Control method of photovoltaic cleaning robot based on visual identification

Abstract

The invention relates to the field of photovoltaic cleaning robot control, in particular to a control method of a photovoltaic cleaning robot based on visual identification. The method comprises the following steps: forming an alignment data packet and a to-be-rechecked set on the basis of completing alignment, reflection suppression and quality scoring by constructing a session configuration packet comprising array topology, component geometry, a security boundary, robot parameters, visual probe parameters and resource budget parameters and driving collaborative acquisition and processing of an image sequence and a pose sequence; reasoning is carried out in combination with the mapping relation, a multi-dimensional result reflecting the dirt category, severity and confidence is generated, candidate cleaning parameters and a value field are constructed accordingly, an execution plan is generated, movement and cleaning instruction issuing are completed, and an execution log is recorded; and completing standard judgment, local reclearing and abnormal typing based on a verification result. According to the invention, through a closed-loop control flow which runs through acquisition, reasoning, decision making and verification, the consistency and rechecking performance of cleaning decision making and execution are improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic cleaning robot control, and more particularly to a control method for a photovoltaic cleaning robot based on vision recognition. Background Technology

[0002] In the field of photovoltaic cleaning robot control, existing control methods typically revolve around task orchestration based on image and pose sequences, and combine robot parameters and visual probe parameters to issue motion and cleaning commands. However, these methods suffer from limitations such as insufficient constraint correlation between the session configuration package and the image / pose sequences, a lack of unified verification loops for alignment, reflection suppression, and quality scoring, and loose data transfer between inference products and subsequent control stages. Existing methods often rely on a single link to complete data acquisition, judgment, and action generation. In scenarios constrained by both safety boundaries and resource budget parameters, issues such as insufficient consistency in the source of alignment data packets and a lack of traceable evidence for updating the set to be verified can arise. This leads to unstable connections between the three-image results and the execution plan, making it difficult to meet the consistency requirements of generating execution plans and issuing motion and cleaning commands based on the three-image results. For the joint processing of aligned data packets, mapping tables, and inference, existing technologies generally suffer from common shortcomings such as inconsistent organizational standards for type graphs, severity graphs, and confidence graphs; lack of stable mapping between action segment index tables and candidate parameter groups; and insufficient granularity in recording benefit items, resource cost items, and risk cost items. These shortcomings make it difficult to form a consistent process of data collection, alignment, inference, scoring, planning, distribution, logging, and verification in the application scenario of photovoltaic cleaning robot control methods. This results in an incomplete correlation link between the execution log and the verification results of the three graphs, further leading to a lack of stable data basis and consistent operational constraints in the generation process of compliance judgment, non-compliance set, partial re-cleaning task, and abnormal classification results. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a control method for a photovoltaic cleaning robot based on vision recognition, comprising:

[0004] The array topology, component geometry, safety boundaries, robot parameters, visual probe parameters, and resource budget parameters are obtained to generate a session configuration package; the session configuration package includes segmentation rules, a mapping table, an exposure strategy table, an action fragment index table, and a classification rule table;

[0005] Image sequences and pose sequences are acquired based on session configuration packets, and alignment, reflection suppression and quality scoring are performed to generate alignment data packets and a set to be reviewed.

[0006] Based on the aligned data packets and mapping table, inference is performed to generate a three-graph result; the three-graph result includes a type graph, a severity graph, and a confidence graph, and the set to be reviewed is updated according to the confidence graph;

[0007] Based on the results of the three graphs, the action fragment index table, and the resource budget parameters, candidate parameter groups are generated, and the benefit items, resource cost items, and risk cost items are calculated to generate net benefit score records and value fields.

[0008] An execution plan is generated based on the value field, and the execution plan includes a component priority queue and a local path action sequence;

[0009] Based on the execution plan, motion commands and cleaning commands are issued, and an execution log is generated;

[0010] Based on the execution log, the image sequence is collected to generate a three-image verification result. The three-image verification result is compared with the three-image result to generate a set of compliance and non-compliance. A local re-cleaning task is generated and an abnormal classification result is generated according to the classification rule table.

[0011] Furthermore, the segmentation rules in the session configuration package include cell grid size and cell number fields; the mapping table includes pixel coordinate and cell number fields; the exposure strategy table includes exposure, gain, shutter speed, and switching condition fields; the motion fragment index table includes type map label, severity map level, and cleaning instruction parameter fields; and the classification rule table includes abnormal classification result trigger and strategy number fields.

[0012] Furthermore, the quality score is obtained by combining the sharpness score, exposure saturation score, reflection ratio score, and motion blur score; when the quality score is less than the quality threshold, the corresponding image frame is written into the set to be reviewed and the sampling parameters are switched according to the exposure strategy table to re-acquire the image sequence.

[0013] Furthermore, the reasoning process based on the aligned data packets and mapping table includes:

[0014] The inference process includes multi-scale feature extraction, scale fusion, and cell aggregation; the type map is composed of the category labels after cell aggregation, the severity map is composed of the level values ​​after cell aggregation, and the confidence map is synthesized from the consistency score and quality score of the multi-scale output.

[0015] Furthermore, the candidate parameter group is obtained by indexing the action segment index table according to the type map label and the severity map level; the benefit item is calculated by the severity map level and the cell area field; the resource cost item is calculated by the water consumption field, electricity consumption field and duration field in the candidate parameter group; the risk cost item is calculated by the confidence map and the set to be reviewed; the net benefit score record is calculated by the benefit item, resource cost item and risk cost item according to the linear composition rule.

[0016] Furthermore, the component priority queue is obtained by aggregating and sorting the net benefit score records of the value field within the component scope; the local path action sequence is generated by sorting the cell net benefit score records of the value field within the component scope, and the cleaning instruction parameter field in the candidate parameter group is bound to the sorted cell.

[0017] Furthermore, the execution plan update trigger condition is a change in the set to be reviewed or a change in the three graph results; the incremental update of the execution plan includes component priority queue rearrangement and replacement of local path action sequences within the component scope.

[0018] Furthermore, the anomaly classification result is determined by the consecutive failure count field, the persistent low value field of the confidence graph, and the parameter upper limit trigger flag field of the cleaning instruction parameter field; the strategy number field generates a degradation control instruction or a remote alarm instruction and writes it into the execution log.

[0019] The key innovations of this invention include:

[0020] (1) Based on the session configuration package, the array topology, component geometry, safety boundary, robot parameters, vision probe parameters and resource budget parameters are uniformly assembled, and the acquisition, alignment, reflection suppression and quality scoring of image sequence and pose sequence are driven within the same session, generating traceable alignment data packets and forming a continuously updated set to be reviewed.

[0021] (2) Based on the alignment data package and mapping table, inference is performed and the three-graph result is generated. The three-graph result consists of a type graph, a severity graph and a confidence graph. The confidence graph is used to perform closed-loop update on the set to be reviewed, so that the inference product and the review object are linked under the same data organization caliber.

[0022] (3) Based on the three graph results, action fragment index table and resource budget parameters, generate candidate parameter groups, calculate benefit items, resource cost items and risk cost items to form net benefit score records and value fields, and then generate an execution plan containing component priority queue and local path action sequence from the value field. After issuing motion instructions and cleaning instructions, the execution log triggers a closed loop link to verify the three graph results, meet the criteria, fail to meet the criteria set, local re-cleaning task and abnormal classification results.

[0023] The following are its main beneficial effects:

[0024] (1) In view of the problem that the constraints between the session configuration package and the image sequence and pose sequence are insufficient in the background technology, and that the alignment, reflection suppression and quality scoring lack a unified review loop, the session configuration package is used to uniformly assemble the array topology, component geometry, safety boundary, robot parameters, vision probe parameters and resource budget parameters, so that the acquisition, alignment, reflection suppression and quality scoring can run in the same session, and the output is solidified into alignment data package and set to be reviewed, so that the subsequent inference based on alignment data package and mapping table has a consistent data source and review entry.

