Casting high-temperature alloy composition design method and system based on OCR and small sample learning

Abstract

The embodiment of the application discloses a casting high-temperature alloy component design method and system based on OCR and small sample learning, comprising: an intelligent visual identification link based on OCR: scanning of paper quality certificates and detection reports of casting high-temperature alloy master alloys is converted into a structured data set with a traceable label through a laminated OCR post-processing scheme; the laminated OCR post-processing scheme comprises sequentially executed page direction automatic correction, adaptive row clustering, chemical priori cost Hungarian matching and structured data output with a traceable label; a small sample credible mapping link: based on the structured data set, a component-mechanical property credible mapping model capable of simultaneously outputting a predicted mean value and a predicted uncertainty is established and trained; and a reverse component design link: based on the component-mechanical property credible mapping model, reverse search is performed in a process feasible region, and an optimized component recommendation list with a statistical confidence interval is output.

Description

Technical Field

[0001] This invention relates to, but is not limited to, the field of alloy composition design technology, and particularly to a method and system for designing the composition of casting high-temperature alloys based on OCR and small-sample learning. Background Technology

[0002] Cast high-temperature alloys are core components of the hot-end parts of aero-engines and gas turbines. The metallurgical quality and service reliability of high-temperature alloy precision castings such as turbine blades, casings, and guide vanes are largely determined by the chemical composition and mechanical properties of their upstream master alloys. As the source raw material for cast high-temperature alloys, even slight fluctuations in the chemical composition of the master alloy have a decisive impact on the mechanical properties and quality of the final castings.

[0003] In the production practice of master alloys, each heat of master alloy requires chemical composition testing and mechanical property testing, and the results are archived by the quality inspection department in the form of certificates of conformity and test reports. However, the industry has long relied on archiving these reports as paper documents or scanned copies. These reports are often scattered in different locations such as workshop ledgers, quality inspection room filing cabinets, and the company's electronic document system, and the content of the reports has not been standardized to form calculable data assets. Although the dozens to hundreds of master alloy reports accumulated over the years constitute data assets with engineering value, the industry's understanding and utilization of their potential value remains insufficient. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned technical problems by providing a method and system for designing the composition of casting high-temperature alloys based on OCR and few-sample learning. This addresses the issue that existing composition-performance data reports for casting high-temperature alloy master alloys are typically paper or scanned documents, resulting in scattered reports that lack computable data assets, thus hindering the systematic management and utilization of master alloy data assets.

[0005] The technical solution of the present invention: In a first aspect, embodiments of the present invention provide a method and system for designing the composition of casting high-temperature alloys based on OCR and few-sample learning, comprising:

[0006] Step 1, the intelligent visual recognition step based on optical character recognition (OCR), includes: scanning the paper certificates of conformity and test reports of the cast high-temperature alloy master alloy, and converting them into a structured dataset with traceability tags through a layered OCR post-processing scheme; the layered OCR post-processing scheme includes the sequential execution of automatic page orientation correction, adaptive line clustering, Hungarian matching with chemical prior costs, and structured data output with traceability tags;

[0007] Step 2, the small sample reliable mapping step, includes: based on the structured dataset, establishing and training a component-mechanical property reliable mapping model that can simultaneously output the prediction mean and prediction uncertainty;

[0008] Step 3, the reverse composition design stage, includes: based on the composition-mechanical property reliable mapping model, performing a reverse search within the process feasible domain allowed by the casting high-temperature alloy technology conditions, and outputting a list of recommended preferred compositions with statistical confidence intervals.

[0009] Optionally, in the casting high-temperature alloy composition design method based on OCR and few-shot learning as described above, the layered OCR post-processing scheme in step 1 includes:

[0010] Step 11, automatic page orientation correction, includes: performing OCR recognition on each scanned page in multiple candidate orientations, calculating the orientation score of each candidate orientation using a combined scoring function based on the average confidence of characters and the proportion of Chinese characters, and determining the candidate orientation with the highest score as the positive orientation;

[0011] Step 12, adaptive row clustering, includes: based on the Y-direction coordinates of the character bounding boxes identified on the scanned page, adaptively determining the neighborhood radius of density clustering with the median height of characters on the scanned page as a reference, and executing the density clustering algorithm to divide the physical rows of the scanned page;

[0012] Step 13, Hungarian matching of chemical prior costs, includes: modeling the alignment problem of element headers and numerical candidates in the same physical row to obtain the optimal allocation problem of bipartite graph, whose cost function includes the normalized distance of the horizontal axis and chemical prior terms;

[0013] Step 14, output structured data with source tags, including: appending source file information and location coordinates to the successfully matched element-component numerical pairs and writing them into the structured dataset.

[0014] Optionally, in the casting high-temperature alloy composition design method based on OCR and few-shot learning as described above, the combined score function in step 11 is defined as:

[0015] score=mean conf ×(1+0.5×ratio zh );

[0016] Where score is the directional score of the candidate direction, and mean conf The ratio represents the average recognition confidence of all characters in the candidate recognition results. zh This represents the percentage of Chinese characters in the candidate recognition results.

[0017] Optionally, in the casting high-temperature alloy composition design method based on OCR and few-shot learning as described above, the adaptive determination of the neighborhood radius in step 12 is as follows:

[0018] Calculate the height of the bounding boxes of all characters on the currently scanned page, and denot the median as h. med The neighborhood radius of density clustering is calculated as ε:

[0019] ε=0.55×h med ;

[0020] Wherein, the minimum number of neighborhood samples in the density clustering algorithm is min samples Set to 2;

[0021] A density clustering algorithm is performed on the Y-axis coordinates of all characters in the currently scanned page. Characters belonging to the same cluster are identified as belonging to the same physical row.

[0022] Optionally, in the casting high-temperature alloy composition design method based on OCR and few-shot learning as described above, the bipartite graph optimal allocation problem modeled in step 13 is represented by a cost matrix, where the element in the i-th row and j-th column of the cost matrix represents the cost of assigning the i-th element header to the j-th numerical candidate.

[0023] The cost function used is specifically defined as follows:

[0024] cost(i,j)=α·Δx norm (i,j)+β·prior_mismatch(value j ∈range i );

[0025] Where, Δx norm (i,j) is the normalized distance between the x-coordinate center of the i-th element header and the x-coordinate center of the j-th numerical candidate; prior_mismatch is a chemical prior, which is the distance between the x-coordinate center of the i-th element header and the j-th numerical candidate. j The reasonable content range of the i-th element in the process. i The value is 0 if the condition is met, otherwise it is 1; α and β are weighting constants.

[0026] After obtaining the cost matrix, the Hungarian algorithm is used to find the globally optimal assignment that minimizes the total cost.

[0027] Optionally, in the casting high-temperature alloy composition design method based on OCR and few-shot learning as described above, step 14 includes:

[0028] Each successfully matched element-component pair is appended with four source tags: source file name, page number, line number, and original character bounding box coordinates. These tags, along with the persistence performance and heat processing status fields, are written into the structured dataset.

[0029] Optionally, in the OCR-based and few-sample learning-based casting high-temperature alloy composition design method described above, step 2 establishes a reliable composition-mechanical property mapping model, and two schemes are formed based on whether there is censored observation in the data:

[0030] Option 1: For cases where all mechanical property data of the alloy are fully observed, Gaussian process regression is used as the mapping model from composition to properties, and the kernel function is the Matérn-5 / 2 kernel function. The hyperparameters of the kernel function are solved by maximizing the marginal likelihood.

[0031] Option 2 involves training an accelerated failure time model in parallel with Gaussian process regression to address censored observations caused by terminating the test before the alloy sample has completely failed. The training objective of this accelerated failure time model is based on a log-likelihood function that includes censored indicator variables.

[0032] Optionally, in the OCR-based and few-shot learning-based method for designing casting high-temperature alloy compositions as described above, the expression for the Matérn-5 / 2 kernel function in step 2 is:

[0033] ;

[0034] Where r = ||x−x′|| is the Euclidean distance between the input feature vectors x and x′; It is the core amplitude; It is a nuclear length scale.

[0035] Optionally, in the casting high-temperature alloy composition design method based on OCR and few-shot learning as described above, the expression for the accelerated failure time model in step 2 is:

[0036] ;

[0037] in, It is an event indicator variable, with 1 for natural failure and 0 for failure under applied force; It is the observed lifespan; is the component vector of the i-th record; f() is the probability density function, and S() is the survival function.

