A method for analyzing the root cause of internal bubble hole defects of a grinding wheel

Abstract

The present application relates to a kind of sand wheel internal bubble hole defect root cause analysis method, to solve the technical problems such as double difficulty efficient fusion, automatic traceability and model self-adaptive optimization in the process of sand wheel bubble hole defect attribution analysis Physical mechanism and data statistics.The technical scheme is analyzed by process specification structure, combined with thermodynamics modeling and historical defect sample statistics, constructs process semantic anchor point sequence, and realizes the mapping from multidimensional CT defect feature to process activation probability by means of deep learning method.Furthermore, by local and cross-process causal reasoning, physical constraint verification, causal path credibility assessment, root cause ordering and report generation are realized, and adaptive iterative optimization of production line is supported.Its beneficial effect lies in realizing the mechanism and data fusion attribution of sand wheel bubble defect, supporting automatic traceability and efficient and accurate optimization, and improving the intelligent level of defect processing in production line.

Description

Technical Field

[0001] This invention relates to the field of quality management and intelligent defect attribution analysis in the grinding wheel manufacturing process, and in particular to a method for root cause analysis of internal bubble porosity defects in grinding wheels. Background Technology

[0002] Quality control and defect attribution in the current grinding wheel manufacturing process have become important research directions in the field of precision manufacturing. Internal air bubble defects in grinding wheels have a significant impact on product performance. Accurately tracing the root causes of these defects and achieving highly reliable root cause sequencing analysis are key technical objectives for ensuring product consistency and reducing scrap rates. Currently, mainstream technical approaches mainly include defect attribution based on expert experience rules, data-driven statistical analysis methods, and intelligent attribution systems based on knowledge graphs and causal reasoning models.

[0003] A common approach in the industry is the defect attribution system based on knowledge graphs. This type of method constructs a multi-level knowledge graph containing raw materials, process parameters, equipment environment, and defect phenomena. Through entity relationship reasoning, path search, or rule engines, it automatically locates the causes of defects. In terms of development trends, some high-end grinding wheel manufacturers have attempted to introduce multivariate statistical analysis, machine learning models, and causal discovery algorithms to conduct data-driven attribution analysis of defect formation, improving the automation and intelligence of defect detection and tracing. Meanwhile, preliminary explorations in literature and engineering practice have explored methods such as causal path tracing that combine process flow modeling with physical mechanisms.

[0004] However, existing technologies still face numerous bottlenecks and shortcomings in practical applications. First, knowledge graph-based attribution methods require long-term accumulation and maintenance of a complete entity-attribute-relationship network, involving massive amounts of heterogeneous data across the entire process, including process parameters, equipment status, and material properties. The construction, updating, and fusion of knowledge graphs are extremely costly, with long maintenance cycles and weak resistance to dynamic process changes. Second, existing causal reasoning models often remain at the level of global static relationship identification or data correlation analysis, failing to fully embed the specific physical mechanisms of the grinding wheel manufacturing process. The causal chain lacks physical explanations of process nodes, resulting in unsatisfactory interpretability and verifiability of the derived root cause conclusions. Furthermore, traditional path search and knowledge reasoning methods suffer from slow reasoning speed, numerous ambiguous paths, and a lack of result credibility evaluation when facing high-dimensional parameter coupling anomalies, dynamic process adjustments, and new defect patterns, making it difficult to meet the actual needs of real-time production line decision-making and continuous self-optimization.

[0005] In practice, production line engineers expect the system to output more than just attributions for deviations in a specific process parameter; they desire comprehensive root cause analysis reports with traceable physical mechanisms, concrete data evidence chains, and highly reliable ranking criteria. However, existing systems suffer from significant shortcomings in the deep integration and dynamic mapping capabilities of causal reasoning and knowledge graphs. Their causal chain analysis is coarse-grained and unable to accurately describe the defect generation process driven by the synergistic failure of process parameters, making it difficult to provide truly effective process intervention suggestions for abnormal batches. Furthermore, with the introduction of new equipment and processes, and product upgrades, the fixed knowledge network struggles to adapt to unknown defect patterns in a timely manner. The dynamic mapping rules between defect characteristics and process flows lack self-learning and continuous optimization mechanisms, limiting the system's intelligent evolution capabilities.

[0006] In summary, existing intelligent attribution analysis technologies for air bubble defects in grinding wheels urgently need to overcome technical bottlenecks such as high-cost reliance on global knowledge graph modeling, lack of physical interpretability in reasoning paths, difficulty in quantitatively identifying causes due to abnormal parameter coupling, and insufficient adaptability to new defect patterns. There is an urgent need to propose a root cause ranking and output method that integrates physical process mechanisms, lightweight dynamic causal framework modeling, multi-source parameter coupling identification, and adaptive learning, to significantly improve the interpretability, real-time performance, and engineering usability of attribution analysis for air bubble defects inside grinding wheels. Summary of the Invention

[0007] This application provides a method for root cause analysis of internal bubble porosity defects in grinding wheels, aiming to solve one of the problems or issues of the prior art mentioned in the background section.

[0008] This application provides a method for root cause analysis of internal bubble porosity defects in grinding wheels, specifically including: S1: Obtain the sequence of key process nodes in the grinding wheel manufacturing process specification. Based on the explanation of the physical mechanism of bubble formation in each process and the statistical correlation pattern of historical defect samples, construct a process semantic anchor sequence encapsulating a set of typical controllable process parameters, sensitivity factors and statistical correlation patterns.

[0009] S2: Input the multidimensional defect feature vector, composed of the bubble pore spatial density heat map, pore diameter spectrum and pore wall microcrack features extracted from the grinding wheel CT scan image, into the pre-trained anchor response predictor, and generate the anchor activation probability distribution representing the significant activation probability of each process based on the process semantic anchor sequence.

[0010] S3: Based on the anchor point activation probability distribution, select high-response process semantic anchor points as the starting point for causal inference. For each high-response process semantic anchor point, call its bound local causal strength evaluator and calculate the local causal contribution score of the characterization parameter to the marginal contribution of bubble formation based on the actual parameter value of the current batch, equipment status log fragments and material batch attributes.

[0011] S4: Identify the abnormal combinations of parameter couplings in the local causal contribution scores, and combine them with the process flow to pass the local causal contribution scores along the process chain to the process semantic anchors with clear physical interpretation paths in the preceding order, thereby generating an initial causal propagation path set containing cross-process causal effect propagation paths.

[0012] S5: Based on preset physical constraint rules, perform process fidelity attenuation verification on each path in the initial causal propagation path set, eliminate physically infeasible paths that violate the upper limit of drying temperature or skip necessary processes, and generate a set of valid causal propagation paths verified by physical constraints.

[0013] S6: Summarize the causal scores at the endpoints of the effective causal propagation path set, and perform weighted fusion calculations by combining the process weights, parameter coupling confidence, and historical verification frequencies of the process semantic anchors along each path to generate a normalized root cause ranking list that represents the confidence of root causes.

[0014] S7: Extract the root cause anchor name, dominant parameter combination, and attribution conclusion composed of physical interpretation for each data in the normalized root cause ranking list, and generate a structured root cause analysis report containing an index of the original data fragments.

[0015] S8: Feed the structured root cause analysis report back to the production line decision system. If a new defect feature pattern is detected that does not match the existing process semantic anchor response, the statistical correlation pattern in the process semantic anchor sequence is updated to optimize the prediction accuracy of the subsequent anchor response predictor.

[0016] The method for root cause analysis of internal bubble porosity defects in grinding wheels provided in this application has the following beneficial effects: (1) The root cause ranking method for internal air bubble defects in grinding wheels provided in this application effectively overcomes the high cost and low flexibility of traditional knowledge graph-based fault diagnosis systems in terms of construction and maintenance. Existing technologies rely on explicitly constructing a global knowledge graph covering equipment, process parameters, and material properties, and performing cross-node path search through complex relational reasoning algorithms. This not only requires a large number of experts to participate in rule definition and data alignment, but also makes it difficult to adapt to scenarios of dynamic adjustment of production line parameters or introduction of new processes. In contrast, this solution abandons the full graph modeling paradigm and instead adopts "process semantic anchor points" as lightweight and modular causal reasoning units. Each anchor point only encapsulates the controllable parameter set, physical sensitivity factors, and historical statistical association patterns under a specific process. In the offline stage, there is no need to perform entity linking or relation fusion, which greatly reduces the threshold for system initialization and updating. During online analysis, the system directly maps the defect feature vector to the probability distribution of activation of each process through a pre-trained anchor point response predictor, skipping the entity recognition and graph matching process in traditional methods, significantly improving the response speed and avoiding the risk of missed detection due to incomplete graphs.

[0017] (2) This scheme, by introducing local causal strength assessment and restricted forward propagation mechanism, realizes the leap from "correlation matching" to "physical explanation-driven" root cause localization, which significantly enhances the technical credibility and engineering operability of the diagnostic results. Unlike general graph neural networks that infer the correlation path based solely on the co-occurrence frequency of data, this method calls the bound neural symbol model for each activated anchor point, combines the actual process parameters, equipment logs and material properties of the current batch, quantifies the marginal causal contribution of each parameter to bubble formation, and automatically identifies nonlinear coupling anomalies between key parameters (such as the synergistic effect of high temperature rapid burning and low binder content causing premature liquid phase precipitation), thereby revealing potential process runaway mechanisms; furthermore, the system implements causal effect forward propagation along the process flow, but strictly limits the propagation span to 1 to 2 anchor points with clear physical explanation paths in the preceding sequence, and embeds verified physical constraint rules (such as irreversible temperature gradient and continuous densification window) in each step of the propagation for verification, to prevent false attributions that violate the manufacturing mechanism. This mechanism ensures that all derivation paths can be traced back to specific process steps and known physical principles, and the generated root cause explanations are highly interpretable, making it easy for field engineers to quickly understand and adopt them.

[0018] (3) This invention constructs a closed-loop, scalable process defect attribution system that balances inference efficiency, accuracy, and industrial adaptability, making it particularly suitable for advanced ceramic product manufacturing scenarios with multiple processes, strong coupling, and high precision requirements. The final attribution list comprehensively considers the process weights of the anchor points along each path (reflecting their influence on product density), parameter coupling confidence, and historical verification frequency, outputting normalized ranking results. Each item includes the root cause anchor point name, dominant parameter combination, physical explanation phrase, and original data index, supporting decision tracing and continuous optimization. The entire process does not require complex graph algorithm search or large-scale parameter fine-tuning, has low computational overhead and low deployment cost, and can complete the entire process from defect image input to root cause ranking output within hundreds of milliseconds, meeting the real-time feedback requirements of the production line. More importantly, this architecture naturally supports incremental expansion—adding a process only requires defining new semantic anchor points and training the corresponding evaluator, without affecting the operation of existing modules, and possesses good process migration capabilities and system robustness.

