Knowledge graph pushing system for process parameters of zirconium-titanium alloy fastener

Abstract

This system, which pushes knowledge graph-based parameters for zirconium-titanium alloy fastener manufacturing processes, involves knowledge graph and intelligent manufacturing technologies. It acquires the temperature and pressure sequences of the first process to generate temporal features, extracts the hidden layer state trend gradient, and constructs a temporal drift vector. The system obtains the second feature vector of the associated process from the knowledge graph, and obtains a dynamic mapping matrix through feature cross-interaction and mapping. This matrix is ​​then used to perform an affine transformation on the first feature vector to obtain a reconstructed feature vector, which is then aggregated and updated to generate the control parameters for the second process. Finally, the system optimizes the parameters based on feedback and updates the preset time window. This system can eliminate the error cascading effect caused by linear aggregation in heterogeneous spaces, effectively solving the industry problem of high yield per process but high scrap rate per assembly.

Description

Technical Field

[0001] This application relates to the fields of knowledge graph and intelligent manufacturing technology, specifically a knowledge graph push system for the preparation process parameters of zirconium-titanium alloy fasteners. Background Technology

[0002] The fabrication of zirconium-titanium alloy fasteners involves multiple consecutive processes such as forging and annealing, with complex physical coupling and transmission between these processes. Existing mainstream process parameter recommendation technologies use knowledge graphs as carriers and graph neural networks as engines, employing graph convolutional networks for node feature aggregation. The core assumption is that the feature vectors of adjacent process nodes lie in the same linear mapping space (feature space isomorphism), thus linear summation or mean pooling can be used for aggregation.

[0003] However, the dynamic accumulation of temperature variation rate and instantaneous strain rate in the upstream processes of zirconium-titanium alloy hot working generates strong nonlinear physical transmission to the optimal process window of the downstream processes. This causes nonlinear distortion of the feature space during cross-process data flow, with upstream and downstream nodes actually existing in heterogeneous feature spaces. Existing technologies attempt to mitigate this by increasing the number of GCN layers or introducing attention mechanisms, but the former is prone to oversmoothing, and the latter is still scalar scaling, neither of which changes the fundamental nature of linear combination of heterogeneous vectors and cannot eliminate mapping errors.

[0004] This defect causes process parameters that appear normal when output by the system, but lead to abnormal grain growth and reduced fatigue life when combined, resulting in an industry-wide problem of high yield for individual processes but high scrap rates for assemblies. Therefore, there is an urgent need for a process parameter push scheme that can sense and correct nonlinear distortions across the process feature space to achieve full-process physical self-consistency. Summary of the Invention

[0005] The purpose of this application is to provide a knowledge graph push system for the manufacturing process parameters of zirconium-titanium alloy fasteners, which can eliminate the error cascade effect caused by heterogeneous spatial linear polymerization and effectively solve the industry problem of high yield of single process and high scrap rate of assembly.

[0006] The objective of this application can be achieved through the following technical solution: Firstly, a knowledge graph push system for zirconium-titanium alloy fastener manufacturing process parameters includes the following modules: The data acquisition module is used to synchronously acquire the absolute temperature sequence and instantaneous pressure sequence of the first process within a preset execution cycle, and generate the time-series feature sequence of the first process within the preset execution cycle. The encoding module is used to extract the hidden layer state sequence of the temporal feature sequence based on a preset encoding network, and generate a temporal drift vector representing the evolution direction of the first process based on the trend gradient of the hidden layer state sequence within a preset time window. The mapping module is used to obtain the second feature vector of the second process that has a topological relationship with the first process in a preset knowledge graph, and to perform feature cross operation on the time drift vector and the second feature vector and then input it into a preset mapping network to output a dynamic mapping matrix. The vector alignment module is used to obtain the first feature vector of the first process in the preset knowledge graph, and perform spatial affine transformation processing on the first feature vector according to the dynamic mapping matrix to generate the aligned first reconstructed feature vector. The parameter generation module is used to perform aggregation update processing on the second feature vector based on the first reconstructed feature vector, and generate process control parameters for the second process based on the updated second feature vector. The update module is used to obtain feedback optimization parameters characterizing the accuracy of the second process control after executing the process control parameters, and update the preset time window accordingly.

[0007] Secondly, the knowledge graph push method for zirconium-titanium alloy fastener manufacturing process parameters includes the following steps: The absolute temperature sequence and instantaneous pressure sequence of the first process within a preset execution cycle are acquired simultaneously, and the time-series feature sequence of the first process within the preset execution cycle is generated. The hidden layer state sequence of the time-series feature sequence is extracted based on a preset coding network, and a time-series drift vector representing the evolution direction of the first process is generated based on the trend gradient of the hidden layer state sequence within a preset time window. A second feature vector of a second process that has a topological relationship with the first process is obtained from a preset knowledge graph. After performing a feature cross operation on the time-series drift vector and the second feature vector, the vector is input into a preset mapping network to output a dynamic mapping matrix. Obtain the first feature vector of the first process in the preset knowledge graph, and perform spatial affine transformation on the first feature vector according to the dynamic mapping matrix to generate the aligned first reconstructed feature vector. Based on the first reconstructed feature vector, the second feature vector is aggregated and updated, and the process control parameters for the second process are generated based on the updated second feature vector. After executing the process control parameters, feedback optimization parameters characterizing the accuracy of the second process control are obtained, and the preset time window is updated accordingly.

[0008] Thirdly, a computer storage medium stores computer-executable instructions, which, when executed, implement the knowledge graph push system for the zirconium-titanium alloy fastener manufacturing process parameters described in the first aspect.

