A method and system for detecting and evaluating the quality of casting surface treatment defects

Abstract

The application discloses a kind of casting surface treatment defect detection and quality evaluation method and system, it is related to casting quality evaluation technical field, comprising: the surface data of the casting to be detected is collected and data is preprocessed, obtains standardization input data set and standardization data subset;For each subset, call corresponding analysis submodel, output quality characteristics and confidence and are summarized as source result;Acquire environmental state information to determine submodel weight, weighted fusion is obtained to evaluation result and overall confidence for source result;Calculate space / feature / time consistency and determine according to combination rule;When meeting review condition, obtain review label backflow update data set, train update model and parameter;According to performance index, dynamically adjust threshold and weight value and be used for subsequent evaluation.The application can fuse multi-source data and environmental information, weight self-adapting and closed-loop updating are parallel, improve accuracy and stability, reduce misjudgment and miss judgment, enhance complex working condition adaptability and long-term reliability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of casting quality evaluation, in particular to a casting surface treatment defect detection and quality evaluation method and system. BACKGROUND

[0002] The quality inspection of casting surface treatment is evolving from experience-based manual inspection to data-driven intelligent evaluation. With the application of visible light imaging, thermal imaging, acoustic detection, ultrasonic detection, three-dimensional point cloud and spectral reflection, surface defect recognition has developed from single-mode threshold and geometric feature methods to target detection, segmentation and anomaly recognition based on deep learning, and has begun to combine multi-source surface detection data for information fusion to improve coverage, sensitivity and repeatability, and promote the landing of online and closed-loop quality control scenarios.

[0003] However, the existing technology generally has three shortcomings. First, common early fusion and late fusion are mostly fixed strategies, which fail to adjust the weight values of each detection data source according to environmental state information such as light, temperature, humidity, dust, vibration and conveying speed, resulting in insufficient robustness under complex working conditions. Second, end-to-end black box models usually do not give quality feature output and confidence information divided by detection data sources, cannot calculate quantifiable consistency indicators, and are difficult to implement interpretable cross-validation, so there is no traceable basis when multi-source results conflict. Third, the existing solutions lack a continuous training update mechanism centered on review processes and updated data sets, making it difficult to adapt to new defect morphologies and concept drift in a timely manner, and there is no combined decision rule based on overall confidence and consistency indicators to eliminate boundary ambiguity, so it is difficult to stably output consistent quality evaluation conclusions.

[0004] The above problems determine that the existing technology cannot achieve the comprehensive technical effects of the present application in adaptive weighted fusion, interpretable consistency determination and closed-loop evolution. SUMMARY

[0005] In view of the above problems, the present application is proposed.

[0006] Therefore, the technical problem solved by the present application is that the existing casting surface treatment defect detection and quality evaluation method has the problems of fixed fusion strategy leading to insufficient robustness, lack of detection data source output and consistency indicators leading to uninterpretable determination, lack of continuous training update driven by review processes and updated data sets to cope with concept drift, and how to implement combined determination based on overall confidence and consistency indicators in real-time scenarios and stably output quality evaluation conclusions.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] In a first aspect, the embodiments of the present application provide a casting surface treatment defect detection and quality evaluation method, comprising: collecting multi-source surface detection data of a casting to be detected and performing data preprocessing to obtain a standardized input data set, the standardized input data set being composed of standardized data subsets corresponding to each detection data source;

[0009] For each standardized data subset, an analysis sub-model corresponding to the detection data source is called to process, and quality feature output and confidence information of the detection data source are obtained, which are merged into a source analysis result set;

[0010] Obtain environmental state information, determine the weight value of each analysis sub-model based on the environmental state information, and perform weighted fusion processing on the source analysis result set to obtain a fusion quality evaluation result and an overall confidence;

[0011] Calculate the consistency index between the source analysis result set, determine the fusion quality evaluation result according to the combination judgment rule of the overall confidence and the consistency index, and output the quality evaluation conclusion;

[0012] When the quality evaluation conclusion meets the preset review condition, trigger the review process, obtain review label data, merge the review label data and historical data into an update data set, and train and update the analysis sub-model and the parameters used for weighted fusion processing and determination processing based on the update data set;

[0013] According to the evaluation performance index, dynamically adjust the combination judgment rule and the weight value, and apply the adjustment result to subsequent quality evaluation.

[0014] As a preferred scheme of the casting surface treatment defect detection and quality evaluation method described in the present application, the collection of multi-source surface detection data of 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.

[0015] As a preferred scheme of the casting surface treatment defect detection and quality evaluation method described in the present application, the data preprocessing includes: performing geometric distortion correction, illumination equalization, reflection highlight suppression, denoising processing, sharpness detection and blurred image rejection, geometric calibration and color calibration on the visible light image data to obtain visible light image standardized data.

[0016] Perform non-uniformity correction, temperature calibration, emissivity setting and reflection background temperature compensation, thermal drift compensation and temperature field smoothing processing on the thermal imaging data to obtain thermal imaging standardized data.

[0017] The acoustic detection data is subjected to DC component removal processing, band-pass filtering processing, frame windowing processing, short-time Fourier transform processing, amplitude normalization processing, and channel synchronization calibration to obtain acoustic detection standardized data;

[0018] The ultrasonic detection data is subjected to time reference calibration, pulse echo denoising processing, coupling condition compensation processing, envelope extraction processing, time-of-flight alignment processing, and amplitude and phase correction processing to obtain ultrasonic detection standardized data;

[0019] The three-dimensional point cloud data is subjected to outlier removal processing, denoising processing, point cloud registration processing, normal vector estimation processing, resampling processing, and scale normalization processing, and is spatially aligned with the visible light image coordinates to obtain three-dimensional point cloud standardized data;

[0020] The spectral reflectance data is subjected to dark current subtraction processing, whiteboard correction processing, band response correction processing, reflectance normalization processing, and illumination drift compensation processing to obtain spectral reflectance standardized data;

[0021] The obtained standardized data is subjected to unified time alignment processing and unified spatial alignment processing, and is subjected to signal-to-noise ratio threshold detection, saturation threshold detection, and abnormal sample removal processing, and is mapped to a predetermined numerical range, and is divided into a visible light image standardized data subset, a thermal imaging standardized data subset, an acoustic detection standardized data subset, an ultrasonic detection standardized data subset, a three-dimensional point cloud standardized data subset, and a spectral reflectance standardized data subset according to the detection data source, and is combined to form a standardized input data set.

[0022] As a preferred scheme of the casting surface treatment defect detection and quality evaluation method, the quality feature output and confidence information of the obtained detection data source includes processing the visible light image standardized data subset by an analysis sub-model corresponding to the visible light image detection data source to output a defect region mask, a defect boundary coordinate, a defect area, a defect length, a defect direction, a surface roughness index, and corresponding confidence information.

[0023] The thermal imaging standardized data subset is processed by an analysis sub-model corresponding to the thermal imaging detection data source to output a thermal anomaly region mask, a temperature extreme value, a temperature difference, a thermal gradient, a thermal anomaly duration, and corresponding confidence information.

[0024] The acoustic detection standardized data subset is processed by an analysis sub-model corresponding to the acoustic detection data source to output a time-frequency energy spectrum feature, a resonance peak frequency, a resonance peak bandwidth, an energy ratio, an envelope amplitude, an acoustic anomaly score, and corresponding confidence information.

[0025] The ultrasonic detection standardization data subset is processed by an analysis sub-model corresponding to the ultrasonic detection data source, and echo amplitude curve, flight time, reflection coefficient, attenuation coefficient, defect depth estimation, interface continuity score and corresponding confidence information are outputted;

[0026] The three-dimensional point cloud standardization data subset is processed by an analysis sub-model corresponding to the three-dimensional point cloud detection data source, and surface curvature distribution, normal vector change rate, pit index, step height, geometric residual, size deviation and corresponding confidence information are outputted;

[0027] The spectral reflection standardization data subset is processed by an analysis sub-model corresponding to the spectral reflection detection data source, and feature band reflectivity, band ratio, material composition index, oxidation degree index, surface pollution index and corresponding confidence information are outputted;

[0028] 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 reflection detection data source are merged according to the detection data source respectively to reserve source identification, spatial position information and time identification, and a source analysis result set is formed.

