Intelligent diagnosis method for steel product quality defects based on multi-source data fusion

Abstract

This invention relates to the field of quality diagnosis technology and provides an intelligent diagnostic method for steel product quality defects based on multi-source data fusion. The method includes the following steps: constructing a unified physical meaning space containing multi-layer effect labels defined based on metallurgical physics principles; mapping multi-source heterogeneous features from surface images, time-series process parameters, spectral data, and infrared thermography to corresponding layers under the multi-layer effect labels according to their physical meaning, and performing standardization transformation to achieve feature alignment; performing online verification of the aligned multi-source features based on a preset process knowledge rule base, detecting and marking conflicts and their levels between features; and initiating a corresponding arbitration decision mechanism based on the conflict level, fusing multi-source evidence, and outputting defect diagnosis results and confidence levels. By mapping multi-source heterogeneous features to a unified semantic space constructed based on metallurgical physics principles, deep alignment of features in physical meaning is achieved, improving the physical interpretability of the diagnostic process.

Description

Technical Field

[0001] This invention relates to the field of quality diagnosis technology, specifically to an intelligent diagnostic method for quality defects in steel products based on multi-source data fusion. Background Technology

[0002] In the steel production process, accurate diagnosis of quality defects is crucial for improving product quality and reducing scrap rates. Modern production lines are equipped with various sensors and detection devices (such as high-definition cameras, spectrometers, infrared thermal imagers, and process parameter sensors) that can collect multi-source heterogeneous data reflecting the product status in real time, including surface images, time-series process parameters, spectral data, and thermal imaging data, providing a data foundation for intelligent diagnosis.

[0003] Currently, defect diagnosis methods based on multi-source data fusion typically employ direct feature vector concatenation or simple data-layer / feature-layer fusion strategies. These methods suffer from the following main shortcomings: First, data from different sources differ significantly in scale, dimension, and physical meaning, making it difficult to achieve true semantic alignment through simple concatenation, which can easily lead to information confusion and difficulties in model learning. Second, during the acquisition, transmission, or interpretation of multi-source data, contradictory conclusions may arise due to equipment errors, operating condition fluctuations, or model limitations. Existing methods lack effective online conflict detection and arbitration mechanisms, potentially resulting in unreliable diagnostic results. Therefore, a multi-source data fusion-based intelligent diagnostic method for steel product quality defects is needed to address these issues. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of this invention is to provide an intelligent diagnostic method for quality defects in steel products based on multi-source data fusion, so as to solve the problems existing in the above-mentioned background technology.

[0005] This invention is implemented as follows: a multi-source data fusion-based intelligent diagnostic method for quality defects in steel products, the method comprising the following steps: Construct a unified physical meaning space, which includes multi-layer effect labels defined based on metallurgical physics principles; Multi-source heterogeneous features from surface images, process parameter timing, spectral data, and infrared thermography are mapped to the corresponding layers under the multi-layer effect label according to their physical meaning, and then standardized and transformed to achieve feature alignment. Based on a pre-defined process knowledge rule base, the aligned multi-source features are validated online to detect and mark conflicts and their levels between features. Based on the level of conflict, the corresponding arbitration decision-making mechanism is initiated, multi-source evidence is integrated, and defect diagnosis results and confidence levels are output. Diagnostic cases are collected and retrospectively analyzed. Based on the performance evaluation results, the process knowledge rule base and feature mapping relationship are dynamically optimized.

[0006] As a further aspect of the present invention, the multilayer effect label includes a thermodynamic effect layer, a mechanical effect layer, and a metallurgical effect layer; the standardization conversion includes converting the original feature into a dimensionless or hierarchical index within its respective effect layer.

[0007] As a further aspect of the present invention, the process knowledge rule base includes at least one of temperature-structure correspondence rules, rolling force-surface morphology association rules, composition-performance matching rules, and spatiotemporal continuity verification rules; the conflict levels are divided into level one, level two, and level three.

[0008] As a further aspect of the present invention, when a first-level conflict is detected, a manual intervention process is triggered; when a second-level conflict is detected, a complete chain of evidence arbitration process is triggered; and when a third-level conflict is detected, a rapid weighted arbitration process based on data source weights is triggered.

[0009] As a further aspect of the present invention, the complete chain of evidence arbitration process includes: gathering characterization evidence directly related to the defect, tracing indirect process evidence of the production process, matching causal physical mechanism evidence to explain the defect, and performing confidence fusion calculation on the three types of evidence.

[0010] As a further aspect of the present invention, the corresponding arbitration decision mechanism routes according to a preset decision tree during execution: it receives conflict level markers and routes to the processing paths of manual intervention, full arbitration, or rapid weighted arbitration, respectively, based on whether the marker is level one, level two, or level three.

[0011] As a further aspect of this invention, the effectiveness evaluation of each rule in the rule base entry adopts a quantitative formula: effectiveness index E=(Nc-α×Nf-β×Nm) / Nt, where E is the effectiveness index, Nc is the number of correct warnings, Nf is the number of false alarms, Nm is the number of missed warnings, α and β are penalty coefficients, and Nt is the total number of rule triggers; the weight or trigger threshold of the rule is adjusted according to the effectiveness index E.

[0012] As a further aspect of the present invention, when mapping multi-source heterogeneous features to corresponding layers, different initial confidence weights are assigned to features from different data sources within the same effect layer.

[0013] As a further aspect of the present invention, the confidence fusion calculation adopts a weighted summation model: final confidence = W1 × direct evidence credibility + W2 × indirect evidence consistency index + W3 × highest physical mechanism conformity, wherein the weight coefficients W1, W2, and W3 are configured according to the conflict level and defect type.

[0014] As a further aspect of the present invention, in the dynamic optimization, the clustering density of similar defects in the physical meaning space is evaluated periodically using a standard sample set of known defect categories, and the transformation parameters or weight allocation of the feature mapping are adjusted accordingly to optimize the alignment effect.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By mapping multi-source heterogeneous features to a unified semantic space constructed based on metallurgical physics principles, deep alignment of features in a physical sense is achieved. This overcomes the drawbacks of traditional simple feature concatenation, enabling subsequent analysis to be based on consistent physical quantities, significantly improving the physical interpretability of the diagnostic process, and laying a solid foundation for generating reliable conclusions.

