Casting surface treatment defect detection and quality evaluation method and system

Abstract

The invention discloses a casting surface treatment defect detection and quality evaluation method and system, and relates to the technical field of casting quality evaluation, and the method comprises the steps: collecting the surface data of a to-be-detected casting, and carrying out the data preprocessing, and obtaining a standardized input data set and a standardized data subset; calling a corresponding analysis sub-model for each subset, outputting quality features and confidence coefficients, and summarizing the quality features and the confidence coefficients into a sub-source result; environment state information is acquired to determine sub-model weights, and weighted fusion is carried out on sub-source results to obtain an evaluation result and an overall confidence coefficient; calculating space / feature / time consistency and judging according to a combination rule; when a re-checking condition is met, obtaining a re-checking label backflow updating data set, and training an updating model and parameters; and dynamically adjusting the threshold value and the weight value according to the performance index for subsequent evaluation. According to the method, multi-source data and environment information can be fused, weight self-adaption and closed-loop updating are parallel, accuracy and stability are improved, misjudgment and missed judgment are reduced, and complex working condition adaptability and long-term reliability are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of casting quality assessment, and in particular to a casting surface treatment defect detection and quality assessment method and system. Background Art

[0002] The quality inspection of post-casting surface treatment is evolving from experience-based manual inspections to data-driven intelligent assessments. With the application of inspection methods such as visible light imaging, thermal imaging, acoustic testing, ultrasonic testing, 3D point clouds, and spectral reflectance, surface defect recognition has evolved from single-modality threshold and geometric feature methods to deep learning-based object detection, segmentation, and anomaly recognition. This technology is also beginning to integrate multi-source surface inspection data for information fusion to improve coverage, sensitivity, and repeatability, promoting the implementation of online, closed-loop quality control scenarios.

[0003] However, existing technologies generally have three shortcomings. First, the common early fusion and late fusion are mostly fixed strategies, which fail to adjust the weight values ​​of each detection data source based on environmental status information such as lighting, temperature, humidity, dust, vibration and conveying speed, resulting in insufficient robustness under complex working conditions. Second, the end-to-end black box model usually does not provide quality feature output and confidence information divided by the source of the detection data, and it is impossible to calculate quantifiable consistency indicators. It is also difficult to implement explainable cross-validation, resulting in a lack of traceability when multi-source results conflict. Third, the existing solution lacks a continuous training and update mechanism centered on the review process and updated data sets. It is difficult to adapt to new defect forms and concept drift in a timely manner, and no combined judgment rules based on overall confidence and consistency indicators have been established to eliminate boundary ambiguity. Therefore, it is difficult to stably output consistent quality assessment conclusions.

[0004] The above problems determine that the existing technology cannot achieve the comprehensive technical effects of the present invention in terms of adaptive weighted fusion, explainable consistency judgment and closed-loop evolution. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problems solved by the present invention are: the existing methods for defect detection and quality assessment of casting surface treatment have the following problems: the fixed fusion strategy leads to insufficient robustness, the lack of output and consistency indicators based on the source of detection data leads to unexplainable judgment, the lack of review process and continuous training updates driven by updated data sets makes it difficult to cope with concept drift, and how to implement combined judgment based on overall confidence and consistency indicators in real-time scenarios and stably output quality assessment conclusions.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, an embodiment of the present invention provides a method for detecting and assessing casting surface treatment defects, comprising collecting multi-source surface inspection data of a casting to be inspected and performing data preprocessing to obtain a standardized input data set, wherein the standardized input data set is composed of standardized data subsets corresponding to each inspection data source; For each standardized data subset, a pre-established analysis sub-model corresponding to the detection data source is called to process the data, and the quality feature output and confidence information of the detection data source are obtained, and then merged into a source analysis result set; Obtain environmental status information, determine the weight value of each analysis sub-model based on the environmental status information, perform weighted fusion processing on the source analysis result set, and obtain the fusion quality assessment result and overall confidence; Calculate the consistency index between the source analysis result sets, judge the fusion quality assessment results based on the combined judgment rules of the overall confidence and consistency index, and output the quality assessment conclusion; When the quality assessment conclusion meets the preset review conditions, the review process is triggered to obtain the review label data, merge the review label data with the historical data into an updated data set, and train and update the analysis sub-model and the parameters used for weighted fusion processing and judgment processing based on the updated data set; The combination decision rules and weight values ​​are dynamically adjusted based on the evaluation performance indicators, and the adjustment results are applied to subsequent quality evaluations.

[0008] As a preferred embodiment of the method for detecting and assessing casting surface treatment defects according to the present invention, the multi-source surface detection data collected from the casting to be inspected includes visible light image data, thermal imaging data, acoustic detection data, ultrasonic detection data, three-dimensional point cloud data, and spectral reflectance data.

[0009] As a preferred embodiment of the method for detecting and assessing casting surface treatment defects according to the present invention, the data preprocessing includes performing geometric distortion correction, illumination balancing, reflection highlight suppression, denoising, clarity detection and blur image removal, geometric calibration, and color calibration on the visible light image data to obtain standardized visible light image data; Perform non-uniformity correction, temperature calibration, emissivity setting and reflected background temperature compensation, thermal drift compensation, and temperature field smoothing on thermal imaging data to obtain standardized thermal imaging data; The acoustic detection data is processed by removing DC components, bandpass filtering, frame dividing and windowing, short-time Fourier transform, amplitude normalization, and channel synchronization calibration to obtain acoustic detection standardized data; Perform time base calibration, pulse echo denoising, coupling condition compensation, envelope extraction, time of flight alignment, amplitude and phase correction on ultrasonic detection data to obtain standardized ultrasonic detection data; The 3D point cloud data is processed by outlier removal, denoising, point cloud registration, normal vector estimation, resampling, and scale normalization, and spatially aligned with the visible light image coordinates to obtain 3D point cloud standardized data; The spectral reflectance data is processed by dark current subtraction, white board correction, band response correction, reflectance normalization, and light drift compensation to obtain spectral reflectance standardized data; The obtained standardized data are subjected to unified time alignment processing and unified spatial alignment processing, and signal-to-noise ratio threshold detection, saturation threshold detection and abnormal sample elimination processing are carried out. Each standardized data is mapped to a predetermined numerical range and divided into visible light image standardized data subset, thermal imaging standardized data subset, acoustic detection standardized data subset, ultrasonic detection standardized data subset, three-dimensional point cloud standardized data subset and spectral reflectance standardized data subset according to the source of the detection data, and merged to form a standardized input data set.

[0010] As a preferred embodiment of the casting surface treatment defect detection and quality assessment method of the present invention, the method of obtaining the quality feature output and confidence information of the detection data source includes processing a subset of the visible light image normalized data using an analysis sub-model corresponding to the visible light image detection data source, and outputting a defect area mask, defect boundary coordinates, defect area, defect length, defect direction, surface roughness index, and corresponding confidence information; The standardized thermal imaging data subset is processed by the analysis sub-model corresponding to the source of the thermal imaging detection data, and the thermal anomaly area mask, temperature extremes, temperature differences, thermal gradients, thermal anomaly duration and corresponding confidence information are output; The standardized acoustic detection data subset is processed by the analysis sub-model corresponding to the acoustic detection data source, and the time-frequency energy spectrum characteristics, formant frequency, formant bandwidth, energy ratio, envelope amplitude, acoustic anomaly score and corresponding confidence information are output; The standardized ultrasonic testing data subset is processed by the analysis sub-model corresponding to the ultrasonic testing data source, and the echo amplitude curve, flight time, reflection coefficient, attenuation coefficient, defect depth estimation, interface continuity score and corresponding confidence information are output; The standardized 3D point cloud data subset is processed by the analysis sub-model corresponding to the 3D point cloud detection data source, and the surface curvature distribution, normal vector change rate, pit index, step height, geometric residual, dimensional deviation and corresponding confidence information are output; The spectral reflectance standardized data subset is processed by the analysis sub-model corresponding to the spectral reflectance detection data source, and the characteristic band reflectance, band ratio, material composition index, oxidation degree index, surface contamination index and corresponding confidence information are output; The quality feature output and confidence information of the visible light image detection data source, the quality feature output and confidence information of the thermal imaging detection data source, the quality feature output and confidence information of the acoustic detection data source, the quality feature output and confidence information of the ultrasonic detection data source, the quality feature output and confidence information of the three-dimensional point cloud detection data source, and the quality feature output and confidence information of the spectral reflectance detection data source are merged according to the source of the detection data, retaining the source identification, spatial location information and time identification to form a source analysis result set.

