Risk field-based active perception closed-loop inspection method, device, system and medium

Abstract

The application discloses a kind of active perception closed loop inspection method, device, system and medium based on risk field. The method uses unified voxel multi-modal risk field data structure, reproducible update mechanism of multi-modal risk field, unified active perception scoring mechanism, quantifiable and reproducible coverage quality model, and coverage evaluation driven additional observation generation and minimum incremental incorporation mechanism. Thus, quadruped robot and its carrying sensor can automatically construct multi-modal risk field in multi-scene environment, and carry out active perception adaptive inspection under the constraint of the risk field, and generate additional observation action when coverage is insufficient, high uncertainty, abnormality is not confirmed, path is blocked or risk mutates, so as to improve the coverage rate of inspection, the reliability of abnormality discovery and the stability of task execution.

Description

Technical Field

[0001] This invention relates to the field of inspection technology for public safety / emergency / industrial infrastructure, and in particular to an active sensing closed-loop inspection method based on a risk field, a computer-readable storage medium, an active sensing closed-loop inspection device based on a risk field, and an active sensing closed-loop inspection system based on a risk field. Background Technology

[0002] In chemical plant areas, warehousing and logistics parks, power substations, subway stations and tunnels, routine safety inspections typically require continuous monitoring of equipment temperature, flammable or toxic gases, spatial obstacles and passage conditions, abnormal acoustic characteristics, and the condition of critical components.

[0003] However, the problems with the related technologies are: (1) Inspection systems mostly adopt preset routes and fixed observation actions, mainly relying on task scripts to complete point collection. This method works well when the environmental structure is stable, but when there is smoke obstruction, sudden changes in lighting, temporary obstacles, or changes in the stacking around the equipment, the fixed viewing angle often leads to the obstruction of key areas, and the observation quality of thermal imaging and visible light decreases, resulting in missed inspections or the need for repeated manual adjustment of the observation position and angle. The inspection efficiency and coverage are difficult to guarantee stably. (2) Viewpoint planning and next best viewpoint methods mostly take visual reconstruction or geometric coverage as the core objective. They usually treat risk factors as additional weights, and most of the work is based on a single or a small number of sensors. If there is a lack of a unified risk expression and cross-sensor consistency update mechanism, it is difficult to achieve stable active perception decision-making and task-level closed-loop supplementary observation during operation. (3) Industry standards and engineering guidelines mainly stipulate the sensor configuration, basic functions and safety requirements that the system should have, but do not provide a method for converting multi-source risk information into a computable risk field for inspection decision-making, nor do they specify how to quantify and verify the inspection coverage and automatically generate supplementary observation actions. Therefore, in actual deployment, the system can often only collect data and provide alarm prompts, and it is difficult to form a reusable algorithm system to adapt to the dynamic changes of different scenarios. (4) The study on gas distribution mapping has proven that mobile robots can construct gas concentration grid maps and use them for source localization, but such methods usually focus on single gas risk factors. Summary of the Invention

[0004] This invention aims to at least partially solve one of the technical problems in the related art. Therefore, The first objective of this invention is to propose an active perception closed-loop inspection method based on a risk field, enabling a quadruped robot and its onboard sensors to automatically construct a multimodal risk field in multiple scenarios and conduct active perception adaptive inspection under the constraints of this risk field. When there is insufficient coverage, high uncertainty, unconfirmed anomalies, blocked paths, or sudden changes in risk, additional observation actions are automatically generated, thereby improving the inspection coverage, the reliability of anomaly detection, and the stability of task execution.

[0005] A second objective of this invention is to provide a computer-readable storage medium.

[0006] The third objective of this invention is to propose an active sensing closed-loop inspection device based on a risk field.

[0007] The fourth objective of this invention is to propose an active sensing closed-loop inspection system based on a risk field.

[0008] To achieve the above objectives, the first aspect of this invention proposes an active perception closed-loop inspection method based on a risk field, comprising: constructing a multimodal risk field, wherein the multimodal risk field maintains occupancy probability, passage risk, thermal risk, gas risk, visibility risk, semantic category probability, and various uncertainty and timeliness fields for each voxel in a unified voxel structure, and fused and outputting a comprehensive risk; acquiring multi-sensor data and poses, and performing reproducible updates to the multimodal risk field based on the multi-sensor data and poses, wherein the reproducible updates include geometric map updates, thermal risk updates, gas risk updates, visibility risk updates, and uncertainty and timeliness updates; acquiring a set of candidate inspection actions for the multimodal risk field, and acquiring the expected information gain and coverage gain of each candidate inspection action, and constraining and penalizing them with risk costs and time costs, in order to apply risk and time costs to each candidate inspection action. The system actively senses, scores, and sorts candidate inspection actions, selecting the highest-scoring one for execution. After execution, it regenerates or filters candidate actions based on the updated risk field and continues decision-making. It obtains the coverage quality evaluation corresponding to each candidate inspection action, and determines coverage compliance when the coverage quality evaluation is not lower than a preset coverage quality lower limit and the corresponding risk uncertainty is not higher than a preset risk uncertainty threshold; otherwise, it determines coverage failure. The coverage quality evaluation consists of distance, viewing angle, sharpness, thermal imaging effectiveness, and gas sampling effectiveness. When the current candidate inspection action fails to meet coverage standards or meets the additional observation triggering conditions, an additional observation sub-task is generated and inserted using minimum incremental cost. The additional observation sub-task reuses the active sensing score and incorporates coverage compliance constraints and uncertainty reduction thresholds when selecting additional observation candidate actions.

[0009] The active perception closed-loop inspection method based on risk fields according to embodiments of the present invention employs a unified voxelized multimodal risk field data structure, a reproducible update mechanism for multimodal risk fields, a unified active perception scoring mechanism, a quantifiable and reproducible coverage quality model, and a coverage evaluation-driven supplementary observation generation and minimum increment incorporation mechanism. This enables a quadruped robot and its onboard sensors to automatically construct multimodal risk fields in multiple scenarios and conduct active perception adaptive inspections under the constraints of these risk fields. Furthermore, it automatically generates supplementary observation actions when there is insufficient coverage, high uncertainty, unconfirmed anomalies, path obstruction, or sudden risk changes. This improves inspection coverage, the reliability of anomaly detection, and the stability of task execution.

[0010] In addition, the active sensing closed-loop inspection method based on the risk field according to the above embodiments of the present invention may also have the following additional technical features: According to one embodiment of the present invention, the thermal risk update includes a joint update of thermal image projection, depth mapping, and weighted index fusion; the gas risk update includes a joint update of hysteresis compensation and spatial diffusion; the visibility risk includes a joint update of the contrast / sharpness index of the visible light image and the effective echo density of the lidar in the forward sector; the passage risk includes a joint update of the slope angle, roughness, and local step height; and the uncertainty and timeliness update includes maintaining the uncertainty of each risk based on observation quality and time decay.

