A performance prediction system for high-strength low-alloy steel for marine engineering

Abstract

This invention relates to the field of data processing technology and discloses a performance prediction system for high-strength low-alloy steel used in marine engineering. The system includes: a marine engineering process functional module for generating continuous temperature process functional objects and process derivative functionals from discrete time-series sensor data; an alloy dynamic topology module for converting microstructure image data of high-strength low-alloy steel into multi-dimensional spatial coordinate point clouds and generating persistent graph data by combining the process derivative functional; a topology constraint registration module for generating topological affinity matrix elements as penalty constraints based on the persistent graph data and generating a functional principal component score vector with microstructure topological space penalty attributes by combining the continuous temperature process functional object; and a marine performance prediction module for generating a persistent landscape vector from the persistent graph data and the functional principal component score vector, outputting a macroscopic mechanical performance prediction value. This invention can reduce the extrapolation prediction bias of small samples and improve the accuracy of mechanical performance prediction.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a performance prediction system for high-strength low-alloy steel used in marine engineering. Background Technology

[0002] Currently, marine engineering places stringent requirements on the mechanical properties of high-strength low-alloy steel. Thermomechanical control processes are crucial in determining its final quality. Complex rolling and nonlinear cooling operations drive intense phase transformation processes within the material. The dynamic evolution of the underlying microstructure directly maps to and determines the steel's low-temperature impact toughness and room-temperature yield strength. Establishing high-fidelity numerical models for mechanical properties has become a critical technological step in shortening the development cycle of new alloys and ensuring quality control in industrial production lines.

[0003] Regarding the aforementioned issues, conventional prediction techniques primarily rely on discretely acquired process characteristic parameters. Technicians typically extract the final rolling temperature at a specific time point or calculate the average cooling rate across the entire process. At the microscopic level, static two-dimensional image raster data is acquired using electron microscopy. During spatial feature analysis, the system pre-determines a fixed static filtering distance tolerance to construct a simple complex structure in multidimensional space. Subsequently, traditional statistical or machine learning regression algorithms are deployed at the computing center. The prediction model directly uses the scalarized process inputs as independent variables, traversing the intermediate physical evolution process to establish a single-step end-to-end mapping fitting mechanism from process parameters directly to macroscopic mechanical properties.

[0004] This analytical model deviates from the objective laws of physical metallurgical evolution. Scalar extraction inevitably filters out transient dynamic information in continuous process trajectories; the nonlinear cooling evolution characteristics caused by latent heat of phase transformation on the time-series curves are completely masked. The static spatial connectivity tolerance setting severs the causal intervention mechanism between macroscopic instantaneous cooling acceleration and microscopic grain boundary network generation density. When performing multimodal data feature compression, the lack of effective constraints from underlying microscopic spatial topology metrics makes the model prone to falling into the ill-conditioned inverse mapping dilemma of data dimensionality reduction. Time-series features with different underlying microstructural evolution paths are often forcibly aliased in similar low-dimensional coordinate systems; this amplifies the deviation in the performance extrapolation of data extrapolation models based on finite sample-driven models when facing new process parameters.

[0005] Therefore, the present invention provides a performance prediction system for high-strength low-alloy steel for marine engineering, in order to overcome the shortcomings of the prior art. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a performance prediction system for high-strength low-alloy steel used in marine engineering. This system solves the problem that existing prediction methods, by severing the physical causal relationship between the dynamic characteristics of continuous processes and microscopic topological evolution, lead to ill-conditioned inverse mappings during dimensionality reduction of multimodal features, resulting in excessively large deviations in the extrapolation prediction of macroscopic mechanical properties under small sample conditions.

[0007] To achieve the above objectives, the present invention provides a performance prediction system for high-strength low-alloy steel used in marine engineering, comprising: The Marine Engineering Process Functional Module is used to generate continuous temperature process functional objects and process derivative functionals from time-series sensor data of discrete thermomechanical control processes. The alloy dynamic topology module is used to convert the microstructure image data of high-strength low-alloy steel into a multi-dimensional spatial coordinate point cloud, and combine the chemical composition properties of high-strength low-alloy steel with the process derivative functional to generate persistent graph data. The topology constraint registration module is used to generate topology affinity matrix elements as penalty constraints based on the persistent graph data, and to generate functional principal component score vectors with microstructure topology space penalty attributes in combination with the continuous temperature process functional object. The ship and marine performance prediction module is used to generate a persistent landscape vector by combining the persistent graph data with the functional principal component score vector with micro-organization topological space penalty attribute, and output the macro-mechanical performance prediction value.

[0008] By adopting the above technical solution, the discrete sensor data is reconstructed into a continuous functional object and the evolution derivative is extracted through functionalization processing. At the same time, a dynamic filtering mechanism is established to intervene in the growth of microscopic topological complexes by process derivatives. In the dimension reduction and registration stage, topological affinity is introduced as a penalty constraint for manifold mapping. Finally, the deduction is performed through a two-stage regression network. Therefore, the physical causal chain between thermomechanical control process, microscopic topological evolution and macroscopic mechanical properties is strictly followed, thereby reducing the extrapolation bias of small samples and improving the prediction accuracy of mechanical properties.

[0009] Preferably, the marine process functional module includes a basis function projection calculation unit, used to: smoothly fit the time-series sensor data of discrete thermomechanical control processes using multi-order spline basis functions; and during the smooth fitting process, introduce a least squares method with a roughness penalty term for parameter estimation to generate the continuous temperature process functional object.

[0010] By adopting the above technical solution, the curve overfitting caused by sensor noise is suppressed by using the second derivative integral penalty term, thus restoring the true physical characteristics of material temperature monotonically decreasing and phase transition recovery.

[0011] Preferably, the marine process functional module includes a differential calculator, used to: perform a first-order derivative operation in the time dimension on the continuous temperature process functional object to obtain a characteristic variable characterizing the instantaneous cooling rate change; and transform the static temperature curve into the process derivative functional containing transient dynamic characteristics by continuous differentiation operation and combining the characteristic variable.

[0012] By adopting the above technical solution, the instantaneous cooling rate mutation law caused by the release of latent heat from phase transformation in the internal structure of high-strength low-alloy steel is extracted, realizing the transformation of static time series data into dynamic process state characteristics.

[0013] Preferably, the alloy dynamic topology module includes an image point cloud processing unit, used to: perform grayscale threshold segmentation and edge contour detection operations on the microstructure image data of the high-strength low-alloy steel sample, extract feature points of large-angle grain boundary distribution boundaries with crystallographic orientation differences and isolated boundaries of precipitates; map each extracted boundary feature point to a preset multidimensional Euclidean space coordinate system to generate the multidimensional space coordinate point cloud.

[0014] By adopting the above technical solution, the pixel features of two-dimensional microscopic tissue images are extracted and transformed into a set of basic geometric vertices that meet the requirements of topological complex operations.

[0015] Preferably, the alloy dynamic topology module includes a simple complex generation unit, used for: extracting the critical time node of phase transformation based on the chemical composition properties of the high-strength low-alloy steel sample; extracting the value of the process derivative functional at the critical time node of phase transformation and performing absolute value calculation to obtain the instantaneous cooling acceleration value; inputting the instantaneous cooling acceleration value as an independent variable into a negative exponential mapping function for mathematical deduction to obtain a dynamic spatial filtering threshold parameter; and setting the connectivity tolerance according to the dynamic spatial filtering threshold parameter, constructing a simple complex structure in the multidimensional spatial coordinate point cloud, and extracting topological invariant features to complete the generation of the persistent graph data. Specifically, during the mathematical deduction calculation, the set dynamic spatial filtering threshold parameter is equal to a pre-set basic filtering threshold constant multiplied by a power function with a natural constant as the base and the negative exponent of the product of the material phase transformation sensitivity coefficient and the instantaneous cooling acceleration value.

[0016] By adopting the above technical solution, a physical mapping intervention mechanism for the process derivative on the microscopic topological geometric connection constraint is established, and the direct influence of thermomechanical control on the local grain boundary network generation density is quantified.