[0029] As a preferred scheme of the casting surface treatment defect detection and quality evaluation method, the fusion quality evaluation result and the overall confidence degree are obtained by acquiring environmental state information and performing standardization processing and time synchronization processing to form an environmental state information vector;

[0030] According to a preset weight determination rule, a weight value is calculated for each analysis sub-model according to the environmental state information vector;

[0031] The source analysis result set is subjected to spatial alignment processing and scale alignment processing in a unified spatial coordinate system, and the source identification, spatial position information and time identification are kept consistent;

[0032] The source analysis result set is subjected to weighted fusion processing according to the calculated weight value, and a fusion quality evaluation result is obtained;

[0033] The confidence information in the source analysis result set and the corresponding weight value are subjected to weighted summation to obtain an overall confidence degree, and the fusion quality evaluation result and the overall confidence degree are taken as inputs for subsequent judgment processing.

[0034] As a preferred scheme of the casting surface treatment defect detection and quality evaluation method, the output quality evaluation conclusion comprises extracting source identification, spatial position information, time identification and quality characteristic output from the set of source analysis results, and comparing in a unified spatial coordinate system and a unified time reference.

[0035] A spatial consistency index is calculated, which is calculated based on the overlap ratio between the defect positioning results of each detection data source;

[0036] A feature consistency index is calculated, which is calculated based on the similarity between the normalized feature vectors of the quality characteristic outputs of each detection data source, and the similarity is calculated by using a preset similarity measure;

[0037] A time consistency index is calculated, which is calculated based on the consistency degree of the time identifications of each detection data source within the same collection period;

[0038] The spatial consistency index, the feature consistency index and the time consistency index are weighted according to a preset consistency weight to obtain a comprehensive consistency index;

[0039] The fusion quality evaluation result is determined according to a combination determination rule, and the combination determination rule comprises:

[0040] When the overall confidence is greater than or equal to a first confidence threshold and the comprehensive consistency index is greater than or equal to a first consistency threshold, if the defect severity result in the fusion quality evaluation result is greater than a defect determination threshold, the quality evaluation conclusion is output as a defect; if the defect severity result in the fusion quality evaluation result is less than or equal to the defect determination threshold, the quality evaluation conclusion is output as qualified;

[0041] When the overall confidence is less than or equal to a second confidence threshold or the comprehensive consistency index is less than or equal to a second consistency threshold, the quality evaluation conclusion is output as pending review;

[0042] For the case that does not satisfy the two conditions, the quality evaluation conclusion is output as 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.

[0043] As a preferred scheme of the casting surface treatment defect detection and quality evaluation method, wherein: 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 comprises pushing the fusion quality evaluation result, the overall confidence, the comprehensive consistency index, and the source identification, the spatial position information, and the time identification to a review terminal, calling 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 reflection standardized data subset for review display.

[0044] According to the preset annotation specification, the to-be-reviewed object is annotated, and review label data is output. The review label data includes defect positioning results, defect severity results, and defect category information.

[0045] The visible light image standardized data subset and the corresponding review label data, the thermal imaging standardized data subset and the corresponding review label data, the acoustic detection standardized data subset and the corresponding review label data, the ultrasonic detection standardized data subset and the corresponding review label data, the three-dimensional point cloud standardized data subset and the corresponding review label data, and the spectral reflection standardized data subset and the corresponding review label data are paired and merged to form an updated data set.

[0046] 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 reflection detection data source are trained and updated to obtain updated analysis sub-model parameters.

[0047] Based on the updated data set, the parameters for weighted fusion processing and the parameters for decision processing are trained and updated. The updated analysis sub-model parameters, the parameters for weighted fusion processing, and the parameters for decision processing are input into subsequent quality evaluation.

[0048] As a preferred scheme of the casting surface treatment defect detection and quality evaluation method, wherein: the dynamic adjustment of the combined decision rule and the weight according to the evaluation performance index comprises: within a time window of a preset length, the evaluation performance index is counted. The evaluation performance index includes correct decision rate, false alarm rate, missed alarm rate, and average processing time delay.

[0049] The evaluation performance index is compared with the corresponding preset target value to obtain a comparison result.

[0050] update the first confidence threshold, the second confidence threshold, the first consistency threshold and the second consistency threshold in the combination determination rule according to the preset adjustment rule according to the comparison result, and update the weight value in the weight determination rule;

[0051] apply the updated first confidence threshold, the second confidence threshold, the first consistency threshold, the second consistency threshold and the weight value to subsequent quality evaluation.

[0052] In a second aspect, an embodiment of the present application provides a cast surface treatment defect detection and quality evaluation system, comprising:

[0053] A data acquisition and preprocessing module acquires multi-source surface detection data of a cast to be detected and performs data preprocessing to obtain a standardized input data set, wherein the standardized input data set is composed of standardized data subsets corresponding to each detection data source;

[0054] A sub-model analysis module calls an analysis sub-model corresponding to each detection data source and pre-established to process each standardized data subset, to obtain quality feature outputs and confidence information of the detection data source, and to merge into a source-by-source analysis result set;

[0055] A fusion and confidence calculation module acquires environmental state information, determines weight values of each analysis sub-model based on the environmental state information, and performs weighted fusion processing on the source-by-source analysis result set to obtain a fusion quality evaluation result and an overall confidence;

[0056] A consistency calculation and combination determination module calculates consistency indicators between the source-by-source analysis result set, determines the fusion quality evaluation result according to a combination determination rule of the overall confidence and the consistency indicators, and outputs a quality evaluation conclusion;

[0057] A review and training update module triggers a review process when the quality evaluation conclusion meets a preset review condition, obtains review label data, merges the review label data and historical data into an update data set, and trains and updates the analysis sub-model and parameters for weighted fusion processing and determination processing based on the update data set;

[0058] A dynamic adjustment module dynamically adjusts the combination determination rule and the weight value according to an evaluation performance indicator, and applies the adjustment result to subsequent quality evaluation.

[0059] The present application has the following advantages: the present application can comprehensively integrate multi-source detection data and environmental state information, dynamically allocate analysis sub-model weights and fuse evaluation results, not only improves the accuracy and stability of cast surface treatment quality determination, but also continuously optimizes model performance through a review and self-adaptive parameter update mechanism, reduces misjudgment and omission, and significantly improves the adaptability and long-term reliability of the detection system under complex working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and all other drawings obtained by those skilled in the art without creative work on the premise of the drawings should also belong to the protection scope of the present application.

[0061] Figure 1 A flow chart of a casting surface treatment defect detection and quality evaluation method is provided for the first embodiment of the present application. DETAILED DESCRIPTION

[0062] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification. Obviously, the described embodiments are only some embodiments of the present application, rather than all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work on the premise should also belong to the protection scope of the present application.

[0063] Embodiment 1, refer to Figure 1 For an embodiment of the present application, a casting surface treatment defect detection and quality evaluation method is provided, which comprises:

[0064] S1: Collecting multi-source surface detection data of the casting to be detected and performing data preprocessing to obtain a standardized input data set, the standardized input data set being composed of standardized data subsets corresponding to each detection data source.

[0065] The multi-source surface detection data of 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.

[0066] The visible light image data is subjected to geometric distortion correction, illumination equalization, reflection highlight suppression, denoising processing, sharpness detection and fuzzy image rejection, geometric calibration and color calibration to obtain visible light image standardized data.

[0067] The thermal imaging data is subjected to non-uniformity correction, temperature calibration, emissivity setting and reflection background temperature compensation, thermal drift compensation and temperature field smoothing processing to obtain thermal imaging standardized data.

[0068] The acoustic detection data is subjected to direct current component removal processing, band-pass filtering processing, frame windowing processing, short-time Fourier transform processing, amplitude normalization processing and channel synchronization calibration to obtain acoustic detection standardized data.