[0011] According to one embodiment of the present invention, the step of actively perceiving, scoring, and ranking each candidate inspection action, selecting the candidate inspection action with the highest score for execution, and regenerating or filtering candidate actions based on the updated risk field after execution and continuing decision-making includes: selecting a region of interest, wherein the region of interest at least satisfies the following conditions: the overall risk is higher than a preset risk threshold, or the overall uncertainty is higher than a preset uncertainty threshold and is adjacent to a key inspection item, or there are unconfirmed voxels or semantic objects, or there are inspection items with insufficient coverage quality; and generating the set of candidate inspection actions based on the region of interest; and processing the candidate inspection actions. The system performs accessibility and security filtering, and obtains the expected information gain and coverage gain of each candidate inspection action that passes the accessibility and security filtering, as well as the risk cost and time cost of each candidate inspection action that passes the accessibility and security filtering. Based on the expected information gain, coverage gain, risk cost, and time cost of each candidate inspection action that passes the accessibility and security filtering, the system performs active perception scoring and ranking of each candidate inspection action that passes the accessibility and security filtering. The system selects the candidate inspection action with the highest score for execution, and after execution, it regenerates or filters candidate actions based on the updated risk field and continues to make decisions.

[0012] According to one embodiment of the present invention, the step of determining that coverage meets the standard when the coverage quality evaluation is not lower than a preset coverage quality lower limit and the corresponding risk uncertainty is not higher than a preset risk uncertainty threshold, and otherwise determining that coverage does not meet the standard, includes: constructing a set of inspection items based on each candidate inspection action, each inspection item including spatial location / semantic object binding information, sensor type, target observation distance range, allowable viewing angle range, minimum sharpness, thermal imaging dynamic range constraint, and minimum residence time for gas sampling; obtaining a coverage quality evaluation for each inspection item based on the distance index, viewing angle index, sharpness index, thermal imaging effectiveness, and gas sampling effectiveness of each inspection item; determining that the current inspection item meets the coverage standard when the coverage quality evaluation is not lower than the preset coverage quality lower limit and the corresponding risk uncertainty is not higher than a preset risk uncertainty threshold, and otherwise determining that the current inspection item does not meet the coverage standard.

[0013] According to one embodiment of the present invention, the additional observation triggering conditions include high uncertainty triggering conditions, unconfirmed anomaly triggering conditions, blocked or unreachable path triggering conditions, and risk mutation triggering conditions.

[0014] According to one embodiment of the present invention, generating additional observation subtasks and inserting them using minimum incremental cost includes: obtaining an additional observation target set, the additional observation target set including a set of inspection items with insufficient coverage, abnormal voxel clusters, or high uncertainty regions; obtaining the additional observation subtasks based on the additional observation target set, the active perception score, the coverage compliance constraint, and the uncertainty reduction threshold; and inserting the additional observation subtasks into the inspection task using the minimum incremental cost.

[0015] According to one embodiment of the present invention, the method includes: after the additional observation subtask is completed, updating the coverage quality and risk field uncertainty of the corresponding inspection item, and making active perception decisions again.

[0016] To achieve the above objectives, a computer-readable storage medium is provided in the second aspect of the present invention, which stores an active sensing closed-loop inspection program based on a risk field. When the active sensing closed-loop inspection program based on a risk field is executed by a processor, it implements the active sensing closed-loop inspection method based on a risk field according to the embodiments of the present invention described above.

[0017] According to embodiments of the present invention, a computer-readable storage medium, by executing a risk field-based active perception closed-loop inspection program stored thereon, enables a quadruped robot and its onboard sensors to automatically construct a multimodal risk field in multiple scenarios, and conduct active perception adaptive inspections under the constraints of this risk field. Furthermore, it automatically generates additional observation actions when there is insufficient coverage, high uncertainty, unconfirmed anomalies, blocked paths, or sudden changes in risk. This improves inspection coverage, the reliability of anomaly detection, and the stability of task execution.

[0018] To achieve the above objectives, the third aspect of the present invention proposes an active perception closed-loop inspection device based on a risk field, comprising: a risk field construction module for constructing a multimodal risk field, wherein the multimodal risk field maintains occupancy probability, passage risk, thermal risk, gas risk, visibility risk, semantic category probability, and various uncertainty and timeliness fields for each voxel in a unified voxel structure, and fuses and outputs a comprehensive risk; a risk field update module for acquiring multi-sensor data and poses, and performing reproducible updates to the multimodal risk field based on the multi-sensor data and poses, wherein the reproducible updates include geometric map updates, thermal risk updates, gas risk updates, visibility risk updates, and uncertainty and timeliness updates; and an active perception decision module for acquiring a set of candidate inspection actions for the multimodal risk field, acquiring the expected information gain and coverage gain of each candidate inspection action, and constraining and penalizing them with risk and time costs, so as to affect each candidate inspection action. The system actively senses, scores, and sorts actions, selecting the candidate inspection action with the highest score for execution. After execution, it regenerates or filters candidate actions based on the updated risk field and continues decision-making. A coverage quality evaluation module obtains the coverage quality evaluation corresponding to each candidate inspection action. If the coverage quality evaluation is not lower than a preset coverage quality lower limit and the corresponding risk uncertainty is not higher than a preset risk uncertainty threshold, the coverage is deemed satisfactory; otherwise, the coverage is deemed unsatisfactory. The coverage quality evaluation consists of distance, viewing angle, sharpness, thermal imaging effectiveness, and gas sampling effectiveness. An additional observation control module generates an additional observation sub-task when the current candidate inspection action's coverage is unsatisfactory or the additional observation triggering condition is met. This additional observation sub-task is inserted using minimum incremental cost. The additional observation sub-task reuses the active sensing score and incorporates coverage satisfaction constraints and uncertainty reduction thresholds when selecting candidate actions for additional observation.

[0019] The active perception closed-loop inspection device based on a risk field, according to embodiments of the present invention, employs a unified voxelized multimodal risk field data structure, a reproducible update mechanism for the multimodal risk field, a unified active perception scoring mechanism, a quantifiable and reproducible coverage quality model, and a coverage evaluation-driven supplementary observation generation and minimum increment incorporation mechanism. This enables the quadruped robot and its onboard sensors to automatically construct a multimodal risk field in multiple scenarios and conduct active perception adaptive inspections under the constraints of this risk field. Furthermore, it automatically generates supplementary observation actions when there is insufficient coverage, high uncertainty, unconfirmed anomalies, path obstruction, or sudden risk changes. This improves inspection coverage, the reliability of anomaly detection, and the stability of task execution.

[0020] To achieve the above objectives, the fourth aspect of the present invention proposes an active perception closed-loop inspection system based on a risk field, which includes a quadruped robot, sensor components, a task management and visualization terminal, and the active perception closed-loop inspection device based on a risk field as described in the above-described embodiments of the present invention.

[0021] According to embodiments of the present invention, the active perception closed-loop inspection system based on a risk field, employing the aforementioned active perception closed-loop inspection device based on a risk field, enables a quadruped robot and its onboard sensors to automatically construct a multimodal risk field in multiple scenarios. Under the constraints of this risk field, it conducts active perception adaptive inspections and automatically generates additional observation actions when there is insufficient coverage, high uncertainty, unconfirmed anomalies, path obstruction, or sudden risk changes. This improves inspection coverage, the reliability of anomaly detection, and the stability of task execution.

[0022] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the proactive sensing closed-loop inspection method based on a risk field according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the data structure of a multimodal risk field according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating an active sensing closed-loop inspection method based on a risk field according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating an active sensing closed-loop inspection method based on a risk field according to another embodiment of the present invention. Figure 5 This is a schematic diagram of the process for generating and incorporating additional observation subtasks according to yet another embodiment of the present invention; Figure 6This is a schematic diagram illustrating the addition of an observation task based on the principle of minimum modification according to an embodiment of the present invention; Figure 7 This is a block diagram of an active sensing closed-loop inspection device based on a risk field according to an embodiment of the present invention. Figure 8 This is a block diagram of a risk field-based active sensing closed-loop inspection system according to an embodiment of the present invention. Detailed Implementation

[0024] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0025] The following describes, with reference to the accompanying drawings, an active sensing closed-loop inspection method, a computer-readable storage medium, an active sensing closed-loop inspection device, and an active sensing closed-loop inspection system based on a risk field, all according to embodiments of the present invention.