[0017] Preferably, the simple complex generation unit is further configured to use the multidimensional spatial coordinate point cloud as a vertex set, perform iterative distance scanning within a distance scale range increasing from 0 to the dynamic spatial filtering threshold parameter, determine whether a connected edge is established between any two discrete data points in the multidimensional spatial coordinate point cloud, generate a simple complex structure in the multidimensional space, and apply a boundary matrix elimination algorithm to track the connectivity and destruction states of the topology generators of the simple complex structure during the increasing scanning distance scale, and construct the persistent graph data that records the lifecycle of micro-organization topological invariants.

[0018] By adopting the above technical solution, the generation and destruction scales of zero-dimensional to multi-dimensional surface topological generators in multi-dimensional surface structures are extracted, thereby realizing the high-dimensional algebraic topological feature recording of the underlying microstructure morphology.

[0019] Preferably, the topology constraint registration module includes a spatial metric calculator, used for: calculating the Wasserstein distance between two persistent graph data for any two high-strength low-alloy steel samples; performing a nonlinear mapping operation on the calculated Wasserstein distance using a Gaussian kernel function to generate the topology affinity matrix elements describing the degree of similarity in the microstructure spatial topology between any two high-strength low-alloy steel samples; and providing the topology affinity matrix elements as the penalty constraint conditions required for subsequent dimensionality reduction operations.

[0020] By adopting the above technical solution, the differences in the distribution morphology of topological generators in the evolution of microstructure morphology among different samples are transformed into a basic metric for manifold mapping alignment.

[0021] Preferably, the topology constraint registration module includes a principal component dimensionality reduction unit, used for: performing functional principal component analysis with a topology penalty term; integrating the penalty constraint conditions into the variance maximization objective function for discretization and solution of the integral equation; extracting the optimal set of orthogonal weight functions; and performing inner product integration on the continuous temperature process functional objects of each high-strength low-alloy steel sample with the extracted set of orthogonal weight functions to generate the functional principal component score vector with microstructure topology spatial penalty attribute. The objective of the functional principal component analysis with the topology penalty term is to maximize the difference between the sum of squares of the functional principal component scores of all samples and the spatial penalty term; this spatial penalty term is equal to the sum of the products of the set topology penalty coefficient and the elements of the topology affinity matrix between any two samples and the squares of the differences in their functional principal component scores.

[0022] By adopting the above technical solution, macroscopic temporal features with different underlying microscopic evolution paths are forced to maintain physical isolation margin in the low-dimensional principal component space, thus solving the inverse ill-conditioned mapping defect of conventional temporal functional dimensionality reduction.

[0023] Preferably, the ship and marine performance prediction module includes a first-stage regression calculation network, used for: performing feature flattening operations on the two-dimensional planar distributed persistent graph data to transform the persistent graph data into persistent landscape vectors that meet the dimensional input requirements and have a fixed length; calling a multi-output support vector regression algorithm, using the functional principal component score vector with micro-organization topological space penalty attribute as the input independent variable, using the corresponding persistent landscape vector as the target output variable to train the parameters of the network multivariate mapping operator, and performing multivariate mapping calculations on the newly input functional principal component score vector to generate the predicted output persistent landscape vector.

[0024] By adopting the above technical solution, the scattered persistent graph distribution is aligned and transformed, and a high-precision intermediate correlation operator is constructed to push forward the process dimension to the organizational evolution topological features.

[0025] Preferably, the ship and marine performance prediction module includes a second-stage regression calculation network for receiving the persistent landscape vector; invoking a Gaussian process regression algorithm using a squared exponential covariance kernel function, taking the persistent landscape vector as the input independent variable to perform a probabilistic regression deduction process, mapping the persistent landscape vector to the predicted macroscopic mechanical performance value, and outputting it. Specifically, the new predicted macroscopic mechanical performance value is equal to the transpose of the covariance column vector between the newly input persistent landscape vector and the persistent landscape vectors of the known sample group, multiplied by the inverse matrix of the sum of the covariance matrices of the known sample group containing the observation noise variance term, and then multiplied by the column vector composed of the macroscopic mechanical performance values ​​of the known sample group.

[0026] Preferably, the high-strength low-alloy steel is a medium-thick plate with a room temperature yield strength in the range of 355 MPa to 690 MPa; the chemical composition of the high-strength low-alloy steel, by mass percentage, includes: carbon content of 0.02% to 0.15%, manganese content of 0.8% to 2.0%, and contains at least one microalloying element selected from niobium, vanadium, and titanium, with a total mass percentage of 0.01% to 0.25%, and the balance being iron and unavoidable impurities; after being treated by the thermomechanical control process, its microstructure includes at least one of acicular ferrite, granular bainite, or lower bainite.

[0027] By employing the aforementioned technical solution, a probabilistic correlation model between microscopic dimensionality reduction features and mechanical indicators is established, completing a closed-loop calculation of the entire physical chain from material process design to macroscopic mechanical verification. This clarifies the underlying carbon equivalent control and microalloying physical basis of the material targeted by the prediction system, as well as the typical microstructure of the thermomechanical control process products. This not only defines the efficient range of performance prediction, ensuring the model has extremely high robustness within a specific high-strength steel domain, but also provides the algorithm for extracting grain boundaries and topological features with a definite metallurgical physical basis, highlighting the system's practical engineering value in predicting the performance of thick cross-section steel plates in marine engineering.

[0028] Furthermore, the discrete time-series sensor data includes multiple time-temperature data pairs and multiple time-rolling force data pairs; the macroscopic mechanical property prediction values ​​include the predicted low-temperature impact energy at -40 degrees Celsius and the predicted room-temperature yield strength.

[0029] By adopting the above technical solution, the specific physical quantity data types of the prediction system's underlying input and final output are clarified, ensuring that the collected multi-source thermomechanical control process characteristics can establish an accurate calculation and docking relationship with the key mechanical assessment indicators under extreme service environments of ships and marine engineering.

[0030] Furthermore, when the first-stage regression calculation network calls the multi-output support vector regression algorithm to train the parameters of the network's multivariate mapping operator, the penalty parameter and kernel function width parameter of the multi-output support vector regression algorithm are both iteratively selected through a set multi-fold cross-validation procedure.

[0031] By adopting the above technical solution, the structural risk error of the first-stage regression calculation network on the model training set is minimized, effectively improving the generalization inference accuracy and stability of the network's multivariate mapping operator when facing new input data that has not been experimentally tested.

[0032] The present invention provides a performance prediction system for high-strength low-alloy steel used in marine engineering, which has the following advantages: This invention uses a marine process functional module to fit and reconstruct discrete thermomechanical control process timing sensor data into a continuous temperature process functional object, and then generates a process derivative functional characterizing the instantaneous cooling rate change through continuous differentiation operations. This overcomes the shortcomings of conventional data feature extraction methods that rely only on discrete endpoints or means, thus losing continuous dynamic information. It can accurately extract and quantify the transient thermodynamic evolution law caused by the release of latent heat of material phase transition, providing a foundation of input data with a complete physical evolution trajectory for subsequent performance extrapolation.

[0033] This invention extracts the instantaneous cooling acceleration value of the process derivative functional in an alloy dynamic topology module, adaptively calculates the dynamic spatial filtering threshold parameter, and uses this to set the distance connectivity tolerance when constructing a simple complex structure. This calculation mechanism directly establishes a mathematical correlation between the macroscopic cooling rate and the geometric connectivity conditions of the underlying microstructure, objectively reflecting the causal law of cooling conditions affecting local nucleus density in physical metallurgy, thereby accurately extracting the microscopic local heterogeneous topological invariant characteristics affected by complex thermomechanical control processes.

