A marine welding data intelligent AI evaluation optimization method based on ray detection

Abstract

The application discloses a kind of marine welding data intelligent AI evaluation piece optimization method based on ray detection.The method first acquires the original data of weld by ray probe, and carries out normalized storage and pre-processing, forms defect feature data;Further in combination with standard test block reference database, the standardization storage and multi-modal feature coding, fusion processing of defect characteristics are carried out, and unified defect feature representation and concrete defect parameters are generated;Subsequently, the defect parameters are classified by confidence evaluation, and the corresponding processing procedure is triggered according to different grades;Finally, evaluation piece report is automatically generated.The application realizes the full-process automation from data acquisition to report output, and significantly improves the accuracy and efficiency of weld defect identification, quantification and rating by deeply integrating the physical characteristics of multi-modal detection data and the abstract semantic features extracted by deep learning model.

Description

Technical Field

[0001] This invention relates to a method for inspecting welds in marine engineering, specifically to a method for obtaining concrete defect parameters from marine weld data based on radiographic inspection. This method uses a three-layer feature processing architecture and a teacher-student model to obtain concrete defect parameters, and uses a four-level quantitative evaluation and a standard test block benchmark comparison mechanism to verify the data and feedback mechanism to achieve automated evaluation. Background Technology

[0002] With the rapid development of marine resource development, non-destructive testing of welds in marine engineering using X-rays has become an indispensable technical means. However, traditional X-ray data processing and interpretation have the following significant drawbacks: they rely heavily on human experience, resulting in low efficiency and strong subjectivity; the processing flow is fragmented, lacking a unified pipeline, leading to low information transmission efficiency and the easy accumulation and amplification of errors; there is a lack of quantitative evaluation standards, making it difficult to guarantee the reliability and consistency of the evaluation results; and when dealing with new types of data or encountering complex scenarios, they cannot learn and optimize themselves through feedback mechanisms, resulting in poor adaptability.

[0003] While existing technologies utilize AI for image recognition, they are mostly limited to single recognition models and fail to construct a complete system from a systems engineering perspective, encompassing data quality control, feature engineering, model training, result verification, and feedback optimization. Therefore, there is an urgent need for an automated, intelligent, and continuously iteratively optimized solution for weld evaluation in marine non-destructive testing. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent AI evaluation method for marine welding data based on radiographic testing. This method enables quantitative evaluation and has automatic optimization capabilities.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The present invention provides an intelligent AI-based image evaluation method for marine welding data based on radiographic testing, comprising the following steps: Step 1: The X-ray probe moves along a preset weld seam scanning trajectory. At each acquisition point during the movement, a data acquisition operation is triggered to obtain weld seam inspection data at the corresponding location. The weld seam inspection data is then sequentially standardized and stored, preprocessed, and feature extracted to form NDT feature data for weld seam inspection. The NDT feature data includes at least preprocessed defect feature data. The acquisition of weld seam inspection data includes: the acquisition of raw data collected during the X-ray inspection process. Step 2: Establish a standard test block benchmark database based on standard test blocks containing known defects, including "defect type - actual size - defect characteristic parameters"; at the same time, perform initial screening and standardization on the preprocessed defect characteristic data in Step 1 to obtain standardized defect characteristic data, and store the standardized defect characteristic data in the standardized defect characteristic data layer of the effective characteristic database. Step 3: Perform feature processing and parameter mapping on the standardized defect feature data layer obtained and stored in the effective feature database in Step 2. The feature processing includes at least multimodal feature encoding, abstract feature extraction, and feature fusion. Based on this, generate defect feature representations and visualized defect parameters for subsequent intelligent image evaluation. The defect feature representation includes at least a unified structured feature vector, an abstract semantic feature vector, and a fused feature vector. The visualized defect parameters are generated based on the fused feature vector through parameter mapping. The unified structured feature vector, abstract semantic feature vector, fused feature vector, and visualized defect parameters are all stored back into the corresponding fields of the effective feature database. Step 4: Analyze the confidence level of the visualized defect parameters generated and stored in the visualized defect feature database in Step 3. Based on the multi-dimensional evaluation results, classify the confidence level of each visualized defect parameter and trigger the corresponding subsequent processing flow according to different confidence levels to improve the reliability of weld inspection results and optimize the allocation of quality inspection resources. Step 5: Based on the confidence level results of the visualized defect parameters obtained in Step 4, generate and output the film review report.