[0009] Compared with the prior art, the beneficial effects of this application are: This application extracts the thermodynamic trajectory of zirconium-titanium alloys into a time-series drift vector, dynamically generates an affine transformation matrix, and achieves precise alignment of upstream node features to the downstream space before GNN message passing. This fundamentally breaks the rigid assumption of linear aggregation, decouples cascade amplification effects, and outputs physically self-consistent process parameters throughout the entire process, suppressing grain coarsening and fatigue failure. Furthermore, it introduces exponential moving average feedback and adaptive time window closed-loop to continuously optimize recommendation accuracy. The application uses case-based reasoning output parameters based on cosine similarity and weighted interpolation, meeting the requirements of high-security domains. All improvements are implemented at the software algorithm layer, requiring no additional hardware or significant increase in computing power, making it highly practical, low-cost, and valuable for widespread application. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the knowledge graph push system for the manufacturing process parameters of zirconium-titanium alloy fasteners in this application. Figure 2 This is a schematic diagram illustrating the steps of the knowledge graph push method for the manufacturing process parameters of zirconium-titanium alloy fasteners in this application. Detailed Implementation

[0011] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but only to illustrate selected embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item has been defined in one figure, it does not need to be further defined and explained in subsequent figures. The terms first, second, etc. are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0012] Zirconium-titanium alloys are widely used in fields with extremely demanding material performance requirements, such as aerospace structural fastening, nuclear reactor pressure pipe fastening, and high-end medical device fastening, due to their excellent corrosion resistance, high-temperature specific strength, and good biocompatibility. These high-end fasteners must withstand the combined effects of high-cycle alternating fatigue loads and complex corrosive media during service, thus requiring extremely stringent requirements on their microstructure (grain size, phase ratio, texture uniformity) and mechanical properties (yield strength, fatigue life, fracture toughness). The fabrication of zirconium-titanium alloy fasteners typically involves multiple consecutive processes, including forging, annealing, and machining. Complex physical coupling and transmission relationships exist between these processes: the execution parameters and process trajectory of upstream processes not only directly determine the material state of that pass but also profoundly influence the optimal process window required by downstream processes through mechanisms such as latent heat release of phase transformation and residual strain accumulation.

[0013] In the field of process parameter recommendation technology, intelligent recommendation schemes using knowledge graphs as structured carriers and graph neural networks as inference engines have become the mainstream technical approach. Such systems typically represent various manufacturing processes (e.g., forging process nodes, annealing process nodes) with nodes in the knowledge graph, and the physical transmission relationships between processes (e.g., forging → annealing process flow edges) with directed edges in the knowledge graph. Furthermore, graph convolutional networks are used to hierarchically aggregate and infer the feature vectors of each process node, ultimately decoding and generating the target process parameters for each process.

[0014] In the node feature aggregation stage of graph convolutional networks, existing technologies generally use linear summation or mean pooling to weight and merge the feature vectors of adjacent nodes. This type of linear aggregation operation mathematically implies a core assumption: the feature vectors of adjacent process nodes in the knowledge graph lie in the same linear mapping space. Therefore, direct linear combination operations can be performed on the feature vectors of upstream and downstream nodes without introducing mathematical errors.

[0015] However, under the conventional continuous batch production conditions of zirconium-titanium alloys, the aforementioned characteristic space isomorphism assumption has a fundamental flaw and is a typical linearization fallacy. The underlying physical reason is that zirconium-titanium alloys exhibit complex phase transformation behavior in an α+β two-phase region during hot working: within the conventional forging temperature range, the alloy exists in a two-phase coexistence region of α and β phases, and the thermodynamic state of the material is highly dependent on the spatiotemporal distribution and historical evolution trajectory of the temperature and stress fields, rather than the absolute temperature value at a single moment.

[0016] Specifically, the dynamic accumulation of temperature change rate and instantaneous strain rate in the upstream forging process within the preset execution cycle will have a strong nonlinear physical implicit transmission effect on the parameters such as the optimal soaking temperature and holding time required by the downstream annealing process through the non-uniform distribution of the two-phase interface energy and the gradient distribution of dislocation density.

[0017] This strong nonlinear cross-process physical coupling means that the key variable determining the initial state of the material in the downstream annealing process, and thus the optimal process window, is not the parameter setting value of the upstream forging process, but rather the dynamic evolution trajectory of temperature and strain rate during the actual execution of that process. When data flows across processes, the physical mapping space of the feature vectors of the upstream process nodes will undergo nonlinear distortion relative to the downstream nodes, i.e., temporal topological drift will occur, so that the feature vectors of the upstream and downstream nodes are actually in heterogeneous feature spaces, rather than the isomorphic Euclidean space assumed by existing technologies.

[0018] To address the aforementioned shortcomings, existing technologies mainly employ two types of strategies, but neither addresses the root cause of the problem: First, increasing the number of graph network layers in the hope that multi-hop information transmission will enable the model to implicitly learn nonlinear mappings. However, this essentially trades increased model complexity for limited accuracy improvement, and multiple stacked layers can lead to oversmoothing, causing node features to become homogeneous. Second, introducing an attention mechanism to differentiate the contributions of different neighboring nodes through learnable scalar attention weights. However, attention weights are essentially still scalar scaling of heterogeneous feature vectors, failing to change the fundamental nature of performing linear combinations on heterogeneous space vectors, and thus cannot eliminate aggregation errors caused by feature space distortion at the root.

[0019] The aforementioned defects cause the following problems in actual engineering production: When the system recommends multi-pass continuous process parameters, the feature mapping error across processes is amplified at each level of the graph. This results in the system continuously outputting a group of erroneous parameters that appear to be within the normal control range for individual process parameters, but after combined execution, microstructural defects such as abnormal grain growth and shortened fatigue crack initiation cycles are induced due to the mismatch of the internal physical state of the material. Ultimately, this leads to fastener failure during assembly or service, forming a common quality bottleneck in the industry where the yield of a single process is high but the scrap rate of the assembly is high.

