A robot nondestructive testing method and system based on artificial intelligence

Abstract

The present invention relates to the field of robotics technology, and discloses a robot non-destructive testing method and system based on artificial intelligence. The method includes: acquiring multimodal sensing data and motion state parameters; constructing a three-dimensional material feature model and analyzing and generating an initial defect feature evaluation model; loading material properties and detection accuracy constraints to generate a comprehensive model, and combining motion parameter simulation to obtain path control data; training a dynamic defect prediction model to obtain a detection path correction model and generate initial detection parameters; obtaining dynamic correction simulation information based on the initial parameters, etc.; constructing a multi-objective optimization model to optimize and obtain optimal detection path parameters; combining real-time model reasoning to generate a safe detection probability, and adjusting the detection strategy to achieve dynamic path matching. The system includes modules such as data acquisition, feature modeling and analysis, comprehensive evaluation and simulation. The present invention improves detection accuracy, efficiency and adaptability, and is suitable for non-destructive testing of complex targets.

Description

Technical Field

[0001] The present invention relates to the field of robotics technology, and in particular to an artificial intelligence-based robot nondestructive testing method and system. Background Art

[0002] In industries such as industrial manufacturing, aerospace, energy, and transportation, nondestructive testing of materials and components is crucial for ensuring product quality and safe operation. Traditional nondestructive testing methods, such as visual inspection, ultrasonic testing, and radiographic testing, suffer from low efficiency, high human error, limited detection range, and insufficient adaptability to complex structures. The advancement of industrial automation and intelligentization is placing higher demands on the accuracy, efficiency, degree of automation, and ability to detect complex targets of nondestructive testing technologies.

[0003] Although existing robotic nondestructive testing technology has achieved automated testing to a certain extent, it still faces many challenges. On the one hand, single-modal sensor data is difficult to fully and accurately reflect the internal defect characteristics of the material, resulting in limited accuracy in defect identification and assessment. On the other hand, detection path planning is often based on fixed rules or simple algorithms, and cannot be adaptively adjusted according to the real-time changes in material properties, the dynamic characteristics of defect distribution, and the motion state of the robot. It is prone to problems such as blind spots in detection, repeated detection, or unstable detection accuracy. In addition, traditional methods lack effective feature fusion and analysis methods when processing multi-dimensional, multi-source heterogeneous data, making it difficult to build accurate material feature models and defect assessment models, which in turn affects the scientific nature and effectiveness of the detection strategy.

[0004] The rapid development of artificial intelligence technologies, such as machine learning, deep learning, and intelligent optimization algorithms, has provided new ideas and methods for solving the above-mentioned problems. Combining artificial intelligence with robotic nondestructive testing is expected to enable efficient processing and analysis of multimodal sensor data, build more accurate material and defect models, and achieve intelligent planning and dynamic adjustment of inspection paths, thereby improving the automation level of inspection and the reliability of test results. However, artificial intelligence-based robotic nondestructive testing technology is still in its developmental stage. Many technical challenges remain to be addressed, such as how to effectively integrate multimodal sensor data, build dynamic defect assessment models, intelligently optimize inspection paths, and adjust inspection strategies based on real-time feedback.

[0005] Therefore, there is an urgent need for an artificial intelligence-based robotic nondestructive testing method and system to overcome the shortcomings of existing technologies, improve the accuracy, efficiency and adaptability of nondestructive testing, and meet the needs of modern industry for high-quality nondestructive testing. Summary of the Invention

[0006] The purpose of the present invention is to provide a robot nondestructive testing method and system based on artificial intelligence to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a robot nondestructive testing method based on artificial intelligence, the method comprising:

[0008] Obtain multimodal sensing data of the target to be detected and the motion state parameters of the robot end effector;

[0009] Perform multi-dimensional scanning on the target and construct a three-dimensional material characteristic model. Combined with multi-modal sensing data, perform dynamic defect characteristic analysis on the three-dimensional material characteristic model to generate an initial defect characteristic assessment model.

[0010] The initial defect feature evaluation model is loaded with material property distribution data and detection accuracy constraints to generate a comprehensive defect feature evaluation model. The detection path dynamic simulation is performed in combination with the motion state parameters to obtain path control simulation data.

[0011] A dynamic defect prediction model is constructed and trained using path control simulation data to obtain a detection path correction model, thereby generating initial detection parameters. The initial detection parameters include at least initial detection point coordinates, initial scanning path, and initial correction response parameters.

[0012] Based on the initial detection parameters, the comprehensive defect feature evaluation model and the motion state parameters, dynamic correction simulation information is obtained;

[0013] Combining the dynamic correction simulation information with the inspection task planning parameters, a multi-objective optimization model is constructed and the planning parameters are iteratively optimized to obtain the optimal inspection path parameters;

[0014] Based on the real-time material feature model, the detection path parameters of the current period are inferred to generate the safe detection probability of the current period. Combined with the optimal detection path parameters of the current period and the actual scanning trajectory, the robot's real-time detection strategy is adjusted to achieve the dynamic path matching goal.

[0015] Preferably, the path control simulation data includes at least material stress distribution, defect clustering area, safety detection boundary set and dynamic correction priority sequence;

[0016] The dynamic correction simulation information at least includes detection point offset, path coverage completeness rate and accuracy stability evaluation index.

[0017] Preferably, the multi-dimensional scanning of the target and construction of a three-dimensional material characteristic model, performing dynamic defect characteristic analysis on the three-dimensional material characteristic model in combination with multimodal sensing data, and generating an initial defect characteristic evaluation model comprises the following steps:

[0018] Perform a composite scan of radiation and sound waves on the target to obtain a sensor data sequence of the internal defect distribution of the material;

[0019] Preprocessing the sensor data sequence to obtain standardized feature fusion data, wherein the preprocessing includes one or more of signal denoising, time domain registration, feature fusion, grid division, and data interpolation;

[0020] Generate a 3D material feature model based on standardized feature fusion data and combined with 3D point cloud reconstruction technology;

[0021] A dynamic defect feature analysis is performed on the three-dimensional material feature model to generate an initial defect feature assessment model, wherein the dynamic defect feature analysis includes at least grid density allocation and feature weight mapping, the defect risk level is divided by grid density allocation, and corresponding sensor data parameters are assigned to each level area.

[0022] Preferably, the method of loading the initial defect feature assessment model with material property distribution data and detection accuracy constraints to generate a comprehensive defect feature assessment model, performing dynamic simulation of the detection path in combination with motion state parameters to obtain path control simulation data includes the following steps:

[0023] Loading material property distribution data into the initial defect feature assessment model to simulate real-time property changes to generate a first defect feature assessment model;

[0024] Loading the first defect feature assessment model with detection accuracy constraints to generate a comprehensive defect feature assessment model, wherein the detection accuracy constraints include a maximum detection angle limit, a minimum resolution threshold, and a sensor response redundancy parameter;

[0025] Based on the comprehensive defect feature evaluation model and motion state parameters, dynamic simulation of the detection path is performed. The specific process includes:

[0026] Construct a multi-objective path planning equation, which includes at least a coverage integrity equation, a trajectory coincidence equation, and a detection efficiency equation. Combined with a comprehensive defect feature evaluation model, the multi-objective path planning equation is numerically solved using a piecewise gradient descent method to obtain material stress distribution, defect clustering areas, a safe detection boundary set, and a dynamic correction priority sequence.

[0027] The generation of the security detection boundary set includes the following steps:

[0028] Based on the defect clustering area, the defect coverage density of each grid cell is calculated;

[0029] Identify areas in the comprehensive defect feature assessment model where the defect coverage density is no greater than a preset threshold and generate a set of safe detection boundaries.