[0038] Optionally, in the casting high-temperature alloy composition design method based on OCR and few-shot learning as described above, step 2 further includes:

[0039] Under the same cross-validation conditions, the outputs of various regression models, including linear regression, kernel regression, support vector regression, random forest, gradient boosting tree, and Gaussian process regression, are compared horizontally. The model with the best prediction accuracy is automatically selected as the base model and used as the input for downstream reverse recommendation. Gaussian process regression is used as the base model for uncertainty estimation by default.

[0040] Optionally, in the OCR-based and few-shot learning-based high-temperature alloy composition design method described above, the reverse search in step 3 includes: executing a reverse search scheme in scenario one, where scenario one is: specifying the target performance to deduce the composition; the reverse search process in scenario one includes:

[0041] The first step is to use Latin hypercube sampling to generate N candidate compositions within the element content range allowed by the alloy technology conditions;

[0042] The second step is to call the trained component-mechanical property reliability mapping model to predict the predicted mean μ(x) and predicted standard deviation σ(x) of each candidate component on the target performance;

[0043] The third step involves using the expected improvement acquisition function to calculate the expected improvement EI score for each candidate component, sorting them by score, and selecting the candidate components with the highest expected improvement EI scores as the recommended output.

[0044] Optionally, in the OCR-based and few-shot learning-based high-temperature alloy composition design method for casting described above, the expression for improving the acquisition function in the third step is:

[0045] ;

[0046] in, , These are the predicted mean and standard deviation of candidate component x by Gaussian process regression, respectively. It is the currently known optimal performance value; It involves exploring and utilizing the equilibrium constant, typically taken as 0.01; ; , These are the cumulative distribution function and probability density function of the standard normal distribution, respectively.

[0047] Optionally, in the OCR-based and few-shot learning-based casting high-temperature alloy composition design method described above, the reverse search in step 3 further includes: executing a reverse search scheme under scenario two, where scenario two involves locking key elements and deducing other elements; the reverse search process in scenario two includes:

[0048] Using the lock values ​​or lock ranges set by process engineers for the content of one or more elements as constraints, sampling and reverse search are performed only on elements for which no lock values ​​or lock ranges are set within the constrained search space.

[0049] Optionally, in the OCR-based and few-shot learning-based casting high-temperature alloy composition design method described above, the reverse search process in scenario two further includes:

[0050] Based on the upper and lower relaxation coefficients set by process engineers for individual elements, the proportion represented by the relaxation coefficient is expanded by extending the sampling range of the element below the lower limit and above the upper limit of the original alloy standard, thereby pushing the content of precious elements outside the alloy standard range for cost control exploration.

[0051] Optionally, in the OCR-based and few-shot learning-based casting high-temperature alloy composition design method described above, the reverse search in step 3 further includes: executing the reverse search scheme under scenario three, where scenario three is: multi-objective frontier search; the reverse search process for scenario three includes:

[0052] For design scenarios that require simultaneous consideration of multiple mechanical performance indicators, the multi-objective genetic algorithm NSGA-II is used to search for the Pareto front solution set in the element content space that optimizes multiple mechanical performance indicators simultaneously.

[0053] The parameters of the NSGA-II algorithm include: N pop =100; Evolutionary generation N gen =50; the mutation uses multinomial mutation, with a mutation probability p. m =1 / d x d x Let η be the element dimension and the distribution index η. m =20; Crossover uses simulated binary crossover, with a crossover probability p. c =0.9, distribution index η c =15.

[0054] Optionally, in the OCR-based and few-shot learning-based casting high-temperature alloy composition design method described above, step 3, before outputting the preferred composition recommendation list, further includes:

[0055] Constructing an out-of-distribution suppression factor based on Mahalanobis distance includes: multiplying the out-of-distribution suppression factor by the original EI score or Pareto ranking score and then re-ranking the data to remove candidate components that deviate from the reasonable distribution range of the training data; the out-of-distribution suppression factor is defined as follows: Where α is the inhibition intensity coefficient, The Mahalanobis distance between candidate component x and the training set;

[0056] Mahalanobis distance The calculation formula is:

[0057] ;

[0058] in, and These are the mean vector and covariance matrix of the training set element vectors, respectively, and T denotes the transpose.

[0059] Optionally, in the OCR-based and few-shot learning-based casting high-temperature alloy composition design method described above, step 3 further includes:

[0060] Using a conformal prediction method, based on the empirical distribution of prediction residuals on the calibration set, a statistically significant confidence interval is recommended for each component; the final width of the confidence interval is the maximum of the width of the posterior interval of the Gaussian process regression and the width of the conformal prediction quantile.

[0061] Secondly, the present invention also provides a casting high-temperature alloy composition design system based on OCR and few-sample learning, the system being used to execute the casting high-temperature alloy composition design method based on OCR and few-sample learning as described above, the system comprising:

[0062] The OCR intelligent visual recognition module is used to perform step 1, converting the scanned paper report into structured data;

[0063] The structured data foundation module is used to store and manage the structured dataset and traceability tags;

[0064] A trusted mapping model library module is used to perform step 2, train, store, and manage the component-mechanical property trusted mapping model;

[0065] The reverse design module is used to execute step 3, which calls the trusted mapping model to perform reverse search based on the engineering constraints or performance goals input by the user, and generates a list of recommended preferred components.

[0066] The visualization front-end module provides a graphical user interface that allows users to import original reports, review structured results, set reverse design parameters, view component recommendations and their confidence intervals, and trace and verify source files.

[0067] The trusted mapping model library module performs step 2 to establish a component-mechanical property trusted mapping model, and forms two schemes based on whether the data has censored observations:

[0068] Option 1: For cases where all mechanical property data of the alloy are fully observed, Gaussian process regression is used as the mapping model from composition to properties, and the kernel function is the Matérn-5 / 2 kernel function. The hyperparameters of the kernel function are solved by maximizing the marginal likelihood.

[0069] Option 2: For censored observations caused by terminating the test before the alloy sample completely fails, in addition to Gaussian process regression, an accelerated failure time model is trained in parallel; the training objective of this accelerated failure time model is constructed based on the log-likelihood function containing censored indicator variables.

[0070] The expression for the Matérn-5 / 2 kernel function is:

[0071] ;

[0072] Where r = ||x−x′|| is the Euclidean distance between the input feature vectors x and x′; It is the core amplitude; It is a nuclear length scale;

[0073] In step 2, the expression for the accelerated failure time model is:

[0074] ;

[0075] in, It is an event indicator variable, with 1 for natural failure and 0 for failure under applied force; It is the observed lifespan; is the component vector of the i-th record; f() is the probability density function, and S() is the survival function.

[0076] Thirdly, the present invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when executed by a processor, is used to implement the casting high-temperature alloy composition design method based on OCR and few-sample learning as described in any one of the above.

[0077] The beneficial effects of this invention are as follows: This invention provides a method and system for designing casting high-temperature alloy compositions based on OCR and small-sample learning, specifically an end-to-end reverse design method and system for master alloy compositions. Using the technical solution provided by this invention, implicit data assets scattered in historical paper certificates and test reports can be automatically structured through OCR intelligent visual recognition, modeled through small-sample reliable mapping, and ultimately transformed into executable master alloy optimal composition recommendations without changing alloy grades or relying on large-scale new experiments. Each recommendation is accompanied by a statistically significant confidence interval, enabling process engineers to robustly and continuously optimize master alloy compositions within the limits of current technology. The technical solution provided by this invention is alloy grade independent; the three core steps in the solution can be easily transferred to various casting high-temperature alloy master alloy systems. Furthermore, this invention has the following beneficial effects:

[0078] Firstly, the composition and performance data of the master alloy, previously scattered in paper certificates of conformity and test reports and long neglected by the industry, are transformed into calculable, traceable, and governable structured data assets through a layered OCR post-processing scheme that combines automatic orientation correction, adaptive row clustering, and Hungarian matching of chemical prior costs. This establishes an end-to-end data governance entry point in the context of master alloy production reports, which have long lacked algorithmically readable formats. Secondly, the explicit introduction of chemical prior costs is one of the key technical details of this invention. The cost function explicitly utilizes the reasonable content range of each element in the casting high-temperature alloy as prior information, giving the layered OCR post-processing scheme a domain-adaptive capability not found in general-purpose OCR tools.