[0026] Figure 1 This is a flowchart illustrating the proactive sensing closed-loop inspection method based on a risk field according to an embodiment of the present invention.

[0027] Specifically, in some embodiments of the present invention, such as Figure 1 As shown, the proactive sensing closed-loop inspection method based on the risk field includes: S101. Construct a multimodal risk field. In a unified voxel structure, the multimodal risk field maintains the occupancy probability, passage risk, thermal risk, gas risk, visibility risk, semantic category probability, and various uncertainty and timeliness fields for each voxel, and then integrates and outputs the comprehensive risk.

[0028] Specifically, in this embodiment of the invention, the workspace is discretized into a set of voxels. The side length of the voxel is Each voxel Corresponding center coordinates The multimodal risk field is stored using a voxel grid structure to balance resolution and computational efficiency, such as... Figure 2 As shown, for each voxel Maintain the following status fields: Occupancy probability / Passage risk: Occupancy probability Traffic risks ; Thermal Risk: Temperature Estimation Temperature variance Heat risk ; Gas Risk: Concentration Estimation Concentration variance Gas risks ; Visibility Risk: Visibility Score ; Semantics and Inspection Items: Semantic Category Probability Vector And the inspection status field that is bound to the inspection item; Uncertainty and Timeliness: Uncertainty for Each Type of Risk , , , and the most recent effective observation time .

[0029] To facilitate planning and comparison, multi-source risk factors are unified and normalized to... For example, in the case of thermal risk, a safety threshold is set within a certain range. With alarm threshold Then the following can be adopted: ; Gas risks can be normalized according to relative concentration limits, such as using the lower explosive limit (LEL) or occupational exposure limit (OEL). background value As a baseline: ; Visibility risk and access risk are also mapped to The calculation method will be given by subsequent definitions.

[0030] Therefore, in the above embodiments of the present invention, the multimodal state risk field is used as a unified intermediate layer so as to map the occupancy probability, passage risk, thermal risk, gas risk, visibility risk, semantic category probability, and various uncertainties and timeliness fields into the same spatial structure. This enables subsequent active sensing, path planning, and additional observation closed-loop inspection to be computable and reproducible on the same representation.

[0031] S102, acquire multi-sensor data and pose, and perform reproducible updates to the multimodal risk field based on the multi-sensor data and pose. The reproducible updates include geometric map updates, thermal risk updates, gas risk updates, visibility risk updates, and uncertainty and timeliness updates.

[0032] Furthermore, in some embodiments of the present invention, thermal risk updates include joint updates of thermal image projection, depth mapping, and weighted index fusion; gas risk updates include joint updates of hysteresis compensation and spatial diffusion; visibility risk updates include joint updates of the contrast / sharpness index of the visible light image and the effective echo density of the lidar in the forward sector; traffic risk updates include joint updates of slope angle, roughness, and local step height; and uncertainty and timeliness updates include maintaining the uncertainty of each risk based on observation quality and time decay.

[0033] Specifically, in this embodiment of the invention, the multimodal risk field is reproducibly updated in the following ways: 1) Thermal risk update (joint update of thermal image projection, depth mapping and weighted index fusion) Assuming that thermal imaging observations are at the pixel level The temperature is given at the location. This invention utilizes the intrinsic parameters of a thermal imaging camera. External reference and robot pose Pixel rays are projected into space, and surface projection with thermal imaging and depth is used, where the depth can be obtained from the lidar point cloud. The image is obtained by projecting it onto the thermal image plane in the coordinate system. For each pixel with effective depth, a point is obtained. And mapped to the corresponding voxel To form a primary voxel temperature observation .

[0034] Establish confidence weights for thermal observations This allows for interpretable effects of observation distance, incident angle, saturation / noise, and visibility degradation on the fusion results. An implementable weighting method is presented here: ; in, The distance from the sensor to the voxel. The angle between the surface normal and the observation direction. This refers to the saturation pixel ratio of the thermal image. This is a visibility risk.

[0035] Based on this weight, voxel temperatures are updated using a weighted exponential fusion method: ; ; in, , This is the fusion step size coefficient.

[0036] It should be noted that when voxels are not effectively observed for a long period of time, their uncertainty is increased through time decay, thereby driving subsequent active sensing to supplement observations.

[0037] 2) Gas risk update (joint update of hysteresis compensation and spatial diffusion) Because gas sensors suffer from response / recovery hysteresis and strong locality, this embodiment of the invention employs a joint update of hysteresis compensation and spatial diffusion for gas risk assessment. It is assumed that the gas sensor, at discrete moments... Output raw readings The sampling period is Δt. The sensor response is modeled as a first-order inertial element, and the current concentration estimate is approximately recovered using a first-order inverse filter. : ; in, This is the sensor time constant. To suppress noise, [the following is used]... Perform a first-order smoothing. , Pick .

[0038] Gas readings after hysteresis compensation Updated to the ground Figure 1 The voxel structure is consistent. Let the gas sensor be located at position [value missing] in the world coordinate system. Regarding its neighborhood radius Endogenous Spatial diffusion update is performed, with weights using a Gaussian kernel: ; in, The diffusion scale.

[0039] Voxel concentration fusion uses: ; ; in, , To update the step size coefficient.

[0040] It should be noted that when the sensor is in a walking state, a dwell validity constraint can be added; only if the dwell time... Or robot speed ≤ At that time, the gas observation was considered a highly reliable update.

[0041] 3) Visibility risk update (joint update of the contrast / sharpness index of the visible light image and the effective echo density of the lidar in the forward sector) and traffic risk update (joint update of slope angle, roughness and local step height) Understandably, visibility risk is used to characterize imaging and laser degradation caused by smoke, dust, or strong / weak light. Specifically, this invention calculates the contrast / sharpness index of the visible light image in each cycle. and the effective echo density of the lidar in the forward sector. Then, the two are normalized into a visibility score: ; in, As a calibration reference, Visibility risk And it is written into the voxels corresponding to the robot's front field of view in an exponential fusion manner.

[0042] Furthermore, this invention also employs a deep learning smoke recognition network to output the smoke probability. and order To enhance robustness to complex smoke morphologies.

[0043] Understandably, traffic risk is used to characterize the impact of factors such as slope, steps, narrowness, and loose ground on quadrupedal gait. Specifically, this invention calculates the slope angle by fitting a ground plane to a point cloud within a local neighborhood. Roughness With local step height And normalized to: ; in, Given by robot capability parameters.

[0044] It should be noted that the passage risk and occupancy probability are used together to generate the planning cost map. Occupied voxels are not passable, while non-occupied voxels are assigned a cost based on the risk.

[0045] 4) Semantic risk and inspection item binding update For scenarios with equipment ledgers or fixed inspection items, this invention incorporates semantic objects, spatial locations, and inspection status into the risk field, enabling computational coverage evaluation and supplementary observation closed-loop inspections. Specifically, it performs instance segmentation on visible light / thermal images and outputs masks and confidence levels. Furthermore, by combining depth information, the target is projected into space to form a set of object centers or surface points, and updated within the risk field voxels. For example, a simple Bayesian update can be used: ; The inspection items are bound to semantic objects or spatial points. Each inspection item records its required sensor type, quality threshold, and verification rules, and is used in subsequent coverage verification.