[0034] This invention incorporates elements of the topological affinity matrix, which characterizes the similarity of microstructure topology, as penalty constraints into the principal component analysis solution function during the dimensionality reduction and prediction stages. It also employs a concatenated network of first-stage multivariate mapping and second-stage Gaussian process regression to output the prediction results. This computational architecture forces the infinite-dimensional process functional to align with the underlying microscopic topology metric during dimensionality reduction, eliminating the ill-conditioned inverse problem of different microstructures being incorrectly mapped to similar dimensionality reduction spaces. This ensures that the mathematical derivation process strictly follows the physical evolution chain where process determines structure and then performance, effectively suppressing prediction bias in small sample datasets. Attached Figure Description

[0035] Figure 1 This is a system framework diagram of the performance prediction system for high-strength low-alloy steel for marine engineering according to the present invention. Figure 2 This is a schematic diagram of the analytical principle of the shipbuilding and marine engineering functional of the present invention; Figure 3 This is a schematic diagram of the simple complex growth of microstructure dynamic filtering in this invention; Figure 4 This is a flowchart of the bidirectional joint manifold embedding of functional principal component analysis with topological penalty terms according to the present invention. Figure 5 This is a network topology diagram of the two-order regression mapping engine of the present invention; Figure 6 This is a flowchart of the performance prediction method of the present invention; Figure 7 This is a graph showing the change of the dynamic spatial filtering threshold parameter of the present invention with instantaneous cooling acceleration; Figure 8 This is a comparison chart showing the actual test value and the predicted value of the room temperature yield strength of the present invention. Detailed Implementation

[0036] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] See Figure 1 The present invention provides a performance prediction system for high-strength low-alloy steel for marine engineering, which may include: a marine engineering functional module, an alloy dynamic topology module, a topology constraint registration module, and a marine performance prediction module.

[0038] Among them, thermomechanical control process refers to the process of strictly controlling temperature, reduction and cooling rate during rolling to refine grains and control phase transformation; high-strength low-alloy steel refers to steel for shipbuilding and marine engineering with high strength and good toughness by adding a small amount of alloying elements to carbon steel.

[0039] As a preferred embodiment of the present invention, in order to obtain accurate physical and metallurgical support for the probability mapping of macroscopic mechanical properties and the extraction of microstructure features (especially for the identification of large-angle grain boundaries with crystallographic orientation differences) in the system, the high-strength low-alloy steel is specifically defined as marine engineering structural steel with a room temperature yield strength in the range of 355MPa to 690MPa, and preferably meets the extreme service requirement of Charpy V-notch impact energy or impact absorption energy of not less than 50J in an environment of -40℃.

[0040] Specifically, its chemical composition, by mass percentage, includes: 0.02%–0.15% carbon (C), 0.8%–2.0% manganese (Mn), and at least one microalloying element selected from niobium (Nb), vanadium (V), and titanium (Ti). The total mass percentage of the microalloying elements is controlled between 0.01% and 0.25%, with the balance being iron (Fe) and unavoidable impurities. In actual industrial production, this high-strength low-alloy steel is typically delivered as medium-thick plates with a thickness greater than 10 mm, used to manufacture the hulls of deep-water offshore platforms, polar vessels, or offshore wind power foundation steel pipe piles. Due to the drastic evolution of the internal carbon equivalent and microalloying elements under nonlinear cooling operations, after thermomechanically controlled process (TMCP) treatment, its microstructure mainly exhibits at least one of acicular ferrite, granular bainite, or lower bainite.

[0041] The performance prediction system for high-strength low-alloy steel used in marine engineering is deployed in a cluster of computing devices, including a central processing unit and a graphics processing unit, or a distributed cloud computing server cluster. The computing device cluster is equipped with a non-volatile memory array for long-term storage of time-series sensor data for thermomechanical control processes, electron backscatter diffraction image data, scanning electron microscope image data, and macroscopic mechanical property test results of known high-strength low-alloy steel samples. The system facilitates data interaction and command calls between the marine engineering functional module, alloy dynamic topology module, topology constraint registration module, and marine performance prediction module through a high-speed data bus and memory sharing mechanism within the computing devices.

[0042] The marine engineering process functional module receives discrete time-series sensor data from thermomechanical control processes via a data input interface. This data specifically includes multiple time-temperature data pairs and multiple time-rolling force data pairs. The module is equipped with a basis function projection calculation unit, which uses multi-order spline basis functions to smoothly fit the discrete time-series sensor data, mapping it to a continuous process functional object. The module also includes a differential calculator, which performs derivative operations on the continuous process functional object, outputting a process derivative functional that characterizes the latent heat release and instantaneous cooling rate of the material's phase transition. After completing the calculation, the module caches the continuous process functional object and the process derivative functional, and then transmits the process derivative functional to the alloy dynamic topology module.

[0043] The alloy dynamic topology module is connected to the marine engineering functional module and the non-volatile memory array via a data bus. The module reads microstructure image data of high-strength low-alloy steel samples from the non-volatile memory array. The microstructure image data is either electron backscatter diffraction (ESD) image data or scanning electron microscope (SEM) image data. The module is equipped with an image point cloud processing unit and a simple complex generation unit. The image point cloud processing unit converts the two-dimensional microstructure image data into a multi-dimensional spatial coordinate point cloud. The simple complex generation unit receives the process derivative functional output from the marine engineering functional module and extracts the absolute value of the process derivative at a specific phase transition critical time node. Based on the absolute value of the process derivative, the simple complex generation unit adaptively calculates the spatial filtering threshold parameter. This spatial filtering threshold parameter directly sets the Euclidean distance connectivity tolerance when constructing a simple complex from the multi-dimensional spatial coordinate point cloud. The alloy dynamic topology module extracts topological invariant features from the simple complexes constructed at different scales, generates persistent graph data recording the creation and destruction coordinates of topology generators in different dimensions, and transmits this persistent graph data to the topology constraint registration module.

[0044] The topology constraint registration module establishes bidirectional data communication links with both the marine engineering functional module and the alloy dynamic topology module. The topology constraint registration module is equipped with a spatial metric calculator and a principal component reduction unit. The spatial metric calculator receives persistent graph data of multiple high-strength low-alloy steel samples output by the alloy dynamic topology module and calculates the spatial distance metric matrix between these persistent graph data. The principal component reduction unit receives continuous process functional objects output by the marine engineering functional module. When extracting the principal component basis of the continuous process functional objects, the principal component reduction unit uses the spatial distance metric matrix output by the spatial metric calculator as a penalty constraint condition in the objective solution function of the principal component analysis. After iterative orthogonal solution operations, the topology constraint registration module outputs a functional principal component score vector with microstructure topological spatial penalty attributes, incorporating topological constraint feature information. This functional principal component score vector with microstructure topological spatial penalty attributes, along with the persistent graph data, is synchronously transmitted to the marine performance prediction module.

[0045] The marine performance prediction module receives functional principal component score vectors with microstructure topological space penalty attributes and persistent graph data from the topology constraint registration module. The module is configured with a regression mapping network containing two-stage cascaded computation logic. The first-stage regression computation network establishes a multivariate mapping operator between the functional principal component score vectors with microstructure topological space penalty attributes and the persistent graph dimensionality-reduced feature vectors, outputting the predicted persistent graph dimensionality-reduced feature vectors. The second-stage regression computation network receives the persistent graph dimensionality-reduced feature vectors output by the first-stage network. This second-stage network executes a probabilistic regression process with a preset covariance kernel function, mapping the persistent graph dimensionality-reduced feature vectors to predicted macroscopic mechanical properties of high-strength low-alloy steel. These predicted macroscopic mechanical properties specifically include low-temperature impact energy test values ​​at -40 degrees Celsius and room-temperature yield strength test values. The module outputs the final calculated macroscopic mechanical property prediction values ​​to a terminal device via a display interface or overwrites them to a non-volatile memory array for later retrieval.

[0046] See Figure 2The marine engineering functional module of this invention addresses the technical problem that existing discrete sampling features cannot capture the nonlinear thermodynamic evolution process in thermomechanical control processes. The module receives discrete time-series sensor data from a thermomechanical control process, containing multiple pairs of time and temperature measurements. This discrete time-series sensor data is acquired by temperature and pressure sensors located on the rolling mill and laminar cooling equipment at a preset fixed frequency. Conventional feature extraction methods directly extract the final rolling temperature or average cooling rate values; such dimensionality reduction operations lose the dynamic characteristics of temperature changing continuously over time. The marine engineering functional module uses functional data analysis methods to reconstruct the discrete time-series sensor data into continuous function objects.