[0030] Preferably, the construction of a dynamic defect prediction model and training it with path control simulation data to obtain a detection path correction model, and then generate initial detection parameters, includes the following steps:

[0031] Construct a dynamic defect prediction model based on a three-dimensional material feature model;

[0032] The dynamic defect prediction model is trained and verified through path control simulation data to obtain the detection path correction model;

[0033] Input the real-time motion state parameters into the detection path correction model to predict the safe detection boundary set;

[0034] Based on the predicted safe detection boundary set, initial detection parameters are generated, wherein the initial detection parameters at least include initial detection point coordinates, an initial scanning path, and initial correction response parameters.

[0035] Preferably, a dynamic defect prediction model is constructed and trained with path control simulation data to obtain a detection path correction model, thereby generating initial detection parameters. This also includes data reconstruction of the path control simulation data, specifically:

[0036] Construct an initial multi-dimensional detection input tensor based on material stress distribution, defect clustering areas, and safe detection boundary sets;

[0037] Normalize and reduce the feature dimensionality of the initial multi-dimensional detection input tensor to generate the final multi-dimensional detection input tensor;

[0038] Based on the dynamic correction priority sequence, the detection path correction label tensor is constructed;

[0039] The final multi-dimensional detection input tensor and the detection path correction label tensor are combined to form a training sample set.

[0040] Preferably, generating initial detection parameters based on the predicted safety detection boundary set includes the following steps:

[0041] Extract the grid cell with the lowest defect coverage density in the predicted safety detection boundary set and generate the coordinates of the initial detection point;

[0042] According to the spatial topological relationship of the predicted safety detection boundary set, the feasible connection structure of the initial scanning path is fitted to generate the initial correction response parameters;

[0043] The material stress gradient direction of the predicted safety detection boundary set is calculated and normalized into a path reference vector, which is the initial scanning path direction.

[0044] Preferably, obtaining dynamic correction simulation information based on initial detection parameters, comprehensive defect feature evaluation model and motion state parameters comprises the following steps:

[0045] Map the initial inspection point coordinates to the comprehensive defect feature evaluation model, match the initial scanning path with the initial correction response parameters, adjust the path node density in the correction area, update the comprehensive defect feature evaluation model, define the dynamic correction trigger conditions, path reconstruction rules and response parameter increments, and generate a dynamic correction simulation model;

[0046] Based on the dynamic correction simulation model, the multi-objective path planning equation is iteratively solved by the piecewise gradient descent method to obtain dynamic correction simulation information, including:

[0047] When the dynamic correction trigger conditions are met, the detection point offset, path coverage completeness, and accuracy stability evaluation indicators are updated, and the multi-objective path planning equation is re-solved until the simulation termination conditions are met;

[0048] The dynamic correction trigger condition includes triggering the response parameter update when the current path coverage completeness rate is not greater than the preset completeness rate threshold; the path reconstruction rule includes adjusting the initial scanning path based on the material stress gradient direction; the response parameter increment is in a piecewise linear relationship with the current precision stability evaluation index.

[0049] Preferably, the construction of a multi-objective optimization model and iterative optimization of planning parameters to obtain optimal detection path parameters includes the following steps:

[0050] A multi-objective optimization model was constructed, where the optimization variables included path node distribution density, scanning angle threshold, and sensor sampling frequency. The optimization objectives included maximizing path coverage efficiency and minimizing detection energy consumption. The constraints included the robot kinematic upper limit and sensor accuracy threshold.

[0051] Perform a preliminary solution to the multi-objective optimization model through a dynamic programming algorithm to generate an initial set of optimization paths;

[0052] Based on the initial optimized path set, the ant colony optimization algorithm is used for global optimization to generate the optimal detection path parameters.

[0053] Preferably, the present invention further includes a robot nondestructive testing system based on artificial intelligence, the system comprising:

[0054] The data acquisition module is used to obtain the multimodal sensing data of the target to be detected and the motion state parameters of the robot end effector;

[0055] The feature modeling and analysis module is used to perform multi-dimensional scanning of the target and build a three-dimensional material feature model. It combines multi-modal sensing data to perform dynamic defect feature analysis on the three-dimensional material feature model and generate an initial defect feature assessment model.

[0056] The comprehensive evaluation and simulation module is used to load material property distribution data and detection accuracy constraints into the initial defect feature evaluation model, generate a comprehensive defect feature evaluation model, and perform dynamic simulation of the detection path in combination with motion state parameters to obtain path control simulation data;

[0057] A parameter generation module is used to construct a dynamic defect prediction model and train it through path control simulation data to obtain a detection path correction model, thereby generating initial detection parameters, wherein the initial detection parameters at least include initial detection point coordinates, initial scanning path, and initial correction response parameters;

[0058] Dynamic correction simulation module, used to obtain dynamic correction simulation information based on initial detection parameters, comprehensive defect feature evaluation model and motion state parameters;

[0059] The path optimization module is used to combine the dynamic correction simulation information with the inspection task planning parameters, build a multi-objective optimization model and iteratively optimize the planning parameters to obtain the optimal inspection path parameters;

[0060] The strategy adjustment module is used to infer the detection path parameters of the current period based on the real-time material feature model, generate the safe detection probability of the current period, combine the optimal detection path parameters of the current period with the actual scanning trajectory, and adjust the robot's real-time detection strategy to achieve the dynamic path matching goal.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] In terms of detection accuracy, by acquiring multimodal sensor data of the target to be detected and combining it with technologies such as X-ray and acoustic wave composite scanning, information on the internal defect distribution of the material can be obtained more comprehensively and accurately. The sensor data is preprocessed through signal denoising, time domain registration, and feature fusion to generate standardized feature fusion data. Based on this, a three-dimensional material feature model is constructed, achieving accurate modeling of material characteristics. Through dynamic defect feature analysis, including grid density allocation and feature weight mapping, corresponding sensor data parameters can be assigned to each area according to the defect risk level, making defect feature assessment more accurate. Loading material property distribution data and detection accuracy constraints to generate a comprehensive defect feature assessment model further considers the real-time changes in material properties and the requirements of detection accuracy, improving the accuracy and reliability of defect assessment.

[0063] In terms of detection efficiency, a multi-objective path planning equation is constructed, including the coverage integrity equation, the trajectory coincidence equation, and the detection efficiency equation. Combined with a comprehensive defect feature evaluation model, this equation is numerically solved using the piecewise gradient descent method to generate an efficient detection path. The multi-objective optimization model is solved using a dynamic programming algorithm and an ant colony optimization algorithm, enabling iterative optimization of the detection path parameters to obtain the optimal detection path parameters. This effectively reduces redundant paths and repeated detection during the detection process, thereby improving detection efficiency. Furthermore, through dynamic simulation and dynamic correction simulation of the detection path, the detection path can be adjusted promptly based on real-time feedback, avoiding detection blind spots and unreasonable paths, further improving detection efficiency.

[0064] In terms of adaptability, the system can adjust detection strategies in real time based on the motion state parameters of the robot's end-effector to achieve dynamic path matching. By constructing a dynamic defect prediction model and training it with path control simulation data, it can predict the set of safe detection boundaries and generate initial detection parameters suitable for different detection scenarios. When the dynamic correction trigger conditions are met during the detection process, the system can automatically update the detection point offset, path coverage completeness, and accuracy stability evaluation indicators, and re-solve the multi-objective path planning equations, enabling the detection system to adapt to complex situations such as changes in material properties and dynamic adjustments to defect distribution, thereby improving the system's adaptability and robustness.