[0020] In summary, there is an urgent need for a novel process parameter recommendation scheme that can break through the rigid assumption of linear aggregation in knowledge graph graph neural networks. Without increasing hardware sensors and computing power consumption, this scheme can achieve a qualitative leap in the process recommendation system, moving from single-process compliance to full-process physical self-consistency by sensing and correcting nonlinear distortions across the process feature space.

[0021] Therefore, such as Figure 1 As shown, this application provides a knowledge graph push system for the manufacturing process parameters of zirconium-titanium alloy fasteners, including the following modules: The data acquisition module is used to synchronously acquire the absolute temperature sequence and instantaneous pressure sequence of the first process within a preset execution cycle, and generate the time-series feature sequence of the first process within the preset execution cycle. The encoding module is used to extract the hidden layer state sequence of the temporal feature sequence based on a preset encoding network, and generate a temporal drift vector representing the evolution direction of the first process based on the trend gradient of the hidden layer state sequence within a preset time window. The mapping module is used to obtain the second feature vector of the second process that has a topological relationship with the first process in a preset knowledge graph, and to perform feature cross operation on the time drift vector and the second feature vector and then input it into a preset mapping network to output a dynamic mapping matrix. The vector alignment module is used to obtain the first feature vector of the first process in the preset knowledge graph, and perform spatial affine transformation processing on the first feature vector according to the dynamic mapping matrix to generate the aligned first reconstructed feature vector. The parameter generation module is used to perform aggregation update processing on the second feature vector based on the first reconstructed feature vector, and generate process control parameters for the second process based on the updated second feature vector. The update module is used to obtain feedback optimization parameters characterizing the accuracy of the second process control after executing the process control parameters, and update the preset time window accordingly.

[0022] In another implementation, such as Figure 2 As shown, this application also provides a method for pushing knowledge graphs of zirconium-titanium alloy fastener manufacturing process parameters, including the following steps: The absolute temperature sequence and instantaneous pressure sequence of the first process within a preset execution cycle are acquired simultaneously, and the time-series feature sequence of the first process within the preset execution cycle is generated. The hidden layer state sequence of the time-series feature sequence is extracted based on a preset coding network, and a time-series drift vector representing the evolution direction of the first process is generated based on the trend gradient of the hidden layer state sequence within a preset time window. A second feature vector of a second process that has a topological relationship with the first process is obtained from a preset knowledge graph. After performing a feature cross operation on the time-series drift vector and the second feature vector, the vector is input into a preset mapping network to output a dynamic mapping matrix. Obtain the first feature vector of the first process in the preset knowledge graph, and perform spatial affine transformation on the first feature vector according to the dynamic mapping matrix to generate the aligned first reconstructed feature vector. Based on the first reconstructed feature vector, the second feature vector is aggregated and updated, and the process control parameters for the second process are generated based on the updated second feature vector. After executing the process control parameters, feedback optimization parameters characterizing the accuracy of the second process control are obtained, and the preset time window is updated accordingly.

[0023] I. Construction of temporal feature sequences; During the data acquisition phase, the system synchronously acquires the absolute temperature sequence and instantaneous pressure sequence of the first process within a preset execution cycle through the Supervisory Control and Data Acquisition (SCADA) system. The absolute temperature sequence records the actual temperature of the first process (forging process) at each sampling moment, using absolute temperature (Kelvin, K) as the unit of measurement rather than Celsius. This is to ensure strict dimensional consistency between subsequent derivative calculations involving thermodynamic dimensional conversion and thermodynamic formulas, avoiding systematic deviations introduced by dimensional conversion.

[0024] The instantaneous pressure sequence records the instantaneous contact pressure actually applied to the workpiece by the forging equipment at each sampling moment, in megapascals (MPa). Its physical meaning is the distribution of the equipment output (force) on the contact surface. Compared with the average load or rated load, the instantaneous pressure sequence can more accurately reflect the instantaneous transfer efficiency of mechanical energy to the workpiece during the forging process, laying the physical basis for subsequent calculation of strain rate.

[0025] In the first step of generating the time-series feature sequence, the system calculates the instantaneous stress sequence of the first process within a preset execution cycle based on the instantaneous pressure sequence and the preset die cross-sectional area: at each sampling time step, the instantaneous pressure value at that moment is divided by the pre-calibrated and recorded effective contact cross-sectional area of ​​the forging die (unit: m²). 2 This allows us to obtain the nominal instantaneous stress value (in MPa) borne by the workpiece at that moment, thereby converting the output (pressure) at the equipment end into the stress borne by the material end, achieving a unified process semantics of physical dimensions.

[0026] After obtaining the instantaneous stress sequence, the numerical difference between adjacent time steps in the instantaneous stress sequence is calculated. This is achieved by subtracting the instantaneous stress value of the immediately preceding time step from the instantaneous stress value of the current time step and dividing by the sampling time interval, thus generating the instantaneous strain rate sequence for the first process. Instantaneous strain rate is a core physical parameter measuring the deformation rate of a material. Its magnitude directly determines the accumulation rate and distribution pattern of dislocation density within the zirconium-titanium alloy, thereby affecting the degree of residual strain energy accumulation at the grain interface. It is one of the core driving variables that exert a physical implicit transmission effect on the annealing process during the forging process.