[0016] The fused features are validated online using a process knowledge rule base to detect logical conflicts or numerical anomalies between different data sources and to classify and label them, thereby improving the robustness of diagnostic decisions. Based on different levels of conflict, a differentiated arbitration decision mechanism is triggered, balancing the rigor and efficiency of the diagnosis and achieving intelligent decision-making. Attached Figure Description

[0017] Figure 1 A flowchart for an intelligent diagnostic method for quality defects in steel products based on multi-source data fusion.

[0018] Figure 2 This is a structural block diagram of an intelligent diagnostic system for quality defects in steel products based on multi-source data fusion. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0021] like Figure 1 As shown in the figure, this embodiment of the invention provides an intelligent diagnostic method for quality defects in steel products based on multi-source data fusion. The method includes the following steps: S100, Construct a unified physical meaning space, which contains multi-layer effect labels defined based on metallurgical physical principles; S200 maps the multi-source heterogeneous features from surface images, process parameter timing, spectral data, and infrared thermography to the corresponding layer under the multi-layer effect label according to their physical meaning, and performs standardization transformation to achieve feature alignment. In step S200, the multi-source heterogeneous features from surface images, process parameter timing, spectral data, and infrared thermography are mapped to the corresponding layers under the multilayer effect label according to their physical meaning, and then standardized to achieve feature alignment. The specific steps are as follows: S201, based on the principles of metallurgical physics, divides the unified physical meaning space into a thermodynamic effect layer, a mechanical effect layer, and a metallurgical effect layer to obtain a structured framework; among them, the thermodynamic effect layer is used to characterize the temperature history and energy state during the production process; the mechanical effect layer is used to characterize the stress, strain, and deformation behavior of the material; and the metallurgical effect layer is used to characterize changes in chemical composition and the resulting microstructure evolution. S202, the texture feature values ​​and geometric features obtained from the surface image are mapped to the mechanical effect layer of the structured framework; the cooling rate data extracted from the process parameter time series are mapped to the thermodynamic effect layer of the structured framework; the spectral peak features obtained from the spectral data are mapped to the metallurgical effect layer of the structured framework; the temperature distribution matrix extracted from the infrared thermogram is mapped to the thermodynamic effect layer of the structured framework; so as to obtain the feature set for the initial classification. S203, For continuous numerical features in the feature set of the initial classification, the linear normalization method of (current value - historical minimum value) / (historical maximum value - historical minimum value) is used for processing. At the same time, for state features in the feature set of the initial classification, they are converted into a finite number of levels of hierarchical index according to a preset threshold to obtain a standardized feature set. S204 Calculate the standard deviation of all pixel values ​​in the infrared thermal image temperature matrix in the standardized feature set, and then divide the standard deviation by the preset maximum allowable temperature standard deviation threshold to obtain the dimensionless temperature index. S205, compare the cooling rate data in the standardized feature set with the preset cooling rate classification threshold to obtain the cooling rate level identifier; wherein, the cooling rate classification threshold has three cooling rate ranges. S206, calculate the contrast and uniformity feature values ​​of texture feature values ​​using the gray-level co-occurrence matrix algorithm, process the ratio of contrast to uniformity feature values ​​according to the linear normalization method in step S203 to obtain the dimensionless index of surface roughness; compare the spectral peak features with the preset standard curve to obtain the elemental content grading index. S207, the dimensionless temperature index, cooling rate level identifier, surface roughness dimensionless index, and element content classification index are organized and aggregated according to the effect layer to complete feature alignment.

[0022] Furthermore, this step maps multi-source heterogeneous data to a thermodynamic, mechanical, and metallurgical effect layer constructed based on metallurgical physics principles, and performs standardization and indexation transformation, achieving unified alignment of multi-source features in a physical sense. This effectively solves the problems of mixed data semantics and inconsistent dimensions in traditional methods, improves the physical interpretability and feature comparability of defect diagnosis, and lays a reliable data foundation for subsequent intelligent analysis.

[0023] It should be noted that the "organization and aggregation" step in this process includes the following three specific sub-operations: Sub-operation 1: The dimensionless temperature index, cooling rate level identifier, surface roughness dimensionless index, and element content classification index generated in steps S204 to S206 are respectively assigned to their respective effect layers according to the mapping relationship defined in steps S201 to S202. Significance: It unifies indicators from different physical sources under the same cognitive framework of "thermal-mechanical-metallurgical", establishing a physical logical thread for subsequent analysis; Sub-operation 2: Combine the features of each effect layer into a multi-dimensional feature vector in a preset, fixed order; Significance: It completes the final transformation from diverse and heterogeneous data to a single, regular numerical vector, that is, the physical realization of "feature alignment"; this vector can be directly input into intelligent diagnostic models such as classification and regression. Sub-operation 3: Completion marker of feature alignment. The completion of the "sorting and converging" step signifies the achievement of "feature alignment," and its marker is as follows: Physical meaning alignment: All data has been mapped to a unified physical effect framework; Numerical scale alignment: All features have been dimensionless through normalization, grading and other operations, and are on comparable orders of magnitude; Data structure alignment: All features have been assembled into standardized feature vectors with fixed dimensions and fixed order.