[0025] (2) In view of the problems of inconsistent organizational standards of type graph, severity graph and confidence graph in the background technology and lack of traceable basis for updating the set to be reviewed, the three graph results are generated by taking the alignment data package and mapping table as input, and the confidence graph is used to perform closed-loop update of the set to be reviewed, so that the inference output and the review object are linked under the same standard, reducing the risk of separation between the three graph results and the review object, thereby providing a consistent input basis for the subsequent generation of candidate parameter groups and construction of value field based on the three graph results, action fragment index table and resource budget parameters.

[0026] (3) In view of the loose connection between the inference products and the subsequent control links in the background technology, the lack of stable mapping between the action segment index table and the candidate parameter group, and the insufficient granularity of the records of benefit items, resource cost items and risk cost items, the candidate parameter group is generated by using the three-graph results, the action segment index table and the resource budget parameters to calculate the benefit items, resource cost items and risk cost items, forming a net benefit score record and value field, and generating an execution plan accordingly. This enables the component priority queue and the local path action sequence to form a feasible distribution link with the motion instructions and cleaning instructions. At the same time, the execution log drives the verification of the three-graph results, the standard determination, the non-standard set, the local re-cleaning task and the abnormal classification results, so that the cleaning instructions and the subsequent review, re-cleaning and classification have a consistent record link and operation boundary. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating a control method for a photovoltaic cleaning robot based on vision recognition, provided in an embodiment of this application. Detailed Implementation

[0028] Example 1: Refer to Figure 1 This is a flowchart illustrating a control method for a photovoltaic cleaning robot based on visual recognition, provided in an embodiment of the present invention. The process may include at least steps S100-S700:

[0029] S100: Obtain array topology, component geometry, safety boundaries, robot parameters, vision probe parameters, and resource budget parameters, and generate a session configuration package;

[0030] S200: Acquire image sequences and pose sequences based on session configuration packets, perform alignment, reflection suppression and quality scoring, and generate alignment data packets and sets to be reviewed;

[0031] S300. Based on the aligned data packets and mapping table, inference is performed to generate a three-graph result; the three-graph result includes a type graph, a severity graph, and a confidence graph, and the set to be reviewed is updated according to the confidence graph;

[0032] S400: Based on the three graph results, action segment index table and resource budget parameters, generate candidate parameter groups, calculate benefit items, resource cost items and risk cost items, and generate net benefit score records and value fields;

[0033] S500. Generate an execution plan based on the value field, the execution plan including a component priority queue and a local path action sequence;

[0034] S600: Based on the execution plan, issue motion instructions and cleaning instructions, and generate an execution log;

[0035] S700. Based on the execution log, collect and verify the image sequence to generate a verification three-image result. Compare the verification three-image result with the three-image result to generate a pass / fail set and a fail set. Generate a local re-cleaning task and generate an abnormal classification result according to the classification rule table.

[0036] Specifically, in S100, the array topology, component geometry, safety boundaries, robot parameters, vision probe parameters, and resource budget parameters are obtained, and a session configuration package is generated.

[0037] This step is executed by the photovoltaic cleaning robot controller, which is installed within the photovoltaic cleaning robot body. The controller performs parameter assembly and conversational configuration within the technical scope of the mobile body's movement and the linkage control of the actuators. Specifically, the controller accesses the array topology from the array configuration file on the power plant operation and maintenance side, the component geometry from the component installation file on the installation and mapping side, the safety boundary from the operation and maintenance safety rule base, the robot parameters and vision probe parameters from the equipment factory calibration file, and the resource budget parameters from the task scheduling side's work quota. The array topology is defined as a combination of data including the row and column indices of components in the array, the connection relationships between adjacent components, the channel travel relationships, and the turning restriction relationships. The component geometry is defined as a combination of data including the component's external boundary, the cleanable area boundary, the component tilt direction information, and the component coordinate system. The safety boundary is defined as... The definition of the parameters is a combination of data for the restricted area boundary, boundary crossing threshold, emergency stop triggering condition, and speed limit area boundary. The robot parameters are defined as a combination of kinematic constraint parameters, drive wheel or track control parameters, minimum turning radius parameters, maximum acceleration parameters, and positioning odometer calibration parameters. The visual probe parameters are defined as a combination of data for the visual probe installation pose, intrinsic parameter calibration parameters, distortion calibration parameters, field of view coverage parameters, and image acquisition timing parameters. The resource budget parameters are defined as a combination of data for the current operation time window parameters, water quota parameters, electricity quota parameters, cleaning agent quota parameters, and verification sampling quota parameters. The array topology, component geometry, safety boundary, robot parameters, vision probe parameters, and resource budget parameters constitute the minimum set of fields in this step. After the controller completes the connection, it performs caliber verification and consistency alignment processing on each field. Calibration verification includes unit system consistency verification, coordinate system consistency verification, and boundary closure verification. Consistency alignment processing includes binding the component index in the array topology with the component coordinate system in the component geometry, projecting the safety boundary onto the component coordinate system and generating an out-of-bounds judgment benchmark, and writing the positioning odometry calibration parameters in the robot parameters and the installation pose in the vision probe parameters into the same session time base.

[0038] After input access is completed, the controller generates a session configuration package and writes it to the session configuration package version record. The session configuration package consists of segmentation rules, a mapping table, an exposure strategy table, an action fragment index table, and a classification rule table. Each table entry is bound to a session number field and includes a generation timestamp field and a verification code field, thus ensuring traceability of configuration references for the same session in subsequent steps. Specifically, the controller generates segmentation rules based on component geometry and array topology. These segmentation rules are used to discretize the cleanable area of ​​a component into a cleanable cell grid. The segmentation rules include a cell grid size field and a cell number field. The cell grid size field is jointly determined by the component's cleanable area scale and the coverage width of the cleaning actuator. The cell number field is generated by concatenating the component index and the grid row and column index in the array topology and written to the segmentation rule record. The controller generates a mapping table based on visual probe parameters and segmentation rules. This mapping table establishes the correspondence between image pixels and clean cells, and includes a pixel coordinate field and a cell number field. The pixel coordinate field comes from the visual probe imaging plane coordinates, and the cell number field comes from the segmentation rule record. The controller performs correction mapping on the pixel coordinates using intrinsic calibration parameters and distortion calibration parameters, and then performs spatial alignment by combining the visual probe mounting pose and component coordinate system, thereby assigning each pixel coordinate field to a unique cell number field and fixing it in the mapping table. The controller generates an exposure strategy table based on visual probe parameters and safety boundaries. This exposure strategy table constrains the exposure control actions and switching trigger conditions for image acquisition. The exposure strategy table includes an exposure field, a gain field, a shutter field, and a switching condition field. The switching condition field is obtained by combining the reflective highlight area ratio threshold, the shadow coverage ratio threshold, and the motion blur indication threshold, and is linked to the speed limit area boundary and emergency stop trigger condition in the safety boundary, so that the image sampling parameter switching and motion state switching are under the same session constraint.

[0039] The controller generates an action segment index table based on resource budget parameters and the cleaning execution mechanism's capability profile. This index table maps the type and severity maps from the subsequent three graphs to cleaning instruction parameters. The index table includes a type map label field, a severity map level field, and a cleaning instruction parameter field. The type map label field uses a pre-defined set of dirt type codes, the severity map level field uses a pre-defined set of severity levels, and the cleaning instruction parameter field includes brush head rotation speed, water pressure, detergent dosage, number of repetitions, and dwell time, and is constrained by water, electricity, and detergent quotas in the resource budget parameters. When generating the index table, the controller simultaneously writes the action segment version number and applicable resource range fields to ensure version consistency when subsequent steps reference the index table. The controller generates a classification rule table based on the security boundary and operation and maintenance rule base. The classification rule table is used for subsequent judgment of abnormal classification results and policy number triggering. The classification rule table includes an abnormal classification result trigger field and a policy number field. The abnormal classification result trigger field registers the threshold of the number of consecutive failures, the threshold of the confidence graph low value persistence field, and the parameter upper limit trigger flag field threshold of the cleaning instruction parameter field. The policy number field registers the degradation control policy identifier and remote alarm policy identifier corresponding to each trigger condition, and is bound to the session number field.