[0027] In the second step of generating the time-series feature sequence, the absolute temperature values ​​in the absolute temperature sequence and the instantaneous strain rate values ​​in the instantaneous strain rate sequence at the same time step are concatenated: the absolute temperature scalar and the instantaneous strain rate scalar of that time step are linked together to form a two-dimensional column vector, forming the concatenated feature for that time step. This concatenation operation is repeated for all time steps, ultimately generating a time-series feature sequence composed of concatenated features from multiple time steps arranged in chronological order.

[0028] This time-series feature sequence forms a matrix with two time steps, where each row carries both temperature information (characterizing thermodynamic state) and strain rate information (characterizing deformation kinetic state) at that moment. This constitutes a complete information carrier for a two-dimensional quantitative description of the thermodynamic evolution state of zirconium-titanium alloy, providing sufficient input data for the subsequent sequence feature extraction of the encoding module.

[0029] At the engineering implementation level, the system also performs necessary preprocessing on the raw sensor data: Kalman filtering is used to smooth and denoise high-frequency random noise in real time to eliminate non-physical high-frequency disturbances introduced by sensor electrical interference or mechanical vibration; spline interpolation algorithm is used to uniformly resample non-uniform time step data caused by non-constant device sampling frequency to a fixed time step to ensure the time equidistance between rows of time series feature sequences, thereby meeting the input requirements of the subsequent long short-term memory network for equidistant time series data.

[0030] II. Extraction of hidden layer state sequences and generation of temporal drift vectors; In extracting the hidden layer state sequence, the preset encoding network is a Long Short-Term Memory (LSTM) network with a forget gate mechanism. The temporal feature sequence is input into the preset encoding network, which processes each concatenated feature vector in the sequence sequentially according to the time step order, and outputs the hidden layer state vector for that time step accordingly, thereby gradually generating a hidden layer state sequence covering all time steps.

[0031] The reason for choosing a Long Short-Term Memory network with a forget gate mechanism instead of a regular Recurrent Neural Network (RNN) is that its forget gate can automatically learn and adjust the degree of forgetting of early information based on the current input and historical state during sequence processing. This allows it to selectively retain information on key physical events in the middle stage (such as the crossing of phase transition temperature ranges) that are most decisive for the final state, effectively alleviating the gradient vanishing problem of regular RNNs when dealing with long sequences. This enables the state vector of the terminal hidden layer to more accurately and robustly reflect the comprehensive trajectory of thermodynamic evolution throughout the entire forging execution cycle.

[0032] In generating the time-series drift vector, the final state vector corresponding to the last time step in the hidden layer state sequence is extracted, as well as the initial state vector corresponding to the starting time step defined by a preset time window. The preset time window, in units of time steps k, determines the time span used for trend gradient calculation: the starting time step is the position corresponding to k time steps backward from the final time step, and the hidden layer state vector at this position is the starting state vector.

[0033] The vector difference between the terminal state vector and the initial state vector is obtained as its trend gradient: subtracting the initial state vector from the terminal state vector yields a difference vector that points in the hidden state space to the evolution direction of the first process from the start of the preset time window to the end of execution, directly representing the directional change of the latent heat release rate of phase change and the trend of residual strain accumulation in the feature space. To eliminate the dimensional interference of the absolute magnitude of the trend gradient on subsequent matrix generation operations, only the trend direction information is retained. The trend gradient is normalized using the second norm and mapped onto a unit hypersphere, outputting a time-series drift vector. The mathematical expression of the above calculation relationship is as follows: , ; in, Indicates the trend gradient. This represents the output time drift vector. Indicates the final time step The corresponding terminal state vector, Indicates the length of the preset time window (unit: time steps). Indicates the starting time step defined by the preset time window ( The corresponding initial state vector, The second norm (L2 norm) of the trend gradient is the arithmetic square root of the sum of the squares of the components of the trend gradient vector.

[0034] The physical meaning of the temporal drift vector lies in the fact that it encodes the direction of the recent thermodynamic evolution trajectory of the first process in the LSTM hidden state space as a unit vector. Essentially, it is a standardized vector quantization of the physical evolution trend at the end of the upstream process execution. This vector will be used in the mapping module to perform a tensor outer product operation with the second feature vector of the downstream process node, generating a feature cross tensor that fully captures the physical coupling relationship across processes, thereby driving the generation of the dynamic mapping matrix.

[0035] III. Feature Cross Operation and Dynamic Mapping Matrix Generation; In determining the second process and obtaining the second feature vector, the system first uses the node identifier of the first node corresponding to the first process in the preset knowledge graph, combined with the graph adjacency matrix stored in the preset knowledge graph, to retrieve all downstream nodes that have preset process flow relationship edges with the first node, and uses these as the second node corresponding to the second process. In the graph structure of the knowledge graph, process flow relationship edges represent the physical action path of the upstream process on the downstream process in the form of directed edges (such as the directed edge of the forging process node → annealing process node). By traversing the outgoing adjacency list of the first node, the downstream process node that immediately follows the first process in the process flow, i.e., the second node, can be accurately located.

[0036] Subsequently, based on the node identifier of the second node, a matrix row indexing operation is performed from the node feature matrix of the preset knowledge graph to extract the corresponding row vector as its second feature vector. This second feature vector is a d-dimensional real-valued vector, and its components carry the standard process baseline information of the second process recorded in the knowledge graph in the form of implicit encoding, such as the semantic embedding of parameters like the rated homogenization temperature range, holding time, and heating rate of the annealing process.

[0037] In performing the feature cross operation, the tensor outer product of the temporal drift vector and the second eigenvector is obtained as the feature cross tensor. The tensor outer product operation combines the d-dimensional temporal drift vector and the d-dimensional second eigenvector to generate a d×d matrix, i.e., the feature cross tensor, where the element value in the i-th row and j-th column is equal to the product of the i-th component of the temporal drift vector and the j-th component of the second eigenvector. Each element of this matrix explicitly encodes the pairwise cross relationship between the upstream process trend component i and the downstream baseline feature component j, so that the nonlinear perturbation effect of the upstream thermodynamic evolution trajectory on the downstream process parameter space of various dimensions can be explicitly captured in the matrix in a high order, overcoming the inherent limitation of low-order fusion methods such as simple splicing or scalar attention weighting, which can only capture linear correlations.