[0079] Secondly, considering the inherent small sample size in production reports of cast high-temperature alloy master alloys, and taking into account the censored observations in some alloy systems, a reliable mapping scheme is established that simultaneously possesses the ability to estimate unbiased performance and quantify reliable uncertainty, forming a reliable foundation for downstream reverse search. Specifically, this invention employs Matérn-5 / 2 kernel Gaussian process regression for complete observation scenarios, and trains an accelerated failure time survival model in parallel for censored observation scenarios. Furthermore, it performs horizontal comparisons of multiple mainstream regression models under the same cross-validation to automatically select the most robust base model.

[0080] Thirdly, the technical solution of this invention proposes three reverse search scenarios that can be invoked independently or in combination: Scenario 1 is to reverse-engineer the composition by specifying the target performance; Scenario 2 is to reverse-engineer other elements by locking key elements, supporting controlled exploration beyond the upper and lower limits of the original alloy standard; Scenario 3 is a multi-objective Pareto front search, using NSGA-II for non-dominated sorting in the element content space. All three scenarios use an out-of-distribution suppression factor based on Mahalanobis distance and a calibrable confidence interval based on conformal prediction for final filtering and statistical guarantee, ensuring that the composition recommendation has both engineering feasibility and calibrable statistical significance.

[0081] Fourth, the essence of the technical solution of this invention, which differs from existing single-step solutions, lies in connecting the above three steps in the order of OCR intelligent visual recognition, small-sample reliable mapping, and reverse composition design, forming a complete closed loop for optimizing the composition of casting high-temperature alloy master alloys. The casting high-temperature alloy composition method and system provided by this invention maintains decoupling from alloy grades and can be easily transferred to various casting high-temperature alloy master alloy systems. Attached Figure Description

[0082] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of the present invention and do not constitute a limitation on the technical solutions of the present invention.

[0083] Figure 1 A schematic diagram of the overall design architecture of a casting high-temperature alloy composition design method based on OCR and few-sample learning provided in an embodiment of the present invention;

[0084] Figure 2 for Figure 1 The illustrated embodiment provides a detailed flowchart of each step in the design method for casting high-temperature alloy composition based on OCR and few-sample learning. Detailed Implementation

[0085] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

[0086] As explained in the background section, the master alloy serves as the primary raw material for casting high-temperature alloys, and its chemical composition significantly impacts the mechanical properties and metallurgical quality of the final casting. However, existing master alloy composition-performance data reports are generally stored as paper documents or scanned copies, resulting in scattered reports and a lack of calculable data assets. This leads to difficulties in the systematic management and utilization of master alloy data assets.

[0087] A key characteristic of cast high-temperature alloy master alloys is the adjustable range of element content within standard allowable limits. The same alloy grade exhibits objective compositional fluctuations across different heats, and these fluctuations correspond precisely to differences in mechanical properties. In other words, there are still identifiable and usable optimal ranges hidden within the compositional range of the master alloy. The inventors of this invention considered that, while maintaining the alloy grade and production process unchanged, systematically identifying and locking onto these optimal ranges could potentially further improve the mechanical properties and overall metallurgical quality of the castings. However, to truly implement this idea in engineering, two unavoidable engineering hurdles must be overcome simultaneously. First, the historically accumulated composition-performance data must be migrated from paper to an algorithmically readable structured database, making it computationally readable. Second, under the real-world condition of generally limited data volume (often only a few dozen heats), a reliable composition-performance mapping must be established, and based on this, a reverse compositional search must be performed, ultimately outputting an engineering-executable recommended list. In addition, for some high-temperature alloys, there are situations in mechanical property testing where the test is terminated before the sample has completely failed (i.e., the observation is censored). If such situations are directly treated as accurate observations, it will systematically distort the modeling target.

[0088] In existing solutions, general optical character recognition (OCR) tools, machine learning prediction models, and Bayesian optimization methods are mostly applied to single stages of composition design. They fail to form an end-to-end composition design solution specifically for the production report of casting high-temperature alloy master alloys, integrating data governance, small-sample modeling, and reverse design. Furthermore, existing solutions also fail to effectively address common problems in production data such as few samples, censored observations, missing values, and out-of-distribution candidate risks.

[0089] Therefore, constructing a complete method and system that starts with automatic identification of paper reports, proceeds through reliable mapping with a small number of samples, and finally leads to reverse recommendation of components has become a technical problem that urgently needs to be solved in this field.

[0090] To address the aforementioned problems and needs, this invention provides a method and system for designing casting high-temperature alloy compositions based on OCR and few-sample learning. Specifically, it is an end-to-end reverse design method and system for master alloy compositions. The technical solution provided by this invention is alloy grade independent, and the three core steps can be easily transferred to various casting high-temperature alloy master alloy systems.

[0091] The present invention provides the following specific embodiments, which can be combined with each other. For the same or similar concepts or processes, they may not be described again in some embodiments.

[0092] Figure 1This is a schematic diagram of the overall design architecture of a casting high-temperature alloy composition design method based on OCR and few-sample learning, provided as an embodiment of the present invention. Figure 1 As shown, the casting high-temperature alloy composition design method provided in this embodiment of the invention includes three interconnected steps, executed sequentially in a head-to-tail manner: an OCR-based intelligent visual recognition step for paper reports, a composition-performance reliable mapping step for small sample data, and a reverse composition design step for the process feasible domain. The operation content of the three steps provided in this embodiment of the invention is as follows:

[0093] Step 1, the intelligent visual recognition step based on OCR, includes: scanning the paper certificates of conformity and test reports of the cast high-temperature alloy master alloy, and converting them into a structured dataset with traceability tags through a layered OCR post-processing scheme; the layered OCR post-processing scheme includes the sequential execution of automatic page orientation correction, adaptive line clustering, Hungarian matching with chemical prior costs, and structured data output with traceability tags;

[0094] Step 2, the small sample reliable mapping step, includes: based on the structured dataset, establishing and training a component-mechanical property reliable mapping model that can simultaneously output the prediction mean and prediction uncertainty;

[0095] Step 3, the reverse composition design stage, includes: based on the composition-mechanical property reliable mapping model, performing a reverse search within the process feasible domain allowed by the casting high-temperature alloy technology conditions, and outputting a list of recommended preferred compositions with statistical confidence intervals.

[0096] It should be noted that, in this embodiment of the invention, by executing the above three steps sequentially in a connected manner, an optimal solution for the composition design of the master alloy for casting high-temperature alloys is formed, creating a complete closed-loop composition design solution that has not been achieved in existing single-step research. The composition design process of this embodiment of the invention is as follows: Figure 1 As shown. The following describes the specific implementation of the three steps in the casting high-temperature alloy composition design method provided by the present invention, which are executed sequentially in a first-to-last manner.

[0097] The following is in conjunction with the appendix Figure 1 and Figure 2 The specific implementation methods of the three core components in the technical solution of this invention will be described in detail.

[0098] Step 1: OCR-based intelligent visual recognition

[0099] In this embodiment of the invention, the OCR-based intelligent visual recognition process addresses typical problems commonly found in paper certificates of conformity and test reports for cast high-temperature alloy master alloys, such as inconsistent scanning directions, inconsistent layouts, large differences in table row heights, and mismatches between multi-column values ​​and element headers. It constructs a layered OCR post-processing scheme that combines automatic orientation correction, adaptive row clustering, and chemical prior cost matching.

[0100] The plan consists of four steps to be executed sequentially, which are described below.

[0101] Step 11: Automatic page orientation correction.

[0102] In this step, each scanned page is fed into the OCR engine along two candidate directions (0 degrees and 180 degrees, or up to four candidate directions if necessary) to obtain two candidate recognition results. A direction score is then calculated for each candidate recognition result. This direction score is calculated from two parts: the first part is the average recognition confidence of all characters in the candidate recognition result, denoted as mean. conf The first part is directly output by the OCR engine; the second part is the proportion of Chinese characters (i.e., Unicode encoding in the range of 0x4E00 to 0x9FFF) in the candidate recognition result, denoted as ratio. zh In this embodiment of the invention, the direction score is set to the former multiplied by one plus 0.5 times the latter. The formula for calculating the direction score is as follows:

[0103] score=mean conf ×(1+0.5×ratio zh );

[0104] The direction with the higher score among the two candidate directions is selected as the positive direction, and the inverted scanned page is automatically flipped. The engineering significance of the combined scoring function used in this step is that after the inverted scanned page is recognized by the OCR engine, a large number of low-confidence garbled characters and a very low proportion of Chinese characters will appear, resulting in a significantly lower score, thus achieving reliable direction determination.