[0069] The ultrasonic detection data is subjected to time reference calibration, pulse echo denoising processing, coupling condition compensation processing, envelope extraction processing, time of flight alignment processing, amplitude and phase correction processing, to obtain ultrasonic detection standardized data;

[0070] The three-dimensional point cloud data is subjected to outlier rejection processing, denoising processing, point cloud registration processing, normal vector estimation processing, resampling processing, scale normalization processing, and is spatially aligned with the visible light image coordinates, to obtain three-dimensional point cloud standardized data;

[0071] The spectral reflectance data is subjected to dark current subtraction processing, whiteboard correction processing, band response correction processing, reflectance normalization processing, and illumination drift compensation processing, to obtain spectral reflectance standardized data;

[0072] The obtained various standardized data is subjected to unified time alignment processing and unified spatial alignment processing, and is subjected to signal-to-noise ratio threshold detection, saturation threshold detection and abnormal sample rejection processing, and is mapped to a predetermined numerical range, and is divided into a visible light image standardized data subset, a thermal imaging standardized data subset, an acoustic detection standardized data subset, an ultrasonic detection standardized data subset, a three-dimensional point cloud standardized data subset and a spectral reflectance standardized data subset according to the source of the detection data, and is combined to form a standardized input data set.

[0073] In the embodiment of the application, in order to ensure the repeatability and traceability of the multi-source surface detection data, a fixed acquisition unit is arranged at the first detection station after the surface treatment of the casting, and the six types of detection data sources share a unified trigger signal and a unified timestamp to complete time synchronization processing, and the encoder of the transmission device is used as a speed reference to trigger acquisition according to the window of the workpiece entering the field of view. A calibration piece and a spatial pose solver are used to establish a spatial alignment process between sensors, to solve the external parameters of each sensor to the workpiece coordinate system, and to check the residual error and the re-projection error, and the channels exceeding the limit value are not entered into the subsequent process. After acquisition is completed, self-checking of the channels is immediately performed, including exposure and clarity check of the optical channel, temperature drift check of the thermal imaging channel, channel consistency check of the acoustic and ultrasonic channels, coverage and hole rate check of the three-dimensional point cloud channel, and illumination stability check of the spectral channel, to generate a quality report for this batch and write it into the data header.

[0074] In the data preprocessing link, the imaging channel eliminates systematic deviations caused by the lens and illumination through geometric distortion correction and illumination equalization, reduces pseudo defects caused by mirror reflection through reflection highlight suppression, and guarantees the signal-to-noise ratio of texture and edge through denoising and clarity detection; the thermal imaging channel strips device errors and material radiation characteristics through non-uniformity correction, temperature calibration and emissivity setting, and then stabilizes the temperature field through reflection background temperature compensation and thermal drift compensation to suppress random noise spots; the acoustic channel enables energy to be concentrated in the frequency band representing cracks and debonding through de-DC component, band-pass filtering, frame windowing and short-time Fourier transform, and guarantees the comparability of array channel outputs through amplitude normalization and channel synchronization calibration; the ultrasonic channel obtains consistent depth scales through time reference calibration and time of flight alignment, and improves the detectability of weak echoes through pulse echo denoising, coupling condition compensation and envelope extraction; the three-dimensional point cloud channel obtains stable geometric quantities through outlier rejection, registration, normal vector estimation and resampling, and forms a bidirectional index of pixels and point clouds through scale normalization and alignment with visible light coordinates; the spectral channel obtains true reflectance spectra through dark current subtraction, whiteboard correction, band response correction and illumination drift compensation, and eliminates differences in different batches of illumination through reflectance normalization.

[0075] Quality gating and numerical mapping are then performed. Quality gating includes signal-to-noise ratio threshold detection, saturation threshold detection and abnormal sample rejection, and the gating threshold is determined by historical statistics and metrological calibration. Data frames and data regions that do not pass the gating are marked as invalid and retain source identification, spatial location information and time identification for tracing. Numerical mapping linearly maps the data that passes the gating to a uniform numerical domain of zero to one, and the mapping function and calibration parameters are recorded together to ensure consistent input scale of subsequent analysis sub-models. If a certain detection data source is missing or does not pass the gating on the current workpiece, an empty standardized data subset is generated and an invalid flag is attached. The weight value of the detection data source in the subsequent link can be set to zero without affecting the continuity of the process.

[0076] It should also be noted that the object of the time synchronization processing and the spatial alignment processing between sensors is the standardized data, and the purpose is to establish the same time reference and uniform spatial reference across 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 is finally composed of visible light image standardized data subset, thermal imaging standardized data subset, acoustic detection standardized data subset, ultrasonic wave detection standardized data subset, three-dimensional point cloud standardized data subset and spectral reflection standardized data subset, and is indexed by batch number, workpiece number, source identification, spatial location information and time identification. The above technical settings ensure that the input data has the same position, same scale and reliable signal-to-noise ratio, and provide a verifiable basis for subsequent calling of analysis sub-models, implementation of adaptive weighted fusion and consistency determination according to detection data sources.

[0077] S2: For each standardized data subset, call the pre-established and corresponding analysis sub-model of the detection data source for processing to obtain the quality characteristic output and confidence information of the detection data source, and merge into a source analysis result set.

[0078] The visible light image standardized 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;

[0079] The thermal imaging standardized data subset is processed by the analysis sub-model corresponding to the thermal imaging detection data source, and the thermal anomaly area mask, temperature extreme value, temperature difference, thermal gradient, thermal anomaly duration and corresponding confidence information are output;

[0080] The acoustic detection standardized data subset is processed by the analysis sub-model corresponding to the acoustic detection data source, and the time-frequency energy spectrum feature, resonance peak frequency, resonance peak bandwidth, energy ratio, envelope amplitude, acoustic anomaly score and corresponding confidence information are output;

[0081] The ultrasonic wave detection standardized data subset is processed by the analysis sub-model corresponding to the ultrasonic wave detection 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;

[0082] The three-dimensional point cloud standardized data subset is processed by the analysis sub-model corresponding to the three-dimensional point cloud detection data source, and the surface curvature distribution, normal vector change rate, pit index, step height, geometric residual, size deviation and corresponding confidence information are output;

[0083] The spectral reflection standardized data subset is processed by the analysis sub-model corresponding to the spectral reflection detection data source, and the characteristic waveband reflectivity, waveband ratio, material composition index, oxidation degree index, surface pollution index and corresponding confidence information are output;

[0084] The quality characteristic output and confidence information of the visible light image detection data source, the quality characteristic output and confidence information of the thermal imaging detection data source, the quality characteristic output and confidence information of the acoustic detection data source, the quality characteristic output and confidence information of the ultrasonic wave detection data source, the quality characteristic output and confidence information of the three-dimensional point cloud detection data source, and the quality characteristic output and confidence information of the spectral reflection detection data source are merged according to the detection data source respectively. The source identifier, spatial position information and time identifier are reserved to form a source analysis result set.

[0085] In the embodiment of the present application, step S2 takes the standardized input data set as input, and calls the pre-established analysis sub-model according to the detection data source. The analysis sub-model has completed parameter training and fixing before deployment, and the output scale is unified through the calibration data in the batch. For the visible light image standardized data subset, 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 quantity composed of texture and gradient stability; the output provides pixel-level position evidence for the subsequent spatial consistency index. For the thermal imaging standardized data subset, the analysis sub-model segments the thermal anomaly area mask on the temperature field, extracts the temperature extreme value, temperature difference and thermal gradient, and forms the confidence information with the time period stability of the abnormal area and the background contrast; the output is used to determine the severity of the residual stress or local overheating after processing. For the acoustic detection standardized data subset, the analysis sub-model identifies the frequency band structure related to cracks or debonding on the time-frequency energy spectrum, quantifies the resonance peak frequency, resonance peak bandwidth, energy ratio and envelope amplitude, and gives confidence information through the spectral peak clarity and channel consistency; the output has a supplementary effect on invisible subsurface defects. For the ultrasonic wave detection standardized data subset, the analysis sub-model analyzes the echo amplitude curve and time of flight on the time axis, calculates the reflection coefficient and attenuation coefficient, combines the time of flight to get the defect depth estimate, and forms the interface continuity score and confidence information with the echo form integrity; the output directly provides the depth dimension, and provides a vertical reference for the severity estimation in the subsequent fusion. For the three-dimensional point cloud standardized data subset, the analysis sub-model calculates the surface curvature distribution, normal vector change rate, pit index, step height and geometric residual in the unified scale geometric field, generates confidence information combined with the consistency of local geometry; the output has high sensitivity to sand eye, strain and step-shaped abnormal shapes. For the spectral reflectance standardized data subset, the analysis sub-model infers the material composition index, oxidation degree index and surface contamination index from the feature band reflectivity and band ratio, and generates confidence information with the spectral shape fitting residual; the output can distinguish between material discoloration and optical artifacts of real defects.