[0038] In generating the dynamic mapping matrix, the preset mapping network consists of a Multi-Layer Perceptron (MLP). The feature cross tensor is flattened into a single matrix in row-major order. A one-dimensional vector is input into a pre-defined mapping network for nonlinear mapping. The hidden layers of a multilayer perceptron further perform nonlinear dimensionality upscaling and feature extraction on the feature cross tensor through a combination of learnable weight matrices and a nonlinear activation function (the first pre-defined activation function is the sigmoid activation function). The network ultimately outputs a... A one-dimensional vector is reshaped into a square matrix with a shape of d×d, thus obtaining the dynamic mapping matrix. The mathematical expression of the above generation process is as follows: ; in, This represents the output dynamic mapping matrix, which has a dimension of d×d; This represents the first preset activation function (Sigmoid function), which constrains the values ​​of each element in the matrix to the interval (0,1) to ensure the numerical stability of the affine transformation; and These are the learnable weight matrix and bias matrix in the preset mapping network, which are optimized through end-to-end gradient backpropagation during the offline training phase of the system. Represents the time drift vector; This represents the second feature vector of the second process; This represents the tensor outer product operation.

[0039] The core value of the dynamic mapping matrix lies in its dynamism and batch-specificity: it is not a globally static weight matrix shared by all batches, but a dynamic transfer operator that evolves in real time in tandem with the specific execution trajectory of the upstream process (carried by the time-series drift vector) and the current baseline state of the downstream process (carried by the second eigenvector) for each batch of material. When different batches of material have different upstream forging execution trajectories due to fluctuations in initial composition, differences in equipment status, etc., the mapping module will generate different dynamic mapping matrices accordingly, thereby providing the vector alignment module with batch-specific nonlinear transformation parameters, achieving batch-level precise adaptation of process parameter recommendations.

[0040] IV. Cross-process feature pre-alignment based on spatial affine transformation; In obtaining the first feature vector, based on the node identifier of the first node, a matrix row indexing operation is performed from the node feature matrix of the preset knowledge graph to extract the corresponding row vector as its first feature vector. This first feature vector is a d-dimensional real-valued vector, and each component carries the standard process parameter encoding information (such as the semantic embedding of parameters like rated forging temperature, forging ratio, and deformation amount) recorded in the knowledge graph for the first process (forging process). It is the original node feature that needs to be transmitted from the first node (i.e., the upstream node) to the second node (i.e., the downstream node) during the message passing stage of the graph neural network.

[0041] In performing spatial affine transformation processing, the dynamic mapping matrix and the first eigenvector are multiplied by matrix multiplication, that is, the d×d dimensional dynamic mapping matrix is ​​multiplied by the d-dimensional first eigenvector to obtain a d-dimensional output vector, called the scaling-rotation vector. Subsequently, a preset dynamic bias vector is superimposed (element-wise added) onto the scaling-rotation vector to generate the aligned first reconstructed eigenvector. The mathematical expression of the entire transformation process is as follows: ; in, This represents the first reconstructed feature vector after alignment; This represents the first feature vector of the first process; Represents a dynamic mapping matrix; This represents the preset dynamic bias vector, which is a d-dimensional vector, and is mapped to the weight matrix of the preset mapping network during the offline training phase of the system. and bias matrix They are jointly optimized through end-to-end gradient backpropagation.

[0042] From a mathematical perspective, the above spatial affine transformation is composed of two geometric operations: matrix multiplication. A linear transformation is performed on the first eigenvector to achieve vector rotation (changing the direction of the eigenvector) and scaling (changing the relative magnitudes of the components of the eigenvector) in high-dimensional space, transforming the first process eigenvector from its original eigencoordinate basis to a coordinate basis corresponding to the second process eigenspace; bias vector terms. Then, a translation compensation is applied to the transformed vector to eliminate the origin offset between the two feature spaces. Together, they constitute a complete affine transformation, forming a mathematical isomorphic mapping with the objective law in materials science that the physical property coordinate basis of zirconium-titanium alloys undergoes nonlinear distortion under residual strain fields. This gives the alignment operation a solid material physics basis.

[0043] Compared to existing technologies, the core breakthrough of this vector alignment module lies in the introduction of a feature pre-alignment step before aggregation. Traditional GCN directly aligns the upstream feature vectors during the message passing phase. The feature is passed and linearly combined with the features of downstream nodes, implicitly assuming... It lies in the same linear feature space as the downstream node. This module uses a dynamic mapping matrix to... Transform into reconstructed feature vectors This mathematically aligns the vectors in the downstream feature space, ensuring that subsequent aggregation operations are effective linear combinations of vectors within the same linear space. This fundamentally eliminates the error cascading divergence problem caused by the forced linear superposition of heterogeneous spaces.

[0044] V. Multi-neighbor aggregation update and process control parameter output; In performing the aggregation update process, the feature vectors corresponding to all neighboring nodes that have a directional connection to the second node (i.e., there is a directed edge pointing to the second node) are first obtained from the preset knowledge graph to form a feature set of neighboring nodes. It should be noted that the neighboring nodes are not limited to the first node (forging process node). In actual engineering scenarios, the factors affecting the optimal process parameters of the second process (annealing process) are multi-dimensional; therefore, neighboring nodes may also include: The upstream nodes connected to the second node via directed edges include raw material attribute nodes (nodes carrying encoded information such as the chemical composition ratios of sponge titanium and sponge zirconium), environmental constraint nodes (nodes carrying information such as the daily workshop temperature, humidity, or atmospheric pressure), and equipment status nodes (nodes carrying information such as the maintenance cycle and current furnace condition of the vacuum annealing furnace). The adjacent node feature set includes the first reconstructed feature vector and the original feature vectors of other adjacent nodes, enabling the aggregation update operation to simultaneously perceive the combined effects of the preparation process, raw materials, environment, and equipment multiphysics.