[0040] Once the session configuration package is generated, the controller writes it to memory and publishes the session number field. This session configuration package is used as input to S200 and invoked by the controller. Within it, the mapping table is referenced as input to S300, the exposure strategy table is referenced as the basis for switching sampling parameters in S200, the action fragment index table is referenced as the basis for indexing candidate parameter groups in S400, and the classification rule table is referenced as the basis for generating abnormal classification results in S700, thus forming a hierarchical input relationship across main steps. The technical effect of this step can be summarized as follows: By assembling and solidifying the array topology, component geometry, safety boundaries, robot parameters, vision probe parameters, and resource budget parameters into a session configuration package, subsequent acquisition, inference, scoring, plan generation, and classification processing can run under the same configuration version.

[0041] S200: Acquire image sequences and pose sequences based on session configuration packets, perform alignment, reflection suppression and quality scoring, and generate alignment data packets and sets to be reviewed;

[0042] This step is executed on-site by the photovoltaic cleaning robot controller. The controller reads the session configuration package generated and stored by S100 as input. The session configuration package includes segmentation rules, a mapping table, an exposure strategy table, a motion segment index table, and a classification rule table, and includes a session number field, a generation timestamp field, and a checksum field. The acquisition objects in this step are image sequences and pose sequences. The image sequences are output by the vision probe installed on the robot body, and the pose sequences are output by the pose acquisition unit of the robot body. The pose acquisition unit consists of an odometry component, an attitude measurement component, and a timestamp component, providing the controller with position data fields, attitude data fields, and pose timestamp fields. When entering this step, the controller first performs a checksum field consistency verification on the session configuration package and writes the session number field into the acquisition session record, thus ensuring that the subsequent image frame number field and pose timestamp field are archived under the same session number field.

[0043] Specifically, the controller generates acquisition trigger conditions based on the component indexes and channel travel relationships in the array topology, and combines these with the speed limit zone boundary, restricted area boundary, and emergency stop trigger conditions in the safety boundary to generate the acquisition window state. The acquisition trigger conditions are composed of a road segment entry marker field, a component arrival marker field, and a speed threshold field. When the acquisition window state satisfies both the road segment entry marker field and the component arrival marker field being true, the controller sends a sampling instruction parameter group to the vision probe and starts image acquisition. The sampling instruction parameter group comes from the exposure strategy table and includes an exposure field, a gain field, and a shutter field. The controller simultaneously writes the sampling instruction parameter group into the sampling parameter record field and binds it to the image frame number field and the acquisition timestamp field. The image sequence consists of consecutive image frames, each image frame containing an original image data field, an acquisition timestamp field, and a sampling parameter record field; the pose sequence consists of consecutive pose records, each pose record containing a position data field, an attitude data field, and a pose timestamp field. In the engineering embodiment, the robot walks along the component row direction, the vision probe faces the component surface to collect image frames, the pose acquisition unit outputs pose records synchronously, and the controller receives and caches the image sequence and pose sequence within the same acquisition time window. The cache index is composed of the session number field and the acquisition timestamp field.

[0044] After the connection is established, the controller performs alignment processing on the image sequence and pose sequence. This alignment processing includes two links: timestamp alignment and coordinate aperture alignment. In the timestamp alignment link, the controller retrieves adjacent pose timestamp fields in the pose sequence based on the acquisition timestamp field of the image frame, and performs interpolation alignment to generate an image pose binding record. The image pose binding record includes an image frame number field, an acquisition timestamp field, a matched pose timestamp field, a position data field, and a pose data field. When the interval between adjacent pose timestamp fields exceeds the pose interval threshold field, the controller writes the corresponding image frame number field into the set to be reviewed and records the reason field for review as a pose gap. In the coordinate alignment link, the controller reads the mapping table in the session configuration package, uses the pixel coordinate field and cell number field in the mapping table to map each pixel coordinate field in the image frame to a unique cell number field, and writes the mapping result into the cell mapping record field; at the same time, it reads the distortion calibration parameter in the visual probe parameters, performs distortion correction processing on the original image data field to generate a corrected image data field, and binds the corrected image data field and the cell mapping record field together to the image frame number field, thus forming the input basis required for subsequent inference.

[0045] In this step, reflection suppression is performed by the controller as part of the image quality control chain. The controller performs highlight region localization and saturation pixel marking processing on the corrected image data field to obtain the reflection mask data field. The reflection mask data field includes a highlight region boundary field and a saturation pixel ratio field. Based on the reflection mask data field, the controller performs local replacement and contrast reshaping processing on the corrected image data field to generate the reflection suppression image data field, and binds the reflection mask data field to the image frame number field for archiving. The local replacement processing adopts a neighborhood consistency constraint. For the saturation pixel marked area, the statistical characteristics of the adjacent non-saturated areas are selected and written into the replacement record field, so that the reflection suppression image data field maintains a spatial caliber consistent with the cell mapping record field. At the same time, the controller reads the switching condition field in the exposure strategy table, compares the saturation pixel ratio field in the reflective mask data field with the switching condition field, and when the switching condition field is met, the controller writes the current sampling parameter record field into the sampling switching history field and sends a new sampling instruction parameter group to the vision probe. The new sampling instruction parameter group also consists of the exposure field, gain field and shutter field. The new image frame generated by resampling uses the same session number field and generates a new image frame number field.

[0046] The quality score is calculated by the controller on the reflection suppression image data field. The quality score is generated by combining the sharpness score, exposure saturation score, reflection percentage score, and motion blur score. The sharpness score is derived from edge intensity statistics and written to the sharpness score record field; the exposure saturation score is derived from the saturated pixel percentage field and written to the exposure saturation score record field; the reflection percentage score is derived from the reflection mask data field and written to the reflection percentage score record field; and the motion blur score is derived from the pose change rate statistics between adjacent image frames and written to the motion blur score record field. After the session configuration packet verification is completed, the controller generates a quality threshold field and writes it to the quality threshold record field. When the quality score is less than the quality threshold field, the controller writes the corresponding image frame number field to the set to be reviewed, records the reason for review as insufficient quality, and triggers sampling parameter switching and resampling based on the switching condition field of the exposure strategy table. The number of resampling attempts is constrained by the review sampling quota parameter in the resource budget parameters. When the upper limit of the review sampling quota parameter is reached, the controller stops resampling and retains the record of the set to be reviewed.

[0047] At the end of this step, the controller encapsulates the results of alignment, glare suppression, and quality scoring into an alignment data packet and generates a set to be reviewed. The alignment data packet includes a session number field, an image sequence field, a pose sequence field, a glare suppression image data field, a cell mapping record field, an image pose binding record field, a quality score record field, and a sampling switching history field. The alignment data packet is used as input to S300, and the cell mapping record field and the mapping table form the inference input in S300. The set to be reviewed includes a session number field, an image frame number field, a review reason field, a quality score record field, and a matching pose timestamp field. The set to be reviewed participates in the update process in S300 along with the confidence map and is continuously archived for reference in subsequent steps. In summary, the technical effect of this step is that through the sampling, alignment, glare suppression, and quality scoring link driven by the session configuration packet, the image sequence and pose sequence form an alignment data packet under the same session number field, and simultaneously form a set to be reviewed for use in subsequent inference and update links.

[0048] S300. Based on the aligned data packets and mapping table, inference is performed to generate a three-graph result; the three-graph result includes a type graph, a severity graph, and a confidence graph, and the set to be reviewed is updated according to the confidence graph;

[0049] This step is executed by the photovoltaic cleaning robot controller on the robot body side. The controller reads the session number field, image sequence field, pose sequence field, reflection suppression image data field, cell mapping record field, image pose binding record field, quality score record field, and sampling switching history field from the alignment data packet output by S200 as inference input. Simultaneously, it reads the mapping table from the session configuration packet output by S100 as the inference caliber input. The mapping table contains pixel coordinate fields and cell number fields. The cell mapping record field is generated by the mapping table in the coordinate caliber alignment link of S200 and bound to the image frame number field. When entering this step, the controller first performs a consistency check between the session number field in the alignment data packet and the session number field in the session configuration packet, and writes the consistency check result into the inference session record field. Then, it extracts the frames to be inferred sequentially from the reflection suppression image data field according to the image frame number field, and combines this with the image pose binding record field to extract the matching pose timestamp field, position data field, and attitude data field, so that the inference input simultaneously includes image information and motion state information at the sampling time.

