Multi-detection standard adaptive cognitive parameter self-optimization method and embodied unmanned system

Abstract

The application relates to the field of intelligent unmanned system sensing control, and provides a multi-detection standard adaptation cognitive parameter self-optimization method and a body-equipped unmanned system. A target detection standard and a historical detection record are acquired, and a standard difference vector is constructed; multi-modal detection data are collected, current detection object features are extracted, cognitive parameter adjustment constraints and body-equipped execution parameter allowable ranges are determined; a source detection standard is selected to calculate a cognitive parameter migration weight, and historical cognitive parameters are migrated and transformed to obtain candidate cognitive parameters; the unmanned system is controlled to execute standard adaptation calibration actions according to the body-equipped execution parameter allowable ranges, feedback data are collected, and a standard adaptation loss is calculated; and the target standard adaptation cognitive parameters are self-optimized in the constraints with the target of reducing the adaptation loss, and are loaded into a defect detection module to execute tasks. Corresponding body-equipped detection unmanned systems are also provided. The application realizes cross-standard cognitive parameter automatic migration and closed-loop self-optimization, and significantly improves standard switching efficiency and detection result consistency.

Description

Technical Field

[0001] This invention relates to the field of perception and control of intelligent unmanned systems, specifically to a method for self-optimization of cognitive parameters of embodied unmanned systems with multi-detection standard adaptation. Background Technology

[0002] In fields such as industrial inspection and infrastructure operation and maintenance, embodied inspection unmanned systems are often used to perform defect identification and condition assessment tasks. With the diversification of application scenarios, unmanned systems often need to switch between different inspection standards, such as from national standards to internal enterprise acceptance standards, or from general industry standards to project-specific customized standards.

[0003] In existing technologies, when detection standards change, technicians typically need to readjust and calibrate the visual acquisition parameters, motion control parameters, and algorithm recognition thresholds of the unmanned system according to the rules of the new standard. This offline debugging method, which relies on human experience, is not only time-consuming but also makes it difficult to ensure that the parameter configurations are accurately matched with the new standard's rules, resulting in low detection efficiency and poor consistency of results for the unmanned system after the standard switch.

[0004] Therefore, there is an urgent need for a technical solution that can automatically and quickly adapt cognitive parameters when switching detection standards. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of low parameter debugging efficiency and poor adaptation accuracy when switching detection standards in the existing technology, and to provide a self-optimization method for cognitive parameters of embodied unmanned systems that adapts to multiple detection standards.

[0006] The objective of this invention can be achieved through the following technical solutions: As a first aspect of the present invention, a method for self-optimization of cognitive parameters adapted to multiple detection standards is provided, comprising the following steps: Acquire the target detection standard corresponding to the current defect detection task and the historical detection records formed by the embodied unmanned system in historical defect detection tasks; Construct a standard difference vector based on the target detection standard and historically applicable detection standards; Collect multimodal detection data of the currently detected object from the embodied unmanned system, and extract the features of the currently detected object; Based on the standard difference vector and the characteristics of the current detection object, the cognitive parameter adjustment constraints corresponding to the target detection standard are determined, and the allowable range of the embodied execution parameters associated with the cognitive parameter adjustment constraints is further determined. Based on historical testing records, source testing standards are selected from historically applicable testing standards that have a standard mapping relationship with the target testing standard. The cognitive parameter migration weight from the source testing standard to the target testing standard is calculated according to the standard difference vector. The historical cognitive parameters corresponding to the source testing standard are then transformed according to the cognitive parameter migration weight to obtain candidate cognitive parameters under the target testing standard. The control system performs standard adaptation calibration actions within the allowable range of the embodied execution parameters and collects calibration feedback data. Based on the calibration feedback data, the standard adaptation loss of candidate cognitive parameters under the target detection standard is calculated. With the goal of reducing the standard adaptation loss, the candidate cognitive parameters are self-optimized within the cognitive parameter adjustment constraints to obtain the target standard-adapted cognitive parameters. The embodied unmanned system loads the target standard adaptation cognitive parameters and performs the current defect detection task based on the target standard adaptation cognitive parameters.

[0007] As a preferred technical solution, the target detection standard includes defect category definition, defect measurement rules, defect level classification rules, and detection result output rules; The historical testing records include historical applicable testing standards, historical testing object attributes, historical cognitive parameters, historical defect identification results, and historical defect review results; The cognitive parameters include at least one of the following: defect category identification parameters, defect boundary understanding parameters, size measurement reasoning parameters, defect level determination parameters, detection standard mapping parameters, re-inspection decision parameters, and detection result expression parameters.

[0008] As a preferred technical solution, the standard difference vector is used to characterize the differences between different detection standards in defect category definition, defect measurement benchmark, defect level threshold, and detection result expression method. The construction process is as follows: The target testing standard and the historically applicable testing standard are parsed separately to obtain the target standard rule description and the historical standard rule description. Extract the defect category field, defect measurement field, defect level field, review requirement field, and test result output field from the target standard rule description and the historical standard rule description respectively; Establish mapping relationships between target testing standards and historically applicable testing standards for defect categories, defect measurement benchmarks, defect level thresholds, and test result expression. The differences between the target testing standard and the historically applicable testing standard in terms of defect category mapping, defect measurement benchmark, defect level threshold, review requirements, and test result expression are calculated to obtain the standard difference vector.

[0009] As a preferred technical solution, the cognitive parameter adjustment constraints include: standard mapping constraints, defect identification constraints, boundary understanding constraints, size measurement reasoning constraints, defect level determination constraints, and re-inspection decision constraints; The standard mapping constraint is used to limit the mapping relationship between different detection standards in terms of defect categories, defect measurement rules, grade thresholds, and result expression methods. The re-inspection decision constraints are used to limit the confidence range, measurement error range, grade boundary distance range, and multi-frame recognition consistency range for triggering re-inspection actions in suspected defect areas.

[0010] As a preferred technical solution, the step of determining the cognitive parameter adjustment constraint corresponding to the target detection standard based on the standard difference vector and the current detection object features is specifically implemented as follows: When the standard difference vector characterizes the target detection standard's requirement for defect size measurement accuracy is higher than that of historically applicable detection standards, the accuracy constraint of the size measurement inference parameters is increased, the threshold tolerance of the defect level judgment parameters is reduced, and the re-inspection trigger range of suspected defect areas is expanded. When the standard difference vector represents the requirement of the target detection standard for the integrity of the defect boundary higher than that of the historically applicable detection standard, the integrity constraint of the defect boundary understanding parameters is increased, and the adjustment step size of the defect boundary judgment threshold is reduced. When the standard difference vector characterizes the defect level classification granularity of the target detection standard, which is higher than that of the historically applicable detection standards, the search step size of the size measurement inference parameters is reduced, and the number of local depth measurements is increased. When the surface reflection, occlusion, or texture interference of the current object being detected exceeds the preset interference threshold, the re-inspection decision parameters and the associated supplementary lighting intensity, shooting distance, and allowable variation range of sensor posture are adjusted.

[0011] As a preferred technical solution, the step of selecting the source detection standard from historically applicable detection standards that have a standard mapping relationship with the target detection standard is specifically implemented as follows: Obtain the historical standard rule descriptions corresponding to multiple historically applicable testing standards; Calculate the standard rule similarity between the target detection standard and each historically applicable detection standard; Based on the similarity of standard rules, the results of historical defect verification, and the attributes of historical test objects, the source standard credibility of each historically applicable test standard is calculated; the source standard credibility is obtained by weighted fusion of standard rule similarity, parameter credibility, and environmental compatibility. The standard rule similarity is the similarity between the semantic embedding vector of the historical standard rule description of the target detection standard and the semantic embedding vector of the historical standard rule description of the currently applicable detection standard. The reliability of the parameter is the ratio of the number of records that passed the review to the total number of test records under the corresponding historical applicable test standard in all historical defect review results. The environmental compatibility refers to the ratio of the number of attribute fields that match the attribute fields of historically applicable detection standards for historically tested objects to the corresponding attribute fields in the current detection object's features, out of the total number of attribute fields. Historically applicable testing standards whose source standard credibility meets the preset credibility threshold are identified as source testing standards.

[0012] As a preferred technical solution, the calculation of the cognitive parameter transfer weight from the source detection standard to the target detection standard based on the standard difference vector is specifically implemented as follows: Calculate the standard domain difference between the source detection standard and the target detection standard based on the standard difference vector; The reliability of parameters corresponding to historical cognitive parameters of the source detection standard is calculated based on the historical defect review results. The parameter reliability calculation process is as follows: extract all historical defect review results corresponding to the source detection standard from the historical detection records, calculate the single reliability score for each review result, and take a weighted average of all single reliability scores to obtain the parameter reliability. The calculation method for the single-item reliability score is as follows: if the defect identification result of the inspection record is completely consistent with the review conclusion, the single-item reliability score is set as the first score value; if the defect category is consistent but the defect size deviation is within the standard allowable tolerance range, the single-item reliability score is set as the second score value; if the defect category is consistent but the defect level judgment deviation is level one, the single-item reliability score is set as the third score value; if the defect category is inconsistent, the single-item reliability score is 0; the first score value, the second score value, and the third score value decrease in that order. Calculate the similarity between the detected objects and the characteristics of the current detected objects based on the attributes of the historical detected objects; Based on the relationship that is negatively correlated with the difference in the standard domain and positively correlated with the reliability of parameters and the similarity of the detected objects, the cognitive parameter transfer weights from the source detection standard to the target detection standard are calculated. The cognitive parameter transfer weights are obtained by nonlinearly fusing parameter confidence, similarity of the detected objects, and difference in the standard domain. The reliability of the parameters is incorporated into the fusion process in the form of a power function. The similarity of the detected objects is fused in the form of a power function, which is the ratio of the number of attribute fields that match the attributes of historical detected objects and the attributes of current detected objects to the total number of attribute fields. The standard domain difference quantity participates in the fusion in the form of a negative exponential function, which is the ratio of the L2 norm of the standard difference vector to the preset maximum difference norm reference value.