[0045] A weighted summation operation is performed on each feature vector in the feature set of adjacent nodes to generate a comprehensive feature vector. Since the first reconstructed feature vector has already been precisely affine aligned to the second process feature space in the vector alignment module, and the feature vectors of other adjacent nodes (such as raw material nodes and equipment nodes) are encoded in the same feature space as the second node (according to the unified embedding specification during knowledge graph construction), the weighted summation of all adjacent node feature vectors at this time is mathematically an effective linear combination of vectors in the same linear space, with strict geometric validity, and no longer introduces the error caused by the forced linear superposition of heterogeneous spaces.

[0046] The comprehensive feature vector is concatenated with the second feature vector to generate a multiphysics joint vector: the comprehensive feature vector (a vector that integrates aggregated information from adjacent processes, raw materials, equipment, etc.) and the second feature vector (the baseline feature vector of the second process node in the current graph state) are concatenated element-wise into a longer vector with a higher dimension, called the multiphysics joint vector. This concatenation operation enables the subsequent network to simultaneously perceive the combined influence of external inputs from neighbors and the joint information of the second process node's own baseline state, laying the information foundation for accurately calculating the cross-coupling terms between feature vectors.

[0047] The multiphysics joint vector is multiplied by a preset update weight matrix to extract the cross-coupling terms between feature vectors: the preset update weight matrix is ​​a learnable rectangular matrix, which is multiplied by the multiphysics joint vector on the left, so that there is an explicit algebraic relationship between the feature components corresponding to different physical quantities in the joint vector (such as the cross-coupling terms between the temperature evolution features from the forging node and the furnace condition features from the equipment node), and outputs a vector of dimension d; then, a nonlinear mapping is performed through a second preset activation function (such as the ReLU function) to filter out insignificant small fluctuations and amplify key feature combinations, and finally outputs the updated second feature vector, completing the aggregation and update of the feature vectors of the annealing process nodes.

[0048] In generating process control parameters, the system employs a cosine similarity retrieval and weighted interpolation method based on historical case reasoning to fully balance the interpretability and engineering stability of parameter recommendations. First, the cosine similarity between the updated second feature vector and each historical feature vector in the preset historical process feature library is obtained. The historical process feature library pre-stores a large number of node feature vectors corresponding to historical production batches and their corresponding final executed process parameters (historical process execution parameters, such as annealing temperature of 780℃, holding time of 90 minutes, etc.). Cosine similarity measures the directional consistency between the currently updated feature vector and each historical feature vector in high-dimensional space, with a value range of [-1, 1]. The closer the value is to 1, the more similar the current process state is to the material state of the corresponding historical batch.

[0049] Subsequently, historical process execution parameters corresponding to historical feature vectors with cosine similarity greater than a preset matching threshold are extracted to form a candidate historical parameter set. Weighted interpolation is then performed on each extracted historical process execution parameter based on its corresponding cosine similarity to output the process control parameters. The specific implementation of weighted interpolation is as follows: the cosine similarity of candidate historical batches is converted into normalized interpolation weight coefficients using a normalized exponential function (Softmax function) (the sum of each weight is 1, and historical cases with higher similarity are assigned larger weights). Each historical process execution parameter is then multiplied by its corresponding interpolation weight coefficient and summed to obtain the final process control parameters.

[0050] The process control parameters output by this method are not copies of any single historical case, but new parameters obtained by interpolation between similar historical experience endpoints according to the similarity ratio. This retains the engineering experience of successful historical cases and can make reasonable adaptive adjustments according to the current batch process status. At the same time, it has good output continuity - when the current batch status changes slightly, the similarity distribution of each historical case changes only slightly, and the output process control parameters are also adjusted only slightly. This avoids the problem of sudden jumps in control parameters caused by slight changes in the ranking of the historical database under the strategy of only taking the most similar case, and significantly improves the stability of process parameter output under normal batch production conditions.

[0051] VI. Feedback optimization of parameter calculation and adaptive updating of preset time windows; To obtain actual mechanical performance values, after executing the process control parameters and at the end of the current preset execution cycle, mechanical performance tests are performed on the batch of fasteners produced according to the process control parameters to obtain actual mechanical performance values ​​characterizing the structural strength of the fasteners prepared within the current preset execution cycle. These actual mechanical performance values ​​can be mechanical testing indicators that quantify the structural strength of the fasteners, such as batch average tensile strength (MPa), minimum elongation (%), and fatigue fracture cycles (times). The specific types of indicators are predetermined based on product quality inspection standards during system engineering deployment. Subsequently, the absolute error between the actual mechanical performance values ​​and the preset mechanical benchmark values ​​is obtained. The absolute error is the absolute value of the difference between the current batch's actual mechanical performance values ​​and the preset mechanical benchmark values ​​(determined by product design specifications). This absolute error quantifies the degree of deviation of the current batch's recommended process control parameters from the final mechanical performance dimension and is the most direct engineering indicator for measuring the accuracy of the recommendation.