[0105] Step 12, Adaptive row clustering;

[0106] In this step, density clustering is performed on all characters identified on the scanned page after orientation correction, based on the Y-coordinate of their bounding boxes. Specifically, the heights of the bounding boxes of all characters on the current scanned page are first calculated, and the median is denoted as h. med The neighborhood radius for density clustering is calculated as follows:

[0107] ε=0.55×h med ;

[0108] The ε calculated by the above formula is the neighborhood radius of density clustering, expressed as min samples =2 is used as the minimum number of neighborhood samples for the density clustering algorithm. A density clustering algorithm (such as DBSCAN) is performed on the Y-coordinates of all characters in the currently scanned page. Characters belonging to the same cluster are considered to belong to the same physical row. Compared to traditional row division methods that round down to a fixed pixel threshold, the key innovation of this step is that the neighborhood radius of the density clustering is adaptively given by the median of the character heights on the page, thus eliminating the impact of font size differences on the stability of row division under different resolution scans.

[0109] Step 13, Hungarian bipartite graph matching of chemical prior costs.

[0110] In this step, the alignment problem between the element headers (typically 11 items such as C, Cr, Co, W, Mo, Ta, Al, Ti, Hf, B, Zr, etc.) in the same physical row and several identified numerical candidates in that physical row is modeled as a bipartite graph optimal allocation problem. In this problem, each element header and each numerical candidate are assigned a cost, and the element in the i-th row and j-th column of the cost matrix represents the cost of assigning the i-th element header to the j-th numerical candidate. The cost function used in this invention considers both horizontal coordinate distance and chemical rationality information, and is specifically defined as follows:

[0111] cost(i,j)=α·Δx norm (i,j)+β·prior_mismatch(value j ∈range i );

[0112] The first term Δx norm (i,j) is the normalized distance between the x-coordinate center of the i-th element header and the x-coordinate center of the j-th numerical candidate; the second term, prior_mismatch, is a chemical prior term, whose value is determined by the following rule: when the numerical candidate value... j The reasonable content range of the alloy process for the i-th element. iWhen the value is outside the specified range, a value of 1 is used; when it falls within the range, a value of 0 is used. α and β are weight constants; for example, α = 1.0 and β = 0.5. After obtaining the cost matrix of 11 rows and several columns, this step calls the Hungarian algorithm (also known as the Kuhn-Munkres algorithm, a classic polynomial-time algorithm for solving bipartite graph optimal assignments) to find the globally optimal assignment that minimizes the sum of the total costs in O(n³) time. The explicit introduction of chemical prior costs gives OCR post-processing a domain-adaptive capability not found in general OCR tools. In other words, even if a numerical candidate is slightly misaligned geometrically with an element header, as long as the latter is chemically unreasonable (e.g., assigning a value of 8.5 to boron, which should be on the order of 0.01), the matching algorithm will prioritize the chemically consistent assignment. The introduction of this chemical prior cost is the key innovation of this step.

[0113] Step 14: Output structured data with traceability tags.

[0114] In this step, each successfully matched element-component value pair is appended with four traceability tags: source file name, page number, line number, and original character bounding box coordinates. These tags, along with other fields such as durability performance and heat treatment status, are written into a structured dataset (such as a CSV file). The traceability chain formed by each structured data entry in the structured dataset allows process engineers to immediately trace back to the specific location in the original PDF for manual verification should any anomalies occur in subsequent model predictions. This forms the auditability basis of this invention.

[0115] In the OCR intelligent recognition process from steps 11 to 14 above, the layered OCR post-processing scheme is as follows: For each scanned page of the paper certificate of conformity and test report, the scanning page orientation is determined based on the scoring function; for each scanned page determined to be in the positive direction, adaptive row clustering is performed using a preset ratio of the median height of all character bounding boxes in the current scanned page as the neighborhood radius to obtain stable physical rows independent of font size; subsequently, the Hungarian bipartite graph matching algorithm is used to solve the globally optimal assignment of element headers and numerical candidates in the same physical row; the matching cost function considers both horizontal coordinate distance and chemical rationality information, and when a candidate value falls outside the reasonable content range of the element in the process, an additional penalty is given and rematching is performed, so that the final matching result is geometrically naturally aligned and materially self-consistent; each piece of structured data generated at the end also includes the source file name, page number, line number and original character box coordinates, forming a complete auditable traceability link.

[0116] Step 2, Small Sample Reliable Mapping

[0117] This small-sample reliable mapping step addresses the common real-world engineering conditions where master alloy composition-performance data often suffers from scarce samples (on the order of dozens of heats) and missing sampling items in a single heat, establishing a reliable mapping from composition to mechanical properties. The specific modeling scheme is developed in two scenarios depending on whether censored observations exist.

[0118] Scenario 1: If all mechanical property test data for the alloy are complete observations (i.e., each test continues until the specimen fractures naturally), this step directly uses Gaussian Process Regression (a nonparametric Bayesian regression method that can simultaneously output the predicted mean and predicted variance in small sample scenarios) as the mapping model from composition to performance. The Gaussian process kernel function used in this embodiment is the Matérn-5 / 2 kernel function, and its kernel function expression is:

[0119] ;

[0120] Where r = ||x−x′|| is the Euclidean distance between the input feature vectors x and x′; It is the core amplitude; It is a nuclear length scale; , All of these are learnable hyperparameters, solved by maximizing the marginal likelihood during the training phase. In one possible specific scheme, for targets with a large order of magnitude difference in lifetime, this invention first performs a logarithmic transformation to base 10 before training, thereby controlling the relative error of subsequent predictions within a reasonable range, and then restores the prediction by the corresponding power of 10 after the prediction.

[0121] Scenario 2: If the alloy exhibits censored observations due to the termination of the test before complete failure (e.g., in a typical 760℃ / 724MPa creep test, the sample breaks under tension after reaching the target number of hours), this step, in addition to Gaussian process regression, uses an accelerated failure time model (an accelerated failure time model, a parametric regression model commonly used in survival analysis) for parallel training. This accelerated failure time model expresses the logarithm of the creep life as a linear function of the input features plus a scale parameter and an error term. Its native log-likelihood form supporting censored observations is:

[0122] ;

[0123] in, It is an event indicator variable, with 1 for natural failure and 0 for failure under applied force; It is the observed lifespan; Let f be the component vector of the i-th record; f() is the probability density function, and S() is the survival function. This processing method prevents the training objective from being systematically distorted by miscounting samples that break under stress as accurate lifetime samples.

[0124] In both scenarios, this step further compares the outputs of various mainstream regression models, including linear regression (such as Ridge Regression, Lasso, Elastic Network), kernel regression, support vector regression, random forest, gradient boosting trees (XGBoost, LightGBM, CatBoost), and Gaussian process regression, using the same cross-validation conditions. It then automatically selects the base model with the best robustness on the given data as the input for downstream reverse recommendation, preventing downstream recommendations from being misled by a non-robust model. Gaussian process regression is consistently used as the default base model in this step because it can simultaneously provide the predicted mean and posterior variance in small sample scenarios, naturally supporting subsequent uncertainty-aware reverse search.

[0125] Step 3, Reverse Component Design

[0126] The reverse component design step is the ultimate goal of this invention. Based on the component-mechanical property reliability mapping model provided in the small-sample reliability mapping step, three independent or combined reverse search scenarios and corresponding search schemes are presented, covering the most common component optimization needs in practical engineering. The key algorithms used in the three scenarios are illustrated below.

[0127] Scenario 1: Back-engineering components based on specified target performance. Process engineers only need to specify the target value of one or more mechanical properties they wish to achieve, such as a creep life of at least a certain number of hours under a specific high-temperature condition. The back-engineering process in this scenario involves the following three steps:

[0128] Step 1: Using Latin hypercube sampling, N candidate components (typically 10000) are generated within the element content range allowed by the alloy's technical specifications. Latin hypercube sampling is a probabilistic sampling method that uniformly covers a high-dimensional interval. The process involves dividing each element content interval into N non-overlapping subintervals, randomly selecting a value from each subinterval, shuffling the order of the N values ​​for each dimension independently, and finally combining the k-th values ​​from each of the N dimensions into a single candidate component. This sampling method offers better spatial uniformity than ordinary random sampling and is particularly suitable for uniformly covering high-dimensional search spaces.