[0046] 5) Uncertainty and Timeliness Update and Output Comprehensive Risk (Uncertainty of each risk is maintained based on observation quality and time decay) To drive proactive sensing, this invention maintains the uncertainty of each risk attribute and takes into account observation quality and time decay.

[0047] Taking thermal risk as an example, it is defined as follows: ; In addition, gas, visibility, and traffic risks can also be addressed in similar ways.

[0048] This invention can further define the comprehensive uncertainty: ; And output the overall risk value For planning purposes: ; in, The weights are 1, For semantic risks, the weights can be adaptively adjusted by scenario configuration or task type. This comprehensive risk, together with the occupancy / passage, forms a risk constraint cost graph, which is shared by proactive perception and path planning.

[0049] Therefore, in the above embodiments of the present invention, the multimodal risk field is reproducibly updated by using multi-sensor data and pose, so that the multimodal risk field can adapt to the problems of occlusion and dynamic changes, thereby improving the reliability of the multimodal risk field.

[0050] S103, obtain the candidate inspection action set of the multimodal risk field, and obtain the expected information gain and coverage gain of each candidate inspection action. Constrain and penalize them with risk cost and time cost to actively perceive, score and rank each candidate inspection action, select the candidate inspection action with the highest score to execute, and regenerate or filter candidate actions based on the updated risk field after execution and continue decision-making.

[0051] Furthermore, in some embodiments of the present invention, such as Figure 3 As shown, each candidate inspection action is actively perceived, scored, and ranked, and the candidate inspection action with the highest score is selected for execution. After execution, candidate actions are regenerated or filtered based on the updated risk field, and further decision-making is carried out, including: S201. Select the region of interest. The region of interest must meet at least the following conditions: the overall risk is higher than the preset risk threshold, or the overall uncertainty is higher than the preset uncertainty threshold and is adjacent to the key inspection item, or there is an abnormal unconfirmed voxel or semantic object, or there is an inspection item with insufficient coverage quality. Generate a set of candidate inspection actions based on the region of interest.

[0052] Specifically, in this embodiment of the invention, the region of interest is first determined. It must meet at least one of the following: 1) Comprehensive risk Above the threshold 2) Overall uncertainty Above the threshold And adjacent to key inspection items; 3) there are abnormal unconfirmed voxels or semantic objects; 4) there are inspection items with insufficient coverage quality.

[0053] Then, focusing on the area of ​​interest By generating candidate observation poses within the reachable space, including position, orientation, and dwell time, for example, by uniformly sampling on a radial ring around the target voxel / object. The system generates several candidate points and samples several orientations for each candidate point to ensure that the sensor's field of view covers the target area. In addition, for gas sampling actions, candidate dwell times can be generated simultaneously.

[0054] S202, perform accessibility and security filtering on the candidate inspection action set, and obtain the expected information gain and coverage gain of each candidate inspection action that passes the accessibility and security filtering, as well as the risk cost and time cost of each candidate inspection action that passes the accessibility and security filtering.

[0055] Specifically, in this embodiment of the invention, for each candidate action in the candidate inspection action set... Accessibility and safety filtering can be performed. For example, a local planner can be used to verify accessibility and apply hard constraints. If there are inaccessible voxels on the path or the path risk score exceeds the budget, the action is rejected. Here, the risk budget can be defined as the upper limit of the path risk score. Risk of the largest monomer For example, if all voxels on the path are required to satisfy... .

[0056] Furthermore, for the retained candidate actions (i.e., candidate inspection actions filtered by accessibility and security), the expected information gain and coverage gain are calculated based on the predicted field of view coverage, and the risk cost and time cost are calculated simultaneously.

[0057] For example, the set of voxels within the candidate action's field of view can be obtained through ray projection / visibility determination. The observed probability is calculated by combining occupancy and visibility estimation. The expected information gain is defined in one implementation as: ; Coverage gain is used to measure the improvement of the inspection item coverage quality by this action, and is defined as follows: ; in, To determine the quality score for the current inspection item, To predict the coverage quality after the action is performed.

[0058] Risk and cost considerations for reaching the path The risk exposure and time energy consumption are defined as follows: ; in, To determine the voxel where the final observation pose of the action is located, Risk weighting at the endpoint. Time cost. Calculated from path length and dwell time.

[0059] S203, based on the expected information gain, coverage gain, risk cost, and time cost of each candidate inspection action that passes the accessibility and security filtering, actively perceive, score, and rank them.

[0060] Specifically, in this embodiment of the invention, the active perception score for each candidate inspection action that passes the accessibility and security filtering is obtained using the following formula: ; in, For proactive perception and scoring.

[0061] S204: Select the candidate inspection action with the highest score to execute, and after execution, regenerate or filter candidate actions based on the updated risk field and continue to make decisions.

[0062] Specifically, in this embodiment of the invention, the selected For the largest number of actions performed, when multiple action scores are close, priority is given to those for incomplete checks. Larger movements are needed to ensure task convergence.

[0063] Therefore, by proactively perceiving and making decisions to reduce uncertainty in the risk field, improve the quality of inspection coverage, and enhance anomaly identification, while being subject to risk budget and compliance constraints, a set of candidate observation actions is generated in each decision cycle based on the current risk field and coverage status. It also selects the next action through a unified evaluation function, achieving reproducible closed-loop decision-making.

[0064] S104. Obtain the coverage quality evaluation corresponding to each candidate inspection action. If the coverage quality evaluation is not lower than the preset coverage quality lower limit and the corresponding risk uncertainty is not higher than the preset risk uncertainty threshold, the coverage is judged to meet the standard. Otherwise, the coverage is judged to fail to meet the standard. The coverage quality evaluation consists of distance, viewing angle, clarity, thermal image effectiveness and gas sampling effectiveness.

[0065] Furthermore, in some embodiments of the present invention, such as Figure 4 As shown, when the coverage quality assessment is not lower than the preset coverage quality lower limit and the corresponding risk uncertainty is not higher than the preset risk uncertainty threshold, the coverage is judged to be up to standard; otherwise, the coverage is judged to be down to standard, including: S301. Based on the inspection task log, construct a set of inspection items. Each inspection item includes spatial location / semantic object binding information, sensor type, target observation distance range, allowable viewing angle range, minimum resolution, thermal imaging dynamic range constraints, and minimum residence time for gas sampling.

[0066] Specifically, in this embodiment of the invention, a set of check items is constructed during task initialization. Each check item includes: spatial location / semantic object binding information, required sensor type, and target observation range. Permissible viewing angle range Minimum resolution Thermal imaging dynamic range constraints, minimum residence time for gas sampling And so on, and maintain a coverage status table to record the current coverage quality score for each inspection item. Whether it meets the standard, whether it needs to be reviewed, the number of attempts and the most recent valid observation time.

[0067] S302, based on the distance index, viewing angle index, sharpness index, thermal imaging effectiveness, and gas sampling effectiveness of each inspection item, obtain the coverage quality evaluation for each inspection item.

[0068] Specifically, in this embodiment of the invention, the coverage quality evaluation output is... It consists of several computable sub-indices: ; Sub-indicators may include: distance indicators Perspective indicators Clarity index Thermal imaging effectiveness Gas sampling effectiveness .