[0011] As a preferred embodiment of the method for detecting and assessing casting surface treatment defects according to the present invention, obtaining the fusion quality assessment result and the overall confidence level includes obtaining environmental state information and performing standardization and time synchronization processing to form an environmental state information vector; According to the preset weight determination rules, the weight value is calculated for each analysis sub-model based on the environmental state information vector; Perform spatial alignment and scale alignment on the source analysis result set in a unified spatial coordinate system to keep the source identification, spatial location information and time identification consistent; Perform weighted fusion processing on the source analysis result set according to the calculated weight value to obtain the fusion quality assessment result; The confidence information in the source analysis result set is weighted and summarized with the corresponding weight value to obtain the overall confidence, and the fusion quality assessment result and the overall confidence are used as input for subsequent judgment processing.

[0012] As a preferred embodiment of the casting surface treatment defect detection and quality assessment method of the present invention, the output quality assessment conclusion includes extracting source identification, spatial location information, time identification, and quality feature output from the source analysis result set, and performing comparison under a unified spatial coordinate system and a unified time reference; Calculating a spatial consistency index, where the spatial consistency index is calculated based on the overlap ratio between defect location results from various detection data sources; Calculating a feature consistency index, the feature consistency index being calculated based on the similarity between normalized feature vectors outputted by the quality features of each detection data source, the similarity being calculated using a preset similarity metric; Calculating a time consistency index, wherein the time consistency index is calculated based on the consistency degree of the time identifiers of the detection data sources within the same collection period; The spatial consistency index, feature consistency index and time consistency index are weighted according to the preset consistency weight to obtain a comprehensive consistency index; The fusion quality assessment result is judged according to the combined judgment rules, and the combined judgment rules include: When the overall confidence is greater than or equal to the first confidence threshold and the comprehensive consistency index is greater than or equal to the first consistency threshold, if the defect severity result in the fusion quality assessment result is greater than the defect judgment threshold, the output quality assessment conclusion is defective; if the defect severity result in the fusion quality assessment result is less than or equal to the defect judgment threshold, the output quality assessment conclusion is qualified; When the overall confidence is less than or equal to the second confidence threshold or the comprehensive consistency index is less than or equal to the second consistency threshold, the output quality assessment conclusion is pending review; If the above two conditions are not met, the output quality assessment conclusion is pending review; wherein, the second confidence threshold is less than the first confidence threshold, and the second consistency threshold is less than the first consistency threshold.

[0013] As a preferred embodiment of the method for detecting and assessing casting surface treatment defects according to the present invention, the training and updating of the analysis sub-model and the parameters for weighted fusion processing and judgment processing based on the updated data set includes pushing the fusion quality assessment results, overall confidence, comprehensive consistency index, source identification, spatial location information, and time identification to the review terminal, and calling the visible light image standardized data subset, thermal imaging standardized data subset, acoustic detection standardized data subset, ultrasonic detection standardized data subset, three-dimensional point cloud standardized data subset, and spectral reflectance standardized data subset for review and display; Label the object to be reviewed according to the preset labeling specification and output review label data, which includes defect location results, defect severity results and defect category information; Pairing the visible light image standardized data subset with the corresponding review label data, the thermal imaging standardized data subset with the corresponding review label data, the acoustic detection standardized data subset with the corresponding review label data, the ultrasonic detection standardized data subset with the corresponding review label data, the 3D point cloud standardized data subset with the corresponding review label data, and the spectral reflectance standardized data subset with the corresponding review label data, and merging them to form an updated data set; Based on the updated data set, the analysis sub-model corresponding to the visible light image detection data source, the analysis sub-model corresponding to the thermal imaging detection data source, the analysis sub-model corresponding to the acoustic detection data source, the analysis sub-model corresponding to the ultrasonic detection data source, the analysis sub-model corresponding to the three-dimensional point cloud detection data source, and the analysis sub-model corresponding to the spectral reflectance detection data source are trained and updated to obtain updated analysis sub-model parameters; The parameters used for weighted fusion processing and the parameters used for decision processing are trained and updated based on the updated data set, and the updated analysis sub-model parameters, parameters used for weighted fusion processing and parameters used for decision processing are put into subsequent quality evaluation.

[0014] As a preferred embodiment of the method for detecting and evaluating casting surface treatment defects according to the present invention, the method further comprises dynamically adjusting the combined judgment rules and weights based on the evaluation performance indicators, comprising statistically evaluating the performance indicators within a time window of a preset length, wherein the evaluation performance indicators include a correct judgment rate, a false alarm rate, a missed alarm rate, and an average processing delay; Comparing the evaluation performance index with the corresponding preset target value to obtain a comparison result; According to the comparison result and the preset adjustment rule, the first confidence threshold, the second confidence threshold, the first consistency threshold and the second consistency threshold in the combination determination rule are updated, and the weight value in the weight determination rule is updated; The updated first confidence threshold, second confidence threshold, first consistency threshold, second consistency threshold and weight value are applied to subsequent quality assessment.

[0015] In a second aspect, an embodiment of the present invention provides a casting surface treatment defect detection and quality assessment system, comprising: Data acquisition and preprocessing module: collects multi-source surface inspection data of the casting to be inspected and performs data preprocessing to obtain a standardized input data set, which is composed of standardized data subsets corresponding to each inspection data source; Sub-model analysis module: For each standardized data subset, a pre-established analysis sub-model corresponding to the detection data source is called to process the data, obtain the quality feature output and confidence information of the detection data source, and merge them into a source analysis result set; Fusion and confidence calculation module: obtains environmental status information, determines the weight value of each analysis sub-model based on the environmental status information, performs weighted fusion processing on the source analysis result set, and obtains the fusion quality assessment result and overall confidence; Consistency calculation and combination judgment module: Calculates the consistency index between the source analysis result sets, judges the fusion quality assessment results based on the overall confidence and the combination judgment rules of the consistency index, and outputs the quality assessment conclusion; Review and training update module: When the quality assessment conclusion meets the preset review conditions, the review process is triggered, the review label data is obtained, the review label data is merged with the historical data into an updated data set, and the analysis sub-model and the parameters used for weighted fusion processing and judgment processing are trained and updated based on the updated data set; Dynamic adjustment module: Dynamically adjusts the combination decision rules and weight values ​​based on the evaluation performance indicators, and applies the adjustment results to subsequent quality evaluations.

[0016] The beneficial effects of the present invention are as follows: the present invention can integrate multi-source detection data and environmental status information, dynamically allocate analysis sub-model weights and integrate evaluation results, which not only improves the accuracy and stability of casting surface treatment quality judgment, but also continuously optimizes model performance through review and adaptive parameter update mechanism, reduces misjudgment and missed judgment, and significantly improves the adaptability and long-term reliability of the detection system under complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive work, among which: Figure 1 This is an overall flow chart of a casting surface treatment defect detection and quality assessment method provided by the first embodiment of the present invention. DETAILED DESCRIPTION

[0018] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0019] Example 1, with reference to Figure 1 , as one embodiment of the present invention, provides a method for detecting and evaluating casting surface treatment defects and quality, comprising: S1: Collect multi-source surface inspection data of the casting to be inspected and perform data preprocessing to obtain a standardized input data set, wherein the standardized input data set is composed of standardized data subsets corresponding to each inspection data source.

[0020] The multi-source surface inspection data collected for the casting to be inspected include visible light image data, thermal imaging data, acoustic inspection data, ultrasonic inspection data, three-dimensional point cloud data and spectral reflectance data.

[0021] Perform geometric distortion correction, illumination balancing, reflection highlight suppression, denoising, clarity detection and blur image removal, geometric calibration and color calibration on visible light image data to obtain standardized visible light image data; Perform non-uniformity correction, temperature calibration, emissivity setting and reflected background temperature compensation, thermal drift compensation, and temperature field smoothing on thermal imaging data to obtain standardized thermal imaging data; The acoustic detection data is processed by removing DC components, bandpass filtering, frame dividing and windowing, short-time Fourier transform, amplitude normalization, and channel synchronization calibration to obtain acoustic detection standardized data; Perform time base calibration, pulse echo denoising, coupling condition compensation, envelope extraction, time of flight alignment, amplitude and phase correction on ultrasonic detection data to obtain standardized ultrasonic detection data; The 3D point cloud data is processed by outlier removal, denoising, point cloud registration, normal vector estimation, resampling, and scale normalization, and spatially aligned with the visible light image coordinates to obtain 3D point cloud standardized data; The spectral reflectance data is processed by dark current subtraction, white board correction, band response correction, reflectance normalization, and light drift compensation to obtain spectral reflectance standardized data; The obtained standardized data are subjected to unified time alignment processing and unified spatial alignment processing, and signal-to-noise ratio threshold detection, saturation threshold detection and abnormal sample elimination processing are carried out. Each standardized data is mapped to a predetermined numerical range and divided into visible light image standardized data subset, thermal imaging standardized data subset, acoustic detection standardized data subset, ultrasonic detection standardized data subset, three-dimensional point cloud standardized data subset and spectral reflectance standardized data subset according to the source of the detection data, and merged to form a standardized input data set.