[0065] In terms of data processing and model building, multi-dimensional scanning and feature fusion of multimodal sensor data were performed to construct a three-dimensional material feature model and a comprehensive defect feature assessment model. This fully leverages the advantages of multi-source heterogeneous data, improving data processing efficiency and model accuracy. By reconstructing the path control simulation data, constructing the initial multi-dimensional detection input tensor, and performing standardization and feature dimensionality reduction, a training sample set was generated. This provided high-quality data support for the training of the dynamic defect prediction model, improving the model's performance and generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a working principle diagram of the artificial intelligence-based robot nondestructive testing method of the present invention;

[0067] Figure 2 A flowchart of the dynamic defect prediction model training and initial detection parameter generation according to the present invention;

[0068] Figure 3 A flowchart of generating initial detection parameters based on a safety detection boundary set according to the present invention;

[0069] Figure 4 This is a flow chart of the dynamic correction simulation information generation and path iterative optimization described in the present invention. DETAILED DESCRIPTION

[0070] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0071] See also Figures 1-4 The present invention relates to a robot nondestructive testing method based on artificial intelligence, and the specific implementation steps are as follows:

[0072] By deploying a variety of sensors at the end of the robot (such as radiation sensors, acoustic sensors, visual sensors, etc.), multimodal sensing data of the target to be detected is collected in real time, including but not limited to the internal structure signals of the material, surface feature information, etc.; at the same time, the motion state parameters of the end effector, such as position, velocity, acceleration, etc., are obtained through the robot control system.

[0073] The target is scanned in multiple dimensions, and the data obtained from the scanning is used to construct a three-dimensional material feature model. The model is then dynamically analyzed for defect characteristics using multimodal sensing data to generate an initial defect feature assessment model.

[0074] The material property distribution data and detection accuracy constraints are loaded into the initial defect feature evaluation model to generate a comprehensive defect feature evaluation model. The detection path dynamic simulation is performed in combination with the motion state parameters to obtain the path control simulation data.

[0075] A dynamic defect prediction model is constructed and trained through path control simulation data to obtain a detection path correction model, and then generate initial detection parameters including initial detection point coordinates, initial scanning path and initial correction response parameters.

[0076] Based on the initial detection parameters, the comprehensive defect feature evaluation model and the motion state parameters, the dynamic correction simulation information is obtained.

[0077] Combining the dynamic correction simulation information with the detection task planning parameters, a multi-objective optimization model is constructed and the planning parameters are iteratively optimized to obtain the optimal detection path parameters.

[0078] Based on the real-time material feature model, the detection path parameters of the current period are inferred to generate the safe detection probability of the current period. Combined with the optimal detection path parameters of the current period and the actual scanning trajectory, the robot's real-time detection strategy is adjusted to achieve the dynamic path matching goal.

[0079] The present invention will be further described below in conjunction with Examples 1 to 5:

[0080] Example 1:

[0081] This embodiment defines in detail the specific composition and generation logic of path control simulation data and dynamic correction simulation information. The following describes the definition, composition, generation principle and role of the two types of data in the detection process:

[0082] Path control simulation data is the key intermediate data generated during the dynamic simulation of the detection path, which is used to provide a basis for subsequent detection path correction and optimization. It at least includes material stress distribution, defect clustering area, safe detection boundary set and dynamic correction priority sequence. Among them, the generation of material stress distribution is based on the simulation calculation of the properties of the detection target material. During the dynamic simulation of the detection path, the system will load the elastic modulus, Poisson's ratio and other property parameters of the material into the comprehensive defect feature evaluation model, and simulate the stress state of the material during the detection process through methods such as finite element analysis, and then obtain the stress numerical distribution of each area. The distribution is based on the grid unit as the basic unit, and each unit corresponds to a stress value, which reflects the distribution of the material's ability to resist deformation. High stress areas may have potential structural weaknesses, which need to be paid special attention to in the detection path planning.

[0083] Defect clustering regions are determined by statistically analyzing the defect features in multimodal sensing data. After generating the initial defect feature assessment model, the system applies a clustering algorithm (such as the DBSCAN algorithm or the K-means algorithm) to the defect features in the model, grouping defects with similar spatial locations and features into the same cluster, thereby identifying areas where defects are concentrated. Each clustering region contains information such as the location coordinates, size, and type of the defect. This information helps determine the degree of damage and development trend of the inspection target. For example, the aggregation of multiple small defects may indicate the impending formation of a larger crack. Therefore, the determination of defect clustering regions can provide direct guidance for the key coverage areas of the inspection path.

[0084] The generation of the safe detection boundary set depends on the calculation and analysis of the defect coverage density. First, the system divides the target area in the comprehensive defect feature assessment model into several grid cells. For each grid cell, the number of internal defects or the proportion of the defect area to the total area of ​​the grid cell is calculated to obtain the defect coverage density. The preset threshold is pre-set according to the detection standard and the characteristics of the target material. For example, for aerospace materials with high reliability requirements, the preset threshold may be set lower to ensure that low defect density areas are also fully detected. When the defect coverage density of a grid cell is not greater than the preset threshold, the cell is judged as a low-risk area, and its boundary is extracted and integrated into a safe detection boundary set. This set defines the safe range of the detection path, which ensures the comprehensiveness of the detection while avoiding excessive detection of low-risk areas, thereby improving detection efficiency.

[0085] The dynamic correction priority sequence is the order of inspection path correction determined according to the defect risk level and inspection requirements. In the initial defect feature assessment model, the defect risk level of each area has been divided (such as high, medium, and low) through grid density allocation and feature weight mapping. The dynamic correction priority sequence first sets a higher correction priority for the inspection path segments corresponding to high-risk areas to ensure that the paths of these areas are adjusted first during the inspection process to deal with defects that may develop rapidly; the second is the medium-risk area, and the last is the low-risk area. The setting of the priority sequence also takes into account the motion state parameters of the robot's end effector. For example, when the robot moves faster in a certain area, in order to avoid missed detection, the priority of the path correction in that area may be appropriately increased.

[0086] Dynamic correction simulation information is data generated based on initial detection parameters, a comprehensive defect feature evaluation model, and motion state parameters to guide the dynamic correction of the detection path. It includes at least detection point offset, path coverage completeness, and accuracy stability evaluation indicators. The detection point offset refers to the spatial deviation between the detection point position of the robot's end effector and the planned initial detection point coordinates during the actual detection process. This deviation is measured and obtained in real time by the robot's position sensor and compared and calculated with the coordinate system in the comprehensive defect feature evaluation model. The size of the detection point offset reflects the accuracy of the robot's motion control and the execution error of the detection path. When the offset exceeds the preset allowable range, the path correction mechanism needs to be triggered to adjust the subsequent detection path.

[0087] The path coverage completeness rate is used to measure the coverage of the target area by the inspection path. It is calculated as the ratio of the area of ​​the area that has been inspected to the total area of ​​the target area. During the dynamic correction simulation process, the system will track the robot's scanning trajectory in real time, mark the scanned area in the comprehensive defect feature evaluation model, and calculate the coverage area. The preset completeness rate threshold is set according to the requirements of the inspection task. For example, for the inspection of key components, the threshold may be set to more than 95%. When the path coverage completeness rate is not greater than the preset threshold, it indicates that there is a coverage gap in the inspection path, and supplementary inspection is required by adjusting the inspection point coordinates, scanning path direction, etc. to ensure the comprehensiveness of the inspection.