[0019] In summary, this solution uses process semantic anchors as the core carrier, integrates physical mechanism modeling and data-driven prediction, and constructs a new paradigm for defect root cause ranking that can achieve high accuracy, interpretability, and high efficiency without the need for global knowledge graph support. It effectively solves the contradiction between modeling complexity, inference reliability, and timeliness in traditional methods, and provides a reliable process quality traceability tool for intelligent manufacturing of high-end abrasives. It is especially suitable for the rapid location and process inversion analysis of bubble-like defects in multi-stage thermoforming products such as grinding wheels. Attached Figure Description

[0020] Figure 1 This is the main flowchart of a method for root cause analysis of internal bubble porosity defects in grinding wheels; Figure 2 This is a sub-flowchart of a method for root cause analysis of internal bubble porosity defects in grinding wheels; Figure 3 This is another sub-flowchart of a method for root cause analysis of internal bubble porosity defects in grinding wheels. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0023] like Figure 1 As shown, this application provides a method for root cause analysis of internal air bubble defects in grinding wheels, specifically including: S1: Obtain the sequence of key process nodes in the grinding wheel manufacturing process specification, and construct the process semantic anchor sequence based on the explanation of the physical mechanism of bubble formation in each process and the statistical correlation pattern of historical defect samples.

[0024] Specifically, the process semantic anchor sequence encapsulates a typical set of controllable process parameters, sensitivity factors, and statistical correlation patterns.

[0025] S2: The multidimensional defect feature vector extracted from the CT scan image of the grinding wheel is input into the pre-trained anchor response predictor, and the anchor activation probability distribution is generated based on the process semantic anchor sequence.

[0026] Specifically, a multidimensional defect feature vector composed of the bubble pore spatial density heat map, pore diameter spectrum, and pore wall microcrack features extracted from the grinding wheel CT scan image is input into a pre-trained anchor response predictor, and an anchor activation probability distribution representing the significant activation probability of each process is generated based on the process semantic anchor sequence.

[0027] S3: Based on the anchor point activation probability distribution, high-response process semantic anchor points are selected. For each high-response process semantic anchor point, its bound local causal strength evaluator is invoked to calculate the local causal contribution score based on the actual parameter values ​​of the current batch, equipment status log fragments, and material batch attributes. The high-response process semantic anchor point serves as the starting point for causal inference, and the local causal contribution score characterizes the marginal contribution of parameters to bubble formation.

[0028] S4: The local causal contribution score is passed along the process chain to the process semantic anchor point with a clear physical interpretation path in the preceding sequence to generate an initial causal propagation path set.

[0029] Specifically, the abnormal combinations of parameter couplings in the local causal contribution scores are identified, and the local causal contribution scores are transmitted along the process chain to the process semantic anchors with clear physical interpretation paths in the forward sequence, in conjunction with the process flow, to generate an initial causal propagation path set containing cross-process causal effect propagation paths.

[0030] S5: Based on preset physical constraint rules, perform process fidelity attenuation verification on each path in the initial causal propagation path set to generate a valid causal propagation path set. The valid causal propagation path set is verified by physical constraints; performing process fidelity attenuation verification on each path in the initial causal propagation path set can eliminate physically infeasible paths that violate the upper limit of drying temperature or skip necessary processes.

[0031] S6: Summarize the causal scores at the endpoints of the effective causal propagation path set, and perform a weighted fusion operation combining the process weights, parameter coupling confidence levels, and historical verification frequencies of the process semantic anchors along each path to generate a normalized root cause ranking list. The normalized root cause ranking list is used to characterize the confidence level of the root causes.

[0032] S7: Extract the root cause anchor name, dominant parameter combination, and attribution conclusion based on physical interpretation for each data point in the normalized root cause ranking list, and generate a structured root cause analysis report. The structured root cause analysis report includes an index of the original data fragments.

[0033] S8: Feed the structured root cause analysis report back to the production line decision system. If a new defect feature pattern is detected that does not match the existing process semantic anchor response, the statistical correlation pattern in the process semantic anchor sequence is updated to optimize the prediction accuracy of the subsequent anchor response predictor.

[0034] Step S1: Obtain the sequence of key process nodes in the grinding wheel manufacturing process specification. Based on the explanation of the physical mechanism of bubble formation in each process and the statistical correlation pattern of historical defect samples, construct a process semantic anchor sequence encapsulating a set of typical controllable process parameters, sensitivity factors, and statistical correlation patterns. Specifically, this includes: S1.1: Perform natural language processing and structured parsing on the grinding wheel manufacturing process specification document to extract key process node sequences such as mixing, forming, drying, sintering and finishing, and map the key process node sequences into an initial process flow topology with temporal logical relationships to establish the time axis benchmark for root cause analysis.

[0035] Natural language processing and structured parsing are performed on the input data of the grinding wheel manufacturing process specification document. The document includes text paragraphs and table records containing step descriptions, parameter thresholds, and process control conditions. The original input condition is the full-process specification text that has been OCR-recognized and encoded in UTF-8 format. A word segmentation engine based on a process terminology dictionary is constructed to decompose the text into a structured word sequence with component, action, parameter, and constraint identifiers. The process name and key control action markers are extracted using a named entity recognition model. A time sequence information extraction algorithm is called to parse the time conditions, sequence indicators, and dependencies associated with each process name, generating a process sequence relationship matrix. A syntactic analysis tree is used to analyze the conditional statements and auxiliary clauses contained in the process description, filtering process nodes related to the physical mechanism of bubble formation and eliminating auxiliary processes that do not have a direct causal relationship. The extracted key process nodes such as mixing, forming, drying, sintering, and finishing are mapped into a graph structure according to the dependency matrix, constructing an initial process flow topology structure with directed edges connecting the nodes, and assigning a weight value based on the process constraints to each edge. By using a process topology weight mapping method, the parsed sequential dependencies are transformed into a timeline benchmark that can be used for subsequent sensitivity factor derivation, thus realizing the temporal expression of the process flow. For example, in a grinding wheel manufacturing process specification document, the mixing process is described as "mixing 120-mesh abrasive with liquid binder at a mass ratio of 3:1," the forming process is described as "applying 30MPa pressure to the mold for forming," the drying process is described as "standing at 60℃ for 12 hours," the sintering process is described as "heating to 850℃ and holding for 1 hour," and the finishing process is described as "cutting the surface to a dimensional tolerance of ±0.01mm using a diamond tool." During parsing, a process terminology dictionary containing node keywords such as "mixing," "forming," "drying," "sintering," and "finishing" is used, and a named entity recognition model accurately extracts the five process nodes. The sequential dependency matrix is ​​formalized as mixing → forming → drying → sintering → finishing. In the dependency matrix, the weights of mixing and forming are 1, forming and drying are 1, drying and sintering are set to 0.9 to represent conditional dependency constraints, and sintering and finishing are 1. An initial process flow topology is generated through graph mapping, with 5 nodes and 4 edges. The weight distribution is used for subsequent sensitivity factor derivation. The initial process flow topology output in this step will serve as the input benchmark for calling the material thermodynamics and kinetics simulation model in the next sub-step, enabling the derivation of bubble nucleation and growth mechanisms for each process node, and ultimately supporting the construction of the defect attribution model.

[0036] S1.2: Based on each process node in the initial process flow topology, the material thermodynamics and kinetics simulation model is called to deduce the bubble nucleation and growth mechanism, and generate a set of sensitivity factors that characterize each process to bubble formation with a clear physical explanation, so as to quantify the theoretical sensitivity of different process fluctuations to defect generation.

[0037] Based on the node identifiers of each process in the initial process flow topology output by step S1.1, the material thermodynamics simulation module is called to load the original set of physical property parameters for the corresponding process, including physical properties such as material composition ratio, solid / liquid phase transition temperature, specific heat capacity and thermal conductivity.

[0038] For each process node, the original set of physical property parameters is coupled with the environmental conditions (temperature curve, humidity curve, atmosphere composition) corresponding to the process node in the thermodynamic equilibrium model to obtain the stability criteria of the solid-liquid-gas three-phase interface.

[0039] Based on the phase interface stability criterion, the bubble nucleation rate formula and growth rate formula are used for iterative calculation in the dynamic simulation module. The bubble nucleation rate calculation formula is as follows:

[0040] Where γ is the interfacial tension, k is the Boltzmann constant, T is the absolute temperature, ΔP is the supercooling or supersaturation driving force, and J 0,ref This serves as a reference nucleation baseline rate.

[0041] By combining the above nucleation rate and growth rate, the change curve of bubble radius over time is obtained through time integration, and the maximum bubble size and bubble number density under the process conditions are calculated.

[0042] Sensitivity analysis was performed on bubble size and number density parameters. The variable perturbation method was used to adjust each process control parameter within ±δ range, recording the relative changes in bubble nucleation and growth output indicators, and calculating the sensitivity coefficient for each parameter. The formula for the sensitivity coefficient is:

[0043] Where y is the original output index value, y′ is the index value after perturbation, and δ is the perturbation ratio.

[0044] Parameters whose sensitivity coefficient values ​​are significantly higher than a preset threshold in each process node are identified as sensitivity factors, and a set of sensitivity factors is output as the physical mechanism input basis for subsequent statistical correlation pattern mining.

[0045] By using the above-mentioned thermodynamic and kinetic simulation coupling processing method, the process nodes obtained from the previous step are mapped into a set of process sensitivity factors with quantitative theoretical sensitivity, thereby achieving the technical transformation effect from process node abstraction to quantifiable physical influence indicators.

[0046] S1.3: Using a historical sample library of internal air bubble defects in grinding wheels, a multidimensional statistical correlation mining algorithm is executed on the typical controllable process parameter set corresponding to the set of sensitivity factors to calculate the mutual information entropy value between parameter deviation and air bubble space density heatmap, pore size spectrum and pore wall microcrack characteristics, and to generate a statistical correlation pattern matrix that characterizes the data-driven law.