[0022] In an embodiment of the present invention, in order to ensure the repeatability and traceability of multi-source surface inspection data in step S1, a fixed acquisition unit is set at the first inspection station after the casting surface treatment. The six types of inspection data sources share a unified trigger signal and a unified timestamp to complete time synchronization processing, and the encoder of the conveyor device is used as a speed reference to trigger the acquisition according to the window where the workpiece enters the field of view. The calibration parts and spatial posture solution are used to establish the spatial alignment processing between sensors, and the external parameters of each sensor to the workpiece coordinate system are solved. The residual and reprojection errors are used for verification, and the channels that exceed the limit value will not enter the subsequent process. After the acquisition is completed, the channel self-test is immediately performed, including exposure and clarity inspection of the optical channel, temperature drift inspection of the thermal imaging channel, channel consistency inspection of the acoustic and ultrasonic channels, coverage and porosity inspection of the three-dimensional point cloud channel, and illumination stability inspection of the spectral channel. A quality report for this batch is generated and written into the data header.

[0023] In the data preprocessing stage, the imaging channel eliminates the systematic deviation caused by the lens and lighting through geometric distortion correction and illumination balance, and suppresses the pseudo defects caused by mirror reflection through reflection highlight suppression. Denoising and clarity detection ensure the signal-to-noise ratio of texture and edge. The thermal imaging channel separates the equipment error and material radiation characteristics through non-uniformity correction, temperature calibration and emissivity setting, and then stabilizes the temperature field with reflection background temperature compensation and thermal drift compensation, and suppresses random noise through smoothing. The acoustic channel is subjected to DC component removal, bandpass filtering, frame windowing and short-time Fourier transform to concentrate the energy on the frequency band that characterizes cracks and debonding, amplitude normalization and channel Synchronous calibration ensures the comparability of the output of each channel in the array; the ultrasonic channel obtains a consistent depth scale through time reference calibration and flight time alignment, pulse echo denoising, coupling condition compensation and envelope extraction improve the detectability of weak echoes, and amplitude and phase correction eliminate probe and coupling differences; the three-dimensional point cloud channel obtains stable geometric quantities through outlier removal, alignment, normal vector estimation and resampling, and the scale is normalized and aligned with visible light coordinates to form a bidirectional index of pixels and point clouds; the spectral channel obtains the true reflectance spectrum through dark current subtraction, whiteboard correction, band response correction and illumination drift compensation, and reflectance normalization eliminates lighting differences between different batches.

[0024] Quality gating and numerical mapping are then performed. Quality gating includes signal-to-noise ratio threshold detection, saturation threshold detection and abnormal sample removal. The gating threshold is determined by historical statistics and metrological calibration. Data frames and data areas that fail are marked as invalid and the source identification, spatial location information and time identification are retained for traceability. Numerical mapping linearly maps the data that passes the gate to a unified numerical domain from zero to one. The mapping function and calibration parameters are recorded together to ensure that the input scale of the subsequent analysis sub-model is consistent. If a certain detection data source is missing on the current workpiece or fails to pass the gating, an empty standardized data subset is generated and an invalid flag is attached. The weight value of the detection data source in the subsequent links can be set to zero without affecting the continuity of the process.

[0025] It should also be noted that the objects of time synchronization processing and spatial alignment processing between sensors are standardized data, with the goal of establishing the same time base and unified spatial reference across all detection data sources, so that the same defect has consistent spatial location information and time identification in the six types of detection data sources. The standardized input data set ultimately consists of a standardized data subset for visible light images, a standardized data subset for thermal imaging, a standardized data subset for acoustic detection, a standardized data subset for ultrasonic detection, a standardized data subset for 3D point clouds, and a standardized data subset for spectral reflectance, and an index structure composed of batch number, workpiece number, source identification, spatial location information, and time identification. The above technical settings ensure that the input data has the same location, scale, and reliable signal-to-noise ratio, providing a verifiable basis for the subsequent invocation of analysis sub-models according to the source of detection data, the implementation of adaptive weighted fusion, and consistency judgment.

[0026] S2: For each standardized data subset, call the pre-established analysis sub-model corresponding to the detection data source for processing, obtain the quality feature output and confidence information of the detection data source, and merge them into a source analysis result set.

[0027] The standardized visible light image data subset is processed by the analysis sub-model corresponding to the visible light image detection data source, and the defect area mask, defect boundary coordinates, defect area, defect length, defect direction, surface roughness index and corresponding confidence information are output; The standardized thermal imaging data subset is processed by the analysis sub-model corresponding to the source of the thermal imaging detection data, and the thermal anomaly area mask, temperature extremes, temperature differences, thermal gradients, thermal anomaly duration and corresponding confidence information are output; The standardized acoustic detection data subset is processed by the analysis sub-model corresponding to the acoustic detection data source, and the time-frequency energy spectrum characteristics, formant frequency, formant bandwidth, energy ratio, envelope amplitude, acoustic anomaly score and corresponding confidence information are output; The standardized ultrasonic testing data subset is processed by the analysis sub-model corresponding to the ultrasonic testing data source, and the echo amplitude curve, flight time, reflection coefficient, attenuation coefficient, defect depth estimation, interface continuity score and corresponding confidence information are output; The standardized 3D point cloud data subset is processed by the analysis sub-model corresponding to the 3D point cloud detection data source, and the surface curvature distribution, normal vector change rate, pit index, step height, geometric residual, dimensional deviation and corresponding confidence information are output; The spectral reflectance standardized data subset is processed by the analysis sub-model corresponding to the spectral reflectance detection data source, and the characteristic band reflectance, band ratio, material composition index, oxidation degree index, surface contamination index and corresponding confidence information are output; The quality feature output and confidence information of the visible light image detection data source, the quality feature output and confidence information of the thermal imaging detection data source, the quality feature output and confidence information of the acoustic detection data source, the quality feature output and confidence information of the ultrasonic detection data source, the quality feature output and confidence information of the three-dimensional point cloud detection data source, and the quality feature output and confidence information of the spectral reflectance detection data source are merged according to the source of the detection data, retaining the source identification, spatial location information and time identification to form a source analysis result set.

[0028] In an embodiment of the present invention, step S2 takes a standardized input data set as input and calls pre-established analysis sub-models according to the source of the detection data. The analysis sub-model has completed parameter training and fixing before deployment, and has completed the consistency of the output scale through the calibration data within the batch. For the subset of standardized data of visible light images, the analysis sub-model outputs the defect area mask and the defect boundary coordinates, calculates the defect area, defect length and defect direction based on the connected domain and boundary tracking, and generates confidence information with the evidence amount composed of texture and gradient stability; this output provides pixel-level position evidence for subsequent spatial consistency indicators. For the subset of standardized data of thermal imaging, the analysis sub-model segments the thermal anomaly area mask in the temperature field, extracts the temperature extremes, temperature differences and thermal gradients, and forms confidence information with the time period stability of the abnormal area and the background contrast; this output is used to determine the severity of residual stress or local overheating after treatment. For a standardized subset of acoustic testing data, the analysis submodel identifies frequency band structures associated with cracks or debonding in the time-frequency energy spectrum, quantifies resonant peak frequency, resonant peak bandwidth, energy ratio, and envelope amplitude, and provides confidence information based on spectral peak clarity and channel consistency. This output complements invisible subsurface defects. For a standardized subset of ultrasonic testing data, the analysis submodel analyzes echo amplitude curves and flight time on the time axis, calculates reflection coefficients and attenuation coefficients, and combines flight time to estimate defect depth. It also uses echo morphological integrity to form an interface continuity score and confidence information. This output directly provides the depth dimension, providing a longitudinal reference for severity estimation in subsequent fusion. For a standardized subset of 3D point cloud data, the analysis submodel calculates surface curvature distribution, normal vector change rate, pitting index, step height, and geometric residual in a uniformly scaled geometric field, generating confidence information based on local geometric consistency. This output is highly sensitive to external anomalies such as sand holes, strains, and steps. For a subset of spectral reflectance standardized data, the analytical sub-model infers material composition indicators, oxidation degree indicators, and surface contamination indicators from the characteristic band reflectance and band ratio, and generates confidence information using the spectral shape fitting residuals; this output can distinguish between optical artifacts of material discoloration and real defects.