[0047] The marine engineering functional module is equipped with a basis function projection calculation unit. This unit acquires discrete time-series sensor data and uses multi-order spline basis functions to smoothly fit the discrete data points. Because the temperature decrease curve in thermomechanical control processes exhibits a non-linear, monotonically decreasing trend accompanied by local phase transition and temperature recovery phenomena, the basis function projection calculation unit specifically selects spline basis functions that can provide local support and possess high-order continuous differentiable properties. During the fitting operation, the basis function projection calculation unit introduces a least-squares method with a roughness penalty term to estimate the coefficient values ​​corresponding to each spline basis function. The roughness penalty term quantifies the degree of curve fluctuation by calculating the square integral of the second derivative of the fitted function, thereby suppressing overfitting caused by sensor measurement noise and ensuring that the generated continuous process functional object conforms to the true thermodynamic laws of materials physics.

[0048] The basis function projection calculation unit calculates and outputs a continuous process functional object through linear combination, and its formula is as follows: ; In the formula, Indicates the first The continuous temperature process functional object of a high-strength low-alloy steel sample (i.e., a smooth curve function equation characterizing the continuous temperature change of the sample over time in a thermomechanical controlled process. For example, discrete time and temperature scatter data collected once per second can be transformed into a continuously differentiable temperature function expression with time as the independent variable through fitting). A positive integer representing the sample number; Represents a time variable; This indicates the total number of spline basis functions used, in uppercase. It is a positive integer; Indicates the first The first high-strength low-alloy steel sample corresponding to the first The coefficients of the spline basis functions are shown in lowercase. The positive integer representing the index of the basis function; Indicates the first Spline basis functions.

[0049] The marine engineering functional module is also equipped with a differential calculator. In thermomechanical control processes, the phase transformation within high-strength low-alloy steel releases latent heat, causing abrupt changes in the instantaneous cooling rate on the macroscopic cooling curve. The differential calculator receives a continuous temperature process functional object output by the basis function projection calculation unit and performs first-order derivative operations along the time variable dimension to obtain characteristic variables characterizing the change in instantaneous cooling rate. Through continuous differentiation operations, the differential calculator transforms the static temperature curve into a process derivative functional describing transient dynamic characteristics.

[0050] The formula for calculating the derivative functional of a process using a differential calculator is as follows: ; In the formula, Indicates the first Process derivative functional of a high-strength low-alloy steel sample; Indicates the first A continuous temperature process functional object for a high-strength low-alloy steel sample. Represents a time variable.

[0051] After completing the aforementioned mathematical analysis steps, the marine engineering process functional module generates a continuous temperature process functional object and a process derivative functional containing complete physical evolution information. The module then transmits and stores this continuous temperature process functional object in a non-volatile memory array via the system's high-speed data bus for subsequent dimensionality reduction calls. Simultaneously, it transmits the process derivative functional to the alloy dynamic topology module. This processing flow achieves a continuous representation of the process trajectory through functionalization techniques, providing a precise time-series continuous analytical data foundation for the subsequent calculation of dynamic filtering parameters for process derivative intervention in microstructure images.

[0052] See Figure 3 The alloy dynamic topology module of this invention receives data output from the marine engineering functional module and extracts underlying spatial features by combining it with microstructure images. Existing topology data analysis methods generally use a globally uniform static filtering distance threshold when constructing simple complexes. This conventional processing method severs the physical and thermodynamic causal relationship between the temporal process state and the microstructure morphology. The alloy dynamic topology module establishes a dynamic mapping mechanism between the process derivative value and the growth distance condition of the spatial topological complex, thereby enabling mathematical representation intervention of the high-dimensional spatial geometric distribution and accurately extracting the local heterogeneous structural features affected by the cooling rate.

[0053] The alloy dynamic topology module is internally equipped with an image point cloudification processing unit. The module reads the microstructure image data of the corresponding high-strength low-alloy steel sample from a non-volatile memory array via the computing device's data bus. The microstructure image data includes electron backscatter diffraction image data or scanning electron microscope image data. The image point cloudification processing unit sequentially performs grayscale thresholding and edge contour detection operations on the input microstructure image data, identifying the distribution boundaries of large-angle grain boundaries (i.e., grain boundaries with a crystallographic orientation difference greater than 15° between adjacent grains) with specific crystallographic orientation differences, as well as isolated precipitate boundary feature points.

[0054] Specifically, regarding the extraction of isolated boundary feature points of precipitates, since niobium, vanadium, and titanium microalloyed carbonitride precipitates in high-strength low-alloy steel typically appear as bright, isolated microparticles with significant differences in atomic number contrast in scanning electron microscopy or backscattered electron images, the image point cloudification processing unit first employs morphological filtering algorithms such as Top-Hat Transform to enhance local bright spots in the microstructure image data and suppress uneven background grain illumination interference; then, it uses a local adaptive threshold segmentation algorithm to binarize the bright spots into independent connected regions; next, it calls a connected region analysis algorithm to calculate the equivalent diameter and roundness geometric features of each independent connected region, and filters out artifact connected regions that do not conform to the true precipitate size (e.g., filtering out coarse impurities or fine electron microscopy scanning noise that are not at the nanometer to submicrometer level) based on preset physical size constraints; finally, it uses edge tracking operators (e.g., Canny edge detector or Moore neighbor tracking algorithm) to extract the outer closed contour pixels of the filtered independent connected regions, and defines these pixels as isolated boundary feature points of precipitates.

[0055] The image point cloud processing unit then maps all extracted boundary feature points to a preset multidimensional Euclidean coordinate system, generating a multidimensional spatial coordinate point cloud. Each discrete data point in the multidimensional spatial coordinate point cloud corresponds to a physical grain boundary or phase boundary pixel in the microscopic tissue image data, thereby transforming the image raster data into a set of basic geometric vertices that meet the requirements of topological operations.

[0056] The alloy dynamic topology module is also equipped with a simple complex generation unit. This unit receives the process derivative functional from the marine process functional module and extracts the corresponding critical phase transformation time node by combining it with the chemical composition properties of the high-strength low-alloy steel sample. The critical phase transformation time node refers to the starting moment when austenite begins to undergo phase transformation during the cooling process. Specifically, the simple complex generation unit calculates the phase transformation start temperature by calling a preset continuous cooling transformation empirical formula and inputting the chemical composition properties. It then locates the time corresponding to this phase transformation start temperature by referring to the continuous temperature process functional object, thus extracting the critical phase transformation time node. The unit extracts a specific value of the process derivative functional at the critical phase transformation time node and performs an absolute value operation on this value to obtain the instantaneous cooling acceleration value. Based on the physical metallurgical mechanism, an increase in the instantaneous cooling acceleration value in the thermomechanical control process leads to the formation of locally high-density crystal nuclei in the high-strength low-alloy steel during the phase transformation process, thereby generating a microscopic grain boundary network with a tighter geometric spacing. To accurately fit this physical evolution law at the computational mechanism level, the simple complex generation unit inputs the instantaneous cooling acceleration value as an independent variable into a set negative exponential mapping function to calculate the dynamic spatial filtering threshold parameter. This dynamic spatial filtering threshold parameter directly defines the maximum Euclidean distance tolerance upper limit for the simple complex generation unit when subsequently processing multidimensional spatial coordinate point clouds to determine whether a connected edge can be established between any two discrete data points. In other words, the simple complex generation unit directly sets this dynamic spatial filtering threshold parameter as the connectivity tolerance. When the distance between any two discrete data points in the multidimensional spatial coordinate point cloud is less than or equal to this connectivity tolerance, a connected edge is determined and established between them.