[0047] For the set of sensitivity factors and their corresponding set of typical controllable process parameters generated in the previous sub-step S1.2, multidimensional defect feature data from the historical grinding wheel internal bubble pore defect sample library is loaded, including bubble space density heatmaps, pore size spectra, and structured description matrices of pore wall microcrack features. Deviation calculation is performed on the set of typical controllable process parameters. Reference values ​​are set from the nominal parameters in the process specification. The deviation expression is centered and standardized to form a dimensionless parameter deviation vector. The parameter deviation vector and the multidimensional defect feature matrix are jointly encoded to form a parameter-defect feature aligned dataset, ensuring that each observed sample has a corresponding record in both the process parameter domain and the defect feature domain. A multidimensional statistical association mining algorithm is called to perform mutual information entropy calculation on the encoded dataset, using the formula:

[0048] Here, X represents the parameter deviation variable, Y represents the defect feature variable, and p(x,y) is the joint probability distribution. Significantly correlated parameter-defect feature pairs are extracted by calculating and arranging the mutual information entropy values ​​of each parameter and each defect feature. The mutual information entropy value sequence is then subjected to a significance test, and correlation terms below a preset significance level are removed to reduce noise influence, retaining stable correlation terms to form a statistical correlation pattern. The correlation terms that pass the test are grouped by process node to form a statistical correlation pattern matrix with sensitivity factors as rows and defect feature indicators as columns. The correlation strength value and significance marker are embedded in the matrix to achieve a quantifiable mapping of the relationship between parameter fluctuations and defect features. Through the above processing method, the sensitivity factors and process parameter set from the previous step are transformed into a statistical correlation pattern matrix with data-driven regularity characteristics, providing a quantitative basis for subsequent process semantic anchor encapsulation.

[0049] For example, regarding the solvent evaporation rate sensitivity factor in the drying stage and the ambient humidity in the typical controllable process parameters of this process, a nominal humidity value of 45% is set, and the humidity observation range in historical samples is 30% to 60%. The deviation vector element is calculated as (observed humidity) 45) / standard deviation. The mean value of the bubble space density heatmap and the main peak frequency of the pore size spectrum of this batch of samples are extracted as defect feature inputs. The joint encoding dataset contains 200 sample entries. The mutual information algorithm is called to calculate the mutual information entropy value on each parameter-defect feature combination, such as mutual information entropy I(humidity; poredensity The value of ) = 0.85, exceeding the significance threshold of 0.5, indicates a stable correlation term. After matrix processing, in the correlation pattern matrix corresponding to the drying stage, the grid points for environmental humidity and bubble space density features are assigned a value of 0.85 and marked as highly significant terms, while the grid points for environmental humidity and pore size spectrum features are assigned a value of 0.62 and marked as moderately significant terms. The resulting matrix serves as the statistical correlation feature input for the drying stage in subsequent process semantic anchor encapsulation, ultimately significantly improving the matching accuracy of the anchor response predictor for bubble defects caused by environmental humidity fluctuations in practical applications.

[0050] S1.4: The typical controllable process parameter set, sensitivity factor set and statistical correlation pattern matrix are encapsulated and processed to construct a process semantic anchor data object containing both physical mechanism constraints and data statistical characteristics, so as to realize the instantiation transformation from abstract process to computable reasoning unit.

[0051] S1.5: Based on the order of the initial process flow topology, all generated process semantic anchor data objects are serialized, arranged, and indexed for registration. The output is a sequence of process semantic anchors encapsulating complete process knowledge and defect mapping relationships, which serves as the sole input source for the subsequent anchor response predictor to perform defect feature matching and probability distribution calculation.

[0052] Step S2: Input the multidimensional defect feature vector, composed of the bubble pore spatial density heat map, pore size spectrum, and pore wall microcrack features extracted from the grinding wheel CT scan image, into the pre-trained anchor response predictor. Based on the process semantic anchor sequence, generate an anchor activation probability distribution representing the significant activation probability of each process. Specifically, this includes: S2.1: Vector splicing and normalization are performed on the bubble pore spatial density heat map, pore diameter spectrum and pore wall microcrack features extracted from the CT scan image of the grinding wheel to generate a standardized multidimensional defect feature vector, which serves as the initial input data for the anchor point response predictor.

[0053] S2.2: Based on the statistical correlation patterns in the process semantic anchor sequence constructed by the preceding main steps, the anchor response predictor model parameters are trained and solidified using a deep neural network architecture to generate a pre-trained anchor response predictor with process semantic mapping capabilities, and to establish a probabilistic mapping benchmark from features to processes.

[0054] Based on the statistical association patterns in the process semantic anchor point sequence constructed in the preceding main steps, the statistical association pattern matrix is ​​used as the label mapping benchmark for training samples. Through data labeling, a one-to-one correspondence is established between each multi-dimensional defect feature vector and its corresponding process semantic anchor point probability distribution. Data segmentation is performed on the labeled training sample set, dividing it into training, validation, and test sets to ensure the generalization ability and validation accuracy during model training. A deep neural network architecture is adopted, defining the number of nodes in the input layer consistent with the dimension of the defect feature vector to ensure the matching between the input features and the network structure. Multiple hidden layers are configured, with the number of nodes and activation function type (e.g., ReLU, Sigmoid) in each layer determined based on the number of process semantic anchor points and the complexity of the statistical association patterns. The mapping relationship between high-order features and process semantics is gradually extracted through the hidden layers. Backpropagation is introduced for gradient calculation and weight update. The Adam optimizer is used with an adaptive learning rate to improve convergence speed, and a difference measure term between the statistical association pattern matrix and the loss function is added. The loss value is calculated using the following formula:

[0055] Where N is the total number of samples, The anchor point activation probability predicted by the model. To determine the true probability values ​​in the statistical association pattern matrix, the accuracy and loss changes of the model across different training epochs are evaluated using a validation set. An early stopping mechanism is implemented to prevent overfitting, and the performance of the final model is tested on a test set. The trained network weight parameters are permanently stored in the parameter file of the anchor response predictor, ensuring that the input features in subsequent inference processes can be stably mapped to the process probability space. Through the above processing method, the process semantic anchor sequence from the previous step is transformed into a pre-trained anchor response predictor with the ability to map features to process probabilities, achieving accurate probabilistic association between defect features and process nodes.

[0056] For example, for a batch of grinding wheel production data, the statistical association pattern matrix contains probability labels for 5 process semantic anchor points, each with a dimension of 5. The defect feature vector, after normalization, has a dimension of 128. The input layer has 128 nodes, the hidden layers have 64 and 32 nodes respectively, using the ReLU activation function, and the output layer has 5 nodes, using Softmax normalization. The 2000 samples are divided into training, validation, and test sets in an 8:1:1 ratio. The Adam optimizer is set with a learning rate of 0.001, a batch size of 32, and 50 training iterations. The loss function is calculated using the squared error formula mentioned above. The validation set accuracy stabilizes after 30 iterations, triggering an early stopping mechanism. The solidified model weight file size is 2.3MB. After deployment to the anchor response predictor, the average loss value on the test set is 0.012, and the deviation between the predicted anchor activation probability and the true label is significantly reduced, meeting the process semantic mapping accuracy requirements.

[0057] S2.3: The generated standardized multidimensional defect feature vector is input into the input layer of the pre-trained anchor response predictor. Through nonlinear feature transformation of the hidden layer, a hidden activation tensor representing the potential response intensity of each process semantic anchor point is generated.

[0058] The normalized multidimensional defect feature vector is loaded into the input layer neural node set of the pre-trained anchor response predictor to ensure that the row and column mapping relationship between each feature component and the input weight matrix is ​​aligned according to the index order fixed during the model training stage.

[0059] The first layer of fully connected operators is performed on the input vector, and the pre-trained parameter matrix stored in the model weight library is called. The matrix product is added to the bias vector to generate a preliminary weighted sum, which serves as the original input signal of the hidden layer.

[0060] The original input signal is processed by a nonlinear activation function, and the nonlinear transformation of the feature response is achieved by using the modified linear unit (ReLU) or the hyperbolic tangent (tanh) function. The role of the nonlinear transformation is to introduce high-order interaction capabilities between multidimensional features in order to capture the complex mapping relationship between the bubble defect feature pattern and the process semantic anchor point.

[0061] Batch normalization is performed on the hidden layer output after nonlinear transformation. The mean and variance of the batch statistics are used to standardize the scale of each node output, eliminate the dimensional differences between feature components, and improve the stability of the model in inference in different batches.

[0062] A multi-head attention mechanism is adopted for the hidden layer output. The normalized feature vector is calculated by multiple independent attention weights to calculate the attention intensity of each of the process semantic anchors. The outputs of all heads are concatenated or weighted and fused to generate a hidden layer activation tensor that represents the potential response intensity from multiple perspectives.

[0063] Through the matrix operations, nonlinear transformations, normalization, and attention fusion processing described above, the standardized defect feature vector from the previous step is transformed into a hidden activation tensor that can be directly input into the soft maximum normalization operation, thereby realizing the construction of a feature mapping structure from the original feature domain to the process response domain.

[0064] For example, the spatial density heatmap of bubble pores, pore size spectrum, and microcrack features of the pore wall extracted from a batch of grinding wheel CT scans are normalized in step S2.1 to form a feature vector of length 128. The number of neural nodes in the input layer is configured to be 128, corresponding to a weight matrix size of 128×256, and a bias vector length of 256. The input vector and the weight matrix are multiplied and the bias is added to generate the original hidden layer signal of length 256. After being transformed by the ReLU function, it enters the batch normalization module to calculate the mean μ and variance σ² of the small batch, and standardizes it using the following formula:

[0065] Where x is the output value of the hidden layer node, μ is the mean of the node in the current mini-batch, and σ is the standard deviation of the node in the current mini-batch. The normalized output is calculated using attention scores from four multi-head attention modules. The attention weights are generated by scaling and softmaxing the dot product of the query vector and the key vector. The output vector of each head is 64 bytes long. The outputs of the four heads are concatenated into a fusion vector of length 256 to form the hidden layer activation tensor, which is used for the soft maximum normalization operation in S2.4. Under this configuration, the output stability of the hidden layer activation tensor on the test set is significantly improved, and the differences in the predictor's response strength to different bubble defect patterns are more clearly defined.

[0066] S2.4: Perform soft maximum normalization operation on the generated hidden layer activation tensor based on the process semantic anchor sequence dimension to generate the anchor activation probability distribution representing the significant activation probability of each process, thus completing the final mapping from the continuous feature space to the discrete process probability space.