[0029] The confidence information of each of the above-mentioned analysis sub-models is obtained based on the mapping function established by the model output probability and historical statistics, and is calibrated once with the calibration frame within the current batch to ensure that the confidence information of different detection data sources is comparable. In order to ensure spatial and temporal traceability, each quality feature output and the corresponding confidence information carry a source identifier, spatial location information and a time identifier. If a certain detection data source is marked as invalid by the quality gate in step S1, the detection data source generates an empty quality feature output and zero-value confidence information in step S2, while retaining the source identifier, spatial location information and time identifier, without interrupting the subsequent process.

[0030] It should also be noted that when the sub-sources are merged into a sub-source analysis result set, they are strictly grouped and stored according to the source of the detection data, and the consistent index relationship of quality feature output, confidence information, source identification, spatial location information and time identification is maintained within the group, so that in the subsequent weighted fusion processing based on environmental state information to determine the weight value, differentiated weight allocation can be implemented for different detection data sources; at the same time, when calculating the consistency index, the defect area mask, defect depth estimation, thermal anomaly area mask, geometric residual and material composition index can be compared item by item under the same spatial location information and time identification to form a quantifiable measure of spatial consistency, feature consistency and time consistency. Through the above settings, step S2 not only provides the discrimination information and reliability measurement of each detection data source, but also ensures that the subsequent fusion and judgment are interpretable and verifiable with structured data organization, meeting the technical objectives of the present invention.

[0031] S3: Obtain environmental status information, determine the weight value of each analysis sub-model based on the environmental status information, perform weighted fusion processing on the source analysis result set, and obtain the fusion quality assessment result and overall confidence.

[0032] Obtain environmental status information and perform standardization and time synchronization to form an environmental status information vector; According to the preset weight determination rules, the weight value is calculated for each analysis sub-model based on the environmental state information vector; Perform spatial alignment and scale alignment on the source analysis result set in a unified spatial coordinate system to keep the source identification, spatial location information and time identification consistent; Perform weighted fusion processing on the source analysis result set according to the calculated weight value to obtain the fusion quality assessment result; The confidence information in the source analysis result set is weighted and summarized with the corresponding weight value to obtain the overall confidence, and the fusion quality assessment result and the overall confidence are used as input for subsequent judgment processing.

[0033] In this embodiment of the present invention, step S3 takes the source analysis results and environmental state information as input. The environmental state information is first normalized and time-synchronized to form an environmental state information vector. The environmental state information vector includes at least elements such as light intensity, ambient temperature, ambient humidity, ambient vibration intensity, airborne dust concentration, ambient electromagnetic interference intensity, and conveyor line speed, and is aligned with the source identifier, spatial location information, and time identifier for the same time period. Based on the environmental state information vector, weights are calculated for each analysis sub-model according to a preset weight determination rule. The weight determination rule can use a table lookup or piecewise function to establish a clear mapping relationship. For example, a low-level weight is assigned to the visible light channel in low-light and high-dust scenarios, a low-level weight is assigned to the thermal imaging channel in high-temperature drift scenarios, and a low-level weight is assigned to the acoustic and ultrasonic channels in high-vibration scenarios. The weights are normalized and used for subsequent fusion. If a detection data source is marked as invalid in step S1, the weight of that detection data source is set to zero, but the source identifier is retained for traceability.

[0034] Subsequently, the source analysis results are spatially aligned and scaled within a unified spatial coordinate system. Spatial alignment maps positioning-related quantities, such as defect region masks, defect boundary coordinates, thermal anomaly region masks, defect depth estimates, and geometric residuals, to a uniform resolution grid within the workpiece coordinate system. Scale unification unifies the measurement dimensions of different channels. For example, metrics such as depth, temperature difference, curvature, and energy ratio are calibrated to the numerical domain of the standardized input dataset, ensuring that the outputs of different inspection data sources are comparable on the same numerical scale.

[0035] In the weighted fusion processing, the source identifier, spatial location information and time identifier are used as indexes to perform fusion operations on multi-channel results at the same location and time period. The positioning results use probability domain weighted averaging to obtain a fused defect probability map, and then threshold segmentation is used to obtain the fused defect positioning results and fused defect boundary coordinates. The severity results use a weighted combination strategy, using the ultrasonic defect depth estimation and the geometric residual of the three-dimensional point cloud as the dominant quantities, the temperature difference and thermal gradient of thermal imaging as auxiliary quantities, and the energy characteristics and reflection coefficient of acoustics and ultrasound as credibility correction quantities to obtain the fused defect severity results. To improve stability, small-scale morphological smoothing and connectivity consistency checks are applied to the fusion results of adjacent positions to avoid isolated noise points from forming misjudgment clusters.

[0036] The overall confidence is obtained by weighted aggregation of the confidence information and corresponding weight values ​​in the source analysis result set. To ensure comparability across batches, the calculation of the overall confidence uses normalized weights and calibrated confidence information, and records metadata such as weight values, the number of detection data sources involved in the fusion, and the quality gate pass rate. If only some detection data sources at a certain location participate in the fusion, the overall confidence automatically considers the reduction factor of the number of participants when summarizing, thereby avoiding inflated conclusions caused by incomplete information. The fusion quality assessment results and the overall confidence serve as inputs for subsequent judgment processing, providing a location-level, dimensionally unified and traceable evidence basis for comprehensive consistency indicators and combined judgment rules.

[0037] It should also be noted that the spatial alignment and scale unification processing in this step are performed on the analysis output results, which is different from the spatial alignment and numerical mapping between sensors for standardized data in step S1. The weight determination rules can be obtained through offline training of historical data and fixed in the system with a version number. When the production line environment changes, only the mapping table needs to be updated without changing the subsequent process. Through this setting, the sensitivity of the weight value to environmental changes and the consistency of the fusion calculation are simultaneously guaranteed, so that the fusion quality assessment results and overall confidence can remain stable and interpretable under complex working conditions.

[0038] S4: Calculate the consistency index between the source analysis result sets, judge the fusion quality assessment results according to the combined judgment rules of the overall confidence and consistency index, and output the quality assessment conclusion.

[0039] Extract source identification, spatial location information, time identification, and quality feature output from the source analysis result set, and compare them under a unified spatial coordinate system and a unified time base; Calculating a spatial consistency index, where the spatial consistency index is calculated based on the overlap ratio between defect location results from various detection data sources; Calculating a feature consistency index, the feature consistency index being calculated based on the similarity between normalized feature vectors outputted by the quality features of each detection data source, the similarity being calculated using a preset similarity metric; Calculating a time consistency index, wherein the time consistency index is calculated based on the consistency degree of the time identifiers of the detection data sources within the same collection period; The spatial consistency index, feature consistency index and time consistency index are weighted according to the preset consistency weight to obtain a comprehensive consistency index; The fusion quality assessment result is judged according to the combined judgment rules, and the combined judgment rules include: When the overall confidence is greater than or equal to the first confidence threshold and the comprehensive consistency index is greater than or equal to the first consistency threshold, if the defect severity result in the fusion quality assessment result is greater than the defect judgment threshold, the output quality assessment conclusion is defective; if the defect severity result in the fusion quality assessment result is less than or equal to the defect judgment threshold, the output quality assessment conclusion is qualified; When the overall confidence is less than or equal to the second confidence threshold or the comprehensive consistency index is less than or equal to the second consistency threshold, the output quality assessment conclusion is pending review; If the above two conditions are not met, the output quality assessment conclusion is pending review; wherein, the second confidence threshold is less than the first confidence threshold, and the second consistency threshold is less than the first consistency threshold.