[0052] In calculating feedback optimization parameters, to avoid overreacting to system parameters due to occasional fluctuations in mechanical properties of a single batch (such as minor differences in raw material composition, instrument errors, and other random factors), an exponential moving average algorithm is used to weight and fuse the absolute error of the current batch with historical smoothing errors to generate feedback optimization parameters for the current preset execution cycle. The exponential moving average algorithm uses a preset smoothing coefficient. (Values ​​range from 0 to 1) This serves as a memory weight for historical errors, ensuring that the feedback optimization parameters smoothly reflect the long-term trend of errors while also assigning relatively high weight to error information from recent batches (weight is...). This achieves a reasonable balance between smoothness and response sensitivity. The calculation relationship of the feedback optimization parameters within the current preset execution cycle is as follows: ; in, This represents the feedback optimization parameters within the current preset execution period t; A larger value indicates a stronger ability of the system to remember historical errors and a lower sensitivity to recent fluctuations; This represents the feedback optimization parameters within the previous preset execution cycle t-1, i.e., the historical smoothing error; This represents the actual mechanical properties of the fasteners produced within the current preset execution cycle t; This indicates the preset mechanical reference value.

[0053] Regarding updating the preset time window, when the feedback optimization parameters within the current preset execution cycle are greater than the preset error threshold, it indicates that the current preset time window length configured by the system is insufficient to enable the trend gradient to accurately capture the key thermodynamic evolution information at the end of the upstream process execution. The time window needs to be extended accordingly to improve the depth of trend perception.

[0054] Specifically, the implementation involves first calculating the deviation ratio of the feedback optimization parameters from the preset error threshold, i.e. This ratio quantitatively characterizes the extent to which the current systematic error exceeds the acceptable range; subsequently, this deviation ratio is used in conjunction with a preset sensitivity coefficient. The product of the two factors is used to calculate the scaling factor for the current time window; with the current preset time window... Multiply by the scaling factor to obtain the updated candidate value, round the candidate value down, and apply the preset time window upper and lower limits. and Extreme value truncation is performed, and the updated preset time window is finally output. This is then applied to the next preset execution cycle. The complete calculation relationship for the above time window update is as follows: ; in, This indicates the updated preset time window; This indicates the preset time window used within the current preset execution cycle; and These represent the minimum and maximum preset time windows, respectively. Together, they constrain the range of time window values ​​within a reasonable physical range, preventing the time window from degenerating to zero (making the trend gradient meaningless) or expanding to too large (introducing early information unrelated to the current batch). This represents the preset sensitivity coefficient, which controls the step size of the time window response error deviation. This represents the preset error threshold, which is the dividing line for judging whether the current recommendation accuracy meets the requirements; Indicates to Perform the floor operation.

[0055] The adaptive physical logic of the above time window update mechanism is as follows: when the system error continues to exceed the preset threshold ( When the deviation ratio is positive and the scaling multiplier is greater than 1, the time window will expand to a larger extent, allowing the encoding module to review the hidden layer states at earlier times in the trend gradient calculation, capturing a deeper thermodynamic evolution hysteresis effect, thereby improving the accuracy of the temporal drift vector in representing the upstream physical state. When the system error is consistently below a preset threshold, the system maintains the current window size without adjustment to avoid unnecessary parameter perturbations. The upper and lower limits of the time window are truncated as a safety protection measure to ensure the numerical stability of the system under extreme error conditions.

[0056] In another embodiment, this application also provides a computer storage medium storing computer-executable instructions, which, when executed, implement the knowledge graph push system for the manufacturing process parameters of the zirconium-titanium alloy fastener.

[0057] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of this application without departing from the spirit and scope of the technical methods of this application.

Claims

1. A knowledge graph pushing system for zirconium-titanium alloy fastener manufacturing process parameters, characterized in that, Includes the following modules: The data acquisition module is used to synchronously acquire the absolute temperature sequence and instantaneous pressure sequence of the first process within a preset execution cycle, and generate the time-series feature sequence of the first process within the preset execution cycle. The encoding module is used to extract the hidden layer state sequence of the temporal feature sequence based on a preset encoding network, and generate a temporal drift vector representing the evolution direction of the first process based on the trend gradient of the hidden layer state sequence within a preset time window. The mapping module is used to obtain the second feature vector of the second process that has a topological relationship with the first process in a preset knowledge graph, and to perform feature cross operation on the time drift vector and the second feature vector and then input it into a preset mapping network to output a dynamic mapping matrix. The vector alignment module is used to obtain the first feature vector of the first process in the preset knowledge graph, and perform spatial affine transformation processing on the first feature vector according to the dynamic mapping matrix to generate the aligned first reconstructed feature vector. The parameter generation module is used to perform aggregation update processing on the second feature vector based on the first reconstructed feature vector, and generate process control parameters for the second process based on the updated second feature vector. The update module is used to obtain feedback optimization parameters characterizing the accuracy of the second process control after executing the process control parameters, and update the preset time window accordingly.

2. The knowledge graph pushing system of process parameters for zirconium-titanium alloy fastener according to claim 1, characterized in that, The process of generating time-series feature sequences includes: The instantaneous stress sequence of the first process within the preset execution cycle is calculated based on the instantaneous pressure sequence and the preset mold cross-sectional area, and the numerical difference between adjacent time steps in the instantaneous stress sequence is calculated to generate the instantaneous strain rate sequence of the first process. The absolute temperature values ​​in the absolute temperature sequence and the instantaneous strain rate values ​​in the instantaneous strain rate sequence at the same time step are spliced ​​together to generate a time-series feature sequence composed of spliced ​​features from multiple time steps.