[0013] As a preferred technical solution, the step of performing a migration transformation on the historical cognitive parameters corresponding to the source detection standard based on the cognitive parameter migration weight to obtain candidate cognitive parameters under the target detection standard is specifically implemented as follows: The transfer sensitivity of various cognitive parameters is determined based on the standard difference vector. The specific calculation is as follows: a sensitivity correlation vector is preset for each type of cognitive parameter; the normalized standard difference vector and the sensitivity correlation vector of each type of cognitive parameter are respectively subjected to vector dot product operation to obtain the transfer sensitivity of the corresponding category of cognitive parameter. Based on the transfer weights and transfer sensitivity of cognitive parameters, transfer transformations are performed on the defect category identification parameters, defect boundary understanding parameters, dimensional measurement reasoning parameters, defect level determination parameters, inspection standard mapping parameters, re-inspection decision parameters, and inspection result expression parameters in the historical cognitive parameters, respectively. k The class cognitive parameters, and their transfer transformation expression are: in, For the transformed th k Class cognitive parameters; For the source detection standard k The original values ​​of the historical cognition parameters; For cognitive parameter transfer weights; For the first k Transfer sensitivity of cognitive parameters; The first one calculated based on the standard difference vector k For the standard adaptation adjustment of the cognitive parameters, for the size measurement inference parameters, the defect measurement benchmark difference component in the standard difference vector is mapped to the parameter offset of the measurement inference model; for the defect level judgment parameters, the defect level threshold difference component is mapped to the adjustment of the level judgment boundary; for the defect category identification parameters, the defect category mapping difference component is mapped to the weight adjustment of the category classifier; for the re-inspection decision parameters, the re-inspection requirement difference component is mapped to the threshold adjustment of the re-inspection triggering condition. The transformed cognitive parameters are projected onto the allowable range of parameters corresponding to the target detection standard to obtain candidate cognitive parameters.

[0014] As a preferred technical solution, the standard adaptation calibration procedure includes: When a suspected defective area meets the re-inspection trigger conditions, the control system reduces its travel speed and re-inspects the suspected defective area. The control system adjusts the pitch or rotation angle of the sensors to obtain multi-angle detection data of suspected defect areas; The control system adjusts the intensity of the supplementary light to reduce the impact of reflections, shadows, or occlusions on defect boundary identification; The control system repeatedly locates the suspected defect location based on mileage and pose data; the control system performs local depth measurement on the suspected defect area to obtain dimensional measurement data of the suspected defect area. The re-inspection triggering conditions include at least one of the following: the defect confidence level is in the preset gray zone, the defect size is close to the defect level threshold, the depth measurement fluctuation exceeds the preset fluctuation threshold, the reflection or occlusion exceeds the preset interference threshold, and the recognition results of multiple frames in the same area are inconsistent.

[0015] As a second aspect of the present invention, an embodied unmanned system is provided, including an airborne multimodal sensor, a defect detection module, a memory, and a processor. The airborne multimodal sensor is communicatively connected to the processor. The memory stores a computer program. When the processor executes the computer program, it implements the cognitive parameter self-optimization method for multi-detection standard adaptation as described above.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1) This invention constructs a standard difference vector that covers multidimensional differences in defect category definition, measurement benchmark, grade threshold and result expression method, transforming unstructured detection standard rules into computable structured representations. It also manages cognitive parameters and embodied execution parameters hierarchically, and calculates cognitive parameter transfer weights that take into account both objective differences between standards and subjective reliability of historical experience by combining the reliability of parameters verified in historical detection records. This allows unmanned systems to quickly obtain high-quality candidate cognitive parameter starting points without having to debug from scratch.

[0017] 2) This invention controls an unmanned system to perform physical-level standard adaptation calibration actions within the allowable range of specific execution parameters. It collects calibration feedback data from a real-world environment and calculates the standard adaptation loss from six dimensions: category mapping, dimensional measurement, grade classification, boundary integrity, result output, and action cost. This constructs a closed-loop self-optimization mechanism based on real physical measurements. This mechanism iteratively optimizes within the cognitive parameter adjustment constraints determined by the standard difference vector and the characteristics of the current detection object, ensuring that the optimization direction conforms to the new standard requirements while avoiding ineffective searches. Finally, by loading the optimized target standard adaptation cognitive parameters into the defect detection module to execute the task, a complete migration-optimization technical closed loop is formed. This significantly shortens the parameter adaptation cycle when switching detection standards and improves the consistency and compliance of detection results under different standard systems. Attached Figure Description

[0018] Figure 1 The flowchart illustrates the self-optimization method for cognitive parameters of an embodied unmanned system adapted to multiple detection standards, as described in this invention. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0020] Example 1 This application provides a method for self-optimization of cognitive parameters in embodied unmanned systems adapting to multiple detection standards. This method addresses the problem of low cognitive parameter adaptation efficiency when switching detection standards by constructing a cognitive parameter transfer mechanism driven by standard differences and a closed-loop self-optimization mechanism based on physical feedback. Figure 1 As shown, the method includes the following steps: Step S110: Obtain the target detection standard corresponding to the current defect detection task and the historical detection records formed by the embodied unmanned system in the historical defect detection tasks. The target detection standard includes the defect category definition, defect measurement rules, defect level classification rules and detection result output rules. The historical detection records include the historical applicable detection standards, historical detection object attributes, historical cognitive parameters, historical defect identification results and historical defect review results.

[0021] Specifically, the target detection standard serves as the compliance benchmark for the detection task. It specifies not only "what to check" (defect categories) but also "how to measure" (measurement rules), "how to judge" (grade classification), and "how to report" (output rules). When the embodied unmanned system enters a new scenario or receives new instructions, the system first parses the standard file bound to the task, transforming it into a machine-readable structured rule set. This step establishes the target boundary for all subsequent cognitive parameter optimizations, ensuring that the unmanned system's defect detection behavior always serves specific standard requirements, rather than general image quality indicators. Simultaneously, the system acquires historical detection records, which constitute the unmanned system's experience knowledge base. The correspondence between historically applicable detection standards and historical cognitive parameters records "what" cognitive parameter configurations were verified as effective under a certain standard in the past; while historical defect verification results provide confidence labels for parameter validity, distinguishing which parameters are truly accurate and which are false positives or omissions. This record structure, encompassing environmental context (object attributes) and verification feedback (verification results), provides a reliable data foundation for subsequent cross-standard migration, avoiding the blind reuse of unverified historical parameters.

[0022] Among them, cognitive parameters include at least one of defect category identification parameters, defect boundary understanding parameters, size measurement reasoning parameters, defect level determination parameters, detection standard mapping parameters, re-inspection decision parameters, and detection result expression parameters.

[0023] These cognitive parameters constitute the core decision variables of the unmanned system's defect detection module, corresponding to different cognitive stages in the defect detection process: Defect category identification parameters determine how the system classifies the detected object into the standard-defined defect category system; defect boundary understanding parameters control the system's judgment logic on the integrity and continuity of the defect contour; dimensional measurement inference parameters affect the system's inference accuracy on the defect's geometric dimensions; defect level determination parameters determine how the system maps measurement results to the standard-specified level system; detection standard mapping parameters manage the semantic correspondence between different standard systems; re-inspection decision parameters control the system's judgment logic on whether to trigger further detection for suspected defect areas; and detection result expression parameters standardize the format and field structure of the final output. Unlike traditional sensor acquisition parameters, cognitive parameters are directly related to the semantic connotation of the detection standard, serving as a bridge connecting physical perception and standard compliance.

[0024] Step S120: Based on the target detection standard and historically applicable detection standards, construct a standard difference vector. The standard difference vector is used to characterize the differences between different detection standards in defect category definition, defect measurement benchmark, defect level threshold and detection result expression method.

[0025] Specifically, the standard difference vector is the core bridge for cross-standard adaptation in this embodiment. It is not simply a text comparison result, but rather a structured vector that quantifies the semantic differences between the old and new standards across four key dimensions. For example, when the new standard changes the measurement benchmark for crack width from maximum width to average width, this vector will produce a significant component in the measurement benchmark dimension. This vector precisely characterizes "where it changed" and "how much it changed," thereby guiding the direction and magnitude of subsequent cognitive parameter adjustments. This enables the unmanned system to understand the technical implications of the standard change, rather than merely perceiving a change in the standard's name.