[0129] Step 2: For each of the N candidate components obtained from sampling, call the Gaussian process regression model trained in the previous section to give the predicted mean and predicted standard deviation of each candidate in terms of target performance, and calculate its 95% confidence interval accordingly.

[0130] Step 3: For all candidate components, calculate the Expected Improvement (EI) score for each component using the Expected Improvement Acquisition Function (EI, a commonly used acquisition function in Bayesian optimization), sort them by score, and select the candidate components with the highest EI scores as the recommended output. The expression for the Expected Improvement Acquisition Function in this step is:

[0131] ;

[0132] in, , These are the predicted mean and standard deviation of candidate component x by Gaussian process regression, respectively. It is the currently known optimal performance value; It involves exploring and utilizing the equilibrium constant, typically taken as 0.01; ; and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively. In other words, EI measures the expected performance improvement of a candidate component relative to the currently known best, taking into account both the level of the candidate prediction mean and the magnitude of the candidate prediction uncertainty. This allows for detailed exploration near known good points while also preserving necessary exploration capabilities in the component regions that have not yet been fully explored.

[0133] Scenario 2: Identify key elements and work backward to deduce other elements.

[0134] In this scenario, process engineers can lock one or more elements that are sensitive to cost or process constraints (typically precious elements such as hafnium, tantalum, and tungsten) at a specified value or range (i.e., set a lock value or lock range for the element), so that only the remaining elements are sampled by Latin hypercube within the constrained search space, and the rest of the process is the same as in scenario one.

[0135] In the specific implementation of constrained search in this embodiment of the invention, process engineers are also allowed to set upper and lower limit relaxation coefficients for individual elements (typically 0 to 0.15, meaning that the sampling range is allowed to be extended by a corresponding proportion below the original standard lower limit and above the upper limit), thereby supporting application scenarios where the content of precious elements is pushed outside the alloy standard range for cost control exploration. This scenario enables the design solution provided by this invention to flexibly serve real engineering constraints such as limited raw materials, alloy downgrading requirements, or cost control.

[0136] Scenario 3: Multi-objective Pareto frontier search.

[0137] When process engineers need to consider multiple mechanical performance indicators simultaneously (typically, optimizing creep life under two temperature conditions: 760℃ / 724MPa and 980℃ / 200MPa), a multi-objective genetic algorithm (NSGA-II, short for Non-dominated Sorting Genetic Algorithm II, a multi-objective evolutionary optimization algorithm based on non-dominated sorting) can be used to search for performance trade-off curves in the elemental content space. The basic concepts of this algorithm and the specific parameters used in this invention are described below.

[0138] Definition of non-dominated relation: For any two candidate components x1 and x2, if x1 is no worse than x2 on all objectives and is strictly better than x2 on at least one objective, then x1 is said to dominate x2. For a set of candidate components, if a candidate component is not dominated by any other candidate component in the set, then that candidate component belongs to the non-dominated solution; the set of all non-dominated solutions is called the Pareto front of the set of candidate components. In engineering semantics, the Pareto front is the optimal trade-off curve among multiple objectives, and process engineers can select specific component recommendations based on actual engineering preferences on this curve.

[0139] The key parameter of the NSGA-II algorithm used in scenario three is: population size N. pop =100; Evolutionary generation N gen =50; the mutation uses multinomial mutation, with a mutation probability p. m =1 / d x (d) x (where η is the element dimension) and the distribution index η m =20; Crossover uses simulated binary crossover (SBX), with a crossover probability p. c =0.9, distribution index η c =15. In each generation, non-dominated sorting, crowding calculation, tournament selection, crossover, mutation and elite retention are performed sequentially on the current population, and finally the Pareto front candidate set is output instead of a single recommendation, so that process engineers can make explicit engineering trade-offs between multiple performance indicators.

[0140] To prevent the reverse search from jumping out of the true support domain of the training data and outputting physically unreliable component points in the above three scenarios, this invention multiplies the final recommendation ranking in all three scenarios by an out-of-distribution suppression factor constructed based on Mahalanobis distance (a distance metric that considers the correlation structure between features, proposed by Mahalanobis in 1936). The analytical expression for this distance is:

[0141] ;

[0142] in, and These are the mean vector and covariance matrix of the training set element vectors, respectively, with T denoteing the transpose. Compared to simple Euclidean distance, Mahalanobis distance explicitly utilizes the training set covariance matrix, correctly considering the correlation structure between elements and avoiding misclassifying reasonable candidates close to the main axis of the training set as out-of-distribution samples. In this embodiment of the invention, the out-of-distribution suppression factor is defined as... , where α is the suppression strength coefficient (e.g., α=0.05). Multiplying this out-of-distribution suppression factor by the original EI score or Pareto ranking score and then re-ranking can effectively eliminate extreme candidate components that deviate from the reasonable distribution range of the training set.

[0143] Beyond out-of-distribution suppression, step 3 of this invention further employs conformal prediction (a method proposed by Vovk et al. in 2005 that calibrates the uncertainty of model prediction output based on empirical residual quantiles) to recommend a 95% confidence interval with calibrable statistical significance for each component. Specifically, this invention performs inductive conformal prediction on the training set: a portion of the training set is reserved as a calibration set, and the absolute value of the prediction residual is calculated for each sample in the calibration set. The 95th quantile qt of the empirical distribution of this absolute residual is taken as the half-width of the conformal prediction interval for that target. The final recommendation interval width is the maximum of the Gaussian process posterior interval (1.96σ) and the conformal quantile qt. Thus, when the component recommendation deviates from the training set, the interval automatically expands with the Gaussian process posterior and maintains the nominal 95% coverage lower bound in any situation, ensuring that the final recommendation not only provides a point estimate but also clearly defines the credible range of that point estimate.

[0144] Based on the OCR- and few-sample learning-based method for designing the composition of high-temperature cast alloys provided in the above embodiments of the invention, such as Figure 2 As shown, embodiments of the present invention also provide a casting high-temperature alloy composition design system based on OCR and few-sample learning for performing the above-described method.

[0145] The system provided in this embodiment of the invention includes: an OCR intelligent visual recognition module, a structured data module, a trusted mapping model library module, a reverse design module, and a visualization front-end module. These five modules are coupled through a unified element content field and a mechanical property field as an interface. Specifically, the OCR intelligent visual recognition module performs step 1, converting scanned copies of paper reports into structured data; the structured data base module stores and manages the structured dataset and traceability tags; the trusted mapping model library module performs step 2, training, storing, and managing the component-mechanical property trusted mapping model; the reverse design module performs step 3, calling the trusted mapping model to perform reverse search based on user-input engineering constraints or performance targets, generating a list of recommended preferred components; and the visualization front-end module provides a graphical user interface, supporting users to import original reports, review structured results, set reverse design parameters, view component recommendations and their confidence intervals, and trace and verify the source files.

[0146] In this embodiment of the invention, the visualization front-end module is implemented based on a web framework (such as Streamlit), enabling process engineers to complete tasks on the shop floor, including importing original reports, reviewing structured results, setting reverse recommendations for target performance or element constraints, displaying the predictive performance and confidence intervals of component recommendations, and tracing the source PDF files. The back-end maintains decoupling from the specific OCR engine, machine learning library, and optimization solver used, allowing deployment in a confidential production environment based on localized open-source components, avoiding dependence on external cloud services.

[0147] This invention provides a method and system for designing casting high-temperature alloy compositions based on OCR and few-sample learning, specifically an end-to-end reverse design method and system for master alloy compositions. Using the technical solution provided by this invention, implicit data assets scattered across historical paper certificates and test reports can be automatically structured through OCR intelligent visual recognition, modeled through few-sample reliable mapping, and ultimately transformed into executable master alloy optimal composition recommendations without changing alloy grades or relying on large-scale new experiments. Each recommendation is accompanied by a statistically significant confidence interval, enabling process engineers to robustly and continuously optimize master alloy compositions within the limits of current technology. The technical solution provided by this invention is alloy grade independent; the three core steps in the solution can be easily transferred to various casting high-temperature alloy master alloy systems. Furthermore, this invention offers the following beneficial effects:

[0148] Firstly, the composition and performance data of the master alloy, previously scattered in paper certificates of conformity and test reports and long neglected by the industry, are transformed into calculable, traceable, and governable structured data assets through a layered OCR post-processing scheme that combines automatic orientation correction, adaptive row clustering, and Hungarian matching of chemical prior costs. This establishes an end-to-end data governance entry point in the context of master alloy production reports, which have long lacked algorithmically readable formats. Secondly, the explicit introduction of chemical prior costs is one of the key technical details of this invention. The cost function explicitly utilizes the reasonable content range of each element in the casting high-temperature alloy as prior information, giving the layered OCR post-processing scheme a domain-adaptive capability not found in general-purpose OCR tools.