[0006] The beneficial effects of this invention are: it achieves end-to-end automated processing from raw data to the final report, greatly improving efficiency and reducing reliance on manual labor. Employing a multimodal data encoding and deep feature fusion strategy, the X-ray inspection data is transformed into structured features through dedicated encoding modules. Abstract semantic features are extracted using a teacher-student model architecture, and then a learnable weight matrix is ​​used to efficiently fuse physical and semantic features, forming composite features that combine concrete attributes and abstract associations. This significantly improves the accuracy of weld defect type identification, size measurement, and grade determination. Detailed Implementation

[0007] The present invention will now be described in detail with reference to specific embodiments.

[0008] The core concept of this invention lies in constructing the film evaluation process as an automated pipeline comprising five core steps: data collection, screening, feature processing, confidence assessment, and report generation. Strict evaluation criteria are established at four key nodes: data screening, feature fusion, parameter mapping, and result evaluation. Through quantitative rules such as comparison with a standard test block benchmark database, verification of physical and statistical consistency of features, and multi-dimensional confidence scoring, the reliability of data and results at each stage is ensured. Combined with multimodal feature fusion and a teacher-student model architecture, the accuracy of defect identification is improved, thus achieving automation and intelligence in the film evaluation process.

[0009] This invention discloses an intelligent AI-based image evaluation method for marine welding data based on radiographic testing, comprising the following steps: Step 1: The X-ray probe moves along a preset weld seam scanning trajectory. At each acquisition point during the movement, a data acquisition operation is triggered to obtain weld seam inspection data at the corresponding location. This weld seam inspection data is then sequentially standardized, preprocessed, and feature extracted to form NDT feature data for weld seam inspection. The NDT feature data includes at least the preprocessed defect feature data. Weld seam inspection data acquisition includes: acquiring the raw data collected during the X-ray inspection process; and acquiring multiple data types at the same acquisition location.

[0010] The specific steps are as follows: Step 101: Place the X-ray probe against the surface of the weld to be sampled, keeping the probe direction parallel to the center line of the weld. By emitting and receiving X-rays, non-destructive testing data is collected from the weld area to obtain the raw scan data.

[0011] Step 102: Standardize and store the original scan data to obtain standardized and stored original scan data.

[0012] Step 103: Preprocess the standardized and stored raw scan data to form array data, image data, and defect feature data extracted based on the array data and image data for subsequent analysis; the preprocessing may include removing invalid data caused by noise interference, equipment malfunction, and acquisition errors, and screening out valid samples that meet the accuracy requirements to reduce the processing cost of subsequent algorithms.

[0013] Step 104: Establish a raw database for weld inspection data management, and uniformly store the array data, image data and corresponding defect feature data obtained after data preprocessing in Step 103.

[0014] The original database adopts a hybrid storage method of relational database and binary data, which supports the storage of structured data such as defect annotation information and equipment parameters, as well as binary storage of image data and array data, to meet the calling needs of subsequent feature analysis, model training and intelligent discrimination steps.

[0015] In one embodiment, the relational database is named the preprocessed defect feature database, and the database creation, writing, and reading operations are implemented through the database driver of the data processing software.

[0016] Step 2: Establish a standard test block benchmark database based on standard test blocks containing known defects, including "defect type - actual size - defect feature parameters". At the same time, perform initial screening and standardization on the preprocessed defect feature data in Step 1 to obtain standardized defect feature data. Store the standardized defect feature data in the standardized defect feature data layer of the effective feature database for subsequent feature analysis, model training and intelligent discrimination steps.

[0017] The initial screening may include operations that meet the requirements of data completeness, data clarity, physical consistency, and data accuracy, specifically by performing the following steps: Step 201: The preprocessed defect feature data is collected and screened at the source to exclude invalid data caused by abnormal environmental conditions, resulting in primary screened defect feature data. Specifically, to remove data with excessive environmental interference, environmental threshold parameters are first set according to the detection environment. These environmental threshold parameters include detection speed, environmental noise level, and real-time temperature deviation. Then, the preprocessed raw data is correlated with the environmental parameters. When the environmental parameters corresponding to the detection data exceed the preset threshold range, the detection data is determined to be invalid and removed, thereby obtaining valid data that meets the environmental requirements. The corresponding defect feature data is output as "primary screened defect feature data". The environmental parameters can be acquired simultaneously or correlated during weld scanning using a X-ray probe.

[0018] In one embodiment, the array data is batch-checked by a data processing software script to determine whether there is a discrepancy with a set threshold. If so, the data with the discrepancy is removed.