[0050] Specifically, the controller executes the inference link, which includes multi-scale feature extraction, scale fusion, and cell aggregation. Multi-scale feature extraction involves the inference module performing convolutional feature extraction on the reflection suppression image data field at different spatial resolutions to generate a multi-scale feature map field. This multi-scale feature map field is bound to the image frame number field and records a scale identifier field. Scale fusion involves the inference module performing a fusion operation on the multi-scale feature map field to generate a fused feature map field. This fused feature map field records a channel fusion record field and a scale fusion record field. Cell aggregation involves the controller merging pixel-level features in the fused feature map field according to the pixel coordinate field and cell number field in the mapping table, calculating a cell feature vector field based on the cell number field. This cell feature vector field is bound to the cell number field and written to the cell aggregation record field. To ensure stable operation of the inference chain under abnormal input conditions, the controller reads the sharpness score record field, exposure saturation score record field, reflection ratio score record field, and motion blur score record field from the quality score record field. During the cell aggregation stage, a quality label field is added to the feature vector field of the cell corresponding to the low-quality frame. When the exposure saturation score record field exceeds the saturation threshold field or the reflection ratio score record field exceeds the reflection threshold field, the controller writes the corresponding image frame number field into the set to be reviewed and records the reason field to be reviewed as high reflection. During the inference output stage, the confidence map output for that frame is registered for weight reduction and written into the confidence weight reduction record field. When the motion blur score record field exceeds the blur threshold field, the controller writes the corresponding image frame number field into the set to be reviewed and records the reason field to be reviewed as motion blur. The sampling parameter record field in the sampling switching history field is associated with the frame and archived for reference in subsequent steps.

[0051] After obtaining the cell feature vector field, the controller generates a type map, a severity map, and a confidence map, and encapsulates them into a three-map result. Specifically, the type map is generated by the inference module performing category discrimination on the cell feature vector field and outputting a category label field. The controller then writes the category label field into the type map label field according to the cell number field to form the type map. The severity map is generated by the inference module performing hierarchical discrimination on the cell feature vector field and outputting a level value field. The controller then writes the level value field into the severity map level field according to the cell number field to form the severity map. The confidence map is generated by the controller performing a synthesis process on the multi-scale output consistency score and quality score record field of the inference output. The multi-scale output consistency score is obtained by the scale fusion stage by statistically analyzing the category discrimination consistency of the same cell number field at different scales and writing it into the consistency score record field. The synthesis process involves the controller jointly mapping the consistency score record field and the quality score record field to the confidence value field and writing it into the confidence map according to the cell number field. The confidence value field is also bound to the image frame number field and the session number field for archiving. Understandably, the three graph results are structurally composed of three parts: a type graph, a severity graph, and a confidence graph. The three parts share a cell number field as an index key, which allows subsequent steps to directly associate the type graph label field, the severity graph level field, and the confidence value field on the same cell number field.

[0052] This step updates the set to be reviewed according to the confidence map after generating the three-image results. The update is triggered by the controller based on the classification rule table and quality threshold record field in the session configuration package. Specifically, the controller extracts the confidence value field corresponding to each cell number field from the confidence map and generates a persistent low value field for the confidence map. The persistent low value field is obtained by counting the number of times the confidence value field of the same cell number field is lower than the confidence threshold field in consecutive image frame number fields. When the confidence value field is lower than the confidence threshold field, the controller writes the corresponding cell number field into the set to be reviewed and records the reason field for review as a low confidence value. When the persistent low value field of the confidence map meets the abnormal classification result trigger field in the classification rule table, the controller combines the corresponding image frame number field and cell number field into the set to be reviewed and appends a placeholder record for the strategy number field. The placeholder record for the strategy number field is used for reference in the abnormal classification result generation link of S700. Furthermore, the controller performs deduplication and merging processing on the set to be reviewed. The deduplication and merging processing uses the session number field, image frame number field, and cell number field as the union key. It performs rule synthesis on the review reason fields under the same union key and writes them into the synthesized reason field, so that the subsequent steps have a consistent reason caliber when reading the set to be reviewed.

[0053] The controller writes the three graph results and the updated set to be reviewed into an inference output package and archives it. The inference output package includes a session number field, a type graph, a severity graph, a confidence graph, a confidence reduction record field, and a cause of composition field. The inference output package is invoked as input to S400. Specifically, the type graph label field in the type graph and the severity graph level field in the severity graph are used in S400 to index the cleaning instruction parameter field according to the action fragment index table. The confidence value field in the confidence graph and the set to be reviewed are used in S400 to calculate the risk cost item and generate a net benefit score record and a value field. In summary, the technical effect of this step is: by aligning the data packet and the mapping table-driven multi-scale inference and cell aggregation links, three graph results are formed, and the set to be reviewed is updated along with the confidence graph, thus ensuring consistent inference input for subsequent scoring and planning links under the cell number field caliber.

[0054] S400: Based on the three graph results, action segment index table and resource budget parameters, generate candidate parameter groups, calculate benefit items, resource cost items and risk cost items, and generate net benefit score records and value fields;

[0055] This step is executed by the photovoltaic cleaning robot controller. The controller takes the three graph results output by S300 as the main input and reads the action segment index table and resource budget parameters from the session configuration package output by S100 as configuration input. The three graph results include a type graph, a severity graph, and a confidence graph, which share a cell number field as an index key. The type graph includes a type graph label field, the severity graph includes a severity graph level field, and the confidence graph includes a confidence value field. The action segment index table includes a type graph label field, a severity graph level field, and a cleaning instruction parameter field. The resource budget parameters include the current operation time window parameters, water quota parameters, electricity quota parameters, cleaning agent quota parameters, and review sampling quota parameters, which are bound in the session configuration package using the session number field. At the beginning of this step, the controller performs a consistency check on the three graph results and the session number field of the session configuration package, and writes the check record into the scoring session record field. Then, it performs a join table reading on the type graph label field, severity graph level field and confidence value field according to the cell number field, thereby forming the cell scoring input record field. The cell scoring input record field serves as the input basis for candidate parameter group generation and cost item calculation.

[0056] Specifically, the controller generates candidate parameter groups, which are obtained by indexing the action segment index table by the type map label field and the severity map level field. The controller reads the corresponding type map label field and severity map level field for each cell number field, and searches for matching records in the action segment index table. When a match is found, the cleaning instruction parameter field is extracted and bound to the cell number field to generate a candidate parameter group record. The candidate parameter group record includes the cell number field, type map label field, severity map level field, and cleaning instruction parameter field, with the addition of a resource consumption estimation field and a duration estimation field. The resource consumption estimation field is obtained by combining the water consumption field, electricity consumption field, and cleaning agent dosage field from the cleaning instruction parameter field. The duration estimation field is calculated jointly by the dwell time parameter, reciprocation number parameter, and the maximum acceleration parameter and minimum turning radius parameter from the robot parameters. The water consumption field, electricity consumption field, cleaning agent dosage field, dwell time parameter, and reciprocation number parameter constitute the minimum set of fields for generating the candidate parameter group in this step. The resource consumption estimation field and duration estimation field serve as direct inputs for subsequent resource cost calculations. When a search fails, the controller writes the cell number field to the set to be reviewed and records the reason for the review as "index missing". At the same time, it associates the version number field of the action fragment index table with the record and writes it to the index missing record field for reference in subsequent abnormal classification results.