[0040] In an embodiment of the present invention, step S4 takes the source analysis result set, fusion quality assessment result and overall confidence as input, and first compares the source identification, spatial position information, time identification and quality feature output position by position under the unified time base after the unified spatial coordinate system and time synchronization processing. The spatial consistency index is calculated according to the overlap ratio of the defect area mask and the defect boundary coordinates. The ultrasonic defect depth estimation and the acoustic energy concentration area are projected onto the workpiece surface to participate in the calculation of the overlap ratio. The geometric residual of the three-dimensional point cloud is divided by the threshold to form an abnormal area to participate in the calculation of the overlap ratio. For the detection data source that only provides regional-level evidence, regional granularity alignment is used, and for the detection data source that only provides point-level evidence, the neighborhood expansion is used to participate in the calculation to ensure that different evidence forms are comparable on the same grid. The feature consistency index is based on the similarity calculation of normalized feature vectors. The normalized feature vectors contain dimensionally unified quantities such as defect depth estimation, temperature difference, thermal gradient, curvature, step height, energy ratio, reflectivity, and material composition indicators after numerical mapping. At any position, test data sources that do not participate in the output are excluded from the similarity calculation for that position and are eliminated from the denominator to avoid introducing zero-value bias. The temporal consistency index calculates the degree of alignment of the time stamps of each test data source within the same acquisition period. The value is determined by a deterministic function composed of the period coverage ratio and the maximum time deviation, ensuring that interval acquisition channels and synchronous acquisition channels are comparable under unified rules.

[0041] The comprehensive consistency index is derived by weighting the spatial consistency index, feature consistency index, and temporal consistency index according to preset consistency weights. These weights are determined offline based on historical data before system deployment and remain constant within a batch. To mitigate the impact of isolated noise points on the comprehensive consistency index, a small-scale consistency smoothing and connectivity check are applied to the comprehensive consistency index after position-level calculation. For scattered anomalies that occur in only a few channels across detection data sources, the comprehensive consistency index is constrained to a lower range, thereby preventing the direct judgment path from being triggered.

[0042] In the combined judgment, the system first checks whether the overall confidence and comprehensive consistency index reach the first confidence threshold and the first consistency threshold. When both are reached, the system determines whether the quality assessment conclusion is defective or qualified based on the comparison of the defect severity result in the fusion quality assessment result and the defect judgment threshold; when the overall confidence is not high or the comprehensive consistency index is not high, the quality assessment conclusion is directly output as pending review. The effect of the above rules is illustrated by a typical case: if the visible light image and the three-dimensional point cloud give the defect area mask and geometric residual at the same position, but the thermal imaging and acoustics show no abnormalities, the spatial consistency index and the feature consistency index can still be maintained in the medium and high range; if the defect severity result in the fusion quality assessment result exceeds the defect judgment threshold, the quality assessment conclusion is defective; if the overall confidence is insufficient due to fewer participating channels, the quality assessment conclusion is pending review. For example, if ultrasound gives a clear estimate of the defect depth but the visible light image has no surface signs, projection and neighborhood expansion are used to participate in the calculation of the spatial consistency index, and the depth, energy ratio and reflection coefficient in the feature consistency index reflect the consistency of evidence; when the comprehensive consistency index and the overall confidence reach the threshold, stable judgment of sub-surface defects can still be achieved.

[0043] It should also be noted that the value ranges of the spatial consistency index, feature consistency index and temporal consistency index are unified to the numerical domain of zero to one, and the number of detection data sources involved in the calculation, the ratio of visible coverage area and the maximum time deviation are recorded as metadata for subsequent review and version tracing. The comprehensive consistency index and the overall confidence are only used as inputs to the combined judgment rules, and are not directly used as substitutes for the defect severity results; the defect severity results are always derived from the fusion quality assessment results to ensure that the sources of evidence in the decision-making chain are clear and separable. The equality judgment in boundary cases has been clearly defined as the closed interval conditions of greater than or equal to and less than or equal to in the combined judgment rules, and the remaining situations are uniformly classified into the path to be reviewed to avoid the threshold gray area. Through the above settings, step S4 forms a complete and verifiable judgment process in terms of evidence alignment, measurement unification and conflict resolution, providing a basis for the stable output of quality assessment conclusions.

[0044] S5: When the quality assessment conclusion meets the preset review conditions, the review process is triggered, the review label data is obtained, the review label data is merged with the historical data into an updated data set, and the analysis sub-model and the parameters used for weighted fusion processing and judgment processing are trained and updated based on the updated data set.

[0045] The fusion quality assessment results, overall confidence, comprehensive consistency indicators, source identification, spatial location information and time identification are pushed to the review terminal, and the standardized data subsets of visible light images, thermal imaging, acoustic detection, ultrasonic detection, 3D point cloud and spectral reflectance are called for review and display; Label the object to be reviewed according to the preset labeling specification and output review label data, which includes defect location results, defect severity results and defect category information; Pairing the visible light image standardized data subset with the corresponding review label data, the thermal imaging standardized data subset with the corresponding review label data, the acoustic detection standardized data subset with the corresponding review label data, the ultrasonic detection standardized data subset with the corresponding review label data, the 3D point cloud standardized data subset with the corresponding review label data, and the spectral reflectance standardized data subset with the corresponding review label data, and merging them to form an updated data set; Based on the updated data set, the analysis sub-model corresponding to the visible light image detection data source, the analysis sub-model corresponding to the thermal imaging detection data source, the analysis sub-model corresponding to the acoustic detection data source, the analysis sub-model corresponding to the ultrasonic detection data source, the analysis sub-model corresponding to the three-dimensional point cloud detection data source, and the analysis sub-model corresponding to the spectral reflectance detection data source are trained and updated to obtain updated analysis sub-model parameters; The parameters used for weighted fusion processing and the parameters used for decision processing are trained and updated based on the updated data set, and the updated analysis sub-model parameters, parameters used for weighted fusion processing and parameters used for decision processing are put into subsequent quality evaluation.

[0046] In an embodiment of the present invention, step S5 takes the quality assessment conclusion, the fusion quality assessment result, the overall confidence, the comprehensive consistency index, the source identification, the spatial location information and the time identification as input. When the quality assessment conclusion is to be reviewed or meets the preset review conditions, the system will synchronously display the visible light image standardized data subset, the thermal imaging standardized data subset, the acoustic detection standardized data subset, the ultrasonic detection standardized data subset, the three-dimensional point cloud standardized data subset and the spectral reflectance standardized data subset on the review terminal, and align them with the same spatial location information and the same time identification. The review terminal marks the position to be reviewed according to the preset marking specification to form review label data. The review label data contains at least defect location results, defect severity results and defect category information, and retains the source identification, spatial location information and time identification to ensure traceability. In order to avoid deviation, in addition to reviewing the samples to be reviewed, the system also extracts samples with an overall confidence close to the threshold in the qualified conclusion and samples with a defect severity result close to the threshold in the defect conclusion at a fixed ratio for review to improve the marking quality and statistical adequacy of the boundary area.

[0047] The system pairs the six standardized data subsets with the corresponding review label data one by one to generate an updated data set. The matching rule is that only the same workpiece number, the same source identifier, the same spatial location information, and the same time identifier can be written. The reviewer number and the marking time of the review label data are also recorded simultaneously. Before writing, the updated data set performs a consistency check. The content includes spatial overlap verification of defect location results and 3D point cloud geometric residuals, range consistency verification of defect severity results and ultrasonic defect depth estimation, and correlation verification of thermal imaging temperature difference and acoustic energy ratio. If any of the checks fail, the data will be returned to the review terminal for re-examination.

[0048] The training and update process consists of two parts: training and updating the analysis sub-model and training and updating the parameters used for weighted fusion processing and decision processing. The analysis sub-model training and update utilizes an incremental strategy, merging the updated dataset with samples drawn from historical data, balanced by category and operating condition, to ensure stable category and operating condition distributions. Using the archived baseline set as an invisible validation set, the system monitors four metrics: correct judgment rate, false alarm rate, missed alarm rate, and average processing latency. If any metric degrades, the update is aborted and rolled back to the previous version. The training and update of the parameters used for weighted fusion processing and decision processing uses the updated dataset as input, refitting the mapping coefficients in the weight determination rules and the thresholds in the combined decision rules. Confidence calibration and consistency threshold calibration are performed on the archived baseline set, ensuring that the overall confidence and comprehensive consistency metrics remain monotonic and interpretable in their correspondence with the true labels. After the update is complete, the system archives metadata using the version number, training time, training data batch number, and key metrics. The system is then launched with low-traffic verification and transitions to full-scale application after operational indicators stabilize.

[0049] It should also be noted that the preset review conditions for the review trigger include the quality assessment conclusion being to be reviewed, samples whose overall confidence is close to the second confidence threshold, and samples whose comprehensive consistency index is close to the second consistency threshold; when a certain detection data source is marked as invalid in step S1, its weight value remains zero during the training update phase and does not participate in parameter learning to prevent the introduction of noise; each sample in the updated data set retains the source identification, spatial location information, and time identification to ensure that any model output in the future can be traced back to the original evidence and review process. Through the above process, the training update completes the parameter update without changing the data and result organization method of steps S1 to S4, ensuring that subsequent batches can obtain improved evaluation performance and stability under the same interface and index structure.