[0024] S300, based on a preset process knowledge rule base, performs online verification of aligned multi-source features, detects and marks conflicts between features and their levels; In step S300, the process knowledge rule base includes at least one of the following: temperature-microstructure correspondence rules, rolling force-surface morphology association rules, composition-performance matching rules, and spatiotemporal continuity verification rules; the conflict level is divided into level one, level two, and level three, specifically including the following sub-steps: Based on the principles of metallurgical physics, the cooling rate data is correlated and matched with the spectral peak characteristics to establish a mapping relationship between the cooling rate data range and the microstructure type characterized by the spectral peak characteristics, thereby obtaining the correspondence between temperature history and expected structure. Based on mechanical principles, the rolling force parameters from the process parameter time sequence are correlated and matched with the dimensionless index of surface roughness to establish the correspondence between the numerical range of rolling force parameters and the range of the dimensionless index of surface roughness, thereby obtaining the correlation between rolling force parameters and surface morphology characteristics. Performance requirements are preset through component-performance matching rules; based on the correlation principle between chemical composition and performance, the element content grading index is matched with the performance requirements to establish a matching relationship between the element content grading index range and the performance index range, thereby obtaining the component-performance index matching relationship. Based on the time-series properties of process parameters and the spatial location-series properties of surface images and infrared thermal imaging data, the temporal spatial logic of data acquisition is obtained. Using the temporal spatial logic of data acquisition, the numerical difference between continuous data points in the process parameter time series and data points in the standardized feature set is calculated and compared with a preset threshold to obtain the spatiotemporal continuity verification standard. Based on the correspondence between temperature history and expected microstructure, the first verification result is obtained by checking whether the cooling rate level identifier and the microstructure represented by the element content grading index in the standardized feature set are consistent. Based on the correlation between rolling force parameter and surface morphology feature, the second verification result is obtained by checking whether the rolling force parameter and the dimensionless surface roughness index in the standardized feature set are within the corresponding range. Based on the matching relationship between composition and performance index, the third verification result is obtained by checking whether the element content grading index meets the preset performance requirements. Based on the spatiotemporal continuity verification standard, the fourth verification result is obtained by checking whether the difference in standardized feature values ​​between adjacent time series or spatial positions in the standardized feature set is within the spatiotemporal continuity verification standard. If any one of the first, second, third, or fourth verification results is negative, then a conflict between features is identified and the conflict between features is marked. Conflicts between features are classified into three levels: violations of fundamental physical laws are marked as Level 1; violations of conventional process association rules are marked as Level 2; and deviations from statistical expectations are marked as Level 3.

[0025] Furthermore, this step constructs a process knowledge base that integrates metallurgy, mechanics, composition, and spatiotemporal rules, performs multi-dimensional online logical verification of standardized features, automatically identifies and grades conflicts based on the severity of violations of physical laws, process conventions, or statistical expectations, thereby achieving proactive discovery and quantitative assessment of contradictions in multi-source data, and improving the reliability and decision robustness of the diagnostic system under real and complex working conditions.

[0026] Based on a pre-defined process knowledge rule base, the aligned multi-source features are validated online to detect and label conflicts and their levels. When a Level 1 conflict is detected, a manual intervention process is triggered; when a Level 2 conflict is detected, a complete chain of evidence arbitration process is triggered; and when a Level 3 conflict is detected, a rapid weighted arbitration process based on data source weights is triggered, which includes the following sub-steps: Based on the classification of the conflict status of the features, we obtain the first-level conflict set, the second-level conflict set, and the third-level conflict set. For each conflict instance in the first-level conflict set, extract the multi-source heterogeneous feature data and standardized feature set corresponding to the conflict feature identifier; integrate the multi-source heterogeneous feature data with the corresponding conflict rule entries to obtain the file to be adjudicated; submit the file to be adjudicated to the manual adjudication interface. Once the manual adjudication interface receives the file to be adjudicated and the set of first-level conflicts, it binds the manual input information corresponding to the manual adjudication interface with the file to be adjudicated to obtain a complete case record, thereby completing the manual intervention process. For each conflict instance in the secondary conflict set, based on the conflict feature identifier, the corresponding standardized feature value is extracted from the standardized feature set as direct characterization evidence; the parameter sequence related to the conflict occurrence time is extracted from the process parameter time series as indirect process evidence; and the physical mechanism description of rule violation association is retrieved from the process knowledge rule base as causal physical mechanism evidence. The credibility of direct evidence is obtained by multiplying the initial credibility weight corresponding to the direct characterization evidence by the statistical distribution credibility coefficient; the numerical difference between the process parameters involved in the indirect process evidence at the time of conflict and the adjacent previous sampling time is calculated, and the absolute value of the numerical difference is compared with the preset continuity threshold to obtain the consistency index; the matching percentage between the rule condition part of the causal physical mechanism evidence and the current standardized feature set is calculated, and the value with the highest matching percentage is taken as the highest conformity. Based on the current set of secondary conflicts, the weighting coefficients are determined; based on the credibility of direct evidence, the consistency index, and the highest degree of compliance, the final confidence level is calculated using a weighted summation model in combination with the weighting coefficients to complete the evidence chain arbitration process. For each conflict instance in the three-level conflict set, the standardized feature values ​​obtained from each data source involved in the conflict instance are multiplied by the initial confidence weight to obtain the weighted feature value; all weighted feature values ​​are summed to obtain the summation result; based on whether the summation result exceeds the preset fast decision threshold, the tendency result of defect diagnosis is obtained, and the fast weighted arbitration process is completed.

[0027] Furthermore, this step initiates differentiated arbitration mechanisms based on the conflict level (Level 1, Level 2, and Level 3), namely manual intervention, complete multi-evidence chain confidence fusion arbitration, and weighted rapid arbitration. This enables tiered and precise decision-making, where serious conflicts are adjudicated by experts, medium conflicts are analyzed in depth, and minor conflicts are handled efficiently. This ensures the reliability of the diagnosis while improving the overall operational efficiency and practicality of the system.

