CN-122008267-A - Control method of photovoltaic cleaning robot based on visual recognition

CN122008267ACN 122008267 ACN122008267 ACN 122008267ACN-122008267-A

Abstract

The invention relates to the field of control of photovoltaic cleaning robots, in particular to a control method of a photovoltaic cleaning robot based on visual identification. The method comprises the steps of constructing a session configuration package comprising array topology, component geometry, safety boundary, robot parameters, vision probe parameters and resource budget parameters, driving collaborative collection and processing of an image sequence and a pose sequence, forming an alignment data package and a set to be rechecked on the basis of finishing alignment, reflection inhibition and quality scoring, reasoning by combining a mapping relation, generating a multidimensional result reflecting dirt category, severity and confidence, constructing candidate cleaning parameters and a value field according to the multidimensional result, generating an execution plan, completing issuing of a motion and cleaning instruction, recording an execution log, and finishing standard reaching judgment, local rechecking and abnormal parting on the basis of a verification result. The invention improves the consistency and rechecking performance of the cleaning decision and execution through the closed-loop control flow of the through acquisition, reasoning, decision and verification.

Inventors

LI HAO
ZHANG CHUNLIN
LI JUNHONG
YAN SHENGLI
LIU MENG
WANG YIZHU
ZHOU XINGYAN
Xiao Dongcai

Assignees

广安职业技术学院
成都远程巨科科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (8)

1. A control method of a photovoltaic cleaning robot based on visual recognition, comprising: Acquiring array topology, component geometry, safety boundary, robot parameters, vision probe parameters and resource budget parameters to generate a session configuration package, wherein the session configuration package comprises a segmentation rule, a mapping table, an exposure strategy table, an action fragment index table and a parting rule table; based on the session configuration package, acquiring an image sequence and a pose sequence, and executing alignment, reflection inhibition and quality scoring to generate an alignment data package and a set to be rechecked; Based on the aligned data packet and the mapping table, reasoning is carried out to generate a three-graph result, wherein the three-graph result comprises a type graph, a severity graph and a confidence graph, and a set to be rechecked is updated according to the confidence graph; Generating candidate parameter sets based on the three-graph result, the action fragment index table and the resource budget parameters, calculating a benefit item, a resource cost item and a risk cost item, and generating a net benefit score record and a value field; generating an execution plan based on the value field, the execution plan including a component priority queue and a sequence of local path actions; generating an execution log based on the execution plan issuing motion instruction and the cleaning instruction; And acquiring a verification image sequence based on the execution log to generate a verification three-graph result, comparing the verification three-graph result with the three-graph result to generate a standard-reaching judgment and non-standard-reaching set, generating a local re-clearing task, and generating an abnormal parting result according to a parting rule table.
2. The method of claim 1, wherein the segmentation rules in the session configuration package include a cell grid size field and a cell number field, the mapping table includes a pixel coordinate field and a cell number field, the exposure policy table includes an exposure field, a gain field, a shutter field, and a switch condition field, the action segment index table includes a type map label field, a severity map level field, and a cleaning instruction parameter field, and the typing rule table includes an abnormal typing result trigger field and a policy number field.
3. The method of claim 1, wherein the quality score is derived from a combination of sharpness score, exposure saturation score, reflectance fraction, and motion blur score, and wherein when the quality score is less than a quality threshold, the corresponding image frame is written into the set to be reviewed and the sampling parameters are switched according to an exposure policy table to re-capture the image sequence.
4. The method of claim 1, wherein the process of reasoning based on the alignment data packet and the mapping table comprises: The reasoning comprises multi-scale feature extraction, scale fusion and cell aggregation, wherein the type graph consists of class labels after cell aggregation, the severity graph consists of level values after cell aggregation, and the confidence graph is obtained by synthesizing a multi-scale output consistency score and a quality score.
5. The method of claim 1, wherein the candidate parameter set is obtained by indexing an action segment index table according to a type graph label and a severity graph level, a benefit item is obtained by calculating a severity graph level and a cell area field, a resource cost item is obtained by calculating a water field, a power field and a duration field in the candidate parameter set, a risk cost item is obtained by calculating a confidence graph and a set to be rechecked, and a net benefit score record is obtained by calculating a benefit item, a resource cost item and a risk cost item according to a linear synthesis rule.
6. The method of claim 2, wherein the component priority queue is derived from aggregate ordering of net gain score records for the value field within the component range, wherein the local path action sequence is generated from ordering of net gain score records for the value field within the component range, and wherein the clean instruction parameter fields in the candidate parameter set are bound for the ordered cells.
7. The method of claim 1, wherein the execution plan update trigger condition is a change in a set to be reviewed or a change in a three-graph result, and wherein the execution plan delta update includes a component priority queue reordering and a component-wide local path action sequence replacement.
8. The method of claim 2, wherein the exception type result is determined by a continuous number of times substandard field, a confidence map low value duration field, and a parameter upper limit trigger flag field of a cleaning instruction parameter field, and wherein a policy number field generates a degradation control instruction or a remote alert instruction and writes to an execution log.