[0050] S6: Dynamically adjust the combination decision rules and weight values ​​based on the evaluation performance indicators, and apply the adjustment results to subsequent quality evaluations.

[0051] Statistically evaluate performance indicators within a time window of a preset length, wherein the performance indicators include correct determination rate, false alarm rate, missed alarm rate, and average processing delay; Comparing the evaluation performance index with the corresponding preset target value to obtain a comparison result; According to the comparison result and the preset adjustment rule, the first confidence threshold, the second confidence threshold, the first consistency threshold and the second consistency threshold in the combination determination rule are updated, and the weight value in the weight determination rule is updated; The updated first confidence threshold, second confidence threshold, first consistency threshold, second consistency threshold and weight value are applied to subsequent quality assessment.

[0052] In an embodiment of the present invention, step S6 performs hierarchical statistics and comparisons on the evaluation performance indicators based on the online samples and review label data within a time window of a preset length. The evaluation performance indicators include correct judgment rate, false alarm rate, missed alarm rate and average processing delay, which are respectively statistically analyzed under the product category, surface treatment process and environmental status layers to avoid structural deviations. Each indicator is compared with the corresponding preset target value to obtain a comparison result. Based on this, the system performs two-level dynamic adjustments: the fast loop is for the weight value, and the slow loop is for the first confidence threshold, the second confidence threshold, the first consistency threshold and the second consistency threshold in the combined judgment rule. The fast loop is executed once for each time window, and the slow loop is triggered only when multiple consecutive time windows deviate from the target to suppress oscillations.

[0053] In the fast loop, the source contribution and source error rate are calculated for each detection data source. The source contribution is defined as the increment of the correct judgment rate of the fusion quality assessment result when the source participates, and the source error rate is defined as the weighted increment of the false alarm rate and the missed alarm rate when the source participates relative to when it does not participate. The weight value of the source is updated proportionally according to the difference between the two. After the update, normalization is performed and projected to the preset upper and lower bounds to ensure that the sum of the weight values ​​and the limit constraints meet the requirements; when a detection data source is invalidated by quality gate in step S1, its weight value remains zero and does not participate in the update. If the average processing delay exceeds the target value, a delay penalty item is imposed on the detection data source with a large delay contribution, so that its weight value is slightly reduced, thereby bringing the overall delay back to the target range while ensuring that the recall rate does not drop significantly.

[0054] In the slow loop, the thresholds are adjusted based on the deviation direction of the false alarm rate and the missed alarm rate: when the false alarm rate is higher than the target and the missed alarm rate is lower than the target, the first confidence threshold or the first consistency threshold is increased to tighten the direct judgment entry; when the missed alarm rate is higher than the target and the false alarm rate is lower than the target, the first confidence threshold or the first consistency threshold is lowered to enhance sensitivity; when both are higher than the target, the fast loop first completes two rounds of weight value rebalancing, and then makes small adjustments to the first confidence threshold and the first consistency threshold in the same direction; the adjustment of the second confidence threshold and the second consistency threshold follows the boundary backoff strategy in the opposite direction to the above, which is used to control the upper and lower bounds of the proportion to be reviewed. Each threshold adjustment sets a maximum change and a hysteresis band. It only takes effect when the deviation direction of multiple consecutive time windows is consistent and the sample size reaches the preset lower limit, preventing frequent switching caused by small sample noise. The defect judgment threshold is determined by the training update process in step S5 and is not adjusted in step S6 to ensure clear responsibilities in the decision chain.

[0055] To ensure traceability, the system records the window number, sample size, layered dimensions, thresholds before and after adjustment, weight values ​​before and after adjustment, evaluation performance indicators and target values, and verification indicators for a time window after going online for each dynamic adjustment. If the verification indicator deteriorates beyond the preset tolerance, the system automatically rolls back to the previous version of the combined judgment rules and weight values, and marks the adjustment as invalid. Through the above settings, step S6 realizes bounded, layered, and rollback closed-loop adaptation of the combined judgment rules and weight values ​​without changing the data structure and interface, thereby maintaining the stability and interpretability of the fusion quality assessment results and overall confidence under complex and changing working conditions, and seamlessly applying the adjustment results to subsequent quality assessments.

[0056] Example 2 is the second embodiment of the present invention, which is different from the previous embodiment in that: If the functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art or the portion of the current technical solution, can be embodied in the form of a software product. The current computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0057] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0058] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, and then editing, interpreting, or processing in another suitable manner as necessary, and then storing it in a computer memory.

[0059] Example 3 is an embodiment of the present invention, which provides a casting surface treatment defect detection and quality assessment system, including a data acquisition and preprocessing module, a sub-model analysis module, a fusion and confidence calculation module, a consistency calculation and combination judgment module, a review and training update module and a dynamic adjustment module.

[0060] Data acquisition and preprocessing module: collects multi-source surface inspection data of the casting to be inspected and performs data preprocessing to obtain a standardized input data set, which is composed of standardized data subsets corresponding to each inspection data source; Sub-model analysis module: For each standardized data subset, a pre-established analysis sub-model corresponding to the detection data source is called to process the data, obtain the quality feature output and confidence information of the detection data source, and merge them into a source analysis result set; Fusion and confidence calculation module: obtains environmental status information, determines the weight value of each analysis sub-model based on the environmental status information, performs weighted fusion processing on the source analysis result set, and obtains the fusion quality assessment result and overall confidence; Consistency calculation and combination judgment module: Calculates the consistency index between the source analysis result sets, judges the fusion quality assessment results based on the overall confidence and the combination judgment rules of the consistency index, and outputs the quality assessment conclusion; Review and training update module: When the quality assessment conclusion meets the preset review conditions, the review process is triggered, the review label data is obtained, the review label data is merged with the historical data into an updated data set, and the analysis sub-model and the parameters used for weighted fusion processing and judgment processing are trained and updated based on the updated data set; Dynamic adjustment module: Dynamically adjusts the combination decision rules and weight values ​​based on the evaluation performance indicators, and applies the adjustment results to subsequent quality evaluations.

[0061] Example 4 is an embodiment of the present invention, which provides a method for detecting and evaluating surface treatment defects of castings. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation / comparative experiments.

[0062] Four common casting types were selected as test subjects: valve bodies (after shot peening), pump casings (after sand blasting), housings (after shot blasting), and brackets (after wire drawing), with 30 pieces of each type, for a total of 120 pieces. A fixed multi-source acquisition unit was installed at the first inspection station after surface treatment, including a visible light camera, infrared thermal imager, acoustic sensor array, ultrasonic roller probe, structured light 3D sensor, and multispectral imaging device. The six types of inspection data sources share a unified trigger and a unified timestamp for time synchronization. Calibration parts are used to solve the external parameters of each sensor to the workpiece coordinate system, enabling spatial alignment between sensors. The acquisition cycle is set according to the production line conveyor speed, and a single piece completes the acquisition of six-source data in one pass at the same station. After acquisition, preprocessing is performed immediately: the visible light channel performs geometric distortion correction, illumination equalization, highlight suppression, and denoising, and blurry frames are removed through clarity detection. The thermal imaging channel performs non-uniformity correction, temperature calibration, emissivity setting, thermal drift compensation, and temperature field smoothing. The acoustic channel performs DC removal, bandpass, frame windowing, and short-time Fourier transform, and performs amplitude normalization and channel synchronization calibration. The ultrasonic channel performs time base calibration, echo denoising, coupling compensation, envelope extraction, time-of-flight alignment, and amplitude and phase correction. The 3D point cloud channel performs outlier removal, registration, normal estimation, and resampling, and spatially aligns with visible light coordinates. The spectral channel performs dark current subtraction, whiteboard and band response correction, and illumination drift compensation. Each channel is then quality-gated (signal-to-noise ratio and saturation threshold), outlier samples are removed, and the gated data is mapped to a unified numerical domain [0, 1]. Six standardized data subsets are formed by source and merged into a standardized input dataset.