[0149] Secondly, considering the inherent small sample size in production reports of cast high-temperature alloy master alloys, and taking into account the censored observations in some alloy systems, a reliable mapping scheme is established that simultaneously possesses the ability to estimate unbiased performance and quantify reliable uncertainty, forming a reliable foundation for downstream reverse search. Specifically, this invention employs Matérn-5 / 2 kernel Gaussian process regression for complete observation scenarios, and trains an accelerated failure time survival model in parallel for censored observation scenarios. Furthermore, it performs horizontal comparisons of multiple mainstream regression models under the same cross-validation to automatically select the most robust base model.

[0150] Thirdly, the technical solution of this invention proposes three reverse search scenarios that can be invoked independently or in combination: Scenario 1 is to reverse-engineer the composition by specifying the target performance; Scenario 2 is to reverse-engineer other elements by locking key elements, supporting controlled exploration beyond the upper and lower limits of the original alloy standard; Scenario 3 is a multi-objective Pareto front search, using NSGA-II for non-dominated sorting in the element content space. All three scenarios use an out-of-distribution suppression factor based on Mahalanobis distance and a calibrable confidence interval based on conformal prediction for final filtering and statistical guarantee, ensuring that the composition recommendation has both engineering feasibility and calibrable statistical significance.

[0151] Fourth, the essence of the technical solution of this invention, which differs from existing single-step solutions, lies in connecting the above three steps in the order of OCR intelligent visual recognition, small-sample reliable mapping, and reverse composition design, forming a complete closed loop for optimizing the composition of casting high-temperature alloy master alloys. The casting high-temperature alloy composition method and system provided by this invention maintains decoupling from alloy grades and can be easily transferred to various casting high-temperature alloy master alloy systems.

[0152] The following examples illustrate the implementation of the casting high-temperature alloy composition design method and system based on OCR and few-sample learning provided by this invention.

[0153] The following implementation examples correspond to three typical engineering reverse design scenarios. Each implementation example sequentially executes the three core steps of the design method provided by this invention: First, in the OCR-based intelligent visual recognition step, the original paper PDF report is converted into a structured dataset; then, in the small-sample reliable mapping step, a component-mechanical property reliable mapping model with uncertainty quantification capability is trained on this structured dataset; finally, in the reverse component design step, a recommended list of engineering-executable components is output.

[0154] Implementation Example 1:

[0155] Scenario 1: K447A master alloy, composition deduced from the specified target durability.

[0156] (1) Intelligent visual recognition process.

[0157] Example 1 uses 48 heats of K447A master alloy certificates and test reports accumulated by a research institution over the years, all archived as scanned PDFs. A typical PDF consists of two to four pages, containing alloy grade information, chemical composition analysis, post-heat treatment creep test results, and quality inspection conclusions. Example 1 first determines the orientation of each scanned page using a combination scoring function, automatically flipping inverted pages. In a typical batch, there are four inverted scanned pages. The orientation scoring function gives a score of 0.83 × 1.42 = 1.18 for positive orientation candidates and only 0.41 × 1.05 = 0.43 for inverted orientation candidates, thus robustly completing orientation correction. After orientation correction, adaptive row clustering is performed on the bounding boxes of the characters identified on each scanned page. The median height of the character bounding boxes identified in a typical chemical composition table is 14.2 pixels. Therefore, a neighborhood radius ε = 0.55 × 14.2 ≈ 7.8 pixels is taken, with a minimum... samples =2 calls DBSCAN, resulting in 11 stable physical rows (corresponding to 11 elements). Then, Hungarian matching with chemical prior cost is performed. In a typical physical row, the element header Hf is located at x=412 pixels, and the numerical candidates identified in the same row include 1.51 (located at x=408 pixels) and 8.43 (located at x=215 pixels).

[0158] If only the horizontal distance is considered, 1.51 clearly wins. However, the cost function in Example 1 also considers chemical rationality: 1.51 falls within the reasonable Hf content range [0.3, 1.8] in K447A, so prior_mismatch=0; while 8.43 far exceeds the reasonable Hf range, so prior_mismatch=1, resulting in an additional β=0.5 penalty, thus prioritizing the physically consistent assignment. This step yielded 137 traceable structured records from 48 PDFs.

[0159] (2) Small sample reliable mapping step.

[0160] In this implementation example 1, 137 structured data points were used as input feature vectors for 11 element contents and several mechanical properties as targets to establish independent regression tasks. Considering that the creep rupture life of the K447A master alloy can terminate under two conditions—natural fracture and stressed fracture—at 760℃ / 724MPa and 980℃ / 200MPa, respectively, this step, following the specific scheme in step 2, trained Gaussian process regression and accelerated failure time survival models for the two conditions. In the Matérn-5 / 2 kernel function used, the length scale... The kernel amplitude is solved by the L-BFGS-B optimizer using a multi-restart search (nrestarts=4) on the marginal likelihood function; the durability lifetime is transformed logarithmically to base 10. Under 5-fold cross-validation, this step simultaneously trains and compares 17 mainstream regression models, among which GPR-Matérn has a mean absolute error of 7.56 hours for leave-one-out cross-validation on a durability lifetime of 760℃, and XGBoost has a mean absolute error of 5.38 hours; accordingly, this implementation example 1 selects XGBoost as the point estimation basis model and GPR-Matérn as the uncertainty estimation basis model, and uses them in parallel.

[0161] (3) Reverse component design stage.

[0162] The process engineer selects 980℃ / 200MPa as the target test condition at the front end and inputs the desired minimum creep life value of 80 hours. The system executes the following process within the element content range allowed by the technical conditions of K447A alloy ingot: S1, generates 10,000 candidate components in the 11-dimensional element range by Latin hypercube sampling in step 3; S2, predicts the mean creep life and 95% confidence interval of each candidate under the condition of 980℃ / 200MPa using the GPR-Matérn model obtained in step (2) above; S3, removes candidates with a predicted lower limit of 80 hours; S4, outputs the top 15 component recommendations by sorting the comprehensive score of EI score and Mahalanobis distance of the training set. The system also provides the accompanying performance prediction of each component recommendation under the condition of 760℃ / 724MPa. In a typical operation, the first component recommendation output has Hf content of 1.62wt%, Ta content of 3.28wt%, and Mo content of 0.72wt%. The predicted average durability lifetime at 980℃ / 200MPa is 88.4 hours, the 95% conformal prediction confidence interval is [78.1, 99.2] hours, and the Mahalanobis distance from the training set is 1.84, indicating that the prediction falls within a reasonable and credible range.

[0163] Implementation Example 2:

[0164] Scenario 2: K447A master alloy, locking in the precious element hafnium to infer other elements

[0165] (1) Intelligent visual recognition process.

[0166] This embodiment 2 directly reuses the 137 structured datasets obtained in embodiment 1, without repeating the PDF recognition and Hungarian matching process; the execution details of its orientation correction, adaptive row clustering and chemical prior cost matching are the same as in embodiment 1, thus demonstrating that the structured data governance results of this invention can be reused in multiple reverse scenarios without repeated collection and manual input.

[0167] (2) Small sample reliable mapping step.

[0168] This embodiment, Example 2, also reuses the GPR-Matérn and XGBoost dual-base model obtained in Example 1, without repeating the training. It is important to note that, since the subsequent inverse search in Example 2 allows pushing the hafnium content outside the alloy standard range, the probability of the recommended composition being located in the sparse region of the training set increases. In this case, Example 2 relies on the calibrable 95% confidence interval jointly provided by the posterior variance of GPR-Matérn and the conformal prediction residual calibration to honestly quantify the uncertainty of out-of-bounds components, avoiding the engineering risk of downstream recommendations that appear precise but are actually baseless.

[0169] (3) Reverse component design stage.