[0019] Step 202: Perform signal quality screening on the "first-stage screening defect feature data" to remove low-quality data that is contaminated by noise or has incomplete signals, and obtain "second-stage screening defect feature data" that has clear, distinguishable and complete signals.

[0020] Specifically: First, set signal quality screening indicators, then set the threshold range for the signal quality screening indicators. The signal quality screening indicators include at least signal-to-noise ratio (SNR) and signal integrity. Then, input the "first-stage screening defect feature data" into the signal quality calculation module (e.g., a program module implemented based on a data processing software language). This module performs the following operations on each piece of input data: calculate the actual SNR and actual signal integrity ratio of the input data; compare them with the threshold range of the signal quality screening indicators, retain the data that meets the screening conditions, output the "second-stage screening defect feature data", and remove the data that does not meet the conditions.

[0021] When the signal-to-noise ratio is greater than a preset threshold (e.g., 10 dB), the defect echo signal is considered to be effectively distinguishable from the noise signal; when the proportion of the effective signal segment to the total signal duration is greater than a preset threshold (e.g., 80%), the signal is considered to have integrity.

[0022] Step 203: Perform feature consistency screening on the "secondary screening defect feature data" to remove abnormal feature data and ensure that the feature results conform to the physical characteristics and statistical laws of weld defects in the "third screening defect feature data".

[0023] Step 2031: Perform feature physical rationality screening, retain feature data within the physical rationality range, and remove abnormal feature data that exceeds the range.

[0024] The following example illustrates the process: Input the defect feature data obtained from the "secondary screening defect feature data" in step 202 into the data processing software to obtain the JSON data of the defect feature data. Then, extract the parameters through the JSON fields of the defect feature data and determine whether the parameters are within the set reasonable range. Data within the set range is retained, and otherwise discarded, thus obtaining defect feature data that meets the physical rationality requirement.

[0025] Taking the physical rationality screening of time-domain features as an example, the peak value of the defect echo is selected as the feature parameter, and a reasonable range is set according to the detection environment and equipment parameters. For example, when the peak value of the defect echo is less than 0.1 V, it is determined to be a noise signal; when the peak value of the defect echo is greater than 5 V, it is determined to be an equipment overload signal; when the peak value of the defect echo is in the range of 0.1 V to 5 V, the feature parameter is determined to be physically rational.

[0026] Step 2032: Within the same inspection area of ​​the same weld, the defect feature data that meets the physical rationality criteria after the physical rationality screening in step 2031 are subjected to feature statistical consistency screening to obtain “three-stage screening defect feature data” with consistent statistical distribution.

[0027] Specifically, for the same inspection area of ​​the same weld, the mean μ and standard deviation σ of the characteristic parameters are calculated based on the pre-processed defect characteristic data that meet the physical rationality requirements. For the multiple sets of defect feature data retained after screening in step 2031, when a certain set of defect feature values ​​exceeds the range of [μ-3σ,μ+3σ], it is determined to be a statistical outlier and is removed, thereby obtaining "three-stage screening defect feature data" with consistent statistical distribution.

[0028] Step 204: Establish a standard test block baseline database. The "three-stage screening defect feature data" obtained after screening in step 203 is compared and screened with the standard test block baseline feature parameters containing the corresponding defect types and sizes to verify whether the detection accuracy of the defect feature data meets the requirements. When the deviation exceeds the allowable range, it is judged as unqualified data and discarded, outputting "four-stage screening defect feature data". A valid feature database is created, storing the "four-stage screening defect feature data" in it, thus obtaining a valid feature database for subsequent analysis.

[0029] Specifically, firstly, based on the test data collected from standard test blocks containing known defect types and sizes, the corresponding defect feature parameters are extracted to establish a standard test block benchmark database of "defect type - actual size - defect feature parameters," and allowable deviation ranges are set for each benchmark feature parameter. Then, the actual collected data of the "three-stage screening defect feature data" processed in step 203 is compared with the corresponding parameters in the standard test block benchmark database in the accuracy deviation verification module of the data processing software. In the accuracy deviation verification module of the data processing software, the "three-stage screening defect feature data" is judged to meet accuracy standards. When the feature parameter deviation is less than the allowable deviation range, the data is judged to meet accuracy standards; otherwise, it is judged to be substandard data and discarded, obtaining "four-stage screening defect feature data," which is then stored in the standardized defect feature data layer of the effective feature database.