3. The knowledge graph pushing system of process parameters for zirconium-titanium alloy fastener according to claim 2, characterized in that, The process of generating the time drift vector includes: The preset encoding network is a long short-term memory network that includes a forget gate mechanism. The temporal feature sequence is input into the preset encoding network, and the hidden layer state vector is output according to the time step order to generate the hidden layer state sequence. extracting an end state vector corresponding to an end time step of the hidden layer state sequence and a start state vector corresponding to a start time step defined by a preset time window, and obtaining a vector difference between the end state vector and the start state vector as a trend gradient thereof performing second norm normalization processing on the trend gradient to output a timing drift vector ; ， ； wherein, denotes the end state vector, denotes a length of a preset time window, denotes the start state vector, denotes a second norm of the trend gradient.

4. The knowledge graph pushing system of process parameters for zirconium-titanium alloy fastener according to claim 1, characterized in that, The process of outputting the dynamic mapping matrix includes: Based on the node identifier and graph adjacency matrix of the first node corresponding to the first process in the preset knowledge graph, the downstream node with the preset process flow relationship edge is obtained as the second node corresponding to the second process. Based on the node identifier of the second node, the corresponding feature vector is extracted from the node feature matrix of the preset knowledge graph as its second feature vector, and the tensor cross product of the temporal drift vector and the second feature vector is obtained as the feature cross tensor. The preset mapping network is composed of a multi-layer perceptron, and the feature cross tensor is flattened and input into the preset mapping network for mapping, and a one-dimensional vector output by the mapping is reshaped into a square matrix form to generate a dynamic mapping matrix ; wherein, represents a first preset activation function, and are a weight matrix and a bias matrix in a preset mapping network, respectively, represents the timing drift vector, represents a second feature vector of a second process, represents a tensor outer product operation.

5. The knowledge graph pushing system of process parameters for zirconium-titanium alloy fastener according to claim 4, characterized in that, The process of generating the first reconstructed feature vector includes: Based on the node identifier of the first node, the corresponding feature vector is extracted from the node feature matrix of the preset knowledge graph as its first feature vector. The dynamic mapping matrix is ​​multiplied with the first eigenvector to output a scaling and rotation vector. A preset dynamic bias vector is then superimposed on the scaling and rotation vector to generate an aligned first reconstructed eigenvector. ; in, This represents the first feature vector of the first process. This represents the preset dynamic bias vector. This represents the dynamic mapping matrix.

6. The knowledge graph push system for the manufacturing process parameters of zirconium-titanium alloy fasteners according to claim 5, characterized in that, The process of updating the second feature vector includes: In a preset knowledge graph, feature vectors corresponding to all neighboring nodes that have a directional connection with the second node are obtained to form a neighboring node feature set, the neighboring node feature set including the first reconstructed feature vector; A weighted summation operation is performed on each feature vector in the feature set of adjacent nodes to generate a comprehensive feature vector, and the comprehensive feature vector is concatenated with the second feature vector to generate a multiphysics joint vector. The multiphysics joint vector is multiplied by a preset update weight matrix to extract the cross-coupling terms between the feature vectors, and the updated second feature vector is output by mapping through a second preset activation function.

7. The knowledge graph push system for the manufacturing process parameters of zirconium-titanium alloy fasteners according to claim 6, characterized in that, The process of generating process control parameters includes: Obtain the cosine similarity between the updated second feature vector and each historical feature vector in the preset historical process feature library; Historical process execution parameters corresponding to historical feature vectors with cosine similarity greater than a preset matching threshold are extracted, and weighted interpolation calculations are performed on each extracted historical process execution parameter based on the corresponding cosine similarity to output process control parameters.

8. The knowledge graph push system for the manufacturing process parameters of zirconium-titanium alloy fasteners according to claim 7, characterized in that, The process of updating the preset time window includes: After executing the process control parameters, the actual mechanical property values ​​characterizing the strength of the fastener structure prepared within the current preset execution cycle are obtained. The actual mechanical performance values ​​and the preset mechanical reference values ​​are obtained. The absolute error between them; Obtain feedback optimization parameters within the current preset execution cycle. ,in, For the preset smoothing coefficient, These are the feedback optimization parameters for the previous preset execution cycle; When the feedback optimization parameter within the current preset execution cycle is greater than the preset error threshold, the preset time window within the current preset execution cycle is... Perform an update to obtain the updated preset time window. And apply it to the next preset execution cycle; ； in, and These are the minimum and maximum preset time windows, respectively. This represents the preset sensitivity coefficient. This indicates the preset error threshold. Indicates to Round down to the nearest integer.

9. A method for pushing knowledge graphs of manufacturing process parameters for zirconium-titanium alloy fasteners, characterized in that, Includes the following steps: The absolute temperature sequence and instantaneous pressure sequence of the first process within a preset execution cycle are acquired simultaneously, and the time-series feature sequence of the first process within the preset execution cycle is generated. The hidden layer state sequence of the time-series feature sequence is extracted based on a preset coding network, and a time-series drift vector representing the evolution direction of the first process is generated based on the trend gradient of the hidden layer state sequence within a preset time window. A second feature vector of a second process that has a topological relationship with the first process is obtained from a preset knowledge graph. After performing a feature cross operation on the time-series drift vector and the second feature vector, the vector is input into a preset mapping network to output a dynamic mapping matrix. Obtain the first feature vector of the first process in the preset knowledge graph, and perform spatial affine transformation on the first feature vector according to the dynamic mapping matrix to generate the aligned first reconstructed feature vector. Based on the first reconstructed feature vector, the second feature vector is aggregated and updated, and the process control parameters for the second process are generated based on the updated second feature vector. After executing the process control parameters, feedback optimization parameters characterizing the accuracy of the second process control are obtained, and the preset time window is updated accordingly.

10. A computer storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, a knowledge graph push system for the manufacturing process parameters of zirconium-titanium alloy fasteners as described in any one of claims 1-8 is implemented.

Images