[0170] In this Example 2, considering the high production cost of hafnium, the process engineer wanted to evaluate how to adjust other elements to maintain mechanical properties while significantly reducing hafnium content. Example 2 locks hafnium at the front end and inputs a specified value Hf=0.50wt% (significantly lower than the K447A standard lower limit of 1.20wt%). Upon receiving this constraint, the system performs Latin hypercube sampling within the process-allowed range of the remaining 10 elements using the reverse search scheme of Scenario 2 (this Example 2 allows for a ±15% extension of the original standard upper and lower limits for other principal elements to support collaborative compensation search in cost-reduction scenarios for expensive elements). The GPR-Matérn model predicts the mean performance of candidate components and the 95% conformal prediction interval. The top 15 components are then recommended based on a combined score of EI score and Mahalanobis distance from the training set. All recommendations output in Example 2 strictly satisfy the hard constraint that hafnium content is always equal to 0.50wt%. In a typical operation, the first component recommendation output by the algorithm was adjusted to 3.42 wt% for Ta, 1.05 wt% for Ti, and slightly adjusted to 0.75 wt% for Mo. The predicted average creep life at 980℃ / 200MPa was 71.6 hours, and the 95% conformal prediction confidence interval was [58.3, 85.9] hours. The lower limit of the prediction interval still met the acceptance threshold of ≥50 hours, and the Mahalanobis distance from the training set was 2.46. Therefore, the process engineer was advised that this out-of-bounds recommendation was a candidate for "medium risk, small-batch trial production verification recommended".

[0171] Implementation Example 3:

[0172] Scenario 3: Another type of master alloy, multi-objective Pareto frontier search

[0173] (1) Intelligent visual recognition process.

[0174] This implementation example 3 uses a cast high-temperature alloy master alloy (K424) as the carrier, and employs 32 heats of K424 master alloy qualification certificates and test reports accumulated over the years by a certain production unit as the raw data. Compared with implementation example 1, the parameters and processes of the OCR-based intelligent visual recognition step in this implementation example 3 remain unchanged, but the reasonable content range of the process used in the chemical prior cost is replaced with the technical conditions of K424 alloy ingots. For example, the reasonable content range of Cr element in K424 is [14.0, 16.0] wt% (K447A is [8.0, 9.0] wt%), and the Hungarian matching priormismatch function data is automatically reset; the rest of the process remains unchanged. This step yields 95 traceable K424 structured records from 32 PDFs.

[0175] (2) Small sample reliable mapping step.

[0176] In this implementation example 3, a composition-performance mapping model is trained on 95 K424 data points according to the specific scheme in step 2. Considering that the creep test data of K424 alloy are all complete observations and there is no censoring caused by tensile fracture, this step only uses Matérn-5 / 2 kernel Gaussian process regression as the base model, and no accelerated failure time survival model is trained. Under the same 5-fold cross-validation, a horizontal comparison of 17 mainstream regression models shows that GPR-Matérn and XGBoost are still the two most robust base models on K424 data; therefore, this implementation example 3 continues the strategy of using both in parallel as in implementation example 1.

[0177] (3) Reverse component design stage.

[0178] In the scenario of Example 3, the process engineer needs to simultaneously consider the crease lifetime under two temperature conditions: 760℃ / 724MPa and 980℃ / 200MPa. Therefore, a multi-objective Pareto front search is selected at the front end. Within the elemental content range allowed by the technical conditions of K424 alloy ingots, the system performs Latin hypercube sampling according to the inverse search scheme of Example 3 and executes NSGA-II multi-objective evolutionary optimization. The key parameters are: population size 100, number of generations 50, mutation probability 1 / d, and simulated binary crossover probability 0.9. After NSGA-II convergence, the system outputs a Pareto front containing 23 candidate components. Each candidate includes the predicted mean of its two objective performances, a 95% conformal prediction interval, and the Mahalanobis distance between the candidate and the training set. The process engineer can select specific component recommendations based on actual engineering preferences (e.g., a preference for higher crease lifetime at 760℃ or a more balanced crease lifetime under the two temperature conditions) on this Pareto front and make explicit engineering trade-offs regarding the credibility of each candidate.

[0179] The three implementation examples mentioned above all start from the same set of invention solutions and cover the three most common engineering reverse engineering requirements in the optimization of the composition of casting high-temperature alloy master alloys. The execution of each implementation example completes the three core steps of AI intelligent visual recognition, small sample reliable mapping, and reverse composition design. The entire process can be completed on ordinary workstations without relying on large-scale computing power and can be directly delivered and used in the workshop.

[0180] While the embodiments disclosed in this invention are as described above, they are merely illustrative of the embodiments to facilitate understanding of the invention and are not intended to limit the invention. Any person skilled in the art to which this invention pertains may make any modifications and variations in the form and details of the implementation without departing from the spirit and scope disclosed herein; however, the scope of patent protection for this invention shall still be determined by the scope defined in the appended claims.

Claims

1. A method for designing the composition of casting high-temperature alloys based on OCR and few-shot learning, characterized in that, include: Step 1, the intelligent visual recognition step based on optical character recognition (OCR), includes: scanning the paper certificates of conformity and test reports of the cast high-temperature alloy master alloy, and converting them into a structured dataset with traceability tags through a layered OCR post-processing scheme; the layered OCR post-processing scheme includes the sequential execution of automatic page orientation correction, adaptive line clustering, Hungarian matching with chemical prior costs, and structured data output with traceability tags; Step 2, the small sample reliable mapping step, includes: based on the structured dataset, establishing and training a component-mechanical property reliable mapping model that can simultaneously output the prediction mean and prediction uncertainty; Step 3, the reverse composition design stage, includes: based on the composition-mechanical property confidence mapping model, performing a reverse search within the process feasibility domain allowed by the casting high-temperature alloy technology conditions, and outputting a list of preferred compositions with statistical confidence intervals; In step 2, a reliable mapping model of composition and mechanical properties is established. Two schemes are formed based on whether censored observations exist in the data: Option 1: For cases where all mechanical property data of the alloy are fully observed, Gaussian process regression is used as the mapping model from composition to properties, and the kernel function is the Matérn-5 / 2 kernel function. The hyperparameters of the kernel function are solved by maximizing the marginal likelihood. Option 2: For censored observations caused by terminating the test before the alloy sample completely fails, in addition to Gaussian process regression, an accelerated failure time model is trained in parallel; the training objective of this accelerated failure time model is constructed based on the log-likelihood function containing censored indicator variables. In step 2, the expression for the Matérn-5 / 2 kernel function is: ； Where r = ||x−x′|| is the Euclidean distance between the input feature vectors x and x′; It is the core amplitude; It is a nuclear length scale; In step 2, the expression for the accelerated failure time model is: ； in, It is an event indicator variable, with 1 for natural failure and 0 for failure under applied force; It is the observed lifespan; is the component vector of the i-th record; f() is the probability density function, and S() is the survival function.

2. The method for designing the composition of casting high-temperature alloys based on OCR and few-sample learning according to claim 1, characterized in that, The layered OCR post-processing scheme in step 1 includes: Step 11, automatic page orientation correction, includes: performing OCR recognition on each scanned page in multiple candidate orientations, calculating the orientation score of each candidate orientation using a combined scoring function based on the average confidence of characters and the proportion of Chinese characters, and determining the candidate orientation with the highest score as the positive orientation; Step 12, adaptive row clustering, includes: based on the Y-direction coordinates of the character bounding boxes identified on the scanned page, adaptively determining the neighborhood radius of density clustering with the median height of characters on the scanned page as a reference, and executing the density clustering algorithm to divide the physical rows of the scanned page; Step 13, Hungarian matching of chemical prior costs, includes: modeling the alignment problem of element headers and numerical candidates in the same physical row to obtain the optimal allocation problem of bipartite graph, whose cost function includes the normalized distance of the horizontal axis and chemical prior terms; Step 14, output structured data with source tags, including: appending source file information and location coordinates to the successfully matched element-component numerical pairs and writing them into the structured dataset.

3. The method for designing the composition of casting high-temperature alloys based on OCR and few-sample learning according to claim 2, characterized in that, The combined scoring function in step 11 is defined as follows: score=mean conf ×(1+0.5×ratio zh ); Where score is the directional score of the candidate direction, and mean conf The ratio represents the average recognition confidence of all characters in the candidate recognition results. zh This represents the percentage of Chinese characters in the candidate recognition results.