[0026] Step S130: Collect multimodal detection data of the current detection object through the airborne sensors of the embodied unmanned system, and extract the features of the current detection object based on the multimodal detection data. The multimodal detection data includes at least image data, depth data, pose data and odometer data. The features of the current detection object include the structural features, material features, surface texture features, suspected defect morphological features, defect spatial location features and motion state features of the unmanned system.

[0027] Specifically, this step embodies the embodied nature of the solution. Unlike fixed cameras, the perception data of an embodied unmanned system is tightly coupled with its own motion state and physical position. Image data provides texture and color information, depth data provides three-dimensional geometric information, pose data provides the sensor's attitude relative to the global coordinate system, and odometry data provides the unmanned system's relative displacement in the environment. The simultaneous acquisition of these four types of data provides indispensable physical input for accurately locating defects, understanding environmental interference, and performing precise physical calibration actions in dynamic environments. The feature extraction process transforms raw sensor data into a high-level semantic description. Structural and material features reflect the inherent properties of the environment, determining the characteristics of light reflection and signal attenuation; surface texture features and suspected defect morphological features are directly related to the detection difficulty; and the spatial location features of defects, combined with the motion state features of the unmanned system, describe the relative relationship between the unmanned system and the defect. These features together constitute a joint state description of the environment and the unmanned system at the current moment, enabling subsequent cognitive parameter constraints to be adapted to local conditions rather than applying a fixed template.

[0028] Step S140: Based on the standard difference vector and the current detection object characteristics, determine the cognitive parameter adjustment constraints corresponding to the target detection standard, and determine the allowable range of the embodied execution parameters associated with the cognitive parameter adjustment constraints. The embodied execution parameters include at least one of the following: travel speed, shooting distance, sensor attitude, supplementary light intensity, re-inspection trigger threshold, and number of local depth measurements.

[0029] Specifically, the cognitive parameter adjustment constraint defines the feasible search space for self-optimization, determined by both standard requirements and the current environmental conditions. Simultaneously, the system also determines the permissible range of embodied execution parameters associated with the cognitive parameter adjustment constraint. Embodied execution parameters are execution variables that the unmanned system can directly control at the physical level, forming a hierarchical collaborative relationship with the cognitive parameters: cognitive parameters determine "how to understand and judge" defects, while embodied execution parameters determine "how to physically act" to obtain higher-quality detection data. For example, if the standard difference vector indicates that the new standard has higher requirements for the detection of minor defects, and the current object features show severe surface reflection, the system will automatically tighten the adjustment range of the defect boundary understanding parameters, while simultaneously expanding the search range of sensor attitude and limiting the upper limit of supplementary lighting intensity. This dynamically generated dual constraint mechanism prevents the optimization process from deviating from the requirements of the new standard and avoids wasting computational resources in physically infeasible or invalid parameter regions, significantly improving the safety and convergence speed of self-optimization.

[0030] Step S150: Based on historical detection records, select source detection standards from historically applicable detection standards that have a standard mapping relationship with the target detection standard. Calculate the cognitive parameter migration weight from the source detection standard to the target detection standard according to the standard difference vector. Then, perform migration transformation on the historical cognitive parameters corresponding to the source detection standard according to the cognitive parameter migration weight to obtain candidate cognitive parameters under the target detection standard.

[0031] Specifically, this step addresses the cold start problem in optimizing cognitive parameters under the new standard. Instead of randomly selecting historical parameters as a starting point, the system intelligently selects the most suitable source detection standard based on the semantic similarity between standards and the actual reliability of historical parameters. The cognitive parameter migration weights further quantify the reference value of the source standard cognitive parameters for the new standard: parameters with smaller differences and higher historical pass rates receive greater migration weights. This allows the unmanned system to start new tasks by building upon existing foundations, significantly shortening the time required to explore optimal parameters from scratch. The migration transformation is not a simple parameter copying but an adaptive adjustment based on weights and standard differences. For cognitive parameters significantly affected by standard changes (such as defect level determination parameters), the transformation amplitude is larger; for more universal cognitive parameters (such as detection standard mapping parameters), more historical values ​​are retained. The generated candidate cognitive parameters represent an initial point highly close to the optimal solution. They inherit the essence of historical experience while initially responding to the differentiated requirements of the new standard, laying a high-quality foundation for subsequent fine-tuning.

[0032] Step S160: Control the embodied unmanned system to perform standard adaptation calibration actions within the allowable range of embodied execution parameters and collect calibration feedback data. Calculate the standard adaptation loss of candidate cognitive parameters under the target detection standard based on the calibration feedback data. With the goal of reducing the standard adaptation loss, self-optimize the candidate cognitive parameters within the cognitive parameter adjustment constraints to obtain the target standard-adapted cognitive parameters.

[0033] Specifically, this is a key step that distinguishes this embodiment from pure algorithm optimization and is a concentrated manifestation of embodied intelligence. Calibration is a proactive detection behavior performed by the unmanned system in the physical world within the allowable range of embodied execution parameters, aiming to obtain the real feedback needed to verify candidate cognitive parameters. For example, to verify whether the boundary understanding parameters of a defect are appropriate, the unmanned system may need to actually adjust the supplementary lighting intensity and retake the image; to confirm the defect size, it may need to decelerate and approach to perform local depth measurements. Calibration feedback data comes from real physical interactions and includes factors that are difficult to simulate, such as environmental noise and sensor nonlinearity, ensuring the authenticity and robustness of the optimization process. Standard adaptation loss is a quantitative indicator that measures the degree of matching between the current cognitive parameters and the new standard. Based on calibration feedback data, it evaluates the total deviation of the unmanned system's detection results from the target detection standard in multiple dimensions, including category mapping, size measurement, level classification, boundary integrity, result output, and action cost, under the current cognitive parameter configuration. Self-optimization is an iterative optimization process. The system uses the loss value to continuously fine-tune the cognitive parameters within the constraint range of cognitive parameter adjustment and repeatedly executes calibration actions and loss calculations until the loss converges to a preset threshold or reaches the maximum number of iterations. Since the starting point (candidate cognitive parameters) has been intelligently transferred and the search space has been reasonably constrained, the optimization process can usually be completed in a very short time, meeting the needs of real-time operation of embodied unmanned systems.

[0034] Step S170: Load the target standard adaptation cognitive parameters into the defect detection module of the embodied unmanned system, and execute the current defect detection task based on the target standard adaptation cognitive parameters.

[0035] Specifically, this is the practical application of the optimized cognitive parameters. The optimized cognitive parameters are injected into the defect detection module of the unmanned system in real time, replacing the original default parameters or old standard parameters. Subsequently, the unmanned system continues to perform the detection task with the new cognitive configuration, outputting detection results that conform to the target detection standard specifications. The entire process requires no manual intervention or downtime for debugging, achieving seamless transition between detection standards and immediate adaptation of cognitive performance.

[0036] Through steps S110 to S170 above, this embodiment establishes a complete technical framework of standard understanding, experience transfer, physical verification, and closed-loop self-optimization. The standard difference vector serves as a semantic bridge connecting old and new standards, ensuring the correct direction of cognitive parameter adjustments; the cross-standard cognitive parameter transfer mechanism utilizes historical experience to accelerate the convergence process; and the closed-loop self-optimization mechanism based on embodied calibration actions completely resolves the gap between simulation and reality by placing the algorithm in a physical feedback loop, guaranteeing the effectiveness and compliance of the final adapted cognitive parameters in the real-world operating environment.

[0037] Example 2 In this embodiment, the construction process of the standard difference vector is further refined. As one implementation method, the standard difference vector is constructed based on the target detection standard and historically applicable detection standards, including: parsing the standard rules of the target detection standard and the historically applicable detection standards respectively to obtain the target standard rule description and the historical standard rule description; extracting the defect category field, defect measurement field, defect level field, review requirement field, and test result output field from the target standard rule description and the historical standard rule description respectively; establishing the defect category mapping relationship, defect measurement benchmark mapping relationship, defect level threshold mapping relationship, and test result expression mapping relationship between the target detection standard and the historically applicable detection standards; calculating the differences between the target detection standard and the historically applicable detection standards in defect category mapping, defect measurement benchmark, defect level threshold, review requirements, and test result expression methods to obtain the standard difference vector.

[0038] Specifically, standard rule parsing is a crucial step in transforming unstructured standard documents into machine-processable structured data. Since inspection standards typically exist in PDF or natural language text format, the system first uses optical character recognition (OCR) and natural language processing (NLP) technologies to locate and extract key semantic fragments. For example, when parsing the clause "concrete surface cracks wider than 0.2mm are classified as Level III defects," the system breaks it down into a structured rule description object, including the defect category field "crack," the measurement field "maximum width," the grade field "Level III," and the threshold "0.2mm." Based on this, establishing five mapping relationships is the core mechanism for eliminating semantic ambiguity across standards. Different standards often name the same physical defect differently; for example, a national standard might call it "crack," while a certain enterprise standard might call it "cracking" or "fissure." Without establishing explicit mapping relationships, the system might misclassify the differences as two completely different defect types, leading to incorrect migration weight calculations. This embodiment automatically aligns "crack" to "fissure" using a pre-built thesaurus or semantic embedding-based similarity matching, ensuring that the difference vector reflects actual rule changes rather than differences in terminology. Specifically, the extraction of the review requirement field and the calculation of the review requirement dimension differences enable the standard difference vector to capture the differences between the old and new standards in the quality verification process. For example, the new standard may require manual review of specific categories of defects, while the old standard only requires automatic judgment. This difference directly affects the migration strategy of the review decision parameters. The final generated standard difference vector is a multi-dimensional structured representation, with each component corresponding to a specific deviation value between the old and new standards in a particular dimension, providing precise quantitative guidance for subsequent cognitive parameter migration.