[0067] like Figure 2 As shown, step S3: Based on the anchor point activation probability distribution, high-response process semantic anchor points are selected as the starting point for causal inference. For each high-response process semantic anchor point, its bound local causal strength evaluator is called. Based on the actual parameter values ​​of the current batch, equipment status log fragments, and material batch attributes, a local causal contribution score is calculated to generate the marginal contribution of the characterization parameters to bubble formation. Specifically, this includes: S3.1: Perform threshold truncation and sorting on the anchor point activation probability distribution to filter out the top set of process semantic anchor points whose probability values ​​exceed the preset confidence threshold, generate a list of high-response process semantic anchor points that represent the starting position of causal reasoning, and establish the target process range for subsequent local evaluation.

[0068] The anchor point activation probability distribution data matrix is ​​loaded and processed. The preset confidence threshold retrieval module is called to compare and calculate the probability values ​​of each process semantic anchor point in the matrix, and generate a threshold judgment vector containing Boolean state markers.

[0069] The threshold determination vector is input to the descending sorting control unit. The anchor point entries that meet the threshold condition are rearranged according to the probability value sorting logic, and the anchor point index list is output in descending order of probability value.

[0070] For the sorted list of anchor indexes, a list truncation algorithm is executed to truncate the first K probability value entries according to the preset top set size parameter K, while retaining the corresponding anchor identifiers.

[0071] A list of high-response process semantic anchors is constructed using the truncated set of anchor identifiers, which establishes the target process range for subsequent calls to the local causality strength evaluator.

[0072] By using threshold comparison and sorting truncation, the probability distribution results from the previous step are transformed into a list of high-response process semantic anchors that characterize the starting position of causal reasoning, thereby achieving precise positioning of the reasoning starting point and optimization of computational load.

[0073] For example, in a grinding wheel defect analysis, the anchor point activation probability distribution is [mixing: 0.32, forming: 0.45, drying: 0.82, sintering: 0.76, finishing: 0.28], the confidence threshold is set to 0.50, and the top set quantity parameter K is 3. The threshold decision vector calculation result is [False, False, True, True, False], and the sorting units are rearranged according to the probability values ​​to obtain the sequence [drying: 0.82, sintering: 0.76]. Based on K=3, the truncation algorithm retains the two items [drying, sintering] to construct a list of high-response process semantic anchor points. If a numerical calculation formula is used, the threshold comparison process can be described as follows:

[0074] in, For the selected set of high-response anchor points, The probability distribution function corresponding to the anchor index. The confidence threshold is T. In this example, T=0.50, and the final output significantly improves the relevance of subsequent local causal assessments.

[0075] S3.2: Based on each anchor identifier in the high-response process semantic anchor list, extract the actual process parameter value sequence, equipment status log fragments, and material batch attribute data corresponding to the current production batch, perform time window alignment and missing value interpolation processing, and generate a standardized multi-source heterogeneous operating condition input vector as the real-time data base for the local causal strength evaluator.

[0076] Under each anchor identifier in the high-response process semantic anchor list, the actual process parameter value sequence of the current production batch is obtained, covering continuous monitoring signals such as temperature, pressure, humidity, and mixing ratio. Simultaneously, equipment status log fragments associated with the batch are extracted, including equipment operation mode, alarm event records, shutdown and maintenance timestamps, to ensure that all dynamic factors that may affect bubble formation are covered.

[0077] The actual process parameter value sequence and equipment status log fragments are aligned with the time windows according to the start and end times of the batches to establish a unified time series index, so that each data source is comparable at the same time scale, and the original sampling frequency information is preserved in the alignment process so that the subsequent model can accurately identify the operating condition fluctuation characteristics.

[0078] Receive material batch attribute data, including static information such as raw material type, particle size distribution, moisture content, and chemical composition ratio. After time window alignment, map it to a fixed time period of the corresponding production batch to supplement the material background layer of the working condition data.

[0079] For multi-source data aligned to time windows, missing value imputation is performed using an imputation algorithm that combines historical pattern matching with physical constraints. The algorithm fills in missing values ​​by utilizing the trend extension of adjacent time periods and physical formula extrapolation. For example, when temperature is missing, the estimated value is calculated based on the heat balance equation.

[0080] Where Q is the input heat, and T is the input heat. prev The value is the temperature at the previous valid moment, Δt is the time interval, m is the material mass, and c is the specific heat capacity. This ensures that the interpolation value is physically reasonable and consistent with the trend of the operating condition.

[0081] The imputed multi-source heterogeneous data is then normalized, and continuous parameters are normalized using Z-score normalization.

[0082] Where x is the original parameter value, μ is the mean of the parameter in the batch, and σ is the standard deviation; one-hot encoding mapping is used for discrete or categorical parameters to ensure that different types of parameters can be directly input into the model in the same numerical space.

[0083] The normalized time series parameter vector, equipment status encoding vector, and material property encoding vector are concatenated by time step to form a standardized multi-source heterogeneous operating condition input vector. A corresponding input vector instance is generated separately for each high-response process semantic anchor point to serve as the real-time data base for the local causal strength evaluator.

[0084] Through the above processing method, the original multi-source heterogeneous data of the high-response process anchor point is transformed into an input vector with a unified time scale, complete numerical coverage, reasonable physical interpolation and standardized performance, so as to realize the accurate calculation of the marginal contribution of the parameters of the subsequent model under multi-source information.

[0085] For example, in a production batch, the real-time temperature data sampling frequency during the mixing stage is 1Hz, with two consecutive missing sampling points. The equipment status log shows that the mixer was in heating mode during this period. The material batch attributes record the raw material mass as 50kg, the specific heat capacity c as 1.2kJ / (kg·K), the input heat Q as 6kJ, and the time interval Δt as 2s. The interpolated temperature value is calculated using the heat balance formula: T = Tprev + (Q×Δt) / (m×c), that is, when Tprev=45.0°C, the interpolated value is 45.0 + (6×2) / (50×1.2) ≈ 45.2°C. Performing Z-score normalization on the temperature sequence, the batch mean μ=46.5°C, and the standard deviation σ=0.8°C, then the normalized value of the current data point is (45.2-46.5) / 0.8≈-1.625. Equipment status is mapped to [0,1,0] using unique thermal encoding, and material properties are encoded as [0.35,0.45,0.20] to represent different component proportions. Normalized temperature, pressure, humidity, equipment status encoding, and material property encoding are concatenated over time steps to form a standardized multi-source heterogeneous operating condition input vector for a single high-response process anchor point. This vector is then used for calculation by the local causal strength estimator. Validation results show a significant improvement in the temporal consistency and physical interpolation stability of the input vector, ensuring higher reliability of the causal contribution score output by the model in this batch.

[0086] S3.3: Utilize a pre-trained local causal strength evaluator to receive standardized multi-source heterogeneous operating condition input vectors, perform neural-symbolic hybrid inference operations, where the neural network part extracts nonlinear feature interaction patterns, and the symbolic logic part embeds physical constraint rules to generate a set of original marginal causal contribution scores that characterize the degree of independent influence of each controllable process parameter on bubble formation.

[0087] The local causal strength evaluator consists of a hybrid neural-symbolic architecture. Its neural network module employs a three-layer fully connected design. The input layer dimension matches the feature dimension of the multi-source heterogeneous operating condition input vector. The number of hidden layer nodes is determined to be 128 and 64 based on the order of magnitude of the process parameters and the interaction complexity. The number of output layer nodes equals the number of typical controllable process parameters to be evaluated, and a sigmoid activation function is used to map the output to initial independent influence scores. The symbolic reasoning module embeds a lightweight knowledge graph containing physical constraints and logical relationships between process parameters, such as equipment interlocks and material balance. A graph attention network is used to constrain and correct the initial scores output by the neural network, verifying logical consistency. The evaluator is trained under supervised supervision using multi-source operating condition data from historical normal and defective batches, along with expert-annotated causal contribution benchmarks. The training objective is to minimize the mean squared error between the predicted score and the benchmark score. Combined with the logical constraint loss term of the knowledge graph, a reliable reasoning model is ultimately solidified, capable of receiving aligned multi-source input vectors and directly outputting the original set of marginal causal contribution scores.

[0088] Based on the standardized multi-source heterogeneous operating condition input vector generated in step S3.2, the weights of the pre-trained local causal strength estimator model are loaded into the neural symbolic hybrid architecture of the inference engine to achieve parallel feature extraction and physical constraint logic operations on the input vector. The multi-source heterogeneous input data is processed through a combination of multi-layer fully connected and convolutional units in the neural network part to extract nonlinear feature interaction patterns. High-order correlation features between different controllable process parameters are captured at the output of each hidden layer node, forming a feature tensor mapping relationship. The symbolic logic part is invoked to group the feature tensors according to parameter categories and embed them into the physical mechanism constraint rule set of grinding wheel manufacturing. Logical exclusion and correction are performed on feature combinations that violate thermodynamic boundaries or kinetic windows to ensure that the causal inference results conform to the physical feasibility of the actual process. After the coupled operation of neural feature extraction and symbolic physical verification is completed, the independent marginal causal contribution score for bubble formation is calculated for each controllable process parameter. The value is quantified using a calculation formula based on mutual information and linear regression residuals, i.e.:

[0089] Among them I pLet denot p be the marginal causal contribution score, MI be the mutual information function, D be the defect feature set, RSS be the sum of squared regression residuals, and n be the number of historical samples for that parameter. In the formula calculation, the mutual information component quantifies the nonlinear dependence between the parameter and defect features, while the regression residual term characterizes the interpretable linear contribution of parameter deviation to the defect under physical constraints. Through this neural symbolic hybrid processing method, the standardized multi-source operating condition vector from the previous step is transformed into a set of original marginal causal contribution scores with physical interpretability and statistical significance, enabling a quantitative assessment of the degree of independent influence of parameters under high-response process semantic anchors.

[0090] For example, in the production of a certain batch of grinding wheels, the high-response process semantic anchor is the sintering process. The standardized multi-source heterogeneous working condition input vector includes the actual heating rate of 4.5℃ / min, the holding time of 2.1 hours, the oxygen partial pressure in the furnace of 0.008MPa, and the fan speed fluctuation data in the equipment status log. This vector is input into a local causality strength evaluator. The neural network part extracts the nonlinear interaction features between the heating rate and the holding time, while the symbolic logic part applies the sintering densification kinetics window rule to exclude infeasible combinations where the heating rate exceeds 5℃ / min and the holding time is less than 2 hours. The mutual information calculation result MI(p,D) is 0.462 corresponding to the heating rate, the regression residual sum of squares RSS(p) is 1.93, and the number of historical samples n is 50. Substituting the values ​​into the formula, the marginal causal contribution score of the heating rate is obtained as 0.501. The calculation process for the holding time and oxygen partial pressure is similar. The parameter contributing the highest score in this batch was the heating rate. The evaluation results showed that its independent physical influence on bubble formation was significantly improved, providing a highly reliable original numerical basis for subsequent identification and correction of parameter coupling effects.