[0057] After the candidate parameter group is generated, the controller calculates the benefit item, resource cost item, and risk cost item and generates a net benefit score record. Specifically, the benefit item is calculated from the severity map level field and the cell area field. The cell area field is converted from the cell grid size field in the segmentation rule and stored in the session configuration package as associated with the cell number field. The controller reads the severity map level field and the cell area field for each cell number field, performs level mapping processing to obtain the level weight field, and combines the level weight field with the cell area field to generate the benefit item record field. The resource cost item is calculated from the water consumption field, electricity consumption field, and duration field in the candidate parameter group. The duration field is generated by the aforementioned duration estimation field. The controller generates a resource constraint record field based on the resource budget parameters. The resource constraint record field includes water consumption quota parameters, electricity consumption quota parameters, and the time window parameters for this operation. For each cell number field, the controller calculates the corresponding water usage ratio field, electricity usage ratio field, and time usage ratio field, and further combines the three into a resource cost item record field and binds it to the cleaning instruction parameter field for archiving. The risk cost item is calculated from the confidence graph and the set to be reviewed. The controller reads the confidence value field from the cell number field of each cell and generates a confidence risk field. At the same time, it checks whether there are any related records with the same cell number field in the set to be reviewed. When a related record exists, the controller maps the reason field to be reviewed to a review weight field and combines it with the confidence risk field to form a risk cost item record field. When no related record exists, the controller generates the risk cost item record field based only on the confidence value field. Understandably, the confidence value field and the reason field to be reviewed constitute the minimum set of fields for calculating the risk cost item in this step, and the review sampling quota parameter constrains the size of the set to be reviewed. When generating the risk cost item record field, the controller simultaneously writes the review quota occupancy field and the review queue number field for reference in subsequent execution plans and verification links.

[0058] After obtaining the benefit item record field, resource cost item record field, and risk cost item record field, the controller generates a net benefit score record and further generates a value field. Specifically, the net benefit score record is calculated from the benefit item, resource cost item, and risk cost item according to a linear composition rule. The linear composition rule is stored in the session configuration package as a score weight record field and bound to the session number field. The controller loads the score weight record field as a composition weight field and calculates the net benefit score value field for each cell number field, thereby generating the net benefit score record. The net benefit score record includes a session number field, a cell number field, a net benefit score value field, a benefit item record field, a resource cost item record field, a risk cost item record field, a cleaning instruction parameter field, and an index missing record field. The value field is generated by the controller by aggregating net income score records according to component scope. The component scope is determined by the array topology and component geometry and has an attribution relationship with the cell number field. The controller writes the net income score value field corresponding to the cell number field under the same component index into the component value distribution record field, and arranges them spatially according to the cell grid size field to generate the value field. Each position in the value field records the cell number field, net income score field, and cleaning instruction parameter field, so that subsequent steps can directly read the value distribution within the component scope and form a component priority queue and local path action sequence.

[0059] To achieve automated closed-loop and auditable operation of this step in the project, the controller synchronously writes version management record fields when generating net income score records and value fields. The version management record fields include session number field, action fragment index table version number field, score weight record field version number field, and resource budget parameter version number field. The version management record fields are also linked with the sampling switching history field and confidence reduction record field. When the action fragment index table version number field or the resource budget parameter version number field changes, the controller triggers the recalculation condition flag field and regenerates candidate parameter groups and net income score records for the affected cell number field. The recalculation process writes the recalculation batch field and recalculation reason field, so that the value field and net income score records maintain a consistent version caliber during the configuration evolution process. In the engineering implementation, the robot walks along the array channel and outputs three-image results in real time. The controller updates the candidate parameter group frame by frame based on the cell number field and accumulates the resource consumption estimation field under the parameter constraints of the current operation time window. If the water consumption ratio field or the electricity consumption ratio field exceeds the quota threshold field, the controller writes the corresponding cell number field into the set to be reviewed and records the reason field for review as resource shortage. At the same time, the resource cost item record field corresponding to the cell number field is marked as high cost state. Then, the net benefit score value field is updated and written into the net benefit score record, so that the value field is updated synchronously with the resource status.

[0060] The output of this step is a net benefit score record and a value field. The controller writes the net benefit score record and value field into a score output package and archives it. The score output package includes a session number field, a net benefit score record, a value field, a version management record field, and a recalculation batch field. The value field is used as input to S500 to generate the execution plan. The net benefit score record participates in the generation of component priority queues and the sorting criteria for local path action sequences in S500. Simultaneously, the cleaning instruction parameter field in the net benefit score record serves as the parameter source for issuing cleaning instructions in S600. The risk cost item record field is referenced as a verification and association basis in the subsequent anomaly classification result chain. In summary, the technical effect of this step is: through the linkage of the three-graph results with the action fragment index table and resource budget parameters in the score chain, a net benefit score record and value field that can be directly referenced by subsequent execution plans are formed, and traceable calculation and output are completed under the constraints of the session number field and the version management record field.

[0061] S500. Generate an execution plan based on the value field, the execution plan including a component priority queue and a local path action sequence;

[0062] This step is executed by the photovoltaic cleaning robot controller on the robot body side. The controller takes the value field output by S400 as the main input, and the net benefit score record output by S400 as the linkage input. It also reads the array topology, component geometry, safety boundary, robot parameters, and resource budget parameters from the session configuration package output by S100 as constraint inputs. The value field is spatially arranged according to the cell grid size field and records the cell number field, net benefit score value field, and cleaning instruction parameter field. The net benefit score record includes the session number field, cell number field, net benefit score value field, benefit item record field, and resource cost field. The system includes a record field, a risk cost record field, and a cleaning instruction parameter field. The array topology includes a component index field, a component adjacency field, and a channel connectivity field. The component geometry includes a component boundary vertex field and a component attitude reference field. The safety boundary includes a restricted area field, an edge buffer distance field, and a turning limit field. The robot parameters include maximum speed parameters, maximum acceleration parameters, minimum turning radius parameters, and a positioning error threshold field. The resource budget parameters include the current operation time window parameters, water quota parameters, electricity quota parameters, cleaning agent quota parameters, and verification sampling quota parameters, and are bound to the session number field. At the beginning of this step, the controller performs a consistency check on the value field and the session number field of the session configuration package, writes the check result to the planned session record field, and establishes an attribution index between the cell number field in the value field and the component index field in the array topology. This allows the plan generation link to directly read the net benefit score field and the cleaning instruction parameter field within the component range.

[0063] Specifically, the controller generates a component priority queue, which is obtained by aggregating and sorting the net profit score records within the component scope of the value field. The controller traverses the component set based on the component index field of the array topology, extracts all cell number fields within the component scope for each component index field, and then reads the corresponding net profit score value field from the value field to form a component-internal score set field. The controller performs aggregation calculations on the component-internal score set field to obtain a component aggregate score field, and writes the component aggregate score field and the component index field into the component score record field. The aggregation calculation involves the controller sorting the component-internal score set field in descending order by the net profit score value field and extracting the preceding score fragment field. The preceding score fragment field is bound to the cell number field and used to generate the component aggregate score field. Simultaneously, the extraction threshold field is written into the plan parameter record field for auditing. The controller further sorts the component scoring record fields according to the component aggregate scoring fields to generate a component priority queue. In the component priority queue, a queue sequence number field and a budget occupancy estimate field are appended to the index field of each component. The budget occupancy estimate field is obtained by aggregating the water consumption, electricity consumption, and duration fields from the cleaning instruction parameter field corresponding to the cell number field within the component's scope. This is then compared with the water consumption quota parameter, electricity consumption quota parameter, and current operation time window parameter in the resource budget parameters to form a component budget constraint marker field. Understandably, the minimum set of fields upon which the component priority queue generation depends includes the component index field, cell number field, and net revenue score value field. The budget occupancy estimate field and the component budget constraint marker field constitute an extended field set, used to provide resource constraint caliber during subsequent local path action sequence generation.

[0064] After the component priority queue is generated, the controller generates a local path action sequence for each component index field in the priority queue. This local path action sequence is generated by sorting the net benefit score records of cells within the component's value field, and then binding the cleaning instruction parameter field from the candidate parameter group to the sorted cells. Specifically, the controller extracts a set of cell number fields within the component's value field from the target component index field, and reads the corresponding net benefit score field and cleaning instruction parameter field from the value field to form a candidate action set field within the component. The controller sorts the candidate action set field within the component according to the net benefit score field to generate a cell access order field, and writes the cell access order field into the sequence number field of the local path action sequence. To ensure that the local path action sequence meets the robot's moving body control constraints, the controller performs boundary constraint filtering on the cell access order field based on the component boundary vertex field of the component geometry and the prohibited area field and edge buffer distance field of the safety boundary. Cell number fields falling into the prohibited area field or exceeding the boundary are written into the set to be reviewed, and the reason for review is recorded as boundary conflict. Simultaneously, a rejection flag is written into the corresponding net benefit score field. The controller further performs action segment splicing processing on the cell access sequence field based on the robot's minimum turning radius and maximum acceleration parameters. The action segment splicing processing converts the spatial relative position of adjacent cell number fields into a local path segment record field, and binds the local path segment record field with the cleaning instruction parameter field to form an action sequence element field. The action sequence element field includes at least a cell number field, a sequence number field, a local path segment record field, and a cleaning instruction parameter field. An estimated duration field and an estimated resource consumption field are appended to each action sequence element field. The estimated duration field and the estimated resource consumption field are respectively mapped from the duration field, water consumption field, and electricity consumption field in the cleaning instruction parameter field.