[0030] The following example illustrates how to establish a standard test block baseline database containing "defect type - actual size - defect characteristic parameters" and link it to the accuracy deviation verification module of the data processing software: First, a branch link table is added to the original database to store the structured information of the baseline data. This table is then exported as a JSON file and stored in the data processing software's project directory for easy retrieval. Finally, the branch link table is queried using the data processing software's database connection library to obtain the baseline data that corresponds one-to-one with the data in the "three-stage screening defect characteristic data" database. Deviation verification is then performed in the data processing software's accuracy deviation verification module to filter out data that passes the deviation verification and discard unqualified data. The branch link table contains the following:

[0031] Step 3: Perform feature processing and parameter mapping on the standardized defect feature data layer obtained and stored in the effective feature database in Step 2. The feature processing includes at least multimodal feature encoding, abstract feature extraction, and feature fusion. Based on this, generate defect feature representations and visualized defect parameters for subsequent intelligent image evaluation. The defect feature representation includes at least a unified structured feature vector, an abstract semantic feature vector, and a fused feature vector; the visualized defect parameters are generated based on the fused feature vector through parameter mapping. All of the aforementioned unified structured feature vector, abstract semantic feature vector, fused feature vector, and visualized defect parameters are stored back into the corresponding fields of the effective feature database.

[0032] Step 301: Multimodal feature encoding and unified structured feature vector generation: Extract the defect feature data corresponding to the same detection sample from the standardized defect feature data layer, encode the feature data, convert it into a feature vector of the same dimension, and then generate a unified structured feature vector of the detection sample and store it in the corresponding field of the effective feature database.

[0033] The same test sample refers to a set of test data associated with the same spatial coordinates or the same acquisition time in weld inspection.

[0034] Step 302: Using the unified mapping of 256-dimensional structured features generated in step 301 and stored in the effective feature database as input, process the data through the knowledge distillation framework to generate an abstract semantic feature vector for defect classification and assessment, and store the abstract semantic feature vector back into the effective feature database.

[0035] Specifically, the knowledge distillation framework includes a teacher model and a student model. During the training phase, a unified mapping to 256-dimensional structured features is used as the common input to both the teacher and student models, and training is performed in conjunction with supervised information corresponding to the unified mapping to 256-dimensional structured features. This supervised information includes: Defect type labels provided by the standard test block reference database; the defect type labels refer to common defect categories: porosity, slag inclusion, incomplete penetration, lack of fusion, and cracks; Defect physical parameter labels obtained from the standard test block benchmark database. These labels refer to defect geometric parameters: predicted depth, length, width, and area of ​​the defect, and the actual peak amplitude of the defect echo.

[0036] During training, the teacher model processes the input vector and outputs a teacher feature representation; the student model processes the same input vector and outputs a student feature representation; the student model is optimized using a joint loss function, which includes at least distillation loss, to constrain the difference between the output features of the student model and the output features of the teacher model.

[0037] After training, the parameters of the student model are fixed, and the student model is used to uniformly map all samples in the effective feature database to 256-dimensional structured features for processing, generating corresponding abstract semantic feature vectors, and storing them in the effective feature database.

[0038] Among these, qualified data is used, and unqualified data is removed, and the above-mentioned various types of data are obtained respectively. From the unified mapping of 256-dimensional structured features stored in the effective feature database, abstract features are extracted and their effectiveness is verified. This requires feature distillation, which can be completed within the teacher-student model.

[0039] The teacher-student model, on the other hand, uses a proven and mature combination of teacher-student models.

[0040] Step 3021: Optimize the large model. The teacher model, as the large model, has the core function of extracting highly discriminative and strongly physically correlated deep abstract features from the original ray data. Two optimizations are needed: one for the weld defect abstract feature extraction function and the other for the 256-dimensional structured feature input (this can be achieved by directly reusing and modifying the official Microsoft open-source code: https: / / github.com / microsoft / Swin-Transformer). Specifically, this includes: The model structure was redesigned to be lightweight and adapted to ray detection data, reducing computational load. First, the original model's classification head fully connected layer was discarded, retaining only the Backbone + Neck feature extraction part. By embedding physical constraints on weld rays, a weld-specific physical loss term is added during the teacher model training phase. This ensures that the extracted features are strongly correlated with the core physical parameters of weld defects. The loss function is defined as follows:

[0041] in, To account for distillation losses, which will need to be integrated with the student model later, we will temporarily represent them using mean squared error (MSE).

[0042] in These are the true values. Enter the elements of the reference feature values ​​for the corresponding dimensions from the 256-dimensional structured feature vector obtained in the preceding steps. It is the model prediction value of the output feature of the teacher model in the corresponding dimension, n is the number of samples, and the entire MSE loss is calculated by averaging the element-wise squared errors between the two high-dimensional vectors.