[0039] Example 3 In this embodiment, the structure and dynamic determination mechanism of cognitive parameter adjustment constraints are further elaborated. As one implementation, cognitive parameter adjustment constraints include standard mapping constraints, defect identification constraints, boundary understanding constraints, dimensional measurement reasoning constraints, defect level determination constraints, and re-inspection decision constraints. Specifically, the standard mapping constraints limit the mapping relationships between different detection standards regarding defect categories, measurement rules, level thresholds, and result expression methods. The re-inspection decision constraints limit the confidence range, measurement error range, level boundary distance range, and multi-frame recognition consistency range for triggering re-inspection actions in suspected defect areas.

[0040] Specifically, the six types of constraints correspond to different cognitive stages in the cognitive parameter system, forming a structured optimization boundary. The standard mapping constraint ensures that the detection standard mapping parameters are adjusted within the range of legal standard correspondences; the defect identification constraint limits the search space of defect category identification parameters, preventing the system from incorrectly classifying a defect into a category not defined by the standard; the boundary understanding constraint controls the adjustment range of the defect boundary understanding parameters, avoiding over-segmentation or under-segmentation; the dimensional measurement inference constraint ensures that the output of the dimensional measurement inference parameters always meets the measurement accuracy requirements specified by the standard; the defect level determination constraint limits the movement range of the level determination boundary, preventing large-scale level misjudgments due to parameter drift; and the re-inspection decision constraint ensures the rationality of re-inspection triggering by limiting the confidence range, measurement error range, level boundary distance range, and multi-frame recognition consistency range, avoiding excessive re-inspection that wastes computing power or insufficient re-inspection that leads to missed detections. This hierarchical constraint structure allows the self-optimization process to independently shrink or expand the search space in each cognitive dimension, ensuring the correctness of the optimization direction while avoiding invalid cross-dimensional searches.

[0041] As one implementation method, the cognitive parameter adjustment constraints corresponding to the target detection standard are determined based on the standard difference vector and the characteristics of the current detection object. This includes: when the standard difference vector indicates that the target detection standard's requirement for defect size measurement accuracy is higher than that of historically applicable detection standards, increasing the accuracy constraint of the size measurement inference parameters, reducing the threshold tolerance of the defect level judgment parameters, and expanding the re-inspection trigger range for suspected defect areas; when the standard difference vector indicates that the target detection standard's requirement for defect boundary integrity is higher than that of historically applicable detection standards, increasing the integrity constraint of the defect boundary understanding parameters and reducing the adjustment step size of the defect boundary judgment threshold; when the standard difference vector indicates that the target detection standard's defect level classification granularity is higher than that of historically applicable detection standards, reducing the search step size of the size measurement inference parameters and increasing the number of local depth measurements; and when the characteristics of the current detection object indicate that surface reflection, occlusion, or texture interference exceeds a preset interference threshold, adjusting the re-inspection decision parameters and the allowable variation range of the associated supplementary lighting intensity, shooting distance, and sensor posture.

[0042] Specifically, these four conditional branching logics illustrate how the constraints are derived. When the measurement accuracy requirement increases, the accuracy constraint of the dimensional measurement inference parameters is tightened, while the threshold tolerance of the defect level determination parameters is reduced, because more accurate measurement results directly affect the accuracy of level determination. When the boundary integrity requirement increases, the integrity constraint of the defect boundary understanding parameters is strengthened, and the adjustment step size is reduced to achieve finer search. When the level division granularity is refined, the search step size is reduced to adapt to a narrower level window, while the number of local depth measurements is increased to eliminate random errors in single-point measurements. When environmental interference is severe, the allowable range of the re-inspection decision parameters and the embodied execution parameters (supplementary light intensity, shooting distance, sensor posture) are adjusted in tandem, reflecting the hierarchical collaborative mechanism of cognitive parameters and embodied execution parameters.

[0043] Example 4 In this embodiment, the specific implementation mechanism of the cross-standard cognitive parameter transfer system is further elaborated. As one implementation method, source detection standards are selected from historically applicable detection standards that have a standard mapping relationship with the target detection standard. This includes: obtaining historical standard rule descriptions corresponding to multiple historically applicable detection standards; calculating the standard rule similarity between the target detection standard and each historically applicable detection standard; calculating the source standard credibility of each historically applicable detection standard based on standard rule similarity, historical defect review results, and historical detection object attributes; and determining historically applicable detection standards whose source standard credibility meets a preset credibility threshold as source detection standards.

[0044] Specifically, the selection of source detection standards is not a simple text matching process, but a multi-dimensional decision-making process that integrates semantic understanding and empirical verification. Standard rule similarity only reflects the degree of closeness between old and new standards at the level of clause expression; however, clause similarity does not equate to the direct reuse of cognitive parameters. For example, two standards may have completely identical definitions of "crack," but if the detection records corresponding to historical standard A show a high false negative rate, or if historical standard A was verified on a dry concrete surface while the current task is on a damp tile surface, then even if standard A has the highest text similarity, the actual reference value of its cognitive parameters is greatly reduced. Therefore, this embodiment introduces historical defect verification results as a truth value verification factor. Only historical records that have been verified accurately by manual or high-precision equipment can be given a higher credibility weight to the corresponding standard. Simultaneously, historical detection object attributes are introduced as an environmental compatibility factor to ensure that the application scenario of the source standard is comparable to the current physical environment.

[0045] Specifically, the credibility of the source standard is calculated using a three-factor weighted fusion model, the expression of which is: in, For the credibility of the source standard, For standard rule similarity, For parameter confidence, For environmental compatibility, , , Preset weighting coefficients and satisfying The specific calculation methods for each factor are as follows: Standard rule similarity The calculation involves encoding the historical standard rule descriptions of the target detection standard and the historical standard rule descriptions of the currently applicable detection standard into semantic embedding vectors, and then calculating the cosine similarity between the two as... The encoding of semantic embedding vectors can be achieved through pre-trained language models, enabling standard terms with similar meanings but different wording (such as "crack" and "fissure") to obtain high similarity scores.

[0046] Parameter reliability The calculation involves extracting all historical defect verification results corresponding to the applicable historical testing standard from historical testing records, and calculating the proportion of records that passed verification (i.e., the defect identification result is consistent with the verification conclusion by manual or high-precision equipment) out of the total number of testing records under that standard. This proportion is used as the basis for the calculation. For example, if there are 200 test records under a certain historical applicable testing standard, and 180 of them have been verified to be accurate, then... =180 / 200=0.9. This factor directly reflects the reliability of historical cognitive parameters in actual operations.

[0047] Environmental compatibility The calculation involves extracting the historical test object attributes (including structured attribute fields such as test object type, material category, surface condition, and ambient temperature and humidity range) corresponding to the historical applicable testing standard. These attributes are then matched item by item with the corresponding attribute fields in the current test object's features. The proportion of matched attribute fields to the total number of attribute fields is calculated as the percentage of the total number of attribute fields. For example, if the attribute field contains a total of 5 items: object type, material category, surface condition, temperature range, and humidity range, and the object type, material category, and surface condition match, then... =3 / 5=0.6. This factor ensures that the application scenarios of the source standard are comparable to the current physical environment, avoiding the transfer of parameter experience from a dry environment to a humid environment.

[0048] The preset values ​​of the weighting coefficients α, β, and γ can be adjusted according to the actual application scenario. As a preferred configuration, α = 0.3、 β = 0.5、 c = 0.2, meaning the parameter confidence level is... It carries the highest weight because the actual verification effect of historical cognitive parameters is the most direct basis for judging whether they are reusable; standard rule similarity. Secondly, it provides semantic-level references; environmental compatibility. As an auxiliary constraint, the preset confidence threshold can be set to 0.7, meaning only... Only historically applicable testing standards with a score ≥ 0.7 are designated as source testing standards. This multi-dimensional evaluation mechanism effectively avoids high-scoring but low-energy source standards misleading the optimization direction and ensures the substantial reliability of the migration starting point.

[0049] As one implementation method, the cognitive parameter migration weight from the source detection standard to the target detection standard is calculated based on the standard difference vector, including: calculating the standard domain difference between the source detection standard and the target detection standard based on the standard difference vector; calculating the parameter reliability of the historical cognitive parameters corresponding to the source detection standard based on the historical defect review results; calculating the similarity of the detected object based on the attributes of the historical detected object and the characteristics of the current detected object; and calculating the cognitive parameter migration weight from the source detection standard to the target detection standard according to the relationship that is negatively correlated with the standard domain difference and positively correlated with the parameter reliability and the similarity of the detected object.