[0086] The confidence information of each analysis sub-model is obtained according to the mapping function established by the model output probability and historical statistics, and is corrected once in the current batch with the calibration frame, so that the confidence information of different detection data sources can be compared. In order to ensure the traceability in space and time, each quality feature output and the corresponding confidence information carries the source identification, spatial position information and time identification. If a detection data source is marked as invalid in step S1 by the quality gating, the detection data source generates empty quality feature output and zero value confidence information in step S2, while retaining the source identification, spatial position information and time identification, without interrupting the subsequent process.

[0087] It should also be noted that when the sub-sources are combined into a sub-source analysis result set, the detection data sources are strictly grouped and stored, the quality characteristic output, confidence information, source identification, spatial position information and time identification are kept consistent index relationship, so that in the subsequent weighted fusion processing based on environmental state information to determine the weight value, different detection data sources can be implemented Differentiated weight allocation; At the same time, in the consistency index calculation, the defect area mask, defect depth estimation, thermal anomaly area mask, geometric residual and material composition index can be compared one by one under the same spatial position information and time identification, forming a quantifiable measurement of spatial consistency, feature consistency and time consistency. Through the above setting, step S2 not only provides discrimination information and reliability measurement of each detection data source, but also ensures the subsequent fusion and determination with explainability and verifiability through structured data organization, meeting the technical target of the present application.

[0088] S3: Obtain environmental state information, determine the weight value of each analysis sub-model based on the environmental state information, and perform weighted fusion processing on the sub-source analysis result set to obtain the fusion quality evaluation result and the overall confidence.

[0089] Obtain environmental state information and perform standardization processing and time synchronization processing to form an environmental state information vector;

[0090] According to the preset weight determination rule, calculate the weight value for each analysis sub-model according to the environmental state information vector;

[0091] Perform spatial alignment processing and scale alignment processing on the sub-source analysis result set under the unified spatial coordinate system to keep the source identification, spatial position information and time identification consistent;

[0092] According to the calculated weight value, perform weighted fusion processing on the sub-source analysis result set to obtain the fusion quality evaluation result;

[0093] According to the confidence information in the sub-source analysis result set and the corresponding weight value, perform weighted summation to obtain the overall confidence, and take the fusion quality evaluation result and the overall confidence as the input of the subsequent determination processing.

[0094] In the embodiment of the present application, step S3 takes the set of source analysis results and the environmental state information as input, first performs standardization processing and time synchronization processing on the environmental state information to form an environmental state information vector. The environmental state information vector at least contains illumination intensity, environmental temperature, environmental humidity, environmental vibration intensity, dust concentration in the air, environmental electromagnetic interference intensity, and conveying line running speed, and is aligned with the source identification, spatial position information, and time identification in the same period. Based on the environmental state information vector, the weight value of each analysis sub-model is calculated according to the preset weight determination rule. The weight determination rule can establish a clear mapping relationship in a lookup table manner or a piecewise function manner, for example, a low-grade weight value of the visible light channel is given in a low light and high dust scene, a low-grade weight value of the thermal imaging channel is given in a high temperature drift scene, and a low-grade weight value of the acoustic and ultrasonic channels is given in a high vibration scene. The weight value after normalization is used for subsequent fusion, and if a certain detection data source is marked as invalid in step S1, the weight value of the detection data source is set to zero, but the source identification is still retained for tracing.

[0095] Subsequently, the set of source analysis results is subjected to spatial position alignment processing and scale unification processing under a unified spatial coordinate system. The spatial position alignment processing maps the results to the unified resolution grid of the workpiece coordinate system for positioning-related quantities such as defect area mask, defect boundary coordinates, thermal anomaly area mask, defect depth estimation, and geometric residual. The scale unification processing unifies the dimensions of different channels, for example, the depth, temperature difference, curvature, and energy ratio are scaled according to the numerical domain of the standardized input data set to ensure that the outputs of different detection data sources are comparable on the same numerical scale.

[0096] In the weighted fusion processing, the fusion operation is performed on the multi-channel results of the same position and the same period indexed by the source identification, spatial position information, and time identification. The positioning type results obtain the fused defect probability map by weighted average in the probability domain, and then obtain the fused defect positioning result and the fused defect boundary coordinates by threshold segmentation. The severity type results use a weighted combination strategy, taking the defect depth estimation of ultrasonic waves and the geometric residual of three-dimensional point clouds as the dominant quantities, taking the temperature difference and thermal gradient of thermal imaging as auxiliary quantities, and taking the energy features and reflection coefficients of acoustic and ultrasonic as credibility correction quantities to obtain the fused defect severity result. To improve stability, morphological smoothing and connectivity consistency test in a small range are applied to the fusion results of adjacent positions to avoid misjudgment clusters formed by isolated noise points.

[0097] The overall confidence is obtained by weighting and aggregating the confidence information in the set of source analysis results and the corresponding weight values. To ensure cross-batch comparability, the calculation of the overall confidence uses normalized weights and calibrated confidence information, and records weight values, the number of detection data sources participating in fusion, quality gate pass rates, and other meta-information. If only part of the detection data sources participate in fusion at a certain position, the overall confidence automatically considers the reduction factor of the number of participants when aggregating, thereby avoiding false high conclusions due to incomplete information. The fusion quality evaluation results and the overall confidence serve as inputs for subsequent decision processing, providing a position-level, dimension-unified, and traceable evidence base for the combination of consistency indicators and decision rules.

[0098] It should also be noted that the spatial position alignment processing and scale unification processing of this step are for analysis output results, which are different from the spatial alignment processing and numerical mapping between sensors for standardized data in step S1. The weight determination rule can be obtained by 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 the above settings, the sensitivity of the weight value to environmental changes and the consistency of the fusion calculation are guaranteed at the same time, so that the fusion quality evaluation results and the overall confidence can remain stable and interpretable under complex working conditions.

[0099] S4: Calculate the consistency indicators between the set of source analysis results, and determine the fusion quality evaluation results according to the combination decision rule of the overall confidence and the consistency indicators, and output the quality evaluation conclusion.

[0100] Extracting source identifiers, spatial position information, time identifiers, and quality characteristics from the set of source analysis results, comparing in a unified spatial coordinate system and a unified time reference;

[0101] Calculating a spatial consistency indicator, which is calculated based on the overlap ratio between the defect positioning results of each detection data source;

[0102] Calculating a feature consistency indicator, which is calculated based on the similarity between the normalized feature vectors of the quality characteristic outputs of each detection data source, and the similarity is calculated using a preset similarity measure;

[0103] Calculating a time consistency indicator, which is calculated based on the consistency degree of the time identifiers of each detection data source within the same collection period;

[0104] Weighting the spatial consistency indicator, the feature consistency indicator, and the time consistency indicator according to a preset consistency weight to obtain a comprehensive consistency indicator;

[0105] The fusion quality evaluation result is determined according to a combination determination rule, and the combination determination rule comprises:

[0106] 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 evaluation result is greater than the defect determination threshold, the quality evaluation conclusion is output as a defect; if the defect severity result in the fusion quality evaluation result is less than or equal to the defect determination threshold, the quality evaluation conclusion is output as qualified.

[0107] 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 quality evaluation conclusion is output as to be reviewed.

[0108] For the case that does not satisfy the two conditions, the quality evaluation conclusion is output as to be reviewed; wherein the second confidence threshold is less than the first confidence threshold, and the second consistency threshold is less than the first consistency threshold.