[0043] This is the physical constraint loss of the weld, representing the error between the features extracted by the computational model and the actual parameters of the weld defect (defect depth, length, area, echo), expressed as the mean absolute error (MAE) loss.

[0044] in, To obtain the true physical parameters of the defect, 256-dimensional fused structured features are read from the effective feature database established in the preceding steps, and the corresponding true physical label of the defect is also read. To predict the physical parameters from the 256-dimensional primary fusion structured features extracted from the model, a small regression head (1-2 fully connected layers) is added after the 256-dimensional features, allowing the model to predict these physical parameters from the 256-dimensional primary fusion structured features.

[0045] These are weighting coefficients used to balance distillation and physical constraints. They need to be calculated and optimized during training. The method involves observing the Pearson correlation coefficient of defect depth and the cosine similarity of feature distillation after several training rounds, and selecting the optimal coefficient based on these two parameters. Values.

[0046] In this way, the teacher model can output deep 256-dimensional abstract features, including spatial features, physical features, and defect semantic features of the ray detection data. An abstract feature database is then established to store these deep 256-dimensional abstract features.

[0047] The core function of the student model is to distill high-quality features from the teacher model, outputting lightweight and easily deployable abstract features. To interface with the deep 256-dimensional abstract feature system, the following optimizations are required: S11. Discard the original model's classification head and retain the lightweight convolution + transformation block.

[0048] S12. Finally, add a fully connected layer to learn the linear relationship between the input and output, and connect a 256-dimensional feature fusion module (concatenation + batch normalization) to ensure that the final output is a lightweight 256-dimensional abstract feature.

[0049] Step 3022: Batch feature extraction. Using the trained student model and fusion module, extract all 256-dimensional fused structured features of all samples at once. Train the teacher model and student model repeatedly and continuously optimize them to obtain the best combined model. Finally, output the lightweight 256-dimensional abstract features and store them in the abstract feature database.

[0050] Step 303: Adaptively weightedly fuse the unified mapping of 256-dimensional structured features obtained in step 301 with the lightweight 256-dimensional abstract features obtained in step 302 to generate a weld defect fusion feature vector for final intelligent discrimination, and store it in the effective feature database.

[0051] Specifically, a unified mapping for the same sample is read from the effective feature database as a 256-dimensional structured feature (F_structured) and a lightweight 256-dimensional abstract feature F_abstract. These two are then fused using a learnable fusion weight matrix W to calculate a "secondary fusion feature vector" F_fused. This final "secondary fusion feature vector" F_fused is denoted as the "weld defect fusion feature vector" and stored in the effective feature database.

[0052] In the formula, I is the identity matrix, used to ensure that the feature dimension remains N after fusion; BatchNorm represents batch normalization operation, used to eliminate differences in feature dimensions and stabilize the feature distribution after fusion. The fusion weight matrix W is learned during model training through the backpropagation algorithm (see Nielsen, M. (2020). In-Depth Neural Networks and Deep Learning (translated by Zhu Xiaohu). People's Posts and Telecommunications Press. pp. 41-56) to achieve adaptive allocation of the contribution of different feature dimensions.

[0053] The specific implementation method is as follows: S21. Extract unified 256-dimensional structured features (denoted as ) from the effective feature database obtained in step 301. ); S22. Extract the lightweight 256-dimensional abstract features of the sample from the student model output (denoted as...). ); S23. Construct a learnable fusion weight matrix W (256×256), and learn the optimal fusion ratio between the two through training to obtain the final "secondary fusion feature vector". :

[0054] Where I is the identity matrix, ensuring the dimensionality remains 256 after fusion; BatchNorm represents batch normalization, eliminating differences in feature dimensions and ensuring stable feature distribution after fusion. The specific construction and training steps of the learnable fusion weight matrix W (256×256) are as follows: S31. First, create a 256×256 matrix, and set all initial values ​​to 0.5, which means that the original 256-dimensional structured features and 256-dimensional abstract semantic features each account for 50% of the weight; S32. Apply Sigmoid activation to W, ensuring that each element ∈(0,1); S33. Add W to the student model training process, and use task loss (such as defect classification loss, physical constraint loss) to backpropagate and update the value of W. During training, the model will automatically learn the optimal ratio between the 256-dimensional structured features uniformly mapped to each dimension and the lightweight 256-dimensional abstract features output by the student model, so that the fused... It performs best in defect detection tasks.