[0063] For each standardized data subset, pre-established analysis sub-models corresponding to the inspection data source are invoked: visible light outputs defect masks, boundaries, and roughness indicators; thermal imaging outputs thermal anomaly masks, temperature differences, and thermal gradients; acoustics outputs time-frequency energy, resonance peaks, energy ratios, and anomaly scores; ultrasound outputs echo amplitude, time of flight, reflection coefficient, attenuation coefficient, defect depth estimation, and interface continuity scores; 3D point clouds output curvature, step height, and geometric residuals; and spectroscopy outputs characteristic band reflectance, band ratios, material composition, and oxidation contamination indicators. All quality feature outputs and confidence information are combined into a source-specific analysis result set based on source identification, spatial location, and time stamp. Environmental state information (lighting, temperature and humidity, vibration, dust, electromagnetic interference, and conveyor speed) is recorded and standardized into an environmental state information vector. Weights for each analysis sub-model are calculated based on pre-set weighting rules. After spatial alignment and scaling of the source-specific analysis result set within a unified spatial coordinate system, weighted fusion is performed to obtain fused defect location and severity results, and the overall confidence is summarized based on the confidence and weight values. Spatial consistency, feature consistency, and temporal consistency are then calculated and weighted to form a comprehensive consistency index. A quality assessment conclusion is then output according to the combined judgment rules. For samples with a "pending review" conclusion and boundary samples, the verification terminal displays the six-source evidence in the same space and time. Verification label data is generated according to standard annotations and merged with historical data to form an updated dataset for incremental training of the analysis sub-model and recalibration of weights and judgment thresholds. Performance evaluation metrics (correct judgment rate, false alarm rate, missed alarm rate, and average processing latency) are statistically analyzed within a sliding window. For scenarios that deviate from the target, a hierarchical, rollback-friendly dynamic adjustment of fast weights and slow thresholds is implemented, and the updated results are directly applied to subsequent batches. For detailed data content, please refer to Table 1.

[0064] Table 1 Data reference table

[0065] As shown in Table 1, the "Defect Recall Rate (%)" and "F1 Score (%)" for all four casting types show a significant improvement for the proposed solution over the baseline solution. For example, the F1 score for the "shell - after shot blasting" model increased from 90.2% to 96.6%, while the recall rate for the "valve body - after shot blasting" model increased from 88.3% to 95.6%. This improvement is not achieved through single-channel threshold optimization but rather stems from the combined effects of weight adaptation driven by environmental state information and source consistency constraints. Specifically, the "Consistency Index [0-1]" improved by approximately 0.12–0.14 for all objects (e.g., 0.74 to 0.88 for the "valve body - after shot blasting" model). This indicates that the cross-correlation of source evidence in a unified space and time scale is higher, reducing the uncertainty in judgment caused by conflicting sources. Correspondingly, the "false alarm rate (%)" dropped from 6.7–8.6% to 3.8–5.1%, indicating that when spatial consistency or feature consistency is insufficient, the combined judgment rule effectively blocks the erroneous direct judgment path and guides it to "pending review", thereby reducing false alarms.

[0066] The improvement in "Overall Confidence [0-1]" (e.g., "Shell - After Shot Blasting" from 0.82 to 0.93) demonstrates that the global confidence, aggregated by weight and confidence, is more monotonically aligned with the true label. This improvement directly supports the significant decrease in "Percentage Awaiting Verification (%)" (e.g., "Pump Casing - After Shot Blasting" from 14.6% to 7.1%). The decrease in the percentage awaiting verification and the increase in the F1 score indicate that the approach is not simply to relax the threshold, but rather to improve the quality of evidence and fusion reliability through consistency metrics. Notably, the "Subsurface Defect Detection Rate (%)" significantly improved across all objects, for example, from 73.2% to 90.1% for "Shell - After Shot Blasting." This demonstrates that the present invention, by utilizing ultrasonic depth information and acoustic energy ratio as the dominant variables for severity estimation, supplemented by auxiliary variables such as 3D geometric residuals and thermal gradients, can reliably identify defects with subtle surface signs but significant deep anomalies. This is a result difficult to achieve using a single modality or fixed fusion strategy.

[0067] From an engineering perspective, the proposed method exhibits an acceptable increase in "Average Processing Latency (ms)" (on the order of 12–15 ms), but this yields a significant improvement in F1 by approximately 6–8 percentage points and a reduction in the proportion of applications awaiting review by approximately half. These improvements significantly enhance the overall performance while still meeting launch requirements within production line cycle times. More representative evidence comes from the "Post-Review Improvement ΔF1 (percentage points)" column: Driven by closed-loop review and updated datasets, the proposed method continues to achieve an F1 gain of 1.9–2.6 percentage points after initial launch. This demonstrates that parameters and submodels are continuously calibrated based on real-world field distributions, a crucial aspect in scenarios involving concept drift, where traditional fixed-parameter approaches struggle to achieve sustained gains.

[0068] Based on the above comparison, the inventiveness of the present invention lies in: first, the use of environmental state information vectors to implement differentiated weights for each analysis sub-model, solving the problem of uneven contributions and unequal interference from multiple sources under complex working conditions; second, the introduction of a position-level spatial-feature-temporal three-dimensional consistency indicator, which combines judgment rules to form an interpretable "direct judgment / pending review" diversion, avoiding threshold gray areas; and third, the construction of a closed-loop mechanism of "review-update dataset-incremental training-dynamic adjustment" to enable the evaluation performance to continue to evolve after going online. The systematic improvement of each indicator in the table has a clear causal relationship with the three technical features mentioned above, demonstrating that the present invention not only surpasses the existing technology in recognition accuracy, but also has outstanding technical effects and application value in terms of robustness, interpretability, and sustainable optimization capabilities.

[0069] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for detecting and evaluating surface defects in castings, characterized in that: include: Collecting multi-source surface inspection data of the casting to be inspected and performing data preprocessing to obtain a standardized input data set, wherein the standardized input data set is composed of standardized data subsets corresponding to each inspection data source; For each standardized data subset, a pre-established analysis sub-model corresponding to the detection data source is called to process the data, and the quality feature output and confidence information of the detection data source are obtained, and then merged into a source analysis result set; Obtain environmental status information, determine the weight value of each analysis sub-model based on the environmental status information, perform weighted fusion processing on the source analysis result set, and obtain the fusion quality assessment result and overall confidence; Calculate the consistency index between the source analysis result sets, judge the fusion quality assessment results based on the combined judgment rules of the overall confidence and consistency index, and output the quality assessment conclusion; When the quality assessment conclusion meets the preset review conditions, the review process is triggered to obtain the review label data, merge the review label data with the historical data into an updated data set, and train and update the analysis sub-model and the parameters used for weighted fusion processing and judgment processing based on the updated data set; The combination decision rules and weight values ​​are dynamically adjusted based on the evaluation performance indicators, and the adjustment results are applied to subsequent quality evaluations.

2. The casting surface treatment defect detection and quality assessment method according to claim 1, characterized in that: The multi-source surface detection data collected for the casting to be detected includes visible light image data, thermal imaging data, acoustic detection data, ultrasonic detection data, three-dimensional point cloud data and spectral reflection data.

3. The casting surface treatment defect detection and quality assessment method according to claim 2, characterized in that: The data preprocessing includes performing geometric distortion correction, illumination balancing, reflection highlight suppression, denoising, clarity detection and blur image removal, geometric calibration and color calibration on the visible light image data to obtain standardized visible light image data; Perform non-uniformity correction, temperature calibration, emissivity setting and reflected background temperature compensation, thermal drift compensation, and temperature field smoothing on thermal imaging data to obtain standardized thermal imaging data; The acoustic detection data is processed by removing DC components, bandpass filtering, frame dividing and windowing, short-time Fourier transform, amplitude normalization, and channel synchronization calibration to obtain acoustic detection standardized data; Perform time base calibration, pulse echo denoising, coupling condition compensation, envelope extraction, time of flight alignment, amplitude and phase correction on ultrasonic detection data to obtain standardized ultrasonic detection data; The 3D point cloud data is processed by outlier removal, denoising, point cloud registration, normal vector estimation, resampling, and scale normalization, and spatially aligned with the visible light image coordinates to obtain 3D point cloud standardized data; The spectral reflectance data is processed by dark current subtraction, white board correction, band response correction, reflectance normalization, and light drift compensation to obtain spectral reflectance standardized data; The obtained standardized data are subjected to unified time alignment processing and unified spatial alignment processing, and signal-to-noise ratio threshold detection, saturation threshold detection and abnormal sample elimination processing are carried out. Each standardized data is mapped to a predetermined numerical range and divided into visible light image standardized data subset, thermal imaging standardized data subset, acoustic detection standardized data subset, ultrasonic detection standardized data subset, three-dimensional point cloud standardized data subset and spectral reflectance standardized data subset according to the source of the detection data, and merged to form a standardized input data set.