[0088] The accuracy stability evaluation index is derived by analyzing the fluctuations in sensor data during the inspection process. When multimodal sensors collect data, they may be affected by factors such as environmental noise and sensor drift, resulting in data fluctuations. The system performs statistical analysis on the real-time data collected by the sensors, calculating statistics such as the standard deviation and coefficient of variation as an accuracy stability evaluation index. This index reflects the stability of the inspection accuracy. A high value indicates significant fluctuations in inspection accuracy, which may affect the accurate identification of defect features. In this case, it is necessary to adjust parameters such as the sensor sampling frequency and inspection angle, or calibrate the sensor to improve the stability of the inspection accuracy.

[0089] In the inspection process, path control simulation data and dynamic correction simulation information are interrelated and work together. Path control simulation data is the basis for the generation of dynamic correction simulation information. For example, the safe detection boundary set provides the basis for the generation of the initial detection point coordinates, and the dynamic correction priority sequence affects the setting of the response parameter increment. Dynamic correction simulation information, on the other hand, is the feedback and adjustment of path control simulation data during the actual inspection process. Through indicators such as detection point offset and path coverage completeness, the system can evaluate the execution effect of the inspection path in real time and dynamically correct the inspection path based on information such as material stress distribution and defect clustering areas in the path control simulation data. This closed-loop feedback mechanism ensures that the robotic nondestructive testing process can adaptively adjust the inspection strategy according to the actual state of the target material and the inspection requirements, thereby improving the accuracy and efficiency of the inspection.

[0090] Example 2:

[0091] This example details the implementation process of performing multi-dimensional scanning on a target, constructing a three-dimensional material feature model, and generating an initial defect feature assessment model based on multimodal sensing data. This process, through the organic integration of composite scanning technology, data preprocessing, three-dimensional reconstruction, and dynamic analysis, enables accurate modeling and preliminary assessment of internal defects in the target. The details are as follows:

[0092] First, a combined X-ray and acoustic wave scan of the target is performed to acquire a sequence of sensor data. In practice, the robot's end effector must be equipped with at least two types of sensors: X-ray sensors (such as X-ray or gamma-ray sensors) and acoustic wave sensors (such as ultrasonic sensors). X-ray sensors transmit radiation through the target material, and based on the varying attenuation of the radiation after passing through the material, they obtain information on the material's internal density distribution, defects such as cracks or pores. Acoustic sensors transmit acoustic signals (such as longitudinal, shear, or surface waves) into the target material and, by receiving reflected, refracted, or scattered acoustic signals, analyze the material's internal interface characteristics, defect location, and size. The two sensors operate synchronously to perform a multi-dimensional scan of the target: X-ray scanning provides the material's overall density distribution and macroscopic defect outlines, while acoustic scanning provides additional information on the material's microstructural characteristics and fine-scale defect boundaries. During the scanning process, the robot's control system precisely controls the sensor's position, scanning angle, and movement speed, ensuring that the scanning path covers the entire target area and avoids missing critical inspection areas. After the scan is completed, the two types of sensors output raw sensor data respectively, forming a sensor data sequence containing information such as time series, spatial coordinates and signal strength. This sequence serves as the basic data for subsequent analysis.

[0093] The sensor data sequence is preprocessed to generate standardized feature-fused data. This preprocessing step includes various processing methods, such as signal denoising, time-domain registration, feature fusion, gridding, and data interpolation. These methods are combined based on the data characteristics and detection requirements. Signal denoising aims to remove noise interference from the raw data. Common methods include Fourier transform filtering and wavelet denoising. For example, for random noise generated by environmental radiation in X-ray sensor data, a wavelet threshold denoising algorithm can be used to retain the effective signal components and suppress noise. For periodic noise generated by sensor vibration in acoustic sensor data, a bandpass filter can be used for frequency domain filtering. Time-domain registration is used to address the problem of asynchrony between multi-sensor data in the temporal dimension. Through timestamp alignment or linear interpolation, the sampling frequencies of the X-ray and acoustic sensor data sequences are unified on the time axis to ensure temporal correspondence of data points during subsequent feature fusion. Feature fusion integrates the features of multimodal sensor data to form a unified feature representation vector. Specifically, the attenuation coefficient characteristics of the ray data and the sound velocity and acoustic impedance characteristics of the acoustic wave data are normalized (e.g., to the interval [0,1]), and then concatenated by dimension to generate a feature vector containing multi-source information. This vector can more comprehensively reflect the internal characteristics of the material. Meshing involves dividing the target material into regular or irregular grid cells in three-dimensional space. The size of the grid is determined based on the required detection accuracy. For example, for critical areas requiring high-precision detection, a smaller grid size (e.g., millimeter level) is used to improve the accuracy of defect location; for non-critical areas, a larger grid size (e.g., centimeter level) can be used to reduce the amount of data processing. Data interpolation is used to fill in areas where data is missing due to scanning blind spots or sensor resolution limitations. Common methods include nearest neighbor interpolation, bilinear interpolation, or kriging interpolation. For example, when ray scanning results in sparse data at the edge of the target due to angle limitations, the missing data can be estimated using the kriging interpolation method using the eigenvalues ​​of adjacent grid cells, thereby improving the spatial continuity and integrity of the data.

[0094] Based on standardized feature fusion data, a 3D material feature model is generated using 3D point cloud reconstruction technology. The core of 3D point cloud reconstruction technology is to convert 2D sensor data into a point cloud in 3D space, thereby constructing a 3D structural model of the target material. The specific steps are as follows: First, based on the spatial position and scanning angle of the sensor, the feature fusion data of each grid cell is mapped to the corresponding position in the 3D coordinate system, generating point cloud data containing X, Y, and Z coordinates and characteristic attributes (such as density and sound velocity). Point cloud filtering algorithms (such as statistical filtering and radius filtering) are then used to remove outliers and improve the quality of the point cloud data. Next, point cloud meshing algorithms (such as Delaunay triangulation or Poisson surface reconstruction) are used to connect the discrete point cloud data into a continuous surface mesh, forming the geometric structure of the 3D material feature model. This model not only intuitively displays the external outline of the target material but also visualizes the distribution of characteristic attributes within the material through color coding or transparency settings, such as using red to indicate high-density areas and blue to indicate low-density areas, making it easier for operators to quickly locate potential defects. In addition, the 3D material feature model supports interactive operations such as rotation, scaling, and cutting, so that the internal structure of the material can be observed from different angles.

[0095] Performing dynamic defect feature analysis on the 3D material feature model to generate an initial defect feature assessment model is a key step in this embodiment. Dynamic defect feature analysis includes two core steps: mesh density allocation and feature weight mapping. Mesh density allocation dynamically adjusts the density of mesh cells in the 3D model based on the defect risk level. First, the system uses a preset defect recognition algorithm (such as a threshold-based segmentation algorithm or a machine learning classification algorithm) to detect defects in each mesh cell in the 3D material feature model, determining the presence and severity of defects. For mesh cells detected with defects, a risk level (e.g., high, medium, or low) is assigned based on the size, type, and number of defects. High-risk cells (e.g., cells containing large cracks or dense pores) are subdivided using a smaller mesh size, for example, by dividing a single mesh cell into eight subcells, to improve detection resolution in that area. Medium-risk cells retain their original mesh size, while low-risk cells have their mesh size increased to reduce computational complexity. Through this adaptive grid density distribution, the system can concentrate computing resources in defect-concentrated areas, improving the accuracy of defect feature analysis, while reducing computational complexity in low-risk areas and improving overall processing efficiency.