[0109] In the embodiment of the application, step S4 takes the set of source analysis results, the fusion quality evaluation result and the overall confidence as input, and first compares the source identifier, the spatial position information, the time identifier and the quality feature output in each position under the unified spatial coordinate system and the unified time reference after 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 defect depth estimation of the ultrasonic wave and the energy accumulation area of the acoustics participate in the calculation of the overlap ratio by projecting onto the workpiece surface, and the geometric residual of the three-dimensional point cloud participates in the calculation of the overlap ratio after the threshold segmentation to form an abnormal area. The detection data sources that only provide area-level evidence are aligned at the area granularity, and the detection data sources that only provide point evidence are expanded in the neighborhood and participate in the calculation, so as to ensure that different evidence forms can be compared on the same grid. The feature consistency index is calculated based on the similarity of the normalized feature vector, and the normalized feature vector includes the dimensionally unified quantities after numerical mapping, such as defect depth estimation, temperature difference, thermal gradient, curvature, step height, energy ratio, reflection coefficient and material composition index; at any position, the detection data source not participating in the output is not included in the similarity calculation of the position and is excluded in the denominator, so as to avoid introducing zero value bias. The time consistency index is calculated according to the alignment degree of the time identifier of each detection data source in the same collection period, and the value is taken according to the deterministic function composed of the period coverage ratio and the maximum time deviation, so as to ensure that the interval collection channel and the synchronous collection channel can be compared under the unified rule.

[0110] The comprehensive consistency index is obtained by weighting the spatial consistency index, the feature consistency index and the time consistency index according to preset consistency weights, the preset consistency weights being determined offline according to historical data before system deployment and remaining unchanged within a batch. In order to suppress the influence of isolated noise points on the comprehensive consistency index, a small range of consistency smoothing and connectivity verification is applied to the comprehensive consistency index after position-level calculation; for scattered abnormalities that only occur in a small number of channels across detection data sources, the comprehensive consistency index is limited to a lower interval, so as not to trigger the direct determination path.

[0111] In the combined determination, the system first checks whether the overall confidence and the comprehensive consistency index reach the first confidence threshold and the first consistency threshold. When both reach, the system determines the quality evaluation conclusion to be defect or qualified according to the comparison between the defect severity result in the fusion quality evaluation result and the defect determination threshold; when the overall confidence is not high or the comprehensive consistency index is not high, the quality evaluation conclusion is directly output as pending review. The effect of the above rules is illustrated by typical cases: if the visible light image and the three-dimensional point cloud give the defect area mask and the geometric residual at the same position, but the thermal imaging and the acoustics do not see abnormalities, then the spatial consistency index and the feature consistency index can still remain in the medium-high interval; if the defect severity result in the fusion quality evaluation result exceeds the defect determination threshold, the quality evaluation conclusion is defect; if the overall confidence is insufficient due to fewer participating channels, the quality evaluation conclusion is pending review. For example, the ultrasonic wave gives an obvious defect depth estimate, while the visible light image has no surface signs, the calculation of the spatial consistency index is performed through projection and neighborhood expansion, and the evidence consistency is reflected by the depth, energy ratio and reflection coefficient in the feature consistency index; when the comprehensive consistency index and the overall confidence reach the threshold, stable determination of subsurface defects can still be achieved.

[0112] It should also be noted that the value ranges of the spatial consistency index, the feature consistency index and the time consistency index are unified to the numerical domain of zero to one, and the number of participating calculation detection data sources, the visible coverage area ratio and the maximum time deviation are recorded as meta information for subsequent review and version tracing. The comprehensive consistency index and the overall confidence are only used as inputs of the combined determination rule, and are not directly used as a substitute for the defect severity result; the defect severity result always comes from the fusion quality evaluation result, ensuring that the evidence source in the decision chain is clear and distinguishable. The equal determination in the boundary condition is explicitly defined as a closed interval condition greater than or equal to and less than or equal to in the combined determination rule, and the remaining cases are uniformly classified into the pending review path, avoiding threshold gray areas. Through the above settings, step S4 forms a complete and verifiable determination process in terms of evidence alignment, metric unification and conflict disposal, providing a basis for stable output of the quality evaluation conclusion.

[0113] S5: when the quality evaluation conclusion meets the preset review condition, triggering the review process, obtaining review label data, merging the review label data and historical data into an updated data set, and training and updating the analysis sub-model and the parameters for weighted fusion processing and decision processing based on the updated data set.

[0114] The fusion quality evaluation result, the overall confidence, the comprehensive consistency index, and the source identifier, the spatial position information, and the time identifier are pushed to the review terminal, and 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 reflection standardized data subset are called to perform review display;

[0115] The object to be reviewed is labeled according to the preset labeling specification, and review label data is output, the review label data including defect positioning results, defect severity results, and defect category information;

[0116] The visible light image standardized data subset and the corresponding review label data, the thermal imaging standardized data subset and the corresponding review label data, the acoustic detection standardized data subset and the corresponding review label data, the ultrasonic detection standardized data subset and the corresponding review label data, the three-dimensional point cloud standardized data subset and the corresponding review label data, and the spectral reflection standardized data subset and the corresponding review label data are paired and merged to form an updated data set;

[0117] 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 reflection detection data source are trained and updated to obtain updated analysis sub-model parameters;

[0118] Based on the updated data set, the parameters for weighted fusion processing and the parameters for decision processing are trained and updated, and the updated analysis sub-model parameters, the parameters for weighted fusion processing, and the parameters for decision processing are input into subsequent quality evaluation.

[0119] In the embodiment of the present application, step S5 takes the quality evaluation conclusion, the fused quality evaluation result, the overall confidence, the comprehensive consistency index, and the source identifier, the spatial position information, and the time identifier as inputs. When the quality evaluation conclusion is to be reviewed or meets the preset review condition, the system synchronously displays the visible light image standardized data subset, the thermal imaging standardized data subset, the acoustic detection standardized data subset, the ultrasonic wave detection standardized data subset, the three-dimensional point cloud standardized data subset, and the spectral reflection standardized data subset on the review terminal, and displays them in alignment with the same spatial position information and the same time identifier. The review terminal labels the position to be reviewed according to the preset labeling specification to form review label data, which at least contains the defect positioning result, the defect severity result, and the defect category information, and retains the source identifier, the spatial position information, and the time identifier to ensure traceability. To avoid bias, the system reviews 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 in addition to the samples to be reviewed, to improve the labeling quality and statistical sufficiency of the boundary area.

[0120] The system pairs the six types of standardized data subsets with the corresponding review label data one by one to generate an updated data set. The pairing rule is that the same workpiece number, the same source identifier, the same spatial position information, and the same time identifier must be completely consistent to be written, and the reviewer number and the labeling time when the review label data is generated are recorded synchronously. The updated data set performs consistency checking before being written, including spatial overlap verification of the defect positioning result and the three-dimensional point cloud geometric residual, range consistency verification of the defect severity result and the ultrasonic wave defect depth estimation, and correlation verification of the thermal imaging temperature difference and the acoustic energy ratio. If any of the checks fails, it is returned to the review terminal for re-examination.

[0121] The training update process is divided into two parts: analysis sub-model training update and parameter training update for weighted fusion processing and decision processing. The analysis sub-model training update adopts an incremental strategy, merges the updated data set with samples evenly extracted by category and working condition from the historical data to ensure stable category distribution and working condition distribution; uses the sealed baseline set as an invisible verification set to monitor four indicators: correct decision rate, false positive rate, false negative rate, and average processing time delay. If any of the indicators deteriorates, the update is stopped and rolled back to the previous version. The training update of the parameters for weighted fusion processing and the parameters for decision processing takes the updated data set as input, re-fits the mapping coefficients in the weight determination rule and the threshold values in the combined decision rule, and completes confidence calibration and consistency threshold calibration on the sealed baseline set, so that the corresponding relationship between the overall confidence and the comprehensive consistency index and the true label remains monotonic and interpretable. After the update is completed, the system archives the version number, the training time, the training data batch number, and the key indicators as meta-information, and goes online in a small flow verification mode. After the running indicators are stable, it is switched to full-scale application.