4. The casting surface treatment defect detection and quality assessment method according to claim 3, characterized in that: Obtaining the quality feature output and confidence information of the detection data source includes processing the visible light image normalized data subset using an analysis sub-model corresponding to the visible light image detection data source, and outputting a defect area mask, defect boundary coordinates, defect area, defect length, defect direction, surface roughness index, and corresponding confidence information; The standardized thermal imaging data subset is processed by the analysis sub-model corresponding to the source of the thermal imaging detection data, and the thermal anomaly area mask, temperature extremes, temperature differences, thermal gradients, thermal anomaly duration and corresponding confidence information are output; The standardized acoustic detection data subset is processed by the analysis sub-model corresponding to the acoustic detection data source, and the time-frequency energy spectrum characteristics, formant frequency, formant bandwidth, energy ratio, envelope amplitude, acoustic anomaly score and corresponding confidence information are output; The standardized ultrasonic testing data subset is processed by the analysis sub-model corresponding to the ultrasonic testing data source, and the echo amplitude curve, flight time, reflection coefficient, attenuation coefficient, defect depth estimation, interface continuity score and corresponding confidence information are output; The standardized 3D point cloud data subset is processed by the analysis sub-model corresponding to the 3D point cloud detection data source, and the surface curvature distribution, normal vector change rate, pit index, step height, geometric residual, dimensional deviation and corresponding confidence information are output; The spectral reflectance standardized data subset is processed by the analysis sub-model corresponding to the spectral reflectance detection data source, and the characteristic band reflectance, band ratio, material composition index, oxidation degree index, surface contamination index and corresponding confidence information are output; The quality feature output and confidence information of the visible light image detection data source, the quality feature output and confidence information of the thermal imaging detection data source, the quality feature output and confidence information of the acoustic detection data source, the quality feature output and confidence information of the ultrasonic detection data source, the quality feature output and confidence information of the three-dimensional point cloud detection data source, and the quality feature output and confidence information of the spectral reflectance detection data source are merged according to the source of the detection data, retaining the source identification, spatial location information and time identification to form a source analysis result set.

5. The casting surface treatment defect detection and quality assessment method according to claim 4, characterized in that: Obtaining the fusion quality assessment result and the overall confidence includes obtaining environmental state information and performing normalization and time synchronization processing to form an environmental state information vector; According to the preset weight determination rules, the weight value is calculated for each analysis sub-model based on the environmental state information vector; Perform spatial alignment and scale alignment on the source analysis result set in a unified spatial coordinate system to keep the source identification, spatial location information and time identification consistent; Perform weighted fusion processing on the source analysis result set according to the calculated weight value to obtain the fusion quality assessment result; The confidence information in the source analysis result set is weighted and summarized with the corresponding weight value to obtain the overall confidence, and the fusion quality assessment result and the overall confidence are used as input for subsequent judgment processing.

6. The casting surface treatment defect detection and quality assessment method according to claim 5, characterized in that: The output quality assessment conclusion includes extracting source identification, spatial location information, time identification and quality feature output from the source analysis result set, and comparing them under a unified spatial coordinate system and a unified time base; Calculating a spatial consistency index, where the spatial consistency index is calculated based on the overlap ratio between defect location results from various detection data sources; Calculating a feature consistency index, the feature consistency index being calculated based on the similarity between normalized feature vectors outputted by the quality features of each detection data source, the similarity being calculated using a preset similarity metric; Calculating a time consistency index, wherein the time consistency index is calculated based on the consistency degree of the time identifiers of the detection data sources within the same collection period; The spatial consistency index, feature consistency index and time consistency index are weighted according to the preset consistency weight to obtain a comprehensive consistency index; The fusion quality assessment result is judged according to the combined judgment rules, and the combined judgment rules include: When the overall confidence is greater than or equal to the first confidence threshold and the comprehensive consistency index is greater than or equal to the first consistency threshold, if the defect severity result in the fusion quality assessment result is greater than the defect judgment threshold, the output quality assessment conclusion is defective; if the defect severity result in the fusion quality assessment result is less than or equal to the defect judgment threshold, the output quality assessment conclusion is qualified; When the overall confidence is less than or equal to the second confidence threshold or the comprehensive consistency index is less than or equal to the second consistency threshold, the output quality assessment conclusion is pending review; If the above two conditions are not met, the output quality assessment conclusion is pending review; wherein, the second confidence threshold is less than the first confidence threshold, and the second consistency threshold is less than the first consistency threshold.

7. The casting surface treatment defect detection and quality assessment method according to claim 6, characterized in that: The training and updating of the analysis sub-model and the parameters for weighted fusion processing and decision processing based on the updated data set includes pushing the fusion quality assessment results, overall confidence, comprehensive consistency index, source identification, spatial location information and time identification to the review terminal, and calling the visible light image standardized data subset, thermal imaging standardized data subset, acoustic detection standardized data subset, ultrasonic detection standardized data subset, three-dimensional point cloud standardized data subset and spectral reflectance standardized data subset for review and display; Label the object to be reviewed according to the preset labeling specification and output review label data, which includes defect location results, defect severity results and defect category information; Pairing the visible light image standardized data subset with the corresponding review label data, the thermal imaging standardized data subset with the corresponding review label data, the acoustic detection standardized data subset with the corresponding review label data, the ultrasonic detection standardized data subset with the corresponding review label data, the 3D point cloud standardized data subset with the corresponding review label data, and the spectral reflectance standardized data subset with the corresponding review label data, and merging them to form an updated data set; Based on the updated data set, the analysis sub-model corresponding to the visible light image detection data source, the analysis sub-model corresponding to the thermal imaging detection data source, the analysis sub-model corresponding to the acoustic detection data source, the analysis sub-model corresponding to the ultrasonic detection data source, the analysis sub-model corresponding to the three-dimensional point cloud detection data source, and the analysis sub-model corresponding to the spectral reflectance detection data source are trained and updated to obtain updated analysis sub-model parameters; The parameters used for weighted fusion processing and the parameters used for decision processing are trained and updated based on the updated data set, and the updated analysis sub-model parameters, parameters used for weighted fusion processing and parameters used for decision processing are put into subsequent quality evaluation.

8. The casting surface treatment defect detection and quality assessment method according to claim 7, characterized in that: The dynamically adjusting the combination decision rules and weights according to the evaluation performance indicators includes statistically evaluating the performance indicators within a time window of a preset length, wherein the evaluation performance indicators include correct decision rate, false alarm rate, missed alarm rate, and average processing delay; Comparing the evaluation performance index with the corresponding preset target value to obtain a comparison result; According to the comparison result and the preset adjustment rule, the first confidence threshold, the second confidence threshold, the first consistency threshold and the second consistency threshold in the combination determination rule are updated, and the weight value in the weight determination rule is updated; The updated first confidence threshold, second confidence threshold, first consistency threshold, second consistency threshold and weight value are applied to subsequent quality assessment.

9. A casting surface treatment defect detection and quality assessment system, used to implement the casting surface treatment defect detection and quality assessment method according to any one of claims 1 to 8, characterized in that: include: Data acquisition and preprocessing module: collects multi-source surface inspection data of the casting to be inspected and performs data preprocessing to obtain a standardized input data set, which is composed of standardized data subsets corresponding to each inspection data source; Sub-model analysis module: For each standardized data subset, a pre-established analysis sub-model corresponding to the detection data source is called to process the data, obtain the quality feature output and confidence information of the detection data source, and merge them into a source analysis result set; Fusion and confidence calculation module: obtains environmental status information, determines the weight value of each analysis sub-model based on the environmental status information, performs weighted fusion processing on the source analysis result set, and obtains the fusion quality assessment result and overall confidence; Consistency calculation and combination judgment module: Calculates the consistency index between the source analysis result sets, judges the fusion quality assessment results based on the overall confidence and the combination judgment rules of the consistency index, and outputs the quality assessment conclusion; Review and training update module: When the quality assessment conclusion meets the preset review conditions, the review process is triggered, the review label data is obtained, the review label data is merged with the historical data into an updated data set, and the analysis sub-model and the parameters used for weighted fusion processing and judgment processing are trained and updated based on the updated data set; Dynamic adjustment module: Dynamically adjusts the combination decision rules and weight values ​​based on the evaluation performance indicators, and applies the adjustment results to subsequent quality evaluations.

Images