[0096] Feature weight mapping is to assign corresponding sensor data parameter weights to areas with different risk levels in order to highlight the importance of features in high-risk areas. Specifically, for grid cells in high-risk areas, the weight values ​​of key features such as the ray attenuation coefficient and the acoustic wave reflection intensity are increased (for example, the weight coefficient is set to 1.5), so that these features occupy a greater proportion of decision-making in subsequent defect assessments; for medium-risk areas, the weight coefficient is set to 1.0 to maintain the importance of the original features; for low-risk areas, the weight coefficient is set to 0.8 to appropriately reduce the impact of the features. The setting of the weight coefficient is determined by a trained machine learning model (such as a random forest or a neural network). The model is trained based on historical detection data and defect assessment results, and can automatically learn the importance of different features in defect identification. Through feature weight mapping, the system can pay more attention to feature changes in high-risk areas when generating the initial defect feature assessment model, thereby improving the ability to identify early defects or potential defects.

[0097] After generating the initial defect signature assessment model, it serves as the foundational data for subsequent processes, loading material property distribution data and inspection accuracy constraints to generate a comprehensive defect signature assessment model. Furthermore, the defect risk classification and feature weighting information in the initial model provide input parameters for the multi-objective path planning equations used in dynamic inspection path simulation. For example, path segments corresponding to high-risk areas will be assigned higher coverage integrity requirements and inspection efficiency priorities.

[0098] Example 3:

[0099] This example describes in detail the process of comprehensively processing the initial defect feature assessment model to generate a comprehensive defect feature assessment model, and then dynamically simulating the inspection path in combination with motion state parameters to obtain path control simulation data. This process achieves dynamic simulation and optimization of the inspection path through steps such as material property simulation, inspection accuracy constraint loading, multi-objective path planning, and boundary set generation. The details are as follows:

[0100] Material property distribution data is loaded into the initial defect signature assessment model to simulate real-time property changes, generating a primary defect signature assessment model. This material property distribution data encompasses material physical parameters such as elastic modulus, density, coefficient of thermal expansion, and Poisson's ratio. These parameters are either pre-set or acquired through real-time sensor data based on the target material type (e.g., metal, composite, ceramic) and specific operating conditions (e.g., temperature and pressure). During the loading process, the system maps the material property parameters according to their spatial distribution to each grid cell in the initial model. For example, regions with large temperature gradients in metal components are assigned temperature-dependent elastic modulus parameters to simulate material property fluctuations under thermal stress. This simulation enables the primary defect signature assessment model to more realistically reflect the material's behavior under actual operating conditions, making subsequent inspection path planning more tailored to practical needs. For example, in aircraft engine blade inspection, the impact of the decrease in material elastic modulus under high temperature on defect development is considered. By loading temperature-related property data, the model can predict the defect's propagation trend under thermal stress and thus adjust key inspection areas.

[0101] Detection accuracy constraints are applied to the first defect feature assessment model to generate a comprehensive defect feature assessment model. Detection accuracy constraints are key parameters for ensuring the reliability of detection results. They include the maximum detection angle limit, the minimum resolution threshold, and sensor response redundancy parameters. The maximum detection angle limit is set based on the sensor's physical properties and detection principle. For example, ultrasonic sensors have an optimal detection angle range. When the detection angle exceeds this range, the intensity of the acoustic reflection signal significantly attenuates, resulting in a decrease in defect recognition rate. Therefore, the detection angle must be limited to the sensor's effective operating range. The minimum resolution threshold is determined based on the detection standard and defect type. For example, for crack defects, the minimum crack length that the sensor can detect must be set (e.g., 0.5 mm) to ensure that defects smaller than this threshold are not ignored in the model. Sensor response redundancy parameters are used to address possible abnormal sensor responses. For example, a redundancy threshold for sensor signal strength is set. When the actual detection signal strength falls below the redundancy threshold, a sensor calibration or switchover to a backup sensor is triggered. When loading the detection accuracy constraints, the system checks each grid unit in the first defect feature evaluation model to ensure that its corresponding detection angle, resolution and sensor response parameters meet the constraints, marks or adjusts the units that do not meet the conditions, and finally generates a comprehensive defect feature evaluation model that includes dual constraints of material properties and detection accuracy.

[0102] The core of dynamic simulation of the detection path based on the comprehensive defect feature evaluation model and motion state parameters is to construct and solve the multi-objective path planning equation. The multi-objective path planning equation includes at least the coverage integrity equation, the trajectory overlap equation and the detection efficiency equation. Each equation describes the different optimization goals of the detection path through mathematical expressions. The coverage integrity equation aims to maximize the coverage area of ​​the target area of ​​the detection path to avoid undetected blind spots. Its expression is usually statistically calculated based on the coverage status of the grid cells (covered or uncovered); the trajectory overlap equation aims to minimize the repeated scanning area of ​​the detection path. By calculating the overlapping area ratio of the path trajectory, redundant detection is reduced and detection efficiency is improved; the detection efficiency equation aims to minimize the detection time or energy consumption. Combined with the motion speed, acceleration and other parameters of the robot end effector, the continuity and smoothness of the path are optimized.

[0103] To solve the multi-objective path planning equation, the system uses a piecewise gradient descent method, dividing the entire inspection area into multiple subareas. The optimal path is then solved within each subarea, and the global path is then obtained by concatenating the paths from each subarea. The specific steps are as follows: First, the grid cells in the integrated defect signature assessment model are grouped according to defect risk level (high, medium, and low). Path planning prioritizes grid cells in the high-risk group to ensure inspection accuracy in critical areas. Second, within each subarea, the coverage integrity equation is used as the primary optimization objective. Combined with the trajectory coincidence equation and the inspection efficiency equation, the coordinates of the path nodes are iteratively updated using a gradient descent algorithm until a local optimal solution is reached. Finally, the local optimal paths of each subarea are smoothed to avoid sudden changes at path junctions, ultimately forming a complete inspection path. The path control simulation data obtained using this method includes material stress distribution, defect clustering areas, a set of safe detection boundaries, and a dynamic correction priority sequence. The material stress distribution, derived from material property parameters in the integrated defect signature model and mechanical analysis during the path planning process, reflects the stress state at each point along the inspection path, providing a mechanical basis for path correction.

[0104] The process of generating the safety detection boundary set is as follows: First, based on the distribution of the defect clustering area, the defect coverage density of each grid cell is calculated. The defect coverage density is calculated as the ratio of the area or number of defects in the cell to the total area of ​​the cell. This ratio reflects the density of defects in the cell. For example, for a grid cell, if it contains 3 pores with sizes of 1mm², 2mm², and 1.5mm², the defect coverage density is (1+2+1.5) / cell area. The preset threshold is set according to the detection standard and the allowable defect level of the target material. For example, for pressure vessel materials, the preset threshold may be set to 5%, that is, when the unit defect coverage density is less than or equal to 5%, the unit is considered to belong to the low-risk area.

[0105] After identifying areas within the comprehensive defect feature assessment model where the defect coverage density does not exceed a preset threshold, the system uses edge detection algorithms (such as the Canny or Sobel algorithms) to extract the boundaries of these low-risk areas and generate a safe inspection boundary set. This set, composed of a series of continuous boundary segments or surfaces, clearly demarcates high-risk areas requiring focused inspection from low-risk areas where inspection density can be appropriately reduced. In inspection path planning, the safe inspection boundary set is used to limit the path's expansion range. Specifically, the inspection path primarily performs intensive scanning within the high-risk areas enclosed by the boundary set, while low-risk areas outside the boundary set can be sparsely scanned or skipped, thereby reducing inspection time and energy consumption while ensuring inspection quality. For example, in pipeline weld inspection, the safe inspection boundary set can be limited to high-risk areas such as the weld and its heat-affected zone, while low-risk areas within the pipeline body can be subject to periodic spot checks.