[0122] It should also be noted that the preset review condition of the review trigger includes that the quality evaluation conclusion is to be reviewed, and the sample with the overall confidence close to the second confidence threshold and the sample with the comprehensive consistency index close to the second consistency threshold; when a certain detection data source is marked as invalid in step S1, its weight value remains zero in the training update stage and does not participate in parameter learning to prevent noise from being introduced; each sample in the updated data set retains the source identification, spatial position information and time identification, so that any model output can be traced back to the original evidence and the review process in the future. Through the above process, the training update completes the parameter update without changing the data and result organization mode of steps S1 to S4, so as to ensure that the subsequent batches can obtain improved evaluation performance and stability under the same interface and index structure.

[0123] S6: dynamically adjusting the combined determination rule and the weight value according to the evaluation performance index, and applying the adjustment result to subsequent quality evaluation.

[0124] The evaluation performance index is counted in a preset length of time window, and the evaluation performance index includes a correct determination rate, a false positive rate, a false negative rate and an average processing delay;

[0125] The evaluation performance index is compared with the corresponding preset target value to obtain a comparison result;

[0126] According to the comparison result, the first confidence threshold, the second confidence threshold, the first consistency threshold and the second consistency threshold in the combined determination rule are updated according to the preset adjustment rule, and the weight value in the weight determination rule is updated;

[0127] The updated first confidence threshold, second confidence threshold, first consistency threshold, second consistency threshold and weight value are applied to subsequent quality evaluation.

[0128] In the embodiment of the application, step S6 is based on online samples and review label data in a preset length of time window to perform hierarchical statistics and comparison on the evaluation performance index. The evaluation performance index includes a correct determination rate, a false positive rate, a false negative rate and an average processing delay, which are counted in product categories, surface treatment processes and environmental state layers to avoid structural bias. Each index is compared with the corresponding preset target value to obtain a comparison result. The system performs two-level dynamic adjustment accordingly: a fast ring for the weight value, and a slow ring for the first confidence threshold, the second confidence threshold, the first consistency threshold and the second consistency threshold in the combined determination rule. The fast ring is executed once every time window, and the slow ring is triggered when a plurality of time windows are deviated from the target in succession to suppress oscillation.

[0129] In the fast loop, the source contribution and the source error rate are calculated for each detection data source. The source contribution is defined as the increment of the correct decision rate of the fusion quality evaluation result when the source participates, and the source error rate is defined as the weighted increment of the false positive rate and the false negative rate when the source participates relative to when the source does not participate. According to the difference between the two, the weight value of the source is updated in proportion, and after the update, normalization is performed and projected to the preset upper and lower limits to ensure that the sum of the weight values and the limit constraints meet the requirements; when a detection data source is quality-gated as invalid 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 term is applied to the detection data source with a large delay contribution, so that its weight value decreases slightly, thereby pressing the overall delay back to the target range while ensuring that the recall rate does not decrease significantly.

[0130] In the slow loop, the threshold is adjusted according to the deviation direction of the false positive rate and the false negative rate: when the false positive rate is higher than the target and the false negative rate is not higher than the target, the first confidence threshold or the first consistency threshold is increased to tighten the direct decision entry; when the false negative rate is higher than the target and the false positive rate is not higher than the target, the first confidence threshold or the first consistency threshold is decreased to increase the sensitivity; when both are higher than the target, the first confidence threshold and the first consistency threshold are adjusted slightly in the same direction after two rounds of weight value reorganization are completed by the fast loop; the adjustment of the second confidence threshold and the second consistency threshold follows the boundary rollback strategy in the opposite direction, which is used to control the upper and lower limits of the proportion to be reviewed. Each threshold adjustment sets a maximum change and a hysteresis band, and only when the deviation direction is consistent for multiple consecutive time windows and the sample size reaches the preset lower limit, the adjustment takes effect, preventing frequent switching caused by small sample noise. The defect judgment threshold is determined by the training update process in step S5, and step S6 does not adjust it, ensuring clear responsibilities in the decision chain.

[0131] To ensure traceability, the system records the window number, sample size, hierarchical dimension, threshold before and after adjustment, weight value before and after adjustment, evaluation performance indicators and target values, and verification indicators of the next time window after going online. If the verification indicators deteriorate beyond the preset tolerance, the system automatically rolls back to the combined decision rule and weight value of the previous version, and marks the adjustment as invalid. Through the above settings, step S6 realizes the bounded, hierarchical, and rollbackable closed-loop self-adaptation of the combined decision rule and weight value without changing the data structure and interface, thereby maintaining the stability and interpretability of the fusion quality evaluation result and the overall confidence in complex and changing working conditions, and seamlessly applying the adjustment results to subsequent quality evaluation.

[0132] Embodiment 2, which is different from the previous embodiment, is a second embodiment of the present application.

[0133] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the current technical solutions can be embodied in the form of software products. The current computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0134] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered a list of executable instructions for implementing logic functions, and can be specifically embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, a system that includes a processor, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device or in conjunction with these instruction execution systems, apparatuses, or devices.

[0135] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection having one or more wires (electrical devices), a portable computer diskette (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CD ROM). In addition, the computer-readable medium can even be paper or other suitable medium on which the program can be printed, because the program can be electronically obtained, for example, by optical scanning of the paper or other medium, followed by editing, interpretation, or necessary processing, if necessary, in other suitable ways, and then stored in a computer memory.

[0136] Embodiment 3, which is an embodiment of the present application, provides a casting surface treatment defect detection and quality evaluation system, which includes a data acquisition and preprocessing module, a sub-model analysis module, a fusion and confidence calculation module, a consistency calculation and combination determination module, a review and training update module, and a dynamic adjustment module.

[0137] The data acquisition and preprocessing module acquires multi-source surface detection data of the castings to be detected and performs data preprocessing to obtain a standardized input data set, wherein the standardized input data set is composed of standardized data subsets corresponding to each detection data source;

[0138] The sub-model analysis module calls an analysis sub-model corresponding to each detection data source for processing to obtain quality feature output and confidence information of the detection data source, and combines the quality feature output and the confidence information into a set of source-by-source analysis results;

[0139] The fusion and confidence calculation module acquires environmental state information, determines weight values of the analysis sub-models based on the environmental state information, and performs weighted fusion processing on the set of source-by-source analysis results to obtain a fusion quality evaluation result and an overall confidence;

[0140] The consistency calculation and combination determination module calculates a consistency index between the set of source-by-source analysis results, determines the fusion quality evaluation result according to a combination determination rule of the overall confidence and the consistency index, and outputs a quality evaluation conclusion;

[0141] The review and training update module triggers a review process when the quality evaluation conclusion meets a preset review condition, obtains review label data, combines the review label data and historical data into an update data set, and performs training update on the analysis sub-models and parameters used for the weighted fusion processing and the determination processing based on the update data set;

[0142] The dynamic adjustment module dynamically adjusts the combination determination rule and the weight values according to an evaluation performance index, and applies the adjusted results to subsequent quality evaluation.

[0143] Embodiment 4 is an embodiment of the present application, which provides a castings surface treatment defect detection and quality evaluation method. In order to verify the beneficial effects of the present application, economic benefit calculation and simulation / contrast experiments are used for scientific demonstration.

[0144] Four types of common castings were selected as test objects: valve body (after shot blasting), pump shell (after sand blasting), shell (after shot blasting), and bracket (after wire drawing), with 30 pieces of each type, totaling 120 pieces. A fixed multi-source acquisition unit was set up at the first detection station after surface treatment, including visible light cameras, infrared thermal imagers, pickup sensor arrays, ultrasonic roller probes, structured light three-dimensional sensors, and multispectral imaging devices. The six types of detection data sources share a unified trigger and a unified timestamp for time synchronization, and the calibration pieces are used to solve the external parameters of each sensor to the workpiece coordinate system, achieving spatial alignment between sensors. The acquisition beat is set according to the production line conveying speed, and the six-source data acquisition of a single piece is completed in one pass at the same station. After acquisition is completed, immediate preprocessing is performed: visible light channels are corrected for geometric distortion, illumination is balanced, highlights are suppressed, and noise is removed, and blurred frames are removed through clarity detection; the thermal image channel implements 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 completes amplitude normalization and channel synchronization calibration; the ultrasonic channel performs time reference calibration, echo denoising, coupling compensation, envelope extraction, time-of-flight alignment, and amplitude and phase correction; the three-dimensional point cloud channel performs outlier rejection, registration, normal estimation, and resampling, and is spatially aligned with the visible light coordinate; the spectral channel completes dark current subtraction, whiteboard and band response correction, and illumination drift compensation. Then, quality gating (signal-to-noise ratio and saturation threshold) and abnormal sample rejection are performed on each channel, and the data that pass the gating are mapped to a unified numerical domain [0, 1], forming six standardized data subsets according to the source and merging them into a standardized input data set.