[0091] S3.4: Execute the parameter coupling effect identification algorithm on the original set of marginal causal contribution scores, calculate the interaction gain coefficient between different parameter combinations, correct the independent contribution of a single parameter to the joint contribution value including interaction terms, and generate a corrected local causal contribution score matrix that characterizes the risk of parameter cooperative failure.

[0092] The original set of marginal causal contribution scores is received as input conditions. The set is a quantitative indicator of the independent effect of each controllable process parameter on the formation of bubble defects under the current high-response process semantic anchor point.

[0093] Construct a parameter combination enumeration set to include all possible binary, ternary, and higher-order combinations, ensuring that the entire combination space is covered during interactive recognition.

[0094] For each set of parameters in the enumeration set, extract its corresponding value in the original score set, and perform covariance matrix construction operation to quantify the statistical dependencies between combinations.

[0095] The mutual information calculation module is invoked to measure the difference between the joint distribution and independent distribution of each parameter combination and to generate the basic interaction gain coefficient.

[0096] The interaction correction score is obtained by performing element-wise matrix multiplication on the interaction gain coefficient and the original marginal score, as shown in the following formula:

[0097] Among them, B i B j For the original independent contribution score, G i,j This is the interaction gain coefficient.

[0098] Perform multi-order recursive correction, using the binary correction result as input to higher-order combinations, iteratively generate a joint contribution matrix containing all interaction terms, and mark the high-risk regions corresponding to cooperative failure modes in the matrix.

[0099] By using the above processing method, the original set of marginal causal contribution scores from the previous step is transformed into a modified local causal contribution score matrix, thereby quantifying the synergistic effect of parameters and identifying potential synergistic failure risks.

[0100] For example, in a batch of grinding wheel sintering process data, the original marginal contribution scores are B(heating rate) = 0.35, B(holding time) = 0.28, and B(oxygen partial pressure) = 0.22. A ternary combination {heating rate, holding time, oxygen partial pressure} is constructed, and the mutual information values ​​G(heating rate, holding time) = 0.15, G(heating rate, oxygen partial pressure) = 0.08, and G(holding time, oxygen partial pressure) = 0.11 are calculated. The corrected scores for the binary combination are calculated using a formula; for example, the combination of heating rate and holding time C = 0.35 + 0.28 + 0.15 = 0.78. This corrected binary score is used as the basic input for the ternary combination, and the independent contribution of the third parameter and the corresponding mutual information are superimposed to generate a joint contribution matrix value of 0.78 + 0.22 + 0.11 = 1.11. In this embodiment, the C value in the matrix increases significantly, indicating a significant synergistic effect between the heating rate and the holding time, which has a high risk impact on defect formation. This high-risk combination is marked for subsequent causal propagation path generation and physical constraint verification.

[0101] S3.5: Based on the corrected local causal contribution score matrix, perform normalization mapping processing, and combine it with the predefined sensitivity factor weights within the process semantic anchor point for weighted fusion to generate the final local causal contribution score that represents the marginal contribution of each parameter combination to the formation of bubble defects under the current high-response process semantic anchor point, thus completing the conversion from operating data to quantitative attribution index.

[0102] like Figure 3As shown, step S4: Identify anomalous combinations of parameter couplings in the local causal contribution scores, and, in conjunction with the process flow, propagate the local causal contribution scores along the process chain to the preceding process semantic anchors with clear physical interpretation paths, generating an initial causal propagation path set containing cross-process causal effect propagation paths. Specifically, this includes: S4.1: Perform multi-dimensional parameter correlation matrix construction on the set of local causal contribution scores output by the semantic anchor point of the current high-response process to identify abnormal parameter coupling combinations with statistically significant deviations and generate a set of parameter coupling anomaly identifiers that characterize parameter interaction failure modes.

[0103] The system receives the corrected local causal contribution score matrix corresponding to the high-response process semantic anchor point as the input data basis, and maps the row and column indices of the matrix to a preset set of controllable process parameters. A multi-dimensional parameter correlation matrix construction process is performed on the mapped matrix, generating a three-dimensional correlation tensor between the various row and column combinations using multiple correlation measures such as Pearson correlation coefficient, mutual information entropy, and partial correlation coefficient. Based on the Pearson correlation coefficient value of each parameter combination in the correlation tensor, the significance test module is called to calculate its significance level p, and the nonlinear correlation relationship is supplemented by the mutual information entropy value, marking statistically significant deviations in the judgment results. For each significant deviation combination, the parameter interaction gain coefficient is derived, selecting combinations with high correlation and deviation directions consistent with defect characteristics, and calculating the interaction gain coefficient according to the following formula:

[0104] in Represents mutual information value, The Pearson correlation coefficient is represented. Combinations with gain coefficients exceeding a preset threshold are marked as parameter coupling anomalies and labeled with a defect triggering mode. Through the above processing method, the local causal contribution score matrix of the previous step is transformed into a parameter coupling anomaly identifier set characterizing parameter interaction failure modes, thereby achieving accurate identification of the starting point of cross-process causal propagation.

[0105] S4.2: Based on the parameter coupling anomaly identifier set and the predefined grinding wheel manufacturing process specification topology, perform the preceding process reachability analysis operation to filter out the preceding process semantic anchor point sequence that has a clear physical interpretation path with the current anomaly combination, and generate a candidate causal transmission target anchor point list.

[0106] S4.3: The local causal contribution score is weighted and mapped using a preset causal effect attenuation function to simulate the energy dissipation characteristics of the physical mechanism in the cross-process transmission process, and to generate a cross-process causal effect intensity value carrying a distance attenuation factor.

[0107] S4.4: Based on the candidate causal transmission target anchor point list, the cross-process causal effect intensity value is recursively assigned along the process flow direction to establish a causal connection relationship from the current abnormal anchor point to the previous related anchor point, and generate a unidirectional causal transmission link containing node causal intensity information.

[0108] The system receives the cross-process causal effect strength value carrying the distance decay factor and the candidate causal transmission target anchor point list as input from the output of sub-step S4.3. It establishes the recursive calculation basis for causal backtracking for each reachable combination of anomaly anchor point and preceding anchor point.

[0109] The current abnormal anchor point identifier is indexed and matched with the candidate preceding anchor point sequence. Based on the directed adjacency matrix of the grinding wheel manufacturing process specification, the set of direct predecessor anchor points for each recursive backtracking is determined, and the corresponding values ​​of the cross-process causal effect intensity are assigned.

[0110] When performing recursive assignment of causal strength for each direct predecessor anchor point, a path cumulative attenuation coefficient is introduced, using the following mapping formula:

[0111] in This represents the current cross-process causal effect strength value. This is the process-related attenuation coefficient. This represents the process distance from the abnormal anchor point to the predecessor anchor point.

[0112] During the recursive process, the updated causal effect strength value is written to the causal strength record of the predecessor anchor point, and the bidirectional binding between the node causal strength information and the path index is maintained to ensure the traceability of the link.

[0113] For anchor points with multiple precursor paths, a superposition and fusion process of causal strength values ​​is performed. The superposition operation must satisfy the cumulative effect of parallel influence of physical interpretation paths, and the following calculation formula is used:

[0114] in This represents the intensity of the causal effect corresponding to different precursor paths.

[0115] The causal strength value obtained from the above fusion is used as the anchor point for backtracking to a previous order until the backtracking depth limit specified in the process specification is reached or there are no more physically reachable paths.

[0116] By recursively backtracking assignment and superimposed fusion processing, the causal effect transmission chain of the current abnormal anchor point is transformed into a unidirectional causal transmission link containing node causal strength information, thereby realizing the quantitative diffusion of abnormal impact to physically reachable preceding processes.

[0117] S4.5: Aggregate and encapsulate all generated unidirectional causal transmission links to form a complete propagation network structure covering multi-source anomaly trigger points, and generate an initial causal propagation path set containing cross-process causal effect propagation paths.

[0118] The causal strength index of the nodes and the process semantic anchor point identifier are loaded into the set of unidirectional causal transmission links. The links of the same source abnormal trigger points are arranged according to the anchor point time sequence and the node sequence merging operation is performed to form a basic link group that is continuously connected across processes.

[0119] The basic link group is processed by splicing node causal strength vectors. The spliced ​​vectors are then weighted and transformed using a preset process flow topology weight matrix to generate a superposition value of causal effects representing the multi-source anomaly trigger points across the entire process range.

[0120] A connectivity optimization algorithm under the process adjacency matrix constraint is applied to the superimposed value matrix to eliminate isolated node combinations or those that do not meet the process connectivity conditions and readjust the link direction to ensure the integrity and reachability of the propagation network structure.

[0121] Causality strength normalization is performed on the connectivity-optimized link set, and the normalized causal contribution score for each link is calculated using the following formula:

[0122] in, To contribute scores to normalized causal relationships, The original causal contribution value of the link is given by the denominator, which is the sum of the contribution values ​​of all links.

[0123] Path index registration is performed on the normalized causal contribution score matrix. The starting abnormal anchor point of each link, the preceding anchor point of the end point, and the intermediate process nodes are used to form a path index table and bind and superimpose causal indicators.

[0124] By using an aggregation and encapsulation process, the unidirectional transmission link result from the previous step is transformed into a propagation network structure that covers multiple sources of anomaly triggering points and has global node causal strength information, thereby realizing the construction of the initial causal propagation path set and the binding of quantitative causal indicators.

[0125] Step S5: Based on preset physical constraint rules, perform process fidelity attenuation verification on each path in the initial causal propagation path set, eliminate physically infeasible paths that violate the upper limit of drying temperature or skip necessary processes, and generate a set of valid causal propagation paths verified by physical constraints. Specifically, this includes: S5.1: Obtain the process semantic anchor sequence and corresponding local causal contribution score of each path in the initial causal propagation path set. Based on the process adjacency matrix in the grinding wheel manufacturing process specification, perform process continuity verification on the process semantic anchor sequence to identify and mark abnormal process jump paths that skip necessary processes.

[0126] S5.2: Extract the actual process parameter values ​​bound to each process semantic anchor point in the path to be verified outside the abnormal process jump path, and use a preset physical mechanism constraint knowledge base to perform thermodynamic boundary matching processing on the actual process parameter values ​​to generate a parameter compliance status identifier containing the upper limit of drying temperature violation criteria.