[0028] S400, based on the conflict level, initiates the corresponding arbitration decision mechanism, integrates multi-source evidence, and outputs defect diagnosis results and confidence levels; S500: Collect diagnostic cases and perform retrospective analysis; dynamically optimize the process knowledge rule base and feature mapping relationship based on the performance evaluation results. In step S500, diagnostic cases are collected and retrospective analysis is performed. Based on the performance evaluation results, the process knowledge rule base and feature mapping relationship are dynamically optimized. The specific steps are as follows: Rule base entries are obtained through the process knowledge rule base, and arbitration decision process records are obtained through the arbitration process; defect diagnosis results, confidence levels, standardized feature sets, rule base entries, arbitration decision process records, and production inspection conclusions are integrated to obtain diagnosis case records; Based on the comparison results between the production inspection conclusion field and the defect diagnosis result field in the diagnostic case records, all diagnostic case records are screened. For cases with consistent conclusions, the corresponding standardized feature set is extracted to obtain the positive sample feature set. For cases with inconsistent conclusions or conflicting records, standardized feature data is extracted from the corresponding standardized feature set as conflict feature data. The system iterates through the diagnostic case records. When a rule base entry is triggered and the diagnostic conclusion is consistent with the production inspection conclusion, it is counted as a correct warning. When a rule base entry is triggered but the diagnostic conclusion is inconsistent, it is counted as a false alarm. When the production inspection conclusion has defects but the rule base entry is not triggered, it is counted as a missed alarm. The number of correct warnings, false alarms, and missed alarms is accumulated, and the efficiency index of each rule base entry is calculated through efficiency evaluation. Based on the efficiency index, the corresponding rule base entry is assigned a high, medium, or low efficiency rating. The elements in the positive sample feature set are grouped according to the defect category. The statistical variance of the features in each group of cases within each effect layer is calculated. The reciprocal of each statistical variance is used to calculate the arithmetic mean to obtain the feature aggregation index. The Euclidean distance between the feature means of different groups is calculated, and the Euclidean distance with the smallest value is taken as the feature discrimination index. The source of conflict feature data is traced back to locate the abnormal occurrence link. For rule base entries corresponding to high efficiency ratings, the cooling rate rating threshold or element content rating threshold is increased by a preset ratio, and the calling priority is increased; for rule base entries corresponding to low efficiency ratings, the cooling rate rating threshold or element content rating threshold is reduced by a preset ratio, and the calling priority is decreased, so as to obtain an adjusted process knowledge rule base. Based on the feature aggregation degree index and the feature discrimination degree index, the parameters involved in the standardization transformation of each defect are adjusted to obtain the adjusted parameters; based on the anomaly occurrence stage, the initial credibility weights of the data sources involved are adjusted to obtain the adjusted credibility weights; the parameters involved in the standardization transformation of each defect include the historical extreme values ​​or hierarchical thresholds of linear normalization. The adjusted parameters, adjusted confidence weights, and adjusted process knowledge rule base are integrated to complete dynamic optimization.

[0029] Furthermore, this step automatically evaluates rule effectiveness and feature representation quality by collecting diagnostic cases in a closed loop and comparing them with real production conclusions. Based on this, the thresholds and priorities of the rule base are dynamically optimized, feature transformation parameters and data source weights are adjusted, enabling the system to continuously self-calibrate and improve during operation, effectively enhancing the adaptive capability, generalization performance and long-term accuracy of the diagnostic model.

[0030] It should be noted that the "integration" mentioned for the first time in this step is essentially a process of data encapsulation and structured storage. It is not an algorithm that creates new information or performs complex transformations, but rather packages multiple independently generated but logically related data items from a complete diagnostic cycle into a standardized case record that can be used for subsequent analysis and traceability, according to a preset, fixed format. Specifically, this "integration" includes the following three distinct sub-operations: Sub-operation 1: Define the input source. This operation involves collecting all relevant data that naturally arises from different modules of the system during a diagnostic task. Sub-operation 2: Create a structured record. This operation associates the above data items and writes them into a unified data structure with fixed fields. Sub-step 3: Create a traceable case. The result of this operation is a self-contained, standardized unit of information. The term "integration," appearing for the first time in this step, specifically refers to the process of linking multiple scattered but logically related data outputs from a single diagnostic event using unique identifiers and storing them in a structured record with a fixed format to create a "data pool" for dynamic optimization. This is the starting point of the entire optimization cycle.

[0031] Meanwhile, the "integration" that appears at the end of this step is a clear and structured configuration management action. Its core is to summarize, format, and deploy the "new configurations" generated by the previous independent optimization steps to generate an updated and complete knowledge system that can be directly used in the next diagnostic cycle. Specifically, this "integration" includes the following three clear sub-operations: Sub-operation 1: Clarify the input items. This operation confirms all optimization results (adjusted processes) to be integrated. Sub-operation 2: Systematic update and replacement. This operation writes or loads new configuration items into a specified storage location or configuration file of the system. Sub-operation 3: Complete the closed loop. This operation confirms that the integration is complete and a traceable version has been created. The word "integration" at the end of this step refers to updating the three core configuration items—the optimized rule base, standardized parameters, and arbitration weights—into the running system in a deterministic manner. This completes the final action of a full "diagnosis-evaluation-optimization" iterative cycle, ensuring that the optimization results are solidified and put into practical application.

[0032] In this embodiment of the invention, an abstract, conceptual physical meaning space is created, serving as a framework for classifying and representing all data. The dimensions of this framework (i.e., multi-layer effect labels) are not mathematically or statistically defined, but rather by the fundamental physical and chemical principles of metallurgy, thermodynamics, and mechanics involved in steel production. This achieves semantic alignment before data fusion, unifying data obtained through different technical means into a description of the physical essence of the production process. Then, the raw or pre-processed features extracted from multi-source heterogeneous data are categorized into corresponding layers of the constructed physical meaning space based on their reflected physical essence. Next, mathematical processing (such as normalization and exponentiation) is applied to the features within each layer to eliminate dimensionality effects and ensure comparability. For example, the temperature distribution of infrared thermography directly corresponds to the thermodynamic effect layer; the peak values ​​of the spectrum correspond to elemental content, belonging to the metallurgical effect layer. In this way, high-dimensional, complex, and heterogeneous raw data can be reduced in dimensionality and normalized into a low-dimensional, homogeneous set of indicators with clear physical interpretations. This significantly reduces the complexity of subsequent calculations and improves the interpretability of the model. Here, feature extraction and transformation algorithms will be developed for each type of data. For example, the temperature matrix of infrared images will be calculated as a temperature uniformity index; the carbon equivalent index will be derived from spectral data through a model, etc.