[0145] For each standardized data subset, call the pre-established and corresponding to the detection data source analysis sub-model: visible light output defect mask and boundary and roughness index; thermal image output thermal anomaly mask, temperature difference and thermal gradient; acoustic output time-frequency energy, resonance peak, energy ratio and anomaly score; ultrasonic output echo amplitude, time of flight, reflection coefficient, attenuation coefficient, defect depth estimation and interface continuity score; three-dimensional point cloud output curvature, step height, geometric residual, etc.; spectral output characteristic band reflectivity, band ratio, material composition and oxidation pollution index. All quality feature outputs and confidence information are merged into a source-specific analysis result set according to source identification, spatial location information, and time identification. Record the environmental state information (light, temperature and humidity, vibration, dust, electromagnetic interference, conveying speed) and standardize it into an environmental state information vector; calculate the weight value of each analysis sub-model according to the pre-set weight determination rule. After spatial position alignment and scale unification of the source-specific analysis result set in the unified spatial coordinate system, weighted fusion is performed to obtain the fusion defect positioning and defect severity result, and the overall confidence is summarized according to the confidence and weight value. Then calculate the spatial consistency, feature consistency and time consistency and weight them as the comprehensive consistency index, and output the quality evaluation conclusion according to the combination judgment rule. For the conclusion of "to be reviewed" and the boundary sample, the review terminal displays the six-source evidence in the same space-time, generates a review label data according to the specification, and merges it with the historical data to form an updated data set for incremental training of the analysis sub-model and recalibration of the weight and judgment threshold. The evaluation performance indicators (correct decision rate, false alarm rate, false alarm rate, average processing delay) are calculated in a sliding window, and the scenes deviating from the target are executed with hierarchical rollback dynamic adjustment of fast weight and slow threshold, and the updated results are directly used for subsequent batches. For specific data content, refer to Table 1.

[0146] Table 1 Data Reference Table

[0147]

[0148] From Table 1, it can be seen that the "defect recall rate (%)" and "F1 score (%)" in the table show a trend that the present application is significantly better than the baseline scheme on the four types of castings: for example, the F1 score of "shell-after shot blasting" is increased from 90.2% to 96.6%, and the recall rate of "valve body-after shot blasting" is increased from 88.3% to 95.6%. This improvement cannot be obtained by single-channel threshold optimization, and it is derived from the joint action of weight value self-adaptation driven by environmental state information and source consistency constraint. Specifically, the "consistency index [0-1]" is increased by about 0.12-0.14 on all objects (e.g. 0.74→0.88 for "valve body-after shot blasting"), indicating that the mutual confirmation degree of multi-source evidence under unified space-time and scale is higher, and the judgment swing caused by multi-source conflict is reduced. Correspondingly, the "false alarm rate (%)" is reduced from 6.7-8.6% to 3.8-5.1%, which shows that when the spatial consistency is insufficient or the feature consistency is insufficient, the combined decision rule effectively blocks the wrong direct decision path and guides it to "to be reviewed", thereby reducing the false alarm.

[0149] The improvement of "overall confidence [0-1]" (e.g. 0.82→0.93 for "shell-after shot blasting") proves that the global reliability after joint aggregation of weight value and confidence has a stronger monotonic alignment relationship with the true label, and this improvement directly supports the significant decrease of "to-be-reviewed ratio (%)" (e.g. 14.6%→7.1% for "pump shell-after shot blasting"). The decrease of to-be-reviewed ratio and the increase of F1 score show that it is not simply to relax the threshold, but to improve the evidence quality and fusion reliability through the consistency index. It is worth noting that the "subsurface defect detection rate (%)" is greatly improved on all objects, such as 73.2%→90.1% for "shell-after shot blasting", which shows that the present application can stably identify defects with obvious surface signs but significant deep abnormalities by taking the ultrasonic depth information and acoustic energy ratio as the dominant quantity for severity estimation, supplemented by three-dimensional geometric residual and thermal gradient as auxiliary quantities, which is an effect that cannot be achieved by a single modality or a fixed fusion strategy.

[0150] From the engineering cost, the "average processing delay (ms)" of the present application has an acceptable increase (about 12-15 ms), but it is exchanged for about 6-8 percentage points of F1 improvement and about half of the decrease of to-be-reviewed ratio, which has a significant comprehensive benefit and still meets the online requirements within the production line beat. More representative evidence comes from the "improved AF1 (percentage points) after review" column: under the closed-loop review and updated dataset driving, the present application continues to obtain an F1 gain of 1.9-2.6 percentage points after the initial online, which shows that the parameters and sub-models can be continuously calibrated according to the true scene distribution, which is particularly critical for the "concept drift" scenario, and the traditional fixed parameter scheme cannot obtain sustained gains.

[0151] Compared with the above, the creativity of the present application is embodied in the following aspects: first, the environmental state information vector is used to implement differentiated weight values for each analysis sub-model, solving the problem of uneven contribution and non-equivalent interference of multiple sources under complex working conditions; second, the three-dimensional consistency index of space-feature-time at the position level is introduced to form an interpretable shunt of "direct judgment / pending review" by combining the judgment rules, avoiding the threshold gray area; third, a closed-loop mechanism of "review-update dataset-incremental training-dynamic adjustment" is constructed, so that the evaluation performance can still evolve after going online. The systematic improvement of each index in the table has a clear causal correspondence with the above three technical features, proving that the present application is not only superior in recognition accuracy, but also has outstanding technical effects and application value in robustness, interpretability and sustainable optimization capability.

[0152] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application 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 application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A method for detecting and evaluating the quality of surface treatment defects in castings, characterized in that, include: Multi-source surface inspection data of the casting to be inspected are collected and preprocessed to obtain a standardized input dataset, which consists of standardized data subsets corresponding to each source of inspection data. For each standardized data subset, a pre-established analysis sub-model corresponding to the source of the detection data is called for processing to obtain the quality characteristic output and confidence information of the source of the detection data, which are then merged into a set of source-specific analysis results. Obtain environmental status information, determine the weight values ​​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 level; Calculate the consistency index among the source analysis results set, judge the fusion quality assessment results according to the combination judgment rule of 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 review label data. The review label data is then merged with historical data to form an updated dataset. Based on the updated dataset, the analysis sub-model and the parameters used for weighted fusion processing and judgment processing are trained and updated. The combination judgment rules and weight values ​​are dynamically adjusted based on the performance evaluation indicators, and the adjustment results are applied to subsequent quality evaluations. The process of obtaining the fusion quality assessment results and overall confidence includes acquiring environmental state information and performing standardized and time-synchronized processing to form an environmental state information vector. Based on the preset weight determination rules, weight values ​​are calculated for each analysis sub-model according to the environmental state information vector; Under a unified spatial coordinate system, the source analysis result set is spatially and scale-aligned to maintain consistency in source identifiers, spatial location information, and time identifiers. The source analysis results are weighted and fused according to the calculated weight values ​​to obtain the fusion quality assessment results. The confidence level is obtained by weighting and summarizing the confidence information and corresponding weight values ​​in the source analysis result set, and the fusion quality assessment result and the overall confidence level are used as inputs for subsequent judgment processing. The output quality assessment conclusions include extracting source identifiers, spatial location information, time identifiers, and quality characteristic outputs from the source analysis result set, and comparing them under a unified spatial coordinate system and a unified time reference. Calculate the spatial consistency index, which is based on the overlap ratio between the defect location results from each detection data source; The feature consistency index is calculated based on the similarity between the normalized feature vectors output by the quality features of each detection data source. The similarity is calculated using a preset similarity metric. Calculate the time consistency index, which is based on the degree of consistency of the time stamps of each detection data source within the same collection period. The spatial consistency index, feature consistency index, and temporal consistency index are weighted according to the preset consistency weights to obtain the comprehensive consistency index. The fusion quality assessment results are determined according to the combination determination rules, which include: When the overall confidence level is greater than or equal to the first confidence level threshold and the overall consistency index is greater than or equal to the first consistency threshold, if the severity of the defect in the fusion quality assessment result is greater than the defect judgment threshold, the output quality assessment conclusion is defect; if the severity of the defect 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 level 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 "to be reviewed"; whereby the second confidence threshold is less than the first confidence threshold, and the second consistency threshold is less than the first consistency threshold.