[0127] For the set of paths to be verified after removing abnormal paths with skipped processes following the continuity verification in step S5.1, the set of actual process parameter values ​​bound to each process semantic anchor point in the path is called to form a parameter data mapping with process node index.

[0128] The parameter data mapping is performed based on the time sequence relationship of the process flow to analyze the thermodynamic boundary conditions, clarify the physical meaning and control range of each parameter, and match the analysis results with the boundary rules of the corresponding process in the pre-set physical mechanism constraint knowledge base.

[0129] For process nodes with temperature parameters, a drying temperature upper limit criterion is constructed during the boundary matching process. The difference between the actual drying temperature value and the upper limit threshold specified in the knowledge base is calculated, and a criterion result identifier is generated.

[0130] The thermodynamic boundary matching method is adopted to generate the parameter compliance status value of the process node by calculating the normalized deviation ratio between the actual parameter value and the upper and lower limits of physical constraints. A status value lower than zero indicates a lower limit violation, and a value higher than one indicates an upper limit violation.

[0131] The compliance status of path-level parameters is synthesized by using the compliance status value of the criterion result identifier and the normalized deviation ratio. The comprehensive parameter compliance result of each path to be verified is output as a set of status identifiers containing the violation criterion of the upper limit of drying temperature.

[0132] Through the boundary matching process described above, the path to be verified in the previous step is transformed into data indicators that can judge the compliance of physical parameters, thereby realizing the automatic detection function for violations of drying temperature in the physical constraint verification of the path.

[0133] For example, in a certain batch of grinding wheel manufacturing processes, the path to be verified includes two consecutive process nodes: drying and sintering. The actual temperature parameter value of the drying process is 720, the upper limit of the drying temperature specified by the material batch attributes is 700, and the physical constraint rules of the drying process in the knowledge base set the upper limit threshold of the drying temperature to 710. Thermodynamic boundary matching processing is performed on the drying process nodes, and the criterion formula for temperature parameters is:

[0134] Where T is the actual drying temperature value, T max This represents the upper limit of the drying process temperature. Substitute the actual value...

[0135] The calculated ΔT is 10, indicating a violation of the upper limit of the drying temperature at this node. According to the normalized deviation ratio formula:

[0136] Where T min The upper limit of the drying process temperature is set to 650. Substituting the numerical value, R is 0.875, which does not meet the condition of being higher than 1, but the criterion of ΔT>0 has triggered a violation flag. The final output path-level parameter compliance status is marked as "Drying temperature upper limit violation". In the subsequent S5.3 physical feasibility filtering, this path is determined to be physically infeasible. This embodiment shows that the method can significantly improve the accuracy of drying temperature violation detection and ensure the physical validity of path screening, avoiding bias introduced by physical violations in subsequent root cause reasoning.

[0137] S5.3: Based on the parameter compliance status identifier and the marking results of the abnormal process jump path, perform a physical feasibility filtering operation on the initial causal propagation path set to eliminate physically infeasible paths that violate the upper limit of drying temperature or have process jump phenomena and generate a preliminary candidate causal propagation path set.

[0138] The parameter compliance status identifier is parsed to establish a mapping index between each initial causal propagation path and its corresponding parameter compliance assessment result.

[0139] The marking results of abnormal paths with process jumps are compared one by one with the parameter compliance assessment results to form a composite path screening matrix that includes path identification, process continuity status, and parameter compliance status.

[0140] In the composite path screening matrix, physical feasibility determination rules are defined. Records with abnormal path continuity status or non-compliant parameter compliance status are directly marked as infeasible paths. The determination rules are implemented through Boolean logic operations, in the following form: The continuity and parameter compliance status values ​​are set to 1 to indicate success and 0 to indicate failure.

[0141] During the execution of the judgment rules, the parameter compliance status identifier and the abnormal path status of the process jump are filtered simultaneously to form a set of candidate paths after eliminating the illegal paths.

[0142] The candidate path set is indexed, reconstructed, and encapsulated to generate a preliminary set of candidate causal propagation paths, which serves as the input for subsequent calculation of the process fidelity attenuation coefficient.

[0143] Through the above filtering operation, the path sequence obtained in the previous step is transformed into a dataset containing only physically constrained feasible paths, thus achieving the expected technical effect of path filtering and credibility improvement.

[0144] For example, in the defect analysis of a batch of grinding wheels, the initial causal propagation path set contains 10 paths. Three paths exhibit process jump anomalies, with a continuity status of 0; two paths have drying stage temperatures exceeding the upper limit of 120℃, resulting in a parameter compliance status of 0; the remaining five paths have both continuity and parameter compliance status values ​​of 1. In the composite path screening matrix, physical feasibility judgment rules are defined, and Boolean logic operations automatically eliminate paths with a continuity or parameter compliance status of 0, ultimately generating five preliminary candidate causal propagation paths. In this scenario, assuming that each path in the candidate path set has a span length of 2 processes corresponding to the subsequent process fidelity attenuation coefficient calculation, and a physical correlation strength of 0.85, the confidence baseline value is calculated using the attenuation function before sorting, significantly improving the accuracy and interpretability of defect attribution analysis.

[0145] S5.4: For each path in the candidate causal propagation path set, calculate the process fidelity attenuation coefficient based on the path span length and the strength of the physical association between processes. Use the process fidelity attenuation coefficient to perform a step-by-step attenuation correction process on the local causal contribution score on the path to generate an attenuated causal contribution score that characterizes the credibility of cross-process causal effects.

[0146] S5.5: Summarize all candidate causal propagation paths that have undergone attenuation correction and their corresponding attenuated causal contribution scores, and perform final validity confirmation on the attenuated causal contribution scores based on a preset effective path determination threshold, so as to generate a set of effective causal propagation paths that have been verified by physical constraints and have quantitative credibility indicators.

[0147] The decayed causal contribution score matrix based on the candidate causal propagation path set is used as the input data object for this step. The preset effective path determination threshold model loading module is called to initialize the path validity determination parameters.

[0148] Each path in the attenuated causal contribution score matrix is ​​traversed path by path to extract three key indicators: path span, physical correlation strength of processes, and attenuation coefficient, which constitute a comprehensive evaluation vector. This comprehensive evaluation vector is then compared with the effective path judgment threshold to generate a preliminary validity status identifier.

[0149] For the set of paths initially identified as valid, a credibility metric mapping is performed. The attenuated causal contribution score is then calculated using a predefined credibility normalization formula, which is:

[0150] Where S is the causal contribution score after attenuation, R is the correction coefficient for the physical correlation strength of the process, and T is the path span length.

[0151] The normalized confidence value and the process fidelity attenuation coefficient are weighted and fused twice to form a path confidence index set containing the final confidence assessment value. The path index information is then combined to construct a traceable data structure.

[0152] A global filtering operation is performed on the path credibility index set to remove path entries with credibility values ​​lower than the effective path determination threshold, thereby obtaining a set of effective causal propagation paths that have been verified by physical constraints and have quantified credibility indicators.

[0153] By using validity verification and credible quantification methods, the decayed causal contribution score matrix from the previous step is transformed into a set of valid causal propagation paths that have been verified by physical constraints and have quantitative credibility, thus providing a high-precision data foundation for subsequent root cause credibility ranking.

[0154] For example, in the analysis of a batch of grinding wheel production, the candidate causal propagation path set contains 12 paths. The causal contribution score of each path after attenuation is calculated to be between 0.35 and 0.87, the path span length ranges from 2 to 5 process nodes, and the process physical correlation strength correction coefficient is between 1.2 and 0.8. The effective path judgment threshold is set to 0.4. The system traverses each path and calculates the confidence value using a formula. For example, if a path has S = 0.65, R = 1.1, and T = 3, the calculated confidence value is 0.17875. After weighting with a process fidelity attenuation coefficient of 0.9, the final confidence value is 0.160875. Based on threshold filtering, only paths with confidence values ​​greater than 0.4 are retained, and finally, 4 effective causal propagation paths verified by physical constraints are output as input for subsequent root cause confidence ranking. The verification results show that this filtering process can significantly improve the stability and interpretability of root cause ranking.

[0155] Step S6: Summarize the causal scores at the endpoints of the effective causal propagation path set, and perform a weighted fusion operation by combining the process weights, parameter coupling confidence levels, and historical verification frequencies of the process semantic anchors along each path to generate a normalized root cause ranking list representing root cause credibility. Specifically, this includes: S6.1: Based on the sequence of process semantic anchor points traversed by each path in the set of effective causal propagation paths, extract the preset process weight coefficients corresponding to each process semantic anchor point, and perform a multiplication or weighted summation operation on the preset process weight coefficients to generate a path process weight factor that characterizes the overall process importance of the path.

[0156] S6.2: Using the path process weight factor as the basic input, and combining it with the parameter coupling confidence value corresponding to the parameter coupling anomaly combination identified in the local causality strength evaluator, the path process weight factor is dynamically corrected to generate a path comprehensive basic score that incorporates parameter interaction uncertainty.

[0157] After receiving the parameter coupling confidence value of the abnormal parameter coupling combination identified in the path process weight factor and the local causality strength evaluator, the path process weight factor is used as the basic input signal to enter the correction calculation module.

[0158] Vectorization of the basic input signal ensures that the path process weight factor and parameter coupling confidence can be fused in the same numerical space.

[0159] The parameter interaction uncertainty correction function is invoked to perform a step-by-step mapping process on the path process weight factor and the corresponding parameter coupling confidence level in a multiplicative scaling manner, forming an intermediate correction vector containing the interaction risk adjustment factor.

[0160] A weighted accumulation strategy is adopted to fuse the elements in the intermediate correction vector. The weighting coefficient depends on the preset weight of the parameter to the sensitivity factor of defect formation, so as to ensure that the influence of different parameters is reflected in the comprehensive score according to the proportion of physical mechanism.

[0161] Numerical smoothing is performed on the fusion results, and high-frequency fluctuations caused by local anomalies are eliminated by Gaussian filtering or exponential moving average to generate the final path comprehensive base score.

[0162] Through the above dynamic correction process, the path process weight factor of the previous step is transformed into a comprehensive basic score of the path that incorporates the uncertainty of parameter interaction, thereby realizing the quantitative characterization of parameter coupling risk in the root cause credibility calculation chain.