[0033] Next, the aligned multi-source features are validated online to detect and label conflicts and their severity. A pre-established set of rules, incorporating process knowledge and expert experience, is used to perform real-time logical or numerical validation on the aligned standardized features. The aim is to identify contradictions between conclusions from different data sources and to assess and classify the degree of contradiction. This enables proactive monitoring of data quality and the reliability of the diagnostic process. It can promptly detect sensor malfunctions, abnormal data acquisition, or model misjudgments, preventing a single data source error from causing overall diagnostic failure. Then, based on the severity of the conflict, different pre-defined decision-making processes are invoked to resolve the issue. A decision router is designed here to call different arbitration submodules based on the input conflict level. Finally, the input, process, output, and ultimate truth of each diagnosis are completely saved, forming a case library. These cases are periodically statistically analyzed to evaluate the performance of the rules and feature transformation methods used in the system. Based on the evaluation results, the rules and transformation parameters are automatically adjusted to continuously improve system performance over time.

[0034] In a preferred embodiment of the present invention, the multilayer effect label includes a thermodynamic effect layer, a mechanical effect layer, and a metallurgical effect layer; the standardization conversion includes converting the original feature into a dimensionless or hierarchical index within its respective effect layer.

[0035] In this embodiment of the invention, the specific composition of the multi-layer effect label is clearly defined, which is the core dimension of the physical meaning space. The thermodynamic layer focuses on energy (temperature, cooling); the mechanical layer focuses on force and deformation (stress, rolling force); and the metallurgical layer focuses on changes in the internal structure of materials (composition, phase transformation, microstructure). The classification system is scientific, complete, and mutually exclusive, basically covering all major defect causal categories. The standardization conversion method is also specified. The dimensionless index refers to a value whose physical unit is eliminated through mathematical processing (such as dividing by the maximum value, standardizing to the [0, 1] interval). The grading index discretizes continuous values ​​into several levels. For temperature uniformity, the variance of the actual temperature distribution can be calculated and divided by a theoretical maximum variance to obtain a dimensionless index of 0-1. For cooling rate, it can be divided into three levels: "fast cooling," "medium cooling," and "slow cooling" according to its value range.

[0036] As a preferred embodiment of the present invention, the process knowledge rule base includes at least one of temperature-microstructure correspondence rules, rolling force-surface morphology association rules, composition-performance matching rules, and spatiotemporal continuity verification rules; the conflict level is divided into level one, level two, and level three.

[0037] In this embodiment of the invention, the temperature-structure correspondence rule is a rule that infers the microstructure (such as ferrite, pearlite, martensite) that steel should form based on the temperature history of hot working (rolling, cooling), and compares it with the actual detected microstructure. The rolling force-surface morphology correlation rule refers to the empirical or theoretical relationship between mechanical process parameters such as rolling force and reduction amount and the final steel plate surface roughness, texture, and other morphological characteristics. The composition-performance matching rule refers to the rule that predicts the mechanical properties (strength, toughness) of steel based on its chemical composition using empirical formulas (such as the carbon equivalent formula) or machine learning models, and compares it with the measured or process model-calculated performance. The spatiotemporal continuity verification rule refers to the rule that checks whether adjacent sensor data transitions smoothly in space, or whether the same sensor data changes continuously in time, to identify abrupt changes, spikes, and other abnormal data. Conflict levels are at least divided into Level 1, Level 2, and Level 3, based on whether the conflict violates natural laws, process specifications, or statistical expectations. For example, Level 1 conflict corresponds to violations of fundamental laws; Level 2 corresponds to violations of conventional process rules; and Level 3 corresponds to minor deviations.

[0038] As a preferred embodiment of the present invention, when a first-level conflict is detected, a manual intervention process is triggered; when a second-level conflict is detected, a complete chain of evidence arbitration process is triggered; and when a third-level conflict is detected, a rapid weighted arbitration process based on data source weights is triggered.

[0039] In this embodiment of the invention, the manual intervention process refers to the system automatically pausing and pushing conflicting data, triggered rules, and other information to the manual operation interface, awaiting expert adjudication. The complete evidence chain arbitration process refers to a series of evidence collection, evaluation, and fusion operations, conducting in-depth and rigorous automated analysis of moderate conflicts, balancing efficiency and accuracy. The rapid weighted arbitration process based on data source weights assigns fixed weights to different data sources. When there is a minor conflict, it directly calculates the weighted score of the conclusions supported by each data source, selecting the one with the highest score. This process is extremely fast and efficient when handling simple conflicts.

[0040] As a preferred embodiment of the present invention, the complete chain of evidence arbitration process includes: gathering characterization evidence directly related to the defect, tracing indirect process evidence of the production process, matching causal physical mechanism evidence to explain the defect, and performing confidence fusion calculation on the three types of evidence.

[0041] In this embodiment of the invention, characterization evidence refers to evidence that directly indicates defect information, such as the shape, size, and location of cracks in surface images. Indirect process evidence refers to evidence tracing the background of the defect's occurrence, such as the abnormal fluctuation frequency of the cooling water valve within 30 seconds before the defect appeared, or the temperature trend at the same location in other slabs of the same batch. Causal physical mechanism evidence refers to the theoretical basis explaining the defect's occurrence. Finally, the credibility or conformity of the above three types of evidence is combined into a final credibility score through a mathematical model, making the decision-making process transparent and traceable.

[0042] As a preferred embodiment of the present invention, the corresponding arbitration decision mechanism routes according to a preset decision tree during execution: it receives conflict level labels and routes to the processing paths of manual intervention, full arbitration, or rapid weighted arbitration according to whether the label is level one, level two, or level three.

[0043] In this embodiment of the invention, the implementation architecture of the arbitration decision-making mechanism is defined as a pre-defined decision tree. The root node receives the conflict level, and the three branches of the decision tree (level 1, level 2, and level 3) point to three different processing modules, namely manual arbitration, full arbitration, and fast arbitration. The logic is clear, the execution efficiency is high, the program is easy to implement and maintain, and a dedicated decision tree engine can be used to implement the routing function.

[0044] As a preferred embodiment of the present invention, the effectiveness evaluation of each rule in the rule base entry adopts a quantitative formula: effectiveness index E=(Nc-α×Nf-β×Nm) / Nt, where E is the effectiveness index, Nc is the number of correct warnings, Nf is the number of false alarms, Nm is the number of missed warnings, α and β are penalty coefficients, and Nt is the total number of rule triggers; the weight or trigger threshold of the rule is adjusted according to the effectiveness index E.