[0069] Distance index: when the observation distance In High scores are awarded for longer intervals; otherwise, scores decay linearly. , The desired distance; Viewpoint index: the angle between the target normal and the line of sight. calculate, Or when Select 1 otherwise select 0; Sharpness metric: Laplace variance express, ; Thermal image validity: by saturation ratio express, ; Gas sampling validity: based on residence time With velocity constraints, If the speed exceeds the limit, then Reduce its ranking.

[0070] S303: When the coverage quality evaluation is not lower than the preset coverage quality lower limit and the corresponding risk uncertainty is not higher than the preset risk uncertainty threshold, the current inspection item is judged to meet the coverage standard; otherwise, the current inspection item is judged to fail to meet the coverage standard.

[0071] Specifically, in this embodiment of the present invention, when Furthermore, when the corresponding risk uncertainty is below the threshold, the inspection item coverage is deemed to be up to standard; otherwise, it is marked as "insufficient coverage requiring additional observation," thereby avoiding false coverage where the quadruped robot reaches the point but the target is occluded, the image is blurred, or the thermal image is invalid.

[0072] Furthermore, in some embodiments of the present invention, risk-constrained path planning is implemented in the following ways: By merging the occupation map and risk field into a planning cost map, and assigning costs to passable voxels: ; in, Based on the cost, As a weight, the cost of an impassable voxel is infinite.

[0073] Optionally, the global planning uses a path planning algorithm, while the local planning can use a dynamic window or trajectory optimization method. The planning output must satisfy hard constraints. and path risk points This helps control the risks of heat exposure, toxic gas exposure, and passage.

[0074] Additionally, when the environment is unknown or the map changes significantly, frontier points can be introduced to expand coverage. It should be noted that the scoring of frontier candidate points not only considers the map increment brought by exploration, but also the coverage benefits and risk constraints associated with the inspection items. This ensures that the robot prioritizes serving the inspection targets while expanding the traversable map, and avoids blindly exploring high-risk areas of smoke / toxic gas.

[0075] Therefore, by simultaneously considering geometrical accessibility, comprehensive risk constraints, and inspection item coverage constraints in the inspection route planning, and by providing an evaluation mechanism for the reproducibility of coverage quality, reaching a point is not equivalent to effective coverage.

[0076] S105 When the coverage of the current candidate inspection action is not up to standard or the additional observation triggering condition is met, an additional observation sub-task is generated, and the additional observation sub-task is incorporated into the current inspection action sequence using the minimum incremental cost. The additional observation sub-task reuses the active perception score, and coverage compliance constraints and uncertainty reduction thresholds are added when selecting additional observation candidate actions.

[0077] Optionally, in some embodiments of the present invention, additional observation triggering conditions include high uncertainty triggering conditions, unconfirmed anomaly triggering conditions, blocked or unreachable path triggering conditions, and risk mutation triggering conditions.

[0078] Specifically, in this embodiment of the invention, the triggering conditions for generating additional observation subtasks include, but are not limited to: 1) Insufficient coverage quality triggering (i.e., the current inspection item coverage is not up to standard): for any inspection item If after one observation If the key sub-indicator falls below the lower limit, additional observations will be triggered; 2) High uncertainty trigger: Voxel set in the vicinity of key equipment. exist And the duration exceeds 1) Trigger additional observations to reduce misjudgments; 2) Trigger for unconfirmed anomalies: When anomalies are detected but the confidence level is insufficient, additional observations are triggered for multi-view / stationary confirmation; 3) Trigger for unconfirmed anomalies: When anomalies are detected but the confidence level is insufficient, additional observations are triggered for multi-view / stationary confirmation; 4) Trigger for blocked / inaccessible paths: When the planned path is blocked by temporary obstacles or the risk exceeds the limit, resulting in inaccessibility, additional observations are triggered for the affected check items to reselect points or change routes; 5) Trigger for sudden risk changes: If the comprehensive risk of a certain area is increased, additional observations are triggered to reduce misjudgments; 6) Trigger for sudden risk changes: If the comprehensive risk of a certain area is increased, additional observations are triggered to reduce misjudgments; 7) Trigger for unconfirmed anomalies: When anomalies are detected but the confidence level is insufficient, additional observations are triggered for multi-view / stationary confirmation; 8) Trigger for sudden risk changes: When the comprehensive risk of a certain area is increased, additional observations are triggered to reduce misjudgments; 9) Trigger for sudden risk changes: When the comprehensive risk of a certain area is increased, additional observations are triggered to reduce misjudgments; 10) Trigger for unconfirmed anomalies: When anomalies are detected but the confidence level is insufficient, additional observations are triggered for multi-view / stationary confirmation; 11) Trigger for sudden risk changes: When the comprehensive risk of a certain area is increased, additional observations are triggered for multi-view / stationary confirmation; 12) Trigger for sudden risk changes: When the comprehensive risk of a certain area is increased, additional observations are triggered for multi-view / stationary confirmation; 13) Trigger for sudden risk changes: When the comprehensive risk of a certain area is increased, additional observations are triggered for multi-view / stationary confirmation; 14) Trigger for sudden risk changes: When the comprehensive risk of a certain area The rise exceeded the threshold in a short period of time. The system can trigger additional observations for rapid verification and output alarm evidence.

[0079] To ensure the task can be terminated, in this embodiment of the invention, a maximum number of additional observation attempts can be set for each inspection item. The maximum additional observation time budget is set, and if the budget is exceeded, the process will proceed to "alarm / manual intervention / record as unreachable".

[0080] More specifically, in some embodiments of the present invention, such as Figure 5 As shown, additional observation subtasks are generated and incorporated into the current inspection action sequence using minimum incremental cost, including: S401, Obtain the additional observation target set, which includes the set of check items with insufficient coverage, anomalous voxel clusters, or regions with high uncertainty.

[0081] Specifically, in this embodiment of the invention, when it is determined that the coverage of the current candidate inspection action is insufficient or that the additional observation triggering condition is met, the set of additional observation targets is first determined. Add observation target set This includes insufficiently covered checklists, anomalous voxel clusters, or regions of high uncertainty.

[0082] S402, based on the set of additional observation targets, and based on the active perception score, coverage compliance constraints, and uncertainty reduction threshold, obtain additional observation sub-tasks.

[0083] Specifically, in this embodiment of the invention, the additional set of observation targets is considered. Following the aforementioned steps S201-S203, a set of candidate observation actions is generated and scored. Coverage compliance constraints and uncertainty reduction thresholds are incorporated into the scoring process; that is, candidate actions must meet the predicted coverage quality requirements. Or the amount of reduction in prediction uncertainty If the threshold is reached, otherwise even Higher values ​​are not selected, thus ensuring that additional observations are not redundant and invalid acquisitions, but rather viewpoints and configurations selected to meet the criteria.

[0084] S403, using the minimum incremental cost, incorporates the additional observation subtask into the current inspection action sequence.

[0085] Specifically, in this embodiment of the invention, it is assumed that the remaining path of the main inspection task is a waypoint sequence. The current pose is Therefore, the additional observation pose is obtained after the additional observation subtask is determined. And the cost of the path planning to reach it, and insert using the minimum incremental cost: for each insertable location Calculate the insertion increment:

[0086] in, To ensure a unified cost function, it must include at least time / path length and risk integral.

[0087] Furthermore, such as Figure 6 As shown, select Minimum insertion point that satisfies the risk budget This serves as the insertion point for the minimum modified path among the remaining paths of the main inspection task. Additional observation subtasks are then inserted into the main inspection task from this minimum modified path insertion point to generate additional observation subtasks (i.e., the new path) along the minimum modified path. ).