Description

Control method of photovoltaic cleaning robot based on visual recognition Technical Field The invention relates to the field of control of photovoltaic cleaning robots, in particular to a control method of a photovoltaic cleaning robot based on visual identification. Background In the field of control of photovoltaic cleaning robots, the existing scheme of the control method of the photovoltaic cleaning robots generally carries out task arrangement around an image sequence and a pose sequence, completes issuing of a motion instruction and a cleaning instruction by combining robot parameters and vision probe parameters, and has the limitations that a session configuration package is not enough in constraint association with the image sequence and the pose sequence, a unified rechecking closed loop is lacking in the alignment and reflection inhibition and quality scoring process, and data receiving between an inference product and a subsequent control link is loose. The existing method is used for completing acquisition, judgment and action generation by relying on a single link, and the problems of insufficient source consistency of an aligned data packet, lack of traceable basis for updating a set to be checked and the like easily occur under the scene of common constraint of a safety boundary and a resource budget parameter, so that the connection between a three-graph result and an execution plan is unstable, and the flow consistency requirement of generating the execution plan around the three-graph result and issuing a motion instruction and a cleaning instruction is difficult to meet. Aiming at the combined processing of an aligned data packet, a mapping table and reasoning, the prior art generally has the defects that the organization caliber of a type chart, a severity chart and a confidence chart is not uniform, the corresponding relation between an action fragment index table and a candidate parameter group lacks stable mapping, the record granularity of a benefit item, a resource cost item and a risk cost item is insufficient, and the like, and the consistency flow of acquisition, alignment, reasoning, scoring, planning, issuing, journaling and verification is difficult to form in the application scene of the photovoltaic cleaning robot control method, so that the correlation link between an execution journal and a verification three-chart result is incomplete, and further the generation process of standard reaching judgment, standard reaching failure collection, local re-cleaning task and abnormal parting result lacks stable data basis and consistent operation constraint is further caused. Disclosure of Invention In order to solve the technical problems, the invention provides a control method of a photovoltaic cleaning robot based on visual identification, which comprises the following steps: Acquiring array topology, component geometry, safety boundary, robot parameters, vision probe parameters and resource budget parameters to generate a session configuration package, wherein the session configuration package comprises a segmentation rule, a mapping table, an exposure strategy table, an action fragment index table and a parting rule table; based on the session configuration package, acquiring an image sequence and a pose sequence, and executing alignment, reflection inhibition and quality scoring to generate an alignment data package and a set to be rechecked; Based on the aligned data packet and the mapping table, reasoning is carried out to generate a three-graph result, wherein the three-graph result comprises a type graph, a severity graph and a confidence graph, and a set to be rechecked is updated according to the confidence graph; Generating candidate parameter sets based on the three-graph result, the action fragment index table and the resource budget parameters, calculating a benefit item, a resource cost item and a risk cost item, and generating a net benefit score record and a value field; generating an execution plan based on the value field, the execution plan including a component priority queue and a sequence of local path actions; generating an execution log based on the execution plan issuing motion instruction and the cleaning instruction; And acquiring a verification image sequence based on the execution log to generate a verification three-graph result, comparing the verification three-graph result with the three-graph result to generate a standard-reaching judgment and non-standard-reaching set, generating a local re-clearing task, and generating an abnormal parting result according to a parting rule table. Further, the segmentation rule in the session configuration packet includes a cell grid size field and a cell number field, the mapping table includes a pixel coordinate field and a cell number field, the exposure policy table includes an exposure field, a gain field, a shutter field and a switching condition field, the action fragment index table includes a type map t