[0055] Step 304: Input the "secondary fusion feature vector" generated in step 303 into the parameter mapping network, output the concrete defect parameters required for the defect detection report, and store the concrete defect parameters in the corresponding fields of the effective feature database. Specifically, the parameter mapping network includes at least one fully connected layer for mapping the "secondary fusion feature vector" to one or more of the following defect description parameters: Defect geometry parameters: including the predicted depth, length, width and area of ​​the defect; Defect type probability: Used to characterize the probability that a defect belongs to a predefined defect category; Defect hazard level rating: Used to comprehensively characterize the severity of a defect.

[0056] Based on the existing 256-dimensional primary fusion structured features in the effective feature database, the abstract features output by the teacher model + student model, and the "secondary fusion feature vector" obtained in step 303, no additional training is required. Through simple feature mapping, the concrete defect features commonly used in the weld detection business can be output, and a concrete defect feature database is created to store the concrete defect features therein. For specific details of the feature mapping, please refer to: Shi, C., Yan, S., Wang, L., Zhu, C., Yu, Y., Zang, X., Liu, A., Zhang, C., & Feng, X. (2025). A Welding Defect Detection Model Based on Hybrid-Enhanced Multi-Granularity Spatiotemporal Representation Learning. Sensors, 25(15), 4656. https: / / doi.org / 10.3390 / s25154656 to obtain the concrete defect features of depth, length, width, area, defect type, probability distribution, and hazard score.

[0057] Step Four: Conduct confidence evaluation on the concrete defect parameters generated and stored in the concrete defect feature database in Step Three, classify the confidence levels of each concrete defect parameter based on the multi-dimensional evaluation results, and trigger corresponding subsequent processing flows according to different confidence levels to improve the reliability of the weld detection results and optimize the quality inspection resource allocation. The specific steps are as follows: Step 401: Prediction stability evaluation of concrete defect parameters: For the same detection sample, while keeping the data in the effective feature database that has been screened in Step Two and is from the same detection sample (i.e., "quadruple-screened defect feature data") unchanged, by introducing different random initialization parameters, different batch inference orders, or different model inference rounds in the feature processing and parameter mapping process described in Step Three, perform K independent predictions (K ≥ 2, for example, 5 times) on the same concrete defect parameter to obtain K predicted values of this concrete defect parameter.

[0058] Calculate the coefficient of variation CV of the K predicted values. CV is equal to the ratio of the standard deviation of the predicted values to their mean.

[0059] Determine the prediction stability score S1 according to the CV value: When CV ≤ 0.05, S1 = 30; When 0.05 < CV ≤ 0.1, S1 = 20; When 0.1 <cv ≤ 0.2时，s1="10；" 当cv>When the value is 0.2, S1 = 0.

[0060] The above thresholds can be adjusted according to the accuracy of the testing equipment, the type of material, or the specific application scenario.

[0061] Step 402: Correlation assessment between visualized defect parameters and feature representations: Select M test samples (M≥10, for example, 100 samples), and for the concrete defect parameter to be evaluated, calculate the Pearson correlation coefficient between the concrete defect parameter and each feature component in the unified structured feature vector or fused feature vector corresponding to each sample, and take the correlation coefficient with the largest absolute value, denoted as r_max.

[0062] The correlation score S2 is determined based on the value of r_max. When |r_max| ≥ 0.85, S2 = 40; When 0.7 ≤ |r_max| < 0.85, S2 = 30; When 0.5 ≤ |r_max| < 0.7, S2 = 20; When |r_max|<0.5, S2 = 0.

[0063] Step 403, Physical rationality assessment of concrete defect parameters: For different types of concrete defect parameters, at least one physical rationality verification rule is predefined. For example: For the defect depth parameter, it is stipulated that its predicted value shall not be greater than the nominal thickness of the base material of the weld being inspected; For the defect echo amplitude parameter, it is stipulated that its predicted value shall not be lower than the system baseline.

[0064] Based on the conformity of the concrete defect parameters to the predefined physical rules, determine the physical rationality score S3: When all physical rules are met, S3 = 30; When only some physical rules are met, S3 = 15; S3 = 0 when it does not conform to any physical rules.

[0065] Step 404: Confidence level comprehensive calculation, classification and automatic processing: This step integrates the evaluation results from steps 401 to 403, and performs confidence level calculation and grading for each concrete defect parameter.

[0066] For each concrete defect parameter, calculate its final confidence score S, S=S1+S2+S3, and store the confidence score as an auxiliary attribute of the concrete defect parameter in the effective feature database.