[0045] In this embodiment of the invention, the numerator of the performance index calculation formula is the effective contribution minus the harmful contribution, and the denominator is the total rule activity. α is greater than β because the cost of false alarms (disrupting production) is usually higher than that of false negatives, making the assessment comprehensive, quantitative, and objective. The weight of a rule refers to the influence of the conclusion supported by the rule in evidence fusion (such as weighted voting). The trigger threshold is the critical value in the rule's judgment condition; lowering the threshold makes the rule more sensitive, and raising the threshold makes it more stringent. For example, if two performance thresholds Thmax (e.g., 0.7) and Thmin (e.g., 0.2) are set, if E > Thmax, then its weight is increased or the threshold is slightly decreased; if E < Thmin, then the weight is decreased or the threshold is increased; if E is negative multiple times consecutively, then the rule is disabled and an alarm is triggered.

[0046] In a preferred embodiment of the present invention, when mapping multi-source heterogeneous features to corresponding layers, different initial confidence weights are assigned to features from different data sources within the same effect layer.

[0047] In this embodiment of the invention, an initial reliability weight is assigned to different data sources during the feature alignment process. The initial reliability weight is a reliability score assigned to each data source based on prior knowledge before operational data has been accumulated. Specifically, it is based on the reliability of the measurement principle, the historical accuracy of the sensor, and the degree of direct correlation with the core physical quantity. For example, for the mechanical index of surface stress, the weight of directly measured ultrasonic flaw detector data should be higher than that of model output indirectly inferred through image texture. This provides a foundation for rapid arbitration and preliminary fusion.

[0048] As a preferred embodiment of the present invention, the confidence fusion calculation adopts a weighted summation model: final confidence = W1 × direct evidence credibility + W2 × indirect evidence consistency index + W3 × physical mechanism highest conformity, wherein the weight coefficients W1, W2, and W3 are configured according to the conflict level and defect type.

[0049] In this embodiment of the invention, the credibility of direct evidence is derived from the quality scores of multiple direct pieces of evidence (such as image clarity and spectral signal-to-noise ratio) or their mutual corroboration. The indirect evidence consistency index refers to the degree of agreement between indirect evidence and with the current hypothesis. The highest physical mechanism conformity score refers to the score of the mechanism that best matches the current evidence among several listed possible physical mechanisms. The weighting coefficients can be obtained based on historical data training or set by experts according to the defect type. For example, for surface cracks, the weight W1 of image evidence (direct) may be higher; for internal performance discrepancies, the weights W2 and W3 of process evidence (indirect) and compositional mechanisms (physical) may be higher.

[0050] In a preferred embodiment of the present invention, the dynamic optimization periodically uses a standard sample set of known defect categories to evaluate the clustering density of similar defects in the physical meaning space, and adjusts the transformation parameters or weight allocation of the feature mapping accordingly to optimize the alignment effect.

[0051] In this embodiment of the invention, when evaluating the clustering density of similar defects in the physical space, all similar defect samples (e.g., all edge cracks) in the standard sample set are transformed through the current feature mapping relationship, and the degree of clustering of their corresponding feature vectors in the physical space is calculated. High density indicates that the physical fingerprints of similar defects are very similar. Here, evaluation metrics of clustering algorithms, such as intra-cluster average distance and silhouette coefficient, can be used to quantify the density. Adjusting the transformation parameters of the feature mapping refers to the coefficients in the feature transformation algorithm, such as the denominator value during normalization and the constant in the exponential calculation formula. In this way, starting from the most direct machine learning goal of making similar defects appear more similar, the front-end feature engineering process can be optimized in reverse, enabling the entire system to achieve data-driven self-improvement in its ability to represent defects.

[0052] like Figure 2 As shown in the figure, this embodiment of the invention also provides an intelligent diagnostic system for quality defects in steel products based on multi-source data fusion, the system comprising: The physical space construction module 100 is used to construct a unified physical meaning space, which includes multi-layer effect labels defined based on metallurgical physical principles. The effect label mapping module 200 is used to map multi-source heterogeneous features from surface images, process parameter time series, spectral data and infrared thermography to the corresponding layer under the multi-layer effect label according to their physical meaning, and perform standardization conversion to achieve feature alignment. The feature conflict detection module 300 is used to perform online verification of aligned multi-source features based on a preset process knowledge rule base, and to detect and mark conflicts between features and their levels. The decision-making mechanism activation module 400 is used to activate the corresponding arbitration decision-making mechanism according to the conflict level, integrate multi-source evidence, and output defect diagnosis results and confidence levels. The dynamic optimization module 500 is used to collect diagnostic cases and perform retrospective analysis, and to dynamically optimize the process knowledge rule base and feature mapping relationship based on the performance evaluation results.

[0053] The above description only details the preferred embodiments of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0054] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0055] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0056] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the disclosure in the specification and embodiments. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

Claims

1. A method for intelligent diagnosis of quality defects of steel products by multi-source data fusion, characterized in that, The method includes the following steps: Construct a unified physical meaning space, which includes multi-layer effect labels defined based on metallurgical physics principles; Multi-source heterogeneous features from surface images, process parameter timing, spectral data, and infrared thermography are mapped to the corresponding layers under the multi-layer effect label according to their physical meaning, and then standardized and transformed to achieve feature alignment. Based on a pre-defined process knowledge rule base, the aligned multi-source features are validated online to detect and mark conflicts and their levels between features. Based on the level of conflict, the corresponding arbitration decision-making mechanism is initiated, multi-source evidence is integrated, and defect diagnosis results and confidence levels are output. Diagnostic cases are collected and retrospectively analyzed. Based on the performance evaluation results, the process knowledge rule base and feature mapping relationship are dynamically optimized.