[0106] The generation of the dynamic correction priority sequence is related to the defect risk level and the complexity of the detection path. In high-risk areas, due to the rapid development of defects and the great impact on structural safety, the corresponding detection path segments are given the highest priority and are adjusted first during the dynamic correction process; the path segments in medium-risk areas have the second highest priority; and the path segments in low-risk areas have the lowest priority. In addition, the complexity of the path (such as the number of path nodes and the number of turns) will also affect the priority. Complex path segments may be given a higher correction priority because they are more prone to execution errors. The dynamic correction priority sequence is stored in the form of a list, with each element corresponding to the identifier and priority value of a path segment. The system corrects the path segments in sequence according to the sequence to ensure the orderliness and efficiency of the detection process.

[0107] During the dynamic simulation of the inspection path, motion state parameters (such as the position, velocity, and acceleration of the robot's end effector) are input into the simulation system in real time to verify the feasibility of the path and the robot's reachability. For example, if the simulated path contains angles that are beyond the range of motion of the robot's joints, the system automatically marks that path segment and triggers a path reconstruction mechanism to adjust the positions or angles of the path nodes until the path satisfies the robot's kinematic constraints. Furthermore, the motion state parameters are used to calculate the execution time and energy consumption of the path, providing data support for solving the inspection efficiency equation.

[0108] Example 4:

[0109] This example describes in detail the complete process of building a dynamic defect prediction model, training the model with path control simulation data to generate a detection path correction model, and generating initial detection parameters based on the model. It also covers the specific implementation steps of data reconstruction. The following details the model construction, data processing, parameter generation, and other aspects:

[0110] A dynamic defect prediction model is constructed based on a three-dimensional material feature model. The three-dimensional material feature model has been generated through multi-dimensional scanning and point cloud reconstruction technology, and contains the geometric structure of the target material, the distribution of internal characteristic properties (such as density and sound velocity), and grid unit division information. The dynamic defect prediction model uses an artificial intelligence algorithm architecture, such as a convolutional neural network (CNN), a recurrent neural network (RNN), or a graph neural network (GNN). Its input is the grid unit feature vector of the three-dimensional model (including standardized multimodal sensor data), and its output is the defect probability value or defect type label of each grid unit. The choice of model architecture depends on the spatial distribution characteristics of the defect characteristics: for defects with local spatial correlation (such as crack propagation), CNN is more suitable for capturing two-dimensional or three-dimensional spatial features; for defect development processes that change over time, RNN or temporal convolutional network (TCN) is more suitable for processing sequence data; for models with complex topological relationships between grid units, GNN can better model the dependencies between nodes. During the model construction process, hyperparameters such as the number of network layers, the number of neurons, and the activation function need to be configured according to the characteristics of the detection target. For example, in aerospace material detection, due to the complexity of defect types, a model structure combining a deep CNN with an attention mechanism can be used to enhance the ability to capture subtle defect features.

[0111] The dynamic defect prediction model is trained and validated using path control simulation data. This process requires data reconstruction. The first step in data reconstruction is to construct an initial multi-dimensional detection input tensor, which integrates the material stress distribution, defect clustering area, and safety detection boundary set information from the path control simulation data. Specifically, the material stress distribution value, the grid cell identifier of the defect clustering area (e.g., belonging to the nth cluster), and the grid cell identifier of the safety detection boundary set (e.g., 0 for non-boundary cells and 1 for boundary cells) are concatenated in grid cell order to form a three-dimensional tensor (with dimensions of the number of grid cells × feature dimension × 1). For example, each grid cell corresponds to a feature vector of length 3, which contains the stress value, cluster number, and boundary identifier. The shape of the initial tensor is (N, 3, 1), where N is the total number of grid cells.

[0112] Standardizing and reducing the dimensionality of the initial multi-dimensional detection input tensor are key steps in data reconstruction. Standardization normalizes the data in each dimension to a distribution with zero mean and unit variance by subtracting the feature mean and dividing by the standard deviation, thus avoiding model training bias caused by differences in feature dimensions. Feature dimensionality reduction utilizes algorithms such as principal component analysis (PCA) or t-distributed stochastic neighbor embedding (t-SNE), for example, reducing the dimensionality of a 3D feature vector to 2D, thereby reducing computational complexity while preserving key information. The resulting multi-dimensional detection input tensor has a shape of (N, 2, 1). This tensor, used as model input data, effectively improves training efficiency and model generalization.

[0113] The inspection path correction label tensor is constructed based on the dynamic correction priority sequence. The dynamic correction priority sequence is a pre-set order of path segment corrections based on the defect risk level and inspection requirements. Each grid cell corresponds to a priority value (e.g., 3 for high-risk areas, 2 for medium-risk areas, and 1 for low-risk areas). The label tensor is constructed by one-hot encoding the priority value of each grid cell and converting it into a binary vector of length 3 (e.g., priority 3 corresponds to [1,0,0]). The label vectors of all grid cells are then concatenated in sequence into a three-dimensional tensor (with dimensions of N×3×1). This label tensor is used to indicate the inspection path correction direction that the model needs to learn during training. For example, grid cells corresponding to high-priority labels need to focus on path adjustment.

[0114] The final multi-dimensional detection input tensor and the detection path correction label tensor are combined to form a training sample set. Each sample consists of a grid cell feature vector in the input tensor and the corresponding label vector. The sample set is divided into a training set (70%), a validation set (20%), and a test set (10%). During the training process, stochastic gradient descent (SGD) or an adaptive optimization algorithm (such as Adam) is used to minimize the loss function (such as the cross-entropy loss function), and the model parameters are updated through the backpropagation algorithm. The validation set is used to monitor overfitting during model training. When the validation loss no longer decreases, the early stopping mechanism is triggered to avoid overtraining of the model. The test set is used to evaluate the generalization ability of the model and calculate indicators such as accuracy and recall, but does not involve the description of specific experimental effect data.

[0115] The detection path correction model is further derived from the trained and verified dynamic defect prediction model. The detection path correction model is a derivative model of the dynamic defect prediction model, and its output is the detection path correction parameters for each grid unit (such as the scanning angle adjustment amount and the path node offset). The real-time motion state parameters (such as the current position, speed, and acceleration of the robot end effector) are input into the detection path correction model. The model combines the characteristic attributes and priority labels of the current grid unit to predict the corresponding set of safe detection boundaries for the unit. For example, when the robot moves to a grid unit in a high-risk area, the model predicts that the safe detection boundary of the unit may be offset due to motion inertia based on the real-time speed parameters, and thus outputs the corresponding boundary adjustment value.

[0116] Generating initial detection parameters based on the predicted safety detection boundary set involves the following specific steps: First, extract the grid cell with the lowest defect coverage density in the safety detection boundary set and use its geometric center coordinates or feature point coordinates (such as the midpoint of a boundary segment) as the initial detection point coordinates. The area with the lowest defect coverage density is generally considered the "weak link" in the detection path, and prioritizing detection points there can increase the probability of discovering potential defects. For example, in a certain area of ​​the safety detection boundary set, by traversing the defect coverage density values ​​of all grid cells, find the cell with the minimum value, and use its center coordinates (x0, y0, z0) as the initial detection point.

[0117] Based on the spatial topology of the predicted set of safety detection boundaries, a feasible connectivity structure for the initial scanning path is fitted. This spatial topology is represented by the adjacency matrix of the grid cells, which records the connectivity between each cell and its neighbors (e.g., its neighbors in six directions: up, down, left, right, front, and back). The fitting process uses a shortest path algorithm from graph theory (such as Dijkstra or A*), starting from the initial detection point and targeting grid cells covering all high-risk areas. The shortest path connecting all detection points is searched, while satisfying the robot's kinematic constraints (e.g., maximum turning angle and minimum straight-line movement distance). During the fitting process, initial correction response parameters are generated, including the path adjustment step size (e.g., 5mm per step) and the angle increment (e.g., no more than 30° per turn). These parameters are used to control the magnitude of the robot's path correction during the scanning process.