[0067] Based on the final confidence score S, the concrete defect parameters are divided into the following confidence levels: Grade A: 90 ≤ S ≤ 100; Grade B: 70 ≤ S < 90; Grade C: 50 ≤ S < 70; Grade D: S < 50.

[0068] The system automatically executes the corresponding processing strategy based on the confidence level of the concrete defect parameters: For Class A parameters, the results are automatically adopted and a test report is generated. For Class B parameters, they are marked as low review priority in the test report; For Class C parameters, mark them as high priority for review in the test report and prompt for manual review; For Class D parameters, the test result is deemed invalid, and a retest process is triggered, returning to step one to the corresponding weld position for data acquisition and processing again.

[0069] Step 5: Based on the confidence level results of the visualized defect parameters obtained in Step 4, generate and output the film review report: For the concrete defect parameters and corresponding test results that are determined to be of confidence level A in step four, the system automatically generates a standardized weld inspection evaluation report. The evaluation report is output as the final test result and does not require manual review.

[0070] The film review report shall include at least the following: Textual information section: includes the predicted depth, length, width and area of ​​the defect, the probability distribution of the defect type, and the comprehensive score of the severity of the defect; Image information section: includes X-ray detection data corresponding to the detection results, used to visually display the response characteristics of the defects.

[0071] To ensure the traceability and integrity of the test results, preferably, the evaluation report includes the model version number, detection and processing timestamps, and digital signature information simultaneously when it is generated, wherein: The model version number is used to identify the feature processing and discrimination model used to generate the review result; The timestamp is used to record the time when the review report was generated; The digital signature is used to prevent the film review report from being tampered with.

[0072] More preferably, the film review report is output in at least two formats, including PDF format for manual review and archiving, and JSON format for system integration, data transmission, or subsequent automated analysis.< / cv>

Claims

1. A method for intelligent AI-based image evaluation and optimization of marine welding data based on radiographic testing, characterized in that, Includes the following steps: Step 1: The X-ray probe moves along a preset weld seam scanning trajectory. At each acquisition point during the movement, a data acquisition operation is triggered to obtain weld seam inspection data at the corresponding location. The weld seam inspection data is then sequentially standardized and stored, preprocessed, and feature extracted to form NDT feature data for weld seam inspection. The NDT feature data includes at least preprocessed defect feature data. The acquisition of weld seam inspection data includes: the acquisition of raw data collected during the X-ray inspection process. Step 2: Establish a standard test block benchmark database based on standard test blocks containing known defects, including "defect type - actual size - defect characteristic parameters"; at the same time, perform initial screening and standardization on the preprocessed defect characteristic data in Step 1 to obtain standardized defect characteristic data, and store the standardized defect characteristic data in the standardized defect characteristic data layer of the effective characteristic database. Step 3: Perform feature processing and parameter mapping on the standardized defect feature data layer obtained and stored in the effective feature database in Step 2. The feature processing includes at least multimodal feature encoding, abstract feature extraction, and feature fusion. Based on this, generate defect feature representations and visualized defect parameters for subsequent intelligent image evaluation. The defect feature representation includes at least a unified structured feature vector, an abstract semantic feature vector, and a fused feature vector. The visualized defect parameters are generated based on the fused feature vector through parameter mapping. The unified structured feature vector, abstract semantic feature vector, fused feature vector, and visualized defect parameters are all stored back into the corresponding fields of the effective feature database. Step 4: Assess the confidence level of the concrete defect parameters generated and stored in the concrete defect feature database in Step 3. Based on the multi-dimensional assessment results, classify the confidence level of each concrete defect parameter and trigger the corresponding subsequent processing procedures according to different confidence levels. Step 5: Based on the confidence level results of the visualized defect parameters obtained in Step 4, generate and output the film review report.

2. The intelligent AI-based image evaluation and optimization method for marine welding data based on radiographic inspection as described in claim 1, characterized in that, Step one specifically includes the following steps: Step 101: Place the X-ray probe against the surface of the weld to be sampled, keeping the probe direction parallel to the center line of the weld. By emitting and receiving X-rays, non-destructive testing data is collected from the weld area to obtain the raw scan data. Step 102: Normalize and store the raw scan data to obtain normalized and stored raw scan data; Step 103: Preprocess the standardized and stored raw scan data to form array data, image data, and defect feature data extracted based on the array data and image data for subsequent analysis; Step 104: Establish a raw database for weld inspection data management, and uniformly store the array data, image data and corresponding defect feature data obtained after data preprocessing in Step 103.