[0088] It should be noted that when there is no insertion point that meets the risk budget, the following equivalent strategies can be adopted: 1) Add the additional observation subtask to the end of the remaining path of the main inspection task as an independent subtask; 2) Trigger an alarm and mark the inspection item as "unsafe to add observation", and output voice.

[0089] Therefore, unlike conventional fixed routes and fixed actions, the additional observation subtask is an adaptive subtask driven by the need for coverage quality, uncertainty, or anomaly confirmation during the execution of the main inspection task. Its generation, insertion, and execution are all controlled by risk constraints, and the principle of minimum modification is used to reduce the disturbance to the main inspection task. Moreover, through coverage status evaluation and additional observation triggering mechanism, the inspection no longer depends on the occasional visibility of fixed routes. It can automatically select new observation angles and complete additional observations when there is occlusion, changes in lighting, or the appearance of temporary obstacles, thereby improving the effective observation ratio of key points and reducing the probability of missed detections. Furthermore, through the active perception and additional observation mechanism, the observation position and acquisition configuration can be automatically adjusted when the observation quality deteriorates, thereby reducing the number of remote personnel interventions and repeated re-inspections.

[0090] Furthermore, in some embodiments of the present invention, such as Figure 5 As shown, the method includes: S404 After the additional observation subtask is completed, update the coverage quality and risk field uncertainty of the corresponding inspection items, and make active perception decisions again.

[0091] Specifically, after the additional observation sub-task is completed, the coverage quality and risk field uncertainty of the corresponding inspection items are updated; if the target is met, the process returns to the main inspection task and continues; if the target is still not met but the budget is not exceeded... Additional observation actions can be generated again; if the limit is exceeded, an alarm / manual review process will be initiated.

[0092] Thus, through a closed loop of active perception scoring, additional observation sub-task execution, and risk field state writing, the robot can adaptively adjust its observation position and data acquisition configuration when faced with occlusion, smoke, obstacles, and sudden risk changes, rather than mechanically executing a fixed script.

[0093] Specifically, in the above embodiments of the present invention, by unifying the risk field and uncertainty intermediate layer, proactive decision-making perception, execution path planning, coverage quality evaluation, and closed-loop inspection are linked to the same spatial representation, thereby improving inspection coverage and coverage stability, enhancing the reliability of anomaly identification and reducing false alarms, realizing the spatial expression of thermal and gas risks and supporting decision-making closed loops, reducing the frequency of manual intervention and improving task adaptability, while enhancing multi-scenario adaptability and engineering deployment capabilities.

[0094] In summary, the active perception closed-loop inspection method based on risk fields according to embodiments of the present invention employs a unified voxelized multimodal risk field data structure, a reproducible update mechanism for the multimodal risk field, a unified active perception scoring mechanism, a quantifiable and reproducible coverage quality model, and a coverage evaluation-driven supplementary observation generation and minimum increment incorporation mechanism. This enables the quadruped robot and its onboard sensors to automatically construct a multimodal risk field in multiple scenarios and conduct active perception adaptive inspections under the constraints of this risk field. Furthermore, it automatically generates supplementary observation actions when there is insufficient coverage, high uncertainty, unconfirmed anomalies, path obstruction, or sudden risk changes. This improves inspection coverage, the reliability of anomaly detection, and the stability of task execution.

[0095] Based on the aforementioned risk field-based active perception closed-loop inspection method of the present invention, the present invention proposes a computer-readable storage medium storing a risk field-based active perception closed-loop inspection program thereon. When the risk field-based active perception closed-loop inspection program is executed by a processor, it implements the risk field-based active perception closed-loop inspection method of the present invention as described above.

[0096] It should be understood that the specific implementation of the computer-readable storage medium in the embodiments of the present invention can be found in the specific implementation of the active sensing closed-loop inspection method based on the risk field in the foregoing embodiments of the present invention. To reduce redundancy, it will not be repeated here.

[0097] In summary, according to embodiments of the present invention, a computer-readable storage medium, by executing a risk field-based active perception closed-loop inspection program stored thereon, enables a quadruped robot and its onboard sensors to automatically construct a multimodal risk field in multiple scenarios, and conduct active perception adaptive inspections under the constraints of this risk field. Furthermore, it automatically generates additional observation actions when there is insufficient coverage, high uncertainty, unconfirmed anomalies, path obstruction, or sudden risk changes. This improves inspection coverage, the reliability of anomaly detection, and the stability of task execution.

[0098] Figure 7 This is a block diagram of a risk field-based active sensing closed-loop inspection device according to an embodiment of the present invention.

[0099] Specifically, in some embodiments of the present invention, such as Figure 7 As shown, the active perception closed-loop inspection device 1000 based on the risk field includes: a risk field construction module 100, a risk field update module 200, an active perception decision module 300, a coverage quality evaluation module 400, and an additional observation and control module 500.

[0100] The risk field construction module 100 is used to construct a multimodal risk field. In a unified voxel structure, the multimodal risk field maintains occupancy probability, passage risk, thermal risk, gas risk, visibility risk, semantic category probability, and various uncertainty and timeliness fields for each voxel, and then fuses and outputs a comprehensive risk. The risk field update module 200 is used to acquire multi-sensor data and poses, and perform reproducible updates to the multimodal risk field based on the multi-sensor data and poses. Reproducible updates include geometric map updates, thermal risk updates, gas risk updates, visibility risk updates, and uncertainty and timeliness updates. The active perception and decision-making module 300 is used to acquire a set of candidate inspection actions for the multimodal risk field, and to acquire the expected information gain and coverage gain of each candidate inspection action. Constraints and penalties are applied using risk and time costs to actively perceive, score, and rank each candidate inspection action, and then select... The highest-scoring candidate inspection action is executed, and after execution, candidate actions are regenerated or screened based on the updated risk field for further decision-making. The coverage quality evaluation module 400 is used to obtain the coverage quality evaluation corresponding to each candidate inspection action. When the coverage quality evaluation is not lower than the preset coverage quality lower limit and the corresponding risk uncertainty is not higher than the preset risk uncertainty threshold, the coverage is judged to be up to standard; otherwise, the coverage is judged to be down to standard. The coverage quality evaluation consists of distance, viewing angle, clarity, thermal image effectiveness, and gas sampling effectiveness. The additional observation control module 500 is used to generate additional observation sub-tasks when the coverage of the current candidate inspection action is down to standard or the additional observation triggering conditions are met. The additional observation sub-tasks are inserted using the minimum incremental cost. The additional observation sub-tasks reuse the active perception score and add coverage compliance constraints and uncertainty reduction thresholds when selecting additional observation candidate actions.

[0101] Furthermore, in some embodiments of the present invention, thermal risk updates include joint updates of thermal image projection, depth mapping, and weighted index fusion; gas risk updates include joint updates of hysteresis compensation and spatial diffusion; visibility risk updates include joint updates of the contrast / sharpness index of the visible light image and the effective echo density of the lidar in the forward sector; traffic risk updates include joint updates of slope angle, roughness, and local step height; and uncertainty and timeliness updates include maintaining the uncertainty of each risk based on observation quality and time decay.