2. The method for detecting and evaluating the surface treatment defects of castings as described in claim 1, characterized in that, The multi-source surface detection data of the casting to be tested includes visible light image data, thermal imaging data, acoustic detection data, ultrasonic detection data, three-dimensional point cloud data, and spectral reflectance data.

3. The method for detecting and evaluating surface treatment defects in castings as described in claim 2, characterized in that, The data preprocessing includes performing geometric distortion correction, illuminance equalization, highlight suppression, noise reduction, sharpness detection and blur removal, geometric calibration and color calibration on the visible light image data to obtain standardized visible light image data. The thermal imaging data is subjected to non-uniformity correction, temperature calibration, emissivity setting and reflection background temperature compensation, thermal drift compensation, and temperature field smoothing to obtain standardized thermal imaging data. The acoustic test data is processed by removing DC components, bandpass filtering, frame windowing, short-time Fourier transform, amplitude normalization, and channel synchronization calibration to obtain standardized acoustic test data. The ultrasonic testing data is subjected to time reference calibration, pulse echo denoising, coupling condition compensation, envelope extraction, time-of-flight alignment, and amplitude and phase correction to obtain standardized ultrasonic testing data. The 3D point cloud data is processed by outlier removal, noise reduction, point cloud registration, normal vector estimation, resampling, and scale normalization, and then spatially aligned with the coordinates of the visible light image to obtain standardized 3D point cloud data. Dark current subtraction, whiteboard correction, band response correction, reflectance normalization, and illumination drift compensation are performed on the spectral reflectance data to obtain standardized spectral reflectance data. The obtained standardized data are subjected to unified time alignment and unified spatial alignment processing, signal-to-noise ratio threshold detection, saturation threshold detection and abnormal sample removal processing, and each standardized data is mapped to a predetermined numerical range. According to the source of the detection data, the data is divided into a subset of standardized data for visible light images, a subset of standardized data for thermal imaging, a subset of standardized data for acoustic detection, a subset of standardized data for ultrasonic detection, a subset of standardized data for 3D point clouds, and a subset of standardized data for spectral reflectance, and then merged to form a standardized input dataset.

4. The method for detecting and evaluating the surface treatment defects of castings as described in claim 3, characterized in that, The quality feature output and confidence information of the obtained detection data source include processing the standardized subset of visible light image data by an analysis sub-model corresponding to the visible light image detection data source, and outputting defect region mask, defect boundary coordinates, defect area, defect length, defect direction, surface roughness index and corresponding confidence information. The standardized subset of thermal imaging data is processed by the analysis sub-model corresponding to the source of thermal imaging detection data, and outputs thermal anomaly region mask, temperature extreme value, temperature difference, thermal gradient, thermal anomaly duration and corresponding confidence information. The standardized subset of acoustic testing data is processed by the analysis sub-model corresponding to the source of the acoustic testing data, and outputs time-frequency energy spectrum characteristics, formant frequency, formant bandwidth, energy ratio, envelope amplitude, acoustic anomaly score and corresponding confidence information; The standardized subset of ultrasonic testing data is processed by the analysis sub-model corresponding to the source of the ultrasonic testing data, and outputs echo amplitude curve, time of flight, reflection coefficient, attenuation coefficient, defect depth estimation, interface continuity score and corresponding confidence information. The standardized subset of 3D point cloud data is processed by the analysis sub-model corresponding to the source of 3D point cloud detection data, and outputs surface curvature distribution, normal vector change rate, pitting index, step height, geometric residual, dimensional deviation and corresponding confidence information. The standardized subset of spectral reflectance data is processed by the analysis sub-model corresponding to the source of the spectral reflectance detection data, and outputs characteristic band reflectance, band ratio, material composition index, oxidation degree index, surface contamination index and corresponding confidence information; The quality feature outputs and confidence information from visible light image detection data sources, thermal imaging detection data sources, acoustic detection data sources, ultrasonic detection data sources, 3D point cloud detection data sources, and spectral reflectance detection data sources are merged according to the data source, retaining the source identifier, spatial location information, and time identifier respectively, to form a set of source-specific analysis results.

5. The method for detecting and evaluating surface treatment defects in castings as described in claim 4, characterized in that, The process of training and updating the analysis sub-model and the parameters used for weighted fusion processing and judgment processing based on the updated dataset includes pushing the fusion quality assessment results, overall confidence, comprehensive consistency index, source identifier, spatial location information and time identifier to the review terminal, and calling the standardized subsets of visible light image data, thermal imaging data, acoustic detection data, ultrasonic detection data, three-dimensional point cloud data, and spectral reflectance data for review and display. The objects to be reviewed are labeled according to the preset labeling specifications, and the review label data is output. The review label data includes defect location results, defect severity results, and defect category information. The standardized subsets of visible light image data, thermal imaging data, acoustic detection data, ultrasonic detection data, 3D point cloud data, and spectral reflectance data were paired and merged to form an updated dataset. Based on the updated dataset, the analysis sub-models corresponding to the visible light image detection data source, the thermal imaging detection data source, the acoustic detection data source, the ultrasonic detection data source, the 3D point cloud detection data source, and the spectral reflectance detection data source are trained and updated to obtain the updated analysis sub-model parameters. The parameters used for weighted fusion processing and decision processing are trained and updated based on the updated dataset. The updated sub-model parameters, the parameters used for weighted fusion processing, and the parameters used for decision processing are then used in subsequent quality assessments.

6. The method for detecting and evaluating the surface treatment defects of castings as described in claim 5, characterized in that, The dynamic adjustment of the combined judgment rules and weights based on the evaluation performance indicators includes statistically analyzing the evaluation performance indicators within a preset time window. The evaluation performance indicators include the correct judgment rate, false alarm rate, false negative rate, and average processing latency. The performance indicators are compared with the corresponding preset target values ​​to obtain the comparison results; Based on the comparison results, the first confidence threshold, second confidence threshold, first consistency threshold, and second consistency threshold in the combination judgment rule are updated according to the preset adjustment rules, and the weight values ​​in the weight determination rule are also updated. The updated first confidence threshold, second confidence threshold, first consistency threshold, second consistency threshold, and weight values ​​will be applied to subsequent quality assessments.

7. A casting surface treatment defect detection and quality assessment system, used to implement the casting surface treatment defect detection and quality assessment method as described in any one of claims 1 to 6, characterized in that, include: Data acquisition and preprocessing module: Acquires multi-source surface inspection data of the casting to be inspected and performs data preprocessing to obtain a standardized input dataset, which consists of standardized data subsets corresponding to each inspection data source; Sub-model analysis module: For each standardized data subset, it calls the pre-established analysis sub-model corresponding to the source of the detection data for processing, and obtains the quality characteristic output and confidence information of the source of the detection data, and merges them into a source-specific analysis result set; Fusion and confidence calculation module: acquires environmental state information, determines the weight values ​​of each analysis sub-model based on the environmental state information, performs weighted fusion processing on the source analysis result set, and obtains the fusion quality assessment result and overall confidence score; Consistency Calculation and Combination Judgment Module: Calculates the consistency index among the sets of source analysis results, judges the fusion quality assessment results according to the combination judgment rules of the overall confidence and 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 to obtain review label data, merge the review label data with historical data into an updated dataset, and train and update the analysis sub-model and the parameters used for weighted fusion processing and judgment processing based on the updated dataset. Dynamic adjustment module: Based on the performance evaluation indicators, the combination judgment rules and weight values ​​are dynamically adjusted, and the adjustment results are applied to subsequent quality evaluation.

Images