4. The method for designing the composition of casting high-temperature alloys based on OCR and few-sample learning according to claim 3, characterized in that, The adaptive method for determining the neighborhood radius in step 12 is as follows: Calculate the height of the bounding boxes of all characters on the currently scanned page, and denot the median as h. med The neighborhood radius of density clustering is calculated as ε: ε=0.55×h med ; Wherein, the minimum number of neighborhood samples in the density clustering algorithm is min samples Set to 2; A density clustering algorithm is performed on the Y-coordinates of all characters in the currently scanned page. Characters belonging to the same cluster are identified as belonging to the same physical row.

5. The method for designing the composition of casting high-temperature alloys based on OCR and few-sample learning according to claim 4, characterized in that, In step 13, the bipartite graph optimal allocation problem is modeled and represented by a cost matrix. The element in the i-th row and j-th column of the cost matrix represents the cost of assigning the i-th element to the j-th numerical candidate. The cost function used is specifically defined as follows: cost(i,j)=α·Δx norm (i,j)+β·prior_mismatch(value j ∈range i ); Where, Δx norm (i,j) is the normalized distance between the x-coordinate center of the i-th element header and the x-coordinate center of the j-th numerical candidate; prior_mismatch is a chemical prior, which is the distance between the x-coordinate center of the i-th element header and the j-th numerical candidate. j The reasonable content range of the i-th element in the process. i The value is 0 if the condition is met, otherwise it is 1; α and β are weighting constants. After obtaining the cost matrix, the Hungarian algorithm is used to find the globally optimal assignment that minimizes the total cost.

6. The method for designing the composition of casting high-temperature alloys based on OCR and few-sample learning according to claim 5, characterized in that, Step 14 includes: Each successfully matched element-component pair is appended with four source tags: source file name, page number, line number, and original character bounding box coordinates. These tags, along with the persistence performance and heat processing status fields, are written into the structured dataset.

7. The method for designing the composition of casting high-temperature alloys based on OCR and few-sample learning according to claim 1, characterized in that, Step 2 also includes: Under the same cross-validation conditions, the outputs of various regression models, including linear regression, kernel regression, support vector regression, random forest, gradient boosting tree, and Gaussian process regression, are compared horizontally. The model with the best prediction accuracy is automatically selected as the base model and used as the input for downstream reverse recommendation. Gaussian process regression is used as the base model for uncertainty estimation by default.

8. The method for designing the composition of casting high-temperature alloys based on OCR and few-sample learning according to claim 1, characterized in that, The reverse search in step 3 includes: executing a reverse search scheme in scenario one, where scenario one involves: specifying the target performance and back-deriving the components; the reverse search process in scenario one includes: The first step is to use Latin hypercube sampling to generate N candidate compositions within the element content range allowed by the alloy technology conditions; The second step is to call the trained component-mechanical property reliability mapping model to predict the predicted mean μ(x) and predicted standard deviation σ(x) of each candidate component on the target performance; The third step involves using the expected improvement acquisition function to calculate the expected improvement EI score for each candidate component, sorting them by score, and selecting the candidate components with the highest expected improvement EI scores as the recommended output.

9. The method for designing the composition of casting high-temperature alloys based on OCR and few-sample learning according to claim 8, characterized in that, In the third step, the expression for improving the acquisition function is: ； in, , These are the predicted mean and standard deviation of candidate component x by Gaussian process regression, respectively. It is the currently known optimal performance value; It is about exploring and utilizing the equilibrium constant; ; , These are the cumulative distribution function and probability density function of the standard normal distribution, respectively.

10. The method for designing the composition of casting high-temperature alloys based on OCR and few-sample learning according to claim 8, characterized in that, The reverse search in step 3 further includes: executing the reverse search scheme under scenario two, which is: locking key elements and deducing other elements from them; the reverse search process in scenario two includes: Using the lock values ​​or lock ranges set by process engineers for the content of one or more elements as constraints, sampling and reverse search are performed only on elements for which no lock values ​​or lock ranges are set within the constrained search space.

11. The method for designing casting high-temperature alloy composition based on OCR and few-sample learning according to claim 10, characterized in that, The reverse search process in scenario two also includes: Based on the upper and lower relaxation coefficients set by process engineers for individual elements, the proportion represented by the relaxation coefficient is expanded by extending the sampling range of the element below the lower limit and above the upper limit of the original alloy standard, thereby pushing the content of precious elements outside the alloy standard range for cost control exploration.

12. The method for designing the composition of casting high-temperature alloys based on OCR and few-sample learning according to claim 10, characterized in that, The reverse search in step 3 further includes: executing the reverse search scheme under scenario three, which is: multi-objective frontier search; the reverse search process for scenario three includes: For design scenarios that require simultaneous consideration of multiple mechanical performance indicators, the multi-objective genetic algorithm NSGA-II is used to search for the Pareto front solution set in the element content space that optimizes multiple mechanical performance indicators simultaneously. The parameters of the NSGA-II algorithm include: N pop =100; Evolutionary generation N gen =50; the mutation uses multinomial mutation, with a mutation probability p. m =1 / d x d x Let η be the element dimension and the distribution index η. m =20; Crossover uses simulated binary crossover, with a crossover probability p. c =0.9, distribution index η c =15.

13. The method for designing the composition of casting high-temperature alloys based on OCR and few-sample learning according to claim 12, characterized in that, Before outputting the recommended list of preferred ingredients, step 3 also includes: Constructing an out-of-distribution suppression factor based on Mahalanobis distance includes: multiplying the out-of-distribution suppression factor by the original EI score or Pareto ranking score and then re-ranking the data to remove candidate components that deviate from the reasonable distribution range of the training data; the out-of-distribution suppression factor is defined as follows: Where α is the inhibition intensity coefficient, The Mahalanobis distance between candidate component x and the training set; Mahalanobis distance The calculation formula is: ； in, and These are the mean vector and covariance matrix of the training set element vectors, respectively, and T denotes the transpose.

14. The method for designing the composition of casting high-temperature alloys based on OCR and few-sample learning according to claim 1, characterized in that, Step 3 also includes: Using a conformal prediction method, based on the empirical distribution of prediction residuals on the calibration set, a statistically significant confidence interval is recommended for each component; the final width of the confidence interval is the maximum of the width of the posterior interval of the Gaussian process regression and the width of the conformal prediction quantile.

15. A casting high-temperature alloy composition design system based on OCR and few-shot learning, characterized in that, The system is used to execute the casting high-temperature alloy composition design method based on OCR and few-sample learning as described in any one of claims 1 to 14, characterized in that the system comprises: The OCR intelligent visual recognition module is used to perform step 1, converting the scanned paper report into structured data; The structured data foundation module is used to store and manage the structured dataset and traceability tags; A trusted mapping model library module is used to perform step 2, train, store, and manage the component-mechanical property trusted mapping model; The reverse design module is used to execute step 3, which calls the trusted mapping model to perform reverse search based on the engineering constraints or performance goals input by the user, and generates a list of recommended preferred components. The visualization front-end module provides a graphical user interface that allows users to import original reports, review structured results, set reverse design parameters, view component recommendations and their confidence intervals, and trace and verify source files. The trusted mapping model library module performs step 2 to establish a component-mechanical property trusted mapping model, and forms two schemes based on whether the data has censored observations: Option 1: For cases where all mechanical property data of the alloy are fully observed, Gaussian process regression is used as the mapping model from composition to properties, and the kernel function is the Matérn-5 / 2 kernel function. The hyperparameters of the kernel function are solved by maximizing the marginal likelihood. Option 2: For censored observations caused by terminating the test before the alloy sample completely fails, in addition to Gaussian process regression, an accelerated failure time model is trained in parallel; the training objective of this accelerated failure time model is constructed based on the log-likelihood function containing censored indicator variables. The expression for the Matérn-5 / 2 kernel function is: ； Where r = ||x−x′|| is the Euclidean distance between the input feature vectors x and x′; It is the core amplitude; It is a nuclear length scale; In step 2, the expression for the accelerated failure time model is: ； in, It is an event indicator variable, with 1 for natural failure and 0 for failure under applied force; It is the observed lifespan; is the component vector of the i-th record; f() is the probability density function, and S() is the survival function.

16. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, is used to implement the casting high-temperature alloy composition design method based on OCR and few-sample learning as described in any one of claims 1 to 14.

Images