[0118] The material stress gradient direction of the predicted safe inspection boundary set is calculated and normalized to a path reference vector. The material stress gradient direction is derived by calculating the gradient of the stress values ​​of each grid cell within the safe inspection boundary set. The direction of the gradient vector points in the direction of the fastest stress increase, and its modulus represents the rate of stress change. Normalization converts the gradient vector into a unit vector, which serves as a reference for the initial scanning path direction. For example, if the calculated stress gradient vector is (Δx, Δy, Δz), the path reference vector is (Δx / |Δr|, Δy / |Δr|, Δz / |Δr|), where |Δr| is the modulus of the gradient vector. The initial scanning path direction extends along this reference vector, ensuring that the inspection path preferentially covers areas of stress concentration and improving the detection efficiency of stress-induced defects.

[0119] Throughout the implementation process, data reconstruction transforms heterogeneous, multi-source path control simulation data into a standard format suitable for model training through tensor construction and processing. The construction and training of dynamic defect prediction models and inspection path correction models achieve intelligent mapping from material characteristics to inspection path corrections. Initial inspection parameter generation combines geometric analysis, topology optimization, and mechanical properties to ensure scientific and feasible parameters. These seamless integrations form an AI-based inspection path pre-planning and correction mechanism, providing critical initial conditions for subsequent dynamic correction simulations and multi-objective optimization.

[0120] Example 5:

[0121] This embodiment describes in detail the specific process of obtaining dynamic correction simulation information based on initial detection parameters and a comprehensive defect feature assessment model, and constructing a multi-objective optimization model to iteratively optimize detection path parameters. The following describes the construction of the dynamic correction simulation model, the iterative solution process, and the application of the multi-objective optimization algorithm.

[0122] To generate dynamic correction simulation information based on the initial detection parameters, the comprehensive defect feature evaluation model and the motion state parameters, it is necessary to first build a dynamic correction simulation model. The specific steps are: map the initial detection point coordinates to the three-dimensional space coordinate system of the comprehensive defect feature evaluation model, match the initial scanning path and the initial correction response parameters (such as step size, angle increment), and adjust the path node density according to the defect risk level of the correction area - increase the node density in high-risk areas (such as reducing the node spacing to 2mm), and reduce the node density in low-risk areas (such as increasing the node spacing to 10mm), thereby updating the path parameters of the comprehensive defect feature evaluation model. Subsequently, define the dynamic correction trigger conditions, path reconstruction rules and response parameter increments: the dynamic correction trigger conditions are set to the current path coverage completeness rate Not greater than the preset completion rate threshold (such as 90%); path reconstruction rules include based on the direction of material stress gradient Adjust the initial scanning path direction so that the new path direction is consistent with The angle does not exceed ; Response parameter increment and current precision stability evaluation index It is a piecewise linear relationship, that is, when hour ,when hour ,when hour (in is the preset increment coefficient, and are the high and low thresholds for accuracy stability).

[0123] Based on the dynamic correction simulation model, the multi-objective path planning equation is iteratively solved by the piecewise gradient descent method. The multi-objective path planning equation includes coverage integrity. , trajectory overlap , detection efficiency The objective functions are as follows (Formula 1 is given only taking the coverage integrity equation as an example):

[0124] (1)

[0125] in, represents the number of covered grid cells, Indicates the total number of grid cells in the target area. During the iterative solution process, the system monitors the path coverage completeness in real time. :when When the trigger response parameter is updated, the current accuracy stability evaluation index Adjust the detection point offset , path coverage completeness rate and precision stability evaluation indicators The calculated value of , and re-solve the multi-objective path planning equation until Or reach the preset number of iterations (e.g. 50 times) and terminate the simulation.

[0126] The steps of building a multi-objective optimization model and iteratively optimizing planning parameters are as follows: The variables of the optimization model include the path node distribution density , Scan angle threshold , sensor sampling frequency , the optimization goal is to maximize the path coverage efficiency Minimize detection energy consumption , the constraints include the upper limit of robot kinematics (such as the maximum joint speed , maximum acceleration ) and sensor accuracy thresholds (such as minimum resolution First, the multi-objective optimization model is preliminarily solved by the dynamic programming algorithm, and the detection area is divided into multiple stages. Each stage corresponds to a set of path node distribution density. and scan angle threshold , by recursively calculating the optimal solution of each stage, an initial optimized path set containing 10-20 paths is generated.

[0127] Based on the initial optimized path set, the ant colony optimization algorithm is used for global optimization. The ant colony optimization algorithm searches for the optimal solution in the initial optimized path set by simulating the pheromone transmission mechanism during the foraging process of ants. The specific process is: each ant represents a detection path, and the ant searches for the optimal solution according to the pheromone concentration during the path search. and heuristic information (such as the inverse of the path length) select the next node, pheromone concentration The quality of the path is updated as the ants pass through it - the pheromone concentration of the path with high coverage efficiency and low energy consumption increases, and vice versa. After iterations (e.g. 100), the optimal detection path parameters are converged to obtain the optimal path node distribution density. , Scan angle threshold , sensor sampling frequency , this set of parameters also satisfies the coverage efficiency (such as 95%) and energy consumption (such as 500J) constraints.

[0128] In the dynamic correction simulation and multi-objective optimization process, the motion state parameters (such as the real-time position of the robot end effector) ,speed ) is fed back to the system in real time to adjust the spatial coordinates of the path nodes and the simulation time step. For example, when the robot moves at a Exceeding the preset threshold When to ensure detection accuracy under high-speed motion; conversely, when the speed is low, the node density can be appropriately reduced to improve efficiency.

[0129] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0130] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A robot nondestructive testing method based on artificial intelligence, characterized in that: The following steps are involved: Obtain multimodal sensing data of the target to be detected and the motion state parameters of the robot end effector; Perform multi-dimensional scanning on the target and construct a three-dimensional material characteristic model. Combined with multi-modal sensing data, perform dynamic defect characteristic analysis on the three-dimensional material characteristic model to generate an initial defect characteristic assessment model. The initial defect feature evaluation model is loaded with material property distribution data and detection accuracy constraints to generate a comprehensive defect feature evaluation model. The detection path dynamic simulation is performed in combination with the motion state parameters to obtain path control simulation data. A dynamic defect prediction model is constructed based on a three-dimensional material feature model, and is trained using path control simulation data to obtain a detection path correction model, thereby generating initial detection parameters. The initial detection parameters include at least initial detection point coordinates, initial scanning path, and initial correction response parameters. Based on the initial detection parameters, the comprehensive defect feature evaluation model and the motion state parameters, dynamic correction simulation information is obtained; Combining the dynamic correction simulation information with the inspection task planning parameters, a multi-objective optimization model is constructed and the planning parameters are iteratively optimized to obtain the optimal inspection path parameters; Based on the real-time material feature model, the detection path parameters of the current period are inferred to generate the safe detection probability of the current period. Combined with the optimal detection path parameters of the current period and the actual scanning trajectory, the robot's real-time detection strategy is adjusted to achieve the dynamic path matching goal.

2. The artificial intelligence-based robot nondestructive testing method according to claim 1, characterized in that: The path control simulation data includes at least material stress distribution, defect clustering area, safety detection boundary set and dynamic correction priority sequence; The dynamic correction simulation information at least includes detection point offset, path coverage completeness rate and accuracy stability evaluation index.