2. The multi-source data fusion-based intelligent diagnosis method of steel product quality defects according to claim 1, characterized in that, The multi-source heterogeneous features from surface images, process parameter timing, spectral data, and infrared thermography are mapped to the corresponding layers under the multilayer effect label according to their physical meaning, and then standardized to achieve feature alignment. The specific steps include the following: S201, based on the principles of metallurgical physics, divides the unified physical meaning space into a thermodynamic effect layer, a mechanical effect layer, and a metallurgical effect layer to obtain a structured framework; among them, the thermodynamic effect layer is used to characterize the temperature history and energy state during the production process; the mechanical effect layer is used to characterize the stress, strain, and deformation behavior of the material; and the metallurgical effect layer is used to characterize changes in chemical composition and the resulting microstructure evolution. S202, the texture feature values ​​and geometric features obtained from the surface image are mapped to the mechanical effect layer of the structured framework; the cooling rate data extracted from the process parameter time series are mapped to the thermodynamic effect layer of the structured framework; the spectral peak features obtained from the spectral data are mapped to the metallurgical effect layer of the structured framework; the temperature distribution matrix extracted from the infrared thermogram is mapped to the thermodynamic effect layer of the structured framework; so as to obtain the feature set for the initial classification. S203, For continuous numerical features in the feature set of the initial classification, a linear normalization method is used to process them. At the same time, for state features in the feature set of the initial classification, a finite number of level indexes are converted according to a preset threshold to obtain a standardized feature set. S204 Calculate the standard deviation of all pixel values ​​in the infrared thermal image temperature matrix in the standardized feature set, and then divide the standard deviation by the preset maximum allowable temperature standard deviation threshold to obtain the dimensionless temperature index. S205, compare the cooling rate data in the standardized feature set with the preset cooling rate classification threshold to obtain the cooling rate level identifier; wherein, the cooling rate classification threshold has three cooling rate ranges. S206, calculate the contrast and uniformity feature values ​​of the texture feature values ​​using the gray-level co-occurrence matrix algorithm, process the ratio of contrast to uniformity feature values ​​according to the linear normalization method in step S203 to obtain the dimensionless index of surface roughness; compare the spectral peak features with the preset standard curve to obtain the element content grading index. S207, the dimensionless temperature index, cooling rate level identifier, surface roughness dimensionless index, and element content classification index are organized and aggregated according to the effect layer to complete feature alignment.

3. The intelligent diagnostic method for quality defects in steel products based on multi-source data fusion according to claim 2, characterized in that, In the process of online verification of aligned multi-source features based on a preset process knowledge rule base, detecting and marking conflicts and their levels between features, the process knowledge rule base includes at least one of temperature-microstructure correspondence rules, rolling force-surface morphology association rules, composition-performance matching rules, and spatiotemporal continuity verification rules; the conflict levels are divided into level one, level two, and level three, specifically including the following sub-steps: Based on the principles of metallurgical physics, the cooling rate data is correlated and matched with the spectral peak characteristics to establish a mapping relationship between the cooling rate data range and the microstructure type characterized by the spectral peak characteristics, thereby obtaining the correspondence between temperature history and expected structure. Based on mechanical principles, the rolling force parameters from the process parameter time sequence are correlated and matched with the dimensionless index of surface roughness to establish the correspondence between the numerical range of rolling force parameters and the range of the dimensionless index of surface roughness, thereby obtaining the correlation between rolling force parameters and surface morphology characteristics. Performance requirements are preset through component-performance matching rules; based on the correlation principle between chemical composition and performance, the element content grading index is matched with the performance requirements to establish a matching relationship between the element content grading index range and the performance index range, thereby obtaining the component-performance index matching relationship. Based on the time-series properties of process parameters and the spatial location-series properties of surface images and infrared thermal imaging data, the temporal spatial logic of data acquisition is obtained. Using the temporal spatial logic of data acquisition, the numerical difference between continuous data points in the process parameter time series and data points in the standardized feature set is calculated and compared with a preset threshold to obtain the spatiotemporal continuity verification standard. Based on the correspondence between temperature history and expected microstructure, the first verification result is obtained by checking whether the cooling rate level identifier and the microstructure represented by the element content grading index in the standardized feature set are consistent. Based on the correlation between rolling force parameter and surface morphology feature, the second verification result is obtained by checking whether the rolling force parameter and the dimensionless surface roughness index in the standardized feature set are within the corresponding range. Based on the matching relationship between composition and performance index, the third verification result is obtained by checking whether the element content grading index meets the preset performance requirements. Based on the spatiotemporal continuity verification standard, the fourth verification result is obtained by checking whether the difference in standardized feature values ​​between adjacent time series or spatial positions in the standardized feature set is within the spatiotemporal continuity verification standard. If any one of the first, second, third, or fourth verification results is negative, then a conflict between features is identified and the conflict between features is marked. Conflicts between features are classified into three levels: violations of fundamental physical laws are marked as Level 1; violations of conventional process association rules are marked as Level 2; and deviations from statistical expectations are marked as Level 3.

4. The intelligent diagnostic method for quality defects in steel products based on multi-source data fusion according to claim 3, characterized in that, Based on a pre-defined process knowledge rule base, the aligned multi-source features are validated online. During the process of detecting and marking conflicts and their levels between features, a manual intervention process is triggered when a first-level conflict is detected. Upon detection of a Level 2 conflict, a complete chain of evidence arbitration process is triggered; upon detection of a Level 3 conflict, a rapid weighted arbitration process based on data source weights is triggered, which includes the following sub-steps: Based on the classification of the conflict status of the features, we obtain the first-level conflict set, the second-level conflict set, and the third-level conflict set. For each conflict instance in the first-level conflict set, extract the multi-source heterogeneous feature data and the standardized feature set corresponding to the conflict feature identifier; The multi-source heterogeneous feature data is integrated with the corresponding conflict rule entries to obtain the file to be adjudicated; the file to be adjudicated is then submitted to the manual adjudication interface. Once the manual adjudication interface receives the file to be adjudicated and the set of first-level conflicts, it binds the manual input information corresponding to the manual adjudication interface with the file to be adjudicated to obtain a complete case record, thereby completing the manual intervention process. For each conflict instance in the secondary conflict set, the corresponding standardized feature value is extracted from the standardized feature set as direct characterization evidence based on the conflict feature identifier. Extract parameter sequences related to the timing of conflict occurrence from the process parameter time series as indirect process evidence; retrieve physical mechanism descriptions of rule violations from the process knowledge rule base as causal physical mechanism evidence; The credibility of direct evidence is obtained by multiplying the initial credibility weight corresponding to the direct characterization evidence by the statistical distribution credibility coefficient. The numerical difference between the process parameters involved in the indirect process evidence at the time of the conflict and the adjacent previous sampling time is calculated, and the absolute value of the numerical difference is compared with a preset continuity threshold to obtain the consistency index. Calculate the percentage of match between the rule condition portion of the causal physical mechanism evidence and the current standardized feature set, and take the value with the highest percentage of match as the highest degree of agreement; Determine the weighting coefficients based on the current set of secondary conflicts; Based on the credibility of direct evidence, consistency index, and highest degree of compliance, the final confidence level is calculated using a weighted summation model with weighting coefficients to complete the evidence chain arbitration process. For each conflict instance in the three-level conflict set, the standardized feature values ​​obtained from each data source involved in the conflict instance are multiplied by the initial confidence weight to obtain the weighted feature value. All weighted feature values ​​are summed to obtain the summation result; based on whether the summation result exceeds a preset fast decision threshold, a tendency result for defect diagnosis is obtained, and the fast weighted arbitration process is completed.