[0102] Furthermore, in some embodiments of the present invention, the proactive perception and decision-making module 300 is further configured to: select a region of interest, wherein the region of interest at least satisfies the following conditions: the overall risk is higher than a preset risk threshold, or the overall uncertainty is higher than a preset uncertainty threshold and is adjacent to a key inspection item, or there is an unconfirmed voxel or semantic object, or there is an inspection item with insufficient coverage quality; and generate a set of candidate inspection actions based on the region of interest; perform accessibility and safety filtering on the set of candidate inspection actions, and obtain the expected information gain and coverage gain of each candidate inspection action that passes the accessibility and safety filtering, as well as the risk cost and time cost of each candidate inspection action that passes the accessibility and safety filtering; perform proactive perception scoring and ranking on each candidate inspection action that passes the accessibility and safety filtering based on the expected information gain, coverage gain, risk cost, and time cost; select the candidate inspection action with the highest score for execution, and after execution, regenerate or filter candidate actions based on the updated risk field and continue decision-making.

[0103] Furthermore, in some embodiments of the present invention, the coverage quality evaluation module 400 is also used to: construct a set of inspection items based on the inspection task log, each inspection item including spatial location / semantic object binding information, sensor type, target observation distance range, allowable viewing angle range, minimum sharpness, thermal imaging dynamic range constraint, and minimum residence time for gas sampling; obtain a coverage quality evaluation for each inspection item based on the distance index, viewing angle index, sharpness index, thermal imaging validity, and gas sampling validity of each inspection item; when the coverage quality evaluation is not lower than a preset coverage quality lower limit and the corresponding risk uncertainty is not higher than a preset risk uncertainty threshold, determine that the current inspection item coverage meets the standard; otherwise, determine that the current inspection item coverage does not meet the standard.

[0104] Furthermore, in some embodiments of the present invention, additional observation triggering conditions include high uncertainty triggering conditions, unconfirmed anomaly triggering conditions, blocked or unreachable path triggering conditions, and risk mutation triggering conditions.

[0105] Furthermore, in some embodiments of the present invention, the additional observation control module 500 is also used to: obtain an additional observation target set, which includes a set of inspection items with insufficient coverage, abnormal voxel clusters, or high uncertainty regions; obtain additional observation sub-tasks based on the additional observation target set, active perception score, coverage compliance constraints, and uncertainty reduction threshold; and insert additional observation sub-tasks into the inspection task using minimum incremental cost.

[0106] Furthermore, in some embodiments of the present invention, the additional observation control module 500 is also used to update the coverage quality and risk field uncertainty of the corresponding inspection item after the additional observation subtask is completed, and to make active perception decisions again.

[0107] It should be understood that the specific implementation of the risk field-based active sensing closed-loop inspection device 1000 in this embodiment of the invention corresponds one-to-one with the specific implementation of the risk field-based active sensing closed-loop inspection method in the foregoing embodiments of the invention. To reduce redundancy, it will not be described again here.

[0108] In summary, the active perception closed-loop inspection device based on a risk field according to embodiments of the present invention employs a unified voxelized multimodal risk field data structure, a reproducible update mechanism for the multimodal risk field, a unified active perception scoring mechanism, a quantifiable and reproducible coverage quality model, and a coverage evaluation-driven supplementary observation generation and minimum increment incorporation mechanism. This enables the quadruped robot and its onboard sensors to automatically construct a multimodal risk field in multiple scenarios and conduct active perception adaptive inspections under the constraints of this risk field. Furthermore, it automatically generates supplementary observation actions when there is insufficient coverage, high uncertainty, unconfirmed anomalies, path obstruction, or sudden risk changes. This improves inspection coverage, the reliability of anomaly detection, and the stability of task execution.

[0109] Figure 8 This is a block diagram of a risk field-based active sensing closed-loop inspection system according to an embodiment of the present invention.

[0110] Specifically, in some embodiments of the present invention, such as Figure 8 As shown, the risk field-based active perception closed-loop inspection system 5000 includes a quadruped robot 2000, a sensor assembly 3000, a task management and visualization terminal 4000, and the risk field-based active perception closed-loop inspection device 1000 described in the above embodiment of the present invention.

[0111] It should be noted that, in this embodiment of the invention, the quadruped robot 2000 includes a motion controller, a chassis status monitoring module, and a safety protection module. It is capable of walking, stopping, and adjusting its posture according to a trajectory or discrete waypoints, and enters a safe state upon falling, getting stuck, or triggering an emergency stop. The sensor assembly 3000 includes a 3D LiDAR, an infrared thermal imaging camera, a visible light camera, a gas detection sensor, an IMU, an acoustic sensor, and a temperature and humidity sensor. The risk field-based active perception closed-loop inspection device 1000 is used to perform multi-sensor time synchronization, extrinsic parameter calibration management, pose estimation and 3D mapping, risk field construction and updating, active perception decision-making, risk-constrained path planning, coverage quality evaluation, generation and insertion of additional observation sub-tasks, and task execution status management. The task management and visualization terminal 4000 is used to issue inspection tasks and inspection item ledgers, configure risk weights and threshold strategies, display risk fields and coverage status, and record alarms and inspection reports.

[0112] Furthermore, to ensure the fusion and reproducibility of multiple sensors, this invention employs a unified timestamp and coordinate system management. For example, it defines a world coordinate system {W}, a robot base coordinate system {B}, and coordinate systems for each sensor, such as lidar {L}, visible light {C}, thermal imaging {T}, and gas {G}. Extrinsic parameters are stored in the form of rigid body transformation matrices, such as... , , , The system is calibrated during system initialization or periodic maintenance. The system interpolates and aligns sensor data using a unified time axis, ensuring that all sensor data within the same fusion cycle correspond to the same robot pose. .

[0113] Specifically, the system data flow executes in a closed loop within each fusion cycle Δt as follows: sensor acquisition → time synchronization and extrinsic parameter transformation → pose estimation and 3D geometric map update → multimodal risk field update → coverage quality and uncertainty assessment → generation of candidate observation actions and proactive perception decision-making → path planning and trajectory generation under risk constraints → execution of actions and writing back new observations to update the risk field and coverage status. The task ends when the coverage termination condition is met or the safety / time limit is reached, and the inspection records, alarm events, and coverage evidence are uploaded to the task management terminal.

[0114] It should be understood that the specific implementation of the risk field-based active perception closed-loop inspection system 5000 in this embodiment of the invention can be found in the specific implementation of the risk field-based active perception closed-loop inspection method in the foregoing embodiments of the invention. To reduce redundancy, it will not be repeated here.

[0115] In summary, the risk-field-based active perception closed-loop inspection system according to embodiments of the present invention, employing the aforementioned risk-field-based active perception closed-loop inspection device, enables a quadruped robot and its onboard sensors to automatically construct a multimodal risk field in multiple scenarios. Under the constraints of this risk field, it conducts active perception adaptive inspections and automatically generates additional observation actions when there is insufficient coverage, high uncertainty, unconfirmed anomalies, path obstruction, or sudden risk changes. This improves inspection coverage, the reliability of anomaly detection, and the stability of task execution.