3. The artificial intelligence-based robot nondestructive testing method according to claim 1, characterized in that: The method of performing multi-dimensional scanning on the target and constructing a three-dimensional material characteristic model, performing dynamic defect characteristic analysis on the three-dimensional material characteristic model in combination with multi-modal sensing data, and generating an initial defect characteristic evaluation model includes the following steps: Perform a composite scan of radiation and sound waves on the target to obtain a sensor data sequence of the internal defect distribution of the material; Preprocessing the sensor data sequence to obtain standardized feature fusion data, wherein the preprocessing includes one or more of signal denoising, time domain registration, feature fusion, grid division, and data interpolation; Generate a 3D material feature model based on standardized feature fusion data and combined with 3D point cloud reconstruction technology; A dynamic defect feature analysis is performed on the three-dimensional material feature model to generate an initial defect feature assessment model, wherein the dynamic defect feature analysis includes at least grid density allocation and feature weight mapping, the defect risk level is divided by grid density allocation, and corresponding sensor data parameters are assigned to each level area.

4. The artificial intelligence-based robot nondestructive testing method according to claim 1, characterized in that: The method of loading the initial defect feature assessment model with material property distribution data and detection accuracy constraints to generate a comprehensive defect feature assessment model, performing dynamic simulation of the detection path in combination with motion state parameters, and obtaining path control simulation data includes the following steps: Loading material property distribution data into the initial defect feature assessment model to simulate real-time property changes to generate a first defect feature assessment model; Loading the first defect feature assessment model with detection accuracy constraints to generate a comprehensive defect feature assessment model, wherein the detection accuracy constraints include a maximum detection angle limit, a minimum resolution threshold, and a sensor response redundancy parameter; Based on the comprehensive defect feature evaluation model and motion state parameters, dynamic simulation of the detection path is performed. The specific process includes: Construct a multi-objective path planning equation, which includes at least a coverage integrity equation, a trajectory coincidence equation, and a detection efficiency equation. Combined with a comprehensive defect feature evaluation model, the multi-objective path planning equation is numerically solved using a piecewise gradient descent method to obtain material stress distribution, defect clustering areas, a safe detection boundary set, and a dynamic correction priority sequence. The generation of the security detection boundary set includes the following steps: Based on the defect clustering area, the defect coverage density of each grid cell is calculated; Identify areas in the comprehensive defect feature assessment model where the defect coverage density is no greater than a preset threshold and generate a set of safe detection boundaries.

5. The artificial intelligence-based robot nondestructive testing method according to claim 1, characterized in that: The method of constructing a dynamic defect prediction model based on the three-dimensional material feature model and training it with path control simulation data to obtain a detection path correction model and then generate initial detection parameters includes the following steps: The dynamic defect prediction model is trained and verified through path control simulation data to obtain the detection path correction model; Input the real-time motion state parameters into the detection path correction model to predict the safe detection boundary set; Based on the predicted safe detection boundary set, initial detection parameters are generated, wherein the initial detection parameters at least include initial detection point coordinates, an initial scanning path, and initial correction response parameters.

6. The artificial intelligence-based robot nondestructive testing method according to claim 5, characterized in that: The dynamic defect prediction model is constructed and trained with the path control simulation data to obtain a detection path correction model, thereby generating initial detection parameters. The method also includes data reconstruction of the path control simulation data, specifically: Construct an initial multi-dimensional detection input tensor based on material stress distribution, defect clustering areas, and safe detection boundary sets; Normalize and reduce the feature dimensionality of the initial multi-dimensional detection input tensor to generate the final multi-dimensional detection input tensor; Based on the dynamic correction priority sequence, the detection path correction label tensor is constructed; The final multi-dimensional detection input tensor and the detection path correction label tensor are combined to form a training sample set.

7. The artificial intelligence-based robot nondestructive testing method according to claim 6, characterized in that: The method of generating initial detection parameters based on the predicted security detection boundary set includes the following steps: Extract the grid cell with the lowest defect coverage density in the predicted safety detection boundary set and generate the coordinates of the initial detection point; According to the spatial topological relationship of the predicted safety detection boundary set, the feasible connection structure of the initial scanning path is fitted to generate the initial correction response parameters; The material stress gradient direction of the predicted safety detection boundary set is calculated and normalized into a path reference vector, which is the initial scanning path direction.

8. The artificial intelligence-based robot nondestructive testing method according to claim 1, characterized in that: The method of obtaining dynamic correction simulation information based on the initial detection parameters, the comprehensive defect feature evaluation model and the motion state parameters includes the following steps: Map the initial inspection point coordinates to the comprehensive defect feature evaluation model, match the initial scanning path with the initial correction response parameters, adjust the path node density in the correction area, update the comprehensive defect feature evaluation model, define the dynamic correction trigger conditions, path reconstruction rules and response parameter increments, and generate a dynamic correction simulation model; Based on the dynamic correction simulation model, the multi-objective path planning equation is iteratively solved by the piecewise gradient descent method to obtain dynamic correction simulation information, including: When the dynamic correction trigger conditions are met, the detection point offset, path coverage completeness, and accuracy stability evaluation indicators are updated, and the multi-objective path planning equation is re-solved until the simulation termination conditions are met; The dynamic correction trigger condition includes triggering the response parameter update when the current path coverage completeness rate is not greater than the preset completeness rate threshold; the path reconstruction rule includes adjusting the initial scanning path based on the material stress gradient direction; the response parameter increment is in a piecewise linear relationship with the current precision stability evaluation index.

9. The artificial intelligence-based robot nondestructive testing method according to claim 1, characterized in that: The multi-objective optimization model is constructed and the planning parameters are iteratively optimized to obtain the optimal detection path parameters, including the following steps: A multi-objective optimization model was constructed, where the optimization variables included path node distribution density, scanning angle threshold, and sensor sampling frequency. The optimization objectives included maximizing path coverage efficiency and minimizing detection energy consumption. The constraints included the robot kinematic upper limit and sensor accuracy threshold. Perform a preliminary solution to the multi-objective optimization model through a dynamic programming algorithm to generate an initial set of optimization paths; Based on the initial optimized path set, the ant colony optimization algorithm is used for global optimization to generate the optimal detection path parameters.

10. A robot nondestructive testing system based on artificial intelligence, characterized in that: include: The data acquisition module is used to obtain the multimodal sensing data of the target to be detected and the motion state parameters of the robot end effector; The feature modeling and analysis module is used to perform multi-dimensional scanning of the target and build a three-dimensional material feature model. It combines multi-modal sensing data to perform dynamic defect feature analysis on the three-dimensional material feature model and generate an initial defect feature assessment model. The comprehensive evaluation and simulation module is used to load material property distribution data and detection accuracy constraints into the initial defect feature evaluation model, generate a comprehensive defect feature evaluation model, and perform dynamic simulation of the detection path in combination with motion state parameters to obtain path control simulation data; A parameter generation module is used to build a dynamic defect prediction model based on the three-dimensional material feature model, and to obtain a detection path correction model through training using path control simulation data, thereby generating initial detection parameters, wherein the initial detection parameters include at least the initial detection point coordinates, the initial scanning path, and the initial correction response parameters; Dynamic correction simulation module, used to obtain dynamic correction simulation information based on initial detection parameters, comprehensive defect feature evaluation model and motion state parameters; The path optimization module is used to combine the dynamic correction simulation information with the inspection task planning parameters, build a multi-objective optimization model and iteratively optimize the planning parameters to obtain the optimal inspection path parameters; The strategy adjustment module is used to infer the detection path parameters of the current period based on the real-time material feature model, generate the safe detection probability of the current period, combine the optimal detection path parameters of the current period with the actual scanning trajectory, and adjust the robot's real-time detection strategy to achieve the dynamic path matching goal.

Images