[0057] The formula for calculating the dynamic spatial filtering threshold parameter of the simple complex generating unit is as follows: ; In the formula, Indicates the first The dynamic spatial filtering threshold parameters for a high-strength low-alloy steel sample, here A positive integer representing the sample number; This represents the pre-set basic filter threshold constant; The material phase transition sensitivity coefficient is used for new material samples where historical test data is lacking. The carbon equivalent (CE) is determined by a simple complex generation unit based on the chemical composition properties of the high-strength low-alloy steel sample. Specifically, the system extracts the mass percentages of carbon (C), manganese (Mn), and microalloying elements from the new material sample to calculate the CE. This CE is then substituted into a phase transformation sensitivity regression equation pre-established by the system based on a known sample database. The mathematical structure of the regression equation is as follows: ,in and These are empirical weighting coefficients; these coefficients are determined values ​​obtained by the system through linear fitting calibration using the least squares method, based on pre-extracted real cooling phase transformation data and carbon equivalent data of known high-strength low-alloy steel samples. Through the above equations, the system can adaptively calculate and obtain the material phase transformation sensitivity coefficient for new samples. ; Indicates the first The value of the process derivative functional of a high-strength low-alloy steel sample at the critical time node of phase transformation. Indicates the critical time point of phase transition; Represents an exponential function with the natural constant as its base; This represents the absolute value operator.

[0058] After obtaining the specific dynamic spatial filtering threshold parameters for each high-strength low-alloy steel sample, the simplex generation unit uses the multidimensional spatial coordinate point cloud as the vertex set and performs iterative distance scanning operations within a distance scale that gradually increases from zero to the dynamic spatial filtering threshold parameters. Following the mathematical rules for constructing the Vitoris-Lipps complex, the simplex generation unit connects all vertices whose Euclidean distance is less than the current scanning distance scale, generating a simplex structure with zero-dimensional to multidimensional surface structures in multidimensional space. The simplex generation unit further applies a boundary matrix elimination algorithm (this algorithm performs standard column elimination on the matrix recording the boundary inclusion relationships between simplexes to identify the generation and closure of topological generators at different distance scales) to track and calculate the changes in the number of zero-dimensional connected components and the number of one-dimensional topological loops in the generated simplex structure as the scanning distance scale increases.

[0059] The simple complex generation unit extracts the generation distance scale corresponding to the first connection of each topological generator, and the disappearance distance scale corresponding to the closure or merging of the topological generator. The simple complex generation unit combines each acquired generation and disappearance distance scale into horizontal and vertical coordinate points on a two-dimensional plane, constructing persistent graph data that records the complete evolutionary lifecycle of the microstructure's topological invariants. Finally, the alloy dynamic topology module transmits this persistent graph data to the topology constraint registration module via a high-speed data communication link within the system, providing a spatial structure metric benchmark condition, including the state of process history intervention, for subsequent joint dimensionality reduction and alignment operations in the manifold space.

[0060] See Figure 4The topology constraint registration module of this invention addresses the ill-conditioned inverse problem in feature dimension compression operations on infinite-dimensional time-series process functional data. Existing functional data dimensionality reduction methods rely solely on the variance distribution of the process time-series features to extract the dimensionality reduction basis. This leads to time-series process input features with different microstructure evolution results being incorrectly mapped to similar low-dimensional spatial coordinates, failing to align the physical trajectory of the time-series process with the underlying data of multi-dimensional spatial microstructure topology features. The topology constraint registration module introduces spatial topology metric constraint penalties into the objective function of the dimensionality reduction operation, establishing a bidirectional joint manifold embedding and cross-feature intervention mechanism for spatiotemporal multimodal data.

[0061] The topology constraint registration module is equipped with a spatial metric calculator. This calculator receives persistent map data corresponding to multiple high-strength low-alloy steel samples output by the alloy dynamic topology module via the system data communication bus. The persistent map data records the lifetime change coordinate set of microstructure topology generators for each high-strength low-alloy steel sample at different scanning distance scales. For any two high-strength low-alloy steel samples, the spatial metric calculator calculates the Wasserstein distance between their persistent map data. The Wasserstein distance quantifies the minimum Euclidean spatial migration cost required to transform the geometric distribution of topology generators in one persistent map data to that of another. The spatial metric calculator further performs a nonlinear mapping operation on the calculated Wasserstein distance using a Gaussian kernel function, generating topological affinity matrix elements that describe the similarity of the microstructure spatial topology between any two high-strength low-alloy steel samples. The larger the value of this topological affinity matrix element, the closer the microstructure spatial topology morphology of the two high-strength low-alloy steel samples.

[0062] The topology constraint registration module is also equipped with a principal component dimensionality reduction unit. This unit receives continuous temperature-process functional objects corresponding to multiple high-strength low-alloy steel samples from the marine engineering functional module. The unit performs functional principal component analysis with a topology penalty term. The goal of this calculation is to find an optimal set of continuous orthogonal weight functions that maximizes the retention of the characteristic variance distribution of the original process data fluctuations after the continuous temperature-process functional objects are projected onto this set. The unit combines the elements of the topological affinity matrix output from the spatial metric calculator as penalty conditions into the objective function for maximizing variance. This algorithm design mandates that for two high-strength low-alloy steel samples with significant morphological distribution differences in the persistent microstructure graph space, their corresponding functional principal component scores must maintain a feature isolation margin of at least a set distance in the reduced low-dimensional feature space.

[0063] The formula for performing functional principal component analysis with topological penalty terms by the principal component dimensionality reduction unit is as follows: ; In the formula, Characterized by weight function To optimize the variables, we need to find the maximum value of the objective function. This indicates the total number of high-strength low-alloy steel samples input into the system. It is a positive integer; This indicates the sample number of the first high-strength low-alloy steel sample. Less than or equal to Positive integers; This indicates the sample number of the second high-strength low-alloy steel. Less than or equal to Positive integers; Indicates the first A continuous temperature process functional object of a high-strength low-alloy steel sample is projected onto a weighting function. The functional principal component scores obtained afterward; Indicates the first A continuous temperature process functional object of a high-strength low-alloy steel sample is projected onto a weighting function. The functional principal component scores obtained afterward; This represents the set topology penalty coefficient; Indicates the first The high-strength low-alloy steel sample and the first The topological affinity matrix elements between high-strength low-alloy steel samples.

[0064] The principal component dimensionality reduction unit iteratively calculates the objective function of the above formula using an integral equation discretization decomposition algorithm to obtain the orthogonal weight function set that maximizes the objective function value. The unit then performs inner product integration operations on the continuous temperature process functional objects of each high-strength low-alloy steel sample with the extracted orthogonal weight function set, generating a dimensionality-reduced functional principal component score vector. After completing the computation, the topology constraint registration module synchronously transmits the functional principal component score vector with microstructure topology space penalty attributes, along with the persistent graph data from the alloy dynamic topology module, to the ship and marine performance prediction module. This dimensionality reduction registration process ensures that the distribution coordinates of the low-dimensional manifold space of the input continuous thermomechanical control process features are strictly controlled by the metric constraints of the internal physical microstructure topology of the high-strength low-alloy steel material during feature dimension compression.

[0065] See Figure 5This invention provides a ship and marine performance prediction module that may include a first-stage regression calculation network and a second-stage regression calculation network. The ship and marine performance prediction module receives data output from a topology constraint registration module and performs causal mapping inference calculations from underlying microscopic features to macroscopic mechanical properties. Existing machine learning prediction models directly map scalar process parameters to performance indicators. This calculation process traverses the physical causal chain of microstructure evolution, resulting in a generalization prediction error greater than 20% on new data samples. The ship and marine performance prediction module constructs a regression mapping network containing two-stage serial calculation logic, strictly following the sequential causal evolution law from time-series process data to microstructure and then to macroscopic mechanical properties in physical metallurgy, thereby achieving numerical output of the mechanical properties of high-strength low-alloy steel.

[0066] The ship and marine performance prediction module is internally configured with a first-stage regression computation network. This network establishes a multivariate mapping operator to transform process features into micro-topological spatial features. It receives functional principal component score vectors with micro-organizational topological spatial penalty attributes and persistent graph data transmitted from the topology constraint registration module. Since the persistent graph data consists of a discrete set of two-dimensional spatial coordinate points, the first-stage regression computation network first performs a feature flattening operation, transforming the persistent graph data into a fixed-length persistent landscape vector (a persistent landscape vector is a one-dimensional continuous piecewise linear function vector with a fixed dimension length, obtained by projecting the discrete two-dimensional generation and destruction coordinate points in the persistent graph data onto the envelope of a hierarchical tent function), to meet the input dimension alignment requirements of machine learning algorithms. The first-stage regression computation network employs a multi-output support vector regression algorithm, using the functional principal component score vectors with micro-organizational topological spatial penalty attributes as independent variables and the corresponding persistent landscape vectors as target variables for network parameter training. The penalty parameters and kernel width parameters of the multi-output support vector regression algorithm are selected iteratively through a set multi-fold cross-validation procedure, so that the structural risk error of the first-stage regression calculation network on the training set is minimized.