[0065] To ensure the execution plan operates in a closed loop under resource budget parameter constraints, the controller performs budget consistency verification simultaneously when generating local path action sequences. Specifically, the controller accumulates the estimated budget usage field for each component in the priority queue according to the queue number field, forming cumulative water usage, cumulative electricity usage, and cumulative time usage fields. These fields are then compared with the water and electricity quota parameters and the current operation time window parameters in the resource budget parameters. When the cumulative water usage, cumulative electricity usage, or cumulative time usage fields trigger an over-limit condition field, the controller writes the over-limit condition field into the budget over-limit flag field of the execution plan and performs truncation processing on the local path action sequence of the current component index field. The truncation processing is accomplished by increasing the truncation threshold field of the cell access order field. Simultaneously, the truncated cell number field is written into the pending review set, and the reason for pending review is recorded as budget truncation. Furthermore, the controller reads the risk cost item record field from the net benefit score record and performs review quota queuing processing on the cell number field with low confidence value field. It maps the review sampling quota parameter to the review queuing upper limit field and adds a review pre-mark field to the queued cell number field in the local path action sequence. This ensures that the subsequent S600 motion command and cleaning command issuance links call the review sampling link and write to the execution log before executing the action sequence element field.

[0066] In one embodiment, after the robot is in the photovoltaic array channel and completes the value field generation of S400, the controller determines the reachable component set field based on the channel connectivity field of the array topology, and generates a component priority queue only for the reachable component set field. When the positioning error threshold field triggers the positioning anomaly flag field, the controller writes the corresponding component index field into the set to be reviewed and records the reason field to be reviewed as a positioning anomaly. It then performs a weight reduction registration on the component aggregation score field of the component index field, writes the weight reduction registration into the component score record field, and subsequently reorders the component priority queue and replaces the local path action sequence of the affected components. In this embodiment, the automatic update trigger condition for the execution plan is recorded as a change in the set to be reviewed or a change in the three-graph result. The trigger timestamp field and trigger reason field are archived in the plan session record field, enabling subsequent execution logs and anomaly classification result links to have a traceable plan version association.

[0067] The output of this step is an execution plan, which includes a component priority queue and local path action sequences. The controller writes the component priority queue into the component priority queue field of the execution plan, writes the local path action sequences into the local path action sequence field of the execution plan, and appends a session number field, a budget overrun flag field, a truncation threshold field, and a pre-review flag field to form a plan output package. The plan output package is called as input to S600. The component priority queue field in S600 is used to determine the order of component-level motion instructions, and the local path action sequence field in S600 is used to generate corresponding motion instructions and cleaning instructions and write them to the execution log. In summary, this step generates an execution plan containing a component priority queue and local path action sequences by connecting the component aggregation sorting and cell sorting links driven by the value field, and forms a directly deployable plan output package under resource budget parameters and safety boundary constraints.

[0068] S600: Based on the execution plan, issue motion instructions and cleaning instructions, and generate an execution log;

[0069] This step is executed online by the photovoltaic cleaning robot controller during the work cycle. The controller takes the execution plan output by S500 as direct input and reads the array topology, component geometry, safety boundary, robot parameters, and resource budget parameters from the session configuration package output by S100 as constraint input. At the same time, it reads the review reason field and cell number field from the review set formed by S200 as review linkage input. The execution plan includes a component priority queue field and a local path action sequence field, which are bound to the session number field. The component priority queue field includes a component index field, a queue number field, a component budget constraint flag field, and a budget occupancy estimate field. The local path action sequence field includes a cell number field, a sequence number field, a local path segment record field, a cleaning instruction parameter field, an estimated duration field, an estimated resource consumption field, and a review pre-mark field. The cleaning instruction parameter field includes a water consumption field, an electricity consumption field, and a duration field, and is consistent with the cleaning instruction parameter field in the action segment index table. Before issuing commands, the controller verifies the consistency between the session number field of the execution plan and the session number field of the session configuration package, and writes the verification result to the session consistency flag field of the execution log. At the same time, it reads the water quota parameter, electricity quota parameter, current operation time window parameter, and review sampling quota parameter from the resource budget parameters and initializes the budget ledger field. The budget ledger field includes the cumulative water usage field, cumulative electricity usage field, cumulative time usage field, and review used quota field. The budget ledger field serves as the status input for the closed-loop operation within this step and is associated with subsequent instruction gating processing.

[0070] Specifically, the controller schedules target component index fields one by one according to the queue number field of the component priority queue field. When scheduling each target component index field, the controller resolves the channel connectivity field and component adjacency relationship field in the array topology into reachability constraint fields, the component boundary vertex field in the component geometry into component range constraint fields, and the prohibited area field, edge buffer distance field, and turn restriction field in the safety boundary into safety constraint fields. The reachability constraint fields, component range constraint fields, and safety constraint fields are then loaded together into a path gating configuration field. Subsequently, the controller extracts the action sequence element field belonging to the target component index field from the local path action sequence field, and generates paired delivery queues of motion instructions and cleaning instructions in sequence according to the sequence number field. The paired delivery queues include motion start timestamp field, cleaning start timestamp field, and action completion timestamp field in the time dimension, and target cell number field and local path segment record field in the spatial dimension. The motion command is obtained by the controller mapping the local path segment record field with the maximum speed parameter, maximum acceleration parameter, and minimum turning radius parameter in the robot parameters. The mapping result is written into the speed level field, acceleration level field, and turning constraint field of the motion command. The cleaning command is obtained by the controller directly writing the cleaning command parameter field in the action sequence element field into the water consumption field, electricity consumption field, and duration field of the cleaning command, and writing the target cell number field into the cell number field of the cleaning command, so that the command object of the cleaning actuator is consistent with the cell scope of the value field. To meet safety constraints, the controller performs boundary checks on the local path segment record field before generating motion commands. The boundary check determines collisions between the local path segment record field and the prohibited area field and edge buffer distance field, and writes the result to the boundary check flag field of the motion command. When the boundary check flag field triggers the out-of-bounds condition field, the controller writes the corresponding target cell number field to the out-of-bounds blocking field of the execution log, moves the action sequence element field to the set to be reviewed and records the reason field to be reviewed as boundary blocking. At the same time, the controller performs pause gating on the subsequent action sequence element field of the target component index field and writes the pause reason field to the execution log.

[0071] Furthermore, the controller performs budget gating during the queue execution process. Budget gating takes the budget ledger field as input and the parameter field of each issued cleaning instruction as incremental input. Specifically, before issuing each cleaning instruction, the controller writes the water consumption, electricity consumption, and duration fields from the cleaning instruction parameter field into the estimated resource consumption field. It then compares the estimated resource consumption field with the water consumption quota, electricity consumption quota, and current operation time window parameters in the resource budget parameters for threshold comparison. When the threshold comparison triggers an over-limit condition field, the controller writes the over-limit condition field into the budget over-limit flag field of the execution log and performs instruction suppression processing on the current target cell number field. This instruction suppression process sets the water consumption, electricity consumption, and duration fields of the cleaning instruction to zero and writes the suppression flag field into the cleaning instruction, thus creating an auditable suppression record at the instruction level. Simultaneously, the controller writes the current target cell number field into the pending review set and records the pending review reason field as budget suppression. If the threshold comparison does not trigger the over-limit condition field, the controller updates the budget ledger field, adds the cumulative water usage field, cumulative electricity usage field, and cumulative time usage field to the water usage field, electricity usage field, and duration field respectively, and writes the updated budget ledger field snapshot to the budget snapshot field of the execution log. The budget snapshot field is associated with the session number field and the target cell number field, thereby supporting the subsequent S700 abnormal classification results to trace back to the specific resource usage status.