[0163] For example, a path containing sintering and drying processes is selected from the set of effective causal propagation paths. The process weight factor for this path is 0.72, and the local causal strength evaluator identifies a parameter coupling confidence level of 0.65 for the combination of "high-temperature rapid firing + low viscosity". The 0.72 and 0.65 are multiplicatively corrected: 0.72 × 0.65 yields an intermediate value of 0.468. Assuming a sensitivity factor weight of 0.6 for the sintering stage and 0.4 for the drying stage, a weighted cumulative processing is performed on the intermediate value: 0.468 × 0.6 + 0.468 × 0.4, resulting in 0.468. Applying an exponential moving average smoothing function with a window length of 3 to this value, the final output path comprehensive base score is 0.465. This value is used as the base score input for weighted historical verification frequency in subsequent S6.3. Verification results show that this correction significantly improves the position of high-interaction-risk paths in the ranking, enhancing the sensitivity to process anomalies in root cause analysis.

[0164] S6.3: Based on the comprehensive basic score of the path, retrieve historical verification records that match the characteristics of the path from the historical defect sample library, count the frequency of historical verifications and convert them into time decay coefficients, apply the time decay coefficients to the comprehensive basic score of the path for weighted adjustment, and generate an original root cause confidence score that represents the root cause probability at the current moment.

[0165] Based on the comprehensive basic score of the input path, load the index of multiple batches of grinding wheel bubble defect cases recorded in the historical defect sample library and their corresponding process semantic anchor sequence and parameter combination to construct a feature retrieval index structure for matching verification.

[0166] The feature similarity retrieval module is invoked, using the process semantic anchor sequence, dominant parameter combination, and local causal contribution score vector of the current path as the retrieval key to perform multidimensional similarity calculation, and matching and scoring the feature vector of the corresponding path in the historical database with the features of this path.

[0167] Select a set of historical verification records whose matching scores exceed a preset similarity threshold, and statistically analyze the time series and frequency of their verification to form a time distribution dataset.

[0168] The time decay coefficient is calculated for the time distribution dataset using an exponential decay model:

[0169] in The time decay coefficient, The decay rate constant is This represents the time difference between the current moment and the moment the historical verification record event occurred.

[0170] The time decay coefficient is used to perform weighted adjustment on the overall basic score of the path, and a corrected credibility matrix is ​​generated through multiplicative weight mapping to eliminate the risk of high credibility of long-term unverified paths.

[0171] The output is the original root cause confidence score matrix representing the root cause probability at the current moment, which serves as the direct input for subsequent normalization and ranking algorithms.

[0172] By using a time decay weighting method, the path comprehensive basic score from the previous step is converted into an original root cause credibility score that includes timeliness information, thereby achieving dynamic correction of root cause credibility to enhance the practical applicability of the current analysis results.

[0173] For example, in a certain batch of grinding wheel defect analysis scenario, the path comprehensive basic score is 0.78. There are 12 verification records matching this path in the historical defect sample library. The most recent verification occurred 15 days before the current batch. The time decay rate constant is set to 0.02. The time difference Δt between the current moment and the most recent verification is taken as 15 days. Substituting these values ​​into the formula, the time decay coefficient λ is calculated as follows:

[0174] The attenuation coefficient was 0.739. Using this coefficient to adjust the overall base score of 0.78, the corrected confidence score was 0.576. This score was directly used as one item in the original root cause confidence score matrix. In the subsequent normalization stage, this significantly improved the ranking priority of recently validated paths in causal ordination and reduced the weight of paths that had not been validated for a long time in the current batch analysis, thereby enhancing the timeliness and accuracy of the analysis results.

[0175] S6.4: Perform global normalization on the original root cause confidence scores generated for all valid causal propagation paths, and use a soft maximum function or the summation reciprocal method to eliminate dimensional differences, mapping the original root cause confidence scores to a probability interval of zero to one, so as to generate a standardized root cause relative confidence probability distribution vector.

[0176] The system receives the original root cause confidence score vector generated by S6.3 as the input data basis, loads each element of the score vector, and establishes a global normalization operation environment to ensure that the subsequent probability mapping process covers the score results of all valid causal propagation paths.

[0177] The original root cause confidence score vector is input into the normalization calculation module, and the soft maximum function is called to implement nonlinear scaling. The specific calculation formula is as follows:

[0178] in, and These are the root cause credibility scores for the i-th and j-th valid causal propagation paths, respectively. Let be the normalized relative confidence probability value of the i-th valid causal propagation path. This represents the total number of effective causal propagation paths.

[0179] The proportion of high-scoring terms is increased through exponential operations, and global scale balance is achieved by summing the denominators.

[0180] The vectors generated by the above probability mapping operation are verified using the sum-of-the-parts method, and a balanced normalized model is constructed. The calculation formula is as follows:

[0181] This method retains the original score in the numerator while accumulating all scores in the denominator to achieve dimensionless calculation and proportional normalization, thus avoiding probability distribution distortion caused by differences in score intervals across different paths.

[0182] The soft maximum value function output vector and the sum reciprocal method output vector are checked for consistency to ensure that the values ​​of each element in the probability interval [0,1] meet the requirements of non-negativity and summation of 1. If the numerical deviation is detected to exceed the preset threshold, the exponential base or denominator compression parameter is adjusted backtrackingly until the normalized scores of all paths meet the probability space constraints.

[0183] By using a global normalization process, the original root cause confidence scores from the previous step are mapped to a probability range of zero to one, generating a standardized root cause relative confidence probability distribution vector. This achieves unified scaling of results across different scoring ranges, providing an accurate and comparable input benchmark for the descending sorting operation in S6.5.

[0184] For example, for a scoring vector containing 5 valid causal propagation paths, the original confidence scores are 3.2, 4.5, 2.8, 5.1, and 3.9. When using the soft-maximum function to calculate the exponent base with the natural constant e, the normalized probability value of the first path is calculated to be approximately 0.0713. Using the inverse sum rule, the numerator is 3.2, and the denominator is 3.2 + 4.5 + 2.8 + 5.1 + 3.9 = 19.5, resulting in approximately 0.1641. Consistency checks reveal that the soft-maximum distribution is biased towards high-scoring paths, while the inverse sum distribution is more balanced. Combining the two forms a composite normalized probability distribution vector. Ultimately, the probability interval values ​​for all 5 paths are non-negative and sum to 1, achieving a uniform scale of root cause confidence across different paths and significantly improving the stability and interpretability of subsequent ranking.

[0185] S6.5: Based on the numerical values ​​in the root cause relative confidence probability distribution vector, perform a descending sorting operation on each effective causal propagation path, extract the root cause anchor name and dominant parameter combination at the end of each path as the index key, and construct a normalized root cause sorting list containing sorting position, standardized score and associated path information.

[0186] Based on the numerical data of the probability distribution vector of the relative confidence of the root cause, the probability sorting engine is invoked to sort the terminal root causes of each valid causal propagation path in descending order, so as to establish a path sorting index structure from high confidence to low confidence.

[0187] Perform endpoint node parsing on the path set after descending order, extract the process semantic anchor point identifier and the bound dominant parameter combination at the path endpoint, and combine these two types of information to form a unique index key for subsequent sorting list construction.

[0188] The unique index key is associated with the normalized root cause relative confidence value and stored to generate an intermediate sorted dataset containing sort position and standardized score fields, ensuring that each record can be uniquely located through the index key.

[0189] For each record in the intermediate sorted dataset, add associated path structure information, including path span, sequence of semantic anchor points of processes along the way, and causal contribution scores of each node, to form a sorted record unit with complete context.

[0190] The aforementioned sorting record units are encapsulated according to a predetermined output data protocol to construct a normalized root cause sorting list containing three types of fields: sorting position, standardized score, and associated path information. This list serves as the direct input source for the subsequent root cause semantic analysis module.

[0191] By using serialization and index mapping, the probability distribution results from the previous step are transformed into a normalized root cause ranking list with searchable and structured fields, thus realizing the mapping transformation from root cause scoring to engineering application decisions.

[0192] For example, in a grinding wheel bubble defect analysis, the input root cause relative confidence probability distribution vectors are 0.82, 0.76, 0.65, and 0.58, respectively, corresponding to the paths: drying → sintering, mixing → drying, forming → drying, and sintering → finishing. A sorting engine is used to generate descending indices 1 to 4. The endpoint anchor point "sintering" of index 1 is bound to the dominant parameter combination "heating rate + holding time" to form an index key; the endpoint anchor point "drying" of index 2 is bound to the dominant parameter combination "solvent evaporation gradient + ambient humidity" to form an index key, and so on. A mapping relationship is established between the index keys and the corresponding probability values, recording fields including sorting position, standardized score (here 0.82, 0.76, 0.65, 0.58), and path information. The causal contribution scores of nodes in the path are filled in according to the process flow order, along with contextual information such as path span (e.g., drying → sintering is span 2) and sensitivity factor weights. The final output normalized root cause ranking list fully records the index key, standardized score, and associated path data of each root cause item, providing raw structural data for the next step of semantic parsing and report generation. During the verification process, the list can be directly used by production line engineers to retrieve the parameter runaway mode of the corresponding process anchor point, significantly improving the pertinence and efficiency of on-site decision-making.

[0193] Step S7: Extract the root cause anchor name, dominant parameter combination, and attribution conclusion based on physical interpretation for each data point in the normalized root cause ranking list, and generate a structured root cause analysis report containing an index of the original data fragments. Specifically, this includes: S7.1: Obtain the root cause record data for each item in the normalized root cause sorting list, and perform semantic mapping processing on the anchor identifiers in the root cause record based on the predefined process semantic anchor dictionary to generate a root cause semantic context set containing standardized process names and corresponding physical mechanism description templates.

[0194] S7.2: Receive the dominant parameter combination data in the root cause semantic context set, and use the parameter coupling anomaly identification rule base to perform logical verification and physical decoding processing on the dominant parameter combination to generate a parameter anomaly behavior description phrase sequence that characterizes the specific process runaway mode.

[0195] S7.3: Extract the sequence of descriptive phrases for abnormal parameter behavior and the corresponding set of root cause semantic contexts. Based on the causal deduction template in the knowledge graph of the formation mechanism of air wheel bubble defects, perform natural language synthesis processing on the sequence of descriptive phrases for abnormal parameter behavior to generate physical explanation phrases with complete causal logic chains.

[0196] S7.4: Load the physical interpretation phrase conclusion and the original multidimensional defect feature vector storage address index, and execute the data tracing and association algorithm to perform bidirectional binding processing between the key feature fingerprints in the physical interpretation phrase conclusion and the original data storage path to generate evidence chain index entries containing access pointers to the original data fragments.

[0197] S7.5: Summarize the root cause semantic context set, physical explanation phrase conclusions, and evidence chain index entries. After processing the summarized data by field encapsulation and format serialization according to the preset industrial report structured protocol, a structured root cause analysis report containing root cause anchor names, dominant parameter combinations, physical explanation phrases, and original data fragment indexes is generated.