[0067] The marine performance prediction module is also equipped with a second-stage regression calculation network. This second-stage network is connected in series with the output of the first-stage network, receiving the persistent landscape vector from the first-stage network and using it as the input of independent variables. The second-stage network employs a Gaussian process regression algorithm, using the persistent landscape vector as the input to perform a probabilistic regression deduction process, constructing a mapping operator from microstructural characteristics to macroscopic mechanical properties. The Gaussian process regression algorithm uses a squared exponential covariance kernel function to quantify the degree of data correlation between any two persistent landscape vectors. The target output variable of the second-stage network is the predicted value of the macroscopic mechanical properties of the high-strength low-alloy steel sample, specifically including the low-temperature impact energy test value at -40℃ and the room temperature yield strength test value.

[0068] For the new high-strength low-alloy steel process design parameters that have not been experimentally tested, the marine performance prediction module performs forward inference calculations sequentially through a first-stage regression calculation network and a second-stage regression calculation network. The second-stage regression calculation network ultimately outputs the posterior mean predicted values ​​of the macroscopic mechanical properties of the new high-strength low-alloy steel sample, and its calculation formula is as follows: ; In the formula, The posterior mean predicted value represents the macroscopic mechanical properties of the new high-strength low-alloy steel sample. This represents the covariance column vector between the persistent landscape vector of the new high-strength low-alloy steel sample and the persistent landscape vector of the known high-strength low-alloy steel samples. This represents the matrix transpose operator. This represents the covariance matrix calculated among the persistent landscape vectors of all known high-strength low-alloy steel samples. This represents the variance of observation noise that follows a normal distribution in Gaussian process regression calculations; Represents the identity matrix; This represents a column vector consisting of the macroscopic mechanical properties of all known high-strength low-alloy steel samples. This represents the matrix inversion operator.

[0069] After completing the above extrapolation calculations, the marine performance prediction module extracts the posterior mean prediction value as the final mechanical performance evaluation index output by the system. The module then sends the posterior mean prediction value to the terminal screen via the display interface of the computing device, or overwrites it into the non-volatile memory array at the system's underlying layer for use by other industrial control software. This two-stage regression calculation architecture transforms the complex materials research and development trial-and-error process into an executable mathematical calculation flow, solving the technical shortcomings of conventional algorithms that suffer from excessive extrapolation prediction bias and violate the thermodynamic principles of materials under extremely small sample datasets.

[0070] See Figure 6 The present invention provides a method for predicting the performance of high-strength low-alloy steel for marine engineering, which may include the following detailed steps.

[0071] In step S1, the marine engineering functional module receives discrete time-series sensor data of the thermomechanical control process through the data input interface. This discrete time-series sensor data contains multiple pairs of time and temperature measurements. The basis function projection calculation unit within the module uses multi-order spline basis functions to smoothly fit the discrete time-series sensor data, generating a continuous temperature process functional object. During this fitting process, the basis function projection calculation unit introduces a least-squares method with a roughness penalty term for parameter estimation. The differential calculator within the module performs first-order time-dimension derivative operations on the continuous temperature process functional object, generating a process derivative functional containing transient dynamic characteristics. The module stores the continuous temperature process functional object in a non-volatile memory array and sends the process derivative functional to the alloy dynamic topology module.

[0072] In step S2, the alloy dynamic topology module reads the microstructure image data of the corresponding high-strength low-alloy steel sample from the non-volatile memory array via the data bus of the computing device. The microstructure image data is either electron backscatter diffraction image data or scanning electron microscope image data. The image point cloud processing unit inside the alloy dynamic topology module performs grayscale thresholding and edge contour detection operations on the microstructure image data to extract feature points of large-angle grain boundary distribution with crystallographic orientation differences and isolated precipitate boundaries. Specifically, for the extraction of isolated precipitate boundary feature points, the image point cloud processing unit sequentially enhances the image contrast of locally bright precipitate particles through morphological filtering, performs adaptive binarization and connected component feature analysis to filter out noise points with inconsistent geometric dimensions, and applies an edge tracking operator to obtain the closed outer contour pixels of the remaining real microalloyed carbonitride particles as isolated precipitate boundary feature points. The image point cloud processing unit maps each extracted boundary feature point to a preset multidimensional Euclidean coordinate system to generate a multidimensional spatial coordinate point cloud.

[0073] In step S3, the simple complex generation unit within the alloy dynamic topology module receives the process derivative functional from the marine process functional module. The simple complex generation unit extracts the critical time node for phase transformation based on the chemical composition properties of the high-strength low-alloy steel sample. It extracts the specific value of the process derivative functional at this critical time node and performs absolute value calculations to obtain the instantaneous cooling acceleration value. The simple complex generation unit uses this instantaneous cooling acceleration value as an independent variable and inputs it into a negative exponential mapping function for mathematical derivation, calculating the dynamic spatial filtering threshold parameter. Using the multidimensional spatial coordinate point cloud as the vertex set, the simple complex generation unit performs iterative distance scanning within a distance scale increasing from 0 to the dynamic spatial filtering threshold parameter, determining whether a connected edge is established between any two discrete data points in the multidimensional spatial coordinate point cloud, thereby generating a simple complex structure with zero-dimensional to multidimensional surface structures in multidimensional space. The simple complex generating unit uses a boundary matrix elimination algorithm to track the connectivity and destruction states of topological generators in the simple complex structure as the scanning distance scale increases, constructing persistent graph data that records the lifecycle of microstructure topological invariants. The alloy dynamic topology module then transmits the persistent graph data to the topology constraint registration module.

[0074] In step S4, the topology constraint registration module receives persistent graph data of multiple known high-strength low-alloy steel samples and corresponding continuous temperature process functional objects from the alloy dynamic topology module and the non-volatile storage array, respectively. The spatial metric calculator within the topology constraint registration module calculates the Wasserstein distance for any two high-strength low-alloy steel samples based on their persistent graph data. The spatial metric calculator further applies a Gaussian kernel function to perform a nonlinear mapping on the calculated Wasserstein distance, generating topological affinity matrix elements that describe the degree of similarity in the microstructure spatial topology between different high-strength low-alloy steel samples. The principal component dimensionality reduction unit within the topology constraint registration module performs functional principal component analysis with a topological penalty term, using the topological affinity matrix elements as penalty constraints and integrating them into the variance maximization objective function for discretization and solution of the integral equation. The principal component dimensionality reduction unit extracts the optimal set of orthogonal weight functions and performs inner product integration with this set of orthogonal weight functions on the continuous temperature process functional objects, generating a functional principal component score vector with microstructure topological spatial penalty attributes. The topology constraint registration module synchronously transmits the functional principal component score vector and persistent graph data to the ship and marine performance prediction module via the data bus.

[0075] In step S5, the first-stage regression computation network configured within the ship and marine performance prediction module receives the functional principal component score vector with micro-organization topological space penalty attributes and persistent graph data transmitted from the topology constraint registration module. The first-stage regression computation network performs feature flattening on the two-dimensional planar distributed persistent graph data, transforming it into persistent landscape vectors of fixed length that meet the dimensional input requirements. The first-stage regression computation network then calls the multi-output support vector regression algorithm, using the functional principal component score vector with micro-organization topological space penalty attributes as the input independent variable and the corresponding persistent landscape vector as the target output variable to train the parameters of the network's multivariate mapping operator. After establishing the multivariate mapping operator, the first-stage regression computation network performs multivariate mapping calculations on the newly input functional principal component score vectors, passing the predicted persistent landscape vector to the second-stage regression computation network.