3. The intelligent AI-based image evaluation and optimization method for marine welding data based on radiographic testing as described in claim 2, characterized in that, Step two specifically includes the following steps: The initial screening includes operations that meet the requirements of data completeness, data clarity, physical consistency, and data accuracy, specifically performed in the following steps: Step 201: The preprocessed defect feature data is collected and screened at the source to exclude invalid data caused by abnormal environmental conditions and obtain the first screening of defect feature data. Step 202: Perform signal quality screening on the "first-stage screening defect feature data" to remove low-quality data that is contaminated by noise or has incomplete signals, and obtain "second-stage screening defect feature data" with clear, distinguishable and complete signals. Step 203: Perform feature consistency screening on the "secondary screening defect feature data" to remove abnormal feature data and ensure that the feature results conform to the physical characteristics and statistical laws of weld defects in the "third screening defect feature data". Step 204: Establish a standard test block baseline database. The "three-stage screening defect feature data" obtained after screening in step 203 is compared and screened with the standard test block baseline feature parameters containing the corresponding defect types and defect sizes to verify whether the detection accuracy of the defect feature data meets the requirements. When the deviation exceeds the allowable range, it is judged as unqualified data and removed. The output is "four-stage screening defect feature data". Create an effective feature database and store the "four-stage screening defect feature data" in it to obtain an effective feature database for subsequent analysis.

4. The intelligent AI-based image evaluation and optimization method for marine welding data based on radiographic inspection as described in claim 3, characterized in that, Step four specifically includes the following steps: Step 401, Evaluation of the predictive stability of the visualized defect parameters: For the same detection sample, while keeping the data stored in the effective feature database after screening in step two unchanged, and originating from the same detection sample, different random initialization parameters, different batch inference orders, or different model inference rounds are introduced in the feature processing and parameter mapping process described in step three to perform K independent predictions on the same concrete defect parameter, thereby obtaining K predicted values ​​of the concrete defect parameter. Calculate the coefficient of variation (CV) for the K predicted values. CV is equal to the ratio of the standard deviation of each predicted value to its mean. Determine the prediction stability score S1 based on the CV values. When CV ≤ 0.05, S1 = 30; When 0.05 < CV ≤ 0.1, S1 = 20; When 0.1 < CV ≤ 0.2, S1 = 10; When CV > 0.2, S1 = 0; Step 402: Correlation assessment between visualized defect parameters and feature representations: Select M detection samples, and for the concrete defect parameter to be evaluated, calculate the Pearson correlation coefficient between the concrete defect parameter and each feature component in the unified structured feature vector or fused feature vector corresponding to each sample, and take the correlation coefficient with the largest absolute value, denoted as r_max; The correlation score S2 is determined based on the value of r_max. When |r_max| ≥ 0.85, S2 = 40; When 0.7 ≤ |r_max| < 0.85, S2 = 30; When 0.5 ≤ |r_max| < 0.7, S2 = 20; When |r_max| < 0.5, S2 = 0; Step 403, Physical rationality assessment of concrete defect parameters: For different types of concrete defect parameters, at least one physical rationality verification rule is predefined: For the defect depth parameter, it is stipulated that its predicted value shall not be greater than the nominal thickness of the base material of the weld being inspected; For the defect echo amplitude parameter, it is stipulated that its predicted value shall not be lower than the system baseline. Based on the conformity of the concrete defect parameters to the predefined physical rules, determine the physical rationality score S3: When all physical rules are met, S3 = 30; When only some physical rules are met, S3 = 15; When it does not conform to any physical rules, S3=0; Step 404: Confidence level comprehensive calculation, classification and automatic processing: This step integrates the evaluation results from steps 401 to 403, and performs confidence level calculation and grading for each concrete defect parameter; For each concrete defect parameter, calculate its final confidence score S, S=S1+S2+S3, and store the confidence score as an auxiliary attribute of the concrete defect parameter in the effective feature database. Based on the final confidence score S, the concrete defect parameters are divided into the following confidence levels: Grade A: 90 ≤ S ≤ 100; Grade B: 70 ≤ S < 90; Grade C: 50 ≤ S < 70; Grade D: S < 50; The system automatically executes the corresponding processing strategy based on the confidence level of the concrete defect parameters: For Class A parameters, the results are automatically adopted and a test report is generated. For Class B parameters, they are marked as low review priority in the test report; For Class C parameters, mark them as high priority for review in the test report and prompt for manual review; For Class D parameters, the test result is deemed invalid, and a retest process is triggered, returning to step one to re-collect and process data at the corresponding weld location.

Images