which involves constructing the graph model and solving for the energy function minimization, specifically includes: S4.1 Constructing a Graph Model ,in The vertex set corresponds to each point in the point cloud dataset. ; For undirected edges between adjacent vertices, constructed based on spatial neighborhood relationships, if two points... and If the Euclidean distance in space is less than a preset threshold, then an edge connection is established; The weight of the edge connecting adjacent points; S4.2 Constructing a multi-label energy function : ； in, Assign a function to the label. The weighting coefficients for balancing the fidelity and smoothing terms have a range of values. ; S4.3 Fidelity Item Posterior probability soft labels using a random forest classifier: ； in, For point Multi-scale feature vectors Points output by the random forest classifier Category tags The posterior probability, Point The assigned category label; S4.4 Smoothing Item The similarity of fast point feature histogram features calculated using the histogram intersection kernel is used as the edge weight: ； in, For indicator functions, when The time value is Otherwise ; For the edge The weight is calculated using the following formula: ； in, For point Fast feature histogram feature vectors, For point Fast feature histogram feature vectors, The dimension of the fast feature histogram, For point No. histogram values ​​of dimension, For point No. histogram values ​​of dimension, Represents the kernel function for histogram intersection; S4.5 passed -expansion Algorithm for solving multi-label energy function Minimize the value to obtain the optimized classification result.

8. The method for autonomous robot inspection and anomaly handling in high-risk environments according to claim 7, characterized in that, The multi-factor energy consumption prediction model in step S5 is as follows: ； in: The total energy consumption required for this predicted abnormal inspection is expressed in milliampere-hours (mAh). The total planned path length, in meters, is calculated from GPS coordinates in robot operation and environmental parameter data. The number of abnormal devices; For the first The terrain slope of each inspection point, in degrees. For the first Load weight of each inspection point; This is an environmental compensation factor, measured in milliampere-hours, based on wind speed. and temperature calculate: ； in, Based on the compensation value, and This is the environmental compensation coefficient. Reference temperature; These are the weighting coefficients to be optimized.

9. A method for autonomous robot inspection and anomaly handling in high-risk environments according to claim 8, characterized in that: Weighting coefficient It was obtained through Bayesian linear regression optimization.

10. A method for autonomous robot inspection and anomaly handling in high-risk environments according to claim 9, characterized in that: Step S5, which involves dynamically planning the inspection path and execution interval, specifically includes: If the current remaining battery power Dynamic path replanning is triggered when the following conditions are met: ； in, This is the current remaining battery level. Set as the safety threshold. 15%; Dynamic path replanning adopts A* Algorithm, energy cost function for: ； in, For a certain edge in the path, This is the planar projection distance of the side, in meters; The elevation difference between the two endpoints is expressed in meters. and These are cost coefficients, with values ​​of 1.0 and 0.5 respectively. The formula for dynamically adjusting the inspection interval is: ； This is the execution interval for this abnormal inspection, in minutes; As the reference interval; This is the adjustment coefficient; This refers to the charging rate, measured in milliampere-hours per minute. The function ensures that the increment of the interval is non-negative.

Images