[0076] Step S6: The second-stage regression calculation network configured within the ship and marine performance prediction module receives the persistent landscape vector output by the first-stage regression calculation network. The second-stage regression calculation network invokes a Gaussian process regression algorithm using the squared exponential covariance kernel function, employing the persistent landscape vector as the input independent variable to perform a probabilistic regression deduction process. The second-stage regression calculation network maps the underlying microscopic characteristic indicators to macroscopic mechanical performance prediction values. These macroscopic mechanical performance prediction values ​​include low-temperature impact energy test values ​​at -40℃ and room temperature yield strength test values. The second-stage regression calculation network outputs the posterior mean prediction values ​​of the new high-strength low-alloy steel sample's macroscopic mechanical properties. The ship and marine performance prediction module extracts these posterior mean prediction values ​​as the final performance evaluation result and outputs them to the terminal screen through the display interface of the computing device, thus completing the complete closed-loop calculation operation from discrete process signal input to mechanical performance indicator output.

[0077] This invention also provides an electronic device that may include a processor, internal memory, a non-volatile memory array, and a network communication interface. The processor, internal memory, non-volatile memory array, and network communication interface are physically connected and transmit electrical signals to each other via a high-speed data communication bus within the device. The high-speed data communication bus adopts a peripheral component interconnection expansion architecture or an industry-standard architecture, providing a stable hardware signal interaction channel.

[0078] The processor employs a multi-core central processing unit (CPU) chip or a graphics processing unit (GPU) chip. The non-volatile memory array uses a solid-state drive (SSD) or hard disk drive (HDD) storage array to persistently store the operating system's underlying code and the computer program code implementing the aforementioned method for predicting the performance of high-strength low-alloy steel for marine engineering. The internal memory uses random access memory (RAM) to provide a temporary data cache environment with a specified operating space capacity for the processor to execute the computer program code. The network communication interface uses an Ethernet card chip with a transmission rate greater than or equal to 1000 Mbps or a fiber optic communication module to establish standard data transmission protocol links with external thermomechanical control process data acquisition systems and microscopic tissue image acquisition equipment, completing the reception and transmission of basic hardware signals.

[0079] When the electronic device is powered on, the processor loads the computer program code and related dependent runtime libraries stored in the non-volatile memory array into the internal memory via a high-speed data communication bus for compilation and execution. While executing the computer program code, the processor instantiates and generates the following modules in the electronic device's internal memory and hardware registers: a marine engineering functional module, an alloy dynamic topology module, a topology constraint registration module, and a marine performance prediction module. Under the processor's computing power scheduling, these four functional modules perform instruction-level data flow and computational resource allocation.

[0080] The processor receives discrete time-series sensor data for thermomechanical control processes and microstructure images of high-strength low-alloy steel samples via a network communication interface, and temporarily stores this data in internal memory. The processor drives the marine engineering functional module to extract the time-series sensor data from internal memory, perform functionalization processing, and calculate the process derivative functional. The processor then drives the alloy dynamic topology module to calculate the dynamic spatial filtering threshold parameters using this process derivative functional and constructs persistent graph data in a multidimensional Euclidean coordinate system to extract microstructure topological generator features. Next, the processor drives the topology constraint registration module to perform functional principal component dimensionality reduction operations with spatial distance constraints. Finally, the processor drives the marine performance prediction module to call the first-stage regression calculation network and the second-stage regression calculation network to perform forward inference, calculating the posterior mean predicted values ​​of the macroscopic mechanical properties of the high-strength low-alloy steel samples. The processor sends the calculated posterior mean predicted values ​​to an external display terminal via the network communication interface, or encapsulates and writes them into a non-volatile memory array for storage.

[0081] This invention also provides a computer-readable storage medium, which may include a non-volatile physical hardware medium capable of storing computer program code. Specifically, the computer-readable storage medium may be a read-only memory, magnetic tape, floppy disk, flash memory storage device, or optical disc data carrier. The physical storage space of the computer-readable storage medium records a sequence of computer program instructions. When this sequence of computer program instructions is loaded and executed by a processor of any hardware computing device capable of instruction reading and logical operation, the hardware computing device executes the entire data processing flow of the aforementioned performance prediction method for high-strength low-alloy steel for marine engineering.

[0082] The computer program instruction sequence recorded in the aforementioned computer-readable storage medium is divided into multiple independent code execution segments according to modular programming specifications and sequentially stored in designated physical sectors of the computer-readable storage medium. This physical structure design enables the hardware computing device to accurately reconstruct and execute the mathematical model calculations and control logic corresponding to the marine engineering functional module, alloy dynamic topology module, topology constraint registration module, and marine performance prediction module line by line after reading the data from the computer-readable storage medium. This allows for the objective physical quantification output of the mechanical properties of high-strength low-alloy steel materials without altering the physical hardware structure of the computing device.

[0083] See Figure 7 This invention provides a specific application embodiment of a performance prediction system for high-strength low-alloy steel used in marine engineering, which may include obtaining the numbered... The first high-strength low-alloy steel sample and its number The discrete thermomechanical control process timing sensor data of the second high-strength low-alloy steel sample is used. The marine engineering process functional module fits the discrete timing sensor data of the two samples into a continuous temperature process functional object. The marine engineering process functional module performs derivative operations on the continuous temperature process functional object to generate a process derivative functional. Based on the chemical composition properties of the material, the critical time node for phase transformation is set. The time interval is 120s. The absolute value of the process derivative of the first high-strength low-alloy steel sample at the critical time node of the phase transformation is set to 15℃ / s, and the absolute value of the process derivative of the second high-strength low-alloy steel sample at the critical time node of the phase transformation is set to 30℃ / s.

[0084] The alloy dynamic topology module receives the absolute value of the above process derivative and applies the dynamic spatial filtering threshold parameter. The calculation. The alloy dynamic topology module sets the basic filter threshold constant. The material phase transition sensitivity coefficient is set to 2.5 μm. It is 0.05. Combined with... Figure 7 The visual information shown Figure 7The horizontal axis represents the instantaneous cooling acceleration, in ℃ / s; the vertical axis represents the dynamic spatial filtering threshold parameter, in μm. Figure 7 The solid line parameter variation curves in the graph illustrate the nonlinear exponential decay of the dynamic spatial filtering threshold parameter with increasing instantaneous cooling acceleration, under a set baseline constant. The alloy dynamic topology module uses a negative exponential mapping operator to derive a dynamic spatial filtering threshold parameter of 1.18 μm for the first high-strength low-alloy steel sample and 0.56 μm for the second high-strength low-alloy steel sample. Using the calculated maximum connection distances of 1.18 μm and 0.56 μm as upper limits, the alloy dynamic topology module constructs simple complex structures in the multidimensional spatial coordinate point cloud and extracts the lifetime coordinates of the topology generators to generate two corresponding persistent graph data.

[0085] The topology constraint registration module receives persistent graph data output from the alloy dynamic topology module and continuous temperature process functional objects output from the marine engineering functional module. The spatial metric calculator within the topology constraint registration module calculates the Wasserstein distance between the persistent graph data of the first and second high-strength low-alloy steel samples, and generates topological affinity matrix elements after Gaussian kernel function mapping. The principal component dimensionality reduction unit will set the topological penalty coefficient. The value is assigned to 0.1. The principal component dimensionality reduction unit will use the topological penalty coefficient. and topological affinity matrix elements The input is fed into the functional principal component analysis objective function with a topological penalty term. The principal component dimensionality reduction unit iteratively extracts the orthogonal weight function set and calculates the functional principal component score for the first high-strength low-alloy steel sample. And the functional principal component scores of the second high-strength low-alloy steel sample. The aforementioned penalty calculation process, which introduces topological features, forcibly increases the distance coordinate isolation between two samples with significant differences in physical organization in low-dimensional space.

[0086] The marine performance prediction module calls the first-stage regression calculation network to establish a multivariate mapping operator that transforms process features into microscopic topological space features, and outputs the corresponding persistent landscape vector. The module then calls the second-stage regression calculation network to calculate the covariance column vector between the persistent landscape vector of the first high-strength low-alloy steel sample and the persistent landscape vectors of known high-strength low-alloy steel samples. The second-stage regression calculation network combines the covariance matrix formed by the persistent landscape vectors of all known samples. Observation noise variance follows a normal distribution identity matrix and a column vector consisting of the known macroscopic mechanical properties of the samples. The simulation process is executed. The second-stage regression calculation network ultimately calculates and outputs the posterior mean predicted value of the room temperature yield strength of the first high-strength low-alloy steel sample. It is 520 MPa.