[0116] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0117] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0118] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0119] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0120] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0121] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0122] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0123] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A proactive sensing closed-loop inspection method based on a risk field, characterized in that, The method includes: A multimodal risk field is constructed, in which each voxel maintains occupancy probability, passage risk, thermal risk, gas risk, visibility risk, semantic category probability, and various uncertainty and timeliness fields in a unified voxel structure, and then integrates and outputs a comprehensive risk. Acquire multi-sensor data and pose, and perform reproducible updates to the multimodal risk field based on the multi-sensor data and pose, wherein the reproducible updates include geometric map updates, thermal risk updates, gas risk updates, visibility risk updates, and uncertainty and timeliness updates; The candidate inspection action set of the multimodal risk field is obtained, and the expected information gain and coverage gain of each candidate inspection action are obtained. Constraints and penalties are imposed with risk cost and time cost to actively perceive, score and rank each candidate inspection action, and the candidate inspection action with the highest score is selected for execution. After execution, candidate actions are regenerated or filtered based on the updated risk field and decision-making continues. The coverage quality evaluation corresponding to each candidate inspection action is obtained. When the coverage quality evaluation is not lower than the preset coverage quality lower limit and the corresponding risk uncertainty is not higher than the preset risk uncertainty threshold, the coverage is judged to be up to standard; otherwise, the coverage is judged to be down to standard. The coverage quality evaluation consists of distance, viewing angle, clarity, thermal imaging effectiveness and gas sampling effectiveness. When the coverage of the current candidate inspection action is not up to standard or the conditions for additional observation are met, an additional observation sub-task is generated, and the additional observation sub-task is incorporated into the current inspection action sequence using the minimum incremental cost. The additional observation sub-task reuses the active perception score, and when selecting additional observation candidate actions, a coverage compliance constraint and a preset threshold for uncertainty reduction are added.

2. The active sensing closed-loop inspection method based on risk field according to claim 1, characterized in that, The thermal risk update includes a joint update of thermal image projection, depth mapping, and weighted index fusion; the gas risk update includes a joint update of hysteresis compensation and spatial diffusion; the visibility risk includes a joint update of the contrast / sharpness index of the visible light image and the effective echo density of the lidar in the forward sector; the passage risk includes a joint update of the slope angle, roughness, and local step height; and the uncertainty and timeliness update includes maintaining the uncertainty of each risk based on observation quality and time decay.

3. The active sensing closed-loop inspection method based on risk field according to claim 1, characterized in that, The process of actively sensing, scoring, and ranking each candidate inspection action, selecting the candidate inspection action with the highest score for execution, and regenerating or filtering candidate actions based on the updated risk field after execution and continuing decision-making includes: Select a region of interest, wherein the region of interest must at least meet the following conditions: the overall risk is higher than a preset risk threshold, or the overall uncertainty is higher than a preset uncertainty threshold and is adjacent to a key inspection item, or there are unconfirmed voxels or semantic objects, or there are inspection items with insufficient coverage quality, and generate the candidate inspection action set based on the region of interest. The candidate inspection action set is subjected to accessibility and security filtering, and the expected information gain and coverage gain of each candidate inspection action that passes the accessibility and security filtering are obtained, as well as the risk cost and time cost of each candidate inspection action that passes the accessibility and security filtering. Based on the expected information gain, coverage gain, risk cost, and time cost of each candidate inspection action that passes the accessibility and security filter, an active perception score and ranking are performed on each candidate inspection action that passes the accessibility and security filter. The candidate inspection action with the highest score is selected for execution, and after execution, candidate actions are regenerated or filtered based on the updated risk field to continue the decision-making process.

4. The active sensing closed-loop inspection method based on risk field according to claim 1, characterized in that, The process of obtaining the coverage quality evaluation corresponding to each candidate inspection action, and determining that coverage meets the standard when the coverage quality evaluation is not lower than a preset coverage quality lower limit and the corresponding risk uncertainty is not higher than a preset risk uncertainty threshold, and otherwise determining that coverage does not meet the standard, includes: Based on the inspection task log, a set of inspection items is constructed. Each inspection item includes spatial location / semantic object binding information, sensor type, target observation distance range, allowable viewing angle range, minimum resolution, thermal imaging dynamic range constraints, and minimum residence time for gas sampling. Based on the distance index, viewing angle index, sharpness index, thermal imaging effectiveness, and gas sampling effectiveness of each inspection item, the coverage quality evaluation of each inspection item is obtained; When the coverage quality evaluation is not lower than the preset coverage quality lower limit and the corresponding risk uncertainty is not higher than the preset risk uncertainty threshold, the current inspection item is judged to meet the coverage standard; otherwise, the current inspection item is judged to fail to meet the coverage standard.

5. The active sensing closed-loop inspection method based on risk field according to claim 1, characterized in that, The additional observation triggering conditions include high uncertainty triggering conditions, unconfirmed anomaly triggering conditions, blocked or unreachable path triggering conditions, and risk mutation triggering conditions.

6. The active sensing closed-loop inspection method based on risk field according to claim 1, characterized in that, The process of generating additional observation subtasks and incorporating them into the current inspection action sequence using minimum incremental cost includes: Obtain an additional set of observation targets, which includes a set of check items with insufficient coverage, anomalous voxel clusters, or regions with high uncertainty; Based on the additional observation target set, and based on the active perception score, the coverage achievement constraint, and the preset threshold of uncertainty reduction, the additional observation sub-task is obtained. The additional observation subtask is incorporated into the current inspection action sequence using the minimum incremental cost.

7. The active sensing closed-loop inspection method based on risk field according to claim 1, characterized in that, The method includes: After the additional observation subtask is completed, the coverage quality and risk field uncertainty of the corresponding inspection items are updated, and active perception decision-making is performed again.

8. A computer-readable storage medium, characterized in that, It stores a risk field-based active perception closed-loop inspection program, which, when executed by the processor, implements the risk field-based active perception closed-loop inspection method as described in any one of claims 1-7.

9. A proactive sensing closed-loop inspection device based on a risk field, characterized in that, The device includes: The risk field construction module is used to construct a multimodal risk field. The multimodal risk field maintains the occupancy probability, passage risk, thermal risk, gas risk, visibility risk, semantic category probability, and various uncertainty and timeliness fields for each voxel in a unified voxel structure, and then integrates and outputs a comprehensive risk. The risk field update module is used to acquire multi-sensor data and pose, and to perform reproducible updates to the multimodal risk field based on the multi-sensor data and pose. The reproducible updates include geometric map updates, thermal risk updates, gas risk updates, visibility risk updates, and uncertainty and timeliness updates. The active perception and decision-making module is used to obtain the candidate inspection action set of the multimodal risk field, obtain the expected information gain and coverage gain of each candidate inspection action, and impose constraints and penalties with risk cost and time cost to actively perceive, score and rank each candidate inspection action, select the candidate inspection action with the highest score for execution, and regenerate or filter candidate actions based on the updated risk field after execution and continue decision-making. The coverage quality evaluation module is used to obtain the coverage quality evaluation corresponding to each candidate inspection action. When the coverage quality evaluation is not lower than the preset coverage quality lower limit and the corresponding risk uncertainty is not higher than the preset risk uncertainty threshold, the coverage is judged to be up to standard; otherwise, the coverage is judged to be down to standard. The coverage quality evaluation consists of distance, viewing angle, clarity, thermal imaging effectiveness and gas sampling effectiveness. The additional observation control module is used to generate additional observation sub-tasks when the coverage of the current candidate inspection action is not up to standard or when the additional observation triggering conditions are met, and to incorporate the additional observation sub-tasks into the current inspection action sequence using the minimum incremental cost. The additional observation sub-tasks reuse the active perception score and add coverage compliance constraints and preset thresholds for uncertainty reduction when selecting additional observation candidate actions.

10. A proactive sensing closed-loop inspection system based on a risk field, characterized in that, The system includes a quadruped robot, sensor components, a task management and visualization terminal, and the active perception closed-loop inspection device based on a risk field as described in claim 9.

Images