5. The intelligent diagnostic method for quality defects in steel products based on multi-source data fusion according to claim 4, characterized in that, In the process of initiating the corresponding arbitration decision mechanism according to the conflict level, integrating multi-source evidence and outputting defect diagnosis results and confidence levels, the corresponding arbitration decision mechanism routes according to a preset decision tree during execution. It receives conflict level labels and routes to the processing path of manual intervention, full arbitration or fast weighted arbitration according to whether the label is level one, level two or level three.

6. The intelligent diagnostic method for quality defects in steel products based on multi-source data fusion according to claim 5, characterized in that, The diagnostic cases are collected and backtracked for analysis. Based on the performance evaluation results, the process knowledge rule base and feature mapping relationship are dynamically optimized. The specific steps are as follows: Rule base entries are obtained through the process knowledge rule base, and arbitration decision process records are obtained through the arbitration process; defect diagnosis results, confidence levels, standardized feature sets, rule base entries, arbitration decision process records, and production inspection conclusions are integrated to obtain diagnosis case records; Based on the comparison results between the production inspection conclusion field and the defect diagnosis result field in the diagnostic case records, all diagnostic case records are screened. For cases with consistent conclusions, the corresponding standardized feature set is extracted to obtain the positive sample feature set. For cases with inconsistent conclusions or conflicting records, standardized feature data is extracted from the corresponding standardized feature set as conflict feature data. Traverse the diagnostic case records. When a rule base entry is triggered and the diagnostic conclusion is consistent with the production inspection conclusion, it is counted as a correct warning. When a rule base entry is triggered but the diagnostic conclusion is inconsistent, it is counted as a false alarm; when the production inspection conclusion is defective but the rule base entry is not triggered, it is counted as a missed alarm; the cumulative number of correct warnings, false alarms and missed alarms is used to calculate the performance index of each rule base entry through performance evaluation, and the corresponding rule base entry is assigned a high, medium or low performance level label based on the performance index. The elements in the positive sample feature set are grouped according to the defect category. The statistical variance of the features in each group of cases within each effect layer is calculated. The reciprocal of each statistical variance is used to calculate the arithmetic mean to obtain the feature aggregation index. The Euclidean distance between the feature means of different groups is calculated, and the Euclidean distance with the smallest value is taken as the feature discrimination index. The source of conflict feature data is traced back to locate the abnormal occurrence link. For rule base entries corresponding to high efficiency ratings, the cooling rate rating threshold or element content rating threshold is increased by a preset ratio, and the calling priority is increased; for rule base entries corresponding to low efficiency ratings, the cooling rate rating threshold or element content rating threshold is reduced by a preset ratio, and the calling priority is decreased, so as to obtain an adjusted process knowledge rule base. Based on the feature aggregation degree index and the feature discrimination degree index, the parameters involved in the standardization transformation of each defect are adjusted to obtain the adjusted parameters; based on the anomaly occurrence stage, the initial credibility weights of the data sources involved are adjusted to obtain the adjusted credibility weights. The adjusted parameters, adjusted confidence weights, and adjusted process knowledge rule base are integrated to complete dynamic optimization.

7. The intelligent diagnostic method for quality defects in steel products based on multi-source data fusion according to claim 6, characterized in that, The parameters involved in the standardization transformation of each defect include the historical extreme value or the hierarchical threshold of linear normalization.

8. The intelligent diagnostic method for quality defects in steel products based on multi-source data fusion according to claim 7, characterized in that, In the process of calculating the performance index of each rule in the rule base entry through performance evaluation by accumulating the number of correct warnings, false alarms, and missed warnings, a quantitative formula is used to evaluate the performance of each rule in the rule base entry. The quantitative formula is as follows: E=(Nc-α×Nf-β×Nm) / Nt; Where E is the efficiency index, Nc is the number of correct alerts, Nf is the number of false alarms, Nm is the number of missed alerts, α and β are the penalty coefficients, and Nt is the total number of rule triggers; Adjust the weights or trigger thresholds of the rules based on the performance index E.

9. The intelligent diagnostic method for quality defects in steel products based on multi-source data fusion according to claim 8, characterized in that, When mapping multi-source heterogeneous features to corresponding layers, different initial confidence weights are assigned to features from different data sources within the same effect layer.

10. The intelligent diagnostic method for quality defects in steel products based on multi-source data fusion according to claim 9, characterized in that, The confidence fusion calculation adopts a weighted summation model: final confidence = W1 × direct evidence credibility + W2 × indirect evidence consistency index + W3 × physical mechanism highest conformity, where the weight coefficients W1, W2, and W3 are configured according to the conflict level and defect type.

11. The intelligent diagnostic method for quality defects in steel products based on multi-source data fusion according to claim 10, characterized in that, In the dynamic optimization, the clustering density of similar defects in the physical meaning space is evaluated periodically using a standard sample set of known defect categories, and the transformation parameters or weight allocation of the feature mapping are adjusted accordingly to optimize the alignment effect.

Images