[0087] See Figure 8 To verify the objective prediction effect of the technical solution of this invention, the system input 500 sets of high-strength low-alloy steel sample data with complete physical experimental test records. The system divided 400 sets of data into a model training dataset and the remaining 100 sets into an independent validation dataset. The system used both a conventional support vector regression algorithm system based solely on scalar parameter mapping and the performance prediction system for high-strength low-alloy steel for marine engineering provided by this invention to calculate the room temperature yield strength of the 100 sets of independent validation data. The system compared the predicted values ​​obtained by the two different methods with the actual room temperature yield strength values ​​obtained by a physical tensile testing machine and generated an error scatter distribution.

[0088] Combination Figure 8 The visual information shown Figure 8 The horizontal axis represents the actual room temperature yield strength in MPa; the vertical axis represents the predicted room temperature yield strength in MPa. Figure 8 The scatter distribution shows that the predicted data points output by the system of this invention, marked with solid squares, are more densely packed on both sides of the ideal corresponding line of the dashed line with a slope of 1, while the predicted data points output by the conventional comparison system, marked with hollow circles, show obvious deviation and divergence. Statistical error calculations show that the average absolute error of the conventional comparison system in predicting room temperature yield strength reaches 45 MPa. The system of this invention, by constructing a bidirectional joint manifold embedding and cross-feature intervention mechanism for spatiotemporal multimodal data, reduces the average absolute error of room temperature yield strength to 12 MPa, and its maximum single-point prediction deviation is strictly controlled within a fluctuation range of 5%. The above data demonstrates that the two-order regression architecture with topological penalty mechanism provided by this invention effectively eliminates the ill-conditioned inverse problem in the functional dimensionality reduction stage, improving the accuracy of the numerical mapping from underlying material characteristics to macroscopic mechanical properties.

[0089] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A performance prediction system for high-strength low-alloy steel used in marine engineering, characterized in that, include: The Marine Engineering Process Functional Module is used to generate continuous temperature process functional objects and process derivative functionals from time-series sensor data of discrete thermomechanical control processes. The alloy dynamic topology module is used to convert the microstructure image data of high-strength low-alloy steel into a multi-dimensional spatial coordinate point cloud, and combine the chemical composition properties of high-strength low-alloy steel with the process derivative functional to generate persistent graph data. The topology constraint registration module is used to generate topology affinity matrix elements as penalty constraints based on the persistent graph data, and to generate functional principal component score vectors with microstructure topology space penalty attributes in combination with the continuous temperature process functional object. The ship and marine performance prediction module is used to generate a persistent landscape vector by combining the persistent graph data with the functional principal component score vector with micro-organization topological space penalty attribute, and output the macro-mechanical performance prediction value.

2. The performance prediction system for high-strength low-alloy steel for marine engineering according to claim 1, characterized in that, The shipbuilding and marine engineering functional module includes a basis function projection calculation unit, used for: Multi-order spline basis functions are used to smoothly fit the time-series sensor data of discrete thermomechanical control processes; In the smooth fitting process, a least squares method with a roughness penalty term is introduced for parameter estimation to generate the continuous temperature process functional object.

3. The performance prediction system for high-strength low-alloy steel for marine engineering according to claim 1, characterized in that, The marine engineering functional module includes a differential calculator, used for: Perform the first-order derivative operation in the time dimension on the continuous temperature process functional object to obtain the feature variables characterizing the instantaneous cooling rate change; By performing continuous differentiation operations and combining the characteristic variables, the static temperature curve is transformed into the process derivative functional that includes transient dynamic characteristics.

4. The performance prediction system for high-strength low-alloy steel for marine engineering according to claim 1, characterized in that, The alloy dynamic topology module includes an image point cloudification processing unit, used for: Gray-scale thresholding and edge contour detection operations were performed on the microstructure image data of the high-strength low-alloy steel sample to extract feature points of large-angle grain boundary distribution with crystallographic orientation differences and isolated boundary of precipitated phase. The extracted boundary feature points are mapped to a preset multidimensional Euclidean coordinate system to generate the multidimensional coordinate point cloud.

5. The performance prediction system for high-strength low-alloy steel for marine engineering according to claim 1, characterized in that, The alloy dynamic topology module includes a simple complex generation unit, used for: The critical time node for phase transformation was extracted based on the chemical composition properties of the high-strength low-alloy steel sample. Extract the value of the process derivative functional at the critical time node of the phase transition and perform absolute value calculation to obtain the instantaneous cooling acceleration value; The instantaneous cooling acceleration value is used as an independent variable and input into the negative exponential mapping function for mathematical deduction to obtain the dynamic spatial filtering threshold parameter; Based on the dynamic spatial filtering threshold parameter, the connectivity tolerance is set, a simple complex structure is constructed in the multidimensional spatial coordinate point cloud, and topological invariant features are extracted to complete the generation of the persistent graph data.

6. The performance prediction system for high-strength low-alloy steel for marine engineering according to claim 5, characterized in that, The simple complex generation unit is further configured to use the multidimensional spatial coordinate point cloud as a vertex set, perform iterative distance scanning within a distance scale range increasing from 0 to the dynamic spatial filtering threshold parameter, determine whether a connected edge is established between any two discrete data points in the multidimensional spatial coordinate point cloud, and generate a simple complex structure in the multidimensional space; apply a boundary matrix elimination algorithm to track the connectivity and destruction states of the topological generators of the simple complex structure during the increasing scanning distance scale, and construct the persistent graph data that records the lifecycle of micro-organized topological invariants.

7. The performance prediction system for high-strength low-alloy steel for marine engineering according to claim 1, characterized in that, The topology constraint registration module includes a spatial metric calculator, used for: For any two high-strength low-alloy steel samples, calculate the Wasserstein distance between the two persistent graph data. A Gaussian kernel function is used to perform a nonlinear mapping operation on the calculated Wasserstein distance to generate the topological affinity matrix elements that describe the degree of similarity in the microstructure spatial topology between any two high-strength low-alloy steel samples. The topological affinity matrix elements are then used as the penalty constraints required for subsequent dimensionality reduction operations.

8. The performance prediction system for high-strength low-alloy steel for marine engineering according to claim 1, characterized in that, The topology constraint registration module includes a principal component dimensionality reduction unit, used for: Perform functional principal component analysis with topological penalty terms, integrate the penalty constraints into the objective function of maximizing variance, discretize the integral equation, and extract the optimal set of orthogonal weight functions. The continuous temperature process functional objects of each high-strength low-alloy steel sample are subjected to inner product integration with the extracted orthogonal weight function set to generate the functional principal component score vector with microstructure topological space penalty attribute.

9. The performance prediction system for high-strength low-alloy steel for marine engineering according to claim 1, characterized in that, The ship and marine performance prediction module includes a first-stage regression calculation network, used for: Perform a feature flattening operation on the persistent graph data with a two-dimensional planar distribution to transform the persistent graph data into a persistent landscape vector with a fixed length that meets the dimensional input requirements; The multi-output support vector regression algorithm is invoked, with the functional principal component score vector with micro-organization topology space penalty attribute as the input independent variable, and the corresponding persistent landscape vector as the target output variable for parameter training of the network multivariate mapping operator. Multivariate mapping calculation is then performed on the newly input functional principal component score vector to generate the predicted output persistent landscape vector.

10. The performance prediction system for high-strength low-alloy steel for marine engineering according to claim 1, characterized in that, The ship and marine performance prediction module includes a second-stage regression calculation network, which receives the persistent landscape vector; calls a Gaussian process regression algorithm with a squared exponential covariance kernel function, takes the persistent landscape vector as an input variable to perform a probabilistic regression deduction process, maps the persistent landscape vector to the predicted values ​​of macroscopic mechanical performance, and outputs them.

Images