[0072] In the linkage link of command issuance, the controller performs review gating processing on the review pre-review flag field. Specifically, when the review pre-review flag field of the action sequence element field is in the triggered state, the controller reads the review sampling quota parameter in the resource budget parameter and performs quota discrimination on the review used quota field in the budget ledger field; when the quota discrimination meets the review trigger condition field, the controller starts the vision acquisition unit to generate a review image frame and writes the review image frame into the set to be reviewed, while incrementing the review used quota field and writing the review trigger record field into the execution log; when the quota discrimination does not meet the review trigger condition field, the controller writes the review not executed flag field into the execution log and continues to issue motion commands and cleaning commands. Understandably, the minimum set of fields on which the review gating processing depends includes the review pre-review flag field, the review sampling quota parameter, and the review used quota field. The review image frame and the review trigger record field constitute an extended set of fields for connection with the S700 verification image sequence acquisition link.

[0073] In one embodiment, the robot enters the target component index field range along the channel connectivity field defined by the array topology. The controller generates motion commands for each target cell number field according to the local path action sequence field and sends them to the drive actuator via the motion control unit. At the same time, a cleaning command is sent to the cleaning actuator. The pose feedback record field returned by the drive actuator and the execution status record field returned by the cleaning actuator are collected by the controller and written to the execution log. The execution status record field includes a command reception timestamp field, a command start timestamp field, a command end timestamp field, an execution completion flag field, and an abnormal stop flag field. When the abnormal stop flag field triggers the stop condition field, the controller writes the current target cell number field to the abnormal stop position field of the execution log, writes the stop condition field to the stop reason field of the execution log, and performs stop gating processing on the remaining action sequence element field of the current target component index field and writes it to the gating status field of the execution log. To ensure closed-loop auditability, the controller generates an instruction sequence number field for each pair of dispatch queues. The instruction sequence number field is bound to the session number field, component index field, and target cell number field. The execution log records the motion instruction field, cleaning instruction field, boundary check flag field, budget overrun flag field, suppression flag field, execution completion flag field, and abnormal shutdown flag field, thus enabling the execution log to have a complete record of the link from planning to instruction to execution.

[0074] The output of this step is an execution log. This execution log is generated by the controller by archiving the aforementioned instructions and feedback information in chronological order. The execution log includes at least the following fields: session number, component index, target cell number, instruction serial number, motion instruction, cleaning instruction, execution status record, budget snapshot, and pending review linkage record. Specifically, the water usage, electricity consumption, and duration fields in the cleaning instruction field are used in subsequent step S700 as linkage inputs for verifying the differences between the three graph results and the actual results. The execution status record field is used in subsequent step S700 to generate the time alignment benchmark for the compliance judgment and non-compliance set. The pending review linkage record field is used in subsequent step S700 to generate a local re-cleaning task and match it with the abnormal classification result trigger field in the classification rule table. In summary, this step achieves the following technical effect: Based on the execution plan, a traceable distribution chain of motion and cleaning instructions is formed under safety and resource budget constraints, and an execution log containing the instruction serial number and budget snapshot fields is generated for subsequent verification and abnormal classification chain invocation.

[0075] S700: Based on the execution log, collect the verification image sequence to generate a verification three-image result, compare the verification three-image result with the three-image result to generate a pass / fail set, generate a local re-cleaning task and generate an abnormal classification result according to the classification rule table;

[0076] This step is executed online by the photovoltaic cleaning robot controller after issuing motion and cleaning commands from S600. The controller uses the execution log output by S600 as the main input, and reads the segmentation rules, mapping table, exposure strategy table, action fragment index table, and classification rule table from the session configuration package output by S100 as configuration input. At the same time, it reads the type map, severity map, and confidence map from the three-map results output by S300 as comparison benchmark input. The execution log includes at least the session number field, component index field, cell number field, command serial number field, motion command field, cleaning command field, execution status record field, budget snapshot field, and re-cleaning number field. The execution status record field includes the command start timestamp field and the command end timestamp field and is associated with the cell number field. Before entering the verification link, the controller performs a consistency check between the session number field in the execution log and the session number field in the session configuration package, and writes the verification result to the session consistency flag field in the execution log. If the consistency check fails, the controller writes the corresponding instruction sequence number field to the verification skip flag field in the execution log and ends the verification branch corresponding to the instruction sequence number field, so that the subsequent verification image sequence and the verification three-image result are all bound to the same session number field.

[0077] Specifically, the controller triggers verification acquisition based on the instruction end timestamp field of the execution log. The trigger window is jointly determined by the instruction end timestamp field of the execution log and the switching condition field of the exposure strategy table, and the trigger window is written to the verification trigger window field of the execution log. Within the trigger window, the controller drives the visual probe to complete the acquisition of the verification image sequence. The visual probe parameters come from the session configuration package and include field of view parameters, installation extrinsic parameters, and sampling parameter fields. The sampling parameter field is loaded from the exposure field, gain field, and shutter field in the exposure strategy table. During the acquisition process, the controller synchronously reads the motion instruction field and execution status record field of the execution log, generates a verification pose index field that corresponds one-to-one with the verification image sequence, and writes it to the verification pose index field of the execution log. When the switching condition field of the exposure strategy table is triggered, the controller switches the sampling parameter field within the same trigger window and writes the switching record to the exposure switching record field of the execution log. To suppress glare interference, the controller performs glare suppression processing frame by frame on the verification image sequence. This glare suppression process creates a glare mask field within the pixel domain and uses it to compensate and fill saturated areas, as well as perform local contrast realignment. The glare percentage score field is then written to the quality score detail field in the execution log. Simultaneously, the controller performs quality scoring on the verification image sequence. This quality score is a combination of sharpness score, exposure saturation score, glare percentage score, and motion blur score fields, and is written to the quality score field in the execution log. When the quality score field is less than a quality threshold, the controller writes the corresponding image frame index field to the set to be reviewed and records the reason for the review as the verification quality threshold trigger. Simultaneously, it resamples the verification image sequence according to the exposure strategy table, switching the sampling parameter fields, and writes the number of resamplings to the verification resampling count field in the execution log. When the quality score field meets the quality threshold, the controller loads the qualified verification image sequence, along with the verification pose index field, into the inference input structure. The controller then aligns and binds the inference input structure with the pixel coordinate field and cell number field in the mapping table, ensuring that the verification image sequence maintains a consistent correspondence with the three-image results at the cell number field granularity.

[0078] In the chain of generating the verification three-image results, the controller calls the inference process corresponding to S300 and replaces its input with the verification image sequence. The inference process includes multi-scale feature extraction, scale fusion, and cell aggregation. Cell aggregation is completed according to the cell grid size field and cell number field in the segmentation rules. The controller aggregates the inference output according to the cell number field to generate the verification three-image results. The verification three-image results include a verification type image, a verification severity image, and a verification confidence image, and are associated one-to-one with the cell number field of the execution log. After generating the verification three-image results, the controller writes the verification type image label field, the verification severity image level field, and the verification confidence field into the verification three-image record field of the execution log, and writes the model version tag field and the mapping table version tag field referenced in the inference process into the version record field of the execution log, thereby forming an auditable association between the verification three-image results and the configuration items in the session configuration package.