[0198] The generated root cause semantic context set, physical explanation phrase conclusion, and evidence chain index entries are used as input data for summary processing. The industrial report structured protocol parsing module is called to perform field alignment verification on the above three types of data sources to ensure the consistency of process semantics, parameter anomaly descriptions, and evidence chain indexes of the same root cause record in terms of timestamps and process sequences.

[0199] The data that has passed the field alignment verification is input into the field encapsulation processor. The process semantic context set is standardized and encapsulated according to the field template preset by the protocol, including information such as process anchor name, process physical mechanism description, and sensitivity factor weight, thereby forming a process description field block with a unified identifier.

[0200] The conclusions of the physical explanation phrases are semantically segmented and encoded, and then encapsulated into causal logic chain field blocks. A mapping relationship is established between these blocks and the corresponding dominant parameter combination field blocks. The dominant parameter combination field blocks contain parameter names, numerical ranges, coupling relationship matrices, etc., to ensure the traceability between the causal logic chain and the abnormal behavior of specific parameters.

[0201] The evidence chain index entries are verified through two-way binding and encapsulated into evidence reference field blocks, which include the original data fragment access pointer, the hash value of the multidimensional defect feature vector, and the data source type. Each evidence reference field block is then cross-referenced with the causal logic chain field block.

[0202] The execution format serialization module combines the above process description field block, dominant parameter combination field block, causal logic chain field block, and evidence citation field block into a complete record unit according to the order rules of the industrial report structured protocol, generating a structured root cause analysis report containing root cause anchor name, dominant parameter combination, physical explanation phrase, and original data fragment index.

[0203] By using field encapsulation and format serialization, the results of the previous step are transformed into structured analysis data with a unified protocol format, containing complete causal chains and traceability information, thereby achieving standardized output of root cause analysis reports and automated integration into production lines.

[0204] Step S8: Feed the structured root cause analysis report back to the production line decision system. If a new defect feature pattern is detected that does not match the existing process semantic anchor response, an update of the statistical correlation pattern in the process semantic anchor sequence is triggered to optimize the prediction accuracy of the subsequent anchor response predictor. Specifically, this includes: S8.1: Parse the normalized root cause sorting list and original data fragment index in the structured root cause analysis report to extract the multidimensional defect feature vector of the current batch and the corresponding process semantic anchor activation probability distribution, and generate a defect mode response record set to be verified.

[0205] S8.2: Based on the defect pattern response record set to be verified, calculate the Euclidean distance between the newly input multidimensional defect feature vector and the known defect pattern prototype in the historical database using a predefined similarity metric algorithm, and generate a sequence of difference coefficients representing the novelty of the pattern.

[0206] S8.3: Based on the outliers in the difference coefficient sequence that exceed the preset threshold, locate the unmatched specific process semantic anchors and their associated sensitivity factors, call the incremental learning engine to fine-tune the weight parameters of the local causal strength evaluator, and generate an updated local causal contribution score calculation model.

[0207] S8.4: Utilize the updated local causal contribution score calculation model to re-analyze historical defect samples, correct the statistical correlation pattern between parameter deviation and bubble characteristics at each process node, and generate a calibrated process semantic anchor statistical correlation knowledge base.

[0208] S8.5: Based on the calibrated process semantic anchor statistical association knowledge base, replace the old mapping rules in the anchor response predictor, perform full regression testing to verify the improvement in prediction accuracy, generate the final version of the anchor response predictor with adaptive optimization capabilities, and deploy it to the production line decision system.

[0209] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0210] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0211] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for root cause analysis of internal bubble porosity defects in a grinding wheel, characterized in that, Specifically, it includes: S1: Obtain the sequence of key process nodes in the grinding wheel manufacturing process specification, and construct the process semantic anchor sequence based on the explanation of the physical mechanism of bubble formation in each process and the statistical correlation pattern of historical defect samples. S2: The multidimensional defect feature vector extracted from the CT scan image of the grinding wheel is input into the pre-trained anchor response predictor, and the anchor activation probability distribution is generated based on the process semantic anchor sequence. S3: Based on the anchor point activation probability distribution, select high-response process semantic anchor points, and for each high-response process semantic anchor point, call its bound local causal strength evaluator to calculate the local causal contribution score based on the actual parameter values ​​of the current batch, equipment status log fragments, and material batch attributes. S4: The local causal contribution score is passed along the process chain to the process semantic anchor point with a clear physical interpretation path in the preceding sequence to generate an initial causal propagation path set; S5: Perform process fidelity attenuation verification on each path in the initial causal propagation path set based on preset physical constraint rules to generate a valid causal propagation path set; S6: Summarize the causal scores at the endpoints of the effective causal propagation path set, and perform weighted fusion calculations by combining the process weights, parameter coupling confidence, and historical verification frequencies of the process semantic anchors along each path to generate a normalized root cause ranking list. S7: Extract the root cause anchor name, dominant parameter combination, and attribution conclusion composed of physical explanation for each data in the normalized root cause ranking list, and generate a structured root cause analysis report.

2. The method for root cause analysis of internal bubble porosity defects in a grinding wheel according to claim 1, characterized in that, The process semantic anchor sequence encapsulates a typical set of controllable process parameters, sensitivity factors, and statistical correlation patterns.

3. The method for root cause analysis of internal bubble porosity defects in a grinding wheel according to claim 1, characterized in that, After step S7, the following also includes: S8: Feed the structured root cause analysis report back to the production line decision system. If a new defect feature pattern is detected that does not match the existing process semantic anchor response, then the statistical correlation pattern in the process semantic anchor sequence is updated.

4. The method for root cause analysis of internal bubble porosity defects in a grinding wheel according to claim 2, characterized in that, Step S2 specifically includes: the multidimensional defect feature vector is composed of a heat map of the spatial density of the bubble pores, a pore size spectrum, and microcrack features of the pore wall.

5. The method for root cause analysis of internal bubble porosity defects in a grinding wheel according to claim 4, characterized in that, Step S2 specifically includes: The space density thermal map of the bubble pores, the pore size spectrum, and the microcrack features of the pore wall are subjected to vector splicing and normalization to generate a multidimensional defect feature vector. Based on the statistical correlation patterns in the process semantic anchor sequence, the anchor response predictor model parameters are trained and solidified using a deep neural network architecture to generate a pre-trained anchor response predictor. The multidimensional defect feature vector is input into the input layer of the pre-trained anchor response predictor. Through the nonlinear feature transformation of the hidden layer, a hidden layer activation tensor representing the potential response intensity of each process semantic anchor point is generated. The generated hidden layer activation tensor is subjected to soft maximum normalization based on the dimension of the process semantic anchor point sequence to generate the anchor point activation probability distribution.

6. The method for root cause analysis of internal bubble porosity defects in a grinding wheel according to claim 2, characterized in that, Step S3 specifically includes: The activation probability distribution of the anchor points is truncated and sorted to filter out the set of top process semantic anchor points whose probability values ​​exceed a preset threshold, and generate a list of high-response process semantic anchor points that represent the starting position of causal reasoning. Based on each anchor identifier in the high-response process semantic anchor list, extract the actual process parameter value sequence, equipment status log fragment and material batch attribute data corresponding to the current production batch, perform time window alignment and missing value interpolation processing, and generate multi-source heterogeneous working condition input vector. The pre-trained local causal strength evaluator receives the multi-source heterogeneous operating condition input vector, performs neural symbolic hybrid inference operations, and generates a set of original marginal causal contribution scores that characterize the degree of independent influence of each typical controllable process parameter on bubble formation. The parameter coupling effect identification algorithm is applied to the original set of marginal causal contribution scores to calculate the interaction gain coefficient between different parameter combinations, correct the independent contribution of a single parameter to a joint contribution value including interaction terms, and generate a corrected local causal contribution score matrix that characterizes the risk of parameter collaborative failure. Normalization mapping is performed based on the corrected local causal contribution score matrix, and weighted fusion is performed in combination with the predefined sensitivity factor weights within the process semantic anchor point to generate the final local causal contribution score that represents the marginal contribution of each parameter combination to the formation of bubble defects under the current high-response process semantic anchor point.

7. The method for root cause analysis of internal bubble porosity defects in a grinding wheel according to claim 6, characterized in that, The neural symbolic hybrid inference operation is performed by receiving the multi-source heterogeneous operating condition input vector using a pre-trained local causal strength evaluator. Specifically, the neural network part extracts the nonlinear feature interaction pattern of the multi-source heterogeneous operating condition input vector, and the symbolic logic part embeds physical constraint rules into the multi-source heterogeneous operating condition input vector.

8. The method for root cause analysis of internal bubble porosity defects in a grinding wheel according to claim 1, characterized in that, Step S6 specifically includes: Based on the sequence of process semantic anchor points traversed by each path in the set of effective causal propagation paths, the preset process weight coefficients corresponding to each process semantic anchor point are extracted, and the preset process weight coefficients are multiplied or weighted summed to generate path process weight factors. Using the path process weighting factor as the basic input, and combining it with the parameter coupling confidence value corresponding to the parameter coupling anomaly combination identified in the local causality strength evaluator, the path process weighting factor is dynamically corrected to generate a path comprehensive basic score; wherein, the parameter coupling anomaly combination is a parameter pair that has a significant coupling relationship between the local causality contribution scores; Based on the comprehensive basic score of the path, the historical defect sample library is called to retrieve historical verification records that match the characteristics of the path. The frequency of historical verification is counted and converted into a time decay coefficient. The time decay coefficient is applied to the comprehensive basic score of the path for weighted adjustment to generate the original root cause credibility score. Global normalization is performed on the original root cause confidence scores generated by all valid causal propagation paths, mapping the original root cause confidence scores to a probability interval of zero to one, and generating a root cause relative confidence probability distribution vector. Based on the numerical values ​​in the root cause relative confidence probability distribution vector, each effective causal propagation path is sorted in descending order. The root cause anchor name and dominant parameter combination at the end of each path are extracted as index keys to construct the normalized root cause sorting list.

9. The method for root cause analysis of internal bubble porosity defects in a grinding wheel according to claim 1, characterized in that, Global normalization is performed on the original root cause confidence scores generated for all valid causal propagation paths, including by using a soft maximum function or the inverse summation method.

10. The method for root cause analysis of internal bubble porosity defects in a grinding wheel according to claim 1, characterized in that, The normalized root cause ranking list includes: ranking position, standardized score, and associated path information.

Images