[0050] Specifically, the cognitive parameter transfer weight is a dynamic adjustment knob connecting historical experience with the requirements of new standards. Its value is determined by three factors: the difference in the standard domain (objective distance, negative correlation), the reliability of the parameters (historical validation effect, positive correlation), and the similarity of the detected objects (environmental compatibility, positive correlation). When calculating the transfer weight, the system employs a non-linear fusion strategy: when the difference in the standard domain is small, the reliability of the parameters is high, and the similarity of the detected objects is high, the transfer weight approaches 1, indicating that historical cognitive parameters can be highly trusted; when the difference in the standard domain is large but the reliability of the parameters is extremely high, the transfer weight is moderately reduced but still retains a certain proportion; and when all three are at low levels, the transfer weight is suppressed to an extremely low level, forcing the system to rely more on subsequent self-optimization processes rather than blind transfer.

[0051] Specifically, the calculation process for parameter reliability is as follows: All historical defect verification results corresponding to the source detection standard are extracted from historical inspection records, and a single-item reliability score is calculated for each verification result. The calculation method for the single-item reliability score is as follows: If the defect identification result of the inspection record is completely consistent with the verification conclusion of manual or high-precision equipment (including matching defect category, defect size, and defect level), the single-item reliability score is 1.0; if the defect category is consistent but the defect size deviation is within the standard allowable tolerance range, the single-item reliability score is 0.7; if the defect category is consistent but the defect level judgment deviation is level one, the single-item reliability score is 0.5; if the defect category is inconsistent, the single-item reliability score is 0. All single-item reliability scores are weighted and averaged to obtain the parameter reliability, the expression of which is: in, N This represents the total number of test records under the source testing standard. For the first i Individual credibility score for each record. For the first i Time decay weight of each record , This is the time interval (in days) since the record was created. l For the preset attenuation coefficient (e.g.) l = 0.001), giving higher weight to recent detection records. This time decay mechanism ensures that the parameter reliability reflects the actual reliability of historical knowledge parameters at the current point in time, avoiding over-reliance on outdated experience caused by factors such as equipment aging and environmental changes.

[0052] Specifically, the nonlinear fusion strategy for cognitive parameter transfer weights uses the following calculation expression: in, For cognitive parameter transfer weights, For parameter confidence, To detect object similarity, For standard domain differences, p , q , k These are the preset adjustment parameters.

[0053] Standard domain differences The calculation method is as follows: normalize the standard difference vector and calculate its L2 norm (Euclidean norm), i.e. ,in For the standard difference vector, The maximum difference norm reference value is preset so that It is normalized to the interval [0, 1].

[0054] Detect object similarity The calculation method is compatible with the aforementioned environment. The same means that the attribute fields of the historical detected object and the current detected object are matched item by item to calculate the matching hit ratio.

[0055] In the above expression, the confidence level of the parameters Similarity to the detected object It participates in the fusion in the form of a power function, exponent p and q Control the intensity of its influence on migration weights. As a preferred configuration, p = 1.2、 q = 0.8, meaning the impact of parameter reliability is slightly stronger than that of object similarity, because the actual validation effect of historical cognitive parameters is the most direct basis for judging whether they can be reused. Standard domain difference. It participates in the fusion in the form of a negative exponential function, with coefficients k Controlling the degree of suppression of migration weights by the amount of difference is a preferred configuration. k = 2.0. The technical effect of this nonlinear fusion strategy is that when Approaching 0 (with minimal inter-standard differences) and and When both are relatively high, W_migrate approaches 1, and W_migrate approaches 1. The migration weights are close to their upper limit; when When the difference is large (significant difference between standards), Rapid decay, even Extremely high, It will also be significantly suppressed, but will not drop to zero, retaining some reference value of highly reliable historical parameters; when , and When all three are at low levels Suppressed to extremely low levels (e.g., below 0.1), the system is forced to rely more on subsequent self-optimization processes rather than blind migration.

[0056] As one implementation method, historical cognitive parameters corresponding to the source detection standard are transformed according to the cognitive parameter transfer weights to obtain candidate cognitive parameters under the target detection standard. This includes: determining the transfer sensitivity of various cognitive parameters based on the standard difference vector; performing transfer transformations on defect category identification parameters, defect boundary understanding parameters, size measurement inference parameters, defect level judgment parameters, detection standard mapping parameters, re-inspection decision parameters, and detection result expression parameters in the historical cognitive parameters according to the cognitive parameter transfer weights and transfer sensitivity; and projecting the transformed cognitive parameters into the parameter allowable range corresponding to the target detection standard to obtain candidate cognitive parameters.

[0057] Specifically, the migration transformation process profoundly reflects the differentiated respect for the physical attributes of different cognitive parameters. Migration sensitivity is a key technical concept, characterizing the degree to which a certain type of cognitive parameter responds to standard changes. For example, detection standard mapping parameters are mainly governed by the standard system architecture and are relatively insensitive to changes in specific measurement rules, belonging to low-sensitivity parameters, and should primarily retain historical values ​​during migration. Conversely, defect level judgment parameters and size measurement inference parameters directly correspond to the judgment rules and measurement benchmarks in the standard, are extremely sensitive to standard changes, belonging to high-sensitivity parameters, and must be significantly reconstructed based on the standard difference vector. Through classification transformation, the system can accurately translate changes in standard semantics into adjustments to the corresponding cognitive parameters. Furthermore, projecting to the allowable parameter range is an indispensable safety fallback mechanism, ensuring that the generated candidate cognitive parameters are not only mathematically sound but also physically executable and safe.

[0058] Specifically, the logic for determining the transfer sensitivity of various cognitive parameters based on the standard difference vector is as follows: The standard difference vector comprises five dimensions: defect category mapping difference component, defect measurement benchmark difference component, defect level threshold difference component, review requirement difference component, and detection result expression difference component. The system pre-defines a sensitivity correlation vector for each type of cognitive parameter, which characterizes the correlation strength between that type of cognitive parameter and each dimension of the standard difference vector. The sensitivity correlation vectors for the seven cognitive parameters are defined as follows: the sensitivity correlation vector for the defect category identification parameter is [1.0, 0.2, 0.3, 0.1, 0.1], indicating that it is mainly affected by differences in defect category mapping; the sensitivity correlation vector for the defect boundary understanding parameter is [0.2, 0.8, 0.4, 0.2, 0.1], indicating that it is mainly affected by differences in defect measurement benchmarks and defect level thresholds; the sensitivity correlation vector for the size measurement inference parameter is [0.1, 1.0, 0.6, 0.1, 0.1], indicating that it is most sensitive to differences in defect measurement benchmarks; the sensitivity correlation vector for the defect level determination parameter is [0.2, 0.6, 1.0, 0.3, 0.2], indicating that it is most sensitive to differences in defect level thresholds; the sensitivity correlation vector for the detection standard mapping parameter is [0.5, 0.3, 0.3, 0.2, ...]. [0.5] indicates that it is affected by both the difference in defect category mapping and the difference in the expression of detection results, but the overall sensitivity is low; the sensitivity correlation vector of the re-inspection decision parameter is [0.2, 0.4, 0.5, 1.0, 0.2], indicating that it is mainly affected by the difference in re-inspection requirements; the sensitivity correlation vector of the detection result expression parameter is [0.1, 0.1, 0.2, 0.2, 1.0], indicating that it is mainly affected by the difference in the expression of detection results.

[0059] The transfer sensitivity of various cognitive parameters is calculated using the following expression: in, For the first k Transfer sensitivity (scalar) of cognitive parameters. For the first k Sensitivity correlation vector of class cognitive parameters This is the normalized standard difference vector (each component divided by the maximum difference reference value of its corresponding dimension, normalized to the [0, 1] interval), where "·" represents the vector dot product operation. Transfer sensitivity. The numerical range is A higher value indicates that the cognitive parameter is more sensitive to changes in the current standard, requiring a greater degree of migration and adjustment. According to Based on the numerical value, the system divides the cognitive parameters into three sensitivity levels: when When the value is below a preset low-sensitivity threshold (e.g., 0.3), it is determined to be a low-sensitivity parameter; when... When the value is between the low sensitivity threshold and the high sensitivity threshold (e.g., 0.7), it is determined to be a medium sensitivity parameter; when... Parameters exceeding the high sensitivity threshold are considered high sensitivity parameters.

[0060] Specifically, the calculation process for transferring historical cognitive parameters based on cognitive parameter transfer weights and transfer sensitivity is as follows: For the first... k The class cognitive parameters, and their transfer transformation expression are: in, For the transformed th k Class cognitive parameters, For the source detection standard k The original values ​​of the historical cognition parameters, For cognitive parameter transfer weights, For the first k Transfer sensitivity of cognitive parameters The first one calculated based on the standard difference vector k Standard adaptation adjustment amount for class cognitive parameters.

[0061] Standard adaptation adjustment amount The calculation method is as follows: ,in For the first k The parameter adaptation mapping function corresponds to the cognitive parameters of different categories. The adaptation mapping function differs for different categories of cognitive parameters: for size measurement inference parameters, The defect measurement baseline difference component in the standard difference vector is mapped to the parameter offset of the measurement inference model. For example, when the measurement baseline changes from "maximum width" to "average width", This refers to the parameter for switching the statistical aggregation method in the inference model; for the defect level determination parameter, The defect level threshold difference component is mapped to an adjustment amount for the level determination boundary. For example, when the level granularity is refined from level three to level five, Insert parameters for adding grade boundary thresholds; for defect category identification parameters, Map the defect category difference components to the weight adjustments of the category classifier; for the re-inspection decision parameters... Map the difference components of the review requirements to the threshold adjustment amount of the review trigger condition.