[0079] The controller performs a difference comparison process. This process takes the verification results of the three graphs and aligns them with the cell number field as input. At the cell number field granularity, it calculates the type consistency judgment field, severity difference judgment field, and confidence consistency judgment field. Specifically, the type consistency judgment field is obtained by comparing the verification type graph with the type graph label field; the severity difference judgment field is obtained by comparing the verification severity graph with the severity graph level field; and the confidence consistency judgment field is obtained by comparing the verification confidence graph with the confidence graph. The controller synthesizes these judgment fields into a pass / fail judgment according to preset judgment rules. The pass / fail judgment includes a pass / fail flag field, a difference category field, and a difference magnitude field, and writes the pass / fail judgment to the execution log's pass / fail judgment field. When the pass / fail flag field is in a non-passing state, the controller writes the corresponding cell number field to a non-passing set. This non-passing set includes the cell number field, component index field, difference category field, difference magnitude field, and re-clearing count field, and is bound to the session number field. When constructing the set of non-compliance, the controller synchronously updates the consecutive non-compliance count field. The consecutive non-compliance count field is obtained by merging the historical compliance judgment fields with the same cell number field in the execution log and writing them into the execution log. The controller also obtains the confidence map low value persistence field based on the verification confidence map and the time-series segment statistics of the confidence map. The confidence map low value persistence field is written into the execution log and used as the input field for subsequent anomaly classification results.

[0080] In the process of generating a partial re-cleaning task, the controller extracts the cell number field from the non-compliant set as the re-cleaning location input, and maps the type map label field and severity map level field associated with the cell number field to the cleaning instruction parameter field according to the action fragment index table, forming the re-cleaning cleaning instruction parameter. The re-cleaning cleaning instruction parameter includes a water consumption field, an electricity consumption field, and a duration field, and its definition is consistent with the cleaning instruction parameter field in the action fragment index table. The controller writes the re-cleaning cleaning instruction parameter into the cleaning instruction field of the execution log and increments the re-cleaning count field. At the same time, it writes the corresponding instruction serial number field into the re-cleaning associated serial number field of the execution log, thereby obtaining the partial re-cleaning task. The partial re-cleaning task includes the cell number field, the re-cleaning cleaning instruction parameter, and the re-cleaning count field, and is bound to the session number field. The controller outputs the partial re-cleaning task as a cross-step connection. After the partial re-cleaning task is written to the execution log, it triggers the execution plan update trigger condition. The execution plan update trigger condition is consistent with the execution plan update trigger condition in the claim. The controller then fills the partial re-cleaning task back into the incremental update process of the execution plan of S500, thereby causing the component priority queue of S500 to be replaced with the local path action sequence. S600 reads the updated execution plan and continues to issue motion instructions and cleaning instructions to form a closed loop operation.

[0081] In the abnormal classification result generation chain, the controller uses the consecutive failure count field, the persistent low value field of the confidence graph, and the parameter upper limit trigger flag field of the cleaning instruction parameter field as the minimum input field set. The parameter upper limit trigger flag field is obtained by the controller comparing the water consumption, electricity consumption, and duration fields of the re-cleaning instruction parameters with the resource budget parameters in the session configuration package using thresholds, and writing the comparison results to the execution log. When the consecutive failure count field, the persistent low value field of the confidence graph, and the parameter upper limit trigger flag field satisfy the abnormal classification result trigger field of the classification rule table, the controller determines that an abnormal classification result has been obtained and writes the abnormal classification result to the abnormal classification result field in the execution log. Simultaneously, the controller reads the strategy number field of the classification rule table and generates a degradation control instruction or remote alarm instruction according to the strategy number field, writing it to the strategy execution record field in the execution log. The strategy execution record field is associated with the session number field, component index field, cell number field, and instruction serial number field, thereby forming a classification chain record for the same cell number field. To maintain the stability of automated operation, the controller synchronously writes the version flag fields of the classification rule table and the exposure strategy table to the version record field of the execution log each time it writes the abnormal classification result field. When the version flag field changes, the version change record is written to the version change record field of the execution log, so that subsequent backtracking and comparison of abnormal classification results are completed within the version boundary.

[0082] The output products of this step include the verification three-graph result, the compliance determination, the non-compliance set, the partial re-clearing task, and the anomaly classification result. The verification three-graph result is fixed through the verification three-graph record field of the execution log, the compliance determination is fixed through the compliance determination field of the execution log, the non-compliance set is fixed through the non-compliance set field of the execution log, the partial re-clearing task is fixed through the re-clearing association serial number field and the re-clearing count field of the execution log, and the anomaly classification result is fixed through the anomaly classification result field and the policy execution record field of the execution log. The partial re-clearing task serves as the incremental update input for the S500 execution plan and is regenerated in the execution log via the S600 instruction delivery link. The anomaly classification result serves as the output record of the remote alarm instruction and is archived in association with the subsequent session number field. In summary, this step achieves the following technical effect: under the constraints of the execution log, it forms aligned records of the verification three-graph result and the compliance determination, and establishes a closed-loop archiving link for the non-compliance set, the partial re-clearing task, and the anomaly classification result.

Claims

1. A control method for a photovoltaic cleaning robot based on vision recognition, characterized in that, include: The array topology, component geometry, safety boundaries, robot parameters, visual probe parameters, and resource budget parameters are obtained to generate a session configuration package; the session configuration package includes segmentation rules, a mapping table, an exposure strategy table, an action fragment index table, and a classification rule table; Image sequences and pose sequences are acquired based on session configuration packets, and alignment, reflection suppression and quality scoring are performed to generate alignment data packets and a set to be reviewed. Based on the aligned data packets and mapping table, inference is performed to generate a three-graph result; the three-graph result includes a type graph, a severity graph, and a confidence graph, and the set to be reviewed is updated according to the confidence graph; Based on the results of the three graphs, the action fragment index table, and the resource budget parameters, candidate parameter groups are generated, and the benefit items, resource cost items, and risk cost items are calculated to generate net benefit score records and value fields. An execution plan is generated based on the value field, and the execution plan includes a component priority queue and a local path action sequence; Based on the execution plan, motion commands and cleaning commands are issued, and an execution log is generated; Based on the execution log, the image sequence is collected to generate a three-image verification result. The three-image verification result is compared with the three-image result to generate a set of compliance and non-compliance. A local re-cleaning task is generated and an abnormal classification result is generated according to the classification rule table.

2. The method according to claim 1, characterized in that, The segmentation rules in the session configuration package include cell grid size and cell number fields; the mapping table includes pixel coordinate and cell number fields; the exposure strategy table includes exposure, gain, shutter speed, and switching condition fields; the motion fragment index table includes type map label, severity map level, and cleaning instruction parameter fields; and the classification rule table includes abnormal classification result trigger and strategy number fields.

3. The method according to claim 1, characterized in that, The quality score is obtained by combining the sharpness score, exposure saturation score, reflection ratio score, and motion blur score. When the quality score is less than the quality threshold, the corresponding image frame is written into the set to be reviewed and the sampling parameters are switched according to the exposure strategy table to re-acquire the image sequence.

4. The method according to claim 1, characterized in that, The reasoning process based on aligned data packets and mapping tables includes: The inference process includes multi-scale feature extraction, scale fusion, and cell aggregation; the type map is composed of the category labels after cell aggregation, the severity map is composed of the level values ​​after cell aggregation, and the confidence map is synthesized from the consistency score and quality score of the multi-scale output.

5. The method according to claim 1, characterized in that, The candidate parameter group is obtained by indexing the action segment index table according to the type map label and severity map level; the benefit item is calculated by the severity map level and cell area field, the resource cost item is calculated by the water consumption field, electricity consumption field and duration field in the candidate parameter group, and the risk cost item is calculated by the confidence map and the set to be reviewed; the net benefit score record is calculated by the benefit item, resource cost item and risk cost item according to the linear composition rule.

6. The method according to claim 2, characterized in that, The component priority queue is obtained by aggregating and sorting the net benefit score records of the value field within the component scope; the local path action sequence is generated by sorting the cell net benefit score records of the value field within the component scope, and the cleaning instruction parameter field in the candidate parameter group is bound to the sorted cell.

7. The method according to claim 1, characterized in that, The execution plan update is triggered by changes in the set to be reviewed or changes in the three graph results; the incremental update of the execution plan includes component priority queue rearrangement and replacement of local path action sequences within the component.

8. The method according to claim 2, characterized in that, The anomaly classification result is determined by the consecutive failure count field, the persistent low value field of the confidence graph, and the parameter upper limit trigger flag field of the cleaning instruction parameter field; the strategy number field generates a degradation control instruction or a remote alarm instruction and writes it to the execution log.

Images