[0062] The physical meaning of the above transfer transformation expression is: the transformed cognitive parameters. It is composed of two weighted and merged parts. Part 1 This indicates that the standard adaptation adjustment amount, modulated by sensitivity, is superimposed on historical parameters, reflecting the migration direction towards the new standard based on historical experience; Part Two This indicates that the original values ​​of historical parameters are retained as safety anchors to prevent excessive parameter shifts when the migration weights are low. Higher (e.g., 0.85) and When the value is high (e.g., a defect level determination parameter with high sensitivity), the transformation result is mainly determined by the standard adaptation adjustment amount. Dominated by historical parameters, they were significantly reconstructed; when Higher but At lower (e.g., low-sensitivity detection standard mapping parameters), Approaching zero, the transformation result is close to the original value of the historical parameters; when At lower levels (e.g., 0.2), regardless of The size and transformation results are based on historical parameters, with only minor adjustments, leaving fine-tuning to the subsequent closed-loop self-optimization process.

[0063] The projection to the allowed parameter range is used to ensure that the transformed cognitive parameters always fall within the legally feasible domain specified by the target detection standard. Specifically, for the first... k Class cognitive parameters, the target detection standard specifies the lower bound of its parameters. and parameter upper bound The expression for the projection operation is: That is when When below the lower bound, take ,when Take when it is higher than the upper limit Otherwise keep Unchanged. After projection operation That is, the th in the final candidate cognitive parameters k The value of the class parameter.

[0064] Example 5 In this embodiment, the specific implementation details of the closed-loop self-optimization mechanism based on physical feedback are further elaborated. As one implementation method, the standard adaptation calibration actions include: when a suspected defect area meets the re-inspection trigger conditions, controlling the embodied unmanned system to reduce its travel speed and re-inspect the suspected defect area; controlling the embodied unmanned system to adjust the sensor pitch or rotation angle to obtain multi-angle detection data of the suspected defect area; controlling the embodied unmanned system to adjust the supplementary lighting intensity to reduce the impact of reflection, shadows, or occlusion on defect boundary recognition; controlling the embodied unmanned system to repeatedly locate the suspected defect position based on mileage data and pose data; controlling the embodied unmanned system to perform local depth measurement on the suspected defect area to obtain size measurement data of the suspected defect area; wherein, the re-inspection trigger conditions include at least one of the following: defect confidence level is in a preset gray zone, defect size is close to a level threshold, depth measurement fluctuation exceeds a preset fluctuation threshold, reflection or occlusion exceeds a preset interference threshold, and inconsistent recognition results across multiple frames in the same area.

[0065] Specifically, these five types of calibration actions are not preset fixed procedures, but rather proactive perception behaviors that the system dynamically decides based on the confidence level of the current candidate cognitive parameters and environmental characteristics. The multi-dimensional judgment logic of the re-inspection trigger conditions ensures the precise triggering of calibration actions: a defect confidence level in a preset gray zone indicates that the system is uncertain about the defect judgment and requires more data support; a defect size close to the grade threshold indicates that a small measurement deviation may lead to a misjudgment of the grade; depth measurement fluctuations exceeding a preset fluctuation threshold indicate unstable measurement data; reflections or occlusions exceeding a preset interference threshold indicate poor current imaging conditions; inconsistent recognition results across multiple frames in the same area indicate insufficient repeatability of the detection results. These conditions cover the main scenarios of cognitive parameter failure, enabling calibration actions to obtain the physical feedback required for verification in a targeted manner.

[0066] As one implementation method, the standard adaptation loss includes defect category mapping loss, size measurement error loss, defect level classification consistency loss, defect boundary integrity loss, detection result output consistency loss, and calibration action cost; self-optimization includes iteratively updating candidate cognitive parameters within the cognitive parameter adjustment constraints through at least one optimization method among Bayesian optimization, constrained reinforcement learning, model predictive control, gradient search, or gradient-free search until the standard adaptation loss meets a preset loss threshold or reaches a preset number of iterations.

[0067] Specifically, the six-dimensional loss function comprehensively covers all the requirements of the detection standard for cognitive parameters. The defect category mapping loss reflects the cognitive parameters' ability to extract semantic features of defects; the dimensional measurement error loss is directly related to the accuracy of the dimensional measurement inference parameters; the defect level classification consistency loss reflects the sensitivity of the parameters to the boundary of defect severity classification; the defect boundary integrity loss assesses the continuity and accuracy of the defect contour; the detection result output consistency loss focuses on the standardization of the output format; and the calibration action cost constrains the resource consumption of physical calibration, preventing over-calibration from affecting operational efficiency. The selection of multiple optimization algorithms allows the self-optimization process to be flexibly configured according to actual computational resources and convergence speed requirements: Bayesian optimization is suitable for efficient exploration of high-dimensional parameter spaces; constrained reinforcement learning is suitable for handling complex constraints; model predictive control is suitable for multi-step look-ahead sequential decision-making; gradient search is suitable for scenarios where the loss function is differentiable; and gradient-free search is suitable for scenarios where the loss function is not differentiable or where discrete variables exist.

[0068] Example 6 In this embodiment, the technical features of the foregoing embodiments are integrated into a specific business application scenario to verify the feasibility of the technical solution of this application and the synergistic effect between the modules. It should be understood that the following scenario is merely a preferred implementation of the technical solution of this application and is not a limitation on the scope of protection of this application. In practical applications, the method of this application is also applicable to other inspection objects such as tunnels, pipelines, and industrial containers, as well as the switching and adaptation between other types of inspection standards such as national standards, industry standards, and enterprise standards.

[0069] To illustrate this application more clearly, the following example illustrates a scenario where an embodied unmanned system switches from a national testing standard to a city-specific customized testing standard for underground utility tunnels. In this scenario, the unmanned system was initially performing routine inspections according to the national testing standard. However, due to project acceptance requirements, it needs to temporarily switch to the city's customized testing standard. This customized standard, based on the national standard, adds rules for measuring seepage area and refines the granularity of crack classification. Using traditional manual offline debugging methods, technicians typically need to stop the system, recalibrate the algorithm thresholds, and repeatedly verify the results, a process that takes approximately two hours and is difficult to guarantee accurate matching with the customized standard. However, using the embodied unmanned system cognitive parameter self-optimization method for multi-test standard adaptation provided in this application, the entire adaptation process can be completed online during the unmanned system's movement, reaching a stable testing state in just about five minutes, significantly improving operational efficiency and result compliance.

[0070] Specifically, the cognitive parameter self-optimization process in this comprehensive application scenario includes the following steps: Step S110: Obtain the target detection standard corresponding to the current defect detection task and the historical detection records generated by the embodied unmanned system in historical defect detection tasks. In this scenario, the system receives a task scheduling instruction and loads a customized detection standard file for a municipal utility tunnel. This standard file contains defect category definitions, defect measurement rules, defect level classification rules, and detection result output rules. Unlike general national detection standards, this customized standard specifically stipulates that seepage defects must be measured in square meters of their projected area, and subdivides the crack width levels from the original three levels to five levels. Through natural language processing and rule parsing modules, the system transforms these unstructured text clauses into a machine-readable structured rule set, establishing the target boundary for this optimization.

[0071] Step S120: Based on the target detection standard and historically applicable detection standards, a standard difference vector is constructed. The system compares the structured rules of the customized detection standard with those of the national detection standard field by field. In the defect category mapping, the two are basically consistent; however, in the defect measurement benchmark dimension, the newly added seepage area measurement in the customized standard significantly increases the difference component in this dimension; in the defect level threshold dimension, the refinement of crack grading granularity narrows the size window between adjacent levels, resulting in a moderate difference component; in the result expression method, the customized standard requires the addition of regional coding to the output fields, producing a slight difference. The final generated standard difference vector accurately quantifies the semantic distance between the old and new standards in multiple key dimensions, providing directional guidance for subsequent cognitive parameter transfer and constraint generation.

[0072] Step S130: Collect multimodal detection data and extract features of the currently detected object. The unmanned system simultaneously collects image, depth, pose, and odometer data during its movement. The feature extraction module analysis reveals obvious water stains and reflective textures on the current tunnel wall, along with high ambient humidity, which are typical surface interference features. Simultaneously, combined with pose data, it is confirmed that the unmanned system is in a curved section with significant curvature, and motion characteristics show slight lateral shaking. This environmental context information is injected into subsequent decision-making processes in real time, ensuring that cognitive parameter optimization is adapted to local conditions.

[0073] Step S140: Based on the standard difference vector and the characteristics of the current detection object, determine the cognitive parameter adjustment constraints corresponding to the target detection standard, and determine the allowable range of the embodied execution parameters. Given that the standard difference vector characterizes a customized standard with requirements for defect size measurement accuracy and grade classification granularity that are higher than the national detection standard, the system automatically derives the following constraints: the accuracy constraint of the size measurement inference parameter is tightened by 20%, the threshold tolerance of the defect grade judgment parameter is reduced by 15%, and the re-inspection trigger range of suspected defect areas is expanded by 15%. Simultaneously, the allowable range of the embodied execution parameters is determined: the upper limit of the unmanned system's travel speed is reduced from 0.5 m / s to 0.3 m / s, the sensor pitch angle search range is widened to avoid specular reflection directions, and the upper limit of supplementary light intensity is reduced by 30% to prevent overexposure from obscuring water seepage textures. This dual mechanism of cognitive parameter constraints and embodied execution parameter constraints responds to the stringent requirements of the new standard while mitigating the optimization risks caused by environmental interference.

[0074] Step S150: Select the source detection standard and perform cognitive parameter migration transformation. The system calculation found that the national detection standard and the customized detection standard have the highest rule similarity, and historical verification results show that the cognitive parameters under this standard are stable in similar pipe gallery environments. Therefore, it is determined as the source detection standard. Based on the standard domain difference, parameter credibility, and similarity of the detection object, the cognitive parameter migration weight is calculated to be 0.85. Subsequently, the cognitive parameters are classified and transformed according to their migration sensitivity: for low-sensitivity detection standard mapping parameters, 90% of the historical values ​​are retained as the benchmark; for high-sensitivity defect level judgment parameters and size measurement inference parameters, a significant pre-adjustment is made based on the standard difference vector. For example, considering the refinement of crack levels, the level boundary threshold in the defect level judgment parameter is pre-adjusted to the corresponding position of the new five-level system; considering the addition of seepage area measurement, the area inference model parameters in the size measurement inference parameter are pre-corrected. All transformed cognitive parameters are projected into the constraint range determined in step S140, generating a set of high-quality candidate cognitive parameters.

[0075] Step S160 involves performing standard adaptation calibration, calculating standard adaptation loss, and performing self-optimization. After loading candidate cognitive parameters, the unmanned system does not immediately engage in formal testing but instead enters a brief calibration mode. For a suspected water seepage area, the unmanned system proactively reduces its travel speed to 0.2 m / s within the allowable range of its specific execution parameters to perform a re-inspection (meeting the re-inspection trigger condition: defect size close to the level threshold), and adjusts the supplementary lighting intensity from the default value to 60% to eliminate highlights on the water surface. Simultaneously, it controls the gimbal to rotate 15 degrees to acquire lateral view data. Based on this calibration feedback data, the system calculates the six-dimensional standard adaptation loss: defect category mapping loss is 0.02 (accurate water seepage identification), size measurement error loss is 0.08 (area measurement slightly too large), defect level classification consistency loss is 0.05 (critical crack classification), defect boundary integrity loss is 0.03 (boundary basically intact), detection result output consistency loss is 0 (correct format), and calibration action cost is 0.01 (small action amplitude). The total loss after weighted fusion was 0.06, which had not yet reached the preset convergence threshold of 0.03. Therefore, within the constraints of cognitive parameter adjustment, the system generated the next round of cognitive parameters to be evaluated through Bayesian optimization: the defect boundary understanding parameter was fine-tuned to improve the boundary extraction capability of dark details, and the confidence threshold in the re-examination decision parameter was further lowered from 0.75 to 0.7 to increase recall. After three rounds of iterative calibration and loss calculation, the total loss decreased to 0.025, meeting the convergence condition, and the final target standard-fit cognitive parameters were output.

[0076] Step S170: Load the target standard-adaptive cognitive parameters and execute the detection task. The optimized cognitive parameters are injected into the defect detection module of the unmanned system in real time. The unmanned system then resumes its normal inspection speed and continues to execute the detection task under the customized standard with the new cognitive configuration. Actual operation shows that the optimized cognitive parameters can accurately identify water seepage defects and accurately infer their area. The crack level determination is also highly consistent with the annotations of human experts, fully meeting the acceptance requirements of the customized standard.

[0077] To further verify the synergistic effect between the cross-standard cognitive parameter transfer mechanism and the calibration feedback closed-loop self-optimization mechanism in the technical solution of this application, three sets of comparative analyses are conducted based on the aforementioned underground utility tunnel inspection scenario. Comparison Group A: Only cross-standard cognitive parameter transfer is performed without calibration feedback closed-loop self-optimization; that is, the transferred and transformed candidate cognitive parameters are directly loaded into the defect detection module to execute the detection task, skipping the self-optimization process in step S160. Comparison Group B: Only calibration feedback closed-loop self-optimization is performed without cross-standard cognitive parameter transfer; that is, starting with randomly initialized cognitive parameters, self-optimization is performed within the cognitive parameter adjustment constraints. Comparison Group C: The complete technical solution of this application is adopted; that is, high-quality candidate cognitive parameters are first obtained through cross-standard cognitive parameter transfer, and then fine-tuned through calibration feedback closed-loop self-optimization.

[0078] Under the same underground utility tunnel inspection environment and the same standard switching conditions (switching from the national inspection standard to a city-specific utility tunnel inspection standard), the typical performance of the three comparison groups is as follows: In comparison group A, the transfer cognitive parameters were not physically verified, and its standard adaptation loss typically remained between 0.12 and 0.18. The losses in dimensional measurement error and grade classification consistency were relatively significant because pure mathematical transfer could not eliminate the virtual-to-real gap caused by sensor nonlinearity and environmental interference. In comparison group B, although the loss could be gradually reduced through closed-loop self-optimization, due to the lack of a high-quality initial starting point, an average of 8 to 12 iterations were required. Calibration is required to converge to a loss threshold below 0.03. A single standard fit takes approximately 15 to 20 minutes, and unstable detection results may occur in the early exploration rounds due to cognitive parameters being far from the optimal region. In contrast, group C benefits from a high-quality starting point (initial loss of approximately 0.06) provided by the transfer mechanism, and only 2 to 4 iterations are needed to converge to below 0.025. A single standard fit takes approximately 3 to 5 minutes, and the detection results in each iteration remain within an acceptable accuracy range. The final fitting accuracy (loss value of 0.02 to 0.03) is significantly better than that of group A, and the convergence speed is significantly better than that of group B.

[0079] The above comparison results fully demonstrate that the cross-standard cognitive parameter transfer mechanism and the calibration feedback closed-loop self-optimization mechanism are not simply a functional superposition, but rather form an irreplaceable synergistic relationship: the transfer mechanism provides a near-optimal initial starting point for closed-loop self-optimization, significantly reducing the search space and iteration count; the closed-loop self-optimization mechanism, through real physical feedback, eliminates the unavoidable virtual-real bias during the transfer process, ensuring the actual detection performance of the final cognitive parameters. Both are indispensable—without transfer, optimization efficiency is low; without closed-loop, transfer accuracy cannot be guaranteed. This synergistic effect enables the technical solution of this application to simultaneously achieve superior technical results compared to any single mechanism in both standard adaptation efficiency and adaptation accuracy, constituting the substantial features and significant progress of this application compared to existing technologies.

[0080] Example 7 This application also provides a embodied detection unmanned system, including an airborne multimodal sensor, a defect detection module, a memory, and a processor. The airborne multimodal sensor is communicatively connected to the processor. The memory stores a computer program, and when the processor executes the computer program, it implements the self-optimization method for cognitive parameters of the embodied unmanned system with multi-detection standard adaptation as described in any one of embodiments 1 to 6 above. The airborne multimodal sensor includes at least an image sensor, a depth sensor, an inertial measurement unit, and an odometer, used to collect image data, depth data, pose data, and odometer data, respectively. The defect detection module is the core functional unit of the unmanned system for performing defect identification, measurement, classification, and result output, and it internally operates the aforementioned cognitive parameters. The processor communicates with each sensor and the defect detection module through a bus or dedicated interface to acquire multimodal detection data in real time and execute the cognitive parameter self-optimization process. The embodied detection unmanned system can be an unmanned platform with autonomous mobility and airborne perception capabilities, such as a pipeline inspection robot, a pipe gallery inspection unmanned vehicle, or a tunnel inspection drone.

[0081] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0082] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for self-optimization of cognitive parameters adapted to multiple detection standards, characterized in that the steps include: include: Acquire the target detection standard corresponding to the current defect detection task and the historical detection records formed by the embodied unmanned system in historical defect detection tasks; Construct a standard difference vector based on the target detection standard and historically applicable detection standards; Collect multimodal detection data of the currently detected object from the embodied unmanned system, and extract the features of the currently detected object; Based on the standard difference vector and the characteristics of the current detection object, the cognitive parameter adjustment constraints corresponding to the target detection standard are determined, and the allowable range of the embodied execution parameters associated with the cognitive parameter adjustment constraints is further determined. Based on historical testing records, source testing standards are selected from historically applicable testing standards that have a standard mapping relationship with the target testing standard. The cognitive parameter migration weight from the source testing standard to the target testing standard is calculated according to the standard difference vector. The historical cognitive parameters corresponding to the source testing standard are then transformed according to the cognitive parameter migration weight to obtain candidate cognitive parameters under the target testing standard. The control system performs standard adaptation calibration actions within the allowable range of the embodied execution parameters and collects calibration feedback data. Based on the calibration feedback data, the standard adaptation loss of candidate cognitive parameters under the target detection standard is calculated. With the goal of reducing the standard adaptation loss, the candidate cognitive parameters are self-optimized within the cognitive parameter adjustment constraints to obtain the target standard-adapted cognitive parameters. The embodied unmanned system loads the target standard adaptation cognitive parameters and performs the current defect detection task based on the target standard adaptation cognitive parameters.

2. The cognitive parameter self-optimization method for multi-detection standard adaptation according to claim 1, characterized in that, The target detection standard includes defect category definition, defect measurement rules, defect level classification rules, and detection result output rules; The historical testing records include historical applicable testing standards, historical testing object attributes, historical cognitive parameters, historical defect identification results, and historical defect review results; The cognitive parameters include at least one of the following: defect category identification parameters, defect boundary understanding parameters, size measurement reasoning parameters, defect level determination parameters, detection standard mapping parameters, re-inspection decision parameters, and detection result expression parameters.

3. The cognitive parameter self-optimization method for multi-detection standard adaptation according to claim 1, characterized in that, The standard difference vector is used to characterize the differences between different detection standards in defect category definition, defect measurement benchmark, defect level threshold, and detection result expression. The construction process is as follows: The target testing standard and the historically applicable testing standard are analyzed separately to obtain the target standard rule description and the historical standard rule description. Extract the defect category field, defect measurement field, defect level field, review requirement field, and test result output field from the target standard rule description and the historical standard rule description respectively; Establish mapping relationships between target testing standards and historically applicable testing standards for defect categories, defect measurement benchmarks, defect level thresholds, and test result expression. The differences between the target testing standard and the historically applicable testing standard in terms of defect category mapping, defect measurement benchmark, defect level threshold, review requirements, and test result expression are calculated to obtain the standard difference vector.

4. The cognitive parameter self-optimization method for multi-detection standard adaptation according to claim 1, characterized in that, The cognitive parameter adjustment constraints include: standard mapping constraints, defect identification constraints, boundary understanding constraints, dimensional measurement reasoning constraints, defect level determination constraints, and re-inspection decision constraints. The standard mapping constraint is used to limit the mapping relationship between different detection standards in terms of defect categories, defect measurement rules, defect level thresholds, and result expression methods. The re-inspection decision constraints are used to limit the confidence range, measurement error range, grade boundary distance range, and multi-frame recognition consistency range for triggering re-inspection actions in suspected defect areas.

5. The cognitive parameter self-optimization method for multi-detection standard adaptation according to claim 1, characterized in that, The specific implementation of determining the cognitive parameter adjustment constraint corresponding to the target detection standard based on the standard difference vector and the current detection object features is as follows: When the standard difference vector characterizes the target detection standard's requirement for defect size measurement accuracy is higher than that of historically applicable detection standards, the accuracy constraint of the size measurement inference parameters is increased, the threshold tolerance of the defect level judgment parameters is reduced, and the re-inspection trigger range of suspected defect areas is expanded. When the standard difference vector represents the requirement of the target detection standard for the integrity of the defect boundary higher than that of the historically applicable detection standard, the integrity constraint of the defect boundary understanding parameters is increased, and the adjustment step size of the defect boundary judgment threshold is reduced. When the standard difference vector characterizes the defect level classification granularity of the target detection standard, which is higher than that of the historically applicable detection standards, the search step size of the size measurement inference parameters is reduced, and the number of local depth measurements is increased. When the surface reflection, occlusion, or texture interference of the current object being detected exceeds the preset interference threshold, the re-inspection decision parameters and the associated supplementary lighting intensity, shooting distance, and allowable variation range of sensor posture are adjusted.

6. The cognitive parameter self-optimization method for multi-detection standard adaptation according to claim 1, characterized in that, The selection of the source detection standard from historically applicable detection standards that have a standard mapping relationship with the target detection standard is specifically implemented as follows: Obtain the historical standard rule descriptions corresponding to multiple historically applicable testing standards; Calculate the standard rule similarity between the target detection standard and each historically applicable detection standard; Based on the similarity of standard rules, the results of historical defect verification, and the attributes of historical test objects, the source standard credibility of each historically applicable test standard is calculated; the source standard credibility is obtained by weighted fusion of standard rule similarity, parameter credibility, and environmental compatibility. The standard rule similarity is the similarity between the semantic embedding vector of the historical standard rule description of the target detection standard and the semantic embedding vector of the historical standard rule description of the currently applicable detection standard. The reliability of the parameter is the ratio of the number of records that passed the review to the total number of test records under the corresponding historical applicable test standard in all historical defect review results. The environmental compatibility refers to the ratio of the number of attribute fields that match the attribute fields of historically applicable detection standards for historically tested objects to the corresponding attribute fields in the current detection object's features, out of the total number of attribute fields. Historically applicable testing standards whose source standard credibility meets the preset credibility threshold are identified as source testing standards.

7. The cognitive parameter self-optimization method for multi-detection standard adaptation according to claim 6, characterized in that, The calculation of the cognitive parameter transfer weights from the source detection standard to the target detection standard based on the standard difference vector is specifically implemented as follows: Calculate the standard domain difference between the source detection standard and the target detection standard based on the standard difference vector; The reliability of parameters corresponding to historical cognitive parameters of the source detection standard is calculated based on the historical defect review results. The parameter reliability calculation process is as follows: extract all historical defect review results corresponding to the source detection standard from the historical detection records, calculate the single reliability score for each review result, and take a weighted average of all single reliability scores to obtain the parameter reliability. The calculation method for the single-item credibility score is as follows: if the defect identification result of the detection record is completely consistent with the review conclusion, the single-item credibility score is set as the first score value. If the defect categories are the same but the defect size deviation is within the standard allowable tolerance range, then the individual reliability score is set as the second score value; If the defect categories are the same but the defect level judgment deviation is level one, then the single-item reliability score is set as the third score value; if the defect categories are inconsistent, then the single-item reliability score is 0; the first score value, the second score value, and the third score value decrease in that order. Calculate the similarity between the detected objects and the characteristics of the current detected objects based on the attributes of the historical detected objects; Based on the relationship that is negatively correlated with the difference in the standard domain and positively correlated with the reliability of parameters and the similarity of the detected objects, the cognitive parameter transfer weights from the source detection standard to the target detection standard are calculated. The cognitive parameter transfer weights are obtained by nonlinearly fusing parameter confidence, similarity of the detected objects, and difference in the standard domain. The reliability of the parameters is incorporated into the fusion process in the form of a power function. The similarity of the detected objects is fused in the form of a power function, which is the ratio of the number of attribute fields that match the attributes of historical detected objects and the attributes of current detected objects to the total number of attribute fields. The standard domain difference quantity participates in the fusion in the form of a negative exponential function, which is the ratio of the L2 norm of the standard difference vector to the preset maximum difference norm reference value.

8. The cognitive parameter self-optimization method for multi-detection standard adaptation according to claim 6, characterized in that, The step of performing a migration transformation on the historical cognitive parameters corresponding to the source detection standard based on the cognitive parameter migration weight to obtain the candidate cognitive parameters under the target detection standard is specifically implemented as follows: The transfer sensitivity of various cognitive parameters is determined based on the standard difference vector. The specific calculation is as follows: a sensitivity correlation vector is preset for each type of cognitive parameter; the normalized standard difference vector and the sensitivity correlation vector of each type of cognitive parameter are respectively subjected to vector dot product operation to obtain the transfer sensitivity of the corresponding category of cognitive parameter. Based on the transfer weights and transfer sensitivity of cognitive parameters, transfer transformations are performed on the defect category identification parameters, defect boundary understanding parameters, dimensional measurement reasoning parameters, defect level judgment parameters, inspection standard mapping parameters, re-inspection decision parameters, and inspection result expression parameters in the historical cognitive parameters, respectively. k The class cognitive parameters, and their transfer transformation expression are: in, For the transformed first k Class cognitive parameters; For the source detection standard, the first k The original values ​​of the historical cognition parameters; For cognitive parameter transfer weights; For the first k Transfer sensitivity of cognitive parameters; The first one calculated based on the standard difference vector k For the standard adaptation adjustment of the cognitive parameters, for the size measurement inference parameters, the defect measurement benchmark difference component in the standard difference vector is mapped to the parameter offset of the measurement inference model; for the defect level judgment parameters, the defect level threshold difference component is mapped to the adjustment of the level judgment boundary; for the defect category identification parameters, the defect category mapping difference component is mapped to the weight adjustment of the category classifier; for the re-inspection decision parameters, the re-inspection requirement difference component is mapped to the threshold adjustment of the re-inspection triggering condition. The transformed cognitive parameters are projected onto the allowable range of parameters corresponding to the target detection standard to obtain candidate cognitive parameters.

9. The cognitive parameter self-optimization method for multi-detection standard adaptation according to claim 1, characterized in that, The standard adaptation calibration steps include: When a suspected defective area meets the re-inspection trigger conditions, the control system reduces its travel speed and re-inspects the suspected defective area. The control system adjusts the pitch or rotation angle of the sensors to obtain multi-angle detection data of suspected defect areas; The control system adjusts the intensity of the supplementary light to reduce the impact of reflections, shadows, or occlusions on defect boundary identification; The control system repeatedly locates the suspected defect location based on mileage and pose data; the control system performs local depth measurement on the suspected defect area to obtain dimensional measurement data of the suspected defect area.

10. A body-worn unmanned system, characterized in that, The invention includes an airborne multimodal sensor, a defect detection module, a memory, and a processor. The airborne multimodal sensor is communicatively connected to the processor. The memory stores a computer program. When the processor executes the computer program, it implements the cognitive parameter self-optimization method for multi-detection standard adaptation as described in any one of claims 1 to 9.

Images